Skip to contentSkip to conversation

IOASID extensions for guest SVA

PATCH Superseded Jacob Pan 8 participants 56 messages Aug 22, 2020 – Sep 24, 2020 Page 1 of 2

Subsystems (3)

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore

Jacob Pan

author
#0 lore raw

v2 cover IOASID extensions for guest SVA

IOASID was introduced in v5.5 as a generic kernel allocator service for both PCIe Process Address Space ID (PASID) and ARM SMMU's Sub Stream ID. In addition to basic ID allocation, ioasid_set was defined as a token that is shared by a group of IOASIDs. This set token can be used for permission checking, but lack of some features to address the following needs by guest Shared Virtual Address (SVA).
- Manage IOASIDs by group, group ownership, quota, etc.
- State synchronization among IOASID users
- Non-identity guest-host IOASID mapping
- Lifecycle management across many users

This patchset introduces the following extensions as solutions to the problems above.
- Redefine and extend IOASID set such that IOASIDs can be managed by groups.
- Add notifications for IOASID state synchronization
- Add reference counting for life cycle alignment among users
- Support ioasid_set private IDs, which can be used as guest IOASIDs Please refer to Documentation/ioasid.rst in enclosed patch 1/9 for more details.

This patchset only included VT-d driver as users of some of the new APIs. VFIO and KVM patches are coming up to fully utilize the APIs introduced here.

You can find this series at:
https://github.com/jacobpan/linux.git ioasid_ext_v2
(VFIO and KVM patches will be available at this branch when published.)

This work is a result of collaboration with many people:
Liu, Yi L <[email protected]>
Wu Hao <[email protected]>
Ashok Raj <[email protected]>
Kevin Tian <[email protected]>

Thanks,

Jacob

Changelog

V2:
- Redesigned ioasid_set APIs, removed set ID
- Added set private ID (SPID) for guest PASID usage.
- Add per ioasid_set notification and priority support.
- Back to use spinlocks and atomic notifications.
- Added async work in VT-d driver to perform teardown outside atomic context

Jacob Pan (9):
docs: Document IO Address Space ID (IOASID) APIs iommu/ioasid: Rename ioasid_set_data() iommu/ioasid: Introduce ioasid_set APIs iommu/ioasid: Add reference couting functions iommu/ioasid: Introduce ioasid_set private ID iommu/ioasid: Introduce notification APIs iommu/vt-d: Listen to IOASID notifications iommu/vt-d: Send IOASID bind/unbind notifications iommu/vt-d: Store guest PASID during bind

 Documentation/ioasid.rst    | 618 ++++++++++++++++++++++++++++++++
 drivers/iommu/intel/iommu.c |  27 +-
 drivers/iommu/intel/pasid.h |   1 +
 drivers/iommu/intel/svm.c   |  97 ++++-
 drivers/iommu/ioasid.c      | 835 ++++++++++++++++++++++++++++++++++++++++++--
 include/linux/intel-iommu.h |   2 +
 include/linux/ioasid.h      | 166 ++++++++-
 7 files changed, 1699 insertions(+), 47 deletions(-)
 create mode 100644 Documentation/ioasid.rst
Signature · 1 line
2.7.4
author
#1 lore raw

PATCH v2 1/9 docs: Document IO Address Space ID (IOASID) APIs

IOASID is used to identify address spaces that can be targeted by device DMA. It is a system-wide resource that is essential to its many users. This document is an attempt to help developers from all vendors navigate the APIs. At this time, ARM SMMU and Intel’s Scalable IO Virtualization (SIOV) enabled platforms are the primary users of IOASID. Examples of how SIOV components interact with IOASID APIs are provided in that many APIs are driven by the requirements from SIOV.

---
 Documentation/ioasid.rst | 618 +++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 618 insertions(+)
 create mode 100644 Documentation/ioasid.rst
Documentation/ioasid.rst new file +618 -0
new file mode 100644
index 000000000000..b6a8cdc885ff
@@ -0,0 +1,618 @@
+.. ioasid:
+
+=====================================
+IO Address Space ID
+=====================================
+
+IOASID is a generic name for PCIe Process Address ID (PASID) or ARM
+SMMU sub-stream ID. An IOASID identifies an address space that DMA
+requests can target.
+
+The primary use cases for IOASID are Shared Virtual Address (SVA) and
+IO Virtual Address (IOVA). However, the requirements for IOASID
+management can vary among hardware architectures.
+
+This document covers the generic features supported by IOASID
+APIs. Vendor-specific use cases are also illustrated with Intel's VT-d
+based platforms as the first example.
+
+.. contents:: :local:
+
+Glossary
+========
+PASID - Process Address Space ID
+
+IOASID - IO Address Space ID (generic term for PCIe PASID and
+sub-stream ID in SMMU)
+
+SVA/SVM - Shared Virtual Addressing/Memory
+
+ENQCMD - New Intel X86 ISA for efficient workqueue submission [1]
+
+DSA - Intel Data Streaming Accelerator [2]
+
+VDCM - Virtual device composition module [3]
+
+SIOV - Intel Scalable IO Virtualization
+
+
+Key Concepts
+============
+
+IOASID Set
+-----------
+An IOASID set is a group of IOASIDs allocated from the system-wide
+IOASID pool. An IOASID set is created and can be identified by a
+token of u64. Refer to IOASID set APIs for more details.
+
+IOASID set is particularly useful for guest SVA where each guest could
+have its own IOASID set for security and efficiency reasons.
+
+IOASID Set Private ID (SPID)
+----------------------------
+SPIDs are introduced as IOASIDs within its set. Each SPID maps to a
+system-wide IOASID but the namespace of SPID is within its IOASID
+set. SPIDs can be used as guest IOASIDs where each guest could do
+IOASID allocation from its own pool and map them to host physical
+IOASIDs. SPIDs are particularly useful for supporting live migration
+where decoupling guest and host physical resources are necessary.
+
+For example, two VMs can both allocate guest PASID/SPID #101 but map to
+different host PASIDs #201 and #202 respectively as shown in the
+diagram below.
+::
+
+ .------------------.    .------------------.
+ |   VM 1           |    |   VM 2           |
+ |                  |    |                  |
+ |------------------|    |------------------|
+ | GPASID/SPID 101  |    | GPASID/SPID 101  |
+ '------------------'    -------------------'     Guest
+ __________|______________________|______________________
+           |                      |               Host
+           v                      v
+ .------------------.    .------------------.
+ | Host IOASID 201  |    | Host IOASID 202  |
+ '------------------'    '------------------'
+ |   IOASID set 1   |    |   IOASID set 2   |
+ '------------------'    '------------------'
+
+Guest PASID is treated as IOASID set private ID (SPID) within an
+IOASID set, mappings between guest and host IOASIDs are stored in the
+set for inquiry.
+
+IOASID APIs
+===========
+To get the IOASID APIs, users must #include <linux/ioasid.h>. These APIs
+serve the following functionalities:
+
+  - IOASID allocation/Free
+  - Group management in the form of ioasid_set
+  - Private data storage and lookup
+  - Reference counting
+  - Event notification in case of state change
+
+IOASID Set Level APIs
+--------------------------
+For use cases such as guest SVA it is necessary to manage IOASIDs at
+a group level. For example, VMs may allocate multiple IOASIDs for
+guest process address sharing (vSVA). It is imperative to enforce
+VM-IOASID ownership such that malicious guest cannot target DMA
+traffic outside its own IOASIDs, or free an active IOASID belong to
+another VM.
+::
+
+ struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota, u32 type)
+
+ int ioasid_adjust_set(struct ioasid_set *set, int quota);
+
+ void ioasid_set_get(struct ioasid_set *set)
+
+ void ioasid_set_put(struct ioasid_set *set)
+
+ void ioasid_set_get_locked(struct ioasid_set *set)
+
+ void ioasid_set_put_locked(struct ioasid_set *set)
+
+ int ioasid_set_for_each_ioasid(struct ioasid_set *sdata,
+                                void (*fn)(ioasid_t id, void *data),
+				void *data)
+
+
+IOASID set concept is introduced to represent such IOASID groups. Each
+IOASID set is created with a token which can be one of the following
+types:
+
+ - IOASID_SET_TYPE_NULL (Arbitrary u64 value)
+ - IOASID_SET_TYPE_MM (Set token is a mm_struct)
+
+The explicit MM token type is useful when multiple users of an IOASID
+set under the same process need to communicate about their shared IOASIDs.
+E.g. An IOASID set created by VFIO for one guest can be associated
+with the KVM instance for the same guest since they share a common mm_struct.
+
+The IOASID set APIs serve the following purposes:
+
+ - Ownership/permission enforcement
+ - Take collective actions, e.g. free an entire set
+ - Event notifications within a set
+ - Look up a set based on token
+ - Quota enforcement
+
+Individual IOASID APIs
+----------------------
+Once an ioasid_set is created, IOASIDs can be allocated from the set.
+Within the IOASID set namespace, set private ID (SPID) is supported. In
+the VM use case, SPID can be used for storing guest PASID.
+
+::
+
+ ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max,
+                       void *private);
+
+ int ioasid_get(struct ioasid_set *set, ioasid_t ioasid);
+
+ void ioasid_put(struct ioasid_set *set, ioasid_t ioasid);
+
+ int ioasid_get_locked(struct ioasid_set *set, ioasid_t ioasid);
+
+ void ioasid_put_locked(struct ioasid_set *set, ioasid_t ioasid);
+
+ void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid,
+                   bool (*getter)(void *));
+
+ ioasid_t ioasid_find_by_spid(struct ioasid_set *set, ioasid_t spid)
+
+ int ioasid_attach_data(struct ioasid_set *set, ioasid_t ioasid,
+                        void *data);
+ int ioasid_attach_spid(struct ioasid_set *set, ioasid_t ioasid,
+                        ioasid_t ssid);
+
+
+Notifications
+-------------
+An IOASID may have multiple users, each user may have hardware context
+associated with an IOASID. When the status of an IOASID changes,
+e.g. an IOASID is being freed, users need to be notified such that the
+associated hardware context can be cleared, flushed, and drained.
+
+::
+
+ int ioasid_register_notifier(struct ioasid_set *set, struct
+                              notifier_block *nb)
+
+ void ioasid_unregister_notifier(struct ioasid_set *set,
+                                 struct notifier_block *nb)
+
+ int ioasid_register_notifier_mm(struct mm_struct *mm, struct
+                                 notifier_block *nb)
+
+ void ioasid_unregister_notifier_mm(struct mm_struct *mm, struct
+                                    notifier_block *nb)
+
+ int ioasid_notify(ioasid_t ioasid, enum ioasid_notify_val cmd,
+                   unsigned int flags)
+
+
+Events
+~~~~~~
+Notification events are pertinent to individual IOASIDs, they can be
+one of the following:
+
+ - ALLOC
+ - FREE
+ - BIND
+ - UNBIND
+
+Ordering
+~~~~~~~~
+Ordering is supported by IOASID notification priorities as the
+following (in ascending order):
+
+::
+
+ enum ioasid_notifier_prios {
+	IOASID_PRIO_LAST,
+	IOASID_PRIO_IOMMU,
+	IOASID_PRIO_DEVICE,
+	IOASID_PRIO_CPU,
+ };
+
+The typical use case is when an IOASID is freed due to an exception, DMA
+source should be quiesced before tearing down other hardware contexts
+in the system. This will reduce the churn in handling faults. DMA work
+submission is performed by the CPU which is granted higher priority than
+devices.
+
+
+Scopes
+~~~~~~
+There are two types of notifiers in IOASID core: system-wide and
+ioasid_set-wide.
+
+System-wide notifier is catering for users that need to handle all
+IOASIDs in the system. E.g. The IOMMU driver handles all IOASIDs.
+
+Per ioasid_set notifier can be used by VM specific components such as
+KVM. After all, each KVM instance only cares about IOASIDs within its
+own set.
+
+
+Atomicity
+~~~~~~~~~
+IOASID notifiers are atomic due to spinlocks used inside the IOASID
+core. For tasks cannot be completed in the notifier handler, async work
+can be submitted to complete the work later as long as there is no
+ordering requirement.
+
+Reference counting
+------------------
+IOASID lifecycle management is based on reference counting. Users of
+IOASID intend to align lifecycle with the IOASID need to hold
+reference of the IOASID. IOASID will not be returned to the pool for
+allocation until all references are dropped. Calling ioasid_free()
+will mark the IOASID as FREE_PENDING if the IOASID has outstanding
+reference. ioasid_get() is not allowed once an IOASID is in the
+FREE_PENDING state.
+
+Event notifications are used to inform users of IOASID status change.
+IOASID_FREE event prompts users to drop their references after
+clearing its context.
+
+For example, on VT-d platform when an IOASID is freed, teardown
+actions are performed on KVM, device driver, and IOMMU driver.
+KVM shall register notifier block with::
+
+ static struct notifier_block pasid_nb_kvm = {
+	.notifier_call = pasid_status_change_kvm,
+	.priority      = IOASID_PRIO_CPU,
+ };
+
+VDCM driver shall register notifier block with::
+
+ static struct notifier_block pasid_nb_vdcm = {
+	.notifier_call = pasid_status_change_vdcm,
+	.priority      = IOASID_PRIO_DEVICE,
+ };
+
+In both cases, notifier blocks shall be registered on the IOASID set
+such that *only* events from the matching VM is received.
+
+If KVM attempts to register notifier block before the IOASID set is
+created for the MM token, the notifier block will be placed on a
+pending list inside IOASID core. Once the token matching IOASID set
+is created, IOASID will register the notifier block automatically.
+IOASID core does not replay events for the existing IOASIDs in the
+set. For IOASID set of MM type, notification blocks can be registered
+on empty sets only. This is to avoid lost events.
+
+IOMMU driver shall register notifier block on global chain::
+
+ static struct notifier_block pasid_nb_vtd = {
+	.notifier_call = pasid_status_change_vtd,
+	.priority      = IOASID_PRIO_IOMMU,
+ };
+
+Custom allocator APIs
+---------------------
+
+::
+
+ int ioasid_register_allocator(struct ioasid_allocator_ops *allocator);
+
+ void ioasid_unregister_allocator(struct ioasid_allocator_ops *allocator);
+
+Allocator Choices
+~~~~~~~~~~~~~~~~~
+IOASIDs are allocated for both host and guest SVA/IOVA usage. However,
+allocators can be different. For example, on VT-d guest PASID
+allocation must be performed via a virtual command interface which is
+emulated by VMM.
+
+IOASID core has the notion of "custom allocator" such that guest can
+register virtual command allocator that precedes the default one.
+
+Namespaces
+~~~~~~~~~~
+IOASIDs are limited system resources that default to 20 bits in
+size. Since each device has its own table, theoretically the namespace
+can be per device also. However, for security reasons sharing PASID
+tables among devices are not good for isolation. Therefore, IOASID
+namespace is system-wide.
+
+There are also other reasons to have this simpler system-wide
+namespace. Take VT-d as an example, VT-d supports shared workqueue
+and ENQCMD[1] where one IOASID could be used to submit work on
+multiple devices that are shared with other VMs. This requires IOASID
+to be system-wide. This is also the reason why guests must use an
+emulated virtual command interface to allocate IOASID from the host.
+
+
+Life cycle
+==========
+This section covers IOASID lifecycle management for both bare-metal
+and guest usages. In bare-metal SVA, MMU notifier is directly hooked
+up with IOMMU driver, therefore the process address space (MM)
+lifecycle is aligned with IOASID.
+
+However, guest MMU notifier is not available to host IOMMU driver,
+when guest MM terminates unexpectedly, the events have to go through
+VFIO and IOMMU UAPI to reach host IOMMU driver. There are also more
+parties involved in guest SVA, e.g. on Intel VT-d platform, IOASIDs
+are used by IOMMU driver, KVM, VDCM, and VFIO.
+
+Native IOASID Life Cycle (VT-d Example)
+---------------------------------------
+
+The normal flow of native SVA code with Intel Data Streaming
+Accelerator(DSA) [2] as example:
+
+1. Host user opens accelerator FD, e.g. DSA driver, or uacce;
+2. DSA driver allocate WQ, do sva_bind_device();
+3. IOMMU driver calls ioasid_alloc(), then bind PASID with device,
+   mmu_notifier_get()
+4. DMA starts by DSA driver userspace
+5. DSA userspace close FD
+6. DSA/uacce kernel driver handles FD.close()
+7. DSA driver stops DMA
+8. DSA driver calls sva_unbind_device();
+9. IOMMU driver does unbind, clears PASID context in IOMMU, flush
+   TLBs. mmu_notifier_put() called.
+10. mmu_notifier.release() called, IOMMU SVA code calls ioasid_free()*
+11. The IOASID is returned to the pool, reclaimed.
+
+::
+
+   * With ENQCMD, PASID used on VT-d is not released in mmu_notifier() but
+     mmdrop(). mmdrop comes after FD close. Should not matter.
+     If the user process dies unexpectedly, Step #10 may come before
+     Step #5, in between, all DMA faults discarded. PRQ responded with
+     code INVALID REQUEST.
+
+During the normal teardown, the following three steps would happen in
+order:
+
+1. Device driver stops DMA request
+2. IOMMU driver unbinds PASID and mm, flush all TLBs, drain in-flight
+   requests.
+3. IOASID freed
+
+Exception happens when process terminates *before* device driver stops
+DMA and call IOMMU driver to unbind. The flow of process exists are as
+follows:
+
+::
+
+   do_exit() {
+	exit_mm() {
+		mm_put();
+		exit_mmap() {
+			intel_invalidate_range() //mmu notifier
+			tlb_finish_mmu()
+			mmu_notifier_release(mm) {
+				intel_iommu_release() {
+   [2]					intel_iommu_teardown_pasid();
+                                        intel_iommu_flush_tlbs();
+				}
+				// tlb_invalidate_range cb removed
+			}
+			unmap_vmas();
+                        free_pgtables(); // IOMMU cannot walk PGT after this
+		};
+	}
+	exit_files(tsk) {
+		close_files() {
+			dsa_close();
+   [1]			dsa_stop_dma();
+                        intel_svm_unbind_pasid(); //nothing to do
+		}
+	}
+   }
+
+   mmdrop() /* some random time later, lazy mm user */ {
+   	mm_free_pgd();
+        destroy_context(mm); {
+   [3]	        ioasid_free();
+	}
+   }
+
+As shown in the list above, step #2 could happen before
+#1. Unrecoverable(UR) faults could happen between #2 and #1.
+
+Also notice that TLB invalidation occurs at mmu_notifier
+invalidate_range callback as well as the release callback. The reason
+is that release callback will delete IOMMU driver from the notifier
+chain which may skip invalidate_range() calls during the exit path.
+
+To avoid unnecessary reporting of UR fault, IOMMU driver shall disable
+fault reporting after free and before unbind.
+
+Guest IOASID Life Cycle (VT-d Example)
+--------------------------------------
+Guest IOASID life cycle starts with guest driver open(), this could be
+uacce or individual accelerator driver such as DSA. At FD open,
+sva_bind_device() is called which triggers a series of actions.
+
+The example below is an illustration of *normal* operations that
+involves *all* the SW components in VT-d. The flow can be simpler if
+no ENQCMD is supported.
+
+::
+
+     VFIO        IOMMU        KVM        VDCM        IOASID       Ref
+   ..................................................................
+   1             ioasid_register_notifier/_mm()
+   2 ioasid_alloc()                                               1
+   3 bind_gpasid()
+   4             iommu_bind()->ioasid_get()                       2
+   5             ioasid_notify(BIND)
+   6                          -> ioasid_get()                     3
+   7                          -> vmcs_update_atomic()
+   8 mdev_write(gpasid)
+   9                                    hpasid=
+   10                                   find_by_spid(gpasid)      4
+   11                                   vdev_write(hpasid)
+   12 -------- GUEST STARTS DMA --------------------------
+   13 -------- GUEST STOPS DMA --------------------------
+   14 mdev_clear(gpasid)
+   15                                   vdev_clear(hpasid)
+   16                                   ioasid_put()               3
+   17 unbind_gpasid()
+   18            iommu_ubind()
+   19            ioasid_notify(UNBIND)
+   20                          -> vmcs_update_atomic()
+   21                          -> ioasid_put()                     2
+   22 ioasid_free()                                                1
+   23            ioasid_put()                                      0
+   24                                                 Reclaimed
+   -------------- New Life Cycle Begin ----------------------------
+   1  ioasid_alloc()                                  ->           1
+
+   Note: IOASID Notification Events: FREE, BIND, UNBIND
+
+Exception cases arise when a guest crashes or a malicious guest
+attempts to cause disruption on the host system. The fault handling
+rules are:
+
+1. IOASID free must *always* succeed.
+2. An inactive period may be required before the freed IOASID is
+   reclaimed. During this period, consumers of IOASID perform cleanup.
+3. Malfunction is limited to the guest owned resources for all
+   programming errors.
+
+The primary source of exception is when the following are out of
+order:
+
+1. Start/Stop of DMA activity
+   (Guest device driver, mdev via VFIO)
+2. Setup/Teardown of IOMMU PASID context, IOTLB, DevTLB flushes
+   (Host IOMMU driver bind/unbind)
+3. Setup/Teardown of VMCS PASID translation table entries (KVM) in
+   case of ENQCMD
+4. Programming/Clearing host PASID in VDCM (Host VDCM driver)
+5. IOASID alloc/free (Host IOASID)
+
+VFIO is the *only* user-kernel interface, which is ultimately
+responsible for exception handlings.
+
+#1 is processed the same way as the assigned device today based on
+device file descriptors and events. There is no special handling.
+
+#3 is based on bind/unbind events emitted by #2.
+
+#4 is naturally aligned with IOASID life cycle in that an illegal
+guest PASID programming would fail in obtaining reference of the
+matching host IOASID.
+
+#5 is similar to #4. The fault will be reported to the user if PASID
+used in the ENQCMD is not set up in VMCS PASID translation table.
+
+Therefore, the remaining out of order problem is between #2 and
+#5. I.e. unbind vs. free. More specifically, free before unbind.
+
+IOASID notifier and refcounting are used to ensure order. Following
+a publisher-subscriber pattern where:
+
+- Publishers: VFIO & IOMMU
+- Subscribers: KVM, VDCM, IOMMU
+
+IOASID notifier is atomic which requires subscribers to do quick
+handling of the event in the atomic context. Workqueue can be used for
+any processing that requires thread context. IOASID reference must be
+acquired before receiving the FREE event. The reference must be
+dropped at the end of the processing in order to return the IOASID to
+the pool.
+
+Let's examine the IOASID life cycle again when free happens *before*
+unbind. This could be a result of misbehaving guests or crash. Assuming
+VFIO cannot enforce unbind->free order. Notice that the setup part up
+until step #12 is identical to the normal case, the flow below starts
+with step 13.
+
+::
+
+     VFIO        IOMMU        KVM        VDCM        IOASID       Ref
+   ..................................................................
+   13 -------- GUEST STARTS DMA --------------------------
+   14 -------- *GUEST MISBEHAVES!!!* ----------------
+   15 ioasid_free()
+   16                                             ioasid_notify(FREE)
+   17                                             mark_ioasid_inactive[1]
+   18                          kvm_nb_handler(FREE)
+   19                          vmcs_update_atomic()
+   20                          ioasid_put_locked()   ->           3
+   21                                   vdcm_nb_handler(FREE)
+   22            iomm_nb_handler(FREE)
+   23 ioasid_free() returns[2]          schedule_work()           2
+   24            schedule_work()        vdev_clear_wk(hpasid)
+   25            teardown_pasid_wk()
+   26                                   ioasid_put() ->           1
+   27            ioasid_put()                                     0
+   28                                                 Reclaimed
+   29 unbind_gpasid()
+   30            iommu_unbind()->ioasid_find() Fails[3]
+   -------------- New Life Cycle Begin ----------------------------
+
+Note:
+
+1. By marking IOASID inactive at step #17, no new references can be
+   held. ioasid_get/find() will return -ENOENT;
+2. After step #23, all events can go out of order. Shall not affect
+   the outcome.
+3. IOMMU driver fails to find private data for unbinding. If unbind is
+   called after the same IOASID is allocated for the same guest again,
+   this is a programming error. The damage is limited to the guest
+   itself since unbind performs permission checking based on the
+   IOASID set associated with the guest process.
+
+KVM PASID Translation Table Updates
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Per VM PASID translation table is maintained by KVM in order to
+support ENQCMD in the guest. The table contains host-guest PASID
+translations to be consumed by CPU ucode. The synchronization of the
+PASID states depends on VFIO/IOMMU driver, where IOCTL and atomic
+notifiers are used. KVM must register IOASID notifier per VM instance
+during launch time. The following events are handled:
+
+1. BIND/UNBIND
+2. FREE
+
+Rules:
+
+1. Multiple devices can bind with the same PASID, this can be different PCI
+   devices or mdevs within the same PCI device. However, only the
+   *first* BIND and *last* UNBIND emit notifications.
+2. IOASID code is responsible for ensuring the correctness of H-G
+   PASID mapping. There is no need for KVM to validate the
+   notification data.
+3. When UNBIND happens *after* FREE, KVM will see error in
+   ioasid_get() even when the reclaim is not done. IOMMU driver will
+   also avoid sending UNBIND if the PASID is already FREE.
+4. When KVM terminates *before* FREE & UNBIND, references will be
+   dropped for all host PASIDs.
+
+VDCM PASID Programming
+~~~~~~~~~~~~~~~~~~~~~~
+VDCM composes virtual devices and exposes them to the guests. When
+the guest allocates a PASID then program it to the virtual device, VDCM
+intercepts the programming attempt then program the matching host
+PASID on to the hardware.
+Conversely, when a device is going away, VDCM must be informed such
+that PASID context on the hardware can be cleared. There could be
+multiple mdevs assigned to different guests in the same VDCM. Since
+the PASID table is shared at PCI device level, lazy clearing is not
+secure. A malicious guest can attack by using newly freed PASIDs that
+are allocated by another guest.
+
+By holding a reference of the PASID until VDCM cleans up the HW context,
+it is guaranteed that PASID life cycles do not cross within the same
+device.
+
+
+Reference
+====================================================
+1. https://software.intel.com/sites/default/files/managed/c5/15/architecture-instruction-set-extensions-programming-reference.pdf
+
+2. https://01.org/blogs/2019/introducing-intel-data-streaming-accelerator
+
+3. https://software.intel.com/en-us/download/intel-data-streaming-accelerator-preliminary-architecture-specification
Signature · 1 line
2.7.4
author
#13 lore raw

PATCH v2 3/9 iommu/ioasid: Introduce ioasid_set APIs

In reply to Jacob Pan

ioasid_set was introduced as an arbitrary token that are shared by a group of IOASIDs. For example, if IOASID #1 and #2 are allocated via the same ioasid_set*, they are viewed as to belong to the same set.

For guest SVA usages, system-wide IOASID resources need to be partitioned such that VMs can have its own quota and being managed separately. ioasid_set is the perfect candidate for meeting such requirements. This patch redefines and extends ioasid_set with the following new fields:
- Quota
- Reference count
- Storage of its namespace
- The token is stored in the new ioasid_set but with optional types

ioasid_set level APIs are introduced that wires up these new data. Existing users of IOASID APIs are converted where a host IOASID set is allocated for bare-metal usage.

---
 drivers/iommu/intel/iommu.c |  27 ++-
 drivers/iommu/intel/pasid.h |   1 +
 drivers/iommu/intel/svm.c   |   8 +-
 drivers/iommu/ioasid.c      | 390 +++++++++++++++++++++++++++++++++++++++++---
 include/linux/ioasid.h      |  82 ++++++++--
 5 files changed, 465 insertions(+), 43 deletions(-)
drivers/iommu/intel/iommu.c +22 -5
index a3a0b5c8921d..5813eeaa5edb 100644
@@ -42,6 +42,7 @@
 #include <linux/crash_dump.h>
 #include <linux/numa.h>
 #include <linux/swiotlb.h>
+#include <linux/ioasid.h>
 #include <asm/irq_remapping.h>
 #include <asm/cacheflush.h>
 #include <asm/iommu.h>
@@ -103,6 +104,9 @@
  */
 #define INTEL_IOMMU_PGSIZES	(~0xFFFUL)
 
+/* PASIDs used by host SVM */
+struct ioasid_set *host_pasid_set;
+
 static inline int agaw_to_level(int agaw)
 {
 	return agaw + 2;
@@ -3103,8 +3107,8 @@ static void intel_vcmd_ioasid_free(ioasid_t ioasid, void *data)
 	 * Sanity check the ioasid owner is done at upper layer, e.g. VFIO
 	 * We can only free the PASID when all the devices are unbound.
 	 */
-	if (ioasid_find(NULL, ioasid, NULL)) {
-		pr_alert("Cannot free active IOASID %d\n", ioasid);
+	if (IS_ERR(ioasid_find(host_pasid_set, ioasid, NULL))) {
+		pr_err("Cannot free IOASID %d, not in system set\n", ioasid);
 		return;
 	}
 	vcmd_free_pasid(iommu, ioasid);
@@ -3288,6 +3292,19 @@ static int __init init_dmars(void)
 	if (ret)
 		goto free_iommu;
 
+	/* PASID is needed for scalable mode irrespective to SVM */
+	if (intel_iommu_sm) {
+		ioasid_install_capacity(intel_pasid_max_id);
+		/* We should not run out of IOASIDs at boot */
+		host_pasid_set = ioasid_alloc_set(NULL, PID_MAX_DEFAULT,
+						IOASID_SET_TYPE_NULL);
+		if (IS_ERR_OR_NULL(host_pasid_set)) {
+			pr_err("Failed to enable host PASID allocator %lu\n",
+				PTR_ERR(host_pasid_set));
+			intel_iommu_sm = 0;
+		}
+	}
+
 	/*
 	 * for each drhd
 	 *   enable fault log
@@ -5149,7 +5166,7 @@ static void auxiliary_unlink_device(struct dmar_domain *domain,
 	domain->auxd_refcnt--;
 
 	if (!domain->auxd_refcnt && domain->default_pasid > 0)
-		ioasid_free(domain->default_pasid);
+		ioasid_free(host_pasid_set, domain->default_pasid);
 }
 
 static int aux_domain_add_dev(struct dmar_domain *domain,
@@ -5167,7 +5184,7 @@ static int aux_domain_add_dev(struct dmar_domain *domain,
 		int pasid;
 
 		/* No private data needed for the default pasid */
-		pasid = ioasid_alloc(NULL, PASID_MIN,
+		pasid = ioasid_alloc(host_pasid_set, PASID_MIN,
 				     pci_max_pasids(to_pci_dev(dev)) - 1,
 				     NULL);
 		if (pasid == INVALID_IOASID) {
@@ -5210,7 +5227,7 @@ static int aux_domain_add_dev(struct dmar_domain *domain,
 	spin_unlock(&iommu->lock);
 	spin_unlock_irqrestore(&device_domain_lock, flags);
 	if (!domain->auxd_refcnt && domain->default_pasid > 0)
-		ioasid_free(domain->default_pasid);
+		ioasid_free(host_pasid_set, domain->default_pasid);
 
 	return ret;
 }
drivers/iommu/intel/pasid.h +1 -0
index c9850766c3a9..ccdc23446015 100644
@@ -99,6 +99,7 @@ static inline bool pasid_pte_is_present(struct pasid_entry *pte)
 }
 
 extern u32 intel_pasid_max_id;
+extern struct ioasid_set *host_pasid_set;
 int intel_pasid_alloc_id(void *ptr, int start, int end, gfp_t gfp);
 void intel_pasid_free_id(int pasid);
 void *intel_pasid_lookup_id(int pasid);
drivers/iommu/intel/svm.c +4 -4
index 37a9beabc0ca..634e191ca2c3 100644
@@ -551,7 +551,7 @@ intel_svm_bind_mm(struct device *dev, int flags, struct svm_dev_ops *ops,
 			pasid_max = intel_pasid_max_id;
 
 		/* Do not use PASID 0, reserved for RID to PASID */
-		svm->pasid = ioasid_alloc(NULL, PASID_MIN,
+		svm->pasid = ioasid_alloc(host_pasid_set, PASID_MIN,
 					  pasid_max - 1, svm);
 		if (svm->pasid == INVALID_IOASID) {
 			kfree(svm);
@@ -568,7 +568,7 @@ intel_svm_bind_mm(struct device *dev, int flags, struct svm_dev_ops *ops,
 		if (mm) {
 			ret = mmu_notifier_register(&svm->notifier, mm);
 			if (ret) {
-				ioasid_free(svm->pasid);
+				ioasid_free(host_pasid_set, svm->pasid);
 				kfree(svm);
 				kfree(sdev);
 				goto out;
@@ -586,7 +586,7 @@ intel_svm_bind_mm(struct device *dev, int flags, struct svm_dev_ops *ops,
 		if (ret) {
 			if (mm)
 				mmu_notifier_unregister(&svm->notifier, mm);
-			ioasid_free(svm->pasid);
+			ioasid_free(host_pasid_set, svm->pasid);
 			kfree(svm);
 			kfree(sdev);
 			goto out;
@@ -655,7 +655,7 @@ static int intel_svm_unbind_mm(struct device *dev, int pasid)
 			kfree_rcu(sdev, rcu);
 
 			if (list_empty(&svm->devs)) {
-				ioasid_free(svm->pasid);
+				ioasid_free(host_pasid_set, svm->pasid);
 				if (svm->mm)
 					mmu_notifier_unregister(&svm->notifier, svm->mm);
 				list_del(&svm->list);
drivers/iommu/ioasid.c +367 -23

449 lines not on this page · raw message

include/linux/ioasid.h +71 -11

110 lines not on this page · raw message

Signature · 1 line
2.7.4
author
#35 lore raw

PATCH v2 5/9 iommu/ioasid: Introduce ioasid_set private ID

In reply to Jacob Pan

When an IOASID set is used for guest SVA, each VM will acquire its ioasid_set for IOASID allocations. IOASIDs within the VM must have a host/physical IOASID backing, mapping between guest and host IOASIDs can be non-identical. IOASID set private ID (SPID) is introduced in this patch to be used as guest IOASID. However, the concept of ioasid_set specific namespace is generic, thus named SPID.

As SPID namespace is within the IOASID set, the IOASID core can provide lookup services at both directions. SPIDs may not be allocated when its IOASID is allocated, the mapping between SPID and IOASID is usually established when a guest page table is bound to a host PASID.

---
 drivers/iommu/ioasid.c | 54 ++++++++++++++++++++++++++++++++++++++++++++++++++
 include/linux/ioasid.h | 12 +++++++++++
 2 files changed, 66 insertions(+)
drivers/iommu/ioasid.c +54 -0
index 5f31d63c75b1..c0aef38a4fde 100644
@@ -21,6 +21,7 @@ enum ioasid_state {
  * struct ioasid_data - Meta data about ioasid
  *
  * @id:		Unique ID
+ * @spid:	Private ID unique within a set
  * @users	Number of active users
  * @state	Track state of the IOASID
  * @set		Meta data of the set this IOASID belongs to
@@ -29,6 +30,7 @@ enum ioasid_state {
  */
 struct ioasid_data {
 	ioasid_t id;
+	ioasid_t spid;
 	struct ioasid_set *set;
 	refcount_t users;
 	enum ioasid_state state;
@@ -326,6 +328,58 @@ int ioasid_attach_data(ioasid_t ioasid, void *data)
 EXPORT_SYMBOL_GPL(ioasid_attach_data);
 
 /**
+ * ioasid_attach_spid - Attach ioasid_set private ID to an IOASID
+ *
+ * @ioasid: the ID to attach
+ * @spid:   the ioasid_set private ID of @ioasid
+ *
+ * For IOASID that is already allocated, private ID within the set can be
+ * attached via this API. Future lookup can be done via ioasid_find.
+ */
+int ioasid_attach_spid(ioasid_t ioasid, ioasid_t spid)
+{
+	struct ioasid_data *ioasid_data;
+	int ret = 0;
+
+	spin_lock(&ioasid_allocator_lock);
+	ioasid_data = xa_load(&active_allocator->xa, ioasid);
+
+	if (!ioasid_data) {
+		pr_err("No IOASID entry %d to attach SPID %d\n",
+			ioasid, spid);
+		ret = -ENOENT;
+		goto done_unlock;
+	}
+	ioasid_data->spid = spid;
+
+done_unlock:
+	spin_unlock(&ioasid_allocator_lock);
+	return ret;
+}
+EXPORT_SYMBOL_GPL(ioasid_attach_spid);
+
+ioasid_t ioasid_find_by_spid(struct ioasid_set *set, ioasid_t spid)
+{
+	struct ioasid_data *entry;
+	unsigned long index;
+
+	if (!xa_load(&ioasid_sets, set->sid)) {
+		pr_warn("Invalid set\n");
+		return INVALID_IOASID;
+	}
+
+	xa_for_each(&set->xa, index, entry) {
+		if (spid == entry->spid) {
+			pr_debug("Found ioasid %lu by spid %u\n", index, spid);
+			refcount_inc(&entry->users);
+			return index;
+		}
+	}
+	return INVALID_IOASID;
+}
+EXPORT_SYMBOL_GPL(ioasid_find_by_spid);
+
+/**
  * ioasid_alloc - Allocate an IOASID
  * @set: the IOASID set
  * @min: the minimum ID (inclusive)
include/linux/ioasid.h +12 -0
index 310abe4187a3..d4b3e83672f6 100644
@@ -73,6 +73,8 @@ bool ioasid_is_active(ioasid_t ioasid);
 
 void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid, bool (*getter)(void *));
 int ioasid_attach_data(ioasid_t ioasid, void *data);
+int ioasid_attach_spid(ioasid_t ioasid, ioasid_t spid);
+ioasid_t ioasid_find_by_spid(struct ioasid_set *set, ioasid_t spid);
 int ioasid_register_allocator(struct ioasid_allocator_ops *allocator);
 void ioasid_unregister_allocator(struct ioasid_allocator_ops *allocator);
 void ioasid_is_in_set(struct ioasid_set *set, ioasid_t ioasid);
@@ -136,5 +138,15 @@ static inline int ioasid_attach_data(ioasid_t ioasid, void *data)
 	return -ENOTSUPP;
 }
 
+staic inline int ioasid_attach_spid(ioasid_t ioasid, ioasid_t spid)
+{
+	return -ENOTSUPP;
+}
+
+static inline ioasid_t ioasid_find_by_spid(struct ioasid_set *set, ioasid_t spid)
+{
+	return -ENOTSUPP;
+}
+
 #endif /* CONFIG_IOASID */
 #endif /* __LINUX_IOASID_H */
Signature · 1 line
2.7.4
author
#27 lore raw

PATCH v2 4/9 iommu/ioasid: Add reference couting functions

In reply to Jacob Pan

There can be multiple users of an IOASID, each user could have hardware contexts associated with the IOASID. In order to align lifecycles, reference counting is introduced in this patch. It is expected that when an IOASID is being freed, each user will drop a reference only after its context is cleared.

---
 drivers/iommu/ioasid.c | 113 +++++++++++++++++++++++++++++++++++++++++++++++++
 include/linux/ioasid.h |   4 ++
 2 files changed, 117 insertions(+)
drivers/iommu/ioasid.c +113 -0
index f73b3dbfc37a..5f31d63c75b1 100644
@@ -717,6 +717,119 @@ int ioasid_set_for_each_ioasid(struct ioasid_set *set,
 EXPORT_SYMBOL_GPL(ioasid_set_for_each_ioasid);
 
 /**
+ * IOASID refcounting rules
+ * - ioasid_alloc() set initial refcount to 1
+ *
+ * - ioasid_free() decrement and test refcount.
+ *     If refcount is 0, ioasid will be freed. Deleted from the system-wide
+ *     xarray as well as per set xarray. The IOASID will be returned to the
+ *     pool and available for new allocations.
+ *
+ *     If recount is non-zero, mark IOASID as IOASID_STATE_FREE_PENDING.
+ *     No new reference can be added. The IOASID is not returned to the pool
+ *     for reuse.
+ *     After free, ioasid_get() will return error but ioasid_find() and other
+ *     non refcount adding APIs will continue to work until the last reference
+ *     is dropped
+ *
+ * - ioasid_get() get a reference on an active IOASID
+ *
+ * - ioasid_put() decrement and test refcount of the IOASID.
+ *     If refcount is 0, ioasid will be freed. Deleted from the system-wide
+ *     xarray as well as per set xarray. The IOASID will be returned to the
+ *     pool and available for new allocations.
+ *     Do nothing if refcount is non-zero.
+ *
+ * - ioasid_find() does not take reference, caller must hold reference
+ *
+ * ioasid_free() can be called multiple times without error until all refs are
+ * dropped.
+ */
+
+int ioasid_get_locked(struct ioasid_set *set, ioasid_t ioasid)
+{
+	struct ioasid_data *data;
+
+	data = xa_load(&active_allocator->xa, ioasid);
+	if (!data) {
+		pr_err("Trying to get unknown IOASID %u\n", ioasid);
+		return -EINVAL;
+	}
+	if (data->state == IOASID_STATE_FREE_PENDING) {
+		pr_err("Trying to get IOASID being freed%u\n", ioasid);
+		return -EBUSY;
+	}
+
+	if (set && data->set != set) {
+		pr_err("Trying to get IOASID not in set%u\n", ioasid);
+		/* data found but does not belong to the set */
+		return -EACCES;
+	}
+	refcount_inc(&data->users);
+
+	return 0;
+}
+EXPORT_SYMBOL_GPL(ioasid_get_locked);
+
+/**
+ * ioasid_get - Obtain a reference of an ioasid
+ * @set
+ * @ioasid
+ *
+ * Check set ownership if @set is non-null.
+ */
+int ioasid_get(struct ioasid_set *set, ioasid_t ioasid)
+{
+	int ret = 0;
+
+	spin_lock(&ioasid_allocator_lock);
+	ret = ioasid_get_locked(set, ioasid);
+	spin_unlock(&ioasid_allocator_lock);
+
+	return ret;
+}
+EXPORT_SYMBOL_GPL(ioasid_get);
+
+void ioasid_put_locked(struct ioasid_set *set, ioasid_t ioasid)
+{
+	struct ioasid_data *data;
+
+	data = xa_load(&active_allocator->xa, ioasid);
+	if (!data) {
+		pr_err("Trying to put unknown IOASID %u\n", ioasid);
+		return;
+	}
+
+	if (set && data->set != set) {
+		pr_err("Trying to drop IOASID not in the set %u\n", ioasid);
+		return;
+	}
+
+	if (!refcount_dec_and_test(&data->users)) {
+		pr_debug("%s: IOASID %d has %d remainning users\n",
+			__func__, ioasid, refcount_read(&data->users));
+		return;
+	}
+	ioasid_do_free(data);
+}
+EXPORT_SYMBOL_GPL(ioasid_put_locked);
+
+/**
+ * ioasid_put - Drop a reference of an ioasid
+ * @set
+ * @ioasid
+ *
+ * Check set ownership if @set is non-null.
+ */
+void ioasid_put(struct ioasid_set *set, ioasid_t ioasid)
+{
+	spin_lock(&ioasid_allocator_lock);
+	ioasid_put_locked(set, ioasid);
+	spin_unlock(&ioasid_allocator_lock);
+}
+EXPORT_SYMBOL_GPL(ioasid_put);
+
+/**
  * ioasid_find - Find IOASID data
  * @set: the IOASID set
  * @ioasid: the IOASID to find
include/linux/ioasid.h +4 -0
index 412d025d440e..310abe4187a3 100644
@@ -76,6 +76,10 @@ int ioasid_attach_data(ioasid_t ioasid, void *data);
 int ioasid_register_allocator(struct ioasid_allocator_ops *allocator);
 void ioasid_unregister_allocator(struct ioasid_allocator_ops *allocator);
 void ioasid_is_in_set(struct ioasid_set *set, ioasid_t ioasid);
+int ioasid_get(struct ioasid_set *set, ioasid_t ioasid);
+int ioasid_get_locked(struct ioasid_set *set, ioasid_t ioasid);
+void ioasid_put(struct ioasid_set *set, ioasid_t ioasid);
+void ioasid_put_locked(struct ioasid_set *set, ioasid_t ioasid);
 int ioasid_set_for_each_ioasid(struct ioasid_set *sdata,
 			       void (*fn)(ioasid_t id, void *data),
 			       void *data);
Signature · 1 line
2.7.4
author
#43 lore raw

PATCH v2 6/9 iommu/ioasid: Introduce notification APIs

In reply to Jacob Pan

Relations among IOASID users largely follow a publisher-subscriber pattern. E.g. to support guest SVA on Intel Scalable I/O Virtualization (SIOV) enabled platforms, VFIO, IOMMU, device drivers, KVM are all users of IOASIDs. When a state change occurs, VFIO publishes the change event that needs to be processed by other users/subscribers.

This patch introduced two types of notifications: global and per ioasid_set. The latter is intended for users who only needs to handle events related to the IOASID of a given set.
For more information, refer to the kernel documentation at Documentation/ioasid.rst.

---
 drivers/iommu/ioasid.c | 280 ++++++++++++++++++++++++++++++++++++++++++++++++-
 include/linux/ioasid.h |  70 +++++++++++++
 2 files changed, 348 insertions(+), 2 deletions(-)
drivers/iommu/ioasid.c +278 -2
index c0aef38a4fde..6ddc09a7fe74 100644
@@ -9,8 +9,35 @@
 #include <linux/spinlock.h>
 #include <linux/xarray.h>
 #include <linux/ioasid.h>
+#include <linux/sched/mm.h>
 
 static DEFINE_XARRAY_ALLOC(ioasid_sets);
+/*
+ * An IOASID could have multiple consumers where each consumeer may have
+ * hardware contexts associated with IOASIDs.
+ * When a status change occurs, such as IOASID is being freed, notifier chains
+ * are used to keep the consumers in sync.
+ * This is a publisher-subscriber pattern where publisher can change the
+ * state of each IOASID, e.g. alloc/free, bind IOASID to a device and mm.
+ * On the other hand, subscribers gets notified for the state change and
+ * keep local states in sync.
+ *
+ * Currently, the notifier is global. A further optimization could be per
+ * IOASID set notifier chain.
+ */
+static ATOMIC_NOTIFIER_HEAD(ioasid_chain);
+
+/* List to hold pending notification block registrations */
+static LIST_HEAD(ioasid_nb_pending_list);
+static DEFINE_SPINLOCK(ioasid_nb_lock);
+struct ioasid_set_nb {
+	struct list_head	list;
+	struct notifier_block	*nb;
+	void			*token;
+	struct ioasid_set	*set;
+	bool			active;
+};
+
 enum ioasid_state {
 	IOASID_STATE_INACTIVE,
 	IOASID_STATE_ACTIVE,
@@ -394,6 +421,7 @@ EXPORT_SYMBOL_GPL(ioasid_find_by_spid);
 ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max,
 		      void *private)
 {
+	struct ioasid_nb_args args;
 	struct ioasid_data *data;
 	void *adata;
 	ioasid_t id = INVALID_IOASID;
@@ -445,8 +473,14 @@ ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max,
 		goto exit_free;
 	}
 	set->nr_ioasids++;
-	goto done_unlock;
+	args.id = id;
+	/* Set private ID is not attached during allocation */
+	args.spid = INVALID_IOASID;
+	args.set = set;
+	atomic_notifier_call_chain(&set->nh, IOASID_ALLOC, &args);
 
+	spin_unlock(&ioasid_allocator_lock);
+	return id;
 exit_free:
 	kfree(data);
 done_unlock:
@@ -479,6 +513,7 @@ static void ioasid_do_free(struct ioasid_data *data)
 
 static void ioasid_free_locked(struct ioasid_set *set, ioasid_t ioasid)
 {
+	struct ioasid_nb_args args;
 	struct ioasid_data *data;
 
 	data = xa_load(&active_allocator->xa, ioasid);
@@ -491,7 +526,16 @@ static void ioasid_free_locked(struct ioasid_set *set, ioasid_t ioasid)
 		pr_warn("Cannot free IOASID %u due to set ownership\n", ioasid);
 		return;
 	}
+
 	data->state = IOASID_STATE_FREE_PENDING;
+	/* Notify all users that this IOASID is being freed */
+	args.id = ioasid;
+	args.spid = data->spid;
+	args.pdata = data->private;
+	args.set = data->set;
+	atomic_notifier_call_chain(&ioasid_chain, IOASID_FREE, &args);
+	/* Notify the ioasid_set for per set users */
+	atomic_notifier_call_chain(&set->nh, IOASID_FREE, &args);
 
 	if (!refcount_dec_and_test(&data->users))
 		return;
@@ -514,6 +558,28 @@ void ioasid_free(struct ioasid_set *set, ioasid_t ioasid)
 }
 EXPORT_SYMBOL_GPL(ioasid_free);
 
+static void ioasid_add_pending_nb(struct ioasid_set *set)
+{
+	struct ioasid_set_nb *curr;
+
+	if (set->type != IOASID_SET_TYPE_MM)
+		return;
+
+	/*
+	 * Check if there are any pending nb requests for the given token, if so
+	 * add them to the notifier chain.
+	 */
+	spin_lock(&ioasid_nb_lock);
+	list_for_each_entry(curr, &ioasid_nb_pending_list, list) {
+		if (curr->token == set->token && !curr->active) {
+			atomic_notifier_chain_register(&set->nh, curr->nb);
+			curr->set = set;
+			curr->active = true;
+		}
+	}
+	spin_unlock(&ioasid_nb_lock);
+}
+
 /**
  * ioasid_alloc_set - Allocate a new IOASID set for a given token
  *
@@ -601,6 +667,13 @@ struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota, int type)
 	sdata->quota = quota;
 	sdata->sid = id;
 	refcount_set(&sdata->ref, 1);
+	ATOMIC_INIT_NOTIFIER_HEAD(&sdata->nh);
+
+	/*
+	 * Check if there are any pending nb requests for the given token, if so
+	 * add them to the notifier chain.
+	 */
+	ioasid_add_pending_nb(sdata);
 
 	/*
 	 * Per set XA is used to store private IDs within the set, get ready
@@ -617,6 +690,30 @@ struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota, int type)
 }
 EXPORT_SYMBOL_GPL(ioasid_alloc_set);
 
+
+/*
+ * ioasid_find_mm_set - Retrieve IOASID set with mm token
+ * Take a reference of the set if found.
+ */
+static struct ioasid_set *ioasid_find_mm_set(struct mm_struct *token)
+{
+	struct ioasid_set *sdata, *set = NULL;
+	unsigned long index;
+
+	spin_lock(&ioasid_allocator_lock);
+
+	xa_for_each(&ioasid_sets, index, sdata) {
+		if (sdata->type == IOASID_SET_TYPE_MM && sdata->token == token) {
+			refcount_inc(&sdata->ref);
+			set = sdata;
+			goto exit_unlock;
+		}
+	}
+exit_unlock:
+	spin_unlock(&ioasid_allocator_lock);
+	return set;
+}
+
 void ioasid_set_get_locked(struct ioasid_set *set)
 {
 	if (WARN_ON(xa_load(&ioasid_sets, set->sid) != set)) {
@@ -638,6 +735,8 @@ EXPORT_SYMBOL_GPL(ioasid_set_get);
 
 void ioasid_set_put_locked(struct ioasid_set *set)
 {
+	struct ioasid_nb_args args = { 0 };
+	struct ioasid_set_nb *curr;
 	struct ioasid_data *entry;
 	unsigned long index;
 
@@ -673,8 +772,24 @@ void ioasid_set_put_locked(struct ioasid_set *set)
 done_destroy:
 	/* Return the quota back to system pool */
 	ioasid_capacity_avail += set->quota;
-	kfree_rcu(set, rcu);
 
+	/* Restore pending status of the set NBs */
+	spin_lock(&ioasid_nb_lock);
+	list_for_each_entry(curr, &ioasid_nb_pending_list, list) {
+		if (curr->token == set->token) {
+			if (curr->active)
+				curr->active = false;
+			else
+				pr_warn("Set token exists but not active!\n");
+		}
+	}
+	spin_unlock(&ioasid_nb_lock);
+
+	args.set = set;
+	atomic_notifier_call_chain(&ioasid_chain, IOASID_SET_FREE, &args);
+
+	kfree_rcu(set, rcu);
+	pr_debug("Set freed %d\n", set->sid);
 	/*
 	 * Token got released right away after the ioasid_set is freed.
 	 * If a new set is created immediately with the newly released token,
@@ -927,6 +1042,167 @@ void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid,
 }
 EXPORT_SYMBOL_GPL(ioasid_find);
 
+int ioasid_register_notifier(struct ioasid_set *set, struct notifier_block *nb)
+{
+	if (set)
+		return atomic_notifier_chain_register(&set->nh, nb);
+	else
+		return atomic_notifier_chain_register(&ioasid_chain, nb);
+}
+EXPORT_SYMBOL_GPL(ioasid_register_notifier);
+
+void ioasid_unregister_notifier(struct ioasid_set *set,
+				struct notifier_block *nb)
+{
+	struct ioasid_set_nb *curr;
+
+	spin_lock(&ioasid_nb_lock);
+	/*
+	 * Pending list is registered with a token without an ioasid_set,
+	 * therefore should not be unregistered directly.
+	 */
+	list_for_each_entry(curr, &ioasid_nb_pending_list, list) {
+		if (curr->nb == nb) {
+			pr_warn("Cannot unregister NB from pending list\n");
+			spin_unlock(&ioasid_nb_lock);
+			return;
+		}
+	}
+	spin_unlock(&ioasid_nb_lock);
+
+	if (set)
+		atomic_notifier_chain_unregister(&set->nh, nb);
+	else
+		atomic_notifier_chain_unregister(&ioasid_chain, nb);
+}
+EXPORT_SYMBOL_GPL(ioasid_unregister_notifier);
+
+int ioasid_register_notifier_mm(struct mm_struct *mm, struct notifier_block *nb)
+{
+	struct ioasid_set_nb *curr;
+	struct ioasid_set *set;
+	int ret = 0;
+
+	if (!mm)
+		return -EINVAL;
+
+	spin_lock(&ioasid_nb_lock);
+
+	/* Check for duplicates, nb is unique per set */
+	list_for_each_entry(curr, &ioasid_nb_pending_list, list) {
+		if (curr->token == mm && curr->nb == nb) {
+			ret = -EBUSY;
+			goto exit_unlock;
+		}
+	}
+
+	/* Check if the token has an existing set */
+	set = ioasid_find_mm_set(mm);
+	if (IS_ERR_OR_NULL(set)) {
+		/* Add to the rsvd list as inactive */
+		curr->active = false;
+	} else {
+		/* REVISIT: Only register empty set for now. Can add an option
+		 * in the future to playback existing PASIDs.
+		 */
+		if (set->nr_ioasids) {
+			pr_warn("IOASID set %d not empty\n", set->sid);
+			ret = -EBUSY;
+			goto exit_unlock;
+		}
+		curr = kzalloc(sizeof(*curr), GFP_ATOMIC);
+		if (!curr) {
+			ret = -ENOMEM;
+			goto exit_unlock;
+		}
+		curr->token = mm;
+		curr->nb = nb;
+		curr->active = true;
+		curr->set = set;
+
+		/* Set already created, add to the notifier chain */
+		atomic_notifier_chain_register(&set->nh, nb);
+		/*
+		 * Do not hold a reference, if the set gets destroyed, the nb
+		 * entry will be marked inactive.
+		 */
+		ioasid_set_put(set);
+	}
+
+	list_add(&curr->list, &ioasid_nb_pending_list);
+
+exit_unlock:
+	spin_unlock(&ioasid_nb_lock);
+	return ret;
+}
+EXPORT_SYMBOL_GPL(ioasid_register_notifier_mm);
+
+void ioasid_unregister_notifier_mm(struct mm_struct *mm, struct notifier_block *nb)
+{
+	struct ioasid_set_nb *curr;
+
+	spin_lock(&ioasid_nb_lock);
+	list_for_each_entry(curr, &ioasid_nb_pending_list, list) {
+		if (curr->token == mm && curr->nb == nb) {
+			list_del(&curr->list);
+			goto exit_free;
+		}
+	}
+	pr_warn("No ioasid set found for mm token %llx\n",  (u64)mm);
+	goto done_unlock;
+
+exit_free:
+	if (curr->active) {
+		pr_debug("mm set active, unregister %llx\n",
+			(u64)mm);
+		atomic_notifier_chain_unregister(&curr->set->nh, nb);
+	}
+	kfree(curr);
+done_unlock:
+	spin_unlock(&ioasid_nb_lock);
+	return;
+}
+EXPORT_SYMBOL_GPL(ioasid_unregister_notifier_mm);
+
+/**
+ * ioasid_notify - Send notification on a given IOASID for status change.
+ *                 Used by publishers when the status change may affect
+ *                 subscriber's internal state.
+ *
+ * @ioasid:	The IOASID to which the notification will send
+ * @cmd:	The notification event
+ * @flags:	Special instructions, e.g. notify with a set or global
+ */
+int ioasid_notify(ioasid_t ioasid, enum ioasid_notify_val cmd, unsigned int flags)
+{
+	struct ioasid_data *ioasid_data;
+	struct ioasid_nb_args args = { 0 };
+	int ret = 0;
+
+	spin_lock(&ioasid_allocator_lock);
+	ioasid_data = xa_load(&active_allocator->xa, ioasid);
+	if (!ioasid_data) {
+		pr_err("Trying to notify unknown IOASID %u\n", ioasid);
+		spin_unlock(&ioasid_allocator_lock);
+		return -EINVAL;
+	}
+
+	args.id = ioasid;
+	args.set = ioasid_data->set;
+	args.pdata = ioasid_data->private;
+	args.spid = ioasid_data->spid;
+	if (flags & IOASID_NOTIFY_ALL) {
+		ret = atomic_notifier_call_chain(&ioasid_chain, cmd, &args);
+	} else if (flags & IOASID_NOTIFY_SET) {
+		ret = atomic_notifier_call_chain(&ioasid_data->set->nh,
+						 cmd, &args);
+	}
+	spin_unlock(&ioasid_allocator_lock);
+
+	return ret;
+}
+EXPORT_SYMBOL_GPL(ioasid_notify);
+
 MODULE_AUTHOR("Jean-Philippe Brucker <[email protected]>");
 MODULE_AUTHOR("Jacob Pan <[email protected]>");
 MODULE_DESCRIPTION("IO Address Space ID (IOASID) allocator");
include/linux/ioasid.h +70 -0

108 lines not on this page · raw message

Signature · 1 line
2.7.4
author
#52 lore raw

PATCH v2 8/9 iommu/vt-d: Send IOASID bind/unbind notifications

In reply to Jacob Pan

On Intel Scalable I/O Virtualization (SIOV) enabled platforms with ENQCMD in use by the guest, KVM must establish a guest-host PASID translation table prior to the issuing of ENQCMD. PASID translation table also depends on the IOMMU PASID entry, which is configured during the guest page table bind time. This patch adds a notification event to the IOMMU driver such that KVM can be notified when a new guest-host PASID mapping is established or demolished.

---
 drivers/iommu/intel/svm.c | 7 +++++++
 1 file changed, 7 insertions(+)
drivers/iommu/intel/svm.c +7 -0
index 600e3ae5b656..d8a5efa75095 100644
@@ -471,6 +471,11 @@ int intel_svm_bind_gpasid(struct iommu_domain *domain, struct device *dev,
 	}
 
 	mutex_unlock(&pasid_mutex);
+	/*
+	 * Notify KVM new host-guest PASID bind is ready. KVM will set up
+	 * PASID translation table to support guest ENQCMD.
+	 */
+	ioasid_notify(data->hpasid, IOASID_BIND, IOASID_NOTIFY_SET);
 	return ret;
 }
 
@@ -510,6 +515,8 @@ int intel_svm_unbind_gpasid(struct device *dev, int pasid)
 				 * and perform cleanup.
 				 */
 				ioasid_attach_data(pasid, NULL);
+				ioasid_notify(pasid, IOASID_UNBIND,
+					IOASID_NOTIFY_SET);
 				kfree(svm);
 			}
 		}
Signature · 1 line
2.7.4
author
#10 lore raw

PATCH v2 2/9 iommu/ioasid: Rename ioasid_set_data()

In reply to Jacob Pan

Rename ioasid_set_data() to ioasid_attach_data() to avoid confusion with struct ioasid_set. ioasid_set is a group of IOASIDs that share a common token.

---
 drivers/iommu/intel/svm.c | 6 +++---
 drivers/iommu/ioasid.c    | 6 +++---
 include/linux/ioasid.h    | 4 ++--
 3 files changed, 8 insertions(+), 8 deletions(-)
drivers/iommu/intel/svm.c +3 -3
index b6972dca2ae0..37a9beabc0ca 100644
@@ -342,7 +342,7 @@ int intel_svm_bind_gpasid(struct iommu_domain *domain, struct device *dev,
 			svm->gpasid = data->gpasid;
 			svm->flags |= SVM_FLAG_GUEST_PASID;
 		}
-		ioasid_set_data(data->hpasid, svm);
+		ioasid_attach_data(data->hpasid, svm);
 		INIT_LIST_HEAD_RCU(&svm->devs);
 		mmput(svm->mm);
 	}
@@ -394,7 +394,7 @@ int intel_svm_bind_gpasid(struct iommu_domain *domain, struct device *dev,
 	list_add_rcu(&sdev->list, &svm->devs);
  out:
 	if (!IS_ERR_OR_NULL(svm) && list_empty(&svm->devs)) {
-		ioasid_set_data(data->hpasid, NULL);
+		ioasid_attach_data(data->hpasid, NULL);
 		kfree(svm);
 	}
 
@@ -437,7 +437,7 @@ int intel_svm_unbind_gpasid(struct device *dev, int pasid)
 				 * the unbind, IOMMU driver will get notified
 				 * and perform cleanup.
 				 */
-				ioasid_set_data(pasid, NULL);
+				ioasid_attach_data(pasid, NULL);
 				kfree(svm);
 			}
 		}
drivers/iommu/ioasid.c +3 -3
index 0f8dd377aada..5f63af07acd5 100644
@@ -258,14 +258,14 @@ void ioasid_unregister_allocator(struct ioasid_allocator_ops *ops)
 EXPORT_SYMBOL_GPL(ioasid_unregister_allocator);
 
 /**
- * ioasid_set_data - Set private data for an allocated ioasid
+ * ioasid_attach_data - Set private data for an allocated ioasid
  * @ioasid: the ID to set data
  * @data:   the private data
  *
  * For IOASID that is already allocated, private data can be set
  * via this API. Future lookup can be done via ioasid_find.
  */
-int ioasid_set_data(ioasid_t ioasid, void *data)
+int ioasid_attach_data(ioasid_t ioasid, void *data)
 {
 	struct ioasid_data *ioasid_data;
 	int ret = 0;
@@ -287,7 +287,7 @@ int ioasid_set_data(ioasid_t ioasid, void *data)
 
 	return ret;
 }
-EXPORT_SYMBOL_GPL(ioasid_set_data);
+EXPORT_SYMBOL_GPL(ioasid_attach_data);
 
 /**
  * ioasid_alloc - Allocate an IOASID
include/linux/ioasid.h +2 -2
index 6f000d7a0ddc..9c44947a68c8 100644
@@ -39,7 +39,7 @@ void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid,
 		  bool (*getter)(void *));
 int ioasid_register_allocator(struct ioasid_allocator_ops *allocator);
 void ioasid_unregister_allocator(struct ioasid_allocator_ops *allocator);
-int ioasid_set_data(ioasid_t ioasid, void *data);
+int ioasid_attach_data(ioasid_t ioasid, void *data);
 
 #else /* !CONFIG_IOASID */
 static inline ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min,
@@ -67,7 +67,7 @@ static inline void ioasid_unregister_allocator(struct ioasid_allocator_ops *allo
 {
 }
 
-static inline int ioasid_set_data(ioasid_t ioasid, void *data)
+static inline int ioasid_attach_data(ioasid_t ioasid, void *data)
 {
 	return -ENOTSUPP;
 }
Signature · 1 line
2.7.4
author
#49 lore raw

PATCH v2 7/9 iommu/vt-d: Listen to IOASID notifications

In reply to Jacob Pan

On Intel Scalable I/O Virtualization (SIOV) enabled platforms, IOMMU driver is one of the users of IOASIDs. In normal flow, callers will perform IOASID allocation, bind, unbind, and free in order. However, for guest SVA, IOASID free could come before unbind as guest is untrusted. This patch registers IOASID notification handler such that IOMMU driver can perform PASID teardown upon receiving an unexpected IOASID free event.

---
 drivers/iommu/intel/svm.c   | 74 ++++++++++++++++++++++++++++++++++++++++++++-
 include/linux/intel-iommu.h |  2 ++
 2 files changed, 75 insertions(+), 1 deletion(-)
drivers/iommu/intel/svm.c +73 -1
index 634e191ca2c3..600e3ae5b656 100644
@@ -95,6 +95,72 @@ static inline bool intel_svm_capable(struct intel_iommu *iommu)
 	return iommu->flags & VTD_FLAG_SVM_CAPABLE;
 }
 
+#define pasid_lock_held() lock_is_held(&pasid_mutex.dep_map)
+static DEFINE_MUTEX(pasid_mutex);
+
+static void intel_svm_free_async_fn(struct work_struct *work)
+{
+	struct intel_svm *svm = container_of(work, struct intel_svm, work);
+	struct intel_svm_dev *sdev;
+
+	/*
+	 * Unbind all devices associated with this PASID which is
+	 * being freed by other users such as VFIO.
+	 */
+	mutex_lock(&pasid_mutex);
+	list_for_each_entry_rcu(sdev, &svm->devs, list, pasid_lock_held()) {
+		/* Does not poison forward pointer */
+		list_del_rcu(&sdev->list);
+		spin_lock(&svm->iommu->lock);
+		intel_pasid_tear_down_entry(svm->iommu, sdev->dev,
+					svm->pasid, true);
+		spin_unlock(&svm->iommu->lock);
+		kfree_rcu(sdev, rcu);
+		/*
+		 * Free before unbind only happens with guest usaged
+		 * host PASIDs. IOASID free will detach private data
+		 * and free the IOASID entry.
+		 */
+		ioasid_put(NULL, svm->pasid);
+		if (list_empty(&svm->devs))
+			kfree(svm);
+	}
+	mutex_unlock(&pasid_mutex);
+}
+
+
+static int pasid_status_change(struct notifier_block *nb,
+				unsigned long code, void *data)
+{
+	struct ioasid_nb_args *args = (struct ioasid_nb_args *)data;
+	struct intel_svm *svm = (struct intel_svm *)args->pdata;
+	int ret = NOTIFY_DONE;
+
+	if (code == IOASID_FREE) {
+		if (!svm)
+			goto done;
+		if (args->id != svm->pasid) {
+			pr_warn("Notify PASID does not match data %d : %d\n",
+				args->id, svm->pasid);
+			goto done;
+		}
+		schedule_work(&svm->work);
+		return NOTIFY_OK;
+	}
+done:
+	return ret;
+}
+
+static struct notifier_block pasid_nb = {
+	.notifier_call = pasid_status_change,
+};
+
+void intel_svm_add_pasid_notifier(void)
+{
+	/* Listen to all PASIDs, not specific to a set */
+	ioasid_register_notifier(NULL, &pasid_nb);
+}
+
 void intel_svm_check(struct intel_iommu *iommu)
 {
 	if (!pasid_supported(iommu))
@@ -221,7 +287,6 @@ static const struct mmu_notifier_ops intel_mmuops = {
 	.invalidate_range = intel_invalidate_range,
 };
 
-static DEFINE_MUTEX(pasid_mutex);
 static LIST_HEAD(global_svm_list);
 
 #define for_each_svm_dev(sdev, svm, d)			\
@@ -342,7 +407,14 @@ int intel_svm_bind_gpasid(struct iommu_domain *domain, struct device *dev,
 			svm->gpasid = data->gpasid;
 			svm->flags |= SVM_FLAG_GUEST_PASID;
 		}
+		svm->iommu = iommu;
+		/*
+		 * Set up cleanup async work in case IOASID core notify us PASID
+		 * is freed before unbind.
+		 */
+		INIT_WORK(&svm->work, intel_svm_free_async_fn);
 		ioasid_attach_data(data->hpasid, svm);
+		ioasid_get(NULL, svm->pasid);
 		INIT_LIST_HEAD_RCU(&svm->devs);
 		mmput(svm->mm);
 	}
include/linux/intel-iommu.h +2 -0
index b1ed2f25f7c0..d36038e6ae0b 100644
@@ -744,6 +744,7 @@ void intel_svm_unbind(struct iommu_sva *handle);
 int intel_svm_get_pasid(struct iommu_sva *handle);
 int intel_svm_page_response(struct device *dev, struct iommu_fault_event *evt,
 			    struct iommu_page_response *msg);
+void intel_svm_add_pasid_notifier(void);
 
 struct svm_dev_ops;
 
@@ -770,6 +771,7 @@ struct intel_svm {
 	int gpasid; /* In case that guest PASID is different from host PASID */
 	struct list_head devs;
 	struct list_head list;
+	struct work_struct work; /* For deferred clean up */
 };
 #else
 static inline void intel_svm_check(struct intel_iommu *iommu) {}
Signature · 1 line
2.7.4
author
#53 lore raw

PATCH v2 9/9 iommu/vt-d: Store guest PASID during bind

In reply to Jacob Pan

IOASID core maintains the guest-host mapping in the form of SPID and IOASID. This patch assigns the guest PASID (if valid) as SPID while binding guest page table with a host PASID. This mapping will be used for lookup and notifications.

---
 drivers/iommu/intel/svm.c | 2 ++
 1 file changed, 2 insertions(+)
drivers/iommu/intel/svm.c +2 -0
index d8a5efa75095..4c958b1aec4c 100644
@@ -406,6 +406,7 @@ int intel_svm_bind_gpasid(struct iommu_domain *domain, struct device *dev,
 		if (data->flags & IOMMU_SVA_GPASID_VAL) {
 			svm->gpasid = data->gpasid;
 			svm->flags |= SVM_FLAG_GUEST_PASID;
+			ioasid_attach_spid(data->hpasid, data->gpasid);
 		}
 		svm->iommu = iommu;
 		/*
@@ -517,6 +518,7 @@ int intel_svm_unbind_gpasid(struct device *dev, int pasid)
 				ioasid_attach_data(pasid, NULL);
 				ioasid_notify(pasid, IOASID_UNBIND,
 					IOASID_NOTIFY_SET);
+				ioasid_attach_spid(pasid, INVALID_IOASID);
 				kfree(svm);
 			}
 		}
Signature · 1 line
2.7.4
#36 lore raw

Re: [PATCH v2 5/9] iommu/ioasid: Introduce ioasid_set private ID

In reply to Jacob Pan
I love your patch! Yet something to improve:

Hi Jacob,

I love your patch! Yet something to improve:

[auto build test ERROR on iommu/next]
[also build test ERROR on linux/master linus/master v5.9-rc1 next-20200821] [If your patch is applied to the wrong git tree, kindly drop us a note. And when submitting patch, we suggest to use '--base' as documented in
https://git-scm.com/docs/git-format-patch]

url:    https://github.com/0day-ci/linux/commits/Jacob-Pan/IOASID-extensions-for-guest-SVA/20200822-123111
base:   https://git.kernel.org/pub/scm/linux/kernel/git/joro/iommu.git next
config: ia64-randconfig-r003-20200822 (attached as .config)
compiler: ia64-linux-gcc (GCC) 9.3.0

reproduce (this is a W=1 build):
wget https://raw.githubusercontent.com/intel/lkp-tests/master/sbin/make.cross -O ~/bin/make.cross chmod +x ~/bin/make.cross
# save the attached .config to linux build tree
COMPILER_INSTALL_PATH=$HOME/0day COMPILER=gcc-9.3.0 make.cross ARCH=ia64

If you fix the issue, kindly add following tag as appropriate

All errors (new ones prefixed by >>):

In file included from include/linux/iommu.h:16,
from include/linux/of_iommu.h:6, from drivers/of/device.c:7:

>> include/linux/ioasid.h:141:6: error: expected ';' before 'inline'
     141 | staic inline int ioasid_attach_spid(ioasid_t ioasid, ioasid_t spid)
         |      ^
         |      ;
--
   In file included from include/linux/iommu.h:16,
                    from drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c:34:
>> include/linux/ioasid.h:141:6: error: expected ';' before 'inline'
     141 | staic inline int ioasid_attach_spid(ioasid_t ioasid, ioasid_t spid)
         |      ^
         |      ;
   drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c:1094:5: warning: no previous prototype for 'amdgpu_ttm_gart_bind' [-Wmissing-prototypes]
    1094 | int amdgpu_ttm_gart_bind(struct amdgpu_device *adev,
         |     ^~~~~~~~~~~~~~~~~~~~
   In file included from drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c:55:
   drivers/gpu/drm/amd/amdgpu/amdgpu.h:190:18: warning: 'sched_policy' defined but not used [-Wunused-const-variable=]
     190 | static const int sched_policy = KFD_SCHED_POLICY_HWS;
         |                  ^~~~~~~~~~~~
   In file included from drivers/gpu/drm/amd/amdgpu/../display/dc/dc_types.h:33,
                    from drivers/gpu/drm/amd/amdgpu/../display/dc/dm_services_types.h:30,
                    from drivers/gpu/drm/amd/amdgpu/../include/dm_pp_interface.h:26,
                    from drivers/gpu/drm/amd/amdgpu/amdgpu.h:65,
                    from drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c:55:
   drivers/gpu/drm/amd/amdgpu/../display/include/fixed31_32.h:76:32: warning: 'dc_fixpt_ln2_div_2' defined but not used [-Wunused-const-variable=]
      76 | static const struct fixed31_32 dc_fixpt_ln2_div_2 = { 1488522236LL };
         |                                ^~~~~~~~~~~~~~~~~~
   drivers/gpu/drm/amd/amdgpu/../display/include/fixed31_32.h:75:32: warning: 'dc_fixpt_ln2' defined but not used [-Wunused-const-variable=]
      75 | static const struct fixed31_32 dc_fixpt_ln2 = { 2977044471LL };
         |                                ^~~~~~~~~~~~
   drivers/gpu/drm/amd/amdgpu/../display/include/fixed31_32.h:74:32: warning: 'dc_fixpt_e' defined but not used [-Wunused-const-variable=]
      74 | static const struct fixed31_32 dc_fixpt_e = { 11674931555LL };
         |                                ^~~~~~~~~~
   drivers/gpu/drm/amd/amdgpu/../display/include/fixed31_32.h:73:32: warning: 'dc_fixpt_two_pi' defined but not used [-Wunused-const-variable=]
      73 | static const struct fixed31_32 dc_fixpt_two_pi = { 26986075409LL };
         |                                ^~~~~~~~~~~~~~~
   drivers/gpu/drm/amd/amdgpu/../display/include/fixed31_32.h:72:32: warning: 'dc_fixpt_pi' defined but not used [-Wunused-const-variable=]
      72 | static const struct fixed31_32 dc_fixpt_pi = { 13493037705LL };
         |                                ^~~~~~~~~~~
   drivers/gpu/drm/amd/amdgpu/../display/include/fixed31_32.h:67:32: warning: 'dc_fixpt_zero' defined but not used [-Wunused-const-variable=]
      67 | static const struct fixed31_32 dc_fixpt_zero = { 0 };
         |                                ^~~~~~~~~~~~~

--
In file included from include/linux/iommu.h:16,
from drivers/gpu/drm/nouveau/include/nvif/os.h:30, from drivers/gpu/drm/nouveau/include/nvkm/core/os.h:4, from drivers/gpu/drm/nouveau/include/nvkm/core/oclass.h:3, from drivers/gpu/drm/nouveau/include/nvkm/core/device.h:4, from drivers/gpu/drm/nouveau/include/nvkm/core/subdev.h:4, from drivers/gpu/drm/nouveau/nvkm/nvfw/acr.c:22:

>> include/linux/ioasid.h:141:6: error: expected ';' before 'inline'
     141 | staic inline int ioasid_attach_spid(ioasid_t ioasid, ioasid_t spid)
         |      ^
         |      ;

drivers/gpu/drm/nouveau/nvkm/nvfw/acr.c:49:1: warning: no previous prototype for 'lsb_header_tail_dump' [-Wmissing-prototypes]
49 | lsb_header_tail_dump(struct nvkm_subdev *subdev,
| ^~~~~~~~~~~~~~~~~~~~
--
In file included from include/linux/iommu.h:16,

                    from drivers/gpu/drm/nouveau/include/nvif/os.h:30,
                    from drivers/gpu/drm/nouveau/include/nvkm/core/os.h:4,
                    from drivers/gpu/drm/nouveau/include/nvkm/core/oclass.h:3,
                    from drivers/gpu/drm/nouveau/include/nvkm/core/device.h:4,
                    from drivers/gpu/drm/nouveau/include/nvkm/core/subdev.h:4,
                    from drivers/gpu/drm/nouveau/include/nvkm/subdev/mc.h:4,
                    from drivers/gpu/drm/nouveau/nvkm/subdev/mc/priv.h:5,
                    from drivers/gpu/drm/nouveau/nvkm/subdev/mc/gp10b.c:24:
>> include/linux/ioasid.h:141:6: error: expected ';' before 'inline'
     141 | staic inline int ioasid_attach_spid(ioasid_t ioasid, ioasid_t spid)
         |      ^
         |      ;
   drivers/gpu/drm/nouveau/nvkm/subdev/mc/gp10b.c:27:1: warning: no previous prototype for 'gp10b_mc_init' [-Wmissing-prototypes]
      27 | gp10b_mc_init(struct nvkm_mc *mc)
         | ^~~~~~~~~~~~~

--
In file included from include/linux/iommu.h:16,
from drivers/gpu/drm/nouveau/include/nvif/os.h:30, from drivers/gpu/drm/nouveau/include/nvkm/core/os.h:4, from drivers/gpu/drm/nouveau/include/nvkm/core/oclass.h:3, from drivers/gpu/drm/nouveau/include/nvkm/core/object.h:4, from drivers/gpu/drm/nouveau/nvkm/subdev/mmu/ummu.h:4, from drivers/gpu/drm/nouveau/nvkm/subdev/mmu/base.c:24:

>> include/linux/ioasid.h:141:6: error: expected ';' before 'inline'
     141 | staic inline int ioasid_attach_spid(ioasid_t ioasid, ioasid_t spid)
         |      ^
         |      ;

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/base.c:65:1: warning: no previous prototype for 'nvkm_mmu_ptp_get' [-Wmissing-prototypes]
65 | nvkm_mmu_ptp_get(struct nvkm_mmu *mmu, u32 size, bool zero)
| ^~~~~~~~~~~~~~~~
--
In file included from include/linux/iommu.h:16,

                    from drivers/gpu/drm/nouveau/include/nvif/os.h:30,
                    from drivers/gpu/drm/nouveau/include/nvkm/core/os.h:4,
                    from drivers/gpu/drm/nouveau/include/nvkm/core/oclass.h:3,
                    from drivers/gpu/drm/nouveau/include/nvkm/core/device.h:4,
                    from drivers/gpu/drm/nouveau/include/nvkm/core/subdev.h:4,
                    from drivers/gpu/drm/nouveau/include/nvkm/subdev/mmu.h:4,
                    from drivers/gpu/drm/nouveau/nvkm/subdev/mmu/priv.h:5,
                    from drivers/gpu/drm/nouveau/nvkm/subdev/mmu/mem.h:3,
                    from drivers/gpu/drm/nouveau/nvkm/subdev/mmu/tu102.c:23:
>> include/linux/ioasid.h:141:6: error: expected ';' before 'inline'
     141 | staic inline int ioasid_attach_spid(ioasid_t ioasid, ioasid_t spid)
         |      ^
         |      ;
   drivers/gpu/drm/nouveau/nvkm/subdev/mmu/tu102.c:31:1: warning: no previous prototype for 'tu102_mmu_kind' [-Wmissing-prototypes]
      31 | tu102_mmu_kind(struct nvkm_mmu *mmu, int *count, u8 *invalid)
         | ^~~~~~~~~~~~~~
--
   In file included from include/linux/iommu.h:16,
                    from drivers/gpu/drm/nouveau/include/nvif/os.h:30,
                    from drivers/gpu/drm/nouveau/include/nvkm/core/os.h:4,
                    from drivers/gpu/drm/nouveau/include/nvkm/core/oclass.h:3,
                    from drivers/gpu/drm/nouveau/include/nvkm/core/device.h:4,
                    from drivers/gpu/drm/nouveau/include/nvkm/core/subdev.h:4,
                    from drivers/gpu/drm/nouveau/include/nvkm/subdev/therm.h:4,
                    from drivers/gpu/drm/nouveau/nvkm/subdev/therm/priv.h:27,
                    from drivers/gpu/drm/nouveau/nvkm/subdev/therm/gt215.c:24:
>> include/linux/ioasid.h:141:6: error: expected ';' before 'inline'
     141 | staic inline int ioasid_attach_spid(ioasid_t ioasid, ioasid_t spid)
         |      ^
         |      ;
   drivers/gpu/drm/nouveau/nvkm/subdev/therm/gt215.c:40:1: warning: no previous prototype for 'gt215_therm_init' [-Wmissing-prototypes]
      40 | gt215_therm_init(struct nvkm_therm *therm)
         | ^~~~~~~~~~~~~~~~
--
   In file included from include/linux/iommu.h:16,
                    from drivers/gpu/drm/nouveau/include/nvif/os.h:30,
                    from drivers/gpu/drm/nouveau/include/nvkm/core/os.h:4,
                    from drivers/gpu/drm/nouveau/include/nvkm/core/oclass.h:3,
                    from drivers/gpu/drm/nouveau/include/nvkm/core/device.h:4,
                    from drivers/gpu/drm/nouveau/include/nvkm/core/subdev.h:4,
                    from drivers/gpu/drm/nouveau/include/nvkm/core/engine.h:5,
                    from drivers/gpu/drm/nouveau/include/nvkm/engine/gr.h:4,
                    from drivers/gpu/drm/nouveau/nvkm/engine/gr/priv.h:5,
                    from drivers/gpu/drm/nouveau/nvkm/engine/gr/gf100.h:27,
                    from drivers/gpu/drm/nouveau/nvkm/engine/gr/gf100.c:24:
>> include/linux/ioasid.h:141:6: error: expected ';' before 'inline'
     141 | staic inline int ioasid_attach_spid(ioasid_t ioasid, ioasid_t spid)
         |      ^
         |      ;
   drivers/gpu/drm/nouveau/nvkm/engine/gr/gf100.c:745:1: warning: no previous prototype for 'gf100_gr_fecs_start_ctxsw' [-Wmissing-prototypes]
     745 | gf100_gr_fecs_start_ctxsw(struct nvkm_gr *base)
         | ^~~~~~~~~~~~~~~~~~~~~~~~~
   drivers/gpu/drm/nouveau/nvkm/engine/gr/gf100.c:760:1: warning: no previous prototype for 'gf100_gr_fecs_stop_ctxsw' [-Wmissing-prototypes]
     760 | gf100_gr_fecs_stop_ctxsw(struct nvkm_gr *base)
         | ^~~~~~~~~~~~~~~~~~~~~~~~
   drivers/gpu/drm/nouveau/nvkm/engine/gr/gf100.c:2036:1: warning: no previous prototype for 'gf100_gr_dtor' [-Wmissing-prototypes]
    2036 | gf100_gr_dtor(struct nvkm_gr *base)
         | ^~~~~~~~~~~~~
--
   In file included from include/linux/iommu.h:16,
                    from drivers/gpu/drm/nouveau/include/nvif/os.h:30,
                    from drivers/gpu/drm/nouveau/include/nvkm/core/os.h:4,
                    from drivers/gpu/drm/nouveau/include/nvkm/core/oclass.h:3,
                    from drivers/gpu/drm/nouveau/include/nvkm/core/device.h:4,
                    from drivers/gpu/drm/nouveau/include/nvkm/core/subdev.h:4,
                    from drivers/gpu/drm/nouveau/include/nvkm/core/engine.h:5,
                    from drivers/gpu/drm/nouveau/include/nvkm/engine/gr.h:4,
                    from drivers/gpu/drm/nouveau/nvkm/engine/gr/priv.h:5,
                    from drivers/gpu/drm/nouveau/nvkm/engine/gr/gf100.h:27,
                    from drivers/gpu/drm/nouveau/nvkm/engine/gr/gk20a.c:22:
>> include/linux/ioasid.h:141:6: error: expected ';' before 'inline'
     141 | staic inline int ioasid_attach_spid(ioasid_t ioasid, ioasid_t spid)
         |      ^
         |      ;

drivers/gpu/drm/nouveau/nvkm/engine/gr/gk20a.c:37:1: warning: no previous prototype for 'gk20a_gr_av_to_init' [-Wmissing-prototypes]
37 | gk20a_gr_av_to_init(struct gf100_gr *gr, const char *path, const char *name,
| ^~~~~~~~~~~~~~~~~~~
drivers/gpu/drm/nouveau/nvkm/engine/gr/gk20a.c:87:1: warning: no previous prototype for 'gk20a_gr_aiv_to_init' [-Wmissing-prototypes]
87 | gk20a_gr_aiv_to_init(struct gf100_gr *gr, const char *path, const char *name,
| ^~~~~~~~~~~~~~~~~~~~
drivers/gpu/drm/nouveau/nvkm/engine/gr/gk20a.c:130:1: warning: no previous prototype for 'gk20a_gr_av_to_method' [-Wmissing-prototypes]
130 | gk20a_gr_av_to_method(struct gf100_gr *gr, const char *path, const char *name,
| ^~~~~~~~~~~~~~~~~~~~~
--
In file included from include/linux/iommu.h:16,
from drivers/gpu/drm/nouveau/include/nvif/os.h:30,
from drivers/gpu/drm/nouveau/include/nvif/object.h:5,
from drivers/gpu/drm/nouveau/include/nvif/client.h:5,
from drivers/gpu/drm/nouveau/nouveau_drv.h:43,
from drivers/gpu/drm/nouveau/nouveau_display.h:5,
from drivers/gpu/drm/nouveau/nouveau_fbcon.h:32,
from drivers/gpu/drm/nouveau/nouveau_display.c:38:

>> include/linux/ioasid.h:141:6: error: expected ';' before 'inline'
     141 | staic inline int ioasid_attach_spid(ioasid_t ioasid, ioasid_t spid)
         |      ^
         |      ;
   drivers/gpu/drm/nouveau/nouveau_display.c: In function 'nouveau_framebuffer_new':
   drivers/gpu/drm/nouveau/nouveau_display.c:338:15: warning: variable 'width' set but not used [-Wunused-but-set-variable]
     338 |  unsigned int width, height, i;
         |               ^~~~~

--
In file included from include/linux/iommu.h:16,
from drivers/gpu/drm/nouveau/include/nvif/os.h:30, from drivers/gpu/drm/nouveau/include/nvif/object.h:5, from drivers/gpu/drm/nouveau/include/nvif/mmu.h:3, from drivers/gpu/drm/nouveau/include/nvif/mem.h:3, from drivers/gpu/drm/nouveau/dispnv50/disp.h:3, from drivers/gpu/drm/nouveau/dispnv50/disp.c:24:

>> include/linux/ioasid.h:141:6: error: expected ';' before 'inline'
     141 | staic inline int ioasid_attach_spid(ioasid_t ioasid, ioasid_t spid)
         |      ^
         |      ;
   drivers/gpu/drm/nouveau/dispnv50/disp.c: In function 'nv50_mstm_cleanup':
   drivers/gpu/drm/nouveau/dispnv50/disp.c:1237:6: warning: variable 'ret' set but not used [-Wunused-but-set-variable]
    1237 |  int ret;
         |      ^~~
   drivers/gpu/drm/nouveau/dispnv50/disp.c: In function 'nv50_mstm_prepare':
   drivers/gpu/drm/nouveau/dispnv50/disp.c:1261:6: warning: variable 'ret' set but not used [-Wunused-but-set-variable]
    1261 |  int ret;
         |      ^~~
   drivers/gpu/drm/nouveau/dispnv50/disp.c: At top level:
   drivers/gpu/drm/nouveau/dispnv50/disp.c:2450:1: warning: no previous prototype for 'nv50_display_create' [-Wmissing-prototypes]
    2450 | nv50_display_create(struct drm_device *dev)
         | ^~~~~~~~~~~~~~~~~~~

..

# https://github.com/0day-ci/linux/commit/09f31e901946399a274ce954bdefa4108e895b33
git remote add linux-review https://github.com/0day-ci/linux
git fetch --no-tags linux-review Jacob-Pan/IOASID-extensions-for-guest-SVA/20200822-123111 git checkout 09f31e901946399a274ce954bdefa4108e895b33
vim +141 include/linux/ioasid.h

140

 > 141	staic inline int ioasid_attach_spid(ioasid_t ioasid, ioasid_t spid)

142 {
143 return -ENOTSUPP;
144 }
145

---
0-DAY CI Kernel Test Service, Intel Corporation
https://lists.01.org/hyperkitty/list/kbuild-all(a)lists.01.org

#37 lore raw

Re: [PATCH v2 5/9] iommu/ioasid: Introduce ioasid_set private ID

In reply to Jacob Pan
I love your patch! Yet something to improve:

Hi Jacob,

I love your patch! Yet something to improve:

[auto build test ERROR on iommu/next]
[also build test ERROR on linux/master linus/master v5.9-rc1 next-20200821] [If your patch is applied to the wrong git tree, kindly drop us a note. And when submitting patch, we suggest to use '--base' as documented in
https://git-scm.com/docs/git-format-patch]

url:    https://github.com/0day-ci/linux/commits/Jacob-Pan/IOASID-extensions-for-guest-SVA/20200822-123111
base:   https://git.kernel.org/pub/scm/linux/kernel/git/joro/iommu.git next
config: mips-randconfig-r015-20200822 (attached as .config)
compiler: clang version 12.0.0 (https://github.com/llvm/llvm-project b587ca93be114d07ec3bf654add97d7872325281)

reproduce (this is a W=1 build):
wget https://raw.githubusercontent.com/intel/lkp-tests/master/sbin/make.cross -O ~/bin/make.cross chmod +x ~/bin/make.cross
# install mips cross compiling tool for clang build
# apt-get install binutils-mips-linux-gnu
# save the attached .config to linux build tree
COMPILER_INSTALL_PATH=$HOME/0day COMPILER=clang make.cross ARCH=mips

If you fix the issue, kindly add following tag as appropriate

All errors (new ones prefixed by >>):

In file included from drivers/of/device.c:7:
In file included from include/linux/of_iommu.h:6:
In file included from include/linux/iommu.h:16:

>> include/linux/ioasid.h:141:1: error: unknown type name 'staic'; did you mean 'static'?
   staic inline int ioasid_attach_spid(ioasid_t ioasid, ioasid_t spid)
   ^~~~~
   static
   In file included from drivers/of/device.c:8:
   include/linux/dma-mapping.h:824:9: warning: implicit conversion from 'unsigned long long' to 'unsigned long' changes value from 18446744073709551615 to 4294967295 [-Wconstant-conversion]
           return DMA_BIT_MASK(32);
           ~~~~~~ ^~~~~~~~~~~~~~~~
   include/linux/dma-mapping.h:139:40: note: expanded from macro 'DMA_BIT_MASK'
   #define DMA_BIT_MASK(n) (((n) == 64) ? ~0ULL : ((1ULL<<(n))-1))
                                          ^~~~~

1 warning and 1 error generated.
--
In file included from drivers/of/platform.c:20:
In file included from include/linux/of_iommu.h:6:
In file included from include/linux/iommu.h:16:

>> include/linux/ioasid.h:141:1: error: unknown type name 'staic'; did you mean 'static'?

staic inline int ioasid_attach_spid(ioasid_t ioasid, ioasid_t spid) ^~~~~
static
1 error generated.

# https://github.com/0day-ci/linux/commit/09f31e901946399a274ce954bdefa4108e895b33
git remote add linux-review https://github.com/0day-ci/linux
git fetch --no-tags linux-review Jacob-Pan/IOASID-extensions-for-guest-SVA/20200822-123111 git checkout 09f31e901946399a274ce954bdefa4108e895b33
vim +141 include/linux/ioasid.h

140

 > 141	staic inline int ioasid_attach_spid(ioasid_t ioasid, ioasid_t spid)

142 {
143 return -ENOTSUPP;
144 }
145

---
0-DAY CI Kernel Test Service, Intel Corporation
https://lists.01.org/hyperkitty/list/kbuild-all(a)lists.01.org

#14 lore raw

Re: [PATCH v2 3/9] iommu/ioasid: Introduce ioasid_set APIs

In reply to Jacob Pan
I love your patch! Perhaps something to improve:

Hi Jacob,

I love your patch! Perhaps something to improve:

[auto build test WARNING on iommu/next]
[also build test WARNING on linux/master linus/master v5.9-rc1 next-20200821] [If your patch is applied to the wrong git tree, kindly drop us a note. And when submitting patch, we suggest to use '--base' as documented in
https://git-scm.com/docs/git-format-patch]

url:    https://github.com/0day-ci/linux/commits/Jacob-Pan/IOASID-extensions-for-guest-SVA/20200822-123111
base:   https://git.kernel.org/pub/scm/linux/kernel/git/joro/iommu.git next
config: x86_64-allyesconfig (attached as .config)
compiler: gcc-9 (Debian 9.3.0-15) 9.3.0
reproduce (this is a W=1 build):
        # save the attached .config to linux build tree
        make W=1 ARCH=x86_64 

If you fix the issue, kindly add following tag as appropriate

All warnings (new ones prefixed by >>):

drivers/iommu/ioasid.c: In function 'ioasid_get_capacity':

>> drivers/iommu/ioasid.c:50:10: warning: old-style function definition [-Wold-style-definition]
      50 | ioasid_t ioasid_get_capacity()
         |          ^~~~~~~~~~~~~~~~~~~
   drivers/iommu/ioasid.c: At top level:
>> drivers/iommu/ioasid.c:577:6: warning: no previous prototype for 'ioasid_set_get' [-Wmissing-prototypes]

577 | void ioasid_set_get(struct ioasid_set *set)
| ^~~~~~~~~~~~~~

# https://github.com/0day-ci/linux/commit/59b6f319b27588b2a8a0268a4f4f09f7be458861
git remote add linux-review https://github.com/0day-ci/linux
git fetch --no-tags linux-review Jacob-Pan/IOASID-extensions-for-guest-SVA/20200822-123111 git checkout 59b6f319b27588b2a8a0268a4f4f09f7be458861
vim +50 drivers/iommu/ioasid.c

    49	
  > 50	ioasid_t ioasid_get_capacity()
    51	{
    52		return ioasid_capacity;
    53	}
    54	EXPORT_SYMBOL_GPL(ioasid_get_capacity);
    55	

---
0-DAY CI Kernel Test Service, Intel Corporation
https://lists.01.org/hyperkitty/list/kbuild-all(a)lists.01.org

maintainer
#2 lore raw

Re: [PATCH v2 1/9] docs: Document IO Address Space ID (IOASID) APIs

In reply to Jacob Pan

Hi Jacob,

Quoted 24 lines · Jacob Pan

IOASID is used to identify address spaces that can be targeted by device DMA. It is a system-wide resource that is essential to its many users. This document is an attempt to help developers from all vendors navigate the APIs. At this time, ARM SMMU and Intel’s Scalable IO Virtualization (SIOV) enabled platforms are the primary users of IOASID. Examples of how SIOV components interact with IOASID APIs are provided in that many APIs are driven by the requirements from SIOV.

---
  Documentation/ioasid.rst | 618 +++++++++++++++++++++++++++++++++++++++++++++++
  1 file changed, 618 insertions(+)

create mode 100644 Documentation/ioasid.rst

Documentation/ioasid.rst new file +2 -0
new file mode 100644
index 000000000000..b6a8cdc885ff
@@ -0,0 +1,618 @@
+.. ioasid:
+

Documentation/ioasid.rst line 12

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+=====================================
+IO Address Space ID
+=====================================
+
+IOASID is a generic name for PCIe Process Address ID (PASID) or ARM
+SMMU sub-stream ID. An IOASID identifies an address space that DMA
+requests can target.
+
+The primary use cases for IOASID are Shared Virtual Address (SVA) and
+IO Virtual Address (IOVA). However, the requirements for IOASID

Can you please elaborate a bit more about how ioasid is used by IOVA?

Quoted 12 lines
Documentation/ioasid.rst new file +12 -0
+management can vary among hardware architectures.
+
+This document covers the generic features supported by IOASID
+APIs. Vendor-specific use cases are also illustrated with Intel's VT-d
+based platforms as the first example.
+
+.. contents:: :local:
+
+Glossary
+========
+PASID - Process Address Space ID
+

Documentation/ioasid.rst line 34

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+IOASID - IO Address Space ID (generic term for PCIe PASID and
+sub-stream ID in SMMU)
+
+SVA/SVM - Shared Virtual Addressing/Memory
+
+ENQCMD - New Intel X86 ISA for efficient workqueue submission [1]
+
+DSA - Intel Data Streaming Accelerator [2]
+
+VDCM - Virtual device composition module [3]

Capitalize the first letter of each word.

Quoted 38 lines
Documentation/ioasid.rst new file +38 -0
+
+SIOV - Intel Scalable IO Virtualization
+
+
+Key Concepts
+============
+
+IOASID Set
+-----------
+An IOASID set is a group of IOASIDs allocated from the system-wide
+IOASID pool. An IOASID set is created and can be identified by a
+token of u64. Refer to IOASID set APIs for more details.
+
+IOASID set is particularly useful for guest SVA where each guest could
+have its own IOASID set for security and efficiency reasons.
+
+IOASID Set Private ID (SPID)
+----------------------------
+SPIDs are introduced as IOASIDs within its set. Each SPID maps to a
+system-wide IOASID but the namespace of SPID is within its IOASID
+set. SPIDs can be used as guest IOASIDs where each guest could do
+IOASID allocation from its own pool and map them to host physical
+IOASIDs. SPIDs are particularly useful for supporting live migration
+where decoupling guest and host physical resources are necessary.
+
+For example, two VMs can both allocate guest PASID/SPID #101 but map to
+different host PASIDs #201 and #202 respectively as shown in the
+diagram below.
+::
+
+ .------------------.    .------------------.
+ |   VM 1           |    |   VM 2           |
+ |                  |    |                  |
+ |------------------|    |------------------|
+ | GPASID/SPID 101  |    | GPASID/SPID 101  |
+ '------------------'    -------------------'     Guest
+ __________|______________________|______________________
+           |                      |               Host

Documentation/ioasid.rst line 82

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+           v                      v
+ .------------------.    .------------------.
+ | Host IOASID 201  |    | Host IOASID 202  |
+ '------------------'    '------------------'
+ |   IOASID set 1   |    |   IOASID set 2   |
+ '------------------'    '------------------'
+
+Guest PASID is treated as IOASID set private ID (SPID) within an
+IOASID set, mappings between guest and host IOASIDs are stored in the
+set for inquiry.

Is there a real IOASID set allocated in the host which represent the SPID?

Quoted 118 lines
Documentation/ioasid.rst new file +118 -0
+
+IOASID APIs
+===========
+To get the IOASID APIs, users must #include <linux/ioasid.h>. These APIs
+serve the following functionalities:
+
+  - IOASID allocation/Free
+  - Group management in the form of ioasid_set
+  - Private data storage and lookup
+  - Reference counting
+  - Event notification in case of state change
+
+IOASID Set Level APIs
+--------------------------
+For use cases such as guest SVA it is necessary to manage IOASIDs at
+a group level. For example, VMs may allocate multiple IOASIDs for
+guest process address sharing (vSVA). It is imperative to enforce
+VM-IOASID ownership such that malicious guest cannot target DMA
+traffic outside its own IOASIDs, or free an active IOASID belong to
+another VM.
+::
+
+ struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota, u32 type)
+
+ int ioasid_adjust_set(struct ioasid_set *set, int quota);
+
+ void ioasid_set_get(struct ioasid_set *set)
+
+ void ioasid_set_put(struct ioasid_set *set)
+
+ void ioasid_set_get_locked(struct ioasid_set *set)
+
+ void ioasid_set_put_locked(struct ioasid_set *set)
+
+ int ioasid_set_for_each_ioasid(struct ioasid_set *sdata,
+                                void (*fn)(ioasid_t id, void *data),
+				void *data)
+
+
+IOASID set concept is introduced to represent such IOASID groups. Each
+IOASID set is created with a token which can be one of the following
+types:
+
+ - IOASID_SET_TYPE_NULL (Arbitrary u64 value)
+ - IOASID_SET_TYPE_MM (Set token is a mm_struct)
+
+The explicit MM token type is useful when multiple users of an IOASID
+set under the same process need to communicate about their shared IOASIDs.
+E.g. An IOASID set created by VFIO for one guest can be associated
+with the KVM instance for the same guest since they share a common mm_struct.
+
+The IOASID set APIs serve the following purposes:
+
+ - Ownership/permission enforcement
+ - Take collective actions, e.g. free an entire set
+ - Event notifications within a set
+ - Look up a set based on token
+ - Quota enforcement
+
+Individual IOASID APIs
+----------------------
+Once an ioasid_set is created, IOASIDs can be allocated from the set.
+Within the IOASID set namespace, set private ID (SPID) is supported. In
+the VM use case, SPID can be used for storing guest PASID.
+
+::
+
+ ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max,
+                       void *private);
+
+ int ioasid_get(struct ioasid_set *set, ioasid_t ioasid);
+
+ void ioasid_put(struct ioasid_set *set, ioasid_t ioasid);
+
+ int ioasid_get_locked(struct ioasid_set *set, ioasid_t ioasid);
+
+ void ioasid_put_locked(struct ioasid_set *set, ioasid_t ioasid);
+
+ void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid,
+                   bool (*getter)(void *));
+
+ ioasid_t ioasid_find_by_spid(struct ioasid_set *set, ioasid_t spid)
+
+ int ioasid_attach_data(struct ioasid_set *set, ioasid_t ioasid,
+                        void *data);
+ int ioasid_attach_spid(struct ioasid_set *set, ioasid_t ioasid,
+                        ioasid_t ssid);
+
+
+Notifications
+-------------
+An IOASID may have multiple users, each user may have hardware context
+associated with an IOASID. When the status of an IOASID changes,
+e.g. an IOASID is being freed, users need to be notified such that the
+associated hardware context can be cleared, flushed, and drained.
+
+::
+
+ int ioasid_register_notifier(struct ioasid_set *set, struct
+                              notifier_block *nb)
+
+ void ioasid_unregister_notifier(struct ioasid_set *set,
+                                 struct notifier_block *nb)
+
+ int ioasid_register_notifier_mm(struct mm_struct *mm, struct
+                                 notifier_block *nb)
+
+ void ioasid_unregister_notifier_mm(struct mm_struct *mm, struct
+                                    notifier_block *nb)
+
+ int ioasid_notify(ioasid_t ioasid, enum ioasid_notify_val cmd,
+                   unsigned int flags)
+
+
+Events
+~~~~~~
+Notification events are pertinent to individual IOASIDs, they can be
+one of the following:

Documentation/ioasid.rst line 210

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+
+ - ALLOC
+ - FREE
+ - BIND
+ - UNBIND
+
+Ordering
+~~~~~~~~
+Ordering is supported by IOASID notification priorities as the
+following (in ascending order):

What does ascending order exactly mean here? LAST->IOMMU->DEVICE...?

Quoted 116 lines
Documentation/ioasid.rst new file +116 -0
+
+::
+
+ enum ioasid_notifier_prios {
+	IOASID_PRIO_LAST,
+	IOASID_PRIO_IOMMU,
+	IOASID_PRIO_DEVICE,
+	IOASID_PRIO_CPU,
+ };
+
+The typical use case is when an IOASID is freed due to an exception, DMA
+source should be quiesced before tearing down other hardware contexts
+in the system. This will reduce the churn in handling faults. DMA work
+submission is performed by the CPU which is granted higher priority than
+devices.
+
+
+Scopes
+~~~~~~
+There are two types of notifiers in IOASID core: system-wide and
+ioasid_set-wide.
+
+System-wide notifier is catering for users that need to handle all
+IOASIDs in the system. E.g. The IOMMU driver handles all IOASIDs.
+
+Per ioasid_set notifier can be used by VM specific components such as
+KVM. After all, each KVM instance only cares about IOASIDs within its
+own set.
+
+
+Atomicity
+~~~~~~~~~
+IOASID notifiers are atomic due to spinlocks used inside the IOASID
+core. For tasks cannot be completed in the notifier handler, async work
+can be submitted to complete the work later as long as there is no
+ordering requirement.
+
+Reference counting
+------------------
+IOASID lifecycle management is based on reference counting. Users of
+IOASID intend to align lifecycle with the IOASID need to hold
+reference of the IOASID. IOASID will not be returned to the pool for
+allocation until all references are dropped. Calling ioasid_free()
+will mark the IOASID as FREE_PENDING if the IOASID has outstanding
+reference. ioasid_get() is not allowed once an IOASID is in the
+FREE_PENDING state.
+
+Event notifications are used to inform users of IOASID status change.
+IOASID_FREE event prompts users to drop their references after
+clearing its context.
+
+For example, on VT-d platform when an IOASID is freed, teardown
+actions are performed on KVM, device driver, and IOMMU driver.
+KVM shall register notifier block with::
+
+ static struct notifier_block pasid_nb_kvm = {
+	.notifier_call = pasid_status_change_kvm,
+	.priority      = IOASID_PRIO_CPU,
+ };
+
+VDCM driver shall register notifier block with::
+
+ static struct notifier_block pasid_nb_vdcm = {
+	.notifier_call = pasid_status_change_vdcm,
+	.priority      = IOASID_PRIO_DEVICE,
+ };
+
+In both cases, notifier blocks shall be registered on the IOASID set
+such that *only* events from the matching VM is received.
+
+If KVM attempts to register notifier block before the IOASID set is
+created for the MM token, the notifier block will be placed on a
+pending list inside IOASID core. Once the token matching IOASID set
+is created, IOASID will register the notifier block automatically.
+IOASID core does not replay events for the existing IOASIDs in the
+set. For IOASID set of MM type, notification blocks can be registered
+on empty sets only. This is to avoid lost events.
+
+IOMMU driver shall register notifier block on global chain::
+
+ static struct notifier_block pasid_nb_vtd = {
+	.notifier_call = pasid_status_change_vtd,
+	.priority      = IOASID_PRIO_IOMMU,
+ };
+
+Custom allocator APIs
+---------------------
+
+::
+
+ int ioasid_register_allocator(struct ioasid_allocator_ops *allocator);
+
+ void ioasid_unregister_allocator(struct ioasid_allocator_ops *allocator);
+
+Allocator Choices
+~~~~~~~~~~~~~~~~~
+IOASIDs are allocated for both host and guest SVA/IOVA usage. However,
+allocators can be different. For example, on VT-d guest PASID
+allocation must be performed via a virtual command interface which is
+emulated by VMM.
+
+IOASID core has the notion of "custom allocator" such that guest can
+register virtual command allocator that precedes the default one.
+
+Namespaces
+~~~~~~~~~~
+IOASIDs are limited system resources that default to 20 bits in
+size. Since each device has its own table, theoretically the namespace
+can be per device also. However, for security reasons sharing PASID
+tables among devices are not good for isolation. Therefore, IOASID
+namespace is system-wide.
+
+There are also other reasons to have this simpler system-wide
+namespace. Take VT-d as an example, VT-d supports shared workqueue
+and ENQCMD[1] where one IOASID could be used to submit work on
+multiple devices that are shared with other VMs. This requires IOASID

Documentation/ioasid.rst line 336

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+to be system-wide. This is also the reason why guests must use an
+emulated virtual command interface to allocate IOASID from the host.
+
+
+Life cycle
+==========
+This section covers IOASID lifecycle management for both bare-metal
+and guest usages. In bare-metal SVA, MMU notifier is directly hooked
+up with IOMMU driver, therefore the process address space (MM)
+lifecycle is aligned with IOASID.

MMU notifier for SVA mainly serves IOMMU cache flushes, right? The IOASID life cycle for bare matal SVA is managed by the device driver through the iommu sva api's iommu_sva_(un)bind_device()?

Quoted 47 lines
Documentation/ioasid.rst new file +47 -0
+
+However, guest MMU notifier is not available to host IOMMU driver,
+when guest MM terminates unexpectedly, the events have to go through
+VFIO and IOMMU UAPI to reach host IOMMU driver. There are also more
+parties involved in guest SVA, e.g. on Intel VT-d platform, IOASIDs
+are used by IOMMU driver, KVM, VDCM, and VFIO.
+
+Native IOASID Life Cycle (VT-d Example)
+---------------------------------------
+
+The normal flow of native SVA code with Intel Data Streaming
+Accelerator(DSA) [2] as example:
+
+1. Host user opens accelerator FD, e.g. DSA driver, or uacce;
+2. DSA driver allocate WQ, do sva_bind_device();
+3. IOMMU driver calls ioasid_alloc(), then bind PASID with device,
+   mmu_notifier_get()
+4. DMA starts by DSA driver userspace
+5. DSA userspace close FD
+6. DSA/uacce kernel driver handles FD.close()
+7. DSA driver stops DMA
+8. DSA driver calls sva_unbind_device();
+9. IOMMU driver does unbind, clears PASID context in IOMMU, flush
+   TLBs. mmu_notifier_put() called.
+10. mmu_notifier.release() called, IOMMU SVA code calls ioasid_free()*
+11. The IOASID is returned to the pool, reclaimed.
+
+::
+
+   * With ENQCMD, PASID used on VT-d is not released in mmu_notifier() but
+     mmdrop(). mmdrop comes after FD close. Should not matter.
+     If the user process dies unexpectedly, Step #10 may come before
+     Step #5, in between, all DMA faults discarded. PRQ responded with
+     code INVALID REQUEST.
+
+During the normal teardown, the following three steps would happen in
+order:
+
+1. Device driver stops DMA request
+2. IOMMU driver unbinds PASID and mm, flush all TLBs, drain in-flight
+   requests.
+3. IOASID freed
+
+Exception happens when process terminates *before* device driver stops
+DMA and call IOMMU driver to unbind. The flow of process exists are as
+follows:
+

Documentation/ioasid.rst line 393

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+::
+
+   do_exit() {
+	exit_mm() {
+		mm_put();
+		exit_mmap() {
+			intel_invalidate_range() //mmu notifier
+			tlb_finish_mmu()
+			mmu_notifier_release(mm) {
+				intel_iommu_release() {

intel_mm_release()

Quoted 17 lines
Documentation/ioasid.rst new file +17 -0
+   [2]					intel_iommu_teardown_pasid();
+                                        intel_iommu_flush_tlbs();
+				}
+				// tlb_invalidate_range cb removed
+			}
+			unmap_vmas();
+                        free_pgtables(); // IOMMU cannot walk PGT after this
+		};
+	}
+	exit_files(tsk) {
+		close_files() {
+			dsa_close();
+   [1]			dsa_stop_dma();
+                        intel_svm_unbind_pasid(); //nothing to do
+		}
+	}
+   }

Documentation/ioasid.rst line 420

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+
+   mmdrop() /* some random time later, lazy mm user */ {
+   	mm_free_pgd();
+        destroy_context(mm); {
+   [3]	        ioasid_free();
+	}
+   }
+
+As shown in the list above, step #2 could happen before
+#1. Unrecoverable(UR) faults could happen between #2 and #1.

The VT-d hardware will ignore UR faults due to the setting of FPD bit of the PASID entry. The software won't see UR faults.

Quoted 189 lines
Documentation/ioasid.rst new file +188 -0
+
+Also notice that TLB invalidation occurs at mmu_notifier
+invalidate_range callback as well as the release callback. The reason
+is that release callback will delete IOMMU driver from the notifier
+chain which may skip invalidate_range() calls during the exit path.
+
+To avoid unnecessary reporting of UR fault, IOMMU driver shall disable
+fault reporting after free and before unbind.
+
+Guest IOASID Life Cycle (VT-d Example)
+--------------------------------------
+Guest IOASID life cycle starts with guest driver open(), this could be
+uacce or individual accelerator driver such as DSA. At FD open,
+sva_bind_device() is called which triggers a series of actions.
+
+The example below is an illustration of *normal* operations that
+involves *all* the SW components in VT-d. The flow can be simpler if
+no ENQCMD is supported.
+
+::
+
+     VFIO        IOMMU        KVM        VDCM        IOASID       Ref
+   ..................................................................
+   1             ioasid_register_notifier/_mm()
+   2 ioasid_alloc()                                               1
+   3 bind_gpasid()
+   4             iommu_bind()->ioasid_get()                       2
+   5             ioasid_notify(BIND)
+   6                          -> ioasid_get()                     3
+   7                          -> vmcs_update_atomic()
+   8 mdev_write(gpasid)
+   9                                    hpasid=
+   10                                   find_by_spid(gpasid)      4
+   11                                   vdev_write(hpasid)
+   12 -------- GUEST STARTS DMA --------------------------
+   13 -------- GUEST STOPS DMA --------------------------
+   14 mdev_clear(gpasid)
+   15                                   vdev_clear(hpasid)
+   16                                   ioasid_put()               3
+   17 unbind_gpasid()
+   18            iommu_ubind()
+   19            ioasid_notify(UNBIND)
+   20                          -> vmcs_update_atomic()
+   21                          -> ioasid_put()                     2
+   22 ioasid_free()                                                1
+   23            ioasid_put()                                      0
+   24                                                 Reclaimed
+   -------------- New Life Cycle Begin ----------------------------
+   1  ioasid_alloc()                                  ->           1
+
+   Note: IOASID Notification Events: FREE, BIND, UNBIND
+
+Exception cases arise when a guest crashes or a malicious guest
+attempts to cause disruption on the host system. The fault handling
+rules are:
+
+1. IOASID free must *always* succeed.
+2. An inactive period may be required before the freed IOASID is
+   reclaimed. During this period, consumers of IOASID perform cleanup.
+3. Malfunction is limited to the guest owned resources for all
+   programming errors.
+
+The primary source of exception is when the following are out of
+order:
+
+1. Start/Stop of DMA activity
+   (Guest device driver, mdev via VFIO)
+2. Setup/Teardown of IOMMU PASID context, IOTLB, DevTLB flushes
+   (Host IOMMU driver bind/unbind)
+3. Setup/Teardown of VMCS PASID translation table entries (KVM) in
+   case of ENQCMD
+4. Programming/Clearing host PASID in VDCM (Host VDCM driver)
+5. IOASID alloc/free (Host IOASID)
+
+VFIO is the *only* user-kernel interface, which is ultimately
+responsible for exception handlings.
+
+#1 is processed the same way as the assigned device today based on
+device file descriptors and events. There is no special handling.
+
+#3 is based on bind/unbind events emitted by #2.
+
+#4 is naturally aligned with IOASID life cycle in that an illegal
+guest PASID programming would fail in obtaining reference of the
+matching host IOASID.
+
+#5 is similar to #4. The fault will be reported to the user if PASID
+used in the ENQCMD is not set up in VMCS PASID translation table.
+
+Therefore, the remaining out of order problem is between #2 and
+#5. I.e. unbind vs. free. More specifically, free before unbind.
+
+IOASID notifier and refcounting are used to ensure order. Following
+a publisher-subscriber pattern where:
+
+- Publishers: VFIO & IOMMU
+- Subscribers: KVM, VDCM, IOMMU
+
+IOASID notifier is atomic which requires subscribers to do quick
+handling of the event in the atomic context. Workqueue can be used for
+any processing that requires thread context. IOASID reference must be
+acquired before receiving the FREE event. The reference must be
+dropped at the end of the processing in order to return the IOASID to
+the pool.
+
+Let's examine the IOASID life cycle again when free happens *before*
+unbind. This could be a result of misbehaving guests or crash. Assuming
+VFIO cannot enforce unbind->free order. Notice that the setup part up
+until step #12 is identical to the normal case, the flow below starts
+with step 13.
+
+::
+
+     VFIO        IOMMU        KVM        VDCM        IOASID       Ref
+   ..................................................................
+   13 -------- GUEST STARTS DMA --------------------------
+   14 -------- *GUEST MISBEHAVES!!!* ----------------
+   15 ioasid_free()
+   16                                             ioasid_notify(FREE)
+   17                                             mark_ioasid_inactive[1]
+   18                          kvm_nb_handler(FREE)
+   19                          vmcs_update_atomic()
+   20                          ioasid_put_locked()   ->           3
+   21                                   vdcm_nb_handler(FREE)
+   22            iomm_nb_handler(FREE)
+   23 ioasid_free() returns[2]          schedule_work()           2
+   24            schedule_work()        vdev_clear_wk(hpasid)
+   25            teardown_pasid_wk()
+   26                                   ioasid_put() ->           1
+   27            ioasid_put()                                     0
+   28                                                 Reclaimed
+   29 unbind_gpasid()
+   30            iommu_unbind()->ioasid_find() Fails[3]
+   -------------- New Life Cycle Begin ----------------------------
+
+Note:
+
+1. By marking IOASID inactive at step #17, no new references can be
+   held. ioasid_get/find() will return -ENOENT;
+2. After step #23, all events can go out of order. Shall not affect
+   the outcome.
+3. IOMMU driver fails to find private data for unbinding. If unbind is
+   called after the same IOASID is allocated for the same guest again,
+   this is a programming error. The damage is limited to the guest
+   itself since unbind performs permission checking based on the
+   IOASID set associated with the guest process.
+
+KVM PASID Translation Table Updates
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Per VM PASID translation table is maintained by KVM in order to
+support ENQCMD in the guest. The table contains host-guest PASID
+translations to be consumed by CPU ucode. The synchronization of the
+PASID states depends on VFIO/IOMMU driver, where IOCTL and atomic
+notifiers are used. KVM must register IOASID notifier per VM instance
+during launch time. The following events are handled:
+
+1. BIND/UNBIND
+2. FREE
+
+Rules:
+
+1. Multiple devices can bind with the same PASID, this can be different PCI
+   devices or mdevs within the same PCI device. However, only the
+   *first* BIND and *last* UNBIND emit notifications.
+2. IOASID code is responsible for ensuring the correctness of H-G
+   PASID mapping. There is no need for KVM to validate the
+   notification data.
+3. When UNBIND happens *after* FREE, KVM will see error in
+   ioasid_get() even when the reclaim is not done. IOMMU driver will
+   also avoid sending UNBIND if the PASID is already FREE.
+4. When KVM terminates *before* FREE & UNBIND, references will be
+   dropped for all host PASIDs.
+
+VDCM PASID Programming
+~~~~~~~~~~~~~~~~~~~~~~
+VDCM composes virtual devices and exposes them to the guests. When
+the guest allocates a PASID then program it to the virtual device, VDCM
+intercepts the programming attempt then program the matching host
+PASID on to the hardware.
+Conversely, when a device is going away, VDCM must be informed such
+that PASID context on the hardware can be cleared. There could be
+multiple mdevs assigned to different guests in the same VDCM. Since
+the PASID table is shared at PCI device level, lazy clearing is not
+secure. A malicious guest can attack by using newly freed PASIDs that
+are allocated by another guest.
+
+By holding a reference of the PASID until VDCM cleans up the HW context,
+it is guaranteed that PASID life cycles do not cross within the same

Documentation/ioasid.rst line 618

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+device.
+
+
+Reference
+====================================================
+1. https://software.intel.com/sites/default/files/managed/c5/15/architecture-instruction-set-extensions-programming-reference.pdf
+
+2. https://01.org/blogs/2019/introducing-intel-data-streaming-accelerator
+
+3. https://software.intel.com/en-us/download/intel-data-streaming-accelerator-preliminary-architecture-specification

Best regards,
baolu
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

maintainer
#15 lore raw

Re: [PATCH v2 3/9] iommu/ioasid: Introduce ioasid_set APIs

In reply to Jacob Pan

Hi Jacob,

Quoted 204 lines · Jacob Pan

ioasid_set was introduced as an arbitrary token that are shared by a group of IOASIDs. For example, if IOASID #1 and #2 are allocated via the same ioasid_set*, they are viewed as to belong to the same set.

For guest SVA usages, system-wide IOASID resources need to be partitioned such that VMs can have its own quota and being managed separately. ioasid_set is the perfect candidate for meeting such requirements. This patch redefines and extends ioasid_set with the following new fields:
- Quota
- Reference count
- Storage of its namespace
- The token is stored in the new ioasid_set but with optional types

ioasid_set level APIs are introduced that wires up these new data. Existing users of IOASID APIs are converted where a host IOASID set is allocated for bare-metal usage.

---
  drivers/iommu/intel/iommu.c |  27 ++-
  drivers/iommu/intel/pasid.h |   1 +
  drivers/iommu/intel/svm.c   |   8 +-
  drivers/iommu/ioasid.c      | 390 +++++++++++++++++++++++++++++++++++++++++---
  include/linux/ioasid.h      |  82 ++++++++--
  5 files changed, 465 insertions(+), 43 deletions(-)
drivers/iommu/intel/iommu.c +22 -5
index a3a0b5c8921d..5813eeaa5edb 100644
@@ -42,6 +42,7 @@
  #include <linux/crash_dump.h>
  #include <linux/numa.h>
  #include <linux/swiotlb.h>
+#include <linux/ioasid.h>
  #include <asm/irq_remapping.h>
  #include <asm/cacheflush.h>
  #include <asm/iommu.h>
@@ -103,6 +104,9 @@
   */
  #define INTEL_IOMMU_PGSIZES	(~0xFFFUL)
 
+/* PASIDs used by host SVM */
+struct ioasid_set *host_pasid_set;
+
  static inline int agaw_to_level(int agaw)
  {
  	return agaw + 2;
@@ -3103,8 +3107,8 @@ static void intel_vcmd_ioasid_free(ioasid_t ioasid, void *data)
  	 * Sanity check the ioasid owner is done at upper layer, e.g. VFIO
  	 * We can only free the PASID when all the devices are unbound.
  	 */
-	if (ioasid_find(NULL, ioasid, NULL)) {
-		pr_alert("Cannot free active IOASID %d\n", ioasid);
+	if (IS_ERR(ioasid_find(host_pasid_set, ioasid, NULL))) {
+		pr_err("Cannot free IOASID %d, not in system set\n", ioasid);
  		return;
  	}
  	vcmd_free_pasid(iommu, ioasid);
@@ -3288,6 +3292,19 @@ static int __init init_dmars(void)
  	if (ret)
  		goto free_iommu;
 
+	/* PASID is needed for scalable mode irrespective to SVM */
+	if (intel_iommu_sm) {
+		ioasid_install_capacity(intel_pasid_max_id);
+		/* We should not run out of IOASIDs at boot */
+		host_pasid_set = ioasid_alloc_set(NULL, PID_MAX_DEFAULT,
+						IOASID_SET_TYPE_NULL);
+		if (IS_ERR_OR_NULL(host_pasid_set)) {
+			pr_err("Failed to enable host PASID allocator %lu\n",
+				PTR_ERR(host_pasid_set));
+			intel_iommu_sm = 0;
+		}
+	}
+
  	/*
  	 * for each drhd
  	 *   enable fault log
@@ -5149,7 +5166,7 @@ static void auxiliary_unlink_device(struct dmar_domain *domain,
  	domain->auxd_refcnt--;
 
  	if (!domain->auxd_refcnt && domain->default_pasid > 0)
-		ioasid_free(domain->default_pasid);
+		ioasid_free(host_pasid_set, domain->default_pasid);
  }
 
  static int aux_domain_add_dev(struct dmar_domain *domain,
@@ -5167,7 +5184,7 @@ static int aux_domain_add_dev(struct dmar_domain *domain,
  		int pasid;
 
  		/* No private data needed for the default pasid */
-		pasid = ioasid_alloc(NULL, PASID_MIN,
+		pasid = ioasid_alloc(host_pasid_set, PASID_MIN,
  				     pci_max_pasids(to_pci_dev(dev)) - 1,
  				     NULL);
  		if (pasid == INVALID_IOASID) {
@@ -5210,7 +5227,7 @@ static int aux_domain_add_dev(struct dmar_domain *domain,
  	spin_unlock(&iommu->lock);
  	spin_unlock_irqrestore(&device_domain_lock, flags);
  	if (!domain->auxd_refcnt && domain->default_pasid > 0)
-		ioasid_free(domain->default_pasid);
+		ioasid_free(host_pasid_set, domain->default_pasid);
 
  	return ret;
  }
drivers/iommu/intel/pasid.h +1 -0
index c9850766c3a9..ccdc23446015 100644
@@ -99,6 +99,7 @@ static inline bool pasid_pte_is_present(struct pasid_entry *pte)
  }
 
  extern u32 intel_pasid_max_id;
+extern struct ioasid_set *host_pasid_set;
  int intel_pasid_alloc_id(void *ptr, int start, int end, gfp_t gfp);
  void intel_pasid_free_id(int pasid);
  void *intel_pasid_lookup_id(int pasid);
drivers/iommu/intel/svm.c +4 -4
index 37a9beabc0ca..634e191ca2c3 100644
@@ -551,7 +551,7 @@ intel_svm_bind_mm(struct device *dev, int flags, struct svm_dev_ops *ops,
  			pasid_max = intel_pasid_max_id;
 
  		/* Do not use PASID 0, reserved for RID to PASID */
-		svm->pasid = ioasid_alloc(NULL, PASID_MIN,
+		svm->pasid = ioasid_alloc(host_pasid_set, PASID_MIN,
  					  pasid_max - 1, svm);
  		if (svm->pasid == INVALID_IOASID) {
  			kfree(svm);
@@ -568,7 +568,7 @@ intel_svm_bind_mm(struct device *dev, int flags, struct svm_dev_ops *ops,
  		if (mm) {
  			ret = mmu_notifier_register(&svm->notifier, mm);
  			if (ret) {
-				ioasid_free(svm->pasid);
+				ioasid_free(host_pasid_set, svm->pasid);
  				kfree(svm);
  				kfree(sdev);
  				goto out;
@@ -586,7 +586,7 @@ intel_svm_bind_mm(struct device *dev, int flags, struct svm_dev_ops *ops,
  		if (ret) {
  			if (mm)
  				mmu_notifier_unregister(&svm->notifier, mm);
-			ioasid_free(svm->pasid);
+			ioasid_free(host_pasid_set, svm->pasid);
  			kfree(svm);
  			kfree(sdev);
  			goto out;
@@ -655,7 +655,7 @@ static int intel_svm_unbind_mm(struct device *dev, int pasid)
  			kfree_rcu(sdev, rcu);
 
  			if (list_empty(&svm->devs)) {
-				ioasid_free(svm->pasid);
+				ioasid_free(host_pasid_set, svm->pasid);
  				if (svm->mm)
  					mmu_notifier_unregister(&svm->notifier, svm->mm);
  				list_del(&svm->list);
drivers/iommu/ioasid.c +22 -3
index 5f63af07acd5..f73b3dbfc37a 100644
@@ -1,22 +1,58 @@
  // SPDX-License-Identifier: GPL-2.0
  /*
   * I/O Address Space ID allocator. There is one global IOASID space, split into
- * subsets. Users create a subset with DECLARE_IOASID_SET, then allocate and
- * free IOASIDs with ioasid_alloc and ioasid_free.
+ * subsets. Users create a subset with ioasid_alloc_set, then allocate/free IDs
+ * with ioasid_alloc and ioasid_free.
   */
-#include <linux/ioasid.h>
  #include <linux/module.h>
  #include <linux/slab.h>
  #include <linux/spinlock.h>
  #include <linux/xarray.h>
+#include <linux/ioasid.h>
+
+static DEFINE_XARRAY_ALLOC(ioasid_sets);
+enum ioasid_state {
+	IOASID_STATE_INACTIVE,
+	IOASID_STATE_ACTIVE,
+	IOASID_STATE_FREE_PENDING,
+};
 
+/**
+ * struct ioasid_data - Meta data about ioasid
+ *
+ * @id:		Unique ID
+ * @users	Number of active users
+ * @state	Track state of the IOASID
+ * @set		Meta data of the set this IOASID belongs to
+ * @private	Private data associated with the IOASID
+ * @rcu		For free after RCU grace period
+ */
  struct ioasid_data {
  	ioasid_t id;
  	struct ioasid_set *set;
+	refcount_t users;
+	enum ioasid_state state;
  	void *private;
  	struct rcu_head rcu;

drivers/iommu/ioasid.c line 46

Quoted 10 lines
drivers/iommu/ioasid.c +8 -0
  };
 
+/* Default to PCIe standard 20 bit PASID */
+#define PCI_PASID_MAX 0x100000
+static ioasid_t ioasid_capacity = PCI_PASID_MAX;
+static ioasid_t ioasid_capacity_avail = PCI_PASID_MAX;
+
+void ioasid_install_capacity(ioasid_t total)
+{
+	ioasid_capacity = ioasid_capacity_avail = total;

Need any check for multiple settings or setting after used?

Quoted 130 lines
drivers/iommu/ioasid.c +86 -11
+}
+EXPORT_SYMBOL_GPL(ioasid_install_capacity);
+
+ioasid_t ioasid_get_capacity()
+{
+	return ioasid_capacity;
+}
+EXPORT_SYMBOL_GPL(ioasid_get_capacity);
+
  /*
   * struct ioasid_allocator_data - Internal data structure to hold information
   * about an allocator. There are two types of allocators:
@@ -306,11 +342,23 @@ ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max,
  {
  	struct ioasid_data *data;
  	void *adata;
-	ioasid_t id;
+	ioasid_t id = INVALID_IOASID;
+
+	spin_lock(&ioasid_allocator_lock);
+	/* Check if the IOASID set has been allocated and initialized */
+	if (WARN_ON(xa_load(&ioasid_sets, set->sid) != set)) {
+		pr_warn("Invalid set\n");
+		goto done_unlock;
+	}
+
+	if (set->quota <= set->nr_ioasids) {
+		pr_err("IOASID set %d out of quota %d\n", set->sid, set->quota);
+		goto done_unlock;
+	}
 
  	data = kzalloc(sizeof(*data), GFP_ATOMIC);
  	if (!data)
-		return INVALID_IOASID;
+		goto done_unlock;
 
  	data->set = set;
  	data->private = private;
@@ -319,7 +367,6 @@ ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max,
  	 * Custom allocator needs allocator data to perform platform specific
  	 * operations.
  	 */
-	spin_lock(&ioasid_allocator_lock);
  	adata = active_allocator->flags & IOASID_ALLOCATOR_CUSTOM ? active_allocator->ops->pdata : data;
  	id = active_allocator->ops->alloc(min, max, adata);
  	if (id == INVALID_IOASID) {
@@ -335,42 +382,339 @@ ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max,
  		goto exit_free;
  	}
  	data->id = id;
+	data->state = IOASID_STATE_ACTIVE;
+	refcount_set(&data->users, 1);
+
+	/* Store IOASID in the per set data */
+	if (xa_err(xa_store(&set->xa, id, data, GFP_ATOMIC))) {
+		pr_err("Failed to ioasid %d in set %d\n", id, set->sid);
+		goto exit_free;
+	}
+	set->nr_ioasids++;
+	goto done_unlock;
 
-	spin_unlock(&ioasid_allocator_lock);
-	return id;
  exit_free:
-	spin_unlock(&ioasid_allocator_lock);
  	kfree(data);
-	return INVALID_IOASID;
+done_unlock:
+	spin_unlock(&ioasid_allocator_lock);
+	return id;
  }
  EXPORT_SYMBOL_GPL(ioasid_alloc);
 
+static void ioasid_do_free(struct ioasid_data *data)
+{
+	struct ioasid_data *ioasid_data;
+	struct ioasid_set *sdata;
+
+	active_allocator->ops->free(data->id, active_allocator->ops->pdata);
+	/* Custom allocator needs additional steps to free the xa element */
+	if (active_allocator->flags & IOASID_ALLOCATOR_CUSTOM) {
+		ioasid_data = xa_erase(&active_allocator->xa, data->id);
+		kfree_rcu(ioasid_data, rcu);
+	}
+
+	sdata = xa_load(&ioasid_sets, data->set->sid);
+	if (!sdata) {
+		pr_err("No set %d for IOASID %d\n", data->set->sid,
+		       data->id);
+		return;
+	}
+	xa_erase(&sdata->xa, data->id);
+	sdata->nr_ioasids--;
+}
+
+static void ioasid_free_locked(struct ioasid_set *set, ioasid_t ioasid)
+{
+	struct ioasid_data *data;
+
+	data = xa_load(&active_allocator->xa, ioasid);
+	if (!data) {
+		pr_err("Trying to free unknown IOASID %u\n", ioasid);
+		return;
+	}
+
+	if (data->set != set) {
+		pr_warn("Cannot free IOASID %u due to set ownership\n", ioasid);
+		return;
+	}
+	data->state = IOASID_STATE_FREE_PENDING;
+
+	if (!refcount_dec_and_test(&data->users))
+		return;
+
+	ioasid_do_free(data);
+}
+
  /**
- * ioasid_free - Free an IOASID
- * @ioasid: the ID to remove
+ * ioasid_free - Drop reference on an IOASID. Free if refcount drops to 0,
+ *               including free from its set and system-wide list.
+ * @set:	The ioasid_set to check permission with. If not NULL, IOASID
+ *		free will fail if the set does not match.
+ * @ioasid:	The IOASID to remove
   */
-void ioasid_free(ioasid_t ioasid)
+void ioasid_free(struct ioasid_set *set, ioasid_t ioasid)
  {
-	struct ioasid_data *ioasid_data;

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 466

Quoted 10 lines
drivers/iommu/ioasid.c +9 -0
+	spin_lock(&ioasid_allocator_lock);
+	ioasid_free_locked(set, ioasid);
+	spin_unlock(&ioasid_allocator_lock);
+}
+EXPORT_SYMBOL_GPL(ioasid_free);
 
+/**
+ * ioasid_alloc_set - Allocate a new IOASID set for a given token
+ *
+ * @token:	Unique token of the IOASID set, cannot be NULL

What's the use of @token? I might be able to find the answer in the code, but i have no idea when i comes here first time. :-)

This line of comment says that token cannot be NULL. It seems not to match the real code, where token could be NULL if type is IOASID_SET_TYPE_NULL.

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 471

Quoted 5 lines
drivers/iommu/ioasid.c +5 -0
+ * @quota:	Quota allowed in this set. Only for new set creation
+ * @flags:	Special requirements
+ *
+ * IOASID can be limited system-wide resource that requires quota management.
+ * If caller does not wish to enforce quota, use IOASID_SET_NO_QUOTA flag.

If you are not going to add NO_QUOTA support this time, I'd suggest you to remove above comment.

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 476

Quoted 5 lines
drivers/iommu/ioasid.c +5 -0
+ *
+ * Token will be stored in the ioasid_set returned. A reference will be taken
+ * upon finding a matching set or newly created set.
+ * IOASID allocation within the set and other per set operations will use
+ * the retured ioasid_set *.

nit: remove *, or you mean the ioasid_set pointer?

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+ *
+ */
+struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota, int type)
+{
+	struct ioasid_set *sdata;
+	unsigned long index;
+	ioasid_t id;
+
+	if (type >= IOASID_SET_TYPE_NR)
+		return ERR_PTR(-EINVAL);

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 493

Quoted 10 lines
drivers/iommu/ioasid.c +6 -3
+
+	/*
+	 * Need to check space available if we share system-wide quota.
+	 * TODO: we may need to support quota free sets in the future.
+	 */
  	spin_lock(&ioasid_allocator_lock);
-	ioasid_data = xa_load(&active_allocator->xa, ioasid);
-	if (!ioasid_data) {
-		pr_err("Trying to free unknown IOASID %u\n", ioasid);
+	if (quota > ioasid_capacity_avail) {

Thinking that ioasid_set itself needs an ioasid, so the check might be

(quota + 1 > ioasid_capacity_avail)?

Quoted 15 lines
drivers/iommu/ioasid.c +7 -5
+		pr_warn("Out of IOASID capacity! ask %d, avail %d\n",
+			quota, ioasid_capacity_avail);
+		sdata = ERR_PTR(-ENOSPC);
  		goto exit_unlock;
  	}
 
-	active_allocator->ops->free(ioasid, active_allocator->ops->pdata);
-	/* Custom allocator needs additional steps to free the xa element */
-	if (active_allocator->flags & IOASID_ALLOCATOR_CUSTOM) {
-		ioasid_data = xa_erase(&active_allocator->xa, ioasid);
-		kfree_rcu(ioasid_data, rcu);
+	/*
+	 * Token is only unique within its types but right now we have only
+	 * mm type. If we have more token types, we have to match type as well.
+	 */

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 513

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+	switch (type) {
+	case IOASID_SET_TYPE_MM:
+		/* Search existing set tokens, reject duplicates */
+		xa_for_each(&ioasid_sets, index, sdata) {
+			if (sdata->token == token &&
+				sdata->type == IOASID_SET_TYPE_MM) {
+				sdata = ERR_PTR(-EEXIST);
+				goto exit_unlock;
+			}
+		}

Do you need to enforce non-NULL token policy here?

Quoted 32 lines
drivers/iommu/ioasid.c +30 -0
+		break;
+	case IOASID_SET_TYPE_NULL:
+		if (!token)
+			break;
+		fallthrough;
+	default:
+		pr_err("Invalid token and IOASID type\n");
+		sdata = ERR_PTR(-EINVAL);
+		goto exit_unlock;
  	}
 
+	/* REVISIT: may support set w/o quota, use system available */
+	if (!quota) {
+		sdata = ERR_PTR(-EINVAL);
+		goto exit_unlock;
+	}
+
+	sdata = kzalloc(sizeof(*sdata), GFP_ATOMIC);
+	if (!sdata) {
+		sdata = ERR_PTR(-ENOMEM);
+		goto exit_unlock;
+	}
+
+	if (xa_alloc(&ioasid_sets, &id, sdata,
+			XA_LIMIT(0, ioasid_capacity_avail - quota),
+			GFP_ATOMIC)) {
+		kfree(sdata);
+		sdata = ERR_PTR(-ENOSPC);
+		goto exit_unlock;
+	}
+
+	sdata->token = token;

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 555

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+	sdata->type = type;
+	sdata->quota = quota;
+	sdata->sid = id;
+	refcount_set(&sdata->ref, 1);
+
+	/*
+	 * Per set XA is used to store private IDs within the set, get ready
+	 * for ioasid_set private ID and system-wide IOASID allocation
+	 * results.
+	 */

I'm not sure that I understood the per-set XA correctly. Is it used to save both private ID and real ID allocated from the system-wide pool? If so, isn't private id might be equal to real ID?

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 557

Quoted 2 lines
drivers/iommu/ioasid.c +2 -0
+	xa_init_flags(&sdata->xa, XA_FLAGS_ALLOC);
+	ioasid_capacity_avail -= quota;

As mentioned above, the ioasid_set consumed one extra ioasid, so

	ioasid_capacity_avail -= (quota + 1);

?

Quoted 24 lines
drivers/iommu/ioasid.c +20 -1
+
  exit_unlock:
  	spin_unlock(&ioasid_allocator_lock);
+
+	return sdata;
  }
-EXPORT_SYMBOL_GPL(ioasid_free);
+EXPORT_SYMBOL_GPL(ioasid_alloc_set);
+
+void ioasid_set_get_locked(struct ioasid_set *set)
+{
+	if (WARN_ON(xa_load(&ioasid_sets, set->sid) != set)) {
+		pr_warn("Invalid set data\n");
+		return;
+	}
+
+	refcount_inc(&set->ref);
+}
+EXPORT_SYMBOL_GPL(ioasid_set_get_locked);
+
+void ioasid_set_get(struct ioasid_set *set)
+{
+	spin_lock(&ioasid_allocator_lock);
+	ioasid_set_get_locked(set);

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 590

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+	spin_unlock(&ioasid_allocator_lock);
+}
+EXPORT_SYMBOL_GPL(ioasid_set_get);
+
+void ioasid_set_put_locked(struct ioasid_set *set)
+{
+	struct ioasid_data *entry;
+	unsigned long index;
+
+	if (WARN_ON(xa_load(&ioasid_sets, set->sid) != set)) {

There are multiple occurences of this line of code, how about defining it as a inline helper?

static inline bool ioasid_is_valid(struct ioasid_set *set)

{
	return xa_load(&ioasid_sets, set->sid) == set;
}
Quoted 29 lines
drivers/iommu/ioasid.c +29 -0
+		pr_warn("Invalid set data\n");
+		return;
+	}
+
+	if (!refcount_dec_and_test(&set->ref)) {
+		pr_debug("%s: IOASID set %d has %d users\n",
+			__func__, set->sid, refcount_read(&set->ref));
+		return;
+	}
+
+	/* The set is already empty, we just destroy the set. */
+	if (xa_empty(&set->xa))
+		goto done_destroy;
+
+	/*
+	 * Free all PASIDs from system-wide IOASID pool, all subscribers gets
+	 * notified and do cleanup of their own.
+	 * Note that some references of the IOASIDs within the set can still
+	 * be held after the free call. This is OK in that the IOASIDs will be
+	 * marked inactive, the only operations can be done is ioasid_put.
+	 * No need to track IOASID set states since there is no reclaim phase.
+	 */
+	xa_for_each(&set->xa, index, entry) {
+		ioasid_free_locked(set, index);
+		/* Free from per set private pool */
+		xa_erase(&set->xa, index);
+	}
+
+done_destroy:

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 629

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+	/* Return the quota back to system pool */
+	ioasid_capacity_avail += set->quota;
+	kfree_rcu(set, rcu);
+
+	/*
+	 * Token got released right away after the ioasid_set is freed.
+	 * If a new set is created immediately with the newly released token,
+	 * it will not allocate the same IOASIDs unless they are reclaimed.
+	 */
+	xa_erase(&ioasid_sets, set->sid);

No. pointer is used after free.

Quoted 32 lines
drivers/iommu/ioasid.c +32 -0
+}
+EXPORT_SYMBOL_GPL(ioasid_set_put_locked);
+
+/**
+ * ioasid_set_put - Drop a reference to the IOASID set. Free all IOASIDs within
+ *                  the set if there are no more users.
+ *
+ * @set:	The IOASID set ID to be freed
+ *
+ * If refcount drops to zero, all IOASIDs allocated within the set will be
+ * freed.
+ */
+void ioasid_set_put(struct ioasid_set *set)
+{
+	spin_lock(&ioasid_allocator_lock);
+	ioasid_set_put_locked(set);
+	spin_unlock(&ioasid_allocator_lock);
+}
+EXPORT_SYMBOL_GPL(ioasid_set_put);
+
+/**
+ * ioasid_adjust_set - Adjust the quota of an IOASID set
+ * @set:	IOASID set to be assigned
+ * @quota:	Quota allowed in this set
+ *
+ * Return 0 on success. If the new quota is smaller than the number of
+ * IOASIDs already allocated, -EINVAL will be returned. No change will be
+ * made to the existing quota.
+ */
+int ioasid_adjust_set(struct ioasid_set *set, int quota)
+{
+	int ret = 0;

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 671

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+
+	spin_lock(&ioasid_allocator_lock);
+	if (set->nr_ioasids > quota) {
+		pr_err("New quota %d is smaller than outstanding IOASIDs %d\n",
+			quota, set->nr_ioasids);
+		ret = -EINVAL;
+		goto done_unlock;
+	}
+
+	if (quota >= ioasid_capacity_avail) {

This check doesn't make sense since you are updating (not asking for) a quota.

if ((quota > set->quota) &&

    (quota - set->quota > ioasid_capacity_avail))

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 677

Quoted 6 lines
drivers/iommu/ioasid.c +6 -0
+		ret = -ENOSPC;
+		goto done_unlock;
+	}
+
+	/* Return the delta back to system pool */
+	ioasid_capacity_avail += set->quota - quota;

ioasid_capacity_avail is defined as a unsigned int, hence this always increase the available capacity value even the caller is asking for a bigger quota?

Quoted 9 lines
drivers/iommu/ioasid.c +9 -0
+
+	/*
+	 * May have a policy to prevent giving all available IOASIDs
+	 * to one set. But we don't enforce here, it should be in the
+	 * upper layers.
+	 */
+	set->quota = quota;
+
+done_unlock:

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 696

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+	spin_unlock(&ioasid_allocator_lock);
+
+	return ret;
+}
+EXPORT_SYMBOL_GPL(ioasid_adjust_set);
+
+/**
+ * ioasid_set_for_each_ioasid - Iterate over all the IOASIDs within the set
+ *
+ * Caller must hold a reference of the set and handles its own locking.

Do you need to hold ioasid_allocator_lock here?

Quoted 131 lines
drivers/iommu/ioasid.c +21 -0
+ */
+int ioasid_set_for_each_ioasid(struct ioasid_set *set,
+			       void (*fn)(ioasid_t id, void *data),
+			       void *data)
+{
+	struct ioasid_data *entry;
+	unsigned long index;
+	int ret = 0;
+
+	if (xa_empty(&set->xa)) {
+		pr_warn("No IOASIDs in the set %d\n", set->sid);
+		return -ENOENT;
+	}
+
+	xa_for_each(&set->xa, index, entry) {
+		fn(index, data);
+	}
+
+	return ret;
+}
+EXPORT_SYMBOL_GPL(ioasid_set_for_each_ioasid);
 
  /**
   * ioasid_find - Find IOASID data
include/linux/ioasid.h +69 -9
index 9c44947a68c8..412d025d440e 100644
@@ -10,8 +10,35 @@ typedef unsigned int ioasid_t;
  typedef ioasid_t (*ioasid_alloc_fn_t)(ioasid_t min, ioasid_t max, void *data);
  typedef void (*ioasid_free_fn_t)(ioasid_t ioasid, void *data);
 
+/* IOASID set types */
+enum ioasid_set_type {
+	IOASID_SET_TYPE_NULL = 1, /* Set token is NULL */
+	IOASID_SET_TYPE_MM,	  /* Set token is a mm_struct,
+				   * i.e. associated with a process
+				   */
+	IOASID_SET_TYPE_NR,
+};
+
+/**
+ * struct ioasid_set - Meta data about ioasid_set
+ * @type:	Token types and other features
+ * @token:	Unique to identify an IOASID set
+ * @xa:		XArray to store ioasid_set private IDs, can be used for
+ *		guest-host IOASID mapping, or just a private IOASID namespace.
+ * @quota:	Max number of IOASIDs can be allocated within the set
+ * @nr_ioasids	Number of IOASIDs currently allocated in the set
+ * @sid:	ID of the set
+ * @ref:	Reference count of the users
+ */
  struct ioasid_set {
-	int dummy;
+	void *token;
+	struct xarray xa;
+	int type;
+	int quota;
+	int nr_ioasids;
+	int sid;
+	refcount_t ref;
+	struct rcu_head rcu;
  };
 
  /**
@@ -29,31 +56,64 @@ struct ioasid_allocator_ops {
  	void *pdata;
  };
 
-#define DECLARE_IOASID_SET(name) struct ioasid_set name = { 0 }
-
  #if IS_ENABLED(CONFIG_IOASID)
+void ioasid_install_capacity(ioasid_t total);
+ioasid_t ioasid_get_capacity(void);
+struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota, int type);
+int ioasid_adjust_set(struct ioasid_set *set, int quota);
+void ioasid_set_get_locked(struct ioasid_set *set);
+void ioasid_set_put_locked(struct ioasid_set *set);
+void ioasid_set_put(struct ioasid_set *set);
+
  ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max,
  		      void *private);
-void ioasid_free(ioasid_t ioasid);
-void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid,
-		  bool (*getter)(void *));
+void ioasid_free(struct ioasid_set *set, ioasid_t ioasid);
+
+bool ioasid_is_active(ioasid_t ioasid);
+
+void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid, bool (*getter)(void *));
+int ioasid_attach_data(ioasid_t ioasid, void *data);
  int ioasid_register_allocator(struct ioasid_allocator_ops *allocator);
  void ioasid_unregister_allocator(struct ioasid_allocator_ops *allocator);
-int ioasid_attach_data(ioasid_t ioasid, void *data);
-
+void ioasid_is_in_set(struct ioasid_set *set, ioasid_t ioasid);
+int ioasid_set_for_each_ioasid(struct ioasid_set *sdata,
+			       void (*fn)(ioasid_t id, void *data),
+			       void *data);
  #else /* !CONFIG_IOASID */
+static inline void ioasid_install_capacity(ioasid_t total)
+{
+}
+
+static inline ioasid_t ioasid_get_capacity(void)
+{
+	return 0;
+}
+
  static inline ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min,
  				    ioasid_t max, void *private)
  {
  	return INVALID_IOASID;
  }
 
-static inline void ioasid_free(ioasid_t ioasid)
+static inline void ioasid_free(struct ioasid_set *set, ioasid_t ioasid)
+{
+}
+
+static inline bool ioasid_is_active(ioasid_t ioasid)
+{
+	return false;
+}
+
+static inline struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota, int type)
+{
+	return ERR_PTR(-ENOTSUPP);
+}
+

include/linux/ioasid.h struct ioasid_allocator_ops { · line 119

Quoted 10 lines
include/linux/ioasid.h +2 -2
+static inline void ioasid_set_put(struct ioasid_set *set)
  {
  }
 
-static inline void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid,
-				bool (*getter)(void *))
+static inline void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid, bool (*getter)(void *))
  {
  	return NULL;
  }

Best regards,
baolu

maintainer
#28 lore raw

Re: [PATCH v2 4/9] iommu/ioasid: Add reference couting functions

In reply to Jacob Pan

Hi Jacob,

Quoted 126 lines · Jacob Pan

There can be multiple users of an IOASID, each user could have hardware contexts associated with the IOASID. In order to align lifecycles, reference counting is introduced in this patch. It is expected that when an IOASID is being freed, each user will drop a reference only after its context is cleared.

---
  drivers/iommu/ioasid.c | 113 +++++++++++++++++++++++++++++++++++++++++++++++++
  include/linux/ioasid.h |   4 ++
  2 files changed, 117 insertions(+)
drivers/iommu/ioasid.c +106 -0
index f73b3dbfc37a..5f31d63c75b1 100644
@@ -717,6 +717,119 @@ int ioasid_set_for_each_ioasid(struct ioasid_set *set,
  EXPORT_SYMBOL_GPL(ioasid_set_for_each_ioasid);
 
  /**
+ * IOASID refcounting rules
+ * - ioasid_alloc() set initial refcount to 1
+ *
+ * - ioasid_free() decrement and test refcount.
+ *     If refcount is 0, ioasid will be freed. Deleted from the system-wide
+ *     xarray as well as per set xarray. The IOASID will be returned to the
+ *     pool and available for new allocations.
+ *
+ *     If recount is non-zero, mark IOASID as IOASID_STATE_FREE_PENDING.
+ *     No new reference can be added. The IOASID is not returned to the pool
+ *     for reuse.
+ *     After free, ioasid_get() will return error but ioasid_find() and other
+ *     non refcount adding APIs will continue to work until the last reference
+ *     is dropped
+ *
+ * - ioasid_get() get a reference on an active IOASID
+ *
+ * - ioasid_put() decrement and test refcount of the IOASID.
+ *     If refcount is 0, ioasid will be freed. Deleted from the system-wide
+ *     xarray as well as per set xarray. The IOASID will be returned to the
+ *     pool and available for new allocations.
+ *     Do nothing if refcount is non-zero.
+ *
+ * - ioasid_find() does not take reference, caller must hold reference
+ *
+ * ioasid_free() can be called multiple times without error until all refs are
+ * dropped.
+ */
+
+int ioasid_get_locked(struct ioasid_set *set, ioasid_t ioasid)
+{
+	struct ioasid_data *data;
+
+	data = xa_load(&active_allocator->xa, ioasid);
+	if (!data) {
+		pr_err("Trying to get unknown IOASID %u\n", ioasid);
+		return -EINVAL;
+	}
+	if (data->state == IOASID_STATE_FREE_PENDING) {
+		pr_err("Trying to get IOASID being freed%u\n", ioasid);
+		return -EBUSY;
+	}
+
+	if (set && data->set != set) {
+		pr_err("Trying to get IOASID not in set%u\n", ioasid);
+		/* data found but does not belong to the set */
+		return -EACCES;
+	}
+	refcount_inc(&data->users);
+
+	return 0;
+}
+EXPORT_SYMBOL_GPL(ioasid_get_locked);
+
+/**
+ * ioasid_get - Obtain a reference of an ioasid
+ * @set
+ * @ioasid
+ *
+ * Check set ownership if @set is non-null.
+ */
+int ioasid_get(struct ioasid_set *set, ioasid_t ioasid)
+{
+	int ret = 0;
+
+	spin_lock(&ioasid_allocator_lock);
+	ret = ioasid_get_locked(set, ioasid);
+	spin_unlock(&ioasid_allocator_lock);
+
+	return ret;
+}
+EXPORT_SYMBOL_GPL(ioasid_get);
+
+void ioasid_put_locked(struct ioasid_set *set, ioasid_t ioasid)
+{
+	struct ioasid_data *data;
+
+	data = xa_load(&active_allocator->xa, ioasid);
+	if (!data) {
+		pr_err("Trying to put unknown IOASID %u\n", ioasid);
+		return;
+	}
+
+	if (set && data->set != set) {
+		pr_err("Trying to drop IOASID not in the set %u\n", ioasid);
+		return;
+	}
+
+	if (!refcount_dec_and_test(&data->users)) {
+		pr_debug("%s: IOASID %d has %d remainning users\n",
+			__func__, ioasid, refcount_read(&data->users));
+		return;
+	}
+	ioasid_do_free(data);
+}
+EXPORT_SYMBOL_GPL(ioasid_put_locked);
+
+/**
+ * ioasid_put - Drop a reference of an ioasid
+ * @set
+ * @ioasid
+ *
+ * Check set ownership if @set is non-null.
+ */
+void ioasid_put(struct ioasid_set *set, ioasid_t ioasid)
+{

drivers/iommu/ioasid.c int ioasid_set_for_each_ioasid(struct ioasid_set *set, · line 835

Quoted 10 lines
drivers/iommu/ioasid.c +7 -0
+	spin_lock(&ioasid_allocator_lock);
+	ioasid_put_locked(set, ioasid);
+	spin_unlock(&ioasid_allocator_lock);
+}
+EXPORT_SYMBOL_GPL(ioasid_put);
+
+/**
   * ioasid_find - Find IOASID data
   * @set: the IOASID set
   * @ioasid: the IOASID to find

Do you need to increase the refcount of the found ioasid and ask the caller to drop it after use? Otherwise, the ioasid might be freed elsewhere.

Best regards,
baolu

Quoted 16 lines
include/linux/ioasid.h +4 -0
index 412d025d440e..310abe4187a3 100644
@@ -76,6 +76,10 @@ int ioasid_attach_data(ioasid_t ioasid, void *data);
  int ioasid_register_allocator(struct ioasid_allocator_ops *allocator);
  void ioasid_unregister_allocator(struct ioasid_allocator_ops *allocator);
  void ioasid_is_in_set(struct ioasid_set *set, ioasid_t ioasid);
+int ioasid_get(struct ioasid_set *set, ioasid_t ioasid);
+int ioasid_get_locked(struct ioasid_set *set, ioasid_t ioasid);
+void ioasid_put(struct ioasid_set *set, ioasid_t ioasid);
+void ioasid_put_locked(struct ioasid_set *set, ioasid_t ioasid);
  int ioasid_set_for_each_ioasid(struct ioasid_set *sdata,
  			       void (*fn)(ioasid_t id, void *data),
  			       void *data);
#4 lore raw

Re: [PATCH v2 1/9] docs: Document IO Address Space ID (IOASID) APIs

In reply to Jacob Pan

Documentation/ioasid.rst

Quoted 17 lines · Jacob Pan

IOASID is used to identify address spaces that can be targeted by device DMA. It is a system-wide resource that is essential to its many users. This document is an attempt to help developers from all vendors navigate the APIs. At this time, ARM SMMU and Intel’s Scalable IO Virtualization (SIOV) enabled platforms are the primary users of IOASID. Examples of how SIOV components interact with IOASID APIs are provided in that many APIs are driven by the requirements from SIOV.

---
 Documentation/ioasid.rst | 618 +++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 618 insertions(+)
 create mode 100644 Documentation/ioasid.rst
Documentation/ioasid.rst +0 -0

Thanks for writing this up. Should it go to Documentation/driver-api/, or Documentation/driver-api/iommu/? I think this also needs to Cc [email protected] and [email protected]

Documentation/ioasid.rst line 8

Quoted 13 lines

new file mode 100644
index 000000000000..b6a8cdc885ff

Documentation/ioasid.rst new file +8 -0
@@ -0,0 +1,618 @@
+.. ioasid:
+
+=====================================
+IO Address Space ID
+=====================================
+
+IOASID is a generic name for PCIe Process Address ID (PASID) or ARM
+SMMU sub-stream ID. An IOASID identifies an address space that DMA

"SubstreamID"

Documentation/ioasid.rst line 12

Quoted 4 lines
Documentation/ioasid.rst new file +4 -0
+requests can target.
+
+The primary use cases for IOASID are Shared Virtual Address (SVA) and
+IO Virtual Address (IOVA). However, the requirements for IOASID

IOVA alone isn't a use case, maybe "multiple IOVA spaces per device"?

Quoted 4 lines
Documentation/ioasid.rst new file +4 -0
+management can vary among hardware architectures.
+
+This document covers the generic features supported by IOASID
+APIs. Vendor-specific use cases are also illustrated with Intel's VT-d

Documentation/ioasid.rst line 26

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+based platforms as the first example.
+
+.. contents:: :local:
+
+Glossary
+========
+PASID - Process Address Space ID
+
+IOASID - IO Address Space ID (generic term for PCIe PASID and
+sub-stream ID in SMMU)

"SubstreamID"

Documentation/ioasid.rst line 30

Quoted 4 lines
Documentation/ioasid.rst new file +4 -0
+
+SVA/SVM - Shared Virtual Addressing/Memory
+
+ENQCMD - New Intel X86 ISA for efficient workqueue submission [1]

Maybe drop the "New", to keep the documentation perennial. It might be good to add internal links here to the specifications URLs at the bottom.

Quoted 6 lines
Documentation/ioasid.rst new file +6 -0
+
+DSA - Intel Data Streaming Accelerator [2]
+
+VDCM - Virtual device composition module [3]
+
+SIOV - Intel Scalable IO Virtualization

Documentation/ioasid.rst line 46

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+
+
+Key Concepts
+============
+
+IOASID Set
+-----------
+An IOASID set is a group of IOASIDs allocated from the system-wide
+IOASID pool. An IOASID set is created and can be identified by a
+token of u64. Refer to IOASID set APIs for more details.

Identified either by an u64 or an mm_struct, right? Maybe just drop the second sentence if it's detailed in the IOASID set section below.

Documentation/ioasid.rst

Quoted 9 lines
Documentation/ioasid.rst new file +9 -0
+
+IOASID set is particularly useful for guest SVA where each guest could
+have its own IOASID set for security and efficiency reasons.
+
+IOASID Set Private ID (SPID)
+----------------------------
+SPIDs are introduced as IOASIDs within its set. Each SPID maps to a
+system-wide IOASID but the namespace of SPID is within its IOASID
+set.

The intro isn't super clear. Perhaps this is simpler:
"Each IOASID set has a private namespace of SPIDs. An SPID maps to a single system-wide IOASID."

Quoted 36 lines

SPIDs can be used as guest IOASIDs where each guest could do

Documentation/ioasid.rst new file +35 -0
+IOASID allocation from its own pool and map them to host physical
+IOASIDs. SPIDs are particularly useful for supporting live migration
+where decoupling guest and host physical resources are necessary.
+
+For example, two VMs can both allocate guest PASID/SPID #101 but map to
+different host PASIDs #201 and #202 respectively as shown in the
+diagram below.
+::
+
+ .------------------.    .------------------.
+ |   VM 1           |    |   VM 2           |
+ |                  |    |                  |
+ |------------------|    |------------------|
+ | GPASID/SPID 101  |    | GPASID/SPID 101  |
+ '------------------'    -------------------'     Guest
+ __________|______________________|______________________
+           |                      |               Host
+           v                      v
+ .------------------.    .------------------.
+ | Host IOASID 201  |    | Host IOASID 202  |
+ '------------------'    '------------------'
+ |   IOASID set 1   |    |   IOASID set 2   |
+ '------------------'    '------------------'
+
+Guest PASID is treated as IOASID set private ID (SPID) within an
+IOASID set, mappings between guest and host IOASIDs are stored in the
+set for inquiry.
+
+IOASID APIs
+===========
+To get the IOASID APIs, users must #include <linux/ioasid.h>. These APIs
+serve the following functionalities:
+
+  - IOASID allocation/Free
+  - Group management in the form of ioasid_set

Documentation/ioasid.rst line 100

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+  - Private data storage and lookup
+  - Reference counting
+  - Event notification in case of state change
+
+IOASID Set Level APIs
+--------------------------
+For use cases such as guest SVA it is necessary to manage IOASIDs at
+a group level. For example, VMs may allocate multiple IOASIDs for
+guest process address sharing (vSVA). It is imperative to enforce
+VM-IOASID ownership such that malicious guest cannot target DMA

"a malicious guest"

Quoted 1 line

+traffic outside its own IOASIDs, or free an active IOASID belong to

"that belongs to"

Documentation/ioasid.rst line 107

Quoted 6 lines
Documentation/ioasid.rst new file +6 -0
+another VM.
+::
+
+ struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota, u32 type)
+
+ int ioasid_adjust_set(struct ioasid_set *set, int quota);

These could be named "ioasid_set_alloc" and "ioasid_set_adjust" to be consistent with the rest of the API.

Documentation/ioasid.rst line 117

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+
+ void ioasid_set_get(struct ioasid_set *set)
+
+ void ioasid_set_put(struct ioasid_set *set)
+
+ void ioasid_set_get_locked(struct ioasid_set *set)
+
+ void ioasid_set_put_locked(struct ioasid_set *set)
+
+ int ioasid_set_for_each_ioasid(struct ioasid_set *sdata,

Might be nicer to keep the same argument names within the API. Here "set" rather than "sdata".

Documentation/ioasid.rst line 119

Quoted 2 lines
Documentation/ioasid.rst new file +2 -0
+                                void (*fn)(ioasid_t id, void *data),
+				void *data)

(alignment)

Documentation/ioasid.rst line 122

Quoted 3 lines
Documentation/ioasid.rst new file +3 -0
+
+
+IOASID set concept is introduced to represent such IOASID groups. Each

Or just "IOASID sets represent such IOASID groups", but might be redundant.

Quoted 8 lines
Documentation/ioasid.rst new file +8 -0
+IOASID set is created with a token which can be one of the following
+types:
+
+ - IOASID_SET_TYPE_NULL (Arbitrary u64 value)
+ - IOASID_SET_TYPE_MM (Set token is a mm_struct)
+
+The explicit MM token type is useful when multiple users of an IOASID
+set under the same process need to communicate about their shared IOASIDs.

Documentation/ioasid.rst line 140

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+E.g. An IOASID set created by VFIO for one guest can be associated
+with the KVM instance for the same guest since they share a common mm_struct.
+
+The IOASID set APIs serve the following purposes:
+
+ - Ownership/permission enforcement
+ - Take collective actions, e.g. free an entire set
+ - Event notifications within a set
+ - Look up a set based on token
+ - Quota enforcement

This paragraph could be earlier in the section

Quoted 19 lines
Documentation/ioasid.rst new file +19 -0
+
+Individual IOASID APIs
+----------------------
+Once an ioasid_set is created, IOASIDs can be allocated from the set.
+Within the IOASID set namespace, set private ID (SPID) is supported. In
+the VM use case, SPID can be used for storing guest PASID.
+
+::
+
+ ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max,
+                       void *private);
+
+ int ioasid_get(struct ioasid_set *set, ioasid_t ioasid);
+
+ void ioasid_put(struct ioasid_set *set, ioasid_t ioasid);
+
+ int ioasid_get_locked(struct ioasid_set *set, ioasid_t ioasid);
+
+ void ioasid_put_locked(struct ioasid_set *set, ioasid_t ioasid);

Documentation/ioasid.rst line 169

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+
+ void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid,
+                   bool (*getter)(void *));
+
+ ioasid_t ioasid_find_by_spid(struct ioasid_set *set, ioasid_t spid)
+
+ int ioasid_attach_data(struct ioasid_set *set, ioasid_t ioasid,
+                        void *data);
+ int ioasid_attach_spid(struct ioasid_set *set, ioasid_t ioasid,
+                        ioasid_t ssid);

s/ssid/spid

Quoted 65 lines
Documentation/ioasid.rst new file +65 -0
+
+
+Notifications
+-------------
+An IOASID may have multiple users, each user may have hardware context
+associated with an IOASID. When the status of an IOASID changes,
+e.g. an IOASID is being freed, users need to be notified such that the
+associated hardware context can be cleared, flushed, and drained.
+
+::
+
+ int ioasid_register_notifier(struct ioasid_set *set, struct
+                              notifier_block *nb)
+
+ void ioasid_unregister_notifier(struct ioasid_set *set,
+                                 struct notifier_block *nb)
+
+ int ioasid_register_notifier_mm(struct mm_struct *mm, struct
+                                 notifier_block *nb)
+
+ void ioasid_unregister_notifier_mm(struct mm_struct *mm, struct
+                                    notifier_block *nb)
+
+ int ioasid_notify(ioasid_t ioasid, enum ioasid_notify_val cmd,
+                   unsigned int flags)
+
+
+Events
+~~~~~~
+Notification events are pertinent to individual IOASIDs, they can be
+one of the following:
+
+ - ALLOC
+ - FREE
+ - BIND
+ - UNBIND
+
+Ordering
+~~~~~~~~
+Ordering is supported by IOASID notification priorities as the
+following (in ascending order):
+
+::
+
+ enum ioasid_notifier_prios {
+	IOASID_PRIO_LAST,
+	IOASID_PRIO_IOMMU,
+	IOASID_PRIO_DEVICE,
+	IOASID_PRIO_CPU,
+ };
+
+The typical use case is when an IOASID is freed due to an exception, DMA
+source should be quiesced before tearing down other hardware contexts
+in the system. This will reduce the churn in handling faults. DMA work
+submission is performed by the CPU which is granted higher priority than
+devices.
+
+
+Scopes
+~~~~~~
+There are two types of notifiers in IOASID core: system-wide and
+ioasid_set-wide.
+
+System-wide notifier is catering for users that need to handle all
+IOASIDs in the system. E.g. The IOMMU driver handles all IOASIDs.

Documentation/ioasid.rst line 244

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+
+Per ioasid_set notifier can be used by VM specific components such as
+KVM. After all, each KVM instance only cares about IOASIDs within its
+own set.
+
+
+Atomicity
+~~~~~~~~~
+IOASID notifiers are atomic due to spinlocks used inside the IOASID
+core. For tasks cannot be completed in the notifier handler, async work

"tasks that cannot be"

Documentation/ioasid.rst line 251

Quoted 7 lines
Documentation/ioasid.rst new file +7 -0
+can be submitted to complete the work later as long as there is no
+ordering requirement.
+
+Reference counting
+------------------
+IOASID lifecycle management is based on reference counting. Users of
+IOASID intend to align lifecycle with the IOASID need to hold

"who intend to"

Quoted 1 line

+reference of the IOASID. IOASID will not be returned to the pool for

"a reference to the IOASID. The IOASID"

Quoted 59 lines
Documentation/ioasid.rst new file +59 -0
+allocation until all references are dropped. Calling ioasid_free()
+will mark the IOASID as FREE_PENDING if the IOASID has outstanding
+reference. ioasid_get() is not allowed once an IOASID is in the
+FREE_PENDING state.
+
+Event notifications are used to inform users of IOASID status change.
+IOASID_FREE event prompts users to drop their references after
+clearing its context.
+
+For example, on VT-d platform when an IOASID is freed, teardown
+actions are performed on KVM, device driver, and IOMMU driver.
+KVM shall register notifier block with::
+
+ static struct notifier_block pasid_nb_kvm = {
+	.notifier_call = pasid_status_change_kvm,
+	.priority      = IOASID_PRIO_CPU,
+ };
+
+VDCM driver shall register notifier block with::
+
+ static struct notifier_block pasid_nb_vdcm = {
+	.notifier_call = pasid_status_change_vdcm,
+	.priority      = IOASID_PRIO_DEVICE,
+ };
+
+In both cases, notifier blocks shall be registered on the IOASID set
+such that *only* events from the matching VM is received.
+
+If KVM attempts to register notifier block before the IOASID set is
+created for the MM token, the notifier block will be placed on a
+pending list inside IOASID core. Once the token matching IOASID set
+is created, IOASID will register the notifier block automatically.
+IOASID core does not replay events for the existing IOASIDs in the
+set. For IOASID set of MM type, notification blocks can be registered
+on empty sets only. This is to avoid lost events.
+
+IOMMU driver shall register notifier block on global chain::
+
+ static struct notifier_block pasid_nb_vtd = {
+	.notifier_call = pasid_status_change_vtd,
+	.priority      = IOASID_PRIO_IOMMU,
+ };
+
+Custom allocator APIs
+---------------------
+
+::
+
+ int ioasid_register_allocator(struct ioasid_allocator_ops *allocator);
+
+ void ioasid_unregister_allocator(struct ioasid_allocator_ops *allocator);
+
+Allocator Choices
+~~~~~~~~~~~~~~~~~
+IOASIDs are allocated for both host and guest SVA/IOVA usage. However,
+allocators can be different. For example, on VT-d guest PASID
+allocation must be performed via a virtual command interface which is
+emulated by VMM.
+

Documentation/ioasid.rst line 321

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+IOASID core has the notion of "custom allocator" such that guest can
+register virtual command allocator that precedes the default one.
+
+Namespaces
+~~~~~~~~~~
+IOASIDs are limited system resources that default to 20 bits in
+size. Since each device has its own table, theoretically the namespace
+can be per device also. However, for security reasons sharing PASID
+tables among devices are not good for isolation. Therefore, IOASID
+namespace is system-wide.

I don't follow this development. Having per-device PASID table would work fine for isolation (assuming no hardware bug necessitating IOMMU groups). If I remember correctly IOASID space was chosen to be OS-wide because it simplifies the management code (single PASID per task), and it is system-wide across VMs only in the case of VT-d scalable mode.

Documentation/ioasid.rst line 325

Quoted 4 lines
Documentation/ioasid.rst new file +4 -0
+
+There are also other reasons to have this simpler system-wide
+namespace. Take VT-d as an example, VT-d supports shared workqueue
+and ENQCMD[1] where one IOASID could be used to submit work on

Maybe use the Sphinx glossary syntax rather than "[1]"
https://www.sphinx-doc.org/en/master/usage/restructuredtext/directives.html#glossary-directive

Quoted 30 lines
Documentation/ioasid.rst new file +30 -0
+multiple devices that are shared with other VMs. This requires IOASID
+to be system-wide. This is also the reason why guests must use an
+emulated virtual command interface to allocate IOASID from the host.
+
+
+Life cycle
+==========
+This section covers IOASID lifecycle management for both bare-metal
+and guest usages. In bare-metal SVA, MMU notifier is directly hooked
+up with IOMMU driver, therefore the process address space (MM)
+lifecycle is aligned with IOASID.
+
+However, guest MMU notifier is not available to host IOMMU driver,
+when guest MM terminates unexpectedly, the events have to go through
+VFIO and IOMMU UAPI to reach host IOMMU driver. There are also more
+parties involved in guest SVA, e.g. on Intel VT-d platform, IOASIDs
+are used by IOMMU driver, KVM, VDCM, and VFIO.
+
+Native IOASID Life Cycle (VT-d Example)
+---------------------------------------
+
+The normal flow of native SVA code with Intel Data Streaming
+Accelerator(DSA) [2] as example:
+
+1. Host user opens accelerator FD, e.g. DSA driver, or uacce;
+2. DSA driver allocate WQ, do sva_bind_device();
+3. IOMMU driver calls ioasid_alloc(), then bind PASID with device,
+   mmu_notifier_get()
+4. DMA starts by DSA driver userspace
+5. DSA userspace close FD

Documentation/ioasid.rst line 365

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+6. DSA/uacce kernel driver handles FD.close()
+7. DSA driver stops DMA
+8. DSA driver calls sva_unbind_device();
+9. IOMMU driver does unbind, clears PASID context in IOMMU, flush
+   TLBs. mmu_notifier_put() called.
+10. mmu_notifier.release() called, IOMMU SVA code calls ioasid_free()*
+11. The IOASID is returned to the pool, reclaimed.
+
+::
+

Use a footnote? https://www.sphinx-doc.org/en/master/usage/restructuredtext/basics.html#footnotes

Documentation/ioasid.rst line 367

Quoted 2 lines
Documentation/ioasid.rst new file +2 -0
+   * With ENQCMD, PASID used on VT-d is not released in mmu_notifier() but
+     mmdrop(). mmdrop comes after FD close. Should not matter.

"comes after FD close, which doesn't make a difference?"
The following might not be necessary since early process termination is described later.

Documentation/ioasid.rst line 369

Quoted 2 lines
Documentation/ioasid.rst new file +2 -0
+     If the user process dies unexpectedly, Step #10 may come before
+     Step #5, in between, all DMA faults discarded. PRQ responded with

PRQ hasn't been defined in this document.

Documentation/ioasid.rst line 381

Quoted 12 lines
Documentation/ioasid.rst new file +12 -0
+     code INVALID REQUEST.
+
+During the normal teardown, the following three steps would happen in
+order:
+
+1. Device driver stops DMA request
+2. IOMMU driver unbinds PASID and mm, flush all TLBs, drain in-flight
+   requests.
+3. IOASID freed
+
+Exception happens when process terminates *before* device driver stops
+DMA and call IOMMU driver to unbind. The flow of process exists are as

"exits"

Documentation/ioasid.rst line 394

Quoted 13 lines
Documentation/ioasid.rst new file +13 -0
+follows:
+
+::
+
+   do_exit() {
+	exit_mm() {
+		mm_put();
+		exit_mmap() {
+			intel_invalidate_range() //mmu notifier
+			tlb_finish_mmu()
+			mmu_notifier_release(mm) {
+				intel_iommu_release() {
+   [2]					intel_iommu_teardown_pasid();

Parentheses might be better than square brackets for step numbers

Quoted 92 lines
Documentation/ioasid.rst new file +92 -0
+                                        intel_iommu_flush_tlbs();
+				}
+				// tlb_invalidate_range cb removed
+			}
+			unmap_vmas();
+                        free_pgtables(); // IOMMU cannot walk PGT after this
+		};
+	}
+	exit_files(tsk) {
+		close_files() {
+			dsa_close();
+   [1]			dsa_stop_dma();
+                        intel_svm_unbind_pasid(); //nothing to do
+		}
+	}
+   }
+
+   mmdrop() /* some random time later, lazy mm user */ {
+   	mm_free_pgd();
+        destroy_context(mm); {
+   [3]	        ioasid_free();
+	}
+   }
+
+As shown in the list above, step #2 could happen before
+#1. Unrecoverable(UR) faults could happen between #2 and #1.
+
+Also notice that TLB invalidation occurs at mmu_notifier
+invalidate_range callback as well as the release callback. The reason
+is that release callback will delete IOMMU driver from the notifier
+chain which may skip invalidate_range() calls during the exit path.
+
+To avoid unnecessary reporting of UR fault, IOMMU driver shall disable
+fault reporting after free and before unbind.
+
+Guest IOASID Life Cycle (VT-d Example)
+--------------------------------------
+Guest IOASID life cycle starts with guest driver open(), this could be
+uacce or individual accelerator driver such as DSA. At FD open,
+sva_bind_device() is called which triggers a series of actions.
+
+The example below is an illustration of *normal* operations that
+involves *all* the SW components in VT-d. The flow can be simpler if
+no ENQCMD is supported.
+
+::
+
+     VFIO        IOMMU        KVM        VDCM        IOASID       Ref
+   ..................................................................
+   1             ioasid_register_notifier/_mm()
+   2 ioasid_alloc()                                               1
+   3 bind_gpasid()
+   4             iommu_bind()->ioasid_get()                       2
+   5             ioasid_notify(BIND)
+   6                          -> ioasid_get()                     3
+   7                          -> vmcs_update_atomic()
+   8 mdev_write(gpasid)
+   9                                    hpasid=
+   10                                   find_by_spid(gpasid)      4
+   11                                   vdev_write(hpasid)
+   12 -------- GUEST STARTS DMA --------------------------
+   13 -------- GUEST STOPS DMA --------------------------
+   14 mdev_clear(gpasid)
+   15                                   vdev_clear(hpasid)
+   16                                   ioasid_put()               3
+   17 unbind_gpasid()
+   18            iommu_ubind()
+   19            ioasid_notify(UNBIND)
+   20                          -> vmcs_update_atomic()
+   21                          -> ioasid_put()                     2
+   22 ioasid_free()                                                1
+   23            ioasid_put()                                      0
+   24                                                 Reclaimed
+   -------------- New Life Cycle Begin ----------------------------
+   1  ioasid_alloc()                                  ->           1
+
+   Note: IOASID Notification Events: FREE, BIND, UNBIND
+
+Exception cases arise when a guest crashes or a malicious guest
+attempts to cause disruption on the host system. The fault handling
+rules are:
+
+1. IOASID free must *always* succeed.
+2. An inactive period may be required before the freed IOASID is
+   reclaimed. During this period, consumers of IOASID perform cleanup.
+3. Malfunction is limited to the guest owned resources for all
+   programming errors.
+
+The primary source of exception is when the following are out of
+order:
+
+1. Start/Stop of DMA activity

Documentation/ioasid.rst line 496

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+   (Guest device driver, mdev via VFIO)
+2. Setup/Teardown of IOMMU PASID context, IOTLB, DevTLB flushes
+   (Host IOMMU driver bind/unbind)
+3. Setup/Teardown of VMCS PASID translation table entries (KVM) in
+   case of ENQCMD
+4. Programming/Clearing host PASID in VDCM (Host VDCM driver)
+5. IOASID alloc/free (Host IOASID)
+
+VFIO is the *only* user-kernel interface, which is ultimately
+responsible for exception handlings.

"handling"

Quoted 52 lines
Documentation/ioasid.rst new file +52 -0
+
+#1 is processed the same way as the assigned device today based on
+device file descriptors and events. There is no special handling.
+
+#3 is based on bind/unbind events emitted by #2.
+
+#4 is naturally aligned with IOASID life cycle in that an illegal
+guest PASID programming would fail in obtaining reference of the
+matching host IOASID.
+
+#5 is similar to #4. The fault will be reported to the user if PASID
+used in the ENQCMD is not set up in VMCS PASID translation table.
+
+Therefore, the remaining out of order problem is between #2 and
+#5. I.e. unbind vs. free. More specifically, free before unbind.
+
+IOASID notifier and refcounting are used to ensure order. Following
+a publisher-subscriber pattern where:
+
+- Publishers: VFIO & IOMMU
+- Subscribers: KVM, VDCM, IOMMU
+
+IOASID notifier is atomic which requires subscribers to do quick
+handling of the event in the atomic context. Workqueue can be used for
+any processing that requires thread context. IOASID reference must be
+acquired before receiving the FREE event. The reference must be
+dropped at the end of the processing in order to return the IOASID to
+the pool.
+
+Let's examine the IOASID life cycle again when free happens *before*
+unbind. This could be a result of misbehaving guests or crash. Assuming
+VFIO cannot enforce unbind->free order. Notice that the setup part up
+until step #12 is identical to the normal case, the flow below starts
+with step 13.
+
+::
+
+     VFIO        IOMMU        KVM        VDCM        IOASID       Ref
+   ..................................................................
+   13 -------- GUEST STARTS DMA --------------------------
+   14 -------- *GUEST MISBEHAVES!!!* ----------------
+   15 ioasid_free()
+   16                                             ioasid_notify(FREE)
+   17                                             mark_ioasid_inactive[1]
+   18                          kvm_nb_handler(FREE)
+   19                          vmcs_update_atomic()
+   20                          ioasid_put_locked()   ->           3
+   21                                   vdcm_nb_handler(FREE)
+   22            iomm_nb_handler(FREE)
+   23 ioasid_free() returns[2]          schedule_work()           2
+   24            schedule_work()        vdev_clear_wk(hpasid)
+   25            teardown_pasid_wk()

Documentation/ioasid.rst line 558

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+   26                                   ioasid_put() ->           1
+   27            ioasid_put()                                     0
+   28                                                 Reclaimed
+   29 unbind_gpasid()
+   30            iommu_unbind()->ioasid_find() Fails[3]
+   -------------- New Life Cycle Begin ----------------------------
+
+Note:
+
+1. By marking IOASID inactive at step #17, no new references can be

Is "inactive" FREE_PENDING?

Quoted 30 lines
Documentation/ioasid.rst new file +30 -0
+   held. ioasid_get/find() will return -ENOENT;
+2. After step #23, all events can go out of order. Shall not affect
+   the outcome.
+3. IOMMU driver fails to find private data for unbinding. If unbind is
+   called after the same IOASID is allocated for the same guest again,
+   this is a programming error. The damage is limited to the guest
+   itself since unbind performs permission checking based on the
+   IOASID set associated with the guest process.
+
+KVM PASID Translation Table Updates
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Per VM PASID translation table is maintained by KVM in order to
+support ENQCMD in the guest. The table contains host-guest PASID
+translations to be consumed by CPU ucode. The synchronization of the
+PASID states depends on VFIO/IOMMU driver, where IOCTL and atomic
+notifiers are used. KVM must register IOASID notifier per VM instance
+during launch time. The following events are handled:
+
+1. BIND/UNBIND
+2. FREE
+
+Rules:
+
+1. Multiple devices can bind with the same PASID, this can be different PCI
+   devices or mdevs within the same PCI device. However, only the
+   *first* BIND and *last* UNBIND emit notifications.
+2. IOASID code is responsible for ensuring the correctness of H-G
+   PASID mapping. There is no need for KVM to validate the
+   notification data.
+3. When UNBIND happens *after* FREE, KVM will see error in

Documentation/ioasid.rst line 598

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+   ioasid_get() even when the reclaim is not done. IOMMU driver will
+   also avoid sending UNBIND if the PASID is already FREE.
+4. When KVM terminates *before* FREE & UNBIND, references will be
+   dropped for all host PASIDs.
+
+VDCM PASID Programming
+~~~~~~~~~~~~~~~~~~~~~~
+VDCM composes virtual devices and exposes them to the guests. When
+the guest allocates a PASID then program it to the virtual device, VDCM
+intercepts the programming attempt then program the matching host

"programs"

Thanks,
Jean

Quoted 23 lines
Documentation/ioasid.rst new file +20 -0
+PASID on to the hardware.
+Conversely, when a device is going away, VDCM must be informed such
+that PASID context on the hardware can be cleared. There could be
+multiple mdevs assigned to different guests in the same VDCM. Since
+the PASID table is shared at PCI device level, lazy clearing is not
+secure. A malicious guest can attack by using newly freed PASIDs that
+are allocated by another guest.
+
+By holding a reference of the PASID until VDCM cleans up the HW context,
+it is guaranteed that PASID life cycles do not cross within the same
+device.
+
+
+Reference
+====================================================
+1. https://software.intel.com/sites/default/files/managed/c5/15/architecture-instruction-set-extensions-programming-reference.pdf
+
+2. https://01.org/blogs/2019/introducing-intel-data-streaming-accelerator
+
+3. https://software.intel.com/en-us/download/intel-data-streaming-accelerator-preliminary-architecture-specification
Signature · 1 line
2.7.4

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

#18 lore raw

Re: [PATCH v2 3/9] iommu/ioasid: Introduce ioasid_set APIs

In reply to Jacob Pan
Quoted 169 lines · Jacob Pan

ioasid_set was introduced as an arbitrary token that are shared by a group of IOASIDs. For example, if IOASID #1 and #2 are allocated via the same ioasid_set*, they are viewed as to belong to the same set.

For guest SVA usages, system-wide IOASID resources need to be partitioned such that VMs can have its own quota and being managed separately. ioasid_set is the perfect candidate for meeting such requirements. This patch redefines and extends ioasid_set with the following new fields:
- Quota
- Reference count
- Storage of its namespace
- The token is stored in the new ioasid_set but with optional types

ioasid_set level APIs are introduced that wires up these new data. Existing users of IOASID APIs are converted where a host IOASID set is allocated for bare-metal usage.

---
 drivers/iommu/intel/iommu.c |  27 ++-
 drivers/iommu/intel/pasid.h |   1 +
 drivers/iommu/intel/svm.c   |   8 +-
 drivers/iommu/ioasid.c      | 390 +++++++++++++++++++++++++++++++++++++++++---
 include/linux/ioasid.h      |  82 ++++++++--
 5 files changed, 465 insertions(+), 43 deletions(-)
drivers/iommu/intel/iommu.c +22 -5
index a3a0b5c8921d..5813eeaa5edb 100644
@@ -42,6 +42,7 @@
 #include <linux/crash_dump.h>
 #include <linux/numa.h>
 #include <linux/swiotlb.h>
+#include <linux/ioasid.h>
 #include <asm/irq_remapping.h>
 #include <asm/cacheflush.h>
 #include <asm/iommu.h>
@@ -103,6 +104,9 @@
  */
 #define INTEL_IOMMU_PGSIZES	(~0xFFFUL)
 
+/* PASIDs used by host SVM */
+struct ioasid_set *host_pasid_set;
+
 static inline int agaw_to_level(int agaw)
 {
 	return agaw + 2;
@@ -3103,8 +3107,8 @@ static void intel_vcmd_ioasid_free(ioasid_t ioasid, void *data)
 	 * Sanity check the ioasid owner is done at upper layer, e.g. VFIO
 	 * We can only free the PASID when all the devices are unbound.
 	 */
-	if (ioasid_find(NULL, ioasid, NULL)) {
-		pr_alert("Cannot free active IOASID %d\n", ioasid);
+	if (IS_ERR(ioasid_find(host_pasid_set, ioasid, NULL))) {
+		pr_err("Cannot free IOASID %d, not in system set\n", ioasid);
 		return;
 	}
 	vcmd_free_pasid(iommu, ioasid);
@@ -3288,6 +3292,19 @@ static int __init init_dmars(void)
 	if (ret)
 		goto free_iommu;
 
+	/* PASID is needed for scalable mode irrespective to SVM */
+	if (intel_iommu_sm) {
+		ioasid_install_capacity(intel_pasid_max_id);
+		/* We should not run out of IOASIDs at boot */
+		host_pasid_set = ioasid_alloc_set(NULL, PID_MAX_DEFAULT,
+						IOASID_SET_TYPE_NULL);
+		if (IS_ERR_OR_NULL(host_pasid_set)) {
+			pr_err("Failed to enable host PASID allocator %lu\n",
+				PTR_ERR(host_pasid_set));
+			intel_iommu_sm = 0;
+		}
+	}
+
 	/*
 	 * for each drhd
 	 *   enable fault log
@@ -5149,7 +5166,7 @@ static void auxiliary_unlink_device(struct dmar_domain *domain,
 	domain->auxd_refcnt--;
 
 	if (!domain->auxd_refcnt && domain->default_pasid > 0)
-		ioasid_free(domain->default_pasid);
+		ioasid_free(host_pasid_set, domain->default_pasid);
 }
 
 static int aux_domain_add_dev(struct dmar_domain *domain,
@@ -5167,7 +5184,7 @@ static int aux_domain_add_dev(struct dmar_domain *domain,
 		int pasid;
 
 		/* No private data needed for the default pasid */
-		pasid = ioasid_alloc(NULL, PASID_MIN,
+		pasid = ioasid_alloc(host_pasid_set, PASID_MIN,
 				     pci_max_pasids(to_pci_dev(dev)) - 1,
 				     NULL);
 		if (pasid == INVALID_IOASID) {
@@ -5210,7 +5227,7 @@ static int aux_domain_add_dev(struct dmar_domain *domain,
 	spin_unlock(&iommu->lock);
 	spin_unlock_irqrestore(&device_domain_lock, flags);
 	if (!domain->auxd_refcnt && domain->default_pasid > 0)
-		ioasid_free(domain->default_pasid);
+		ioasid_free(host_pasid_set, domain->default_pasid);
 
 	return ret;
 }
drivers/iommu/intel/pasid.h +1 -0
index c9850766c3a9..ccdc23446015 100644
@@ -99,6 +99,7 @@ static inline bool pasid_pte_is_present(struct pasid_entry *pte)
 }
 
 extern u32 intel_pasid_max_id;
+extern struct ioasid_set *host_pasid_set;
 int intel_pasid_alloc_id(void *ptr, int start, int end, gfp_t gfp);
 void intel_pasid_free_id(int pasid);
 void *intel_pasid_lookup_id(int pasid);
drivers/iommu/intel/svm.c +4 -4
index 37a9beabc0ca..634e191ca2c3 100644
@@ -551,7 +551,7 @@ intel_svm_bind_mm(struct device *dev, int flags, struct svm_dev_ops *ops,
 			pasid_max = intel_pasid_max_id;
 
 		/* Do not use PASID 0, reserved for RID to PASID */
-		svm->pasid = ioasid_alloc(NULL, PASID_MIN,
+		svm->pasid = ioasid_alloc(host_pasid_set, PASID_MIN,
 					  pasid_max - 1, svm);
 		if (svm->pasid == INVALID_IOASID) {
 			kfree(svm);
@@ -568,7 +568,7 @@ intel_svm_bind_mm(struct device *dev, int flags, struct svm_dev_ops *ops,
 		if (mm) {
 			ret = mmu_notifier_register(&svm->notifier, mm);
 			if (ret) {
-				ioasid_free(svm->pasid);
+				ioasid_free(host_pasid_set, svm->pasid);
 				kfree(svm);
 				kfree(sdev);
 				goto out;
@@ -586,7 +586,7 @@ intel_svm_bind_mm(struct device *dev, int flags, struct svm_dev_ops *ops,
 		if (ret) {
 			if (mm)
 				mmu_notifier_unregister(&svm->notifier, mm);
-			ioasid_free(svm->pasid);
+			ioasid_free(host_pasid_set, svm->pasid);
 			kfree(svm);
 			kfree(sdev);
 			goto out;
@@ -655,7 +655,7 @@ static int intel_svm_unbind_mm(struct device *dev, int pasid)
 			kfree_rcu(sdev, rcu);
 
 			if (list_empty(&svm->devs)) {
-				ioasid_free(svm->pasid);
+				ioasid_free(host_pasid_set, svm->pasid);
 				if (svm->mm)
 					mmu_notifier_unregister(&svm->notifier, svm->mm);
 				list_del(&svm->list);
drivers/iommu/ioasid.c +0 -1
index 5f63af07acd5..f73b3dbfc37a 100644
@@ -1,22 +1,58 @@
 // SPDX-License-Identifier: GPL-2.0
 /*
  * I/O Address Space ID allocator. There is one global IOASID space, split into
- * subsets. Users create a subset with DECLARE_IOASID_SET, then allocate and

drivers/iommu/ioasid.c line 11

Quoted 10 lines
drivers/iommu/ioasid.c +3 -2
- * free IOASIDs with ioasid_alloc and ioasid_free.
+ * subsets. Users create a subset with ioasid_alloc_set, then allocate/free IDs
+ * with ioasid_alloc and ioasid_free.
  */
-#include <linux/ioasid.h>
 #include <linux/module.h>
 #include <linux/slab.h>
 #include <linux/spinlock.h>
 #include <linux/xarray.h>
+#include <linux/ioasid.h>

Spurious change (best keep the includes in alphabetical order)

drivers/iommu/ioasid.c line 13

Quoted 2 lines
drivers/iommu/ioasid.c +2 -0
+
+static DEFINE_XARRAY_ALLOC(ioasid_sets);

I'd prefer keeping all static variables together

Quoted 5 lines
drivers/iommu/ioasid.c +5 -0
+enum ioasid_state {
+	IOASID_STATE_INACTIVE,
+	IOASID_STATE_ACTIVE,
+	IOASID_STATE_FREE_PENDING,
+};

drivers/iommu/ioasid.c line 28

Quoted 10 lines
drivers/iommu/ioasid.c +9 -0
 
+/**
+ * struct ioasid_data - Meta data about ioasid
+ *
+ * @id:		Unique ID
+ * @users	Number of active users
+ * @state	Track state of the IOASID
+ * @set		Meta data of the set this IOASID belongs to
+ * @private	Private data associated with the IOASID
+ * @rcu		For free after RCU grace period

nit: it would be nicer to follow the struct order

Quoted 31 lines
drivers/iommu/ioasid.c +20 -0
+ */
 struct ioasid_data {
 	ioasid_t id;
 	struct ioasid_set *set;
+	refcount_t users;
+	enum ioasid_state state;
 	void *private;
 	struct rcu_head rcu;
 };
 
+/* Default to PCIe standard 20 bit PASID */
+#define PCI_PASID_MAX 0x100000
+static ioasid_t ioasid_capacity = PCI_PASID_MAX;
+static ioasid_t ioasid_capacity_avail = PCI_PASID_MAX;
+
+void ioasid_install_capacity(ioasid_t total)
+{
+	ioasid_capacity = ioasid_capacity_avail = total;
+}
+EXPORT_SYMBOL_GPL(ioasid_install_capacity);
+
+ioasid_t ioasid_get_capacity()
+{
+	return ioasid_capacity;
+}
+EXPORT_SYMBOL_GPL(ioasid_get_capacity);
+
 /*
  * struct ioasid_allocator_data - Internal data structure to hold information
  * about an allocator. There are two types of allocators:

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 350

Quoted 10 lines
drivers/iommu/ioasid.c +6 -1
@@ -306,11 +342,23 @@ ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max,
 {
 	struct ioasid_data *data;
 	void *adata;
-	ioasid_t id;
+	ioasid_t id = INVALID_IOASID;
+
+	spin_lock(&ioasid_allocator_lock);
+	/* Check if the IOASID set has been allocated and initialized */
+	if (WARN_ON(xa_load(&ioasid_sets, set->sid) != set)) {
+		pr_warn("Invalid set\n");

WARN_ON() is sufficient

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 355

Quoted 5 lines
drivers/iommu/ioasid.c +5 -0
+		goto done_unlock;
+	}
+
+	if (set->quota <= set->nr_ioasids) {
+		pr_err("IOASID set %d out of quota %d\n", set->sid, set->quota);

As this can be called directly by userspace via VFIO, I wonder if we should remove non-bug error messages like this one to avoid leaking internal IDs, or at least rate-limit them. We already have a few, perhaps we should deal with them before the VFIO_IOMMU_ALLOC_PASID patches land?

Quoted 19 lines
drivers/iommu/ioasid.c +3 -2
+		goto done_unlock;
+	}
 
 	data = kzalloc(sizeof(*data), GFP_ATOMIC);
 	if (!data)
-		return INVALID_IOASID;
+		goto done_unlock;
 
 	data->set = set;
 	data->private = private;
@@ -319,7 +367,6 @@ ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max,
 	 * Custom allocator needs allocator data to perform platform specific
 	 * operations.
 	 */
-	spin_lock(&ioasid_allocator_lock);
 	adata = active_allocator->flags & IOASID_ALLOCATOR_CUSTOM ? active_allocator->ops->pdata : data;
 	id = active_allocator->ops->alloc(min, max, adata);
 	if (id == INVALID_IOASID) {

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 390

Quoted 9 lines
drivers/iommu/ioasid.c +6 -0
@@ -335,42 +382,339 @@ ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max,
 		goto exit_free;
 	}
 	data->id = id;
+	data->state = IOASID_STATE_ACTIVE;
+	refcount_set(&data->users, 1);
+
+	/* Store IOASID in the per set data */
+	if (xa_err(xa_store(&set->xa, id, data, GFP_ATOMIC))) {
+		pr_err("Failed to ioasid %d in set %d\n", id, set->sid);

"Failed to store"
Don't we need to call active_allocator->ops->free()?

And I need to think about this more, but do you see any issue with revoking here the data that we published into the xarray above through alloc()? We might need to free data in an RCU callback.

Quoted 24 lines
drivers/iommu/ioasid.c +14 -4
+		goto exit_free;
+	}
+	set->nr_ioasids++;
+	goto done_unlock;
 
-	spin_unlock(&ioasid_allocator_lock);
-	return id;
 exit_free:
-	spin_unlock(&ioasid_allocator_lock);
 	kfree(data);
-	return INVALID_IOASID;
+done_unlock:
+	spin_unlock(&ioasid_allocator_lock);
+	return id;
 }
 EXPORT_SYMBOL_GPL(ioasid_alloc);
 
+static void ioasid_do_free(struct ioasid_data *data)
+{
+	struct ioasid_data *ioasid_data;
+	struct ioasid_set *sdata;
+
+	active_allocator->ops->free(data->id, active_allocator->ops->pdata);
+	/* Custom allocator needs additional steps to free the xa element */

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 420

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+	if (active_allocator->flags & IOASID_ALLOCATOR_CUSTOM) {
+		ioasid_data = xa_erase(&active_allocator->xa, data->id);
+		kfree_rcu(ioasid_data, rcu);
+	}
+
+	sdata = xa_load(&ioasid_sets, data->set->sid);
+	if (!sdata) {
+		pr_err("No set %d for IOASID %d\n", data->set->sid,
+		       data->id);
+		return;

I don't think we're allowed to fail at this point. If we need more sanity-check on the parameters, it should be before we start removing from the active_allocator above. Otherwise this should be a WARN

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 423

Quoted 3 lines
drivers/iommu/ioasid.c +3 -0
+	}
+	xa_erase(&sdata->xa, data->id);
+	sdata->nr_ioasids--;

Would be nicer to perform the cleanup in the order opposite from ioasid_alloc()

Quoted 39 lines
drivers/iommu/ioasid.c +32 -4
+}
+
+static void ioasid_free_locked(struct ioasid_set *set, ioasid_t ioasid)
+{
+	struct ioasid_data *data;
+
+	data = xa_load(&active_allocator->xa, ioasid);
+	if (!data) {
+		pr_err("Trying to free unknown IOASID %u\n", ioasid);
+		return;
+	}
+
+	if (data->set != set) {
+		pr_warn("Cannot free IOASID %u due to set ownership\n", ioasid);
+		return;
+	}
+	data->state = IOASID_STATE_FREE_PENDING;
+
+	if (!refcount_dec_and_test(&data->users))
+		return;
+
+	ioasid_do_free(data);
+}
+
 /**
- * ioasid_free - Free an IOASID
- * @ioasid: the ID to remove
+ * ioasid_free - Drop reference on an IOASID. Free if refcount drops to 0,
+ *               including free from its set and system-wide list.
+ * @set:	The ioasid_set to check permission with. If not NULL, IOASID
+ *		free will fail if the set does not match.
+ * @ioasid:	The IOASID to remove
  */
-void ioasid_free(ioasid_t ioasid)
+void ioasid_free(struct ioasid_set *set, ioasid_t ioasid)
 {
-	struct ioasid_data *ioasid_data;
+	spin_lock(&ioasid_allocator_lock);
+	ioasid_free_locked(set, ioasid);

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 468

Quoted 10 lines
drivers/iommu/ioasid.c +9 -0
+	spin_unlock(&ioasid_allocator_lock);
+}
+EXPORT_SYMBOL_GPL(ioasid_free);
 
+/**
+ * ioasid_alloc_set - Allocate a new IOASID set for a given token
+ *
+ * @token:	Unique token of the IOASID set, cannot be NULL
+ * @quota:	Quota allowed in this set. Only for new set creation
+ * @flags:	Special requirements

There is no @flags, but @type is missing

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 471

Quoted 3 lines
drivers/iommu/ioasid.c +3 -0
+ *
+ * IOASID can be limited system-wide resource that requires quota management.
+ * If caller does not wish to enforce quota, use IOASID_SET_NO_QUOTA flag.

The flag isn't introduced in this patch. How about passing @quota == 0 in this case? For now I'm fine with leaving this as TODO and returning -EINVAL.

Quoted 36 lines
drivers/iommu/ioasid.c +24 -8
+ *
+ * Token will be stored in the ioasid_set returned. A reference will be taken
+ * upon finding a matching set or newly created set.
+ * IOASID allocation within the set and other per set operations will use
+ * the retured ioasid_set *.
+ *
+ */
+struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota, int type)
+{
+	struct ioasid_set *sdata;
+	unsigned long index;
+	ioasid_t id;
+
+	if (type >= IOASID_SET_TYPE_NR)
+		return ERR_PTR(-EINVAL);
+
+	/*
+	 * Need to check space available if we share system-wide quota.
+	 * TODO: we may need to support quota free sets in the future.
+	 */
 	spin_lock(&ioasid_allocator_lock);
-	ioasid_data = xa_load(&active_allocator->xa, ioasid);
-	if (!ioasid_data) {
-		pr_err("Trying to free unknown IOASID %u\n", ioasid);
+	if (quota > ioasid_capacity_avail) {
+		pr_warn("Out of IOASID capacity! ask %d, avail %d\n",
+			quota, ioasid_capacity_avail);
+		sdata = ERR_PTR(-ENOSPC);
 		goto exit_unlock;
 	}
 
-	active_allocator->ops->free(ioasid, active_allocator->ops->pdata);
-	/* Custom allocator needs additional steps to free the xa element */
-	if (active_allocator->flags & IOASID_ALLOCATOR_CUSTOM) {
-		ioasid_data = xa_erase(&active_allocator->xa, ioasid);
-		kfree_rcu(ioasid_data, rcu);

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 509

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+	/*
+	 * Token is only unique within its types but right now we have only
+	 * mm type. If we have more token types, we have to match type as well.
+	 */
+	switch (type) {
+	case IOASID_SET_TYPE_MM:
+		/* Search existing set tokens, reject duplicates */
+		xa_for_each(&ioasid_sets, index, sdata) {
+			if (sdata->token == token &&
+				sdata->type == IOASID_SET_TYPE_MM) {

Should be aligned at the "if ("

According to the function doc, shouldn't we take a reference to the set in this case, and return it?
"A reference will be taken upon finding a matching set or newly created set."

However it might be better to separate the two functionalities into ioasid_alloc_set() and ioasid_find_set(). Because two modules can want to work on the same set for an mm, but they won't pass the same @quota, right? So it'd make more sense for one of them (VFIO) to alloc the set and the other to reuse it.

Quoted 7 lines
drivers/iommu/ioasid.c +7 -0
+				sdata = ERR_PTR(-EEXIST);
+				goto exit_unlock;
+			}
+		}
+		break;
+	case IOASID_SET_TYPE_NULL:
+		if (!token)

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 526

Quoted 10 lines
drivers/iommu/ioasid.c +8 -0
+			break;
+		fallthrough;
+	default:
+		pr_err("Invalid token and IOASID type\n");
+		sdata = ERR_PTR(-EINVAL);
+		goto exit_unlock;
 	}
 
+	/* REVISIT: may support set w/o quota, use system available */
+	if (!quota) {

Maybe move this next to the other quota check above

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 538

Quoted 12 lines
drivers/iommu/ioasid.c +12 -0
+		sdata = ERR_PTR(-EINVAL);
+		goto exit_unlock;
+	}
+
+	sdata = kzalloc(sizeof(*sdata), GFP_ATOMIC);
+	if (!sdata) {
+		sdata = ERR_PTR(-ENOMEM);
+		goto exit_unlock;
+	}
+
+	if (xa_alloc(&ioasid_sets, &id, sdata,
+			XA_LIMIT(0, ioasid_capacity_avail - quota),

Why this limit? sid could just be an arbitrary u32 (xa_limit_32b)

Quoted 8 lines
drivers/iommu/ioasid.c +8 -0
+			GFP_ATOMIC)) {
+		kfree(sdata);
+		sdata = ERR_PTR(-ENOSPC);
+		goto exit_unlock;
+	}
+
+	sdata->token = token;
+	sdata->type = type;

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 556

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+	sdata->quota = quota;
+	sdata->sid = id;
+	refcount_set(&sdata->ref, 1);
+
+	/*
+	 * Per set XA is used to store private IDs within the set, get ready
+	 * for ioasid_set private ID and system-wide IOASID allocation
+	 * results.
+	 */
+	xa_init_flags(&sdata->xa, XA_FLAGS_ALLOC);

Since it's only used for storage, you could use xa_init()

Quoted 4 lines
drivers/iommu/ioasid.c +2 -0
+	ioasid_capacity_avail -= quota;
+
 exit_unlock:
 	spin_unlock(&ioasid_allocator_lock);

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 569

Quoted 10 lines
drivers/iommu/ioasid.c +8 -1
+
+	return sdata;
 }
-EXPORT_SYMBOL_GPL(ioasid_free);
+EXPORT_SYMBOL_GPL(ioasid_alloc_set);
+
+void ioasid_set_get_locked(struct ioasid_set *set)
+{
+	if (WARN_ON(xa_load(&ioasid_sets, set->sid) != set)) {
+		pr_warn("Invalid set data\n");

WARN_ON() is sufficient

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 575

Quoted 6 lines
drivers/iommu/ioasid.c +6 -0
+		return;
+	}
+
+	refcount_inc(&set->ref);
+}
+EXPORT_SYMBOL_GPL(ioasid_set_get_locked);

Why is this function public, is it for an iterator? Might want to add a lockdep_assert_held() annotation.

Quoted 6 lines
drivers/iommu/ioasid.c +6 -0
+
+void ioasid_set_get(struct ioasid_set *set)
+{
+	spin_lock(&ioasid_allocator_lock);
+	ioasid_set_get_locked(set);
+	spin_unlock(&ioasid_allocator_lock);

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 591

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+}
+EXPORT_SYMBOL_GPL(ioasid_set_get);
+
+void ioasid_set_put_locked(struct ioasid_set *set)
+{
+	struct ioasid_data *entry;
+	unsigned long index;
+
+	if (WARN_ON(xa_load(&ioasid_sets, set->sid) != set)) {
+		pr_warn("Invalid set data\n");

WARN_ON() is sufficient

Quoted 30 lines
drivers/iommu/ioasid.c +30 -0
+		return;
+	}
+
+	if (!refcount_dec_and_test(&set->ref)) {
+		pr_debug("%s: IOASID set %d has %d users\n",
+			__func__, set->sid, refcount_read(&set->ref));
+		return;
+	}
+
+	/* The set is already empty, we just destroy the set. */
+	if (xa_empty(&set->xa))
+		goto done_destroy;
+
+	/*
+	 * Free all PASIDs from system-wide IOASID pool, all subscribers gets
+	 * notified and do cleanup of their own.
+	 * Note that some references of the IOASIDs within the set can still
+	 * be held after the free call. This is OK in that the IOASIDs will be
+	 * marked inactive, the only operations can be done is ioasid_put.
+	 * No need to track IOASID set states since there is no reclaim phase.
+	 */
+	xa_for_each(&set->xa, index, entry) {
+		ioasid_free_locked(set, index);
+		/* Free from per set private pool */
+		xa_erase(&set->xa, index);
+	}
+
+done_destroy:
+	/* Return the quota back to system pool */
+	ioasid_capacity_avail += set->quota;

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 631

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+	kfree_rcu(set, rcu);
+
+	/*
+	 * Token got released right away after the ioasid_set is freed.
+	 * If a new set is created immediately with the newly released token,
+	 * it will not allocate the same IOASIDs unless they are reclaimed.
+	 */
+	xa_erase(&ioasid_sets, set->sid);
+}
+EXPORT_SYMBOL_GPL(ioasid_set_put_locked);

Same comment as ioasid_set_get_locked

Quoted 36 lines
drivers/iommu/ioasid.c +36 -0
+
+/**
+ * ioasid_set_put - Drop a reference to the IOASID set. Free all IOASIDs within
+ *                  the set if there are no more users.
+ *
+ * @set:	The IOASID set ID to be freed
+ *
+ * If refcount drops to zero, all IOASIDs allocated within the set will be
+ * freed.
+ */
+void ioasid_set_put(struct ioasid_set *set)
+{
+	spin_lock(&ioasid_allocator_lock);
+	ioasid_set_put_locked(set);
+	spin_unlock(&ioasid_allocator_lock);
+}
+EXPORT_SYMBOL_GPL(ioasid_set_put);
+
+/**
+ * ioasid_adjust_set - Adjust the quota of an IOASID set
+ * @set:	IOASID set to be assigned
+ * @quota:	Quota allowed in this set
+ *
+ * Return 0 on success. If the new quota is smaller than the number of
+ * IOASIDs already allocated, -EINVAL will be returned. No change will be
+ * made to the existing quota.
+ */
+int ioasid_adjust_set(struct ioasid_set *set, int quota)
+{
+	int ret = 0;
+
+	spin_lock(&ioasid_allocator_lock);
+	if (set->nr_ioasids > quota) {
+		pr_err("New quota %d is smaller than outstanding IOASIDs %d\n",
+			quota, set->nr_ioasids);
+		ret = -EINVAL;

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 677

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+		goto done_unlock;
+	}
+
+	if (quota >= ioasid_capacity_avail) {
+		ret = -ENOSPC;
+		goto done_unlock;
+	}
+
+	/* Return the delta back to system pool */
+	ioasid_capacity_avail += set->quota - quota;

I think this is correct as long as the above check is fixed (as pointed out by Baolu). A check that quota >= 0 could be nice too.

Quoted 19 lines
drivers/iommu/ioasid.c +19 -0
+
+	/*
+	 * May have a policy to prevent giving all available IOASIDs
+	 * to one set. But we don't enforce here, it should be in the
+	 * upper layers.
+	 */
+	set->quota = quota;
+
+done_unlock:
+	spin_unlock(&ioasid_allocator_lock);
+
+	return ret;
+}
+EXPORT_SYMBOL_GPL(ioasid_adjust_set);
+
+/**
+ * ioasid_set_for_each_ioasid - Iterate over all the IOASIDs within the set
+ *
+ * Caller must hold a reference of the set and handles its own locking.

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 706

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+ */
+int ioasid_set_for_each_ioasid(struct ioasid_set *set,
+			       void (*fn)(ioasid_t id, void *data),
+			       void *data)
+{
+	struct ioasid_data *entry;
+	unsigned long index;
+	int ret = 0;
+
+	if (xa_empty(&set->xa)) {

Who calls this function? Empty xa may be a normal use-case if the caller just uses it for sweeping, so pr_warn() could be problematic. The returned value also isn't particularly accurate if concurrent ioasid_alloc/free are allowed, so I'd drop this.

Quoted 24 lines
drivers/iommu/ioasid.c +11 -0
+		pr_warn("No IOASIDs in the set %d\n", set->sid);
+		return -ENOENT;
+	}
+
+	xa_for_each(&set->xa, index, entry) {
+		fn(index, data);
+	}
+
+	return ret;
+}
+EXPORT_SYMBOL_GPL(ioasid_set_for_each_ioasid);
 
 /**
  * ioasid_find - Find IOASID data
include/linux/ioasid.h +2 -0
index 9c44947a68c8..412d025d440e 100644
@@ -10,8 +10,35 @@ typedef unsigned int ioasid_t;
 typedef ioasid_t (*ioasid_alloc_fn_t)(ioasid_t min, ioasid_t max, void *data);
 typedef void (*ioasid_free_fn_t)(ioasid_t ioasid, void *data);
 
+/* IOASID set types */
+enum ioasid_set_type {

include/linux/ioasid.h typedef unsigned int ioasid_t; · line 24

Quoted 10 lines
include/linux/ioasid.h +10 -0
+	IOASID_SET_TYPE_NULL = 1, /* Set token is NULL */
+	IOASID_SET_TYPE_MM,	  /* Set token is a mm_struct,
+				   * i.e. associated with a process
+				   */
+	IOASID_SET_TYPE_NR,
+};
+
+/**
+ * struct ioasid_set - Meta data about ioasid_set
+ * @type:	Token types and other features

nit: doesn't follow struct order

Quoted 25 lines
include/linux/ioasid.h +16 -1
+ * @token:	Unique to identify an IOASID set
+ * @xa:		XArray to store ioasid_set private IDs, can be used for
+ *		guest-host IOASID mapping, or just a private IOASID namespace.
+ * @quota:	Max number of IOASIDs can be allocated within the set
+ * @nr_ioasids	Number of IOASIDs currently allocated in the set
+ * @sid:	ID of the set
+ * @ref:	Reference count of the users
+ */
 struct ioasid_set {
-	int dummy;
+	void *token;
+	struct xarray xa;
+	int type;
+	int quota;
+	int nr_ioasids;
+	int sid;
+	refcount_t ref;
+	struct rcu_head rcu;
 };
 
 /**
@@ -29,31 +56,64 @@ struct ioasid_allocator_ops {
 	void *pdata;
 };
 

include/linux/ioasid.h struct ioasid_allocator_ops { · line 66

Quoted 10 lines
include/linux/ioasid.h +7 -2
-#define DECLARE_IOASID_SET(name) struct ioasid_set name = { 0 }
-
 #if IS_ENABLED(CONFIG_IOASID)
+void ioasid_install_capacity(ioasid_t total);
+ioasid_t ioasid_get_capacity(void);
+struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota, int type);
+int ioasid_adjust_set(struct ioasid_set *set, int quota);
+void ioasid_set_get_locked(struct ioasid_set *set);
+void ioasid_set_put_locked(struct ioasid_set *set);
+void ioasid_set_put(struct ioasid_set *set);

These three functions need a stub for !CONFIG_IOASID

include/linux/ioasid.h struct ioasid_allocator_ops { · line 72

Quoted 9 lines
include/linux/ioasid.h +4 -3
+
 ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max,
 		      void *private);
-void ioasid_free(ioasid_t ioasid);
-void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid,
-		  bool (*getter)(void *));
+void ioasid_free(struct ioasid_set *set, ioasid_t ioasid);
+
+bool ioasid_is_active(ioasid_t ioasid);

Not implemented by this series?

include/linux/ioasid.h struct ioasid_allocator_ops { · line 74

Quoted 2 lines
include/linux/ioasid.h +2 -0
+
+void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid, bool (*getter)(void *));

Spurious change

include/linux/ioasid.h struct ioasid_allocator_ops { · line 42

Quoted 4 lines
include/linux/ioasid.h +1 -1
+int ioasid_attach_data(ioasid_t ioasid, void *data);
 int ioasid_register_allocator(struct ioasid_allocator_ops *allocator);
 void ioasid_unregister_allocator(struct ioasid_allocator_ops *allocator);
-int ioasid_attach_data(ioasid_t ioasid, void *data);

Spurious change?

include/linux/ioasid.h struct ioasid_allocator_ops { · line 78

Quoted 2 lines
include/linux/ioasid.h +1 -1
-
+void ioasid_is_in_set(struct ioasid_set *set, ioasid_t ioasid);

Not implemented here

include/linux/ioasid.h struct ioasid_allocator_ops { · line 81

Quoted 3 lines
include/linux/ioasid.h +3 -0
+int ioasid_set_for_each_ioasid(struct ioasid_set *sdata,
+			       void (*fn)(ioasid_t id, void *data),
+			       void *data);

Needs a stub for !CONFIG_IOASID

Quoted 28 lines
include/linux/ioasid.h +20 -1
 #else /* !CONFIG_IOASID */
+static inline void ioasid_install_capacity(ioasid_t total)
+{
+}
+
+static inline ioasid_t ioasid_get_capacity(void)
+{
+	return 0;
+}
+
 static inline ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min,
 				    ioasid_t max, void *private)
 {
 	return INVALID_IOASID;
 }
 
-static inline void ioasid_free(ioasid_t ioasid)
+static inline void ioasid_free(struct ioasid_set *set, ioasid_t ioasid)
+{
+}
+
+static inline bool ioasid_is_active(ioasid_t ioasid)
+{
+	return false;
+}
+
+static inline struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota, int type)
+{

include/linux/ioasid.h struct ioasid_allocator_ops { · line 116

Quoted 10 lines
include/linux/ioasid.h +5 -2
+	return ERR_PTR(-ENOTSUPP);
+}
+
+static inline void ioasid_set_put(struct ioasid_set *set)
 {
 }
 
-static inline void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid,
-				bool (*getter)(void *))
+static inline void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid, bool (*getter)(void *))

Spurious change

Thanks,
Jean

Quoted 6 lines
include/linux/ioasid.h +0 -0
 {
 	return NULL;
 }
Signature · 1 line
2.7.4

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

Reviewed-by on v2 2/9
#11 lore raw

Re: [PATCH v2 2/9] iommu/ioasid: Rename ioasid_set_data()

In reply to Jacob Pan
Quoted 5 lines · Jacob Pan

Rename ioasid_set_data() to ioasid_attach_data() to avoid confusion with struct ioasid_set. ioasid_set is a group of IOASIDs that share a common token.

Quoted 92 lines
---
 drivers/iommu/intel/svm.c | 6 +++---
 drivers/iommu/ioasid.c    | 6 +++---
 include/linux/ioasid.h    | 4 ++--
 3 files changed, 8 insertions(+), 8 deletions(-)
drivers/iommu/intel/svm.c +3 -3
index b6972dca2ae0..37a9beabc0ca 100644
@@ -342,7 +342,7 @@ int intel_svm_bind_gpasid(struct iommu_domain *domain, struct device *dev,
 			svm->gpasid = data->gpasid;
 			svm->flags |= SVM_FLAG_GUEST_PASID;
 		}
-		ioasid_set_data(data->hpasid, svm);
+		ioasid_attach_data(data->hpasid, svm);
 		INIT_LIST_HEAD_RCU(&svm->devs);
 		mmput(svm->mm);
 	}
@@ -394,7 +394,7 @@ int intel_svm_bind_gpasid(struct iommu_domain *domain, struct device *dev,
 	list_add_rcu(&sdev->list, &svm->devs);
  out:
 	if (!IS_ERR_OR_NULL(svm) && list_empty(&svm->devs)) {
-		ioasid_set_data(data->hpasid, NULL);
+		ioasid_attach_data(data->hpasid, NULL);
 		kfree(svm);
 	}
 
@@ -437,7 +437,7 @@ int intel_svm_unbind_gpasid(struct device *dev, int pasid)
 				 * the unbind, IOMMU driver will get notified
 				 * and perform cleanup.
 				 */
-				ioasid_set_data(pasid, NULL);
+				ioasid_attach_data(pasid, NULL);
 				kfree(svm);
 			}
 		}
drivers/iommu/ioasid.c +3 -3
index 0f8dd377aada..5f63af07acd5 100644
@@ -258,14 +258,14 @@ void ioasid_unregister_allocator(struct ioasid_allocator_ops *ops)
 EXPORT_SYMBOL_GPL(ioasid_unregister_allocator);
 
 /**
- * ioasid_set_data - Set private data for an allocated ioasid
+ * ioasid_attach_data - Set private data for an allocated ioasid
  * @ioasid: the ID to set data
  * @data:   the private data
  *
  * For IOASID that is already allocated, private data can be set
  * via this API. Future lookup can be done via ioasid_find.
  */
-int ioasid_set_data(ioasid_t ioasid, void *data)
+int ioasid_attach_data(ioasid_t ioasid, void *data)
 {
 	struct ioasid_data *ioasid_data;
 	int ret = 0;
@@ -287,7 +287,7 @@ int ioasid_set_data(ioasid_t ioasid, void *data)
 
 	return ret;
 }
-EXPORT_SYMBOL_GPL(ioasid_set_data);
+EXPORT_SYMBOL_GPL(ioasid_attach_data);
 
 /**
  * ioasid_alloc - Allocate an IOASID
include/linux/ioasid.h +2 -2
index 6f000d7a0ddc..9c44947a68c8 100644
@@ -39,7 +39,7 @@ void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid,
 		  bool (*getter)(void *));
 int ioasid_register_allocator(struct ioasid_allocator_ops *allocator);
 void ioasid_unregister_allocator(struct ioasid_allocator_ops *allocator);
-int ioasid_set_data(ioasid_t ioasid, void *data);
+int ioasid_attach_data(ioasid_t ioasid, void *data);
 
 #else /* !CONFIG_IOASID */
 static inline ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min,
@@ -67,7 +67,7 @@ static inline void ioasid_unregister_allocator(struct ioasid_allocator_ops *allo
 {
 }
 
-static inline int ioasid_set_data(ioasid_t ioasid, void *data)
+static inline int ioasid_attach_data(ioasid_t ioasid, void *data)
 {
 	return -ENOTSUPP;
 }
Signature · 1 line
2.7.4
reviewer
#19 lore raw

Re: [PATCH v2 3/9] iommu/ioasid: Introduce ioasid_set APIs

In reply to Jean-Philippe Brucker
Quoted 10 lines · Jean-Philippe Brucker
Quoted 9 lines

+/**
+ * struct ioasid_data - Meta data about ioasid
+ *
+ * @id: Unique ID
+ * @users Number of active users
+ * @state Track state of the IOASID
+ * @set Meta data of the set this IOASID belongs to
+ * @private Private data associated with the IOASID
+ * @rcu For free after RCU grace period

nit: it would be nicer to follow the struct order

and use a ':' after each struct member name, as is done for @id:

Signature · 1 line
~Randy
reviewer
#21 lore raw

Re: [PATCH v2 3/9] iommu/ioasid: Introduce ioasid_set APIs

In reply to Jean-Philippe Brucker
Quoted 10 lines · Jean-Philippe Brucker
Quoted 3 lines

+/**
+ * struct ioasid_set - Meta data about ioasid_set
+ * @type: Token types and other features

nit: doesn't follow struct order

Quoted 5 lines

+ * @token: Unique to identify an IOASID set
+ * @xa: XArray to store ioasid_set private IDs, can be used for
+ * guest-host IOASID mapping, or just a private IOASID namespace.
+ * @quota: Max number of IOASIDs can be allocated within the set
+ * @nr_ioasids Number of IOASIDs currently allocated in the set

* @nr_ioasids: Number of IOASIDs currently allocated in the set

Quoted 14 lines
Quoted 14 lines

+ * @sid: ID of the set
+ * @ref: Reference count of the users
+ */
struct ioasid_set {

include/linux/ioasid.h +8 -1
-	int dummy;
+	void *token;
+	struct xarray xa;
+	int type;
+	int quota;
+	int nr_ioasids;
+	int sid;
+	refcount_t ref;
+	struct rcu_head rcu;
 };
Signature · 1 line
~Randy
#30 lore raw

Re: [PATCH v2 4/9] iommu/ioasid: Add reference couting functions

In reply to Jacob Pan
Quoted 38 lines · Jacob Pan

There can be multiple users of an IOASID, each user could have hardware contexts associated with the IOASID. In order to align lifecycles, reference counting is introduced in this patch. It is expected that when an IOASID is being freed, each user will drop a reference only after its context is cleared.

---
 drivers/iommu/ioasid.c | 113 +++++++++++++++++++++++++++++++++++++++++++++++++
 include/linux/ioasid.h |   4 ++
 2 files changed, 117 insertions(+)
drivers/iommu/ioasid.c +18 -0
index f73b3dbfc37a..5f31d63c75b1 100644
@@ -717,6 +717,119 @@ int ioasid_set_for_each_ioasid(struct ioasid_set *set,
 EXPORT_SYMBOL_GPL(ioasid_set_for_each_ioasid);
 
 /**
+ * IOASID refcounting rules
+ * - ioasid_alloc() set initial refcount to 1
+ *
+ * - ioasid_free() decrement and test refcount.
+ *     If refcount is 0, ioasid will be freed. Deleted from the system-wide
+ *     xarray as well as per set xarray. The IOASID will be returned to the
+ *     pool and available for new allocations.
+ *
+ *     If recount is non-zero, mark IOASID as IOASID_STATE_FREE_PENDING.
+ *     No new reference can be added. The IOASID is not returned to the pool
+ *     for reuse.
+ *     After free, ioasid_get() will return error but ioasid_find() and other
+ *     non refcount adding APIs will continue to work until the last reference
+ *     is dropped
+ *
+ * - ioasid_get() get a reference on an active IOASID
+ *
+ * - ioasid_put() decrement and test refcount of the IOASID.

drivers/iommu/ioasid.c int ioasid_set_for_each_ioasid(struct ioasid_set *set, · line 747

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+ *     If refcount is 0, ioasid will be freed. Deleted from the system-wide
+ *     xarray as well as per set xarray. The IOASID will be returned to the
+ *     pool and available for new allocations.
+ *     Do nothing if refcount is non-zero.
+ *
+ * - ioasid_find() does not take reference, caller must hold reference
+ *
+ * ioasid_free() can be called multiple times without error until all refs are
+ * dropped.
+ */

Since you already document this in ioasid.rst, I'm not sure the comment is necessary. Maybe the doc for _free/_put would be better in the function's documentation.

Quoted 15 lines
drivers/iommu/ioasid.c +15 -0
+
+int ioasid_get_locked(struct ioasid_set *set, ioasid_t ioasid)
+{
+	struct ioasid_data *data;
+
+	data = xa_load(&active_allocator->xa, ioasid);
+	if (!data) {
+		pr_err("Trying to get unknown IOASID %u\n", ioasid);
+		return -EINVAL;
+	}
+	if (data->state == IOASID_STATE_FREE_PENDING) {
+		pr_err("Trying to get IOASID being freed%u\n", ioasid);
+		return -EBUSY;
+	}
+

drivers/iommu/ioasid.c int ioasid_set_for_each_ioasid(struct ioasid_set *set, · line 772

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+	if (set && data->set != set) {
+		pr_err("Trying to get IOASID not in set%u\n", ioasid);
+		/* data found but does not belong to the set */
+		return -EACCES;
+	}
+	refcount_inc(&data->users);
+
+	return 0;
+}
+EXPORT_SYMBOL_GPL(ioasid_get_locked);

Is it necessary to export the *_locked variant? Who'd call them and how would they acquire the lock?

drivers/iommu/ioasid.c int ioasid_set_for_each_ioasid(struct ioasid_set *set, · line 777

Quoted 5 lines
drivers/iommu/ioasid.c +5 -0
+
+/**
+ * ioasid_get - Obtain a reference of an ioasid
+ * @set
+ * @ioasid

Can be dropped. The doc checker will throw a warning, though.

drivers/iommu/ioasid.c int ioasid_set_for_each_ioasid(struct ioasid_set *set, · line 783

Quoted 6 lines
drivers/iommu/ioasid.c +6 -0
+ *
+ * Check set ownership if @set is non-null.
+ */
+int ioasid_get(struct ioasid_set *set, ioasid_t ioasid)
+{
+	int ret = 0;

No need to initialize ret

Quoted 57 lines
drivers/iommu/ioasid.c +49 -0
+
+	spin_lock(&ioasid_allocator_lock);
+	ret = ioasid_get_locked(set, ioasid);
+	spin_unlock(&ioasid_allocator_lock);
+
+	return ret;
+}
+EXPORT_SYMBOL_GPL(ioasid_get);
+
+void ioasid_put_locked(struct ioasid_set *set, ioasid_t ioasid)
+{
+	struct ioasid_data *data;
+
+	data = xa_load(&active_allocator->xa, ioasid);
+	if (!data) {
+		pr_err("Trying to put unknown IOASID %u\n", ioasid);
+		return;
+	}
+
+	if (set && data->set != set) {
+		pr_err("Trying to drop IOASID not in the set %u\n", ioasid);
+		return;
+	}
+
+	if (!refcount_dec_and_test(&data->users)) {
+		pr_debug("%s: IOASID %d has %d remainning users\n",
+			__func__, ioasid, refcount_read(&data->users));
+		return;
+	}
+	ioasid_do_free(data);
+}
+EXPORT_SYMBOL_GPL(ioasid_put_locked);
+
+/**
+ * ioasid_put - Drop a reference of an ioasid
+ * @set
+ * @ioasid
+ *
+ * Check set ownership if @set is non-null.
+ */
+void ioasid_put(struct ioasid_set *set, ioasid_t ioasid)
+{
+	spin_lock(&ioasid_allocator_lock);
+	ioasid_put_locked(set, ioasid);
+	spin_unlock(&ioasid_allocator_lock);
+}
+EXPORT_SYMBOL_GPL(ioasid_put);
+
+/**
  * ioasid_find - Find IOASID data
  * @set: the IOASID set
  * @ioasid: the IOASID to find
include/linux/ioasid.h +0 -0
index 412d025d440e..310abe4187a3 100644

include/linux/ioasid.h int ioasid_attach_data(ioasid_t ioasid, void *data); · line 82

Quoted 7 lines
include/linux/ioasid.h +4 -0
@@ -76,6 +76,10 @@ int ioasid_attach_data(ioasid_t ioasid, void *data);
 int ioasid_register_allocator(struct ioasid_allocator_ops *allocator);
 void ioasid_unregister_allocator(struct ioasid_allocator_ops *allocator);
 void ioasid_is_in_set(struct ioasid_set *set, ioasid_t ioasid);
+int ioasid_get(struct ioasid_set *set, ioasid_t ioasid);
+int ioasid_get_locked(struct ioasid_set *set, ioasid_t ioasid);
+void ioasid_put(struct ioasid_set *set, ioasid_t ioasid);
+void ioasid_put_locked(struct ioasid_set *set, ioasid_t ioasid);

Please also add the stubs for !CONFIG_IOASID.

Thanks,
Jean

Quoted 6 lines
include/linux/ioasid.h +0 -0
 int ioasid_set_for_each_ioasid(struct ioasid_set *sdata,
 			       void (*fn)(ioasid_t id, void *data),
 			       void *data);
Signature · 1 line
2.7.4
#29 lore raw

Re: [PATCH v2 4/9] iommu/ioasid: Add reference couting functions

In reply to Lu Baolu
Quoted 10 lines · Lu Baolu

Hi Jacob,

Quoted 7 lines · Jacob Pan

There can be multiple users of an IOASID, each user could have hardware contexts associated with the IOASID. In order to align lifecycles, reference counting is introduced in this patch. It is expected that when an IOASID is being freed, each user will drop a reference only after its context is cleared.

[...]

Quoted 8 lines
Quoted 4 lines

+/**
* ioasid_find - Find IOASID data
* @set: the IOASID set
* @ioasid: the IOASID to find

Do you need to increase the refcount of the found ioasid and ask the caller to drop it after use? Otherwise, the ioasid might be freed elsewhere.

ioasid_find() takes a getter function as parameter, which ensures that the returned data is valid. It fetches the IOASID data under rcu_read_lock() and calls the getter on the private data (for example mmget_not_zero() for bare-metal SVA). Given that, I don't think returning with a reference to the IOASID is necessary. The IOASID may be freed once ioasid_find() returns but not the returned data.

Thanks,
Jean
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

#38 lore raw

Re: [PATCH v2 5/9] iommu/ioasid: Introduce ioasid_set private ID

In reply to Jacob Pan
Quoted 39 lines · Jacob Pan

When an IOASID set is used for guest SVA, each VM will acquire its ioasid_set for IOASID allocations. IOASIDs within the VM must have a host/physical IOASID backing, mapping between guest and host IOASIDs can be non-identical. IOASID set private ID (SPID) is introduced in this patch to be used as guest IOASID. However, the concept of ioasid_set specific namespace is generic, thus named SPID.

As SPID namespace is within the IOASID set, the IOASID core can provide lookup services at both directions. SPIDs may not be allocated when its IOASID is allocated, the mapping between SPID and IOASID is usually established when a guest page table is bound to a host PASID.

---
 drivers/iommu/ioasid.c | 54 ++++++++++++++++++++++++++++++++++++++++++++++++++
 include/linux/ioasid.h | 12 +++++++++++
 2 files changed, 66 insertions(+)
drivers/iommu/ioasid.c +2 -0
index 5f31d63c75b1..c0aef38a4fde 100644
@@ -21,6 +21,7 @@ enum ioasid_state {
  * struct ioasid_data - Meta data about ioasid
  *
  * @id:		Unique ID
+ * @spid:	Private ID unique within a set
  * @users	Number of active users
  * @state	Track state of the IOASID
  * @set		Meta data of the set this IOASID belongs to
@@ -29,6 +30,7 @@ enum ioasid_state {
  */
 struct ioasid_data {
 	ioasid_t id;
+	ioasid_t spid;
 	struct ioasid_set *set;
 	refcount_t users;
 	enum ioasid_state state;

drivers/iommu/ioasid.c int ioasid_attach_data(ioasid_t ioasid, void *data) · line 337

Quoted 10 lines
drivers/iommu/ioasid.c +7 -0
@@ -326,6 +328,58 @@ int ioasid_attach_data(ioasid_t ioasid, void *data)
 EXPORT_SYMBOL_GPL(ioasid_attach_data);
 
 /**
+ * ioasid_attach_spid - Attach ioasid_set private ID to an IOASID
+ *
+ * @ioasid: the ID to attach
+ * @spid:   the ioasid_set private ID of @ioasid
+ *
+ * For IOASID that is already allocated, private ID within the set can be
+ * attached via this API. Future lookup can be done via ioasid_find.

via ioasid_find_by_spid()?

Quoted 14 lines
drivers/iommu/ioasid.c +14 -0
+ */
+int ioasid_attach_spid(ioasid_t ioasid, ioasid_t spid)
+{
+	struct ioasid_data *ioasid_data;
+	int ret = 0;
+
+	spin_lock(&ioasid_allocator_lock);
+	ioasid_data = xa_load(&active_allocator->xa, ioasid);
+
+	if (!ioasid_data) {
+		pr_err("No IOASID entry %d to attach SPID %d\n",
+			ioasid, spid);
+		ret = -ENOENT;
+		goto done_unlock;

drivers/iommu/ioasid.c int ioasid_attach_data(ioasid_t ioasid, void *data) · line 361

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+	}
+	ioasid_data->spid = spid;
+
+done_unlock:
+	spin_unlock(&ioasid_allocator_lock);
+	return ret;
+}
+EXPORT_SYMBOL_GPL(ioasid_attach_spid);
+
+ioasid_t ioasid_find_by_spid(struct ioasid_set *set, ioasid_t spid)

Maybe add a bit of documentation as this is public-facing.

drivers/iommu/ioasid.c int ioasid_attach_data(ioasid_t ioasid, void *data) · line 374

Quoted 13 lines
drivers/iommu/ioasid.c +13 -0
+{
+	struct ioasid_data *entry;
+	unsigned long index;
+
+	if (!xa_load(&ioasid_sets, set->sid)) {
+		pr_warn("Invalid set\n");
+		return INVALID_IOASID;
+	}
+
+	xa_for_each(&set->xa, index, entry) {
+		if (spid == entry->spid) {
+			pr_debug("Found ioasid %lu by spid %u\n", index, spid);
+			refcount_inc(&entry->users);

Nothing prevents ioasid_free() from concurrently dropping the refcount to zero and calling ioasid_do_free(). The caller will later call ioasid_put() on a stale/reallocated index.

Quoted 26 lines
drivers/iommu/ioasid.c +8 -0
+			return index;
+		}
+	}
+	return INVALID_IOASID;
+}
+EXPORT_SYMBOL_GPL(ioasid_find_by_spid);
+
+/**
  * ioasid_alloc - Allocate an IOASID
  * @set: the IOASID set
  * @min: the minimum ID (inclusive)
include/linux/ioasid.h +2 -0
index 310abe4187a3..d4b3e83672f6 100644
@@ -73,6 +73,8 @@ bool ioasid_is_active(ioasid_t ioasid);
 
 void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid, bool (*getter)(void *));
 int ioasid_attach_data(ioasid_t ioasid, void *data);
+int ioasid_attach_spid(ioasid_t ioasid, ioasid_t spid);
+ioasid_t ioasid_find_by_spid(struct ioasid_set *set, ioasid_t spid);
 int ioasid_register_allocator(struct ioasid_allocator_ops *allocator);
 void ioasid_unregister_allocator(struct ioasid_allocator_ops *allocator);
 void ioasid_is_in_set(struct ioasid_set *set, ioasid_t ioasid);
@@ -136,5 +138,15 @@ static inline int ioasid_attach_data(ioasid_t ioasid, void *data)
 	return -ENOTSUPP;

include/linux/ioasid.h static inline int ioasid_attach_data(ioasid_t ioasid, void *data) · line 148

Quoted 10 lines
include/linux/ioasid.h +8 -0
 }
 
+staic inline int ioasid_attach_spid(ioasid_t ioasid, ioasid_t spid)
+{
+	return -ENOTSUPP;
+}
+
+static inline ioasid_t ioasid_find_by_spid(struct ioasid_set *set, ioasid_t spid)
+{
+	return -ENOTSUPP;

INVALID_IOASID

Thanks,
Jean

Quoted 7 lines
include/linux/ioasid.h +2 -0
+}
+
 #endif /* CONFIG_IOASID */
 #endif /* __LINUX_IOASID_H */
Signature · 1 line
2.7.4

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

#44 lore raw

Re: [PATCH v2 6/9] iommu/ioasid: Introduce notification APIs

In reply to Jacob Pan
Quoted 25 lines · Jacob Pan

Relations among IOASID users largely follow a publisher-subscriber pattern. E.g. to support guest SVA on Intel Scalable I/O Virtualization (SIOV) enabled platforms, VFIO, IOMMU, device drivers, KVM are all users of IOASIDs. When a state change occurs, VFIO publishes the change event that needs to be processed by other users/subscribers.

This patch introduced two types of notifications: global and per ioasid_set. The latter is intended for users who only needs to handle events related to the IOASID of a given set.
For more information, refer to the kernel documentation at Documentation/ioasid.rst.

---
 drivers/iommu/ioasid.c | 280 ++++++++++++++++++++++++++++++++++++++++++++++++-
 include/linux/ioasid.h |  70 +++++++++++++
 2 files changed, 348 insertions(+), 2 deletions(-)
drivers/iommu/ioasid.c +0 -0
index c0aef38a4fde..6ddc09a7fe74 100644

drivers/iommu/ioasid.c line 16

Quoted 8 lines
drivers/iommu/ioasid.c +3 -0
@@ -9,8 +9,35 @@
 #include <linux/spinlock.h>
 #include <linux/xarray.h>
 #include <linux/ioasid.h>
+#include <linux/sched/mm.h>
 
 static DEFINE_XARRAY_ALLOC(ioasid_sets);
+/*
+ * An IOASID could have multiple consumers where each consumeer may have

consumer

drivers/iommu/ioasid.c line 26

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+ * hardware contexts associated with IOASIDs.
+ * When a status change occurs, such as IOASID is being freed, notifier chains
+ * are used to keep the consumers in sync.
+ * This is a publisher-subscriber pattern where publisher can change the
+ * state of each IOASID, e.g. alloc/free, bind IOASID to a device and mm.
+ * On the other hand, subscribers gets notified for the state change and
+ * keep local states in sync.
+ *
+ * Currently, the notifier is global. A further optimization could be per
+ * IOASID set notifier chain.

The patch adds both

drivers/iommu/ioasid.c line 28

Quoted 2 lines
drivers/iommu/ioasid.c +2 -0
+ */
+static ATOMIC_NOTIFIER_HEAD(ioasid_chain);

"ioasid_notifier" may be clearer

Quoted 24 lines
drivers/iommu/ioasid.c +13 -0
+
+/* List to hold pending notification block registrations */
+static LIST_HEAD(ioasid_nb_pending_list);
+static DEFINE_SPINLOCK(ioasid_nb_lock);
+struct ioasid_set_nb {
+	struct list_head	list;
+	struct notifier_block	*nb;
+	void			*token;
+	struct ioasid_set	*set;
+	bool			active;
+};
+
 enum ioasid_state {
 	IOASID_STATE_INACTIVE,
 	IOASID_STATE_ACTIVE,
@@ -394,6 +421,7 @@ EXPORT_SYMBOL_GPL(ioasid_find_by_spid);
 ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max,
 		      void *private)
 {
+	struct ioasid_nb_args args;
 	struct ioasid_data *data;
 	void *adata;
 	ioasid_t id = INVALID_IOASID;

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 479

Quoted 8 lines
drivers/iommu/ioasid.c +4 -1
@@ -445,8 +473,14 @@ ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max,
 		goto exit_free;
 	}
 	set->nr_ioasids++;
-	goto done_unlock;
+	args.id = id;
+	/* Set private ID is not attached during allocation */
+	args.spid = INVALID_IOASID;
+	args.set = set;

args.pdata is uninitialized

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 480

Quoted 1 line
drivers/iommu/ioasid.c +1 -0
+	atomic_notifier_call_chain(&set->nh, IOASID_ALLOC, &args);

No global notification?

Quoted 85 lines
drivers/iommu/ioasid.c +48 -0
+	spin_unlock(&ioasid_allocator_lock);
+	return id;
 exit_free:
 	kfree(data);
 done_unlock:
@@ -479,6 +513,7 @@ static void ioasid_do_free(struct ioasid_data *data)
 
 static void ioasid_free_locked(struct ioasid_set *set, ioasid_t ioasid)
 {
+	struct ioasid_nb_args args;
 	struct ioasid_data *data;
 
 	data = xa_load(&active_allocator->xa, ioasid);
@@ -491,7 +526,16 @@ static void ioasid_free_locked(struct ioasid_set *set, ioasid_t ioasid)
 		pr_warn("Cannot free IOASID %u due to set ownership\n", ioasid);
 		return;
 	}
+
 	data->state = IOASID_STATE_FREE_PENDING;
+	/* Notify all users that this IOASID is being freed */
+	args.id = ioasid;
+	args.spid = data->spid;
+	args.pdata = data->private;
+	args.set = data->set;
+	atomic_notifier_call_chain(&ioasid_chain, IOASID_FREE, &args);
+	/* Notify the ioasid_set for per set users */
+	atomic_notifier_call_chain(&set->nh, IOASID_FREE, &args);
 
 	if (!refcount_dec_and_test(&data->users))
 		return;
@@ -514,6 +558,28 @@ void ioasid_free(struct ioasid_set *set, ioasid_t ioasid)
 }
 EXPORT_SYMBOL_GPL(ioasid_free);
 
+static void ioasid_add_pending_nb(struct ioasid_set *set)
+{
+	struct ioasid_set_nb *curr;
+
+	if (set->type != IOASID_SET_TYPE_MM)
+		return;
+
+	/*
+	 * Check if there are any pending nb requests for the given token, if so
+	 * add them to the notifier chain.
+	 */
+	spin_lock(&ioasid_nb_lock);
+	list_for_each_entry(curr, &ioasid_nb_pending_list, list) {
+		if (curr->token == set->token && !curr->active) {
+			atomic_notifier_chain_register(&set->nh, curr->nb);
+			curr->set = set;
+			curr->active = true;
+		}
+	}
+	spin_unlock(&ioasid_nb_lock);
+}
+
 /**
  * ioasid_alloc_set - Allocate a new IOASID set for a given token
  *
@@ -601,6 +667,13 @@ struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota, int type)
 	sdata->quota = quota;
 	sdata->sid = id;
 	refcount_set(&sdata->ref, 1);
+	ATOMIC_INIT_NOTIFIER_HEAD(&sdata->nh);
+
+	/*
+	 * Check if there are any pending nb requests for the given token, if so
+	 * add them to the notifier chain.
+	 */
+	ioasid_add_pending_nb(sdata);
 
 	/*
 	 * Per set XA is used to store private IDs within the set, get ready
@@ -617,6 +690,30 @@ struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota, int type)
 }
 EXPORT_SYMBOL_GPL(ioasid_alloc_set);
 
+
+/*
+ * ioasid_find_mm_set - Retrieve IOASID set with mm token
+ * Take a reference of the set if found.
+ */
+static struct ioasid_set *ioasid_find_mm_set(struct mm_struct *token)
+{

drivers/iommu/ioasid.c struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota, int type) · line 709

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+	struct ioasid_set *sdata, *set = NULL;
+	unsigned long index;
+
+	spin_lock(&ioasid_allocator_lock);
+
+	xa_for_each(&ioasid_sets, index, sdata) {
+		if (sdata->type == IOASID_SET_TYPE_MM && sdata->token == token) {
+			refcount_inc(&sdata->ref);
+			set = sdata;
+			goto exit_unlock;

Or just break

Quoted 32 lines
drivers/iommu/ioasid.c +16 -1
+		}
+	}
+exit_unlock:
+	spin_unlock(&ioasid_allocator_lock);
+	return set;
+}
+
 void ioasid_set_get_locked(struct ioasid_set *set)
 {
 	if (WARN_ON(xa_load(&ioasid_sets, set->sid) != set)) {
@@ -638,6 +735,8 @@ EXPORT_SYMBOL_GPL(ioasid_set_get);
 
 void ioasid_set_put_locked(struct ioasid_set *set)
 {
+	struct ioasid_nb_args args = { 0 };
+	struct ioasid_set_nb *curr;
 	struct ioasid_data *entry;
 	unsigned long index;
 
@@ -673,8 +772,24 @@ void ioasid_set_put_locked(struct ioasid_set *set)
 done_destroy:
 	/* Return the quota back to system pool */
 	ioasid_capacity_avail += set->quota;
-	kfree_rcu(set, rcu);
 
+	/* Restore pending status of the set NBs */
+	spin_lock(&ioasid_nb_lock);
+	list_for_each_entry(curr, &ioasid_nb_pending_list, list) {
+		if (curr->token == set->token) {
+			if (curr->active)
+				curr->active = false;
+			else

drivers/iommu/ioasid.c void ioasid_set_put_locked(struct ioasid_set *set) · line 792

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+				pr_warn("Set token exists but not active!\n");
+		}
+	}
+	spin_unlock(&ioasid_nb_lock);
+
+	args.set = set;
+	atomic_notifier_call_chain(&ioasid_chain, IOASID_SET_FREE, &args);
+
+	kfree_rcu(set, rcu);
+	pr_debug("Set freed %d\n", set->sid);

set might have been freed

drivers/iommu/ioasid.c void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid, · line 1045

Quoted 8 lines
drivers/iommu/ioasid.c +1 -0
 	/*
 	 * Token got released right away after the ioasid_set is freed.
 	 * If a new set is created immediately with the newly released token,
@@ -927,6 +1042,167 @@ void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid,
 }
 EXPORT_SYMBOL_GPL(ioasid_find);
 
+int ioasid_register_notifier(struct ioasid_set *set, struct notifier_block *nb)

Maybe add a bit of documentation on the difference with the _mm variant, as well as the @set parameter.

Will this be used by anyone at first? We could introduce only the _mm functions for now.

Quoted 45 lines
drivers/iommu/ioasid.c +45 -0
+{
+	if (set)
+		return atomic_notifier_chain_register(&set->nh, nb);
+	else
+		return atomic_notifier_chain_register(&ioasid_chain, nb);
+}
+EXPORT_SYMBOL_GPL(ioasid_register_notifier);
+
+void ioasid_unregister_notifier(struct ioasid_set *set,
+				struct notifier_block *nb)
+{
+	struct ioasid_set_nb *curr;
+
+	spin_lock(&ioasid_nb_lock);
+	/*
+	 * Pending list is registered with a token without an ioasid_set,
+	 * therefore should not be unregistered directly.
+	 */
+	list_for_each_entry(curr, &ioasid_nb_pending_list, list) {
+		if (curr->nb == nb) {
+			pr_warn("Cannot unregister NB from pending list\n");
+			spin_unlock(&ioasid_nb_lock);
+			return;
+		}
+	}
+	spin_unlock(&ioasid_nb_lock);
+
+	if (set)
+		atomic_notifier_chain_unregister(&set->nh, nb);
+	else
+		atomic_notifier_chain_unregister(&ioasid_chain, nb);
+}
+EXPORT_SYMBOL_GPL(ioasid_unregister_notifier);
+
+int ioasid_register_notifier_mm(struct mm_struct *mm, struct notifier_block *nb)
+{
+	struct ioasid_set_nb *curr;
+	struct ioasid_set *set;
+	int ret = 0;
+
+	if (!mm)
+		return -EINVAL;
+
+	spin_lock(&ioasid_nb_lock);
+

drivers/iommu/ioasid.c void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid, · line 1100

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+	/* Check for duplicates, nb is unique per set */
+	list_for_each_entry(curr, &ioasid_nb_pending_list, list) {
+		if (curr->token == mm && curr->nb == nb) {
+			ret = -EBUSY;
+			goto exit_unlock;
+		}
+	}
+
+	/* Check if the token has an existing set */
+	set = ioasid_find_mm_set(mm);

Seems to be a deadlock here, as ioasid_find_mm_set() grabs ioasid_allocator_lock while holding ioasid_nb_lock, and ioasid_set_put/get_locked() grabs ioasid_nb_lock while holding ioasid_allocator_lock.

drivers/iommu/ioasid.c void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid, · line 1101

Quoted 1 line
drivers/iommu/ioasid.c +1 -0
+	if (IS_ERR_OR_NULL(set)) {

Looks a bit off, maybe we can check !set since ioasid_find_mm_set() doesn't return errors.

drivers/iommu/ioasid.c void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid, · line 1103

Quoted 2 lines
drivers/iommu/ioasid.c +2 -0
+		/* Add to the rsvd list as inactive */
+		curr->active = false;

curr isn't valid here

drivers/iommu/ioasid.c void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid, · line 1113

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+	} else {
+		/* REVISIT: Only register empty set for now. Can add an option
+		 * in the future to playback existing PASIDs.
+		 */
+		if (set->nr_ioasids) {
+			pr_warn("IOASID set %d not empty\n", set->sid);
+			ret = -EBUSY;
+			goto exit_unlock;
+		}
+		curr = kzalloc(sizeof(*curr), GFP_ATOMIC);

As a side-note, I think there's too much atomic allocation in this file, I'd like to try and rework the locking once it stabilizes and I find some time. Do you remember why ioasid_allocator_lock needed to be a spinlock?

Quoted 34 lines
drivers/iommu/ioasid.c +34 -0
+		if (!curr) {
+			ret = -ENOMEM;
+			goto exit_unlock;
+		}
+		curr->token = mm;
+		curr->nb = nb;
+		curr->active = true;
+		curr->set = set;
+
+		/* Set already created, add to the notifier chain */
+		atomic_notifier_chain_register(&set->nh, nb);
+		/*
+		 * Do not hold a reference, if the set gets destroyed, the nb
+		 * entry will be marked inactive.
+		 */
+		ioasid_set_put(set);
+	}
+
+	list_add(&curr->list, &ioasid_nb_pending_list);
+
+exit_unlock:
+	spin_unlock(&ioasid_nb_lock);
+	return ret;
+}
+EXPORT_SYMBOL_GPL(ioasid_register_notifier_mm);
+
+void ioasid_unregister_notifier_mm(struct mm_struct *mm, struct notifier_block *nb)
+{
+	struct ioasid_set_nb *curr;
+
+	spin_lock(&ioasid_nb_lock);
+	list_for_each_entry(curr, &ioasid_nb_pending_list, list) {
+		if (curr->token == mm && curr->nb == nb) {
+			list_del(&curr->list);

drivers/iommu/ioasid.c void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid, · line 1157

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+			goto exit_free;
+		}
+	}
+	pr_warn("No ioasid set found for mm token %llx\n",  (u64)mm);
+	goto done_unlock;
+
+exit_free:
+	if (curr->active) {
+		pr_debug("mm set active, unregister %llx\n",
+			(u64)mm);

%px shows raw pointers, but I would drop this altogether or use %p.

Quoted 7 lines
drivers/iommu/ioasid.c +7 -0
+		atomic_notifier_chain_unregister(&curr->set->nh, nb);
+	}
+	kfree(curr);
+done_unlock:
+	spin_unlock(&ioasid_nb_lock);
+	return;
+}

drivers/iommu/ioasid.c void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid, · line 1174

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+EXPORT_SYMBOL_GPL(ioasid_unregister_notifier_mm);
+
+/**
+ * ioasid_notify - Send notification on a given IOASID for status change.
+ *                 Used by publishers when the status change may affect
+ *                 subscriber's internal state.
+ *
+ * @ioasid:	The IOASID to which the notification will send
+ * @cmd:	The notification event
+ * @flags:	Special instructions, e.g. notify with a set or global

Describe valid values for @cmd and @flags? I guess this function shouldn't accept IOASID_ALLOC, IOASID_FREE etc

Quoted 15 lines
drivers/iommu/ioasid.c +15 -0
+ */
+int ioasid_notify(ioasid_t ioasid, enum ioasid_notify_val cmd, unsigned int flags)
+{
+	struct ioasid_data *ioasid_data;
+	struct ioasid_nb_args args = { 0 };
+	int ret = 0;
+
+	spin_lock(&ioasid_allocator_lock);
+	ioasid_data = xa_load(&active_allocator->xa, ioasid);
+	if (!ioasid_data) {
+		pr_err("Trying to notify unknown IOASID %u\n", ioasid);
+		spin_unlock(&ioasid_allocator_lock);
+		return -EINVAL;
+	}
+

drivers/iommu/ioasid.c void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid, · line 1199

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+	args.id = ioasid;
+	args.set = ioasid_data->set;
+	args.pdata = ioasid_data->private;
+	args.spid = ioasid_data->spid;
+	if (flags & IOASID_NOTIFY_ALL) {
+		ret = atomic_notifier_call_chain(&ioasid_chain, cmd, &args);
+	} else if (flags & IOASID_NOTIFY_SET) {
+		ret = atomic_notifier_call_chain(&ioasid_data->set->nh,
+						 cmd, &args);
+	}

else ret = -EINVAL?
What about allowing both flags?

Quoted 32 lines
drivers/iommu/ioasid.c +6 -0
+	spin_unlock(&ioasid_allocator_lock);
+
+	return ret;
+}
+EXPORT_SYMBOL_GPL(ioasid_notify);
+
 MODULE_AUTHOR("Jean-Philippe Brucker <[email protected]>");
 MODULE_AUTHOR("Jacob Pan <[email protected]>");
 MODULE_DESCRIPTION("IO Address Space ID (IOASID) allocator");
include/linux/ioasid.h +2 -0
index d4b3e83672f6..572111cd3b4b 100644
@@ -23,6 +23,7 @@ enum ioasid_set_type {
  * struct ioasid_set - Meta data about ioasid_set
  * @type:	Token types and other features
  * @token:	Unique to identify an IOASID set
+ * @nh:		Notifier for IOASID events within the set
  * @xa:		XArray to store ioasid_set private IDs, can be used for
  *		guest-host IOASID mapping, or just a private IOASID namespace.
  * @quota:	Max number of IOASIDs can be allocated within the set
@@ -32,6 +33,7 @@ enum ioasid_set_type {
  */
 struct ioasid_set {
 	void *token;
+	struct atomic_notifier_head nh;
 	struct xarray xa;
 	int type;
 	int quota;
@@ -56,6 +58,49 @@ struct ioasid_allocator_ops {
 	void *pdata;
 };

include/linux/ioasid.h struct ioasid_allocator_ops { · line 69

Quoted 10 lines
include/linux/ioasid.h +9 -0
 
+/* Notification data when IOASID status changed */
+enum ioasid_notify_val {
+	IOASID_ALLOC = 1,
+	IOASID_FREE,
+	IOASID_BIND,
+	IOASID_UNBIND,
+	IOASID_SET_ALLOC,
+	IOASID_SET_FREE,
+};

May be nicer to prefix these with IOASID_NOTIFY_

Quoted 8 lines
include/linux/ioasid.h +8 -0
+
+#define IOASID_NOTIFY_ALL BIT(0)
+#define IOASID_NOTIFY_SET BIT(1)
+/**
+ * enum ioasid_notifier_prios - IOASID event notification order
+ *
+ * When status of an IOASID changes, users might need to take actions to
+ * reflect the new state. For example, when an IOASID is freed due to

include/linux/ioasid.h struct ioasid_allocator_ops { · line 87

Quoted 10 lines
include/linux/ioasid.h +10 -0
+ * exception, the hardware context in virtual CPU, DMA device, and IOMMU
+ * shall be cleared and drained. Order is required to prevent life cycle
+ * problems.
+ */
+enum ioasid_notifier_prios {
+	IOASID_PRIO_LAST,
+	IOASID_PRIO_DEVICE,
+	IOASID_PRIO_IOMMU,
+	IOASID_PRIO_CPU,
+};

Not used by this patch, can be added later

Quoted 37 lines
include/linux/ioasid.h +24 -0
+
+/**
+ * struct ioasid_nb_args - Argument provided by IOASID core when notifier
+ * is called.
+ * @id:		The IOASID being notified
+ * @spid:	The set private ID associated with the IOASID
+ * @set:	The IOASID set of @id
+ * @pdata:	The private data attached to the IOASID
+ */
+struct ioasid_nb_args {
+	ioasid_t id;
+	ioasid_t spid;
+	struct ioasid_set *set;
+	void *pdata;
+};
+
 #if IS_ENABLED(CONFIG_IOASID)
 void ioasid_install_capacity(ioasid_t total);
 ioasid_t ioasid_get_capacity(void);
@@ -75,8 +120,16 @@ void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid, bool (*getter)(void *
 int ioasid_attach_data(ioasid_t ioasid, void *data);
 int ioasid_attach_spid(ioasid_t ioasid, ioasid_t spid);
 ioasid_t ioasid_find_by_spid(struct ioasid_set *set, ioasid_t spid);
+
+int ioasid_register_notifier(struct ioasid_set *set,
+			struct notifier_block *nb);
+void ioasid_unregister_notifier(struct ioasid_set *set,
+				struct notifier_block *nb);
+
 int ioasid_register_allocator(struct ioasid_allocator_ops *allocator);
 void ioasid_unregister_allocator(struct ioasid_allocator_ops *allocator);
+
+int ioasid_notify(ioasid_t ioasid, enum ioasid_notify_val cmd, unsigned int flags);
 void ioasid_is_in_set(struct ioasid_set *set, ioasid_t ioasid);
 int ioasid_get(struct ioasid_set *set, ioasid_t ioasid);
 int ioasid_get_locked(struct ioasid_set *set, ioasid_t ioasid);

include/linux/ioasid.h void ioasid_put_locked(struct ioasid_set *set, ioasid_t ioasid); · line 142

Quoted 5 lines
include/linux/ioasid.h +2 -0
@@ -85,6 +138,9 @@ void ioasid_put_locked(struct ioasid_set *set, ioasid_t ioasid);
 int ioasid_set_for_each_ioasid(struct ioasid_set *sdata,
 			       void (*fn)(ioasid_t id, void *data),
 			       void *data);
+int ioasid_register_notifier_mm(struct mm_struct *mm, struct notifier_block *nb);
+void ioasid_unregister_notifier_mm(struct mm_struct *mm, struct notifier_block *nb);

These need stubs for !CONFIG_IOASID

include/linux/ioasid.h static inline void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid, bool (* · line 183

Quoted 9 lines
include/linux/ioasid.h +2 -0
+
 #else /* !CONFIG_IOASID */
 static inline void ioasid_install_capacity(ioasid_t total)
 {
@@ -124,6 +180,20 @@ static inline void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid, bool (*
 	return NULL;
 }
 
+static inline int ioasid_register_notifier(struct notifier_block *nb)

Missing set argument

Thanks,
Jean

Quoted 19 lines
include/linux/ioasid.h +13 -0
+{
+	return -ENOTSUPP;
+}
+
+static inline void ioasid_unregister_notifier(struct notifier_block *nb)
+{
+}
+
+static inline int ioasid_notify(ioasid_t ioasid, enum ioasid_notify_val cmd, unsigned int flags)
+{
+	return -ENOTSUPP;
+}
+
 static inline int ioasid_register_allocator(struct ioasid_allocator_ops *allocator)
 {
 	return -ENOTSUPP;
Signature · 1 line
2.7.4
#5 lore raw

Re: [PATCH v2 1/9] docs: Document IO Address Space ID (IOASID) APIs

In reply to Jean-Philippe Brucker

Hi Jacob,

Quoted 23 lines · Jean-Philippe Brucker

Documentation/ioasid.rst

Quoted 17 lines · Jacob Pan

IOASID is used to identify address spaces that can be targeted by device DMA. It is a system-wide resource that is essential to its many users. This document is an attempt to help developers from all vendors navigate the APIs. At this time, ARM SMMU and Intel’s Scalable IO Virtualization (SIOV) enabled platforms are the primary users of IOASID. Examples of how SIOV components interact with IOASID APIs are provided in that many APIs are driven by the requirements from SIOV.

---
 Documentation/ioasid.rst | 618 +++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 618 insertions(+)
 create mode 100644 Documentation/ioasid.rst
Documentation/ioasid.rst +0 -0

Thanks for writing this up. Should it go to Documentation/driver-api/, or Documentation/driver-api/iommu/? I think this also needs to Cc [email protected] and [email protected]

Quoted 15 lines

Documentation/ioasid.rst line 8

Quoted 13 lines

new file mode 100644
index 000000000000..b6a8cdc885ff

Documentation/ioasid.rst new file +8 -0
@@ -0,0 +1,618 @@
+.. ioasid:
+
+=====================================
+IO Address Space ID
+=====================================
+
+IOASID is a generic name for PCIe Process Address ID (PASID) or ARM
+SMMU sub-stream ID. An IOASID identifies an address space that DMA

"SubstreamID"

On ARM if we don't use PASIDs we have streamids (SID) which can also identify address spaces that DMA requests can target. So maybe this definition is not sufficient.

Quoted 53 lines

Documentation/ioasid.rst line 12

Quoted 5 lines
Documentation/ioasid.rst new file +4 -0
+requests can target.
+
+The primary use cases for IOASID are Shared Virtual Address (SVA) and
+IO Virtual Address (IOVA). However, the requirements for IOASID

IOVA alone isn't a use case, maybe "multiple IOVA spaces per device"?

Quoted 4 lines
Documentation/ioasid.rst new file +4 -0
+management can vary among hardware architectures.
+
+This document covers the generic features supported by IOASID
+APIs. Vendor-specific use cases are also illustrated with Intel's VT-d

Documentation/ioasid.rst line 26

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+based platforms as the first example.
+
+.. contents:: :local:
+
+Glossary
+========
+PASID - Process Address Space ID
+
+IOASID - IO Address Space ID (generic term for PCIe PASID and
+sub-stream ID in SMMU)

"SubstreamID"

Documentation/ioasid.rst line 30

Quoted 4 lines
Documentation/ioasid.rst new file +4 -0
+
+SVA/SVM - Shared Virtual Addressing/Memory
+
+ENQCMD - New Intel X86 ISA for efficient workqueue submission [1]

Maybe drop the "New", to keep the documentation perennial. It might be good to add internal links here to the specifications URLs at the bottom.

Quoted 6 lines
Documentation/ioasid.rst new file +6 -0
+
+DSA - Intel Data Streaming Accelerator [2]
+
+VDCM - Virtual device composition module [3]
+
+SIOV - Intel Scalable IO Virtualization

Documentation/ioasid.rst line 46

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+
+
+Key Concepts
+============
+
+IOASID Set
+-----------
+An IOASID set is a group of IOASIDs allocated from the system-wide
+IOASID pool. An IOASID set is created and can be identified by a
+token of u64. Refer to IOASID set APIs for more details.

Identified either by an u64 or an mm_struct, right? Maybe just drop the second sentence if it's detailed in the IOASID set section below.

Quoted 13 lines

Documentation/ioasid.rst

Quoted 9 lines
Documentation/ioasid.rst new file +9 -0
+
+IOASID set is particularly useful for guest SVA where each guest could
+have its own IOASID set for security and efficiency reasons.
+
+IOASID Set Private ID (SPID)
+----------------------------
+SPIDs are introduced as IOASIDs within its set. Each SPID maps to a
+system-wide IOASID but the namespace of SPID is within its IOASID
+set.

The intro isn't super clear. Perhaps this is simpler:
"Each IOASID set has a private namespace of SPIDs. An SPID maps to a single system-wide IOASID."

or, "within an ioasid set, each ioasid can be associated with an alias ID, named SPID."

Quoted 32 lines
Quoted 32 lines

SPIDs can be used as guest IOASIDs where each guest could do

Documentation/ioasid.rst new file +30 -0
+IOASID allocation from its own pool and map them to host physical
+IOASIDs. SPIDs are particularly useful for supporting live migration
+where decoupling guest and host physical resources are necessary.
+
+For example, two VMs can both allocate guest PASID/SPID #101 but map to
+different host PASIDs #201 and #202 respectively as shown in the
+diagram below.
+::
+
+ .------------------.    .------------------.
+ |   VM 1           |    |   VM 2           |
+ |                  |    |                  |
+ |------------------|    |------------------|
+ | GPASID/SPID 101  |    | GPASID/SPID 101  |
+ '------------------'    -------------------'     Guest
+ __________|______________________|______________________
+           |                      |               Host
+           v                      v
+ .------------------.    .------------------.
+ | Host IOASID 201  |    | Host IOASID 202  |
+ '------------------'    '------------------'
+ |   IOASID set 1   |    |   IOASID set 2   |
+ '------------------'    '------------------'
+
+Guest PASID is treated as IOASID set private ID (SPID) within an
+IOASID set, mappings between guest and host IOASIDs are stored in the
+set for inquiry.
+
+IOASID APIs
+===========
Quoted 8 lines

Documentation/ioasid.rst line 93

Quoted 8 lines
Documentation/ioasid.rst new file +8 -0
+To get the IOASID APIs, users must #include <linux/ioasid.h>. These APIs
+serve the following functionalities:
+
+  - IOASID allocation/Free
+  - Group management in the form of ioasid_set
+  - Private data storage and lookup
+  - Reference counting
+  - Event notification in case of state change

(a)

Quoted 5 lines

Documentation/ioasid.rst line 98

Quoted 5 lines
Documentation/ioasid.rst new file +5 -0
+
+IOASID Set Level APIs
+--------------------------
+For use cases such as guest SVA it is necessary to manage IOASIDs at
+a group level. For example, VMs may allocate multiple IOASIDs for

I would use the introduced ioasid_set terminology instead of "group".

Quoted 5 lines

Documentation/ioasid.rst line 100

Quoted 2 lines
Documentation/ioasid.rst new file +2 -0
+guest process address sharing (vSVA). It is imperative to enforce
+VM-IOASID ownership such that malicious guest cannot target DMA

"a malicious guest"

Quoted 8 lines
Quoted 1 line

+traffic outside its own IOASIDs, or free an active IOASID belong to

"that belongs to"

Documentation/ioasid.rst line 105

Quoted 4 lines
Documentation/ioasid.rst new file +4 -0
+another VM.
+::
+
+ struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota, u32 type)

what is this void *token? also the type may be explained here.

Quoted 32 lines
Quoted 2 lines

+
+ int ioasid_adjust_set(struct ioasid_set *set, int quota);

These could be named "ioasid_set_alloc" and "ioasid_set_adjust" to be consistent with the rest of the API.

Documentation/ioasid.rst line 117

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+
+ void ioasid_set_get(struct ioasid_set *set)
+
+ void ioasid_set_put(struct ioasid_set *set)
+
+ void ioasid_set_get_locked(struct ioasid_set *set)
+
+ void ioasid_set_put_locked(struct ioasid_set *set)
+
+ int ioasid_set_for_each_ioasid(struct ioasid_set *sdata,

Might be nicer to keep the same argument names within the API. Here "set" rather than "sdata".

Documentation/ioasid.rst line 119

Quoted 2 lines
Documentation/ioasid.rst new file +2 -0
+                                void (*fn)(ioasid_t id, void *data),
+				void *data)

(alignment)

Documentation/ioasid.rst line 122

Quoted 3 lines
Documentation/ioasid.rst new file +3 -0
+
+
+IOASID set concept is introduced to represent such IOASID groups. Each

Or just "IOASID sets represent such IOASID groups", but might be redundant.

Quoted 2 lines

Documentation/ioasid.rst line 124

Quoted 2 lines
Documentation/ioasid.rst new file +2 -0
+IOASID set is created with a token which can be one of the following
+types:

I think this explanation should happen before the above function prototypes

Quoted 10 lines
Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+
+ - IOASID_SET_TYPE_NULL (Arbitrary u64 value)
+ - IOASID_SET_TYPE_MM (Set token is a mm_struct)
+
+The explicit MM token type is useful when multiple users of an IOASID
+set under the same process need to communicate about their shared IOASIDs.
+E.g. An IOASID set created by VFIO for one guest can be associated
+with the KVM instance for the same guest since they share a common mm_struct.
+
+The IOASID set APIs serve the following purposes:
Quoted 8 lines
Quoted 6 lines

+
+ - Ownership/permission enforcement
+ - Take collective actions, e.g. free an entire set
+ - Event notifications within a set
+ - Look up a set based on token
+ - Quota enforcement

This paragraph could be earlier in the section

yes this is a kind of repetition of (a), above

Quoted 46 lines
Quoted 20 lines
Documentation/ioasid.rst new file +19 -0
+
+Individual IOASID APIs
+----------------------
+Once an ioasid_set is created, IOASIDs can be allocated from the set.
+Within the IOASID set namespace, set private ID (SPID) is supported. In
+the VM use case, SPID can be used for storing guest PASID.
+
+::
+
+ ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max,
+                       void *private);
+
+ int ioasid_get(struct ioasid_set *set, ioasid_t ioasid);
+
+ void ioasid_put(struct ioasid_set *set, ioasid_t ioasid);
+
+ int ioasid_get_locked(struct ioasid_set *set, ioasid_t ioasid);
+
+ void ioasid_put_locked(struct ioasid_set *set, ioasid_t ioasid);

Documentation/ioasid.rst line 169

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+
+ void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid,
+                   bool (*getter)(void *));
+
+ ioasid_t ioasid_find_by_spid(struct ioasid_set *set, ioasid_t spid)
+
+ int ioasid_attach_data(struct ioasid_set *set, ioasid_t ioasid,
+                        void *data);
+ int ioasid_attach_spid(struct ioasid_set *set, ioasid_t ioasid,
+                        ioasid_t ssid);

s/ssid/spid>

Quoted 14 lines
Documentation/ioasid.rst new file +14 -0
+
+
+Notifications
+-------------
+An IOASID may have multiple users, each user may have hardware context
+associated with an IOASID. When the status of an IOASID changes,
+e.g. an IOASID is being freed, users need to be notified such that the
+associated hardware context can be cleared, flushed, and drained.
+
+::
+
+ int ioasid_register_notifier(struct ioasid_set *set, struct
+                              notifier_block *nb)
+
Quoted 8 lines

Documentation/ioasid.rst line 191

Quoted 8 lines

+ void ioasid_unregister_notifier(struct ioasid_set *set,

Documentation/ioasid.rst new file +7 -0
+                                 struct notifier_block *nb)
+
+ int ioasid_register_notifier_mm(struct mm_struct *mm, struct
+                                 notifier_block *nb)
+
+ void ioasid_unregister_notifier_mm(struct mm_struct *mm, struct
+                                    notifier_block *nb)

the mm_struct prototypes may be justified

Quoted 3 lines

Documentation/ioasid.rst line 194

Quoted 3 lines
Documentation/ioasid.rst new file +3 -0
+
+ int ioasid_notify(ioasid_t ioasid, enum ioasid_notify_val cmd,
+                   unsigned int flags)

this one is not obvious either.

Quoted 17 lines
Quoted 17 lines
Documentation/ioasid.rst new file +17 -0
+
+
+Events
+~~~~~~
+Notification events are pertinent to individual IOASIDs, they can be
+one of the following:
+
+ - ALLOC
+ - FREE
+ - BIND
+ - UNBIND
+
+Ordering
+~~~~~~~~
+Ordering is supported by IOASID notification priorities as the
+following (in ascending order):
+
Quoted 8 lines

Documentation/ioasid.rst line 219

Quoted 8 lines

+::

Documentation/ioasid.rst new file +7 -0
+
+ enum ioasid_notifier_prios {
+	IOASID_PRIO_LAST,
+	IOASID_PRIO_IOMMU,
+	IOASID_PRIO_DEVICE,
+	IOASID_PRIO_CPU,
+ };

Maybe:
when registered, notifiers are assigned a priority that affect the call order. Notifiers with CPU priority get called before notifiers with device priority and so on.

Quoted 58 lines
Quoted 15 lines
Documentation/ioasid.rst new file +15 -0
+
+The typical use case is when an IOASID is freed due to an exception, DMA
+source should be quiesced before tearing down other hardware contexts
+in the system. This will reduce the churn in handling faults. DMA work
+submission is performed by the CPU which is granted higher priority than
+devices.
+
+
+Scopes
+~~~~~~
+There are two types of notifiers in IOASID core: system-wide and
+ioasid_set-wide.
+
+System-wide notifier is catering for users that need to handle all
+IOASIDs in the system. E.g. The IOMMU driver handles all IOASIDs.

Documentation/ioasid.rst line 244

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+
+Per ioasid_set notifier can be used by VM specific components such as
+KVM. After all, each KVM instance only cares about IOASIDs within its
+own set.
+
+
+Atomicity
+~~~~~~~~~
+IOASID notifiers are atomic due to spinlocks used inside the IOASID
+core. For tasks cannot be completed in the notifier handler, async work

"tasks that cannot be"

Documentation/ioasid.rst line 251

Quoted 7 lines
Documentation/ioasid.rst new file +7 -0
+can be submitted to complete the work later as long as there is no
+ordering requirement.
+
+Reference counting
+------------------
+IOASID lifecycle management is based on reference counting. Users of
+IOASID intend to align lifecycle with the IOASID need to hold

"who intend to"

Quoted 1 line

+reference of the IOASID. IOASID will not be returned to the pool for

"a reference to the IOASID. The IOASID"

Quoted 16 lines
Documentation/ioasid.rst new file +16 -0
+allocation until all references are dropped. Calling ioasid_free()
+will mark the IOASID as FREE_PENDING if the IOASID has outstanding
+reference. ioasid_get() is not allowed once an IOASID is in the
+FREE_PENDING state.
+
+Event notifications are used to inform users of IOASID status change.
+IOASID_FREE event prompts users to drop their references after
+clearing its context.
+
+For example, on VT-d platform when an IOASID is freed, teardown
+actions are performed on KVM, device driver, and IOMMU driver.
+KVM shall register notifier block with::
+
+ static struct notifier_block pasid_nb_kvm = {
+	.notifier_call = pasid_status_change_kvm,
+	.priority      = IOASID_PRIO_CPU,
Quoted 8 lines

Documentation/ioasid.rst line 276

Quoted 8 lines

+ };

Documentation/ioasid.rst new file +7 -0
+
+VDCM driver shall register notifier block with::
+
+ static struct notifier_block pasid_nb_vdcm = {
+	.notifier_call = pasid_status_change_vdcm,
+	.priority      = IOASID_PRIO_DEVICE,
+ };

not sure those code snippets are really useful. Maybe simply say who is supposed to use each prio.

Quoted 6 lines

Documentation/ioasid.rst line 282

Quoted 6 lines
Documentation/ioasid.rst new file +6 -0
+
+In both cases, notifier blocks shall be registered on the IOASID set
+such that *only* events from the matching VM is received.
+
+If KVM attempts to register notifier block before the IOASID set is
+created for the MM token, the notifier block will be placed on a

using the MM token

Quoted 2 lines

Documentation/ioasid.rst line 284

Quoted 2 lines
Documentation/ioasid.rst new file +2 -0
+pending list inside IOASID core. Once the token matching IOASID set
+is created, IOASID will register the notifier block automatically.

Is this implementation mandated? Can't you enforce the ioasid_set to be created before the notifier gets registered?

Quoted 52 lines
Quoted 27 lines
Documentation/ioasid.rst new file +27 -0
+IOASID core does not replay events for the existing IOASIDs in the
+set. For IOASID set of MM type, notification blocks can be registered
+on empty sets only. This is to avoid lost events.
+
+IOMMU driver shall register notifier block on global chain::
+
+ static struct notifier_block pasid_nb_vtd = {
+	.notifier_call = pasid_status_change_vtd,
+	.priority      = IOASID_PRIO_IOMMU,
+ };
+
+Custom allocator APIs
+---------------------
+
+::
+
+ int ioasid_register_allocator(struct ioasid_allocator_ops *allocator);
+
+ void ioasid_unregister_allocator(struct ioasid_allocator_ops *allocator);
+
+Allocator Choices
+~~~~~~~~~~~~~~~~~
+IOASIDs are allocated for both host and guest SVA/IOVA usage. However,
+allocators can be different. For example, on VT-d guest PASID
+allocation must be performed via a virtual command interface which is
+emulated by VMM.
+

Documentation/ioasid.rst line 321

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+IOASID core has the notion of "custom allocator" such that guest can
+register virtual command allocator that precedes the default one.
+
+Namespaces
+~~~~~~~~~~
+IOASIDs are limited system resources that default to 20 bits in
+size. Since each device has its own table, theoretically the namespace
+can be per device also. However, for security reasons sharing PASID
+tables among devices are not good for isolation. Therefore, IOASID
+namespace is system-wide.

I don't follow this development. Having per-device PASID table would work fine for isolation (assuming no hardware bug necessitating IOMMU groups). If I remember correctly IOASID space was chosen to be OS-wide because it simplifies the management code (single PASID per task), and it is system-wide across VMs only in the case of VT-d scalable mode.

Documentation/ioasid.rst line 325

Quoted 4 lines
Documentation/ioasid.rst new file +4 -0
+
+There are also other reasons to have this simpler system-wide
+namespace. Take VT-d as an example, VT-d supports shared workqueue
+and ENQCMD[1] where one IOASID could be used to submit work on

Maybe use the Sphinx glossary syntax rather than "[1]"
https://www.sphinx-doc.org/en/master/usage/restructuredtext/directives.html#glossary-directive

Quoted 11 lines

Documentation/ioasid.rst line 336

Quoted 11 lines
Documentation/ioasid.rst new file +11 -0
+multiple devices that are shared with other VMs. This requires IOASID
+to be system-wide. This is also the reason why guests must use an
+emulated virtual command interface to allocate IOASID from the host.
+
+
+Life cycle
+==========
+This section covers IOASID lifecycle management for both bare-metal
+and guest usages. In bare-metal SVA, MMU notifier is directly hooked
+up with IOMMU driver, therefore the process address space (MM)
+lifecycle is aligned with IOASID.

therefore the IOASID lifecyle matches the process address space (MM) lifecyle?

Quoted 2 lines

Documentation/ioasid.rst line 338

Quoted 2 lines
Documentation/ioasid.rst new file +2 -0
+
+However, guest MMU notifier is not available to host IOMMU driver,

the guest MMU notifier

Quoted 1 line
Quoted 1 line

+when guest MM terminates unexpectedly, the events have to go through

the guest MM

Quoted 39 lines
Quoted 16 lines
Documentation/ioasid.rst new file +16 -0
+VFIO and IOMMU UAPI to reach host IOMMU driver. There are also more
+parties involved in guest SVA, e.g. on Intel VT-d platform, IOASIDs
+are used by IOMMU driver, KVM, VDCM, and VFIO.
+
+Native IOASID Life Cycle (VT-d Example)
+---------------------------------------
+
+The normal flow of native SVA code with Intel Data Streaming
+Accelerator(DSA) [2] as example:
+
+1. Host user opens accelerator FD, e.g. DSA driver, or uacce;
+2. DSA driver allocate WQ, do sva_bind_device();
+3. IOMMU driver calls ioasid_alloc(), then bind PASID with device,
+   mmu_notifier_get()
+4. DMA starts by DSA driver userspace
+5. DSA userspace close FD

Documentation/ioasid.rst line 365

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+6. DSA/uacce kernel driver handles FD.close()
+7. DSA driver stops DMA
+8. DSA driver calls sva_unbind_device();
+9. IOMMU driver does unbind, clears PASID context in IOMMU, flush
+   TLBs. mmu_notifier_put() called.
+10. mmu_notifier.release() called, IOMMU SVA code calls ioasid_free()*
+11. The IOASID is returned to the pool, reclaimed.
+
+::
+

Use a footnote? https://www.sphinx-doc.org/en/master/usage/restructuredtext/basics.html#footnotes

Documentation/ioasid.rst line 367

Quoted 2 lines
Documentation/ioasid.rst new file +2 -0
+   * With ENQCMD, PASID used on VT-d is not released in mmu_notifier() but
+     mmdrop(). mmdrop comes after FD close. Should not matter.

"comes after FD close, which doesn't make a difference?"
The following might not be necessary since early process termination is described later.

Documentation/ioasid.rst line 369

Quoted 2 lines
Documentation/ioasid.rst new file +2 -0
+     If the user process dies unexpectedly, Step #10 may come before
+     Step #5, in between, all DMA faults discarded. PRQ responded with
Quoted 6 lines

PRQ hasn't been defined in this document.

Documentation/ioasid.rst line 373

Quoted 4 lines
Documentation/ioasid.rst new file +4 -0
+     code INVALID REQUEST.
+
+During the normal teardown, the following three steps would happen in
+order:

can't this be illustrated in the above 1-11 sequence, just adding NORMAL TEARDONW before #7?

Quoted 6 lines

Documentation/ioasid.rst line 379

Quoted 6 lines
Documentation/ioasid.rst new file +6 -0
+
+1. Device driver stops DMA request
+2. IOMMU driver unbinds PASID and mm, flush all TLBs, drain in-flight
+   requests.
+3. IOASID freed
+

Then you can just focus on abnormal termination

Quoted 2 lines

Documentation/ioasid.rst line 381

Quoted 2 lines
Documentation/ioasid.rst new file +2 -0
+Exception happens when process terminates *before* device driver stops
+DMA and call IOMMU driver to unbind. The flow of process exists are as

Can't this be explained with something simpler looking at the steps 1-11?

Quoted 44 lines

"exits"

Documentation/ioasid.rst line 394

Quoted 13 lines
Documentation/ioasid.rst new file +13 -0
+follows:
+
+::
+
+   do_exit() {
+	exit_mm() {
+		mm_put();
+		exit_mmap() {
+			intel_invalidate_range() //mmu notifier
+			tlb_finish_mmu()
+			mmu_notifier_release(mm) {
+				intel_iommu_release() {
+   [2]					intel_iommu_teardown_pasid();

Parentheses might be better than square brackets for step numbers

Quoted 25 lines
Documentation/ioasid.rst new file +25 -0
+                                        intel_iommu_flush_tlbs();
+				}
+				// tlb_invalidate_range cb removed
+			}
+			unmap_vmas();
+                        free_pgtables(); // IOMMU cannot walk PGT after this
+		};
+	}
+	exit_files(tsk) {
+		close_files() {
+			dsa_close();
+   [1]			dsa_stop_dma();
+                        intel_svm_unbind_pasid(); //nothing to do
+		}
+	}
+   }
+
+   mmdrop() /* some random time later, lazy mm user */ {
+   	mm_free_pgd();
+        destroy_context(mm); {
+   [3]	        ioasid_free();
+	}
+   }
+
+As shown in the list above, step #2 could happen before
Quoted 8 lines

Documentation/ioasid.rst line 427

Quoted 8 lines
Documentation/ioasid.rst new file +8 -0
+#1. Unrecoverable(UR) faults could happen between #2 and #1.
+
+Also notice that TLB invalidation occurs at mmu_notifier
+invalidate_range callback as well as the release callback. The reason
+is that release callback will delete IOMMU driver from the notifier
+chain which may skip invalidate_range() calls during the exit path.
+
+To avoid unnecessary reporting of UR fault, IOMMU driver shall disable

UR?

Quoted 52 lines
Quoted 52 lines
Documentation/ioasid.rst new file +52 -0
+fault reporting after free and before unbind.
+
+Guest IOASID Life Cycle (VT-d Example)
+--------------------------------------
+Guest IOASID life cycle starts with guest driver open(), this could be
+uacce or individual accelerator driver such as DSA. At FD open,
+sva_bind_device() is called which triggers a series of actions.
+
+The example below is an illustration of *normal* operations that
+involves *all* the SW components in VT-d. The flow can be simpler if
+no ENQCMD is supported.
+
+::
+
+     VFIO        IOMMU        KVM        VDCM        IOASID       Ref
+   ..................................................................
+   1             ioasid_register_notifier/_mm()
+   2 ioasid_alloc()                                               1
+   3 bind_gpasid()
+   4             iommu_bind()->ioasid_get()                       2
+   5             ioasid_notify(BIND)
+   6                          -> ioasid_get()                     3
+   7                          -> vmcs_update_atomic()
+   8 mdev_write(gpasid)
+   9                                    hpasid=
+   10                                   find_by_spid(gpasid)      4
+   11                                   vdev_write(hpasid)
+   12 -------- GUEST STARTS DMA --------------------------
+   13 -------- GUEST STOPS DMA --------------------------
+   14 mdev_clear(gpasid)
+   15                                   vdev_clear(hpasid)
+   16                                   ioasid_put()               3
+   17 unbind_gpasid()
+   18            iommu_ubind()
+   19            ioasid_notify(UNBIND)
+   20                          -> vmcs_update_atomic()
+   21                          -> ioasid_put()                     2
+   22 ioasid_free()                                                1
+   23            ioasid_put()                                      0
+   24                                                 Reclaimed
+   -------------- New Life Cycle Begin ----------------------------
+   1  ioasid_alloc()                                  ->           1
+
+   Note: IOASID Notification Events: FREE, BIND, UNBIND
+
+Exception cases arise when a guest crashes or a malicious guest
+attempts to cause disruption on the host system. The fault handling
+rules are:
+
+1. IOASID free must *always* succeed.
+2. An inactive period may be required before the freed IOASID is
+   reclaimed. During this period, consumers of IOASID perform cleanup.
Quoted 8 lines

Documentation/ioasid.rst line 487

Quoted 8 lines

+3. Malfunction is limited to the guest owned resources for all

Documentation/ioasid.rst new file +7 -0
+   programming errors.
+
+The primary source of exception is when the following are out of
+order:
+
+1. Start/Stop of DMA activity
+   (Guest device driver, mdev via VFIO)

please explain the meaning of what is inside (): initiator?

Quoted 22 lines

Documentation/ioasid.rst line 496

Quoted 9 lines

+2. Setup/Teardown of IOMMU PASID context, IOTLB, DevTLB flushes

Documentation/ioasid.rst new file +8 -0
+   (Host IOMMU driver bind/unbind)
+3. Setup/Teardown of VMCS PASID translation table entries (KVM) in
+   case of ENQCMD
+4. Programming/Clearing host PASID in VDCM (Host VDCM driver)
+5. IOASID alloc/free (Host IOASID)
+
+VFIO is the *only* user-kernel interface, which is ultimately
+responsible for exception handlings.

"handling"

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+
+#1 is processed the same way as the assigned device today based on
+device file descriptors and events. There is no special handling.
+
+#3 is based on bind/unbind events emitted by #2.
+
+#4 is naturally aligned with IOASID life cycle in that an illegal
+guest PASID programming would fail in obtaining reference of the
+matching host IOASID.
+
Quoted 8 lines

Documentation/ioasid.rst line 514

Quoted 8 lines
Documentation/ioasid.rst new file +8 -0
+#5 is similar to #4. The fault will be reported to the user if PASID
+used in the ENQCMD is not set up in VMCS PASID translation table.
+
+Therefore, the remaining out of order problem is between #2 and
+#5. I.e. unbind vs. free. More specifically, free before unbind.
+
+IOASID notifier and refcounting are used to ensure order. Following
+a publisher-subscriber pattern where:

with the following actors:

Quoted 3 lines
Quoted 3 lines

+

+- Publishers: VFIO & IOMMU
+- Subscribers: KVM, VDCM, IOMMU

this may be introduced before.

Quoted 4 lines

Documentation/ioasid.rst

Quoted 4 lines
Documentation/ioasid.rst new file +4 -0
+
+IOASID notifier is atomic which requires subscribers to do quick
+handling of the event in the atomic context. Workqueue can be used for
+any processing that requires thread context.

repetition of what was said before.
IOASID reference must be

Quoted 71 lines
Quoted 27 lines
Documentation/ioasid.rst new file +27 -0
+acquired before receiving the FREE event. The reference must be
+dropped at the end of the processing in order to return the IOASID to
+the pool.
+
+Let's examine the IOASID life cycle again when free happens *before*
+unbind. This could be a result of misbehaving guests or crash. Assuming
+VFIO cannot enforce unbind->free order. Notice that the setup part up
+until step #12 is identical to the normal case, the flow below starts
+with step 13.
+
+::
+
+     VFIO        IOMMU        KVM        VDCM        IOASID       Ref
+   ..................................................................
+   13 -------- GUEST STARTS DMA --------------------------
+   14 -------- *GUEST MISBEHAVES!!!* ----------------
+   15 ioasid_free()
+   16                                             ioasid_notify(FREE)
+   17                                             mark_ioasid_inactive[1]
+   18                          kvm_nb_handler(FREE)
+   19                          vmcs_update_atomic()
+   20                          ioasid_put_locked()   ->           3
+   21                                   vdcm_nb_handler(FREE)
+   22            iomm_nb_handler(FREE)
+   23 ioasid_free() returns[2]          schedule_work()           2
+   24            schedule_work()        vdev_clear_wk(hpasid)
+   25            teardown_pasid_wk()

Documentation/ioasid.rst line 558

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+   26                                   ioasid_put() ->           1
+   27            ioasid_put()                                     0
+   28                                                 Reclaimed
+   29 unbind_gpasid()
+   30            iommu_unbind()->ioasid_find() Fails[3]
+   -------------- New Life Cycle Begin ----------------------------
+
+Note:
+
+1. By marking IOASID inactive at step #17, no new references can be

Is "inactive" FREE_PENDING?

Quoted 31 lines
Documentation/ioasid.rst new file +31 -0
+   held. ioasid_get/find() will return -ENOENT;
+2. After step #23, all events can go out of order. Shall not affect
+   the outcome.
+3. IOMMU driver fails to find private data for unbinding. If unbind is
+   called after the same IOASID is allocated for the same guest again,
+   this is a programming error. The damage is limited to the guest
+   itself since unbind performs permission checking based on the
+   IOASID set associated with the guest process.
+
+KVM PASID Translation Table Updates
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Per VM PASID translation table is maintained by KVM in order to
+support ENQCMD in the guest. The table contains host-guest PASID
+translations to be consumed by CPU ucode. The synchronization of the
+PASID states depends on VFIO/IOMMU driver, where IOCTL and atomic
+notifiers are used. KVM must register IOASID notifier per VM instance
+during launch time. The following events are handled:
+
+1. BIND/UNBIND
+2. FREE
+
+Rules:
+
+1. Multiple devices can bind with the same PASID, this can be different PCI
+   devices or mdevs within the same PCI device. However, only the
+   *first* BIND and *last* UNBIND emit notifications.
+2. IOASID code is responsible for ensuring the correctness of H-G
+   PASID mapping. There is no need for KVM to validate the
+   notification data.
+3. When UNBIND happens *after* FREE, KVM will see error in
+   ioasid_get() even when the reclaim is not done. IOMMU driver will
Quoted 8 lines

Documentation/ioasid.rst line 590

Quoted 8 lines
Documentation/ioasid.rst new file +8 -0
+   also avoid sending UNBIND if the PASID is already FREE.
[… more lines in the raw message]

Lines of this file are not shown. View the raw message.

The rest of this diff is not shown. View the raw message.

programs as well

Quoted 21 lines
Quoted 1 line

+intercepts the programming attempt then program the matching host

"programs"

Thanks,
Jean

Quoted 14 lines
Documentation/ioasid.rst new file +14 -0

Lines of this file are not shown. View the raw message.

The rest of this diff is not shown. View the raw message.

Quoted 8 lines
Quoted 8 lines

+====================================================
+1. https://software.intel.com/sites/default/files/managed/c5/15/architecture-instruction-set-extensions-programming-reference.pdf

Documentation/ioasid.rst new file +4 -0

Lines of this file are not shown. View the raw message.

The rest of this diff is not shown. View the raw message.

Signature · 1 line
2.7.4

Thanks

Eric

6 lines not shown · raw message

Parts of this message are not shown here. View the raw message.

#3 lore raw

Re: [PATCH v2 1/9] docs: Document IO Address Space ID (IOASID) APIs

In reply to Lu Baolu

Hi Baolu,

Thanks for the review!

Quoted 33 lines · On Sun, 23 Aug 2020 15:05:08 +0800 Lu Baolu <[email protected]> wrote:

Hi Jacob,

Quoted 30 lines · Jacob Pan

IOASID is used to identify address spaces that can be targeted by device DMA. It is a system-wide resource that is essential to its many users. This document is an attempt to help developers from all vendors navigate the APIs. At this time, ARM SMMU and Intel’s Scalable IO Virtualization (SIOV) enabled platforms are the primary users of IOASID. Examples of how SIOV components interact with IOASID APIs are provided in that many APIs are driven by the requirements from SIOV.

---
  Documentation/ioasid.rst | 618

+++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 618 insertions(+) create mode 100644 Documentation/ioasid.rst

Documentation/ioasid.rst new file +7 -0
new file mode 100644
index 000000000000..b6a8cdc885ff
@@ -0,0 +1,618 @@
+.. ioasid:
+
+=====================================
+IO Address Space ID
+=====================================
+
+IOASID is a generic name for PCIe Process Address ID (PASID) or ARM
Quoted 9 lines
Quoted 6 lines
Documentation/ioasid.rst new file +4 -0
+SMMU sub-stream ID. An IOASID identifies an address space that DMA
+requests can target.
+
+The primary use cases for IOASID are Shared Virtual Address (SVA)

and +IO Virtual Address (IOVA). However, the requirements for IOASID

Can you please elaborate a bit more about how ioasid is used by IOVA?

Good point, I will add a paragraph for IOVA usage. Something like this:
"For IOVA, IOASID #0 is typically used for DMA request without PASID. However, some architectures such as VT-d also offers the flexibility of using any PASID for DMA request without PASID. For example, on VT-d PASID #0 is used for PCI device RID2PASID and for SIOV each auxilary domain also allocates a non-zero default PASID for DMA request w/o PASID. PASID #0, is reserved and not allocated from any ioasid_set."

Quoted 16 lines
Quoted 16 lines
Documentation/ioasid.rst new file +4 -0
+management can vary among hardware architectures.
+
+This document covers the generic features supported by IOASID
+APIs. Vendor-specific use cases are also illustrated with Intel's

VT-d +based platforms as the first example.

Documentation/ioasid.rst new file +11 -0
+
+.. contents:: :local:
+
+Glossary
+========
+PASID - Process Address Space ID
+
+IOASID - IO Address Space ID (generic term for PCIe PASID and
+sub-stream ID in SMMU)
+
+SVA/SVM - Shared Virtual Addressing/Memory
Quoted 9 lines

Documentation/ioasid.rst line 34

Quoted 6 lines
Documentation/ioasid.rst new file +6 -0
+
+ENQCMD - New Intel X86 ISA for efficient workqueue submission [1]
+
+DSA - Intel Data Streaming Accelerator [2]
+
+VDCM - Virtual device composition module [3]  

Capitalize the first letter of each word.

will do.

Quoted 43 lines
Quoted 43 lines
Documentation/ioasid.rst new file +14 -0
+
+SIOV - Intel Scalable IO Virtualization
+
+
+Key Concepts
+============
+
+IOASID Set
+-----------
+An IOASID set is a group of IOASIDs allocated from the system-wide
+IOASID pool. An IOASID set is created and can be identified by a
+token of u64. Refer to IOASID set APIs for more details.
+
+IOASID set is particularly useful for guest SVA where each guest

could +have its own IOASID set for security and efficiency reasons.

Documentation/ioasid.rst new file +8 -0
+
+IOASID Set Private ID (SPID)
+----------------------------
+SPIDs are introduced as IOASIDs within its set. Each SPID maps to a
+system-wide IOASID but the namespace of SPID is within its IOASID
+set. SPIDs can be used as guest IOASIDs where each guest could do
+IOASID allocation from its own pool and map them to host physical
+IOASIDs. SPIDs are particularly useful for supporting live

migration +where decoupling guest and host physical resources are necessary. +
+For example, two VMs can both allocate guest PASID/SPID #101 but map to +different host PASIDs #201 and #202 respectively as shown in the +diagram below.
+::

Documentation/ioasid.rst new file +14 -0
+
+ .------------------.    .------------------.
+ |   VM 1           |    |   VM 2           |
+ |                  |    |                  |
+ |------------------|    |------------------|
+ | GPASID/SPID 101  |    | GPASID/SPID 101  |
+ '------------------'    -------------------'     Guest
+ __________|______________________|______________________
+           |                      |               Host
+           v                      v
+ .------------------.    .------------------.
+ | Host IOASID 201  |    | Host IOASID 202  |
+ '------------------'    '------------------'
+ |   IOASID set 1   |    |   IOASID set 2   |
Quoted 9 lines
Quoted 5 lines

+ '------------------' '------------------'

Documentation/ioasid.rst new file +3 -0
+
+Guest PASID is treated as IOASID set private ID (SPID) within an
+IOASID set, mappings between guest and host IOASIDs are stored in

the +set for inquiry.

Is there a real IOASID set allocated in the host which represent the SPID?

SPIDs are not allocated from the host IOASID set, but the backing host IOASID of the SPID is. So the same SPID # can belong to different IOASID set.

SPIDs are allocated by VMM and given to the host, IOASID code in the host just stores it in the ioasid_set.

Quoted 124 lines
Quoted 124 lines
Documentation/ioasid.rst new file +4 -0
+
+IOASID APIs
+===========
+To get the IOASID APIs, users must #include <linux/ioasid.h>.

These APIs +serve the following functionalities:

Documentation/ioasid.rst new file +10 -0
+
+  - IOASID allocation/Free
+  - Group management in the form of ioasid_set
+  - Private data storage and lookup
+  - Reference counting
+  - Event notification in case of state change
+
+IOASID Set Level APIs
+--------------------------
+For use cases such as guest SVA it is necessary to manage IOASIDs

at +a group level. For example, VMs may allocate multiple IOASIDs for +guest process address sharing (vSVA). It is imperative to enforce +VM-IOASID ownership such that malicious guest cannot target DMA +traffic outside its own IOASIDs, or free an active IOASID belong to +another VM.
+::
+
+ struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota, u32 type) +
+ int ioasid_adjust_set(struct ioasid_set *set, int quota);

Documentation/ioasid.rst new file +11 -0
+
+ void ioasid_set_get(struct ioasid_set *set)
+
+ void ioasid_set_put(struct ioasid_set *set)
+
+ void ioasid_set_get_locked(struct ioasid_set *set)
+
+ void ioasid_set_put_locked(struct ioasid_set *set)
+
+ int ioasid_set_for_each_ioasid(struct ioasid_set *sdata,
+                                void (*fn)(ioasid_t id, void

*data),

Documentation/ioasid.rst new file +4 -0
+				void *data)
+
+
+IOASID set concept is introduced to represent such IOASID groups.

Each +IOASID set is created with a token which can be one of the following +types:

Documentation/ioasid.rst new file +5 -0
+
+ - IOASID_SET_TYPE_NULL (Arbitrary u64 value)
+ - IOASID_SET_TYPE_MM (Set token is a mm_struct)
+
+The explicit MM token type is useful when multiple users of an

IOASID +set under the same process need to communicate about their shared IOASIDs. +E.g. An IOASID set created by VFIO for one guest can be associated +with the KVM instance for the same guest since they share a common mm_struct. +

Documentation/ioasid.rst new file +11 -0
+The IOASID set APIs serve the following purposes:
+
+ - Ownership/permission enforcement
+ - Take collective actions, e.g. free an entire set
+ - Event notifications within a set
+ - Look up a set based on token
+ - Quota enforcement
+
+Individual IOASID APIs
+----------------------
+Once an ioasid_set is created, IOASIDs can be allocated from the

set. +Within the IOASID set namespace, set private ID (SPID) is supported. In +the VM use case, SPID can be used for storing guest PASID. +
+::
+
+ ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max,

Documentation/ioasid.rst new file +14 -0
+                       void *private);
+
+ int ioasid_get(struct ioasid_set *set, ioasid_t ioasid);
+
+ void ioasid_put(struct ioasid_set *set, ioasid_t ioasid);
+
+ int ioasid_get_locked(struct ioasid_set *set, ioasid_t ioasid);
+
+ void ioasid_put_locked(struct ioasid_set *set, ioasid_t ioasid);
+
+ void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid,
+                   bool (*getter)(void *));
+
+ ioasid_t ioasid_find_by_spid(struct ioasid_set *set, ioasid_t

spid) +
+ int ioasid_attach_data(struct ioasid_set *set, ioasid_t ioasid,

Documentation/ioasid.rst new file +8 -0
+                        void *data);
+ int ioasid_attach_spid(struct ioasid_set *set, ioasid_t ioasid,
+                        ioasid_t ssid);
+
+
+Notifications
+-------------
+An IOASID may have multiple users, each user may have hardware

context +associated with an IOASID. When the status of an IOASID changes, +e.g. an IOASID is being freed, users need to be notified such that the +associated hardware context can be cleared, flushed, and drained. +
+::

Documentation/ioasid.rst new file +20 -0
+
+ int ioasid_register_notifier(struct ioasid_set *set, struct
+                              notifier_block *nb)
+
+ void ioasid_unregister_notifier(struct ioasid_set *set,
+                                 struct notifier_block *nb)
+
+ int ioasid_register_notifier_mm(struct mm_struct *mm, struct
+                                 notifier_block *nb)
+
+ void ioasid_unregister_notifier_mm(struct mm_struct *mm, struct
+                                    notifier_block *nb)
+
+ int ioasid_notify(ioasid_t ioasid, enum ioasid_notify_val cmd,
+                   unsigned int flags)
+
+
+Events
+~~~~~~
+Notification events are pertinent to individual IOASIDs, they can

be +one of the following:
+
+ - ALLOC
+ - FREE
+ - BIND

Quoted 9 lines

Documentation/ioasid.rst line 210

Quoted 6 lines

+ - UNBIND

Documentation/ioasid.rst new file +5 -0
+
+Ordering
+~~~~~~~~
+Ordering is supported by IOASID notification priorities as the
+following (in ascending order):  

What does ascending order exactly mean here? LAST->IOMMU->DEVICE...?

Yes. CPU has the highest priority and will get notified first.

Quoted 122 lines
Quoted 122 lines
Documentation/ioasid.rst new file +11 -0
+
+::
+
+ enum ioasid_notifier_prios {
+	IOASID_PRIO_LAST,
+	IOASID_PRIO_IOMMU,
+	IOASID_PRIO_DEVICE,
+	IOASID_PRIO_CPU,
+ };
+
+The typical use case is when an IOASID is freed due to an

exception, DMA +source should be quiesced before tearing down other hardware contexts +in the system. This will reduce the churn in handling faults. DMA work +submission is performed by the CPU which is granted higher priority than +devices.

Documentation/ioasid.rst new file +11 -0
+
+
+Scopes
+~~~~~~
+There are two types of notifiers in IOASID core: system-wide and
+ioasid_set-wide.
+
+System-wide notifier is catering for users that need to handle all
+IOASIDs in the system. E.g. The IOMMU driver handles all IOASIDs.
+
+Per ioasid_set notifier can be used by VM specific components such

as +KVM. After all, each KVM instance only cares about IOASIDs within its +own set.

Documentation/ioasid.rst new file +6 -0
+
+
+Atomicity
+~~~~~~~~~
+IOASID notifiers are atomic due to spinlocks used inside the IOASID
+core. For tasks cannot be completed in the notifier handler, async

work +can be submitted to complete the work later as long as there is no +ordering requirement.

Documentation/ioasid.rst new file +4 -0
+
+Reference counting
+------------------
+IOASID lifecycle management is based on reference counting. Users

of +IOASID intend to align lifecycle with the IOASID need to hold +reference of the IOASID. IOASID will not be returned to the pool for +allocation until all references are dropped. Calling ioasid_free() +will mark the IOASID as FREE_PENDING if the IOASID has outstanding +reference. ioasid_get() is not allowed once an IOASID is in the +FREE_PENDING state.

Documentation/ioasid.rst new file +2 -0
+
+Event notifications are used to inform users of IOASID status

change. +IOASID_FREE event prompts users to drop their references after +clearing its context.

Documentation/ioasid.rst new file +18 -0
+
+For example, on VT-d platform when an IOASID is freed, teardown
+actions are performed on KVM, device driver, and IOMMU driver.
+KVM shall register notifier block with::
+
+ static struct notifier_block pasid_nb_kvm = {
+	.notifier_call = pasid_status_change_kvm,
+	.priority      = IOASID_PRIO_CPU,
+ };
+
+VDCM driver shall register notifier block with::
+
+ static struct notifier_block pasid_nb_vdcm = {
+	.notifier_call = pasid_status_change_vdcm,
+	.priority      = IOASID_PRIO_DEVICE,
+ };
+
+In both cases, notifier blocks shall be registered on the IOASID

set +such that *only* events from the matching VM is received.

Documentation/ioasid.rst new file +7 -0
+
+If KVM attempts to register notifier block before the IOASID set is
+created for the MM token, the notifier block will be placed on a
+pending list inside IOASID core. Once the token matching IOASID set
+is created, IOASID will register the notifier block automatically.
+IOASID core does not replay events for the existing IOASIDs in the
+set. For IOASID set of MM type, notification blocks can be

registered +on empty sets only. This is to avoid lost events.

Documentation/ioasid.rst new file +14 -0
+
+IOMMU driver shall register notifier block on global chain::
+
+ static struct notifier_block pasid_nb_vtd = {
+	.notifier_call = pasid_status_change_vtd,
+	.priority      = IOASID_PRIO_IOMMU,
+ };
+
+Custom allocator APIs
+---------------------
+
+::
+
+ int ioasid_register_allocator(struct ioasid_allocator_ops

*allocator); +
+ void ioasid_unregister_allocator(struct ioasid_allocator_ops *allocator); +

Documentation/ioasid.rst new file +3 -0
+Allocator Choices
+~~~~~~~~~~~~~~~~~
+IOASIDs are allocated for both host and guest SVA/IOVA usage.

However, +allocators can be different. For example, on VT-d guest PASID +allocation must be performed via a virtual command interface which is +emulated by VMM.

Documentation/ioasid.rst new file +2 -0
+
+IOASID core has the notion of "custom allocator" such that guest

can +register virtual command allocator that precedes the default one. +

Documentation/ioasid.rst new file +4 -0
+Namespaces
+~~~~~~~~~~
+IOASIDs are limited system resources that default to 20 bits in
+size. Since each device has its own table, theoretically the

namespace +can be per device also. However, for security reasons sharing PASID +tables among devices are not good for isolation. Therefore, IOASID +namespace is system-wide.

Documentation/ioasid.rst new file +5 -0
+
+There are also other reasons to have this simpler system-wide
+namespace. Take VT-d as an example, VT-d supports shared workqueue
+and ENQCMD[1] where one IOASID could be used to submit work on
+multiple devices that are shared with other VMs. This requires

IOASID +to be system-wide. This is also the reason why guests must use an +emulated virtual command interface to allocate IOASID from the host. +

Documentation/ioasid.rst new file +3 -0
+
+Life cycle
+==========
Quoted 9 lines
Quoted 4 lines
Documentation/ioasid.rst new file +2 -0
+This section covers IOASID lifecycle management for both bare-metal
+and guest usages. In bare-metal SVA, MMU notifier is directly

hooked +up with IOMMU driver, therefore the process address space (MM) +lifecycle is aligned with IOASID.

MMU notifier for SVA mainly serves IOMMU cache flushes, right? The IOASID life cycle for bare matal SVA is managed by the device driver through the iommu sva api's iommu_sva_(un)bind_device()?

True that lifecycle between iommu and device are aligned by sva APIs. But between mm/cpu and iommu, it depends on mmu_notifier.release(). In case process terminates unexpectedly.

Quoted 55 lines
Quoted 55 lines
Documentation/ioasid.rst new file +3 -0
+
+However, guest MMU notifier is not available to host IOMMU driver,
+when guest MM terminates unexpectedly, the events have to go

through +VFIO and IOMMU UAPI to reach host IOMMU driver. There are also more +parties involved in guest SVA, e.g. on Intel VT-d platform, IOASIDs +are used by IOMMU driver, KVM, VDCM, and VFIO.

Documentation/ioasid.rst new file +19 -0
+
+Native IOASID Life Cycle (VT-d Example)
+---------------------------------------
+
+The normal flow of native SVA code with Intel Data Streaming
+Accelerator(DSA) [2] as example:
+
+1. Host user opens accelerator FD, e.g. DSA driver, or uacce;
+2. DSA driver allocate WQ, do sva_bind_device();
+3. IOMMU driver calls ioasid_alloc(), then bind PASID with device,
+   mmu_notifier_get()
+4. DMA starts by DSA driver userspace
+5. DSA userspace close FD
+6. DSA/uacce kernel driver handles FD.close()
+7. DSA driver stops DMA
+8. DSA driver calls sva_unbind_device();
+9. IOMMU driver does unbind, clears PASID context in IOMMU, flush
+   TLBs. mmu_notifier_put() called.
+10. mmu_notifier.release() called, IOMMU SVA code calls

ioasid_free()* +11. The IOASID is returned to the pool, reclaimed.

Documentation/ioasid.rst new file +4 -0
+
+::
+
+   * With ENQCMD, PASID used on VT-d is not released in

mmu_notifier() but

Documentation/ioasid.rst new file +2 -0
+     mmdrop(). mmdrop comes after FD close. Should not matter.
+     If the user process dies unexpectedly, Step #10 may come

before
+ Step #5, in between, all DMA faults discarded. PRQ responded
with

Documentation/ioasid.rst new file +3 -0
+     code INVALID REQUEST.
+
+During the normal teardown, the following three steps would happen

in +order:
+
+1. Device driver stops DMA request
+2. IOMMU driver unbinds PASID and mm, flush all TLBs, drain in-flight

Documentation/ioasid.rst new file +4 -0
+   requests.
+3. IOASID freed
+
+Exception happens when process terminates *before* device driver

stops +DMA and call IOMMU driver to unbind. The flow of process exists are as +follows:

Documentation/ioasid.rst new file +5 -0
+
+::
+
+   do_exit() {
+	exit_mm() {
Quoted 8 lines

Documentation/ioasid.rst line 393

Quoted 6 lines
Documentation/ioasid.rst new file +6 -0
+		mm_put();
+		exit_mmap() {
+			intel_invalidate_range() //mmu notifier
+			tlb_finish_mmu()
+			mmu_notifier_release(mm) {
+				intel_iommu_release() {  

intel_mm_release()

Good catch,

Quoted 25 lines
Quoted 25 lines

+ [2]
intel_iommu_teardown_pasid();

Documentation/ioasid.rst new file +6 -0
+                                        intel_iommu_flush_tlbs();
+				}
+				// tlb_invalidate_range cb removed
+			}
+			unmap_vmas();
+                        free_pgtables(); // IOMMU cannot walk PGT

after this

Documentation/ioasid.rst new file +15 -0
+		};
+	}
+	exit_files(tsk) {
+		close_files() {
+			dsa_close();
+   [1]			dsa_stop_dma();
+                        intel_svm_unbind_pasid(); //nothing to do
+		}
+	}
+   }
+
+   mmdrop() /* some random time later, lazy mm user */ {
+   	mm_free_pgd();
+        destroy_context(mm); {
+   [3]	        ioasid_free();
Quoted 9 lines

Documentation/ioasid.rst line 420

Quoted 5 lines
Documentation/ioasid.rst new file +5 -0
+	}
+   }
+
+As shown in the list above, step #2 could happen before
+#1. Unrecoverable(UR) faults could happen between #2 and #1.  

The VT-d hardware will ignore UR faults due to the setting of FPD bit of the PASID entry. The software won't see UR faults.

Yes, here I should note that.
"Fault processing is disabled by the IOMMU driver in #2, therefore the UR fault never reaches the driver."

Quoted 211 lines
Quoted 211 lines
Documentation/ioasid.rst new file +3 -0
+
+Also notice that TLB invalidation occurs at mmu_notifier
+invalidate_range callback as well as the release callback. The

reason +is that release callback will delete IOMMU driver from the notifier +chain which may skip invalidate_range() calls during the exit path. +
+To avoid unnecessary reporting of UR fault, IOMMU driver shall disable +fault reporting after free and before unbind.

Documentation/ioasid.rst new file +4 -0
+
+Guest IOASID Life Cycle (VT-d Example)
+--------------------------------------
+Guest IOASID life cycle starts with guest driver open(), this

could be +uacce or individual accelerator driver such as DSA. At FD open, +sva_bind_device() is called which triggers a series of actions. +

Documentation/ioasid.rst new file +2 -0
+The example below is an illustration of *normal* operations that
+involves *all* the SW components in VT-d. The flow can be simpler

if +no ENQCMD is supported.

Documentation/ioasid.rst new file +4 -0
+
+::
+
+     VFIO        IOMMU        KVM        VDCM        IOASID

Ref

Documentation/ioasid.rst new file +17 -0
+   ..................................................................
+   1             ioasid_register_notifier/_mm()
+   2 ioasid_alloc()                                               1
+   3 bind_gpasid()
+   4             iommu_bind()->ioasid_get()                       2
+   5             ioasid_notify(BIND)
+   6                          -> ioasid_get()                     3
+   7                          -> vmcs_update_atomic()
+   8 mdev_write(gpasid)
+   9                                    hpasid=
+   10                                   find_by_spid(gpasid)      4
+   11                                   vdev_write(hpasid)
+   12 -------- GUEST STARTS DMA --------------------------
+   13 -------- GUEST STOPS DMA --------------------------
+   14 mdev_clear(gpasid)
+   15                                   vdev_clear(hpasid)
+   16                                   ioasid_put()

3

Documentation/ioasid.rst new file +5 -0
+   17 unbind_gpasid()
+   18            iommu_ubind()
+   19            ioasid_notify(UNBIND)
+   20                          -> vmcs_update_atomic()
+   21                          -> ioasid_put()

2
+ 22 ioasid_free()
1
+ 23 ioasid_put()
0

Documentation/ioasid.rst new file +3 -0
+   24                                                 Reclaimed
+   -------------- New Life Cycle Begin ----------------------------
+   1  ioasid_alloc()                                  ->

1 +

Documentation/ioasid.rst new file +9 -0
+   Note: IOASID Notification Events: FREE, BIND, UNBIND
+
+Exception cases arise when a guest crashes or a malicious guest
+attempts to cause disruption on the host system. The fault handling
+rules are:
+
+1. IOASID free must *always* succeed.
+2. An inactive period may be required before the freed IOASID is
+   reclaimed. During this period, consumers of IOASID perform

cleanup. +3. Malfunction is limited to the guest owned resources for all

Documentation/ioasid.rst new file +27 -0
+   programming errors.
+
+The primary source of exception is when the following are out of
+order:
+
+1. Start/Stop of DMA activity
+   (Guest device driver, mdev via VFIO)
+2. Setup/Teardown of IOMMU PASID context, IOTLB, DevTLB flushes
+   (Host IOMMU driver bind/unbind)
+3. Setup/Teardown of VMCS PASID translation table entries (KVM) in
+   case of ENQCMD
+4. Programming/Clearing host PASID in VDCM (Host VDCM driver)
+5. IOASID alloc/free (Host IOASID)
+
+VFIO is the *only* user-kernel interface, which is ultimately
+responsible for exception handlings.
+
+#1 is processed the same way as the assigned device today based on
+device file descriptors and events. There is no special handling.
+
+#3 is based on bind/unbind events emitted by #2.
+
+#4 is naturally aligned with IOASID life cycle in that an illegal
+guest PASID programming would fail in obtaining reference of the
+matching host IOASID.
+
+#5 is similar to #4. The fault will be reported to the user if

PASID +used in the ENQCMD is not set up in VMCS PASID translation table. +

Documentation/ioasid.rst new file +11 -0
+Therefore, the remaining out of order problem is between #2 and
+#5. I.e. unbind vs. free. More specifically, free before unbind.
+
+IOASID notifier and refcounting are used to ensure order. Following
+a publisher-subscriber pattern where:
+
+- Publishers: VFIO & IOMMU
+- Subscribers: KVM, VDCM, IOMMU
+
+IOASID notifier is atomic which requires subscribers to do quick
+handling of the event in the atomic context. Workqueue can be used

for +any processing that requires thread context. IOASID reference must be +acquired before receiving the FREE event. The reference must be +dropped at the end of the processing in order to return the IOASID to +the pool.

Documentation/ioasid.rst new file +2 -0
+
+Let's examine the IOASID life cycle again when free happens

*before* +unbind. This could be a result of misbehaving guests or crash. Assuming +VFIO cannot enforce unbind->free order. Notice that the setup part up +until step #12 is identical to the normal case, the flow below starts +with step 13.

Documentation/ioasid.rst new file +4 -0
+
+::
+
+     VFIO        IOMMU        KVM        VDCM        IOASID

Ref

Documentation/ioasid.rst new file +5 -0
+   ..................................................................
+   13 -------- GUEST STARTS DMA --------------------------
+   14 -------- *GUEST MISBEHAVES!!!* ----------------
+   15 ioasid_free()
+   16

ioasid_notify(FREE)
+ 17
mark_ioasid_inactive[1]

Documentation/ioasid.rst new file +22 -0
+   18                          kvm_nb_handler(FREE)
+   19                          vmcs_update_atomic()
+   20                          ioasid_put_locked()   ->           3
+   21                                   vdcm_nb_handler(FREE)
+   22            iomm_nb_handler(FREE)
+   23 ioasid_free() returns[2]          schedule_work()           2
+   24            schedule_work()        vdev_clear_wk(hpasid)
+   25            teardown_pasid_wk()
+   26                                   ioasid_put() ->           1
+   27            ioasid_put()                                     0
+   28                                                 Reclaimed
+   29 unbind_gpasid()
+   30            iommu_unbind()->ioasid_find() Fails[3]
+   -------------- New Life Cycle Begin ----------------------------
+
+Note:
+
+1. By marking IOASID inactive at step #17, no new references can be
+   held. ioasid_get/find() will return -ENOENT;
+2. After step #23, all events can go out of order. Shall not affect
+   the outcome.
+3. IOMMU driver fails to find private data for unbinding. If

unbind is
+ called after the same IOASID is allocated for the same guest
again,

Documentation/ioasid.rst new file +9 -0
+   this is a programming error. The damage is limited to the guest
+   itself since unbind performs permission checking based on the
+   IOASID set associated with the guest process.
+
+KVM PASID Translation Table Updates
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Per VM PASID translation table is maintained by KVM in order to
+support ENQCMD in the guest. The table contains host-guest PASID
+translations to be consumed by CPU ucode. The synchronization of

the +PASID states depends on VFIO/IOMMU driver, where IOCTL and atomic +notifiers are used. KVM must register IOASID notifier per VM instance +during launch time. The following events are handled:

Documentation/ioasid.rst new file +7 -0
+
+1. BIND/UNBIND
+2. FREE
+
+Rules:
+
+1. Multiple devices can bind with the same PASID, this can be

different PCI

Documentation/ioasid.rst new file +7 -0
+   devices or mdevs within the same PCI device. However, only the
+   *first* BIND and *last* UNBIND emit notifications.
+2. IOASID code is responsible for ensuring the correctness of H-G
+   PASID mapping. There is no need for KVM to validate the
+   notification data.
+3. When UNBIND happens *after* FREE, KVM will see error in
+   ioasid_get() even when the reclaim is not done. IOMMU driver

will

Documentation/ioasid.rst new file +8 -0
+   also avoid sending UNBIND if the PASID is already FREE.
+4. When KVM terminates *before* FREE & UNBIND, references will be
+   dropped for all host PASIDs.
+
+VDCM PASID Programming
+~~~~~~~~~~~~~~~~~~~~~~
+VDCM composes virtual devices and exposes them to the guests. When
+the guest allocates a PASID then program it to the virtual device,

VDCM +intercepts the programming attempt then program the matching host +PASID on to the hardware.

Documentation/ioasid.rst new file +5 -0
+Conversely, when a device is going away, VDCM must be informed such
+that PASID context on the hardware can be cleared. There could be
+multiple mdevs assigned to different guests in the same VDCM. Since
+the PASID table is shared at PCI device level, lazy clearing is not
+secure. A malicious guest can attack by using newly freed PASIDs

that +are allocated by another guest.

Documentation/ioasid.rst new file +2 -0
+
+By holding a reference of the PASID until VDCM cleans up the HW

context, +it is guaranteed that PASID life cycles do not cross within the same +device.

Documentation/ioasid.rst new file +5 -0
+
+
+Reference
+====================================================
+1.

https://software.intel.com/sites/default/files/managed/c5/15/architecture-instruction-set-extensions-programming-reference.pdf
+ +2.
https://01.org/blogs/2019/introducing-intel-data-streaming-accelerator
+ +3.

Quoted 8 lines
Quoted 1 line

https://software.intel.com/en-us/download/intel-data-streaming-accelerator-preliminary-architecture-specification

Best regards,
baolu
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

[Jacob Pan]
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

#9 lore raw

Re: [PATCH v2 1/9] docs: Document IO Address Space ID (IOASID) APIs

In reply to Jean-Philippe Brucker

Hi Jean,

Thanks for the review!

Quoted 18 lines · On Mon, 24 Aug 2020 12:32:39 +0200 Jean-Philippe Brucker…
Quoted 17 lines · Jacob Pan

IOASID is used to identify address spaces that can be targeted by device DMA. It is a system-wide resource that is essential to its many users. This document is an attempt to help developers from all vendors navigate the APIs. At this time, ARM SMMU and Intel’s Scalable IO Virtualization (SIOV) enabled platforms are the primary users of IOASID. Examples of how SIOV components interact with IOASID APIs are provided in that many APIs are driven by the requirements from SIOV.

---
 Documentation/ioasid.rst | 618

+++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 618 insertions(+) create mode 100644 Documentation/ioasid.rst

Quoted 7 lines

Documentation/ioasid.rst

Quoted 1 line
Documentation/ioasid.rst +0 -0

Thanks for writing this up. Should it go to Documentation/driver-api/, or Documentation/driver-api/iommu/? I think this also needs to Cc [email protected] and [email protected]

Good point, I think Documentation/driver-api/ is good for now as there are no other IOMMU docs.
Will CC Jon also.

Quoted 18 lines
Quoted 13 lines

new file mode 100644
index 000000000000..b6a8cdc885ff

Documentation/ioasid.rst new file +8 -0
@@ -0,0 +1,618 @@
+.. ioasid:
+
+=====================================
+IO Address Space ID
+=====================================
+
+IOASID is a generic name for PCIe Process Address ID (PASID) or ARM
+SMMU sub-stream ID. An IOASID identifies an address space that
Quoted 1 line

DMA

"SubstreamID"

Quoted 8 lines
Quoted 5 lines
Documentation/ioasid.rst new file +3 -0
+requests can target.
+
+The primary use cases for IOASID are Shared Virtual Address (SVA)

and +IO Virtual Address (IOVA). However, the requirements for IOASID

IOVA alone isn't a use case, maybe "multiple IOVA spaces per device"?

Yes, I meant guest IOVA for mdev which has "multiple IOVA spaces per device" based on aux domain. I will add this to the IOVA case description.

"The primary use cases for IOASID are Shared Virtual Address (SVA) and multiple IOVA spaces per device. However, the requirements for IOASID management can vary among hardware architectures.

For baremetal IOVA, IOASID #0 is used for DMA request without PASID. Even though some architectures such as VT-d also offers the flexibility of using any PASIDs for DMA request without PASID. PASID #0 is reserved and not allocated from any ioasid_set.

Multiple IOVA spaces per device are mapped to auxiliary domains which can be used for mediated device assignment with and without a virtual IOMMU (vIOMMU). An IOASID is allocated for each auxiliary domain as default PASID. Without vIOMMU, default IOASID is used for DMA map/unmap APIs. With vIOMMU, default IOASID is used for guest IOVA where DMA request with PASID is required for the device. The reason is that there is only one PASID #0 per device, e.g. VT-d, RID_PASID is per PCI device.
"

Quoted 8 lines
Quoted 8 lines
Documentation/ioasid.rst new file +4 -0
+management can vary among hardware architectures.
+
+This document covers the generic features supported by IOASID
+APIs. Vendor-specific use cases are also illustrated with Intel's

VT-d +based platforms as the first example.

Documentation/ioasid.rst new file +3 -0
+
+.. contents:: :local:
+
Quoted 9 lines

Documentation/ioasid.rst line 26

Quoted 6 lines
Documentation/ioasid.rst new file +6 -0
+Glossary
+========
+PASID - Process Address Space ID
+
+IOASID - IO Address Space ID (generic term for PCIe PASID and
+sub-stream ID in SMMU)  

"SubstreamID"

will fix.

Quoted 9 lines

Documentation/ioasid.rst line 30

Quoted 4 lines
Documentation/ioasid.rst new file +4 -0
+
+SVA/SVM - Shared Virtual Addressing/Memory
+
+ENQCMD - New Intel X86 ISA for efficient workqueue submission [1]  

Maybe drop the "New", to keep the documentation perennial. It might be good to add internal links here to the specifications URLs at the bottom.

Good idea

Quoted 11 lines
Quoted 11 lines
Documentation/ioasid.rst new file +11 -0
+
+DSA - Intel Data Streaming Accelerator [2]
+
+VDCM - Virtual device composition module [3]
+
+SIOV - Intel Scalable IO Virtualization
+
+
+Key Concepts
+============
+
Quoted 9 lines

Documentation/ioasid.rst line 46

Quoted 5 lines
Documentation/ioasid.rst new file +5 -0
+IOASID Set
+-----------
+An IOASID set is a group of IOASIDs allocated from the system-wide
+IOASID pool. An IOASID set is created and can be identified by a
+token of u64. Refer to IOASID set APIs for more details.  

Identified either by an u64 or an mm_struct, right? Maybe just drop the second sentence if it's detailed in the IOASID set section below.

Sounds good.

Quoted 5 lines
Quoted 5 lines
Documentation/ioasid.rst new file +2 -0
+
+IOASID set is particularly useful for guest SVA where each guest

could +have its own IOASID set for security and efficiency reasons.

Documentation/ioasid.rst new file +2 -0
+
+IOASID Set Private ID (SPID)
Quoted 9 lines

Documentation/ioasid.rst

Quoted 4 lines

+----------------------------

Documentation/ioasid.rst new file +3 -0
+SPIDs are introduced as IOASIDs within its set. Each SPID maps to a
+system-wide IOASID but the namespace of SPID is within its IOASID
+set.  

The intro isn't super clear. Perhaps this is simpler:
"Each IOASID set has a private namespace of SPIDs. An SPID maps to a single system-wide IOASID."

Sounds better, thanks for the rewrite.

Quoted 41 lines
Quoted 41 lines

SPIDs can be used as guest IOASIDs where each guest could do

Documentation/ioasid.rst new file +2 -0
+IOASID allocation from its own pool and map them to host physical
+IOASIDs. SPIDs are particularly useful for supporting live

migration +where decoupling guest and host physical resources are necessary. +
+For example, two VMs can both allocate guest PASID/SPID #101 but map to +different host PASIDs #201 and #202 respectively as shown in the +diagram below.
+::

Documentation/ioasid.rst new file +18 -0
+
+ .------------------.    .------------------.
+ |   VM 1           |    |   VM 2           |
+ |                  |    |                  |
+ |------------------|    |------------------|
+ | GPASID/SPID 101  |    | GPASID/SPID 101  |
+ '------------------'    -------------------'     Guest
+ __________|______________________|______________________
+           |                      |               Host
+           v                      v
+ .------------------.    .------------------.
+ | Host IOASID 201  |    | Host IOASID 202  |
+ '------------------'    '------------------'
+ |   IOASID set 1   |    |   IOASID set 2   |
+ '------------------'    '------------------'
+
+Guest PASID is treated as IOASID set private ID (SPID) within an
+IOASID set, mappings between guest and host IOASIDs are stored in

the +set for inquiry.

Documentation/ioasid.rst new file +4 -0
+
+IOASID APIs
+===========
+To get the IOASID APIs, users must #include <linux/ioasid.h>.

These APIs +serve the following functionalities:

Documentation/ioasid.rst new file +8 -0
+
+  - IOASID allocation/Free
+  - Group management in the form of ioasid_set
+  - Private data storage and lookup
+  - Reference counting
+  - Event notification in case of state change
+
+IOASID Set Level APIs
Quoted 9 lines
Quoted 6 lines

+--------------------------
+For use cases such as guest SVA it is necessary to manage IOASIDs at +a group level. For example, VMs may allocate multiple IOASIDs for +guest process address sharing (vSVA). It is imperative to enforce +VM-IOASID ownership such that malicious guest cannot target DMA

"a malicious guest"

got it

Quoted 5 lines
Quoted 2 lines

+traffic outside its own IOASIDs, or free an active IOASID belong to

"that belongs to"

got it

Quoted 10 lines
Quoted 6 lines
Documentation/ioasid.rst new file +4 -0
+another VM.
+::
+
+ struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota,

u32 type) +
+ int ioasid_adjust_set(struct ioasid_set *set, int quota);

These could be named "ioasid_set_alloc" and "ioasid_set_adjust" to be consistent with the rest of the API.

right

Quoted 5 lines
Quoted 5 lines
Documentation/ioasid.rst new file +5 -0
+
+ void ioasid_set_get(struct ioasid_set *set)
+
+ void ioasid_set_put(struct ioasid_set *set)
+
Quoted 9 lines

Documentation/ioasid.rst line 117

Quoted 5 lines

+ void ioasid_set_get_locked(struct ioasid_set *set)

Documentation/ioasid.rst new file +4 -0
+
+ void ioasid_set_put_locked(struct ioasid_set *set)
+
+ int ioasid_set_for_each_ioasid(struct ioasid_set *sdata,  

Might be nicer to keep the same argument names within the API. Here "set" rather than "sdata".

yes, will do.

Quoted 6 lines

Documentation/ioasid.rst line 119

Quoted 3 lines

+ void (*fn)(ioasid_t id, void
*data),

Documentation/ioasid.rst new file +1 -0
+				void *data)  

(alignment)

it looked aligned in emacs and generated html doc. might just the email? I have been having smtp issues, i am using gmail smtp this time.

Quoted 8 lines
Quoted 4 lines
Documentation/ioasid.rst new file +3 -0
+
+
+IOASID set concept is introduced to represent such IOASID groups.

Each

Or just "IOASID sets represent such IOASID groups", but might be redundant.

right, simpler is good

Quoted 12 lines
Quoted 12 lines

+IOASID set is created with a token which can be one of the following +types:

Documentation/ioasid.rst new file +5 -0
+
+ - IOASID_SET_TYPE_NULL (Arbitrary u64 value)
+ - IOASID_SET_TYPE_MM (Set token is a mm_struct)
+
+The explicit MM token type is useful when multiple users of an

IOASID +set under the same process need to communicate about their shared IOASIDs. +E.g. An IOASID set created by VFIO for one guest can be associated +with the KVM instance for the same guest since they share a common mm_struct. +
+The IOASID set APIs serve the following purposes:

Quoted 9 lines
Quoted 6 lines

+
+ - Ownership/permission enforcement
+ - Take collective actions, e.g. free an entire set
+ - Event notifications within a set
+ - Look up a set based on token
+ - Quota enforcement

This paragraph could be earlier in the section

Make sense. I will move it before listing the APIs.

Quoted 24 lines
Quoted 24 lines
Documentation/ioasid.rst new file +4 -0
+
+Individual IOASID APIs
+----------------------
+Once an ioasid_set is created, IOASIDs can be allocated from the

set. +Within the IOASID set namespace, set private ID (SPID) is supported. In +the VM use case, SPID can be used for storing guest PASID. +
+::
+
+ ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max,

Documentation/ioasid.rst new file +13 -0
+                       void *private);
+
+ int ioasid_get(struct ioasid_set *set, ioasid_t ioasid);
+
+ void ioasid_put(struct ioasid_set *set, ioasid_t ioasid);
+
+ int ioasid_get_locked(struct ioasid_set *set, ioasid_t ioasid);
+
+ void ioasid_put_locked(struct ioasid_set *set, ioasid_t ioasid);
+
+ void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid,
+                   bool (*getter)(void *));
+
Quoted 9 lines

Documentation/ioasid.rst line 169

Quoted 6 lines

+ ioasid_t ioasid_find_by_spid(struct ioasid_set *set, ioasid_t spid) +
+ int ioasid_attach_data(struct ioasid_set *set, ioasid_t ioasid,

Documentation/ioasid.rst new file +3 -0
+                        void *data);
+ int ioasid_attach_spid(struct ioasid_set *set, ioasid_t ioasid,
+                        ioasid_t ssid);  

s/ssid/spid

got it

Quoted 70 lines
Quoted 70 lines
Documentation/ioasid.rst new file +5 -0
+
+
+Notifications
+-------------
+An IOASID may have multiple users, each user may have hardware

context +associated with an IOASID. When the status of an IOASID changes, +e.g. an IOASID is being freed, users need to be notified such that the +associated hardware context can be cleared, flushed, and drained. +
+::

Documentation/ioasid.rst new file +20 -0
+
+ int ioasid_register_notifier(struct ioasid_set *set, struct
+                              notifier_block *nb)
+
+ void ioasid_unregister_notifier(struct ioasid_set *set,
+                                 struct notifier_block *nb)
+
+ int ioasid_register_notifier_mm(struct mm_struct *mm, struct
+                                 notifier_block *nb)
+
+ void ioasid_unregister_notifier_mm(struct mm_struct *mm, struct
+                                    notifier_block *nb)
+
+ int ioasid_notify(ioasid_t ioasid, enum ioasid_notify_val cmd,
+                   unsigned int flags)
+
+
+Events
+~~~~~~
+Notification events are pertinent to individual IOASIDs, they can

be +one of the following:

Documentation/ioasid.rst new file +21 -0
+
+ - ALLOC
+ - FREE
+ - BIND
+ - UNBIND
+
+Ordering
+~~~~~~~~
+Ordering is supported by IOASID notification priorities as the
+following (in ascending order):
+
+::
+
+ enum ioasid_notifier_prios {
+	IOASID_PRIO_LAST,
+	IOASID_PRIO_IOMMU,
+	IOASID_PRIO_DEVICE,
+	IOASID_PRIO_CPU,
+ };
+
+The typical use case is when an IOASID is freed due to an

exception, DMA +source should be quiesced before tearing down other hardware contexts +in the system. This will reduce the churn in handling faults. DMA work +submission is performed by the CPU which is granted higher priority than +devices.

Documentation/ioasid.rst new file +11 -0
+
+
+Scopes
+~~~~~~
+There are two types of notifiers in IOASID core: system-wide and
+ioasid_set-wide.
+
+System-wide notifier is catering for users that need to handle all
+IOASIDs in the system. E.g. The IOMMU driver handles all IOASIDs.
+
+Per ioasid_set notifier can be used by VM specific components such

as +KVM. After all, each KVM instance only cares about IOASIDs within its +own set.
+

Quoted 9 lines
Quoted 6 lines
Documentation/ioasid.rst new file +5 -0
+
+Atomicity
+~~~~~~~~~
+IOASID notifiers are atomic due to spinlocks used inside the IOASID
+core. For tasks cannot be completed in the notifier handler, async

work

"tasks that cannot be"

got it

Quoted 10 lines
Quoted 7 lines
Documentation/ioasid.rst new file +6 -0
+can be submitted to complete the work later as long as there is no
+ordering requirement.
+
+Reference counting
+------------------
+IOASID lifecycle management is based on reference counting. Users

of +IOASID intend to align lifecycle with the IOASID need to hold

"who intend to"

got it

Quoted 5 lines
Quoted 2 lines

+reference of the IOASID. IOASID will not be returned to the pool for

"a reference to the IOASID. The IOASID"

got it

Quoted 68 lines
Quoted 68 lines
Documentation/ioasid.rst new file +6 -0
+allocation until all references are dropped. Calling ioasid_free()
+will mark the IOASID as FREE_PENDING if the IOASID has outstanding
+reference. ioasid_get() is not allowed once an IOASID is in the
+FREE_PENDING state.
+
+Event notifications are used to inform users of IOASID status

change. +IOASID_FREE event prompts users to drop their references after +clearing its context.

Documentation/ioasid.rst new file +18 -0
+
+For example, on VT-d platform when an IOASID is freed, teardown
+actions are performed on KVM, device driver, and IOMMU driver.
+KVM shall register notifier block with::
+
+ static struct notifier_block pasid_nb_kvm = {
+	.notifier_call = pasid_status_change_kvm,
+	.priority      = IOASID_PRIO_CPU,
+ };
+
+VDCM driver shall register notifier block with::
+
+ static struct notifier_block pasid_nb_vdcm = {
+	.notifier_call = pasid_status_change_vdcm,
+	.priority      = IOASID_PRIO_DEVICE,
+ };
+
+In both cases, notifier blocks shall be registered on the IOASID

set +such that *only* events from the matching VM is received.

Documentation/ioasid.rst new file +7 -0
+
+If KVM attempts to register notifier block before the IOASID set is
+created for the MM token, the notifier block will be placed on a
+pending list inside IOASID core. Once the token matching IOASID set
+is created, IOASID will register the notifier block automatically.
+IOASID core does not replay events for the existing IOASIDs in the
+set. For IOASID set of MM type, notification blocks can be

registered +on empty sets only. This is to avoid lost events.

Documentation/ioasid.rst new file +14 -0
+
+IOMMU driver shall register notifier block on global chain::
+
+ static struct notifier_block pasid_nb_vtd = {
+	.notifier_call = pasid_status_change_vtd,
+	.priority      = IOASID_PRIO_IOMMU,
+ };
+
+Custom allocator APIs
+---------------------
+
+::
+
+ int ioasid_register_allocator(struct ioasid_allocator_ops

*allocator); +
+ void ioasid_unregister_allocator(struct ioasid_allocator_ops *allocator); +

Documentation/ioasid.rst new file +3 -0
+Allocator Choices
+~~~~~~~~~~~~~~~~~
+IOASIDs are allocated for both host and guest SVA/IOVA usage.

However, +allocators can be different. For example, on VT-d guest PASID +allocation must be performed via a virtual command interface which is +emulated by VMM.

Documentation/ioasid.rst new file +2 -0
+
+IOASID core has the notion of "custom allocator" such that guest

can +register virtual command allocator that precedes the default one. +

Documentation/ioasid.rst new file +4 -0
+Namespaces
+~~~~~~~~~~
+IOASIDs are limited system resources that default to 20 bits in
+size. Since each device has its own table, theoretically the

namespace +can be per device also. However, for security reasons sharing PASID +tables among devices are not good for isolation.

Quoted 9 lines
Quoted 1 line

Therefore, IOASID +namespace is system-wide.

I don't follow this development. Having per-device PASID table would work fine for isolation (assuming no hardware bug necessitating IOMMU groups). If I remember correctly IOASID space was chosen to be OS-wide because it simplifies the management code (single PASID per task), and it is system-wide across VMs only in the case of VT-d scalable mode.

You are right, system-wide namespace is chosen for simplicity and enqcmd. I will fix that.

Quoted 8 lines

Documentation/ioasid.rst line 325

Quoted 4 lines
Documentation/ioasid.rst new file +4 -0
+
+There are also other reasons to have this simpler system-wide
+namespace. Take VT-d as an example, VT-d supports shared workqueue
+and ENQCMD[1] where one IOASID could be used to submit work on  

Maybe use the Sphinx glossary syntax rather than "[1]"
https://www.sphinx-doc.org/en/master/usage/restructuredtext/directives.html#glossary-directive

I will look into that, thanks!

Quoted 35 lines
Quoted 35 lines

+multiple devices that are shared with other VMs. This requires IOASID +to be system-wide. This is also the reason why guests must use an +emulated virtual command interface to allocate IOASID from the host. +

Documentation/ioasid.rst new file +5 -0
+
+Life cycle
+==========
+This section covers IOASID lifecycle management for both bare-metal
+and guest usages. In bare-metal SVA, MMU notifier is directly

hooked +up with IOMMU driver, therefore the process address space (MM) +lifecycle is aligned with IOASID.

Documentation/ioasid.rst new file +3 -0
+
+However, guest MMU notifier is not available to host IOMMU driver,
+when guest MM terminates unexpectedly, the events have to go

through +VFIO and IOMMU UAPI to reach host IOMMU driver. There are also more +parties involved in guest SVA, e.g. on Intel VT-d platform, IOASIDs +are used by IOMMU driver, KVM, VDCM, and VFIO.

Documentation/ioasid.rst new file +18 -0
+
+Native IOASID Life Cycle (VT-d Example)
+---------------------------------------
+
+The normal flow of native SVA code with Intel Data Streaming
+Accelerator(DSA) [2] as example:
+
+1. Host user opens accelerator FD, e.g. DSA driver, or uacce;
+2. DSA driver allocate WQ, do sva_bind_device();
+3. IOMMU driver calls ioasid_alloc(), then bind PASID with device,
+   mmu_notifier_get()
+4. DMA starts by DSA driver userspace
+5. DSA userspace close FD
+6. DSA/uacce kernel driver handles FD.close()
+7. DSA driver stops DMA
+8. DSA driver calls sva_unbind_device();
+9. IOMMU driver does unbind, clears PASID context in IOMMU, flush
+   TLBs. mmu_notifier_put() called.
Quoted 9 lines

Documentation/ioasid.rst line 64

Quoted 5 lines

+10. mmu_notifier.release() called, IOMMU SVA code calls ioasid_free()* +11. The IOASID is returned to the pool, reclaimed.

Documentation/ioasid.rst new file +3 -0
+
+::
+  

Use a footnote?
https://www.sphinx-doc.org/en/master/usage/restructuredtext/basics.html#footnotes

ditto

Quoted 8 lines
Quoted 3 lines

+ * With ENQCMD, PASID used on VT-d is not released in
mmu_notifier() but
+ mmdrop(). mmdrop comes after FD close. Should not matter.

"comes after FD close, which doesn't make a difference?" The following might not be necessary since early process termination is described later.

yes, it is redundant. I will remove it.

Quoted 7 lines
Quoted 4 lines

+ If the user process dies unexpectedly, Step #10 may come
before
+ Step #5, in between, all DMA faults discarded. PRQ responded
with

PRQ hasn't been defined in this document.

will remove

Quoted 8 lines
Quoted 8 lines
Documentation/ioasid.rst new file +3 -0
+     code INVALID REQUEST.
+
+During the normal teardown, the following three steps would happen

in +order:
+
+1. Device driver stops DMA request
+2. IOMMU driver unbinds PASID and mm, flush all TLBs, drain in-flight

Quoted 9 lines
Quoted 6 lines
Documentation/ioasid.rst new file +4 -0
+   requests.
+3. IOASID freed
+
+Exception happens when process terminates *before* device driver

stops +DMA and call IOMMU driver to unbind. The flow of process exists are as

"exits"

will fix

Quoted 8 lines
Quoted 8 lines
Documentation/ioasid.rst new file +8 -0
+follows:
+
+::
+
+   do_exit() {
+	exit_mm() {
+		mm_put();
+		exit_mmap() {
Quoted 9 lines
Quoted 6 lines
Documentation/ioasid.rst new file +5 -0
+			intel_invalidate_range() //mmu notifier
+			tlb_finish_mmu()
+			mmu_notifier_release(mm) {
+				intel_iommu_release() {
+   [2]

intel_iommu_teardown_pasid();

Parentheses might be better than square brackets for step numbers

yes, it gets highlight as well. thanks!

Quoted 103 lines
Quoted 103 lines
Documentation/ioasid.rst new file +6 -0
+                                        intel_iommu_flush_tlbs();
+				}
+				// tlb_invalidate_range cb removed
+			}
+			unmap_vmas();
+                        free_pgtables(); // IOMMU cannot walk PGT

after this

Documentation/ioasid.rst new file +23 -0
+		};
+	}
+	exit_files(tsk) {
+		close_files() {
+			dsa_close();
+   [1]			dsa_stop_dma();
+                        intel_svm_unbind_pasid(); //nothing to do
+		}
+	}
+   }
+
+   mmdrop() /* some random time later, lazy mm user */ {
+   	mm_free_pgd();
+        destroy_context(mm); {
+   [3]	        ioasid_free();
+	}
+   }
+
+As shown in the list above, step #2 could happen before
+#1. Unrecoverable(UR) faults could happen between #2 and #1.
+
+Also notice that TLB invalidation occurs at mmu_notifier
+invalidate_range callback as well as the release callback. The

reason +is that release callback will delete IOMMU driver from the notifier +chain which may skip invalidate_range() calls during the exit path. +
+To avoid unnecessary reporting of UR fault, IOMMU driver shall disable +fault reporting after free and before unbind.

Documentation/ioasid.rst new file +4 -0
+
+Guest IOASID Life Cycle (VT-d Example)
+--------------------------------------
+Guest IOASID life cycle starts with guest driver open(), this

could be +uacce or individual accelerator driver such as DSA. At FD open, +sva_bind_device() is called which triggers a series of actions. +

Documentation/ioasid.rst new file +2 -0
+The example below is an illustration of *normal* operations that
+involves *all* the SW components in VT-d. The flow can be simpler

if +no ENQCMD is supported.

Documentation/ioasid.rst new file +4 -0
+
+::
+
+     VFIO        IOMMU        KVM        VDCM        IOASID

Ref

Documentation/ioasid.rst new file +17 -0
+   ..................................................................
+   1             ioasid_register_notifier/_mm()
+   2 ioasid_alloc()                                               1
+   3 bind_gpasid()
+   4             iommu_bind()->ioasid_get()                       2
+   5             ioasid_notify(BIND)
+   6                          -> ioasid_get()                     3
+   7                          -> vmcs_update_atomic()
+   8 mdev_write(gpasid)
+   9                                    hpasid=
+   10                                   find_by_spid(gpasid)      4
+   11                                   vdev_write(hpasid)
+   12 -------- GUEST STARTS DMA --------------------------
+   13 -------- GUEST STOPS DMA --------------------------
+   14 mdev_clear(gpasid)
+   15                                   vdev_clear(hpasid)
+   16                                   ioasid_put()

3

Documentation/ioasid.rst new file +5 -0
+   17 unbind_gpasid()
+   18            iommu_ubind()
+   19            ioasid_notify(UNBIND)
+   20                          -> vmcs_update_atomic()
+   21                          -> ioasid_put()

2
+ 22 ioasid_free()
1
+ 23 ioasid_put()
0

Documentation/ioasid.rst new file +3 -0
+   24                                                 Reclaimed
+   -------------- New Life Cycle Begin ----------------------------
+   1  ioasid_alloc()                                  ->

1 +

Documentation/ioasid.rst new file +9 -0
+   Note: IOASID Notification Events: FREE, BIND, UNBIND
+
+Exception cases arise when a guest crashes or a malicious guest
+attempts to cause disruption on the host system. The fault handling
+rules are:
+
+1. IOASID free must *always* succeed.
+2. An inactive period may be required before the freed IOASID is
+   reclaimed. During this period, consumers of IOASID perform

cleanup. +3. Malfunction is limited to the guest owned resources for all

Documentation/ioasid.rst new file +10 -0
+   programming errors.
+
+The primary source of exception is when the following are out of
+order:
+
+1. Start/Stop of DMA activity
+   (Guest device driver, mdev via VFIO)
+2. Setup/Teardown of IOMMU PASID context, IOTLB, DevTLB flushes
+   (Host IOMMU driver bind/unbind)
+3. Setup/Teardown of VMCS PASID translation table entries (KVM) in
Quoted 9 lines

Documentation/ioasid.rst line 496

Quoted 6 lines
Documentation/ioasid.rst new file +6 -0
+   case of ENQCMD
+4. Programming/Clearing host PASID in VDCM (Host VDCM driver)
+5. IOASID alloc/free (Host IOASID)
+
+VFIO is the *only* user-kernel interface, which is ultimately
+responsible for exception handlings.  

"handling"

got it

Quoted 60 lines
Quoted 60 lines
Documentation/ioasid.rst new file +11 -0
+
+#1 is processed the same way as the assigned device today based on
+device file descriptors and events. There is no special handling.
+
+#3 is based on bind/unbind events emitted by #2.
+
+#4 is naturally aligned with IOASID life cycle in that an illegal
+guest PASID programming would fail in obtaining reference of the
+matching host IOASID.
+
+#5 is similar to #4. The fault will be reported to the user if

PASID +used in the ENQCMD is not set up in VMCS PASID translation table. +

Documentation/ioasid.rst new file +11 -0
+Therefore, the remaining out of order problem is between #2 and
+#5. I.e. unbind vs. free. More specifically, free before unbind.
+
+IOASID notifier and refcounting are used to ensure order. Following
+a publisher-subscriber pattern where:
+
+- Publishers: VFIO & IOMMU
+- Subscribers: KVM, VDCM, IOMMU
+
+IOASID notifier is atomic which requires subscribers to do quick
+handling of the event in the atomic context. Workqueue can be used

for +any processing that requires thread context. IOASID reference must be +acquired before receiving the FREE event. The reference must be +dropped at the end of the processing in order to return the IOASID to +the pool.

Documentation/ioasid.rst new file +2 -0
+
+Let's examine the IOASID life cycle again when free happens

*before* +unbind. This could be a result of misbehaving guests or crash. Assuming +VFIO cannot enforce unbind->free order. Notice that the setup part up +until step #12 is identical to the normal case, the flow below starts +with step 13.

Documentation/ioasid.rst new file +4 -0
+
+::
+
+     VFIO        IOMMU        KVM        VDCM        IOASID

Ref

Documentation/ioasid.rst new file +5 -0
+   ..................................................................
+   13 -------- GUEST STARTS DMA --------------------------
+   14 -------- *GUEST MISBEHAVES!!!* ----------------
+   15 ioasid_free()
+   16

ioasid_notify(FREE)
+ 17
mark_ioasid_inactive[1]

Documentation/ioasid.rst new file +13 -0
+   18                          kvm_nb_handler(FREE)
+   19                          vmcs_update_atomic()
+   20                          ioasid_put_locked()   ->           3
+   21                                   vdcm_nb_handler(FREE)
+   22            iomm_nb_handler(FREE)
+   23 ioasid_free() returns[2]          schedule_work()           2
+   24            schedule_work()        vdev_clear_wk(hpasid)
+   25            teardown_pasid_wk()
+   26                                   ioasid_put() ->           1
+   27            ioasid_put()                                     0
+   28                                                 Reclaimed
+   29 unbind_gpasid()
+   30            iommu_unbind()->ioasid_find() Fails[3]
Quoted 9 lines
Quoted 6 lines
Documentation/ioasid.rst new file +5 -0
+   -------------- New Life Cycle Begin ----------------------------
+
+Note:
+
+1. By marking IOASID inactive at step #17, no new references can

be

Is "inactive" FREE_PENDING?

yes, will fix.

Quoted 70 lines
Quoted 43 lines
Documentation/ioasid.rst new file +4 -0
+   held. ioasid_get/find() will return -ENOENT;
+2. After step #23, all events can go out of order. Shall not affect
+   the outcome.
+3. IOMMU driver fails to find private data for unbinding. If

unbind is
+ called after the same IOASID is allocated for the same guest
again,

Documentation/ioasid.rst new file +9 -0
+   this is a programming error. The damage is limited to the guest
+   itself since unbind performs permission checking based on the
+   IOASID set associated with the guest process.
+
+KVM PASID Translation Table Updates
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Per VM PASID translation table is maintained by KVM in order to
+support ENQCMD in the guest. The table contains host-guest PASID
+translations to be consumed by CPU ucode. The synchronization of

the +PASID states depends on VFIO/IOMMU driver, where IOCTL and atomic +notifiers are used. KVM must register IOASID notifier per VM instance +during launch time. The following events are handled:

Documentation/ioasid.rst new file +7 -0
+
+1. BIND/UNBIND
+2. FREE
+
+Rules:
+
+1. Multiple devices can bind with the same PASID, this can be

different PCI

Documentation/ioasid.rst new file +7 -0
+   devices or mdevs within the same PCI device. However, only the
+   *first* BIND and *last* UNBIND emit notifications.
+2. IOASID code is responsible for ensuring the correctness of H-G
+   PASID mapping. There is no need for KVM to validate the
+   notification data.
+3. When UNBIND happens *after* FREE, KVM will see error in
+   ioasid_get() even when the reclaim is not done. IOMMU driver

will

Documentation/ioasid.rst new file +8 -0
+   also avoid sending UNBIND if the PASID is already FREE.
+4. When KVM terminates *before* FREE & UNBIND, references will be
+   dropped for all host PASIDs.
+
+VDCM PASID Programming
+~~~~~~~~~~~~~~~~~~~~~~
+VDCM composes virtual devices and exposes them to the guests. When
+the guest allocates a PASID then program it to the virtual device,
Quoted 2 lines

VDCM +intercepts the programming attempt then program the matching host

"programs"

Thanks,
Jean

Quoted 19 lines
Documentation/ioasid.rst new file +6 -0
+PASID on to the hardware.
+Conversely, when a device is going away, VDCM must be informed such
+that PASID context on the hardware can be cleared. There could be
+multiple mdevs assigned to different guests in the same VDCM. Since
+the PASID table is shared at PCI device level, lazy clearing is not
+secure. A malicious guest can attack by using newly freed PASIDs

that +are allocated by another guest.

Documentation/ioasid.rst new file +2 -0
+
+By holding a reference of the PASID until VDCM cleans up the HW

context, +it is guaranteed that PASID life cycles do not cross within the same +device.

Documentation/ioasid.rst new file +5 -0
+
+
+Reference
+====================================================
+1.

https://software.intel.com/sites/default/files/managed/c5/15/architecture-instruction-set-extensions-programming-reference.pdf
+ +2.
https://01.org/blogs/2019/introducing-intel-data-streaming-accelerator

Quoted 8 lines
Quoted 4 lines

+ +3.
https://software.intel.com/en-us/download/intel-data-streaming-accelerator-preliminary-architecture-specification
-- 2.7.4

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

[Jacob Pan]
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

Reviewed-by on v2 2/9
#12 lore raw

Re: [PATCH v2 2/9] iommu/ioasid: Rename ioasid_set_data()

In reply to Jacob Pan

Hi jacob,

Quoted 5 lines · Jacob Pan

Rename ioasid_set_data() to ioasid_attach_data() to avoid confusion with struct ioasid_set. ioasid_set is a group of IOASIDs that share a common token.

Eric

Quoted 82 lines
---
 drivers/iommu/intel/svm.c | 6 +++---
 drivers/iommu/ioasid.c    | 6 +++---
 include/linux/ioasid.h    | 4 ++--
 3 files changed, 8 insertions(+), 8 deletions(-)
drivers/iommu/intel/svm.c +3 -3
index b6972dca2ae0..37a9beabc0ca 100644
@@ -342,7 +342,7 @@ int intel_svm_bind_gpasid(struct iommu_domain *domain, struct device *dev,
 			svm->gpasid = data->gpasid;
 			svm->flags |= SVM_FLAG_GUEST_PASID;
 		}
-		ioasid_set_data(data->hpasid, svm);
+		ioasid_attach_data(data->hpasid, svm);
 		INIT_LIST_HEAD_RCU(&svm->devs);
 		mmput(svm->mm);
 	}
@@ -394,7 +394,7 @@ int intel_svm_bind_gpasid(struct iommu_domain *domain, struct device *dev,
 	list_add_rcu(&sdev->list, &svm->devs);
  out:
 	if (!IS_ERR_OR_NULL(svm) && list_empty(&svm->devs)) {
-		ioasid_set_data(data->hpasid, NULL);
+		ioasid_attach_data(data->hpasid, NULL);
 		kfree(svm);
 	}
 
@@ -437,7 +437,7 @@ int intel_svm_unbind_gpasid(struct device *dev, int pasid)
 				 * the unbind, IOMMU driver will get notified
 				 * and perform cleanup.
 				 */
-				ioasid_set_data(pasid, NULL);
+				ioasid_attach_data(pasid, NULL);
 				kfree(svm);
 			}
 		}
drivers/iommu/ioasid.c +3 -3
index 0f8dd377aada..5f63af07acd5 100644
@@ -258,14 +258,14 @@ void ioasid_unregister_allocator(struct ioasid_allocator_ops *ops)
 EXPORT_SYMBOL_GPL(ioasid_unregister_allocator);
 
 /**
- * ioasid_set_data - Set private data for an allocated ioasid
+ * ioasid_attach_data - Set private data for an allocated ioasid
  * @ioasid: the ID to set data
  * @data:   the private data
  *
  * For IOASID that is already allocated, private data can be set
  * via this API. Future lookup can be done via ioasid_find.
  */
-int ioasid_set_data(ioasid_t ioasid, void *data)
+int ioasid_attach_data(ioasid_t ioasid, void *data)
 {
 	struct ioasid_data *ioasid_data;
 	int ret = 0;
@@ -287,7 +287,7 @@ int ioasid_set_data(ioasid_t ioasid, void *data)
 
 	return ret;
 }
-EXPORT_SYMBOL_GPL(ioasid_set_data);
+EXPORT_SYMBOL_GPL(ioasid_attach_data);
 
 /**
  * ioasid_alloc - Allocate an IOASID
include/linux/ioasid.h +1 -1
index 6f000d7a0ddc..9c44947a68c8 100644
@@ -39,7 +39,7 @@ void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid,
 		  bool (*getter)(void *));
 int ioasid_register_allocator(struct ioasid_allocator_ops *allocator);
 void ioasid_unregister_allocator(struct ioasid_allocator_ops *allocator);
-int ioasid_set_data(ioasid_t ioasid, void *data);
+int ioasid_attach_data(ioasid_t ioasid, void *data);
 
 #else /* !CONFIG_IOASID */
 static inline ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min,

include/linux/ioasid.h static inline void ioasid_unregister_allocator(struct ioasid_allocator_ops *allo · line 73

Quoted 8 lines
include/linux/ioasid.h +1 -1
@@ -67,7 +67,7 @@ static inline void ioasid_unregister_allocator(struct ioasid_allocator_ops *allo
 {
 }
 
-static inline int ioasid_set_data(ioasid_t ioasid, void *data)
+static inline int ioasid_attach_data(ioasid_t ioasid, void *data)
 {
 	return -ENOTSUPP;
 }

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

#24 lore raw

Re: [PATCH v2 3/9] iommu/ioasid: Introduce ioasid_set APIs

In reply to Jacob Pan

Hi Jacob,

Quoted 1 line · Jacob Pan

ioasid_set was introduced as an arbitrary token that are shared by a

that is

Quoted 2 lines

group of IOASIDs. For example, if IOASID #1 and #2 are allocated via the same ioasid_set*, they are viewed as to belong to the same set.

two IOASIDs allocated via the same ioasid_set pointer belong to the same set?

Quoted 3 lines

For guest SVA usages, system-wide IOASID resources need to be partitioned such that VMs can have its own quota and being managed

their own

Quoted 9 lines

separately. ioasid_set is the perfect candidate for meeting such requirements. This patch redefines and extends ioasid_set with the following new fields:
- Quota
- Reference count
- Storage of its namespace
- The token is stored in the new ioasid_set but with optional types

ioasid_set level APIs are introduced that wires up these new data.

that wire

Quoted 36 lines

Existing users of IOASID APIs are converted where a host IOASID set is allocated for bare-metal usage.

---
 drivers/iommu/intel/iommu.c |  27 ++-
 drivers/iommu/intel/pasid.h |   1 +
 drivers/iommu/intel/svm.c   |   8 +-
 drivers/iommu/ioasid.c      | 390 +++++++++++++++++++++++++++++++++++++++++---
 include/linux/ioasid.h      |  82 ++++++++--
 5 files changed, 465 insertions(+), 43 deletions(-)
drivers/iommu/intel/iommu.c +4 -0
index a3a0b5c8921d..5813eeaa5edb 100644
@@ -42,6 +42,7 @@
 #include <linux/crash_dump.h>
 #include <linux/numa.h>
 #include <linux/swiotlb.h>
+#include <linux/ioasid.h>
 #include <asm/irq_remapping.h>
 #include <asm/cacheflush.h>
 #include <asm/iommu.h>
@@ -103,6 +104,9 @@
  */
 #define INTEL_IOMMU_PGSIZES	(~0xFFFUL)
 
+/* PASIDs used by host SVM */
+struct ioasid_set *host_pasid_set;
+
 static inline int agaw_to_level(int agaw)
 {
 	return agaw + 2;

drivers/iommu/intel/iommu.c static void intel_vcmd_ioasid_free(ioasid_t ioasid, void *data) · line 3111

Quoted 7 lines
drivers/iommu/intel/iommu.c +2 -2
@@ -3103,8 +3107,8 @@ static void intel_vcmd_ioasid_free(ioasid_t ioasid, void *data)
 	 * Sanity check the ioasid owner is done at upper layer, e.g. VFIO
 	 * We can only free the PASID when all the devices are unbound.
 	 */
-	if (ioasid_find(NULL, ioasid, NULL)) {
-		pr_alert("Cannot free active IOASID %d\n", ioasid);
+	if (IS_ERR(ioasid_find(host_pasid_set, ioasid, NULL))) {
+		pr_err("Cannot free IOASID %d, not in system set\n", ioasid);

not sure the change in the trace is worth. Also you may be more explicit like IOASID %d to be freed cannot be found in the system ioasid set. shouldn't it be rate_limited as it is originated from user space?

drivers/iommu/intel/iommu.c static int __init init_dmars(void) · line 3299

Quoted 12 lines
drivers/iommu/intel/iommu.c +5 -0
 		return;
 	}
 	vcmd_free_pasid(iommu, ioasid);
@@ -3288,6 +3292,19 @@ static int __init init_dmars(void)
 	if (ret)
 		goto free_iommu;
 
+	/* PASID is needed for scalable mode irrespective to SVM */
+	if (intel_iommu_sm) {
+		ioasid_install_capacity(intel_pasid_max_id);
+		/* We should not run out of IOASIDs at boot */
+		host_pasid_set = ioasid_alloc_set(NULL, PID_MAX_DEFAULT,

s/PID_MAX_DEFAULT/intel_pasid_max_id?

drivers/iommu/intel/iommu.c static int __init init_dmars(void) · line 3300

Quoted 1 line
drivers/iommu/intel/iommu.c +1 -0
+						IOASID_SET_TYPE_NULL);

as suggested by jean-Philippe ioasid_set_alloc

drivers/iommu/intel/iommu.c static int __init init_dmars(void) · line 3303

Quoted 3 lines
drivers/iommu/intel/iommu.c +3 -0
+		if (IS_ERR_OR_NULL(host_pasid_set)) {
+			pr_err("Failed to enable host PASID allocator %lu\n",
+				PTR_ERR(host_pasid_set));

does not sound like the correct error message? failed to allocate the system ioasid_set?

Quoted 17 lines
drivers/iommu/intel/iommu.c +5 -1
+			intel_iommu_sm = 0;
+		}
+	}
+
 	/*
 	 * for each drhd
 	 *   enable fault log
@@ -5149,7 +5166,7 @@ static void auxiliary_unlink_device(struct dmar_domain *domain,
 	domain->auxd_refcnt--;
 
 	if (!domain->auxd_refcnt && domain->default_pasid > 0)
-		ioasid_free(domain->default_pasid);
+		ioasid_free(host_pasid_set, domain->default_pasid);
 }
 
 static int aux_domain_add_dev(struct dmar_domain *domain,

drivers/iommu/intel/iommu.c static int aux_domain_add_dev(struct dmar_domain *domain, · line 5189

Quoted 7 lines
drivers/iommu/intel/iommu.c +1 -1
@@ -5167,7 +5184,7 @@ static int aux_domain_add_dev(struct dmar_domain *domain,
 		int pasid;
 
 		/* No private data needed for the default pasid */
-		pasid = ioasid_alloc(NULL, PASID_MIN,
+		pasid = ioasid_alloc(host_pasid_set, PASID_MIN,
 				     pci_max_pasids(to_pci_dev(dev)) - 1,
 				     NULL);

don't you want to ioasid_set_put() the ioasid_set in intel_iommu_free_dmars()?

Quoted 67 lines
drivers/iommu/intel/iommu.c +1 -1
 		if (pasid == INVALID_IOASID) {
@@ -5210,7 +5227,7 @@ static int aux_domain_add_dev(struct dmar_domain *domain,
 	spin_unlock(&iommu->lock);
 	spin_unlock_irqrestore(&device_domain_lock, flags);
 	if (!domain->auxd_refcnt && domain->default_pasid > 0)
-		ioasid_free(domain->default_pasid);
+		ioasid_free(host_pasid_set, domain->default_pasid);
 
 	return ret;
 }
drivers/iommu/intel/pasid.h +1 -0
index c9850766c3a9..ccdc23446015 100644
@@ -99,6 +99,7 @@ static inline bool pasid_pte_is_present(struct pasid_entry *pte)
 }
 
 extern u32 intel_pasid_max_id;
+extern struct ioasid_set *host_pasid_set;
 int intel_pasid_alloc_id(void *ptr, int start, int end, gfp_t gfp);
 void intel_pasid_free_id(int pasid);
 void *intel_pasid_lookup_id(int pasid);
drivers/iommu/intel/svm.c +4 -4
index 37a9beabc0ca..634e191ca2c3 100644
@@ -551,7 +551,7 @@ intel_svm_bind_mm(struct device *dev, int flags, struct svm_dev_ops *ops,
 			pasid_max = intel_pasid_max_id;
 
 		/* Do not use PASID 0, reserved for RID to PASID */
-		svm->pasid = ioasid_alloc(NULL, PASID_MIN,
+		svm->pasid = ioasid_alloc(host_pasid_set, PASID_MIN,
 					  pasid_max - 1, svm);
 		if (svm->pasid == INVALID_IOASID) {
 			kfree(svm);
@@ -568,7 +568,7 @@ intel_svm_bind_mm(struct device *dev, int flags, struct svm_dev_ops *ops,
 		if (mm) {
 			ret = mmu_notifier_register(&svm->notifier, mm);
 			if (ret) {
-				ioasid_free(svm->pasid);
+				ioasid_free(host_pasid_set, svm->pasid);
 				kfree(svm);
 				kfree(sdev);
 				goto out;
@@ -586,7 +586,7 @@ intel_svm_bind_mm(struct device *dev, int flags, struct svm_dev_ops *ops,
 		if (ret) {
 			if (mm)
 				mmu_notifier_unregister(&svm->notifier, mm);
-			ioasid_free(svm->pasid);
+			ioasid_free(host_pasid_set, svm->pasid);
 			kfree(svm);
 			kfree(sdev);
 			goto out;
@@ -655,7 +655,7 @@ static int intel_svm_unbind_mm(struct device *dev, int pasid)
 			kfree_rcu(sdev, rcu);
 
 			if (list_empty(&svm->devs)) {
-				ioasid_free(svm->pasid);
+				ioasid_free(host_pasid_set, svm->pasid);
 				if (svm->mm)
 					mmu_notifier_unregister(&svm->notifier, svm->mm);
 				list_del(&svm->list);
drivers/iommu/ioasid.c +0 -0
index 5f63af07acd5..f73b3dbfc37a 100644

drivers/iommu/ioasid.c line 4

Quoted 4 lines
drivers/iommu/ioasid.c +0 -1
@@ -1,22 +1,58 @@
 // SPDX-License-Identifier: GPL-2.0
 /*
  * I/O Address Space ID allocator. There is one global IOASID space, split into
- * subsets. Users create a subset with DECLARE_IOASID_SET, then allocate and

I would try to avoid using new terms: s/subset_ioset_set

drivers/iommu/ioasid.c line 4

Quoted 2 lines
drivers/iommu/ioasid.c +1 -1
- * free IOASIDs with ioasid_alloc and ioasid_free.
+ * subsets. Users create a subset with ioasid_alloc_set, then allocate/free IDs

here also and ioasid_set_alloc

Quoted 13 lines
drivers/iommu/ioasid.c +7 -1
+ * with ioasid_alloc and ioasid_free.
  */
-#include <linux/ioasid.h>
 #include <linux/module.h>
 #include <linux/slab.h>
 #include <linux/spinlock.h>
 #include <linux/xarray.h>
+#include <linux/ioasid.h>
+
+static DEFINE_XARRAY_ALLOC(ioasid_sets);
+enum ioasid_state {
+	IOASID_STATE_INACTIVE,
+	IOASID_STATE_ACTIVE,

drivers/iommu/ioasid.c line 26

Quoted 10 lines
drivers/iommu/ioasid.c +9 -0
+	IOASID_STATE_FREE_PENDING,
+};
 
+/**
+ * struct ioasid_data - Meta data about ioasid
+ *
+ * @id:		Unique ID
+ * @users	Number of active users
+ * @state	Track state of the IOASID
+ * @set		Meta data of the set this IOASID belongs to

s/Meta data of the set this IOASID belongs to/ioasid_set the asid belongs to

drivers/iommu/ioasid.c line 27

Quoted 1 line
drivers/iommu/ioasid.c +1 -0
+ * @private	Private data associated with the IOASID

I would have expected to find the private asid somewhere

Quoted 83 lines
drivers/iommu/ioasid.c +45 -5
+ * @rcu		For free after RCU grace period
+ */
 struct ioasid_data {
 	ioasid_t id;
 	struct ioasid_set *set;
+	refcount_t users;
+	enum ioasid_state state;
 	void *private;
 	struct rcu_head rcu;
 };
 
+/* Default to PCIe standard 20 bit PASID */
+#define PCI_PASID_MAX 0x100000
+static ioasid_t ioasid_capacity = PCI_PASID_MAX;
+static ioasid_t ioasid_capacity_avail = PCI_PASID_MAX;
+
+void ioasid_install_capacity(ioasid_t total)
+{
+	ioasid_capacity = ioasid_capacity_avail = total;
+}
+EXPORT_SYMBOL_GPL(ioasid_install_capacity);
+
+ioasid_t ioasid_get_capacity()
+{
+	return ioasid_capacity;
+}
+EXPORT_SYMBOL_GPL(ioasid_get_capacity);
+
 /*
  * struct ioasid_allocator_data - Internal data structure to hold information
  * about an allocator. There are two types of allocators:
@@ -306,11 +342,23 @@ ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max,
 {
 	struct ioasid_data *data;
 	void *adata;
-	ioasid_t id;
+	ioasid_t id = INVALID_IOASID;
+
+	spin_lock(&ioasid_allocator_lock);
+	/* Check if the IOASID set has been allocated and initialized */
+	if (WARN_ON(xa_load(&ioasid_sets, set->sid) != set)) {
+		pr_warn("Invalid set\n");
+		goto done_unlock;
+	}
+
+	if (set->quota <= set->nr_ioasids) {
+		pr_err("IOASID set %d out of quota %d\n", set->sid, set->quota);
+		goto done_unlock;
+	}
 
 	data = kzalloc(sizeof(*data), GFP_ATOMIC);
 	if (!data)
-		return INVALID_IOASID;
+		goto done_unlock;
 
 	data->set = set;
 	data->private = private;
@@ -319,7 +367,6 @@ ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max,
 	 * Custom allocator needs allocator data to perform platform specific
 	 * operations.
 	 */
-	spin_lock(&ioasid_allocator_lock);
 	adata = active_allocator->flags & IOASID_ALLOCATOR_CUSTOM ? active_allocator->ops->pdata : data;
 	id = active_allocator->ops->alloc(min, max, adata);
 	if (id == INVALID_IOASID) {
@@ -335,42 +382,339 @@ ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max,
 		goto exit_free;
 	}
 	data->id = id;
+	data->state = IOASID_STATE_ACTIVE;
+	refcount_set(&data->users, 1);
+
+	/* Store IOASID in the per set data */
+	if (xa_err(xa_store(&set->xa, id, data, GFP_ATOMIC))) {
+		pr_err("Failed to ioasid %d in set %d\n", id, set->sid);
+		goto exit_free;
+	}
+	set->nr_ioasids++;
+	goto done_unlock;
 
-	spin_unlock(&ioasid_allocator_lock);
-	return id;
 exit_free:

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 404

Quoted 10 lines
drivers/iommu/ioasid.c +4 -2
-	spin_unlock(&ioasid_allocator_lock);
 	kfree(data);
-	return INVALID_IOASID;
+done_unlock:
+	spin_unlock(&ioasid_allocator_lock);
+	return id;
 }
 EXPORT_SYMBOL_GPL(ioasid_alloc);
 
+static void ioasid_do_free(struct ioasid_data *data)

do_free_locked?

Quoted 334 lines
drivers/iommu/ioasid.c +300 -13
+{
+	struct ioasid_data *ioasid_data;
+	struct ioasid_set *sdata;
+
+	active_allocator->ops->free(data->id, active_allocator->ops->pdata);
+	/* Custom allocator needs additional steps to free the xa element */
+	if (active_allocator->flags & IOASID_ALLOCATOR_CUSTOM) {
+		ioasid_data = xa_erase(&active_allocator->xa, data->id);
+		kfree_rcu(ioasid_data, rcu);
+	}
+
+	sdata = xa_load(&ioasid_sets, data->set->sid);
+	if (!sdata) {
+		pr_err("No set %d for IOASID %d\n", data->set->sid,
+		       data->id);
+		return;
+	}
+	xa_erase(&sdata->xa, data->id);
+	sdata->nr_ioasids--;
+}
+
+static void ioasid_free_locked(struct ioasid_set *set, ioasid_t ioasid)
+{
+	struct ioasid_data *data;
+
+	data = xa_load(&active_allocator->xa, ioasid);
+	if (!data) {
+		pr_err("Trying to free unknown IOASID %u\n", ioasid);
+		return;
+	}
+
+	if (data->set != set) {
+		pr_warn("Cannot free IOASID %u due to set ownership\n", ioasid);
+		return;
+	}
+	data->state = IOASID_STATE_FREE_PENDING;
+
+	if (!refcount_dec_and_test(&data->users))
+		return;
+
+	ioasid_do_free(data);
+}
+
 /**
- * ioasid_free - Free an IOASID
- * @ioasid: the ID to remove
+ * ioasid_free - Drop reference on an IOASID. Free if refcount drops to 0,
+ *               including free from its set and system-wide list.
+ * @set:	The ioasid_set to check permission with. If not NULL, IOASID
+ *		free will fail if the set does not match.
+ * @ioasid:	The IOASID to remove
  */
-void ioasid_free(ioasid_t ioasid)
+void ioasid_free(struct ioasid_set *set, ioasid_t ioasid)
 {
-	struct ioasid_data *ioasid_data;
+	spin_lock(&ioasid_allocator_lock);
+	ioasid_free_locked(set, ioasid);
+	spin_unlock(&ioasid_allocator_lock);
+}
+EXPORT_SYMBOL_GPL(ioasid_free);
 
+/**
+ * ioasid_alloc_set - Allocate a new IOASID set for a given token
+ *
+ * @token:	Unique token of the IOASID set, cannot be NULL
+ * @quota:	Quota allowed in this set. Only for new set creation
+ * @flags:	Special requirements
+ *
+ * IOASID can be limited system-wide resource that requires quota management.
+ * If caller does not wish to enforce quota, use IOASID_SET_NO_QUOTA flag.
+ *
+ * Token will be stored in the ioasid_set returned. A reference will be taken
+ * upon finding a matching set or newly created set.
+ * IOASID allocation within the set and other per set operations will use
+ * the retured ioasid_set *.
+ *
+ */
+struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota, int type)
+{
+	struct ioasid_set *sdata;
+	unsigned long index;
+	ioasid_t id;
+
+	if (type >= IOASID_SET_TYPE_NR)
+		return ERR_PTR(-EINVAL);
+
+	/*
+	 * Need to check space available if we share system-wide quota.
+	 * TODO: we may need to support quota free sets in the future.
+	 */
 	spin_lock(&ioasid_allocator_lock);
-	ioasid_data = xa_load(&active_allocator->xa, ioasid);
-	if (!ioasid_data) {
-		pr_err("Trying to free unknown IOASID %u\n", ioasid);
+	if (quota > ioasid_capacity_avail) {
+		pr_warn("Out of IOASID capacity! ask %d, avail %d\n",
+			quota, ioasid_capacity_avail);
+		sdata = ERR_PTR(-ENOSPC);
 		goto exit_unlock;
 	}
 
-	active_allocator->ops->free(ioasid, active_allocator->ops->pdata);
-	/* Custom allocator needs additional steps to free the xa element */
-	if (active_allocator->flags & IOASID_ALLOCATOR_CUSTOM) {
-		ioasid_data = xa_erase(&active_allocator->xa, ioasid);
-		kfree_rcu(ioasid_data, rcu);
+	/*
+	 * Token is only unique within its types but right now we have only
+	 * mm type. If we have more token types, we have to match type as well.
+	 */
+	switch (type) {
+	case IOASID_SET_TYPE_MM:
+		/* Search existing set tokens, reject duplicates */
+		xa_for_each(&ioasid_sets, index, sdata) {
+			if (sdata->token == token &&
+				sdata->type == IOASID_SET_TYPE_MM) {
+				sdata = ERR_PTR(-EEXIST);
+				goto exit_unlock;
+			}
+		}
+		break;
+	case IOASID_SET_TYPE_NULL:
+		if (!token)
+			break;
+		fallthrough;
+	default:
+		pr_err("Invalid token and IOASID type\n");
+		sdata = ERR_PTR(-EINVAL);
+		goto exit_unlock;
 	}
 
+	/* REVISIT: may support set w/o quota, use system available */
+	if (!quota) {
+		sdata = ERR_PTR(-EINVAL);
+		goto exit_unlock;
+	}
+
+	sdata = kzalloc(sizeof(*sdata), GFP_ATOMIC);
+	if (!sdata) {
+		sdata = ERR_PTR(-ENOMEM);
+		goto exit_unlock;
+	}
+
+	if (xa_alloc(&ioasid_sets, &id, sdata,
+			XA_LIMIT(0, ioasid_capacity_avail - quota),
+			GFP_ATOMIC)) {
+		kfree(sdata);
+		sdata = ERR_PTR(-ENOSPC);
+		goto exit_unlock;
+	}
+
+	sdata->token = token;
+	sdata->type = type;
+	sdata->quota = quota;
+	sdata->sid = id;
+	refcount_set(&sdata->ref, 1);
+
+	/*
+	 * Per set XA is used to store private IDs within the set, get ready
+	 * for ioasid_set private ID and system-wide IOASID allocation
+	 * results.
+	 */
+	xa_init_flags(&sdata->xa, XA_FLAGS_ALLOC);
+	ioasid_capacity_avail -= quota;
+
 exit_unlock:
 	spin_unlock(&ioasid_allocator_lock);
+
+	return sdata;
 }
-EXPORT_SYMBOL_GPL(ioasid_free);
+EXPORT_SYMBOL_GPL(ioasid_alloc_set);
+
+void ioasid_set_get_locked(struct ioasid_set *set)
+{
+	if (WARN_ON(xa_load(&ioasid_sets, set->sid) != set)) {
+		pr_warn("Invalid set data\n");
+		return;
+	}
+
+	refcount_inc(&set->ref);
+}
+EXPORT_SYMBOL_GPL(ioasid_set_get_locked);
+
+void ioasid_set_get(struct ioasid_set *set)
+{
+	spin_lock(&ioasid_allocator_lock);
+	ioasid_set_get_locked(set);
+	spin_unlock(&ioasid_allocator_lock);
+}
+EXPORT_SYMBOL_GPL(ioasid_set_get);
+
+void ioasid_set_put_locked(struct ioasid_set *set)
+{
+	struct ioasid_data *entry;
+	unsigned long index;
+
+	if (WARN_ON(xa_load(&ioasid_sets, set->sid) != set)) {
+		pr_warn("Invalid set data\n");
+		return;
+	}
+
+	if (!refcount_dec_and_test(&set->ref)) {
+		pr_debug("%s: IOASID set %d has %d users\n",
+			__func__, set->sid, refcount_read(&set->ref));
+		return;
+	}
+
+	/* The set is already empty, we just destroy the set. */
+	if (xa_empty(&set->xa))
+		goto done_destroy;
+
+	/*
+	 * Free all PASIDs from system-wide IOASID pool, all subscribers gets
+	 * notified and do cleanup of their own.
+	 * Note that some references of the IOASIDs within the set can still
+	 * be held after the free call. This is OK in that the IOASIDs will be
+	 * marked inactive, the only operations can be done is ioasid_put.
+	 * No need to track IOASID set states since there is no reclaim phase.
+	 */
+	xa_for_each(&set->xa, index, entry) {
+		ioasid_free_locked(set, index);
+		/* Free from per set private pool */
+		xa_erase(&set->xa, index);
+	}
+
+done_destroy:
+	/* Return the quota back to system pool */
+	ioasid_capacity_avail += set->quota;
+	kfree_rcu(set, rcu);
+
+	/*
+	 * Token got released right away after the ioasid_set is freed.
+	 * If a new set is created immediately with the newly released token,
+	 * it will not allocate the same IOASIDs unless they are reclaimed.
+	 */
+	xa_erase(&ioasid_sets, set->sid);
+}
+EXPORT_SYMBOL_GPL(ioasid_set_put_locked);
+
+/**
+ * ioasid_set_put - Drop a reference to the IOASID set. Free all IOASIDs within
+ *                  the set if there are no more users.
+ *
+ * @set:	The IOASID set ID to be freed
+ *
+ * If refcount drops to zero, all IOASIDs allocated within the set will be
+ * freed.
+ */
+void ioasid_set_put(struct ioasid_set *set)
+{
+	spin_lock(&ioasid_allocator_lock);
+	ioasid_set_put_locked(set);
+	spin_unlock(&ioasid_allocator_lock);
+}
+EXPORT_SYMBOL_GPL(ioasid_set_put);
+
+/**
+ * ioasid_adjust_set - Adjust the quota of an IOASID set
+ * @set:	IOASID set to be assigned
+ * @quota:	Quota allowed in this set
+ *
+ * Return 0 on success. If the new quota is smaller than the number of
+ * IOASIDs already allocated, -EINVAL will be returned. No change will be
+ * made to the existing quota.
+ */
+int ioasid_adjust_set(struct ioasid_set *set, int quota)
+{
+	int ret = 0;
+
+	spin_lock(&ioasid_allocator_lock);
+	if (set->nr_ioasids > quota) {
+		pr_err("New quota %d is smaller than outstanding IOASIDs %d\n",
+			quota, set->nr_ioasids);
+		ret = -EINVAL;
+		goto done_unlock;
+	}
+
+	if (quota >= ioasid_capacity_avail) {
+		ret = -ENOSPC;
+		goto done_unlock;
+	}
+
+	/* Return the delta back to system pool */
+	ioasid_capacity_avail += set->quota - quota;
+
+	/*
+	 * May have a policy to prevent giving all available IOASIDs
+	 * to one set. But we don't enforce here, it should be in the
+	 * upper layers.
+	 */
+	set->quota = quota;
+
+done_unlock:
+	spin_unlock(&ioasid_allocator_lock);
+
+	return ret;
+}
+EXPORT_SYMBOL_GPL(ioasid_adjust_set);
+
+/**
+ * ioasid_set_for_each_ioasid - Iterate over all the IOASIDs within the set
+ *
+ * Caller must hold a reference of the set and handles its own locking.
+ */
+int ioasid_set_for_each_ioasid(struct ioasid_set *set,
+			       void (*fn)(ioasid_t id, void *data),
+			       void *data)
+{
+	struct ioasid_data *entry;
+	unsigned long index;
+	int ret = 0;
+
+	if (xa_empty(&set->xa)) {
+		pr_warn("No IOASIDs in the set %d\n", set->sid);
+		return -ENOENT;
+	}
+
+	xa_for_each(&set->xa, index, entry) {
+		fn(index, data);
+	}
+
+	return ret;
+}
+EXPORT_SYMBOL_GPL(ioasid_set_for_each_ioasid);
 
 /**
  * ioasid_find - Find IOASID data
include/linux/ioasid.h +0 -0
index 9c44947a68c8..412d025d440e 100644

include/linux/ioasid.h typedef unsigned int ioasid_t; · line 16

Quoted 7 lines
include/linux/ioasid.h +4 -0
@@ -10,8 +10,35 @@ typedef unsigned int ioasid_t;
 typedef ioasid_t (*ioasid_alloc_fn_t)(ioasid_t min, ioasid_t max, void *data);
 typedef void (*ioasid_free_fn_t)(ioasid_t ioasid, void *data);
 
+/* IOASID set types */
+enum ioasid_set_type {
+	IOASID_SET_TYPE_NULL = 1, /* Set token is NULL */
+	IOASID_SET_TYPE_MM,	  /* Set token is a mm_struct,

s/mm_struct/mm_struct pointer

include/linux/ioasid.h typedef unsigned int ioasid_t; · line 24

Quoted 8 lines
include/linux/ioasid.h +8 -0
+				   * i.e. associated with a process
+				   */
+	IOASID_SET_TYPE_NR,
+};
+
+/**
+ * struct ioasid_set - Meta data about ioasid_set
+ * @type:	Token types and other features

token type. Why "and other features"

Quoted 6 lines
include/linux/ioasid.h +6 -0
+ * @token:	Unique to identify an IOASID set
+ * @xa:		XArray to store ioasid_set private IDs, can be used for
+ *		guest-host IOASID mapping, or just a private IOASID namespace.
+ * @quota:	Max number of IOASIDs can be allocated within the set
+ * @nr_ioasids	Number of IOASIDs currently allocated in the set
+ * @sid:	ID of the set

include/linux/ioasid.h typedef unsigned int ioasid_t; · line 39

Quoted 10 lines
include/linux/ioasid.h +8 -1
+ * @ref:	Reference count of the users
+ */
 struct ioasid_set {
-	int dummy;
+	void *token;
+	struct xarray xa;
+	int type;
+	int quota;
+	int nr_ioasids;
+	int sid;

nit id? sid has a special meaning on ARM.

Quoted 6 lines
include/linux/ioasid.h +2 -0
+	refcount_t ref;
+	struct rcu_head rcu;
 };
 
 /**

include/linux/ioasid.h struct ioasid_allocator_ops { · line 63

Quoted 10 lines
include/linux/ioasid.h +4 -2
@@ -29,31 +56,64 @@ struct ioasid_allocator_ops {
 	void *pdata;
 };
 
-#define DECLARE_IOASID_SET(name) struct ioasid_set name = { 0 }
-
 #if IS_ENABLED(CONFIG_IOASID)
+void ioasid_install_capacity(ioasid_t total);
+ioasid_t ioasid_get_capacity(void);
+struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota, int type);
+int ioasid_adjust_set(struct ioasid_set *set, int quota);

ioasid_set_adjust_quota

include/linux/ioasid.h struct ioasid_allocator_ops { · line 64

Quoted 1 line
include/linux/ioasid.h +1 -0
+void ioasid_set_get_locked(struct ioasid_set *set);

as mentionned during the Plumber uConf, the set_get is unfortunate. Globally I wonder if we shouldn't rename "set" into "pool" or something alike.

Quoted 54 lines
include/linux/ioasid.h +36 -6
+void ioasid_set_put_locked(struct ioasid_set *set);
+void ioasid_set_put(struct ioasid_set *set);
+
 ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max,
 		      void *private);
-void ioasid_free(ioasid_t ioasid);
-void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid,
-		  bool (*getter)(void *));
+void ioasid_free(struct ioasid_set *set, ioasid_t ioasid);
+
+bool ioasid_is_active(ioasid_t ioasid);
+
+void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid, bool (*getter)(void *));
+int ioasid_attach_data(ioasid_t ioasid, void *data);
 int ioasid_register_allocator(struct ioasid_allocator_ops *allocator);
 void ioasid_unregister_allocator(struct ioasid_allocator_ops *allocator);
-int ioasid_attach_data(ioasid_t ioasid, void *data);
-
+void ioasid_is_in_set(struct ioasid_set *set, ioasid_t ioasid);
+int ioasid_set_for_each_ioasid(struct ioasid_set *sdata,
+			       void (*fn)(ioasid_t id, void *data),
+			       void *data);
 #else /* !CONFIG_IOASID */
+static inline void ioasid_install_capacity(ioasid_t total)
+{
+}
+
+static inline ioasid_t ioasid_get_capacity(void)
+{
+	return 0;
+}
+
 static inline ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min,
 				    ioasid_t max, void *private)
 {
 	return INVALID_IOASID;
 }
 
-static inline void ioasid_free(ioasid_t ioasid)
+static inline void ioasid_free(struct ioasid_set *set, ioasid_t ioasid)
+{
+}
+
+static inline bool ioasid_is_active(ioasid_t ioasid)
+{
+	return false;
+}
+
+static inline struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota, int type)
+{
+	return ERR_PTR(-ENOTSUPP);
+}
+

include/linux/ioasid.h struct ioasid_allocator_ops { · line 119

Quoted 10 lines
include/linux/ioasid.h +2 -2
+static inline void ioasid_set_put(struct ioasid_set *set)
 {
 }
 
-static inline void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid,
-				bool (*getter)(void *))
+static inline void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid, bool (*getter)(void *))
 {
 	return NULL;
 }

I felt very difficult to review this patch. Could you split it into several ones? maybe introduce the a dummy host_pasid_set and update the call sites accordingling.

You introduce ownership checking, quota checking, ioasid state, ref counting, ioasid type handling (whereas existing is NULL) so I have the feeling that could ease the review process by adopting a more incremental approach.

Thanks

Eric

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

#32 lore raw

Re: [PATCH v2 4/9] iommu/ioasid: Add reference couting functions

In reply to Jacob Pan

Hi Jacob,

Quoted 19 lines · Jacob Pan

There can be multiple users of an IOASID, each user could have hardware contexts associated with the IOASID. In order to align lifecycles, reference counting is introduced in this patch. It is expected that when an IOASID is being freed, each user will drop a reference only after its context is cleared.

---
 drivers/iommu/ioasid.c | 113 +++++++++++++++++++++++++++++++++++++++++++++++++
 include/linux/ioasid.h |   4 ++
 2 files changed, 117 insertions(+)
drivers/iommu/ioasid.c +0 -0
index f73b3dbfc37a..5f31d63c75b1 100644
@@ -717,6 +717,119 @@ int ioasid_set_for_each_ioasid(struct ioasid_set *set,
 EXPORT_SYMBOL_GPL(ioasid_set_for_each_ioasid);
 

drivers/iommu/ioasid.c int ioasid_set_for_each_ioasid(struct ioasid_set *set, · line 728

Quoted 10 lines
drivers/iommu/ioasid.c +9 -0
 /**
+ * IOASID refcounting rules
+ * - ioasid_alloc() set initial refcount to 1
+ *
+ * - ioasid_free() decrement and test refcount.
+ *     If refcount is 0, ioasid will be freed. Deleted from the system-wide
+ *     xarray as well as per set xarray. The IOASID will be returned to the
+ *     pool and available for new allocations.
+ *
+ *     If recount is non-zero, mark IOASID as IOASID_STATE_FREE_PENDING.

s/recount/refcount

drivers/iommu/ioasid.c int ioasid_set_for_each_ioasid(struct ioasid_set *set, · line 729

Quoted 1 line
drivers/iommu/ioasid.c +1 -0
+ *     No new reference can be added. The IOASID is not returned to the pool

can be taken

drivers/iommu/ioasid.c int ioasid_set_for_each_ioasid(struct ioasid_set *set, · line 741

Quoted 12 lines
drivers/iommu/ioasid.c +12 -0
+ *     for reuse.
+ *     After free, ioasid_get() will return error but ioasid_find() and other
+ *     non refcount adding APIs will continue to work until the last reference
+ *     is dropped
+ *
+ * - ioasid_get() get a reference on an active IOASID
+ *
+ * - ioasid_put() decrement and test refcount of the IOASID.
+ *     If refcount is 0, ioasid will be freed. Deleted from the system-wide
+ *     xarray as well as per set xarray. The IOASID will be returned to the
+ *     pool and available for new allocations.
+ *     Do nothing if refcount is non-zero.

I would drop this last sentence

drivers/iommu/ioasid.c int ioasid_set_for_each_ioasid(struct ioasid_set *set, · line 743

Quoted 2 lines
drivers/iommu/ioasid.c +2 -0
+ *
+ * - ioasid_find() does not take reference, caller must hold reference

So can you use ioasid_find() once the state is IOASID_STATE_FREE_PENDING? According to Jean's reply, the "IOASID may be freed once ioasid_find() returns but not the returned data." So holding a ref is not mandated right?

Quoted 11 lines
drivers/iommu/ioasid.c +11 -0
+ *
+ * ioasid_free() can be called multiple times without error until all refs are
+ * dropped.
+ */
+
+int ioasid_get_locked(struct ioasid_set *set, ioasid_t ioasid)
+{
+	struct ioasid_data *data;
+
+	data = xa_load(&active_allocator->xa, ioasid);
+	if (!data) {

drivers/iommu/ioasid.c int ioasid_set_for_each_ioasid(struct ioasid_set *set, · line 764

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+		pr_err("Trying to get unknown IOASID %u\n", ioasid);
+		return -EINVAL;
+	}
+	if (data->state == IOASID_STATE_FREE_PENDING) {
+		pr_err("Trying to get IOASID being freed%u\n", ioasid);
+		return -EBUSY;
+	}
+
+	if (set && data->set != set) {
+		pr_err("Trying to get IOASID not in set%u\n", ioasid);

maybe try to normalize your traces using always the same formatting for ioasids and ioasid sets. Also we can understand %u is the id of the set.

drivers/iommu/ioasid.c int ioasid_set_for_each_ioasid(struct ioasid_set *set, · line 765

Quoted 1 line
drivers/iommu/ioasid.c +1 -0
+		/* data found but does not belong to the set */

you can get rid of this comment

Quoted 29 lines
drivers/iommu/ioasid.c +29 -0
+		return -EACCES;
+	}
+	refcount_inc(&data->users);
+
+	return 0;
+}
+EXPORT_SYMBOL_GPL(ioasid_get_locked);
+
+/**
+ * ioasid_get - Obtain a reference of an ioasid
+ * @set
+ * @ioasid
+ *
+ * Check set ownership if @set is non-null.
+ */
+int ioasid_get(struct ioasid_set *set, ioasid_t ioasid)
+{
+	int ret = 0;
+
+	spin_lock(&ioasid_allocator_lock);
+	ret = ioasid_get_locked(set, ioasid);
+	spin_unlock(&ioasid_allocator_lock);
+
+	return ret;
+}
+EXPORT_SYMBOL_GPL(ioasid_get);
+
+void ioasid_put_locked(struct ioasid_set *set, ioasid_t ioasid)
+{

drivers/iommu/ioasid.c int ioasid_set_for_each_ioasid(struct ioasid_set *set, · line 804

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+	struct ioasid_data *data;
+
+	data = xa_load(&active_allocator->xa, ioasid);
+	if (!data) {
+		pr_err("Trying to put unknown IOASID %u\n", ioasid);
+		return;
+	}
+
+	if (set && data->set != set) {
+		pr_err("Trying to drop IOASID not in the set %u\n", ioasid);

was set%u above

Quoted 37 lines
drivers/iommu/ioasid.c +28 -0
+		return;
+	}
+
+	if (!refcount_dec_and_test(&data->users)) {
+		pr_debug("%s: IOASID %d has %d remainning users\n",
+			__func__, ioasid, refcount_read(&data->users));
+		return;
+	}
+	ioasid_do_free(data);
+}
+EXPORT_SYMBOL_GPL(ioasid_put_locked);
+
+/**
+ * ioasid_put - Drop a reference of an ioasid
+ * @set
+ * @ioasid
+ *
+ * Check set ownership if @set is non-null.
+ */
+void ioasid_put(struct ioasid_set *set, ioasid_t ioasid)
+{
+	spin_lock(&ioasid_allocator_lock);
+	ioasid_put_locked(set, ioasid);
+	spin_unlock(&ioasid_allocator_lock);
+}
+EXPORT_SYMBOL_GPL(ioasid_put);
+
+/**
  * ioasid_find - Find IOASID data
  * @set: the IOASID set
  * @ioasid: the IOASID to find
include/linux/ioasid.h +0 -0
index 412d025d440e..310abe4187a3 100644

include/linux/ioasid.h int ioasid_attach_data(ioasid_t ioasid, void *data); · line 85

Quoted 10 lines
include/linux/ioasid.h +4 -0
@@ -76,6 +76,10 @@ int ioasid_attach_data(ioasid_t ioasid, void *data);
 int ioasid_register_allocator(struct ioasid_allocator_ops *allocator);
 void ioasid_unregister_allocator(struct ioasid_allocator_ops *allocator);
 void ioasid_is_in_set(struct ioasid_set *set, ioasid_t ioasid);
+int ioasid_get(struct ioasid_set *set, ioasid_t ioasid);
+int ioasid_get_locked(struct ioasid_set *set, ioasid_t ioasid);
+void ioasid_put(struct ioasid_set *set, ioasid_t ioasid);
+void ioasid_put_locked(struct ioasid_set *set, ioasid_t ioasid);
 int ioasid_set_for_each_ioasid(struct ioasid_set *sdata,
 			       void (*fn)(ioasid_t id, void *data),
 			       void *data);

Thanks

Eric

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

#40 lore raw

Re: [PATCH v2 5/9] iommu/ioasid: Introduce ioasid_set private ID

In reply to Jacob Pan

Hi Jacob,

Quoted 39 lines · Jacob Pan

When an IOASID set is used for guest SVA, each VM will acquire its ioasid_set for IOASID allocations. IOASIDs within the VM must have a host/physical IOASID backing, mapping between guest and host IOASIDs can be non-identical. IOASID set private ID (SPID) is introduced in this patch to be used as guest IOASID. However, the concept of ioasid_set specific namespace is generic, thus named SPID.

As SPID namespace is within the IOASID set, the IOASID core can provide lookup services at both directions. SPIDs may not be allocated when its IOASID is allocated, the mapping between SPID and IOASID is usually established when a guest page table is bound to a host PASID.

---
 drivers/iommu/ioasid.c | 54 ++++++++++++++++++++++++++++++++++++++++++++++++++
 include/linux/ioasid.h | 12 +++++++++++
 2 files changed, 66 insertions(+)
drivers/iommu/ioasid.c +2 -0
index 5f31d63c75b1..c0aef38a4fde 100644
@@ -21,6 +21,7 @@ enum ioasid_state {
  * struct ioasid_data - Meta data about ioasid
  *
  * @id:		Unique ID
+ * @spid:	Private ID unique within a set
  * @users	Number of active users
  * @state	Track state of the IOASID
  * @set		Meta data of the set this IOASID belongs to
@@ -29,6 +30,7 @@ enum ioasid_state {
  */
 struct ioasid_data {
 	ioasid_t id;
+	ioasid_t spid;
 	struct ioasid_set *set;
 	refcount_t users;
 	enum ioasid_state state;

drivers/iommu/ioasid.c int ioasid_attach_data(ioasid_t ioasid, void *data) · line 337

Quoted 10 lines
drivers/iommu/ioasid.c +7 -0
@@ -326,6 +328,58 @@ int ioasid_attach_data(ioasid_t ioasid, void *data)
 EXPORT_SYMBOL_GPL(ioasid_attach_data);
 
 /**
+ * ioasid_attach_spid - Attach ioasid_set private ID to an IOASID
+ *
+ * @ioasid: the ID to attach
+ * @spid:   the ioasid_set private ID of @ioasid
+ *
+ * For IOASID that is already allocated, private ID within the set can be
+ * attached via this API. Future lookup can be done via ioasid_find.

I would remove "For IOASID that is already allocated, private ID within the set can be attached via this API"

drivers/iommu/ioasid.c int ioasid_attach_data(ioasid_t ioasid, void *data) · line 344

Quoted 7 lines
drivers/iommu/ioasid.c +7 -0
+ */
+int ioasid_attach_spid(ioasid_t ioasid, ioasid_t spid)
+{
+	struct ioasid_data *ioasid_data;
+	int ret = 0;
+
+	spin_lock(&ioasid_allocator_lock);

We keep on saying the SPID is local to an IOASID set but we don't check any IOASID set contains this ioasid. It looks a bit weird to me.

drivers/iommu/ioasid.c int ioasid_attach_data(ioasid_t ioasid, void *data) · line 353

Quoted 9 lines
drivers/iommu/ioasid.c +9 -0
+	ioasid_data = xa_load(&active_allocator->xa, ioasid);
+
+	if (!ioasid_data) {
+		pr_err("No IOASID entry %d to attach SPID %d\n",
+			ioasid, spid);
+		ret = -ENOENT;
+		goto done_unlock;
+	}
+	ioasid_data->spid = spid;

is there any way/need to remove an spid binding?

Quoted 52 lines
drivers/iommu/ioasid.c +29 -0
+
+done_unlock:
+	spin_unlock(&ioasid_allocator_lock);
+	return ret;
+}
+EXPORT_SYMBOL_GPL(ioasid_attach_spid);
+
+ioasid_t ioasid_find_by_spid(struct ioasid_set *set, ioasid_t spid)
+{
+	struct ioasid_data *entry;
+	unsigned long index;
+
+	if (!xa_load(&ioasid_sets, set->sid)) {
+		pr_warn("Invalid set\n");
+		return INVALID_IOASID;
+	}
+
+	xa_for_each(&set->xa, index, entry) {
+		if (spid == entry->spid) {
+			pr_debug("Found ioasid %lu by spid %u\n", index, spid);
+			refcount_inc(&entry->users);
+			return index;
+		}
+	}
+	return INVALID_IOASID;
+}
+EXPORT_SYMBOL_GPL(ioasid_find_by_spid);
+
+/**
  * ioasid_alloc - Allocate an IOASID
  * @set: the IOASID set
  * @min: the minimum ID (inclusive)
include/linux/ioasid.h +4 -0
index 310abe4187a3..d4b3e83672f6 100644
@@ -73,6 +73,8 @@ bool ioasid_is_active(ioasid_t ioasid);
 
 void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid, bool (*getter)(void *));
 int ioasid_attach_data(ioasid_t ioasid, void *data);
+int ioasid_attach_spid(ioasid_t ioasid, ioasid_t spid);
+ioasid_t ioasid_find_by_spid(struct ioasid_set *set, ioasid_t spid);
 int ioasid_register_allocator(struct ioasid_allocator_ops *allocator);
 void ioasid_unregister_allocator(struct ioasid_allocator_ops *allocator);
 void ioasid_is_in_set(struct ioasid_set *set, ioasid_t ioasid);
@@ -136,5 +138,15 @@ static inline int ioasid_attach_data(ioasid_t ioasid, void *data)
 	return -ENOTSUPP;
 }
 
+staic inline int ioasid_attach_spid(ioasid_t ioasid, ioasid_t spid)
+{

include/linux/ioasid.h static inline int ioasid_attach_data(ioasid_t ioasid, void *data) · line 152

Quoted 10 lines
include/linux/ioasid.h +8 -0
+	return -ENOTSUPP;
+}
+
+static inline ioasid_t ioasid_find_by_spid(struct ioasid_set *set, ioasid_t spid)
+{
+	return -ENOTSUPP;
+}
+
 #endif /* CONFIG_IOASID */
 #endif /* __LINUX_IOASID_H */

Thanks

Eric

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

#6 lore raw

Re: [PATCH v2 1/9] docs: Document IO Address Space ID (IOASID) APIs

In reply to Auger Eric

Hi Eric,

Quoted 42 lines · On Thu, 27 Aug 2020 18:21:07 +0200 Auger Eric <[email protected]> wrote:

Hi Jacob,

Quoted 24 lines · Jean-Philippe Brucker

Documentation/ioasid.rst

Quoted 18 lines · Jacob Pan

IOASID is used to identify address spaces that can be targeted by device DMA. It is a system-wide resource that is essential to its many users. This document is an attempt to help developers from all vendors navigate the APIs. At this time, ARM SMMU and Intel’s Scalable IO Virtualization (SIOV) enabled platforms are the primary users of IOASID. Examples of how SIOV components interact with IOASID APIs are provided in that many APIs are driven by the requirements from SIOV.

---
 Documentation/ioasid.rst | 618

+++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 618 insertions(+) create mode 100644 Documentation/ioasid.rst

Documentation/ioasid.rst +0 -0

Thanks for writing this up. Should it go to Documentation/driver-api/, or Documentation/driver-api/iommu/? I think this also needs to Cc [email protected] and [email protected]

Quoted 16 lines
Quoted 12 lines

new file mode 100644
index 000000000000..b6a8cdc885ff

Documentation/ioasid.rst new file +7 -0
@@ -0,0 +1,618 @@
+.. ioasid:
+
+=====================================
+IO Address Space ID
+=====================================
+
+IOASID is a generic name for PCIe Process Address ID (PASID) or
Quoted 2 lines

ARM +SMMU sub-stream ID. An IOASID identifies an address space that DMA

"SubstreamID"

Quoted 4 lines

On ARM if we don't use PASIDs we have streamids (SID) which can also identify address spaces that DMA requests can target. So maybe this definition is not sufficient.

According to SMMU spec, the SubstreamID is equivalent to PASID. My understanding is that SID is equivalent to PCI requester ID that identifies stage 2. Do you plan to use IOASID for stage 2? IOASID is mostly for SVA and DMA request w/ PASID.

Quoted 68 lines
Quoted 55 lines
Quoted 6 lines
Documentation/ioasid.rst new file +3 -0
+requests can target.
+
+The primary use cases for IOASID are Shared Virtual Address (SVA)

and +IO Virtual Address (IOVA). However, the requirements for IOASID

IOVA alone isn't a use case, maybe "multiple IOVA spaces per device"?

Quoted 5 lines
Documentation/ioasid.rst new file +4 -0
+management can vary among hardware architectures.
+
+This document covers the generic features supported by IOASID
+APIs. Vendor-specific use cases are also illustrated with Intel's

VT-d +based platforms as the first example.

Documentation/ioasid.rst line 26

Quoted 9 lines
Documentation/ioasid.rst new file +9 -0
+
+.. contents:: :local:
+
+Glossary
+========
+PASID - Process Address Space ID
+
+IOASID - IO Address Space ID (generic term for PCIe PASID and
+sub-stream ID in SMMU)  

"SubstreamID"

Quoted 5 lines
Documentation/ioasid.rst new file +4 -0
+
+SVA/SVM - Shared Virtual Addressing/Memory
+
+ENQCMD - New Intel X86 ISA for efficient workqueue submission

[1]

Maybe drop the "New", to keep the documentation perennial. It might be good to add internal links here to the specifications URLs at the bottom.

Quoted 6 lines
Documentation/ioasid.rst new file +6 -0
+
+DSA - Intel Data Streaming Accelerator [2]
+
+VDCM - Virtual device composition module [3]
+
+SIOV - Intel Scalable IO Virtualization

Documentation/ioasid.rst line 46

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+
+
+Key Concepts
+============
+
+IOASID Set
+-----------
+An IOASID set is a group of IOASIDs allocated from the system-wide
+IOASID pool. An IOASID set is created and can be identified by a
+token of u64. Refer to IOASID set APIs for more details.  

Identified either by an u64 or an mm_struct, right? Maybe just drop the second sentence if it's detailed in the IOASID set section below.

Quoted 13 lines
Quoted 9 lines
Documentation/ioasid.rst new file +2 -0
+
+IOASID set is particularly useful for guest SVA where each guest

could +have its own IOASID set for security and efficiency reasons.

Documentation/ioasid.rst new file +4 -0
+
+IOASID Set Private ID (SPID)
+----------------------------
+SPIDs are introduced as IOASIDs within its set. Each SPID maps to

a +system-wide IOASID but the namespace of SPID is within its IOASID +set.

The intro isn't super clear. Perhaps this is simpler:
"Each IOASID set has a private namespace of SPIDs. An SPID maps to a single system-wide IOASID."

Quoted 2 lines

or, "within an ioasid set, each ioasid can be associated with an alias ID, named SPID."

I don't have strong opinion, I feel it is good to explain the relationship between SPID and IOASID in both directions, how about add? " Conversely, each IOASID is associated with an alias ID, named SPID."

Quoted 33 lines
Quoted 33 lines
Quoted 33 lines

SPIDs can be used as guest IOASIDs where each guest could do

Documentation/ioasid.rst new file +2 -0
+IOASID allocation from its own pool and map them to host physical
+IOASIDs. SPIDs are particularly useful for supporting live

migration +where decoupling guest and host physical resources are necessary. +
+For example, two VMs can both allocate guest PASID/SPID #101 but map to +different host PASIDs #201 and #202 respectively as shown in the +diagram below.
+::

Documentation/ioasid.rst new file +18 -0
+
+ .------------------.    .------------------.
+ |   VM 1           |    |   VM 2           |
+ |                  |    |                  |
+ |------------------|    |------------------|
+ | GPASID/SPID 101  |    | GPASID/SPID 101  |
+ '------------------'    -------------------'     Guest
+ __________|______________________|______________________
+           |                      |               Host
+           v                      v
+ .------------------.    .------------------.
+ | Host IOASID 201  |    | Host IOASID 202  |
+ '------------------'    '------------------'
+ |   IOASID set 1   |    |   IOASID set 2   |
+ '------------------'    '------------------'
+
+Guest PASID is treated as IOASID set private ID (SPID) within an
+IOASID set, mappings between guest and host IOASIDs are stored in

the +set for inquiry.

Documentation/ioasid.rst new file +4 -0
+
+IOASID APIs
+===========
+To get the IOASID APIs, users must #include <linux/ioasid.h>.
Quoted 8 lines
Quoted 7 lines

Documentation/ioasid.rst line 93

Quoted 7 lines

These APIs +serve the following functionalities:

Documentation/ioasid.rst new file +6 -0
+
+  - IOASID allocation/Free
+  - Group management in the form of ioasid_set
+  - Private data storage and lookup
+  - Reference counting
+  - Event notification in case of state change  

(a)

got it

Quoted 7 lines
Quoted 6 lines
Quoted 6 lines
Documentation/ioasid.rst new file +4 -0
+
+IOASID Set Level APIs
+--------------------------
+For use cases such as guest SVA it is necessary to manage IOASIDs

at +a group level. For example, VMs may allocate multiple IOASIDs for

I would use the introduced ioasid_set terminology instead of "group".

Right, we already introduced it.

Quoted 8 lines
Quoted 8 lines

Documentation/ioasid.rst line 100

Quoted 2 lines
Documentation/ioasid.rst new file +2 -0
+guest process address sharing (vSVA). It is imperative to enforce
+VM-IOASID ownership such that malicious guest cannot target DMA  

"a malicious guest"

Quoted 2 lines

+traffic outside its own IOASIDs, or free an active IOASID belong to

Quoted 8 lines
Quoted 7 lines

"that belongs to"

Quoted 5 lines
Documentation/ioasid.rst new file +4 -0
+another VM.
+::
+
+ struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota,

u32 type)

what is this void *token? also the type may be explained here.

token is explained in the text following API list. I can move it up.

Quoted 34 lines
Quoted 34 lines
Quoted 2 lines

+
+ int ioasid_adjust_set(struct ioasid_set *set, int quota);

These could be named "ioasid_set_alloc" and "ioasid_set_adjust" to be consistent with the rest of the API.

Documentation/ioasid.rst line 117

Quoted 10 lines
Documentation/ioasid.rst new file +10 -0
+
+ void ioasid_set_get(struct ioasid_set *set)
+
+ void ioasid_set_put(struct ioasid_set *set)
+
+ void ioasid_set_get_locked(struct ioasid_set *set)
+
+ void ioasid_set_put_locked(struct ioasid_set *set)
+
+ int ioasid_set_for_each_ioasid(struct ioasid_set *sdata,  

Might be nicer to keep the same argument names within the API. Here "set" rather than "sdata".

Documentation/ioasid.rst line 119

Quoted 3 lines

+ void (*fn)(ioasid_t id, void
*data),

Documentation/ioasid.rst new file +1 -0
+				void *data)  

(alignment)

Quoted 4 lines
Documentation/ioasid.rst new file +3 -0
+
+
+IOASID set concept is introduced to represent such IOASID groups.

Each

Or just "IOASID sets represent such IOASID groups", but might be redundant.

Quoted 4 lines
Quoted 2 lines
Quoted 2 lines

+IOASID set is created with a token which can be one of the following +types:

I think this explanation should happen before the above function prototypes

ditto.

Quoted 19 lines
Quoted 10 lines
Quoted 10 lines
Documentation/ioasid.rst new file +5 -0
+
+ - IOASID_SET_TYPE_NULL (Arbitrary u64 value)
+ - IOASID_SET_TYPE_MM (Set token is a mm_struct)
+
+The explicit MM token type is useful when multiple users of an

IOASID +set under the same process need to communicate about their shared IOASIDs. +E.g. An IOASID set created by VFIO for one guest can be associated +with the KVM instance for the same guest since they share a common mm_struct. +
+The IOASID set APIs serve the following purposes:

Quoted 8 lines
Quoted 6 lines

+
+ - Ownership/permission enforcement
+ - Take collective actions, e.g. free an entire set
+ - Event notifications within a set
+ - Look up a set based on token
+ - Quota enforcement

This paragraph could be earlier in the section

Quoted 1 line

yes this is a kind of repetition of (a), above

I meant to highlight on what the APIs do such that readers don't need to read the code instead.

Quoted 25 lines
Quoted 25 lines
Quoted 25 lines
Documentation/ioasid.rst new file +4 -0
+
+Individual IOASID APIs
+----------------------
+Once an ioasid_set is created, IOASIDs can be allocated from the

set. +Within the IOASID set namespace, set private ID (SPID) is supported. In +the VM use case, SPID can be used for storing guest PASID. +
+::
+
+ ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max,

Documentation/ioasid.rst new file +13 -0
+                       void *private);
+
+ int ioasid_get(struct ioasid_set *set, ioasid_t ioasid);
+
+ void ioasid_put(struct ioasid_set *set, ioasid_t ioasid);
+
+ int ioasid_get_locked(struct ioasid_set *set, ioasid_t ioasid);
+
+ void ioasid_put_locked(struct ioasid_set *set, ioasid_t ioasid);
+
+ void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid,
+                   bool (*getter)(void *));
+
Quoted 8 lines
Quoted 8 lines

Documentation/ioasid.rst line 169

Quoted 6 lines

+ ioasid_t ioasid_find_by_spid(struct ioasid_set *set, ioasid_t spid) +
+ int ioasid_attach_data(struct ioasid_set *set, ioasid_t ioasid,

Documentation/ioasid.rst new file +3 -0
+                        void *data);
+ int ioasid_attach_spid(struct ioasid_set *set, ioasid_t ioasid,
+                        ioasid_t ssid);  

s/ssid/spid>

got it

Quoted 15 lines
Quoted 15 lines
Quoted 15 lines
Documentation/ioasid.rst new file +5 -0
+
+
+Notifications
+-------------
+An IOASID may have multiple users, each user may have hardware

context +associated with an IOASID. When the status of an IOASID changes, +e.g. an IOASID is being freed, users need to be notified such that the +associated hardware context can be cleared, flushed, and drained. +
+::

Documentation/ioasid.rst new file +5 -0
+
+ int ioasid_register_notifier(struct ioasid_set *set, struct
+                              notifier_block *nb)
+
+ void ioasid_unregister_notifier(struct ioasid_set *set,
Quoted 8 lines
Quoted 7 lines

Documentation/ioasid.rst line 191

Quoted 7 lines
Documentation/ioasid.rst new file +7 -0
+                                 struct notifier_block *nb)
+
+ int ioasid_register_notifier_mm(struct mm_struct *mm, struct
+                                 notifier_block *nb)
+
+ void ioasid_unregister_notifier_mm(struct mm_struct *mm, struct
+                                    notifier_block *nb)  

the mm_struct prototypes may be justified

This is the mm type token, i.e.
- IOASID_SET_TYPE_MM (Set token is a mm_struct)
I am not sure if it is better to keep the explanation in code or in this document, certainly don't want to duplicate.

Quoted 4 lines
Quoted 3 lines

Documentation/ioasid.rst line 194

Quoted 3 lines
Documentation/ioasid.rst new file +3 -0
+
+ int ioasid_notify(ioasid_t ioasid, enum ioasid_notify_val cmd,
+                   unsigned int flags)  

this one is not obvious either.

Here I just wanted to list the API functions, perhaps readers can check out the code comments?

Quoted 22 lines
Quoted 22 lines
Quoted 22 lines
Documentation/ioasid.rst new file +5 -0
+
+
+Events
+~~~~~~
+Notification events are pertinent to individual IOASIDs, they can

be +one of the following:

Documentation/ioasid.rst new file +16 -0
+
+ - ALLOC
+ - FREE
+ - BIND
+ - UNBIND
+
+Ordering
+~~~~~~~~
+Ordering is supported by IOASID notification priorities as the
+following (in ascending order):
+
+::
+
+ enum ioasid_notifier_prios {
+	IOASID_PRIO_LAST,
+	IOASID_PRIO_IOMMU,
Quoted 8 lines
Quoted 3 lines

Documentation/ioasid.rst line 219

Quoted 3 lines
Documentation/ioasid.rst new file +3 -0
+	IOASID_PRIO_DEVICE,
+	IOASID_PRIO_CPU,
+ };  

Maybe:
when registered, notifiers are assigned a priority that affect the call order. Notifiers with CPU priority get called before notifiers with device priority and so on.

Sounds good.

Quoted 62 lines
Quoted 62 lines
Quoted 24 lines
Documentation/ioasid.rst new file +2 -0
+
+The typical use case is when an IOASID is freed due to an

exception, DMA +source should be quiesced before tearing down other hardware contexts +in the system. This will reduce the churn in handling faults. DMA work +submission is performed by the CPU which is granted higher priority than +devices.

Documentation/ioasid.rst new file +11 -0
+
+
+Scopes
+~~~~~~
+There are two types of notifiers in IOASID core: system-wide and
+ioasid_set-wide.
+
+System-wide notifier is catering for users that need to handle all
+IOASIDs in the system. E.g. The IOMMU driver handles all IOASIDs.
+
+Per ioasid_set notifier can be used by VM specific components

such as +KVM. After all, each KVM instance only cares about IOASIDs within its +own set.

Documentation/ioasid.rst new file +5 -0
+
+
+Atomicity
+~~~~~~~~~
+IOASID notifiers are atomic due to spinlocks used inside the
Quoted 2 lines

IOASID +core. For tasks cannot be completed in the notifier handler, async work

"tasks that cannot be"

Quoted 7 lines
Documentation/ioasid.rst new file +6 -0
+can be submitted to complete the work later as long as there is no
+ordering requirement.
+
+Reference counting
+------------------
+IOASID lifecycle management is based on reference counting. Users

of +IOASID intend to align lifecycle with the IOASID need to hold

"who intend to"

Quoted 2 lines

+reference of the IOASID. IOASID will not be returned to the pool for

"a reference to the IOASID. The IOASID"

Quoted 18 lines
Documentation/ioasid.rst new file +6 -0
+allocation until all references are dropped. Calling ioasid_free()
+will mark the IOASID as FREE_PENDING if the IOASID has outstanding
+reference. ioasid_get() is not allowed once an IOASID is in the
+FREE_PENDING state.
+
+Event notifications are used to inform users of IOASID status

change. +IOASID_FREE event prompts users to drop their references after +clearing its context.

Documentation/ioasid.rst new file +10 -0
+
+For example, on VT-d platform when an IOASID is freed, teardown
+actions are performed on KVM, device driver, and IOMMU driver.
+KVM shall register notifier block with::
+
+ static struct notifier_block pasid_nb_kvm = {
+	.notifier_call = pasid_status_change_kvm,
+	.priority      = IOASID_PRIO_CPU,
+ };
+
Quoted 8 lines
Quoted 6 lines

Documentation/ioasid.rst line 276

Quoted 6 lines
Documentation/ioasid.rst new file +6 -0
+VDCM driver shall register notifier block with::
+
+ static struct notifier_block pasid_nb_vdcm = {
+	.notifier_call = pasid_status_change_vdcm,
+	.priority      = IOASID_PRIO_DEVICE,
+ };  

not sure those code snippets are really useful. Maybe simply say who is supposed to use each prio.

Agreed, not all the bits in the snippets are explained. I will explain KVM and VDCM need to use priority to ensure call order.

Quoted 8 lines
Quoted 7 lines
Quoted 7 lines
Documentation/ioasid.rst new file +2 -0
+
+In both cases, notifier blocks shall be registered on the IOASID

set +such that *only* events from the matching VM is received.

Documentation/ioasid.rst new file +2 -0
+
+If KVM attempts to register notifier block before the IOASID set

is +created for the MM token, the notifier block will be placed on a

using the MM token

sounds good

Quoted 57 lines
Quoted 3 lines
Quoted 3 lines

+pending list inside IOASID core. Once the token matching IOASID set +is created, IOASID will register the notifier block automatically.

Is this implementation mandated? Can't you enforce the ioasid_set to be created before the notifier gets registered?

Quoted 52 lines
Quoted 34 lines
Documentation/ioasid.rst new file +2 -0
+IOASID core does not replay events for the existing IOASIDs in the
+set. For IOASID set of MM type, notification blocks can be

registered +on empty sets only. This is to avoid lost events.

Documentation/ioasid.rst new file +14 -0
+
+IOMMU driver shall register notifier block on global chain::
+
+ static struct notifier_block pasid_nb_vtd = {
+	.notifier_call = pasid_status_change_vtd,
+	.priority      = IOASID_PRIO_IOMMU,
+ };
+
+Custom allocator APIs
+---------------------
+
+::
+
+ int ioasid_register_allocator(struct ioasid_allocator_ops

*allocator); +
+ void ioasid_unregister_allocator(struct ioasid_allocator_ops *allocator); +

Documentation/ioasid.rst new file +3 -0
+Allocator Choices
+~~~~~~~~~~~~~~~~~
+IOASIDs are allocated for both host and guest SVA/IOVA usage.

However, +allocators can be different. For example, on VT-d guest PASID +allocation must be performed via a virtual command interface which is +emulated by VMM.

Documentation/ioasid.rst new file +2 -0
+
+IOASID core has the notion of "custom allocator" such that guest

can +register virtual command allocator that precedes the default one. +

Documentation/ioasid.rst new file +4 -0
+Namespaces
+~~~~~~~~~~
+IOASIDs are limited system resources that default to 20 bits in
+size. Since each device has its own table, theoretically the
Quoted 3 lines

namespace +can be per device also. However, for security reasons sharing PASID +tables among devices are not good for isolation. Therefore, IOASID +namespace is system-wide.

I don't follow this development. Having per-device PASID table would work fine for isolation (assuming no hardware bug necessitating IOMMU groups). If I remember correctly IOASID space was chosen to be OS-wide because it simplifies the management code (single PASID per task), and it is system-wide across VMs only in the case of VT-d scalable mode.

Documentation/ioasid.rst line 325

Quoted 4 lines
Documentation/ioasid.rst new file +4 -0
+
+There are also other reasons to have this simpler system-wide
+namespace. Take VT-d as an example, VT-d supports shared workqueue
+and ENQCMD[1] where one IOASID could be used to submit work on  

Maybe use the Sphinx glossary syntax rather than "[1]"
https://www.sphinx-doc.org/en/master/usage/restructuredtext/directives.html#glossary-directive

Quoted 13 lines
Quoted 11 lines
Quoted 11 lines

+multiple devices that are shared with other VMs. This requires IOASID +to be system-wide. This is also the reason why guests must use an +emulated virtual command interface to allocate IOASID from the host. +

Documentation/ioasid.rst new file +4 -0
+
+Life cycle
+==========
+This section covers IOASID lifecycle management for both

bare-metal +and guest usages. In bare-metal SVA, MMU notifier is directly hooked +up with IOMMU driver, therefore the process address space (MM) +lifecycle is aligned with IOASID.

therefore the IOASID lifecyle matches the process address space (MM) lifecyle?

Sounds good.

Quoted 57 lines
Quoted 3 lines
Quoted 3 lines
Documentation/ioasid.rst new file +2 -0
+
+However, guest MMU notifier is not available to host IOMMU

driver,

the guest MMU notifier

Quoted 2 lines
Quoted 2 lines

+when guest MM terminates unexpectedly, the events have to go through

the guest MM

Quoted 42 lines
Quoted 23 lines

+VFIO and IOMMU UAPI to reach host IOMMU driver. There are also more +parties involved in guest SVA, e.g. on Intel VT-d platform, IOASIDs +are used by IOMMU driver, KVM, VDCM, and VFIO.

Documentation/ioasid.rst new file +19 -0
+
+Native IOASID Life Cycle (VT-d Example)
+---------------------------------------
+
+The normal flow of native SVA code with Intel Data Streaming
+Accelerator(DSA) [2] as example:
+
+1. Host user opens accelerator FD, e.g. DSA driver, or uacce;
+2. DSA driver allocate WQ, do sva_bind_device();
+3. IOMMU driver calls ioasid_alloc(), then bind PASID with device,
+   mmu_notifier_get()
+4. DMA starts by DSA driver userspace
+5. DSA userspace close FD
+6. DSA/uacce kernel driver handles FD.close()
+7. DSA driver stops DMA
+8. DSA driver calls sva_unbind_device();
+9. IOMMU driver does unbind, clears PASID context in IOMMU, flush
+   TLBs. mmu_notifier_put() called.
+10. mmu_notifier.release() called, IOMMU SVA code calls

ioasid_free()* +11. The IOASID is returned to the pool, reclaimed.

Documentation/ioasid.rst line 64

Quoted 3 lines
Documentation/ioasid.rst new file +3 -0
+
+::
+  

Use a footnote?
https://www.sphinx-doc.org/en/master/usage/restructuredtext/basics.html#footnotes

Quoted 3 lines

+ * With ENQCMD, PASID used on VT-d is not released in
mmu_notifier() but
+ mmdrop(). mmdrop comes after FD close. Should not matter.

"comes after FD close, which doesn't make a difference?" The following might not be necessary since early process termination is described later.

Quoted 4 lines

+ If the user process dies unexpectedly, Step #10 may come
before
+ Step #5, in between, all DMA faults discarded. PRQ responded
with

Quoted 6 lines

PRQ hasn't been defined in this document.

Quoted 4 lines
Documentation/ioasid.rst new file +3 -0
+     code INVALID REQUEST.
+
+During the normal teardown, the following three steps would

happen in +order:

can't this be illustrated in the above 1-11 sequence, just adding NORMAL TEARDONW before #7?

Quoted 8 lines
Quoted 7 lines

Documentation/ioasid.rst line 379

Quoted 7 lines

+
+1. Device driver stops DMA request
+2. IOMMU driver unbinds PASID and mm, flush all TLBs, drain in-flight

Documentation/ioasid.rst new file +3 -0
+   requests.
+3. IOASID freed
+  

Then you can just focus on abnormal termination

Yes, will refer to the steps starting #7. These can be removed.

Quoted 5 lines
Quoted 3 lines
Quoted 3 lines

+Exception happens when process terminates *before* device driver stops +DMA and call IOMMU driver to unbind. The flow of process exists are as

Can't this be explained with something simpler looking at the steps 1-11?

It meant to be educational given this level of details. Simpler steps are labeled with (1) (2) (3). Perhaps these labels didn't stand out right? I will use the steps in the 1-11 sequence.

Quoted 48 lines
Quoted 48 lines

"exits"

Quoted 13 lines
Documentation/ioasid.rst new file +13 -0
+follows:
+
+::
+
+   do_exit() {
+	exit_mm() {
+		mm_put();
+		exit_mmap() {
+			intel_invalidate_range() //mmu notifier
+			tlb_finish_mmu()
+			mmu_notifier_release(mm) {
+				intel_iommu_release() {
+   [2]
Quoted 1 line

intel_iommu_teardown_pasid();

Parentheses might be better than square brackets for step numbers

Quoted 28 lines
Documentation/ioasid.rst new file +6 -0
+                                        intel_iommu_flush_tlbs();
+				}
+				// tlb_invalidate_range cb removed
+			}
+			unmap_vmas();
+                        free_pgtables(); // IOMMU cannot walk PGT

after this

Documentation/ioasid.rst new file +21 -0
+		};
+	}
+	exit_files(tsk) {
+		close_files() {
+			dsa_close();
+   [1]			dsa_stop_dma();
+                        intel_svm_unbind_pasid(); //nothing to do
+		}
+	}
+   }
+
+   mmdrop() /* some random time later, lazy mm user */ {
+   	mm_free_pgd();
+        destroy_context(mm); {
+   [3]	        ioasid_free();
+	}
+   }
+
+As shown in the list above, step #2 could happen before
+#1. Unrecoverable(UR) faults could happen between #2 and #1.
+
Quoted 8 lines
Quoted 7 lines
Quoted 7 lines
Documentation/ioasid.rst new file +2 -0
+Also notice that TLB invalidation occurs at mmu_notifier
+invalidate_range callback as well as the release callback. The

reason +is that release callback will delete IOMMU driver from the notifier +chain which may skip invalidate_range() calls during the exit path. +
+To avoid unnecessary reporting of UR fault, IOMMU driver shall disable

UR?

Unrecoverable, mentioned in the previous paragraph.

Quoted 106 lines
Quoted 63 lines
Quoted 63 lines
Documentation/ioasid.rst new file +5 -0
+fault reporting after free and before unbind.
+
+Guest IOASID Life Cycle (VT-d Example)
+--------------------------------------
+Guest IOASID life cycle starts with guest driver open(), this

could be +uacce or individual accelerator driver such as DSA. At FD open, +sva_bind_device() is called which triggers a series of actions. +

Documentation/ioasid.rst new file +2 -0
+The example below is an illustration of *normal* operations that
+involves *all* the SW components in VT-d. The flow can be simpler

if +no ENQCMD is supported.

Documentation/ioasid.rst new file +4 -0
+
+::
+
+     VFIO        IOMMU        KVM        VDCM        IOASID

Ref

Documentation/ioasid.rst new file +3 -0
+   ..................................................................
+   1             ioasid_register_notifier/_mm()
+   2 ioasid_alloc()

1 line not shown · raw message

Documentation/ioasid.rst new file +2 -0
+   3 bind_gpasid()
+   4             iommu_bind()->ioasid_get()

1 line not shown · raw message

Documentation/ioasid.rst new file +2 -0
+   5             ioasid_notify(BIND)
+   6                          -> ioasid_get()

1 line not shown · raw message

Documentation/ioasid.rst new file +4 -0
+   7                          -> vmcs_update_atomic()
+   8 mdev_write(gpasid)
[… more lines in the raw message]

Lines of this file are not shown. View the raw message.

The rest of this diff is not shown. View the raw message.

1 line not shown · raw message

Documentation/ioasid.rst new file +6 -0

Lines of this file are not shown. View the raw message.

The rest of this diff is not shown. View the raw message.

1 line not shown · raw message

Documentation/ioasid.rst new file +5 -0

Lines of this file are not shown. View the raw message.

The rest of this diff is not shown. View the raw message.

5 lines not shown · raw message

Documentation/ioasid.rst new file +7 -2

Lines of this file are not shown. View the raw message.

The rest of this diff is not shown. View the raw message.

1 line not shown · raw message

Documentation/ioasid.rst new file +4 -0

Lines of this file are not shown. View the raw message.

The rest of this diff is not shown. View the raw message.

1 line not shown · raw message

Quoted 8 lines

8 lines not shown · raw message

35 lines not shown · raw message

Quoted 7 lines

7 lines not shown · raw message

137 lines not shown · raw message

Parts of this message are not shown here. View the raw message.

#46 lore raw

Re: [PATCH v2 6/9] iommu/ioasid: Introduce notification APIs

In reply to Jacob Pan

Hi Jacob,

Quoted 25 lines · Jacob Pan

Relations among IOASID users largely follow a publisher-subscriber pattern. E.g. to support guest SVA on Intel Scalable I/O Virtualization (SIOV) enabled platforms, VFIO, IOMMU, device drivers, KVM are all users of IOASIDs. When a state change occurs, VFIO publishes the change event that needs to be processed by other users/subscribers.

This patch introduced two types of notifications: global and per ioasid_set. The latter is intended for users who only needs to handle events related to the IOASID of a given set.
For more information, refer to the kernel documentation at Documentation/ioasid.rst.

---
 drivers/iommu/ioasid.c | 280 ++++++++++++++++++++++++++++++++++++++++++++++++-
 include/linux/ioasid.h |  70 +++++++++++++
 2 files changed, 348 insertions(+), 2 deletions(-)
drivers/iommu/ioasid.c +0 -0
index c0aef38a4fde..6ddc09a7fe74 100644

drivers/iommu/ioasid.c line 16

Quoted 8 lines
drivers/iommu/ioasid.c +3 -0
@@ -9,8 +9,35 @@
 #include <linux/spinlock.h>
 #include <linux/xarray.h>
 #include <linux/ioasid.h>
+#include <linux/sched/mm.h>
 
 static DEFINE_XARRAY_ALLOC(ioasid_sets);
+/*
+ * An IOASID could have multiple consumers where each consumeer may have

can have multiple consumers

drivers/iommu/ioasid.c line 18

Quoted 2 lines
drivers/iommu/ioasid.c +2 -0
+ * hardware contexts associated with IOASIDs.
+ * When a status change occurs, such as IOASID is being freed, notifier chains

s/such as IOASID is being freed/, like on IOASID deallocation,

Quoted 36 lines
drivers/iommu/ioasid.c +23 -0
+ * are used to keep the consumers in sync.
+ * This is a publisher-subscriber pattern where publisher can change the
+ * state of each IOASID, e.g. alloc/free, bind IOASID to a device and mm.
+ * On the other hand, subscribers gets notified for the state change and
+ * keep local states in sync.
+ *
+ * Currently, the notifier is global. A further optimization could be per
+ * IOASID set notifier chain.
+ */
+static ATOMIC_NOTIFIER_HEAD(ioasid_chain);
+
+/* List to hold pending notification block registrations */
+static LIST_HEAD(ioasid_nb_pending_list);
+static DEFINE_SPINLOCK(ioasid_nb_lock);
+struct ioasid_set_nb {
+	struct list_head	list;
+	struct notifier_block	*nb;
+	void			*token;
+	struct ioasid_set	*set;
+	bool			active;
+};
+
 enum ioasid_state {
 	IOASID_STATE_INACTIVE,
 	IOASID_STATE_ACTIVE,
@@ -394,6 +421,7 @@ EXPORT_SYMBOL_GPL(ioasid_find_by_spid);
 ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max,
 		      void *private)
 {
+	struct ioasid_nb_args args;
 	struct ioasid_data *data;
 	void *adata;
 	ioasid_t id = INVALID_IOASID;
@@ -445,8 +473,14 @@ ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max,
 		goto exit_free;
 	}

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 483

Quoted 10 lines
drivers/iommu/ioasid.c +7 -1
 	set->nr_ioasids++;
-	goto done_unlock;
+	args.id = id;
+	/* Set private ID is not attached during allocation */
+	args.spid = INVALID_IOASID;
+	args.set = set;
+	atomic_notifier_call_chain(&set->nh, IOASID_ALLOC, &args);
 
+	spin_unlock(&ioasid_allocator_lock);
+	return id;

spurious change

Quoted 12 lines
drivers/iommu/ioasid.c +1 -0
 exit_free:
 	kfree(data);
 done_unlock:
@@ -479,6 +513,7 @@ static void ioasid_do_free(struct ioasid_data *data)
 
 static void ioasid_free_locked(struct ioasid_set *set, ioasid_t ioasid)
 {
+	struct ioasid_nb_args args;
 	struct ioasid_data *data;
 
 	data = xa_load(&active_allocator->xa, ioasid);

drivers/iommu/ioasid.c static void ioasid_free_locked(struct ioasid_set *set, ioasid_t ioasid) · line 529

Quoted 4 lines
drivers/iommu/ioasid.c +1 -0
@@ -491,7 +526,16 @@ static void ioasid_free_locked(struct ioasid_set *set, ioasid_t ioasid)
 		pr_warn("Cannot free IOASID %u due to set ownership\n", ioasid);
 		return;
 	}
+

spurious new line

drivers/iommu/ioasid.c static void ioasid_free_locked(struct ioasid_set *set, ioasid_t ioasid) · line 541

Quoted 12 lines
drivers/iommu/ioasid.c +8 -0
 	data->state = IOASID_STATE_FREE_PENDING;
+	/* Notify all users that this IOASID is being freed */
+	args.id = ioasid;
+	args.spid = data->spid;
+	args.pdata = data->private;
+	args.set = data->set;
+	atomic_notifier_call_chain(&ioasid_chain, IOASID_FREE, &args);
+	/* Notify the ioasid_set for per set users */
+	atomic_notifier_call_chain(&set->nh, IOASID_FREE, &args);
 
 	if (!refcount_dec_and_test(&data->users))
 		return;

Shouldn't we call the notifier only when ref count == 0?

Quoted 130 lines
drivers/iommu/ioasid.c +89 -1
@@ -514,6 +558,28 @@ void ioasid_free(struct ioasid_set *set, ioasid_t ioasid)
 }
 EXPORT_SYMBOL_GPL(ioasid_free);
 
+static void ioasid_add_pending_nb(struct ioasid_set *set)
+{
+	struct ioasid_set_nb *curr;
+
+	if (set->type != IOASID_SET_TYPE_MM)
+		return;
+
+	/*
+	 * Check if there are any pending nb requests for the given token, if so
+	 * add them to the notifier chain.
+	 */
+	spin_lock(&ioasid_nb_lock);
+	list_for_each_entry(curr, &ioasid_nb_pending_list, list) {
+		if (curr->token == set->token && !curr->active) {
+			atomic_notifier_chain_register(&set->nh, curr->nb);
+			curr->set = set;
+			curr->active = true;
+		}
+	}
+	spin_unlock(&ioasid_nb_lock);
+}
+
 /**
  * ioasid_alloc_set - Allocate a new IOASID set for a given token
  *
@@ -601,6 +667,13 @@ struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota, int type)
 	sdata->quota = quota;
 	sdata->sid = id;
 	refcount_set(&sdata->ref, 1);
+	ATOMIC_INIT_NOTIFIER_HEAD(&sdata->nh);
+
+	/*
+	 * Check if there are any pending nb requests for the given token, if so
+	 * add them to the notifier chain.
+	 */
+	ioasid_add_pending_nb(sdata);
 
 	/*
 	 * Per set XA is used to store private IDs within the set, get ready
@@ -617,6 +690,30 @@ struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota, int type)
 }
 EXPORT_SYMBOL_GPL(ioasid_alloc_set);
 
+
+/*
+ * ioasid_find_mm_set - Retrieve IOASID set with mm token
+ * Take a reference of the set if found.
+ */
+static struct ioasid_set *ioasid_find_mm_set(struct mm_struct *token)
+{
+	struct ioasid_set *sdata, *set = NULL;
+	unsigned long index;
+
+	spin_lock(&ioasid_allocator_lock);
+
+	xa_for_each(&ioasid_sets, index, sdata) {
+		if (sdata->type == IOASID_SET_TYPE_MM && sdata->token == token) {
+			refcount_inc(&sdata->ref);
+			set = sdata;
+			goto exit_unlock;
+		}
+	}
+exit_unlock:
+	spin_unlock(&ioasid_allocator_lock);
+	return set;
+}
+
 void ioasid_set_get_locked(struct ioasid_set *set)
 {
 	if (WARN_ON(xa_load(&ioasid_sets, set->sid) != set)) {
@@ -638,6 +735,8 @@ EXPORT_SYMBOL_GPL(ioasid_set_get);
 
 void ioasid_set_put_locked(struct ioasid_set *set)
 {
+	struct ioasid_nb_args args = { 0 };
+	struct ioasid_set_nb *curr;
 	struct ioasid_data *entry;
 	unsigned long index;
 
@@ -673,8 +772,24 @@ void ioasid_set_put_locked(struct ioasid_set *set)
 done_destroy:
 	/* Return the quota back to system pool */
 	ioasid_capacity_avail += set->quota;
-	kfree_rcu(set, rcu);
 
+	/* Restore pending status of the set NBs */
+	spin_lock(&ioasid_nb_lock);
+	list_for_each_entry(curr, &ioasid_nb_pending_list, list) {
+		if (curr->token == set->token) {
+			if (curr->active)
+				curr->active = false;
+			else
+				pr_warn("Set token exists but not active!\n");
+		}
+	}
+	spin_unlock(&ioasid_nb_lock);
+
+	args.set = set;
+	atomic_notifier_call_chain(&ioasid_chain, IOASID_SET_FREE, &args);
+
+	kfree_rcu(set, rcu);
+	pr_debug("Set freed %d\n", set->sid);
 	/*
 	 * Token got released right away after the ioasid_set is freed.
 	 * If a new set is created immediately with the newly released token,
@@ -927,6 +1042,167 @@ void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid,
 }
 EXPORT_SYMBOL_GPL(ioasid_find);
 
+int ioasid_register_notifier(struct ioasid_set *set, struct notifier_block *nb)
+{
+	if (set)
+		return atomic_notifier_chain_register(&set->nh, nb);
+	else
+		return atomic_notifier_chain_register(&ioasid_chain, nb);
+}
+EXPORT_SYMBOL_GPL(ioasid_register_notifier);
+
+void ioasid_unregister_notifier(struct ioasid_set *set,
+				struct notifier_block *nb)
+{
+	struct ioasid_set_nb *curr;
+
+	spin_lock(&ioasid_nb_lock);
+	/*
+	 * Pending list is registered with a token without an ioasid_set,

drivers/iommu/ioasid.c void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid, · line 1071

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+	 * therefore should not be unregistered directly.
+	 */
+	list_for_each_entry(curr, &ioasid_nb_pending_list, list) {
+		if (curr->nb == nb) {
+			pr_warn("Cannot unregister NB from pending list\n");
+			spin_unlock(&ioasid_nb_lock);
+			return;
+		}
+	}
+	spin_unlock(&ioasid_nb_lock);

is it safe to release the lock here? What does prevent another NB to be added to ioasid_nb_pending_list after that?

Quoted 142 lines
drivers/iommu/ioasid.c +134 -0
+
+	if (set)
+		atomic_notifier_chain_unregister(&set->nh, nb);
+	else
+		atomic_notifier_chain_unregister(&ioasid_chain, nb);
+}
+EXPORT_SYMBOL_GPL(ioasid_unregister_notifier);
+
+int ioasid_register_notifier_mm(struct mm_struct *mm, struct notifier_block *nb)
+{
+	struct ioasid_set_nb *curr;
+	struct ioasid_set *set;
+	int ret = 0;
+
+	if (!mm)
+		return -EINVAL;
+
+	spin_lock(&ioasid_nb_lock);
+
+	/* Check for duplicates, nb is unique per set */
+	list_for_each_entry(curr, &ioasid_nb_pending_list, list) {
+		if (curr->token == mm && curr->nb == nb) {
+			ret = -EBUSY;
+			goto exit_unlock;
+		}
+	}
+
+	/* Check if the token has an existing set */
+	set = ioasid_find_mm_set(mm);
+	if (IS_ERR_OR_NULL(set)) {
+		/* Add to the rsvd list as inactive */
+		curr->active = false;
+	} else {
+		/* REVISIT: Only register empty set for now. Can add an option
+		 * in the future to playback existing PASIDs.
+		 */
+		if (set->nr_ioasids) {
+			pr_warn("IOASID set %d not empty\n", set->sid);
+			ret = -EBUSY;
+			goto exit_unlock;
+		}
+		curr = kzalloc(sizeof(*curr), GFP_ATOMIC);
+		if (!curr) {
+			ret = -ENOMEM;
+			goto exit_unlock;
+		}
+		curr->token = mm;
+		curr->nb = nb;
+		curr->active = true;
+		curr->set = set;
+
+		/* Set already created, add to the notifier chain */
+		atomic_notifier_chain_register(&set->nh, nb);
+		/*
+		 * Do not hold a reference, if the set gets destroyed, the nb
+		 * entry will be marked inactive.
+		 */
+		ioasid_set_put(set);
+	}
+
+	list_add(&curr->list, &ioasid_nb_pending_list);
+
+exit_unlock:
+	spin_unlock(&ioasid_nb_lock);
+	return ret;
+}
+EXPORT_SYMBOL_GPL(ioasid_register_notifier_mm);
+
+void ioasid_unregister_notifier_mm(struct mm_struct *mm, struct notifier_block *nb)
+{
+	struct ioasid_set_nb *curr;
+
+	spin_lock(&ioasid_nb_lock);
+	list_for_each_entry(curr, &ioasid_nb_pending_list, list) {
+		if (curr->token == mm && curr->nb == nb) {
+			list_del(&curr->list);
+			goto exit_free;
+		}
+	}
+	pr_warn("No ioasid set found for mm token %llx\n",  (u64)mm);
+	goto done_unlock;
+
+exit_free:
+	if (curr->active) {
+		pr_debug("mm set active, unregister %llx\n",
+			(u64)mm);
+		atomic_notifier_chain_unregister(&curr->set->nh, nb);
+	}
+	kfree(curr);
+done_unlock:
+	spin_unlock(&ioasid_nb_lock);
+	return;
+}
+EXPORT_SYMBOL_GPL(ioasid_unregister_notifier_mm);
+
+/**
+ * ioasid_notify - Send notification on a given IOASID for status change.
+ *                 Used by publishers when the status change may affect
+ *                 subscriber's internal state.
+ *
+ * @ioasid:	The IOASID to which the notification will send
+ * @cmd:	The notification event
+ * @flags:	Special instructions, e.g. notify with a set or global
+ */
+int ioasid_notify(ioasid_t ioasid, enum ioasid_notify_val cmd, unsigned int flags)
+{
+	struct ioasid_data *ioasid_data;
+	struct ioasid_nb_args args = { 0 };
+	int ret = 0;
+
+	spin_lock(&ioasid_allocator_lock);
+	ioasid_data = xa_load(&active_allocator->xa, ioasid);
+	if (!ioasid_data) {
+		pr_err("Trying to notify unknown IOASID %u\n", ioasid);
+		spin_unlock(&ioasid_allocator_lock);
+		return -EINVAL;
+	}
+
+	args.id = ioasid;
+	args.set = ioasid_data->set;
+	args.pdata = ioasid_data->private;
+	args.spid = ioasid_data->spid;
+	if (flags & IOASID_NOTIFY_ALL) {
+		ret = atomic_notifier_call_chain(&ioasid_chain, cmd, &args);
+	} else if (flags & IOASID_NOTIFY_SET) {
+		ret = atomic_notifier_call_chain(&ioasid_data->set->nh,
+						 cmd, &args);
+	}
+	spin_unlock(&ioasid_allocator_lock);
+
+	return ret;
+}
+EXPORT_SYMBOL_GPL(ioasid_notify);
+
 MODULE_AUTHOR("Jean-Philippe Brucker <[email protected]>");
 MODULE_AUTHOR("Jacob Pan <[email protected]>");
 MODULE_DESCRIPTION("IO Address Space ID (IOASID) allocator");
include/linux/ioasid.h +0 -0
index d4b3e83672f6..572111cd3b4b 100644

include/linux/ioasid.h enum ioasid_set_type { · line 26

Quoted 4 lines
include/linux/ioasid.h +1 -0
@@ -23,6 +23,7 @@ enum ioasid_set_type {
  * struct ioasid_set - Meta data about ioasid_set
  * @type:	Token types and other features
  * @token:	Unique to identify an IOASID set
+ * @nh:		Notifier for IOASID events within the set

list of notifiers private to that set?

Quoted 100 lines
include/linux/ioasid.h +62 -0
  * @xa:		XArray to store ioasid_set private IDs, can be used for
  *		guest-host IOASID mapping, or just a private IOASID namespace.
  * @quota:	Max number of IOASIDs can be allocated within the set
@@ -32,6 +33,7 @@ enum ioasid_set_type {
  */
 struct ioasid_set {
 	void *token;
+	struct atomic_notifier_head nh;
 	struct xarray xa;
 	int type;
 	int quota;
@@ -56,6 +58,49 @@ struct ioasid_allocator_ops {
 	void *pdata;
 };
 
+/* Notification data when IOASID status changed */
+enum ioasid_notify_val {
+	IOASID_ALLOC = 1,
+	IOASID_FREE,
+	IOASID_BIND,
+	IOASID_UNBIND,
+	IOASID_SET_ALLOC,
+	IOASID_SET_FREE,
+};
+
+#define IOASID_NOTIFY_ALL BIT(0)
+#define IOASID_NOTIFY_SET BIT(1)
+/**
+ * enum ioasid_notifier_prios - IOASID event notification order
+ *
+ * When status of an IOASID changes, users might need to take actions to
+ * reflect the new state. For example, when an IOASID is freed due to
+ * exception, the hardware context in virtual CPU, DMA device, and IOMMU
+ * shall be cleared and drained. Order is required to prevent life cycle
+ * problems.
+ */
+enum ioasid_notifier_prios {
+	IOASID_PRIO_LAST,
+	IOASID_PRIO_DEVICE,
+	IOASID_PRIO_IOMMU,
+	IOASID_PRIO_CPU,
+};
+
+/**
+ * struct ioasid_nb_args - Argument provided by IOASID core when notifier
+ * is called.
+ * @id:		The IOASID being notified
+ * @spid:	The set private ID associated with the IOASID
+ * @set:	The IOASID set of @id
+ * @pdata:	The private data attached to the IOASID
+ */
+struct ioasid_nb_args {
+	ioasid_t id;
+	ioasid_t spid;
+	struct ioasid_set *set;
+	void *pdata;
+};
+
 #if IS_ENABLED(CONFIG_IOASID)
 void ioasid_install_capacity(ioasid_t total);
 ioasid_t ioasid_get_capacity(void);
@@ -75,8 +120,16 @@ void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid, bool (*getter)(void *
 int ioasid_attach_data(ioasid_t ioasid, void *data);
 int ioasid_attach_spid(ioasid_t ioasid, ioasid_t spid);
 ioasid_t ioasid_find_by_spid(struct ioasid_set *set, ioasid_t spid);
+
+int ioasid_register_notifier(struct ioasid_set *set,
+			struct notifier_block *nb);
+void ioasid_unregister_notifier(struct ioasid_set *set,
+				struct notifier_block *nb);
+
 int ioasid_register_allocator(struct ioasid_allocator_ops *allocator);
 void ioasid_unregister_allocator(struct ioasid_allocator_ops *allocator);
+
+int ioasid_notify(ioasid_t ioasid, enum ioasid_notify_val cmd, unsigned int flags);
 void ioasid_is_in_set(struct ioasid_set *set, ioasid_t ioasid);
 int ioasid_get(struct ioasid_set *set, ioasid_t ioasid);
 int ioasid_get_locked(struct ioasid_set *set, ioasid_t ioasid);
@@ -85,6 +138,9 @@ void ioasid_put_locked(struct ioasid_set *set, ioasid_t ioasid);
 int ioasid_set_for_each_ioasid(struct ioasid_set *sdata,
 			       void (*fn)(ioasid_t id, void *data),
 			       void *data);
+int ioasid_register_notifier_mm(struct mm_struct *mm, struct notifier_block *nb);
+void ioasid_unregister_notifier_mm(struct mm_struct *mm, struct notifier_block *nb);
+
 #else /* !CONFIG_IOASID */
 static inline void ioasid_install_capacity(ioasid_t total)
 {
@@ -124,6 +180,20 @@ static inline void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid, bool (*
 	return NULL;
 }
 
+static inline int ioasid_register_notifier(struct notifier_block *nb)
+{
+	return -ENOTSUPP;
+}
+
+static inline void ioasid_unregister_notifier(struct notifier_block *nb)
+{

include/linux/ioasid.h static inline void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid, bool (* · line 199

Quoted 10 lines
include/linux/ioasid.h +7 -0
+}
+
+static inline int ioasid_notify(ioasid_t ioasid, enum ioasid_notify_val cmd, unsigned int flags)
+{
+	return -ENOTSUPP;
+}
+
 static inline int ioasid_register_allocator(struct ioasid_allocator_ops *allocator)
 {
 	return -ENOTSUPP;

Thanks

Eric

#50 lore raw

Re: [PATCH v2 7/9] iommu/vt-d: Listen to IOASID notifications

In reply to Jacob Pan

Hi Jacob,

Quoted 19 lines · Jacob Pan

On Intel Scalable I/O Virtualization (SIOV) enabled platforms, IOMMU driver is one of the users of IOASIDs. In normal flow, callers will perform IOASID allocation, bind, unbind, and free in order. However, for guest SVA, IOASID free could come before unbind as guest is untrusted. This patch registers IOASID notification handler such that IOMMU driver can perform PASID teardown upon receiving an unexpected IOASID free event.

---
 drivers/iommu/intel/svm.c   | 74 ++++++++++++++++++++++++++++++++++++++++++++-
 include/linux/intel-iommu.h |  2 ++
 2 files changed, 75 insertions(+), 1 deletion(-)
drivers/iommu/intel/svm.c +0 -0
index 634e191ca2c3..600e3ae5b656 100644

drivers/iommu/intel/svm.c static inline bool intel_svm_capable(struct intel_iommu *iommu) · line 98

Quoted 4 lines
drivers/iommu/intel/svm.c +1 -0
@@ -95,6 +95,72 @@ static inline bool intel_svm_capable(struct intel_iommu *iommu)
 	return iommu->flags & VTD_FLAG_SVM_CAPABLE;
 }
 
+#define pasid_lock_held() lock_is_held(&pasid_mutex.dep_map)

put after the pasid_mutex definition?

Quoted 12 lines
drivers/iommu/intel/svm.c +12 -0
+static DEFINE_MUTEX(pasid_mutex);
+
+static void intel_svm_free_async_fn(struct work_struct *work)
+{
+	struct intel_svm *svm = container_of(work, struct intel_svm, work);
+	struct intel_svm_dev *sdev;
+
+	/*
+	 * Unbind all devices associated with this PASID which is
+	 * being freed by other users such as VFIO.
+	 */
+	mutex_lock(&pasid_mutex);

drivers/iommu/intel/svm.c static inline bool intel_svm_capable(struct intel_iommu *iommu) · line 120

Quoted 10 lines
drivers/iommu/intel/svm.c +10 -0
+	list_for_each_entry_rcu(sdev, &svm->devs, list, pasid_lock_held()) {
+		/* Does not poison forward pointer */
+		list_del_rcu(&sdev->list);
+		spin_lock(&svm->iommu->lock);
+		intel_pasid_tear_down_entry(svm->iommu, sdev->dev,
+					svm->pasid, true);
+		spin_unlock(&svm->iommu->lock);
+		kfree_rcu(sdev, rcu);
+		/*
+		 * Free before unbind only happens with guest usaged

usaged?

Quoted 82 lines
drivers/iommu/intel/svm.c +49 -1
+		 * host PASIDs. IOASID free will detach private data
+		 * and free the IOASID entry.
+		 */
+		ioasid_put(NULL, svm->pasid);
+		if (list_empty(&svm->devs))
+			kfree(svm);
+	}
+	mutex_unlock(&pasid_mutex);
+}
+
+
+static int pasid_status_change(struct notifier_block *nb,
+				unsigned long code, void *data)
+{
+	struct ioasid_nb_args *args = (struct ioasid_nb_args *)data;
+	struct intel_svm *svm = (struct intel_svm *)args->pdata;
+	int ret = NOTIFY_DONE;
+
+	if (code == IOASID_FREE) {
+		if (!svm)
+			goto done;
+		if (args->id != svm->pasid) {
+			pr_warn("Notify PASID does not match data %d : %d\n",
+				args->id, svm->pasid);
+			goto done;
+		}
+		schedule_work(&svm->work);
+		return NOTIFY_OK;
+	}
+done:
+	return ret;> +}
+
+static struct notifier_block pasid_nb = {
+	.notifier_call = pasid_status_change,
+};
+
+void intel_svm_add_pasid_notifier(void)
+{
+	/* Listen to all PASIDs, not specific to a set */
+	ioasid_register_notifier(NULL, &pasid_nb);
+}
+
 void intel_svm_check(struct intel_iommu *iommu)
 {
 	if (!pasid_supported(iommu))
@@ -221,7 +287,6 @@ static const struct mmu_notifier_ops intel_mmuops = {
 	.invalidate_range = intel_invalidate_range,
 };
 
-static DEFINE_MUTEX(pasid_mutex);
 static LIST_HEAD(global_svm_list);
 
 #define for_each_svm_dev(sdev, svm, d)			\
@@ -342,7 +407,14 @@ int intel_svm_bind_gpasid(struct iommu_domain *domain, struct device *dev,
 			svm->gpasid = data->gpasid;
 			svm->flags |= SVM_FLAG_GUEST_PASID;
 		}
+		svm->iommu = iommu;
+		/*
+		 * Set up cleanup async work in case IOASID core notify us PASID
+		 * is freed before unbind.
+		 */
+		INIT_WORK(&svm->work, intel_svm_free_async_fn);
 		ioasid_attach_data(data->hpasid, svm);
+		ioasid_get(NULL, svm->pasid);
 		INIT_LIST_HEAD_RCU(&svm->devs);
 		mmput(svm->mm);
 	}
include/linux/intel-iommu.h +1 -0
index b1ed2f25f7c0..d36038e6ae0b 100644
@@ -744,6 +744,7 @@ void intel_svm_unbind(struct iommu_sva *handle);
 int intel_svm_get_pasid(struct iommu_sva *handle);
 int intel_svm_page_response(struct device *dev, struct iommu_fault_event *evt,
 			    struct iommu_page_response *msg);
+void intel_svm_add_pasid_notifier(void);
 
 struct svm_dev_ops;
 

include/linux/intel-iommu.h struct intel_svm { · line 777

Quoted 7 lines
include/linux/intel-iommu.h +1 -0
@@ -770,6 +771,7 @@ struct intel_svm {
 	int gpasid; /* In case that guest PASID is different from host PASID */
 	struct list_head devs;
 	struct list_head list;
+	struct work_struct work; /* For deferred clean up */
 };
 #else
 static inline void intel_svm_check(struct intel_iommu *iommu) {}

Thanks

Eric

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

#54 lore raw

Re: [PATCH v2 9/9] iommu/vt-d: Store guest PASID during bind

In reply to Jacob Pan

Hi Jacob,

Quoted 15 lines · Jacob Pan

IOASID core maintains the guest-host mapping in the form of SPID and IOASID. This patch assigns the guest PASID (if valid) as SPID while binding guest page table with a host PASID. This mapping will be used for lookup and notifications.

---
 drivers/iommu/intel/svm.c | 2 ++
 1 file changed, 2 insertions(+)
drivers/iommu/intel/svm.c +0 -0
index d8a5efa75095..4c958b1aec4c 100644

drivers/iommu/intel/svm.c int intel_svm_bind_gpasid(struct iommu_domain *domain, struct device *dev, · line 409

Quoted 4 lines
drivers/iommu/intel/svm.c +1 -0
@@ -406,6 +406,7 @@ int intel_svm_bind_gpasid(struct iommu_domain *domain, struct device *dev,
 		if (data->flags & IOMMU_SVA_GPASID_VAL) {
 			svm->gpasid = data->gpasid;
 			svm->flags |= SVM_FLAG_GUEST_PASID;
+			ioasid_attach_spid(data->hpasid, data->gpasid);

don't you want to handle the returned value?

drivers/iommu/intel/svm.c int intel_svm_unbind_gpasid(struct device *dev, int pasid) · line 521

Quoted 8 lines
drivers/iommu/intel/svm.c +1 -0
 		}
 		svm->iommu = iommu;
 		/*
@@ -517,6 +518,7 @@ int intel_svm_unbind_gpasid(struct device *dev, int pasid)
 				ioasid_attach_data(pasid, NULL);
 				ioasid_notify(pasid, IOASID_UNBIND,
 					IOASID_NOTIFY_SET);
+				ioasid_attach_spid(pasid, INVALID_IOASID);

So this answers my previous question ;-) but won't it enter the if (!ioasid_data) path and fail to reset the spid?

Eric

drivers/iommu/intel/svm.c int intel_svm_unbind_gpasid(struct device *dev, int pasid) · line 524

Quoted 4 lines
drivers/iommu/intel/svm.c +0 -0
 				kfree(svm);
 			}
 		}

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

#16 lore raw

Re: [PATCH v2 3/9] iommu/ioasid: Introduce ioasid_set APIs

In reply to Lu Baolu
Quoted 221 lines · On Mon, 24 Aug 2020 10:24:11 +0800 Lu Baolu <[email protected]> wrote:

Hi Jacob,

Quoted 218 lines · Jacob Pan

ioasid_set was introduced as an arbitrary token that are shared by a group of IOASIDs. For example, if IOASID #1 and #2 are allocated via the same ioasid_set*, they are viewed as to belong to the same set.

For guest SVA usages, system-wide IOASID resources need to be partitioned such that VMs can have its own quota and being managed separately. ioasid_set is the perfect candidate for meeting such requirements. This patch redefines and extends ioasid_set with the following new fields:
- Quota
- Reference count
- Storage of its namespace
- The token is stored in the new ioasid_set but with optional types

ioasid_set level APIs are introduced that wires up these new data. Existing users of IOASID APIs are converted where a host IOASID set is allocated for bare-metal usage.

---
  drivers/iommu/intel/iommu.c |  27 ++-
  drivers/iommu/intel/pasid.h |   1 +
  drivers/iommu/intel/svm.c   |   8 +-
  drivers/iommu/ioasid.c      | 390

+++++++++++++++++++++++++++++++++++++++++---
include/linux/ioasid.h | 82 ++++++++-- 5 files changed, 465
insertions(+), 43 deletions(-)

(quoted diff) +0 -0

b/drivers/iommu/intel/iommu.c index a3a0b5c8921d..5813eeaa5edb 100644 --- a/drivers/iommu/intel/iommu.c
+++ b/drivers/iommu/intel/iommu.c

drivers/iommu/intel/iommu.c +4 -0
@@ -42,6 +42,7 @@
  #include <linux/crash_dump.h>
  #include <linux/numa.h>
  #include <linux/swiotlb.h>
+#include <linux/ioasid.h>
  #include <asm/irq_remapping.h>
  #include <asm/cacheflush.h>
  #include <asm/iommu.h>
@@ -103,6 +104,9 @@
   */
  #define INTEL_IOMMU_PGSIZES	(~0xFFFUL)
 
+/* PASIDs used by host SVM */
+struct ioasid_set *host_pasid_set;
+
  static inline int agaw_to_level(int agaw)
  {
  	return agaw + 2;
@@ -3103,8 +3107,8 @@ static void intel_vcmd_ioasid_free(ioasid_t
 ioasid, void *data)
  	 * Sanity check the ioasid owner is done at upper layer,

e.g. VFIO
* We can only free the PASID when all the devices are
unbound. */

drivers/iommu/intel/iommu.c +2 -2
-	if (ioasid_find(NULL, ioasid, NULL)) {
-		pr_alert("Cannot free active IOASID %d\n", ioasid);
+	if (IS_ERR(ioasid_find(host_pasid_set, ioasid, NULL))) {
+		pr_err("Cannot free IOASID %d, not in system

set\n", ioasid); return;
}
vcmd_free_pasid(iommu, ioasid);

(quoted diff) +13 -0
@@ -3288,6 +3292,19 @@ static int __init init_dmars(void)
  	if (ret)
  		goto free_iommu;
 
+	/* PASID is needed for scalable mode irrespective to SVM */
+	if (intel_iommu_sm) {
+		ioasid_install_capacity(intel_pasid_max_id);
+		/* We should not run out of IOASIDs at boot */
+		host_pasid_set = ioasid_alloc_set(NULL,
 PID_MAX_DEFAULT,
+
 IOASID_SET_TYPE_NULL);
+		if (IS_ERR_OR_NULL(host_pasid_set)) {
+			pr_err("Failed to enable host PASID
 allocator %lu\n",
+				PTR_ERR(host_pasid_set));
+			intel_iommu_sm = 0;
+		}
+	}
+

/*
* for each drhd
* enable fault log

drivers/iommu/intel/iommu.c +3 -4
@@ -5149,7 +5166,7 @@ static void auxiliary_unlink_device(struct
 dmar_domain *domain, domain->auxd_refcnt--;
 
  	if (!domain->auxd_refcnt && domain->default_pasid > 0)
-		ioasid_free(domain->default_pasid);
+		ioasid_free(host_pasid_set, domain->default_pasid);
  }
 
  static int aux_domain_add_dev(struct dmar_domain *domain,
@@ -5167,7 +5184,7 @@ static int aux_domain_add_dev(struct
 dmar_domain *domain, int pasid;
 
  		/* No private data needed for the default pasid */
-		pasid = ioasid_alloc(NULL, PASID_MIN,
+		pasid = ioasid_alloc(host_pasid_set, PASID_MIN,
  				     pci_max_pasids(to_pci_dev(dev))
- 1, NULL);
  		if (pasid == INVALID_IOASID) {
@@ -5210,7 +5227,7 @@ static int aux_domain_add_dev(struct
 dmar_domain *domain, spin_unlock(&iommu->lock);
  	spin_unlock_irqrestore(&device_domain_lock, flags);
  	if (!domain->auxd_refcnt && domain->default_pasid > 0)
-		ioasid_free(domain->default_pasid);
+		ioasid_free(host_pasid_set, domain->default_pasid);
 
  	return ret;
  }
(quoted diff) +0 -0

b/drivers/iommu/intel/pasid.h index c9850766c3a9..ccdc23446015 100644 --- a/drivers/iommu/intel/pasid.h
+++ b/drivers/iommu/intel/pasid.h

drivers/iommu/intel/pasid.h +1 -0
@@ -99,6 +99,7 @@ static inline bool pasid_pte_is_present(struct
 pasid_entry *pte) }
 
  extern u32 intel_pasid_max_id;
+extern struct ioasid_set *host_pasid_set;
  int intel_pasid_alloc_id(void *ptr, int start, int end, gfp_t
 gfp); void intel_pasid_free_id(int pasid);
  void *intel_pasid_lookup_id(int pasid);
drivers/iommu/intel/svm.c +1 -1
index 37a9beabc0ca..634e191ca2c3 100644
@@ -551,7 +551,7 @@ intel_svm_bind_mm(struct device *dev, int
 flags, struct svm_dev_ops *ops, pasid_max = intel_pasid_max_id;
 
  		/* Do not use PASID 0, reserved for RID to PASID
 */
-		svm->pasid = ioasid_alloc(NULL, PASID_MIN,
+		svm->pasid = ioasid_alloc(host_pasid_set,
 PASID_MIN, pasid_max - 1, svm);
  		if (svm->pasid == INVALID_IOASID) {
  			kfree(svm);
drivers/iommu/intel/svm.c +2 -2
@@ -568,7 +568,7 @@ intel_svm_bind_mm(struct device *dev, int
 flags, struct svm_dev_ops *ops, if (mm) {
  			ret =
 mmu_notifier_register(&svm->notifier, mm); if (ret) {
-				ioasid_free(svm->pasid);
+				ioasid_free(host_pasid_set,
 svm->pasid); kfree(svm);
  				kfree(sdev);
  				goto out;
@@ -586,7 +586,7 @@ intel_svm_bind_mm(struct device *dev, int
 flags, struct svm_dev_ops *ops, if (ret) {
  			if (mm)
  				mmu_notifier_unregister(&svm->notifier,
 mm);
-			ioasid_free(svm->pasid);
+			ioasid_free(host_pasid_set, svm->pasid);
  			kfree(svm);
  			kfree(sdev);

goto out;

drivers/iommu/intel/svm.c +1 -1
@@ -655,7 +655,7 @@ static int intel_svm_unbind_mm(struct device
 *dev, int pasid) kfree_rcu(sdev, rcu);
 
  			if (list_empty(&svm->devs)) {
-				ioasid_free(svm->pasid);
+				ioasid_free(host_pasid_set,
 svm->pasid); if (svm->mm)
  					mmu_notifier_unregister(&svm->notifier,
 svm->mm); list_del(&svm->list);
drivers/iommu/ioasid.c +0 -0
index 5f63af07acd5..f73b3dbfc37a 100644
@@ -1,22 +1,58 @@
  // SPDX-License-Identifier: GPL-2.0
  /*
   * I/O Address Space ID allocator. There is one global IOASID

space, split into
- * subsets. Users create a subset with DECLARE_IOASID_SET, then allocate and
- * free IOASIDs with ioasid_alloc and ioasid_free.
+ * subsets. Users create a subset with ioasid_alloc_set, then allocate/free IDs
+ * with ioasid_alloc and ioasid_free.
*/

drivers/iommu/ioasid.c +20 -1
-#include <linux/ioasid.h>
  #include <linux/module.h>
  #include <linux/slab.h>
  #include <linux/spinlock.h>
  #include <linux/xarray.h>
+#include <linux/ioasid.h>
+
+static DEFINE_XARRAY_ALLOC(ioasid_sets);
+enum ioasid_state {
+	IOASID_STATE_INACTIVE,
+	IOASID_STATE_ACTIVE,
+	IOASID_STATE_FREE_PENDING,
+};
 
+/**
+ * struct ioasid_data - Meta data about ioasid
+ *
+ * @id:		Unique ID
+ * @users	Number of active users
+ * @state	Track state of the IOASID
+ * @set		Meta data of the set this IOASID belongs to
+ * @private	Private data associated with the IOASID
+ * @rcu		For free after RCU grace period
+ */
  struct ioasid_data {
  	ioasid_t id;
  	struct ioasid_set *set;
+	refcount_t users;
+	enum ioasid_state state;
  	void *private;
  	struct rcu_head rcu;
  };
Quoted 11 lines

drivers/iommu/ioasid.c line 46

Quoted 8 lines
drivers/iommu/ioasid.c +8 -0
+/* Default to PCIe standard 20 bit PASID */
+#define PCI_PASID_MAX 0x100000
+static ioasid_t ioasid_capacity = PCI_PASID_MAX;
+static ioasid_t ioasid_capacity_avail = PCI_PASID_MAX;
+
+void ioasid_install_capacity(ioasid_t total)
+{
+	ioasid_capacity = ioasid_capacity_avail = total;  

Need any check for multiple settings or setting after used?

Good point, capacity can only be set once. will add check to prevent setting after use.

Quoted 149 lines
Quoted 149 lines
drivers/iommu/ioasid.c +9 -0
+}
+EXPORT_SYMBOL_GPL(ioasid_install_capacity);
+
+ioasid_t ioasid_get_capacity()
+{
+	return ioasid_capacity;
+}
+EXPORT_SYMBOL_GPL(ioasid_get_capacity);
+
  /*
   * struct ioasid_allocator_data - Internal data structure to hold

information
* about an allocator. There are two types of allocators:

drivers/iommu/ioasid.c +4 -1
@@ -306,11 +342,23 @@ ioasid_t ioasid_alloc(struct ioasid_set *set,
 ioasid_t min, ioasid_t max, {
  	struct ioasid_data *data;
  	void *adata;
-	ioasid_t id;
+	ioasid_t id = INVALID_IOASID;
+
+	spin_lock(&ioasid_allocator_lock);
+	/* Check if the IOASID set has been allocated and

initialized */

drivers/iommu/ioasid.c +7 -0
+	if (WARN_ON(xa_load(&ioasid_sets, set->sid) != set)) {
+		pr_warn("Invalid set\n");
+		goto done_unlock;
+	}
+
+	if (set->quota <= set->nr_ioasids) {
+		pr_err("IOASID set %d out of quota %d\n",

set->sid, set->quota);

drivers/iommu/ioasid.c +3 -1
+		goto done_unlock;
+	}
 
  	data = kzalloc(sizeof(*data), GFP_ATOMIC);
  	if (!data)
-		return INVALID_IOASID;
+		goto done_unlock;
 
  	data->set = set;
  	data->private = private;
@@ -319,7 +367,6 @@ ioasid_t ioasid_alloc(struct ioasid_set *set,
 ioasid_t min, ioasid_t max,
  	 * Custom allocator needs allocator data to perform

platform specific
* operations.
*/

drivers/iommu/ioasid.c +0 -1
-	spin_lock(&ioasid_allocator_lock);
  	adata = active_allocator->flags &

IOASID_ALLOCATOR_CUSTOM ? active_allocator->ops->pdata : data; id = active_allocator->ops->alloc(min, max, adata); if (id == INVALID_IOASID) { @@ -335,42 +382,339 @@ ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, goto exit_free; }
data->id = id;

drivers/iommu/ioasid.c +6 -0
+	data->state = IOASID_STATE_ACTIVE;
+	refcount_set(&data->users, 1);
+
+	/* Store IOASID in the per set data */
+	if (xa_err(xa_store(&set->xa, id, data, GFP_ATOMIC))) {
+		pr_err("Failed to ioasid %d in set %d\n", id,

set->sid);

drivers/iommu/ioasid.c +13 -4
+		goto exit_free;
+	}
+	set->nr_ioasids++;
+	goto done_unlock;
 
-	spin_unlock(&ioasid_allocator_lock);
-	return id;
  exit_free:
-	spin_unlock(&ioasid_allocator_lock);
  	kfree(data);
-	return INVALID_IOASID;
+done_unlock:
+	spin_unlock(&ioasid_allocator_lock);
+	return id;
  }
  EXPORT_SYMBOL_GPL(ioasid_alloc);
 
+static void ioasid_do_free(struct ioasid_data *data)
+{
+	struct ioasid_data *ioasid_data;
+	struct ioasid_set *sdata;
+
+	active_allocator->ops->free(data->id,

active_allocator->ops->pdata);

drivers/iommu/ioasid.c +1 -0
+	/* Custom allocator needs additional steps to free the xa

element */

drivers/iommu/ioasid.c +2 -0
+	if (active_allocator->flags & IOASID_ALLOCATOR_CUSTOM) {
+		ioasid_data = xa_erase(&active_allocator->xa,

data->id);

drivers/iommu/ioasid.c +14 -0
+		kfree_rcu(ioasid_data, rcu);
+	}
+
+	sdata = xa_load(&ioasid_sets, data->set->sid);
+	if (!sdata) {
+		pr_err("No set %d for IOASID %d\n", data->set->sid,
+		       data->id);
+		return;
+	}
+	xa_erase(&sdata->xa, data->id);
+	sdata->nr_ioasids--;
+}
+
+static void ioasid_free_locked(struct ioasid_set *set, ioasid_t

ioasid) +{

drivers/iommu/ioasid.c +5 -0
+	struct ioasid_data *data;
+
+	data = xa_load(&active_allocator->xa, ioasid);
+	if (!data) {
+		pr_err("Trying to free unknown IOASID %u\n",

ioasid);

drivers/iommu/ioasid.c +5 -0
+		return;
+	}
+
+	if (data->set != set) {
+		pr_warn("Cannot free IOASID %u due to set

ownership\n", ioasid);

drivers/iommu/ioasid.c +11 -2
+		return;
+	}
+	data->state = IOASID_STATE_FREE_PENDING;
+
+	if (!refcount_dec_and_test(&data->users))
+		return;
+
+	ioasid_do_free(data);
+}
+
  /**
- * ioasid_free - Free an IOASID
- * @ioasid: the ID to remove
+ * ioasid_free - Drop reference on an IOASID. Free if refcount

drops to 0,
+ * including free from its set and system-wide list.
+ * @set: The ioasid_set to check permission with. If not
NULL, IOASID
+ * free will fail if the set does not match.
+ * @ioasid: The IOASID to remove
*/

drivers/iommu/ioasid.c +6 -2
-void ioasid_free(ioasid_t ioasid)
+void ioasid_free(struct ioasid_set *set, ioasid_t ioasid)
  {
-	struct ioasid_data *ioasid_data;
+	spin_lock(&ioasid_allocator_lock);
+	ioasid_free_locked(set, ioasid);
+	spin_unlock(&ioasid_allocator_lock);
+}
+EXPORT_SYMBOL_GPL(ioasid_free);
Quoted 8 lines
Quoted 4 lines

+/**
+ * ioasid_alloc_set - Allocate a new IOASID set for a given token
+ *
+ * @token: Unique token of the IOASID set, cannot be NULL

What's the use of @token? I might be able to find the answer in the code, but i have no idea when i comes here first time. :-)

You are right, I need to clarify here. How about:
* @token: An optional arbitrary number that can be associated
with the
* IOASID set. @token can be NULL if the type is
* IOASID_SET_TYPE_NULL
* @quota: Quota allowed in this set
* @type: The type of the token used to create the IOASID set

Quoted 4 lines

This line of comment says that token cannot be NULL. It seems not to match the real code, where token could be NULL if type is IOASID_SET_TYPE_NULL.

Good catch, it was a relic from previous versions.

Quoted 9 lines
Quoted 8 lines

+ * @quota: Quota allowed in this set. Only for new set
creation
+ * @flags: Special requirements
+ *
+ * IOASID can be limited system-wide resource that requires quota management.
+ * If caller does not wish to enforce quota, use IOASID_SET_NO_QUOTA flag.

Quoted 3 lines

If you are not going to add NO_QUOTA support this time, I'd suggest you to remove above comment.

Yes, will remove.

Quoted 10 lines
Quoted 7 lines

+ *
+ * Token will be stored in the ioasid_set returned. A reference will be taken
+ * upon finding a matching set or newly created set.
+ * IOASID allocation within the set and other per set operations will use
+ * the retured ioasid_set *.

nit: remove *, or you mean the ioasid_set pointer?

Sounds good

Quoted 26 lines
Quoted 22 lines

+ *
+ */
+struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota, int type) +{

drivers/iommu/ioasid.c +9 -0
+	struct ioasid_set *sdata;
+	unsigned long index;
+	ioasid_t id;
+
+	if (type >= IOASID_SET_TYPE_NR)
+		return ERR_PTR(-EINVAL);
+
+	/*
+	 * Need to check space available if we share system-wide

quota.

drivers/iommu/ioasid.c +1 -0
+	 * TODO: we may need to support quota free sets in the

future.

drivers/iommu/ioasid.c +1 -3
+	 */
  	spin_lock(&ioasid_allocator_lock);
-	ioasid_data = xa_load(&active_allocator->xa, ioasid);
-	if (!ioasid_data) {
-		pr_err("Trying to free unknown IOASID %u\n",

ioasid);

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 493

Quoted 1 line
drivers/iommu/ioasid.c +1 -0
+	if (quota > ioasid_capacity_avail) {  

Thinking that ioasid_set itself needs an ioasid, so the check might be

Quoted 2 lines

(quota + 1 > ioasid_capacity_avail)?

ioasid_set does not need an ioasid, the set ID is allocated from a different XArray: ioasid_sets.

Quoted 26 lines
Quoted 26 lines
drivers/iommu/ioasid.c +1 -0
+		pr_warn("Out of IOASID capacity! ask %d, avail

%d\n",

drivers/iommu/ioasid.c +2 -1
+			quota, ioasid_capacity_avail);
+		sdata = ERR_PTR(-ENOSPC);
  		goto exit_unlock;
  	}
 
-	active_allocator->ops->free(ioasid,

active_allocator->ops->pdata);

drivers/iommu/ioasid.c +0 -1
-	/* Custom allocator needs additional steps to free the xa

element */

drivers/iommu/ioasid.c +0 -2
-	if (active_allocator->flags & IOASID_ALLOCATOR_CUSTOM) {
-		ioasid_data = xa_erase(&active_allocator->xa,

ioasid);

drivers/iommu/ioasid.c +2 -1
-		kfree_rcu(ioasid_data, rcu);
+	/*
+	 * Token is only unique within its types but right now we

have only

drivers/iommu/ioasid.c +1 -0
+	 * mm type. If we have more token types, we have to match

type as well.

drivers/iommu/ioasid.c +6 -0
+	 */
+	switch (type) {
+	case IOASID_SET_TYPE_MM:
+		/* Search existing set tokens, reject duplicates */
+		xa_for_each(&ioasid_sets, index, sdata) {
+			if (sdata->token == token &&
Quoted 9 lines

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 513

Quoted 6 lines
drivers/iommu/ioasid.c +1 -0
+				sdata->type == IOASID_SET_TYPE_MM)

{

drivers/iommu/ioasid.c +4 -0
+				sdata = ERR_PTR(-EEXIST);
+				goto exit_unlock;
+			}
+		}  

Do you need to enforce non-NULL token policy here?

yes, mm type cannot have NULL token. NULL type must have NULL token. This design has been flipped a couple of times. Thanks for catching this.

Quoted 4 lines

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 517

Quoted 4 lines
drivers/iommu/ioasid.c +4 -0
+		break;
+	case IOASID_SET_TYPE_NULL:
+		if (!token)
+			break;

should not be NULL

Quoted 46 lines
Quoted 39 lines
drivers/iommu/ioasid.c +6 -0
+		fallthrough;
+	default:
+		pr_err("Invalid token and IOASID type\n");
+		sdata = ERR_PTR(-EINVAL);
+		goto exit_unlock;
  	}
 
+	/* REVISIT: may support set w/o quota, use system

available */

drivers/iommu/ioasid.c +27 -0
+	if (!quota) {
+		sdata = ERR_PTR(-EINVAL);
+		goto exit_unlock;
+	}
+
+	sdata = kzalloc(sizeof(*sdata), GFP_ATOMIC);
+	if (!sdata) {
+		sdata = ERR_PTR(-ENOMEM);
+		goto exit_unlock;
+	}
+
+	if (xa_alloc(&ioasid_sets, &id, sdata,
+			XA_LIMIT(0, ioasid_capacity_avail - quota),
+			GFP_ATOMIC)) {
+		kfree(sdata);
+		sdata = ERR_PTR(-ENOSPC);
+		goto exit_unlock;
+	}
+
+	sdata->token = token;
+	sdata->type = type;
+	sdata->quota = quota;
+	sdata->sid = id;
+	refcount_set(&sdata->ref, 1);
+
+	/*
+	 * Per set XA is used to store private IDs within the set,

get ready

drivers/iommu/ioasid.c +1 -0
+	 * for ioasid_set private ID and system-wide IOASID

allocation

drivers/iommu/ioasid.c

Quoted 2 lines
drivers/iommu/ioasid.c +2 -0
+	 * results.
+	 */  

I'm not sure that I understood the per-set XA correctly. Is it used to save both private ID and real ID allocated from the system-wide pool? If so, isn't private id might be equal to real ID?

Quoted 9 lines

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 557

Quoted 2 lines
drivers/iommu/ioasid.c +2 -0
+	xa_init_flags(&sdata->xa, XA_FLAGS_ALLOC);
+	ioasid_capacity_avail -= quota;  

As mentioned above, the ioasid_set consumed one extra ioasid, so

	ioasid_capacity_avail -= (quota + 1);

?

Same explanation, make sense?

Quoted 38 lines
Quoted 24 lines
drivers/iommu/ioasid.c +20 -1
+
  exit_unlock:
  	spin_unlock(&ioasid_allocator_lock);
+
+	return sdata;
  }
-EXPORT_SYMBOL_GPL(ioasid_free);
+EXPORT_SYMBOL_GPL(ioasid_alloc_set);
+
+void ioasid_set_get_locked(struct ioasid_set *set)
+{
+	if (WARN_ON(xa_load(&ioasid_sets, set->sid) != set)) {
+		pr_warn("Invalid set data\n");
+		return;
+	}
+
+	refcount_inc(&set->ref);
+}
+EXPORT_SYMBOL_GPL(ioasid_set_get_locked);
+
+void ioasid_set_get(struct ioasid_set *set)
+{
+	spin_lock(&ioasid_allocator_lock);
+	ioasid_set_get_locked(set);

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 590

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+	spin_unlock(&ioasid_allocator_lock);
+}
+EXPORT_SYMBOL_GPL(ioasid_set_get);
+
+void ioasid_set_put_locked(struct ioasid_set *set)
+{
+	struct ioasid_data *entry;
+	unsigned long index;
+
+	if (WARN_ON(xa_load(&ioasid_sets, set->sid) != set)) {  

There are multiple occurences of this line of code, how about defining it as a inline helper?

Quoted 5 lines

static inline bool ioasid_is_valid(struct ioasid_set *set)

{
	return xa_load(&ioasid_sets, set->sid) == set;
}

Sounds good. perhaps the name should be
ioasid_set_is_valid()?

Quoted 42 lines
Quoted 42 lines
drivers/iommu/ioasid.c +7 -0
+		pr_warn("Invalid set data\n");
+		return;
+	}
+
+	if (!refcount_dec_and_test(&set->ref)) {
+		pr_debug("%s: IOASID set %d has %d users\n",
+			__func__, set->sid,

refcount_read(&set->ref));

drivers/iommu/ioasid.c +9 -0
+		return;
+	}
+
+	/* The set is already empty, we just destroy the set. */
+	if (xa_empty(&set->xa))
+		goto done_destroy;
+
+	/*
+	 * Free all PASIDs from system-wide IOASID pool, all

subscribers gets

drivers/iommu/ioasid.c +2 -0
+	 * notified and do cleanup of their own.
+	 * Note that some references of the IOASIDs within the set

can still

drivers/iommu/ioasid.c +1 -0
+	 * be held after the free call. This is OK in that the

IOASIDs will be

drivers/iommu/ioasid.c +1 -0
+	 * marked inactive, the only operations can be done is

ioasid_put.

drivers/iommu/ioasid.c +1 -0
+	 * No need to track IOASID set states since there is no

reclaim phase.

drivers/iommu/ioasid.c +14 -0
+	 */
+	xa_for_each(&set->xa, index, entry) {
+		ioasid_free_locked(set, index);
+		/* Free from per set private pool */
+		xa_erase(&set->xa, index);
+	}
+
+done_destroy:
+	/* Return the quota back to system pool */
+	ioasid_capacity_avail += set->quota;
+	kfree_rcu(set, rcu);
+
+	/*
+	 * Token got released right away after the ioasid_set is

freed.

Quoted 9 lines

drivers/iommu/ioasid.c

Quoted 6 lines
drivers/iommu/ioasid.c +1 -0
+	 * If a new set is created immediately with the newly

released token,

drivers/iommu/ioasid.c +1 -0
+	 * it will not allocate the same IOASIDs unless they are

reclaimed.

drivers/iommu/ioasid.c +2 -0
+	 */
+	xa_erase(&ioasid_sets, set->sid);  

No. pointer is used after free.

Right, will move it before kfree_rcu.

Quoted 51 lines
Quoted 41 lines
drivers/iommu/ioasid.c +5 -0
+}
+EXPORT_SYMBOL_GPL(ioasid_set_put_locked);
+
+/**
+ * ioasid_set_put - Drop a reference to the IOASID set. Free all

IOASIDs within
+ * the set if there are no more users.
+ *
+ * @set: The IOASID set ID to be freed
+ *
+ * If refcount drops to zero, all IOASIDs allocated within the set will be
+ * freed.
+ */

drivers/iommu/ioasid.c +14 -0
+void ioasid_set_put(struct ioasid_set *set)
+{
+	spin_lock(&ioasid_allocator_lock);
+	ioasid_set_put_locked(set);
+	spin_unlock(&ioasid_allocator_lock);
+}
+EXPORT_SYMBOL_GPL(ioasid_set_put);
+
+/**
+ * ioasid_adjust_set - Adjust the quota of an IOASID set
+ * @set:	IOASID set to be assigned
+ * @quota:	Quota allowed in this set
+ *
+ * Return 0 on success. If the new quota is smaller than the

number of
+ * IOASIDs already allocated, -EINVAL will be returned. No change will be
+ * made to the existing quota.
+ */

drivers/iommu/ioasid.c +7 -0
+int ioasid_adjust_set(struct ioasid_set *set, int quota)
+{
+	int ret = 0;
+
+	spin_lock(&ioasid_allocator_lock);
+	if (set->nr_ioasids > quota) {
+		pr_err("New quota %d is smaller than outstanding

IOASIDs %d\n",

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 671

Quoted 6 lines
drivers/iommu/ioasid.c +6 -0
+			quota, set->nr_ioasids);
+		ret = -EINVAL;
+		goto done_unlock;
+	}
+
+	if (quota >= ioasid_capacity_avail) {  

This check doesn't make sense since you are updating (not asking for) a quota.

Quoted 3 lines

if ((quota > set->quota) &&

     (quota - set->quota > ioasid_capacity_avail))

Good point, will fix.

Thanks a lot!

Quoted 175 lines

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 677

Quoted 6 lines
drivers/iommu/ioasid.c +6 -0
+		ret = -ENOSPC;
+		goto done_unlock;
+	}
+
+	/* Return the delta back to system pool */
+	ioasid_capacity_avail += set->quota - quota;  

ioasid_capacity_avail is defined as a unsigned int, hence this always increase the available capacity value even the caller is asking for a bigger quota?

Quoted 19 lines
drivers/iommu/ioasid.c +3 -0
+
+	/*
+	 * May have a policy to prevent giving all available

IOASIDs

drivers/iommu/ioasid.c +1 -0
+	 * to one set. But we don't enforce here, it should be in

the

drivers/iommu/ioasid.c +13 -0
+	 * upper layers.
+	 */
+	set->quota = quota;
+
+done_unlock:
+	spin_unlock(&ioasid_allocator_lock);
+
+	return ret;
+}
+EXPORT_SYMBOL_GPL(ioasid_adjust_set);
+
+/**
+ * ioasid_set_for_each_ioasid - Iterate over all the IOASIDs
Quoted 4 lines

within the set
+ *
+ * Caller must hold a reference of the set and handles its own locking.

Do you need to hold ioasid_allocator_lock here?

Quoted 138 lines

+ */

drivers/iommu/ioasid.c +20 -0
+int ioasid_set_for_each_ioasid(struct ioasid_set *set,
+			       void (*fn)(ioasid_t id, void *data),
+			       void *data)
+{
+	struct ioasid_data *entry;
+	unsigned long index;
+	int ret = 0;
+
+	if (xa_empty(&set->xa)) {
+		pr_warn("No IOASIDs in the set %d\n", set->sid);
+		return -ENOENT;
+	}
+
+	xa_for_each(&set->xa, index, entry) {
+		fn(index, data);
+	}
+
+	return ret;
+}
+EXPORT_SYMBOL_GPL(ioasid_set_for_each_ioasid);
 
  /**
   * ioasid_find - Find IOASID data
include/linux/ioasid.h +0 -0
index 9c44947a68c8..412d025d440e 100644
@@ -10,8 +10,35 @@ typedef unsigned int ioasid_t;
  typedef ioasid_t (*ioasid_alloc_fn_t)(ioasid_t min, ioasid_t max,

void *data); typedef void (*ioasid_free_fn_t)(ioasid_t ioasid, void *data);

include/linux/ioasid.h +14 -0
+/* IOASID set types */
+enum ioasid_set_type {
+	IOASID_SET_TYPE_NULL = 1, /* Set token is NULL */
+	IOASID_SET_TYPE_MM,	  /* Set token is a mm_struct,
+				   * i.e. associated with a process
+				   */
+	IOASID_SET_TYPE_NR,
+};
+
+/**
+ * struct ioasid_set - Meta data about ioasid_set
+ * @type:	Token types and other features
+ * @token:	Unique to identify an IOASID set
+ * @xa:		XArray to store ioasid_set private IDs, can

be used for
+ * guest-host IOASID mapping, or just a private
IOASID namespace.
+ * @quota: Max number of IOASIDs can be allocated within
the set
+ * @nr_ioasids Number of IOASIDs currently allocated in the
set
+ * @sid: ID of the set
+ * @ref: Reference count of the users
+ */
struct ioasid_set {

include/linux/ioasid.h +11 -3
-	int dummy;
+	void *token;
+	struct xarray xa;
+	int type;
+	int quota;
+	int nr_ioasids;
+	int sid;
+	refcount_t ref;
+	struct rcu_head rcu;
  };
 
  /**
@@ -29,31 +56,64 @@ struct ioasid_allocator_ops {
  	void *pdata;
  };
 
-#define DECLARE_IOASID_SET(name) struct ioasid_set name = { 0 }
-
  #if IS_ENABLED(CONFIG_IOASID)
+void ioasid_install_capacity(ioasid_t total);
+ioasid_t ioasid_get_capacity(void);
+struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota,

int type); +int ioasid_adjust_set(struct ioasid_set *set, int quota); +void ioasid_set_get_locked(struct ioasid_set *set);

include/linux/ioasid.h +3 -0
+void ioasid_set_put_locked(struct ioasid_set *set);
+void ioasid_set_put(struct ioasid_set *set);
+
  ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min,

ioasid_t max, void *private);

include/linux/ioasid.h +5 -3
-void ioasid_free(ioasid_t ioasid);
-void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid,
-		  bool (*getter)(void *));
+void ioasid_free(struct ioasid_set *set, ioasid_t ioasid);
+
+bool ioasid_is_active(ioasid_t ioasid);
+
+void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid, bool

(*getter)(void *)); +int ioasid_attach_data(ioasid_t ioasid, void *data); int ioasid_register_allocator(struct ioasid_allocator_ops *allocator); void ioasid_unregister_allocator(struct ioasid_allocator_ops *allocator); -int ioasid_attach_data(ioasid_t ioasid, void *data); -

include/linux/ioasid.h +13 -0
+void ioasid_is_in_set(struct ioasid_set *set, ioasid_t ioasid);
+int ioasid_set_for_each_ioasid(struct ioasid_set *sdata,
+			       void (*fn)(ioasid_t id, void *data),
+			       void *data);
  #else /* !CONFIG_IOASID */
+static inline void ioasid_install_capacity(ioasid_t total)
+{
+}
+
+static inline ioasid_t ioasid_get_capacity(void)
+{
+	return 0;
+}
+
  static inline ioasid_t ioasid_alloc(struct ioasid_set *set,

ioasid_t min, ioasid_t max, void *private)
{
return INVALID_IOASID;
}

include/linux/ioasid.h +1 -1
-static inline void ioasid_free(ioasid_t ioasid)
+static inline void ioasid_free(struct ioasid_set *set, ioasid_t

ioasid) +{

include/linux/ioasid.h +8 -0
+}
+
+static inline bool ioasid_is_active(ioasid_t ioasid)
+{
+	return false;
+}
+
+static inline struct ioasid_set *ioasid_alloc_set(void *token,

ioasid_t quota, int type) +{

include/linux/ioasid.h +4 -0
+	return ERR_PTR(-ENOTSUPP);
+}
+
+static inline void ioasid_set_put(struct ioasid_set *set)
  {
  }
Quoted 15 lines
Quoted 8 lines

-static inline void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid,

include/linux/ioasid.h +1 -1
-				bool (*getter)(void *))
+static inline void *ioasid_find(struct ioasid_set *set, ioasid_t

ioasid, bool (*getter)(void *)) {
return NULL;
}

Best regards,
baolu
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

[Jacob Pan]

maintainer
#17 lore raw

Re: [PATCH v2 3/9] iommu/ioasid: Introduce ioasid_set APIs

Hi,

Quoted 462 lines · Jacob Pan
Quoted 221 lines · On Mon, 24 Aug 2020 10:24:11 +0800 Lu Baolu <[email protected]> wrote:

Hi Jacob,

Quoted 218 lines · Jacob Pan

ioasid_set was introduced as an arbitrary token that are shared by a group of IOASIDs. For example, if IOASID #1 and #2 are allocated via the same ioasid_set*, they are viewed as to belong to the same set.

For guest SVA usages, system-wide IOASID resources need to be partitioned such that VMs can have its own quota and being managed separately. ioasid_set is the perfect candidate for meeting such requirements. This patch redefines and extends ioasid_set with the following new fields:
- Quota
- Reference count
- Storage of its namespace
- The token is stored in the new ioasid_set but with optional types

ioasid_set level APIs are introduced that wires up these new data. Existing users of IOASID APIs are converted where a host IOASID set is allocated for bare-metal usage.

---
   drivers/iommu/intel/iommu.c |  27 ++-
   drivers/iommu/intel/pasid.h |   1 +
   drivers/iommu/intel/svm.c   |   8 +-
   drivers/iommu/ioasid.c      | 390

+++++++++++++++++++++++++++++++++++++++++---
include/linux/ioasid.h | 82 ++++++++-- 5 files changed, 465
insertions(+), 43 deletions(-)

(quoted diff) +0 -0

b/drivers/iommu/intel/iommu.c index a3a0b5c8921d..5813eeaa5edb 100644 --- a/drivers/iommu/intel/iommu.c
+++ b/drivers/iommu/intel/iommu.c

drivers/iommu/intel/iommu.c +4 -0
@@ -42,6 +42,7 @@
   #include <linux/crash_dump.h>
   #include <linux/numa.h>
   #include <linux/swiotlb.h>
+#include <linux/ioasid.h>
   #include <asm/irq_remapping.h>
   #include <asm/cacheflush.h>
   #include <asm/iommu.h>
@@ -103,6 +104,9 @@
    */
   #define INTEL_IOMMU_PGSIZES	(~0xFFFUL)
 
+/* PASIDs used by host SVM */
+struct ioasid_set *host_pasid_set;
+
   static inline int agaw_to_level(int agaw)
   {
   	return agaw + 2;
@@ -3103,8 +3107,8 @@ static void intel_vcmd_ioasid_free(ioasid_t
 ioasid, void *data)
   	 * Sanity check the ioasid owner is done at upper layer,

e.g. VFIO
* We can only free the PASID when all the devices are
unbound. */

drivers/iommu/intel/iommu.c +2 -2
-	if (ioasid_find(NULL, ioasid, NULL)) {
-		pr_alert("Cannot free active IOASID %d\n", ioasid);
+	if (IS_ERR(ioasid_find(host_pasid_set, ioasid, NULL))) {
+		pr_err("Cannot free IOASID %d, not in system
set\n", ioasid); return;
   	}
   	vcmd_free_pasid(iommu, ioasid);
(quoted diff) +13 -0
@@ -3288,6 +3292,19 @@ static int __init init_dmars(void)
   	if (ret)
   		goto free_iommu;
 
+	/* PASID is needed for scalable mode irrespective to SVM */
+	if (intel_iommu_sm) {
+		ioasid_install_capacity(intel_pasid_max_id);
+		/* We should not run out of IOASIDs at boot */
+		host_pasid_set = ioasid_alloc_set(NULL,
 PID_MAX_DEFAULT,
+
 IOASID_SET_TYPE_NULL);
+		if (IS_ERR_OR_NULL(host_pasid_set)) {
+			pr_err("Failed to enable host PASID
 allocator %lu\n",
+				PTR_ERR(host_pasid_set));
+			intel_iommu_sm = 0;
+		}
+	}
+

/*
* for each drhd
* enable fault log

drivers/iommu/intel/iommu.c +3 -4
@@ -5149,7 +5166,7 @@ static void auxiliary_unlink_device(struct
 dmar_domain *domain, domain->auxd_refcnt--;
 
   	if (!domain->auxd_refcnt && domain->default_pasid > 0)
-		ioasid_free(domain->default_pasid);
+		ioasid_free(host_pasid_set, domain->default_pasid);
   }
 
   static int aux_domain_add_dev(struct dmar_domain *domain,
@@ -5167,7 +5184,7 @@ static int aux_domain_add_dev(struct
 dmar_domain *domain, int pasid;
 
   		/* No private data needed for the default pasid */
-		pasid = ioasid_alloc(NULL, PASID_MIN,
+		pasid = ioasid_alloc(host_pasid_set, PASID_MIN,
   				     pci_max_pasids(to_pci_dev(dev))
- 1, NULL);
   		if (pasid == INVALID_IOASID) {
@@ -5210,7 +5227,7 @@ static int aux_domain_add_dev(struct
 dmar_domain *domain, spin_unlock(&iommu->lock);
   	spin_unlock_irqrestore(&device_domain_lock, flags);
   	if (!domain->auxd_refcnt && domain->default_pasid > 0)
-		ioasid_free(domain->default_pasid);
+		ioasid_free(host_pasid_set, domain->default_pasid);
 
   	return ret;
   }
(quoted diff) +0 -0

b/drivers/iommu/intel/pasid.h index c9850766c3a9..ccdc23446015 100644 --- a/drivers/iommu/intel/pasid.h
+++ b/drivers/iommu/intel/pasid.h

drivers/iommu/intel/pasid.h +1 -0
@@ -99,6 +99,7 @@ static inline bool pasid_pte_is_present(struct
 pasid_entry *pte) }
 
   extern u32 intel_pasid_max_id;
+extern struct ioasid_set *host_pasid_set;
   int intel_pasid_alloc_id(void *ptr, int start, int end, gfp_t
 gfp); void intel_pasid_free_id(int pasid);
   void *intel_pasid_lookup_id(int pasid);
drivers/iommu/intel/svm.c +1 -1
index 37a9beabc0ca..634e191ca2c3 100644
@@ -551,7 +551,7 @@ intel_svm_bind_mm(struct device *dev, int
 flags, struct svm_dev_ops *ops, pasid_max = intel_pasid_max_id;
 
   		/* Do not use PASID 0, reserved for RID to PASID
 */
-		svm->pasid = ioasid_alloc(NULL, PASID_MIN,
+		svm->pasid = ioasid_alloc(host_pasid_set,
 PASID_MIN, pasid_max - 1, svm);
   		if (svm->pasid == INVALID_IOASID) {
   			kfree(svm);
drivers/iommu/intel/svm.c +2 -2
@@ -568,7 +568,7 @@ intel_svm_bind_mm(struct device *dev, int
 flags, struct svm_dev_ops *ops, if (mm) {
   			ret =
 mmu_notifier_register(&svm->notifier, mm); if (ret) {
-				ioasid_free(svm->pasid);
+				ioasid_free(host_pasid_set,
 svm->pasid); kfree(svm);
   				kfree(sdev);
   				goto out;
@@ -586,7 +586,7 @@ intel_svm_bind_mm(struct device *dev, int
 flags, struct svm_dev_ops *ops, if (ret) {
   			if (mm)
   				mmu_notifier_unregister(&svm->notifier,
 mm);
-			ioasid_free(svm->pasid);
+			ioasid_free(host_pasid_set, svm->pasid);
   			kfree(svm);
   			kfree(sdev);
   			goto out;
drivers/iommu/intel/svm.c +1 -1
@@ -655,7 +655,7 @@ static int intel_svm_unbind_mm(struct device
 *dev, int pasid) kfree_rcu(sdev, rcu);
 
   			if (list_empty(&svm->devs)) {
-				ioasid_free(svm->pasid);
+				ioasid_free(host_pasid_set,
 svm->pasid); if (svm->mm)
   					mmu_notifier_unregister(&svm->notifier,
 svm->mm); list_del(&svm->list);
drivers/iommu/ioasid.c +0 -0
index 5f63af07acd5..f73b3dbfc37a 100644
@@ -1,22 +1,58 @@
   // SPDX-License-Identifier: GPL-2.0
   /*
    * I/O Address Space ID allocator. There is one global IOASID

space, split into
- * subsets. Users create a subset with DECLARE_IOASID_SET, then allocate and
- * free IOASIDs with ioasid_alloc and ioasid_free.
+ * subsets. Users create a subset with ioasid_alloc_set, then allocate/free IDs
+ * with ioasid_alloc and ioasid_free.

    */
drivers/iommu/ioasid.c +20 -1
-#include <linux/ioasid.h>
   #include <linux/module.h>
   #include <linux/slab.h>
   #include <linux/spinlock.h>
   #include <linux/xarray.h>
+#include <linux/ioasid.h>
+
+static DEFINE_XARRAY_ALLOC(ioasid_sets);
+enum ioasid_state {
+	IOASID_STATE_INACTIVE,
+	IOASID_STATE_ACTIVE,
+	IOASID_STATE_FREE_PENDING,
+};
 
+/**
+ * struct ioasid_data - Meta data about ioasid
+ *
+ * @id:		Unique ID
+ * @users	Number of active users
+ * @state	Track state of the IOASID
+ * @set		Meta data of the set this IOASID belongs to
+ * @private	Private data associated with the IOASID
+ * @rcu		For free after RCU grace period
+ */
   struct ioasid_data {
   	ioasid_t id;
   	struct ioasid_set *set;
+	refcount_t users;
+	enum ioasid_state state;
   	void *private;
   	struct rcu_head rcu;
   };
Quoted 11 lines

drivers/iommu/ioasid.c line 46

Quoted 8 lines
drivers/iommu/ioasid.c +8 -0
+/* Default to PCIe standard 20 bit PASID */
+#define PCI_PASID_MAX 0x100000
+static ioasid_t ioasid_capacity = PCI_PASID_MAX;
+static ioasid_t ioasid_capacity_avail = PCI_PASID_MAX;
+
+void ioasid_install_capacity(ioasid_t total)
+{
+	ioasid_capacity = ioasid_capacity_avail = total;

Need any check for multiple settings or setting after used?

Good point, capacity can only be set once. will add check to prevent setting after use.

Quoted 149 lines
Quoted 149 lines
drivers/iommu/ioasid.c +9 -0
+}
+EXPORT_SYMBOL_GPL(ioasid_install_capacity);
+
+ioasid_t ioasid_get_capacity()
+{
+	return ioasid_capacity;
+}
+EXPORT_SYMBOL_GPL(ioasid_get_capacity);
+
   /*
    * struct ioasid_allocator_data - Internal data structure to hold

information
* about an allocator. There are two types of allocators:

drivers/iommu/ioasid.c +4 -1
@@ -306,11 +342,23 @@ ioasid_t ioasid_alloc(struct ioasid_set *set,
 ioasid_t min, ioasid_t max, {
   	struct ioasid_data *data;
   	void *adata;
-	ioasid_t id;
+	ioasid_t id = INVALID_IOASID;
+
+	spin_lock(&ioasid_allocator_lock);
+	/* Check if the IOASID set has been allocated and

initialized */

drivers/iommu/ioasid.c +7 -0
+	if (WARN_ON(xa_load(&ioasid_sets, set->sid) != set)) {
+		pr_warn("Invalid set\n");
+		goto done_unlock;
+	}
+
+	if (set->quota <= set->nr_ioasids) {
+		pr_err("IOASID set %d out of quota %d\n",

set->sid, set->quota);

drivers/iommu/ioasid.c +3 -1
+		goto done_unlock;
+	}
 
   	data = kzalloc(sizeof(*data), GFP_ATOMIC);
   	if (!data)
-		return INVALID_IOASID;
+		goto done_unlock;
 
   	data->set = set;
   	data->private = private;
@@ -319,7 +367,6 @@ ioasid_t ioasid_alloc(struct ioasid_set *set,
 ioasid_t min, ioasid_t max,
   	 * Custom allocator needs allocator data to perform

platform specific
* operations.
*/

drivers/iommu/ioasid.c +0 -1
-	spin_lock(&ioasid_allocator_lock);
   	adata = active_allocator->flags &

IOASID_ALLOCATOR_CUSTOM ? active_allocator->ops->pdata : data; id = active_allocator->ops->alloc(min, max, adata); if (id == INVALID_IOASID) { @@ -335,42 +382,339 @@ ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, goto exit_free; }
data->id = id;

drivers/iommu/ioasid.c +6 -0
+	data->state = IOASID_STATE_ACTIVE;
+	refcount_set(&data->users, 1);
+
+	/* Store IOASID in the per set data */
+	if (xa_err(xa_store(&set->xa, id, data, GFP_ATOMIC))) {
+		pr_err("Failed to ioasid %d in set %d\n", id,

set->sid);

drivers/iommu/ioasid.c +13 -4
+		goto exit_free;
+	}
+	set->nr_ioasids++;
+	goto done_unlock;
 
-	spin_unlock(&ioasid_allocator_lock);
-	return id;
   exit_free:
-	spin_unlock(&ioasid_allocator_lock);
   	kfree(data);
-	return INVALID_IOASID;
+done_unlock:
+	spin_unlock(&ioasid_allocator_lock);
+	return id;
   }
   EXPORT_SYMBOL_GPL(ioasid_alloc);
 
+static void ioasid_do_free(struct ioasid_data *data)
+{
+	struct ioasid_data *ioasid_data;
+	struct ioasid_set *sdata;
+
+	active_allocator->ops->free(data->id,

active_allocator->ops->pdata);

drivers/iommu/ioasid.c +1 -0
+	/* Custom allocator needs additional steps to free the xa

element */

drivers/iommu/ioasid.c +2 -0
+	if (active_allocator->flags & IOASID_ALLOCATOR_CUSTOM) {
+		ioasid_data = xa_erase(&active_allocator->xa,

data->id);

drivers/iommu/ioasid.c +14 -0
+		kfree_rcu(ioasid_data, rcu);
+	}
+
+	sdata = xa_load(&ioasid_sets, data->set->sid);
+	if (!sdata) {
+		pr_err("No set %d for IOASID %d\n", data->set->sid,
+		       data->id);
+		return;
+	}
+	xa_erase(&sdata->xa, data->id);
+	sdata->nr_ioasids--;
+}
+
+static void ioasid_free_locked(struct ioasid_set *set, ioasid_t

ioasid) +{

drivers/iommu/ioasid.c +5 -0
+	struct ioasid_data *data;
+
+	data = xa_load(&active_allocator->xa, ioasid);
+	if (!data) {
+		pr_err("Trying to free unknown IOASID %u\n",

ioasid);

drivers/iommu/ioasid.c +5 -0
+		return;
+	}
+
+	if (data->set != set) {
+		pr_warn("Cannot free IOASID %u due to set

ownership\n", ioasid);

drivers/iommu/ioasid.c +11 -2
+		return;
+	}
+	data->state = IOASID_STATE_FREE_PENDING;
+
+	if (!refcount_dec_and_test(&data->users))
+		return;
+
+	ioasid_do_free(data);
+}
+
   /**
- * ioasid_free - Free an IOASID
- * @ioasid: the ID to remove
+ * ioasid_free - Drop reference on an IOASID. Free if refcount

drops to 0,
+ * including free from its set and system-wide list.
+ * @set: The ioasid_set to check permission with. If not
NULL, IOASID
+ * free will fail if the set does not match.
+ * @ioasid: The IOASID to remove

    */
drivers/iommu/ioasid.c +6 -2
-void ioasid_free(ioasid_t ioasid)
+void ioasid_free(struct ioasid_set *set, ioasid_t ioasid)
   {
-	struct ioasid_data *ioasid_data;
+	spin_lock(&ioasid_allocator_lock);
+	ioasid_free_locked(set, ioasid);
+	spin_unlock(&ioasid_allocator_lock);
+}
+EXPORT_SYMBOL_GPL(ioasid_free);
Quoted 8 lines
Quoted 4 lines

+/**
+ * ioasid_alloc_set - Allocate a new IOASID set for a given token
+ *
+ * @token: Unique token of the IOASID set, cannot be NULL

What's the use of @token? I might be able to find the answer in the code, but i have no idea when i comes here first time. :-)

You are right, I need to clarify here. How about:
* @token: An optional arbitrary number that can be associated
with the
* IOASID set. @token can be NULL if the type is
* IOASID_SET_TYPE_NULL
* @quota: Quota allowed in this set
* @type: The type of the token used to create the IOASID set

Quoted 4 lines

This line of comment says that token cannot be NULL. It seems not to match the real code, where token could be NULL if type is IOASID_SET_TYPE_NULL.

Good catch, it was a relic from previous versions.

Quoted 9 lines
Quoted 8 lines

+ * @quota: Quota allowed in this set. Only for new set
creation
+ * @flags: Special requirements
+ *
+ * IOASID can be limited system-wide resource that requires quota management.
+ * If caller does not wish to enforce quota, use IOASID_SET_NO_QUOTA flag.

Quoted 3 lines

If you are not going to add NO_QUOTA support this time, I'd suggest you to remove above comment.

Yes, will remove.

Quoted 10 lines
Quoted 7 lines

+ *
+ * Token will be stored in the ioasid_set returned. A reference will be taken
+ * upon finding a matching set or newly created set.
+ * IOASID allocation within the set and other per set operations will use
+ * the retured ioasid_set *.

nit: remove *, or you mean the ioasid_set pointer?

Sounds good

Quoted 26 lines
Quoted 22 lines

+ *
+ */
+struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota, int type) +{

drivers/iommu/ioasid.c +9 -0
+	struct ioasid_set *sdata;
+	unsigned long index;
+	ioasid_t id;
+
+	if (type >= IOASID_SET_TYPE_NR)
+		return ERR_PTR(-EINVAL);
+
+	/*
+	 * Need to check space available if we share system-wide

quota.

drivers/iommu/ioasid.c +1 -0
+	 * TODO: we may need to support quota free sets in the

future.

drivers/iommu/ioasid.c +1 -3
+	 */
   	spin_lock(&ioasid_allocator_lock);
-	ioasid_data = xa_load(&active_allocator->xa, ioasid);
-	if (!ioasid_data) {
-		pr_err("Trying to free unknown IOASID %u\n",

ioasid);

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 493

Quoted 1 line
drivers/iommu/ioasid.c +1 -0
+	if (quota > ioasid_capacity_avail) {

Thinking that ioasid_set itself needs an ioasid, so the check might be

Quoted 4 lines
Quoted 2 lines

(quota + 1 > ioasid_capacity_avail)?

ioasid_set does not need an ioasid, the set ID is allocated from a different XArray: ioasid_sets.

Get it now! Thanks!

Quoted 97 lines
Quoted 27 lines
Quoted 27 lines
drivers/iommu/ioasid.c +1 -0
+		pr_warn("Out of IOASID capacity! ask %d, avail

%d\n",

drivers/iommu/ioasid.c +2 -1
+			quota, ioasid_capacity_avail);
+		sdata = ERR_PTR(-ENOSPC);
   		goto exit_unlock;
   	}
 
-	active_allocator->ops->free(ioasid,

active_allocator->ops->pdata);

drivers/iommu/ioasid.c +0 -1
-	/* Custom allocator needs additional steps to free the xa

element */

drivers/iommu/ioasid.c +0 -2
-	if (active_allocator->flags & IOASID_ALLOCATOR_CUSTOM) {
-		ioasid_data = xa_erase(&active_allocator->xa,

ioasid);

drivers/iommu/ioasid.c +2 -1
-		kfree_rcu(ioasid_data, rcu);
+	/*
+	 * Token is only unique within its types but right now we

have only

drivers/iommu/ioasid.c +1 -0
+	 * mm type. If we have more token types, we have to match

type as well.

drivers/iommu/ioasid.c +6 -0
+	 */
+	switch (type) {
+	case IOASID_SET_TYPE_MM:
+		/* Search existing set tokens, reject duplicates */
+		xa_for_each(&ioasid_sets, index, sdata) {
+			if (sdata->token == token &&
Quoted 9 lines

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 513

Quoted 6 lines
drivers/iommu/ioasid.c +1 -0
+				sdata->type == IOASID_SET_TYPE_MM)

{

drivers/iommu/ioasid.c +4 -0
+				sdata = ERR_PTR(-EEXIST);
+				goto exit_unlock;
+			}
+		}

Do you need to enforce non-NULL token policy here?

yes, mm type cannot have NULL token. NULL type must have NULL token. This design has been flipped a couple of times. Thanks for catching this.

Quoted 4 lines

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 517

Quoted 4 lines
drivers/iommu/ioasid.c +4 -0
+		break;
+	case IOASID_SET_TYPE_NULL:
+		if (!token)
+			break;

should not be NULL

Quoted 51 lines
Quoted 39 lines
drivers/iommu/ioasid.c +6 -0
+		fallthrough;
+	default:
+		pr_err("Invalid token and IOASID type\n");
+		sdata = ERR_PTR(-EINVAL);
+		goto exit_unlock;
   	}
 
+	/* REVISIT: may support set w/o quota, use system

available */

drivers/iommu/ioasid.c +27 -0
+	if (!quota) {
+		sdata = ERR_PTR(-EINVAL);
+		goto exit_unlock;
+	}
+
+	sdata = kzalloc(sizeof(*sdata), GFP_ATOMIC);
+	if (!sdata) {
+		sdata = ERR_PTR(-ENOMEM);
+		goto exit_unlock;
+	}
+
+	if (xa_alloc(&ioasid_sets, &id, sdata,
+			XA_LIMIT(0, ioasid_capacity_avail - quota),
+			GFP_ATOMIC)) {
+		kfree(sdata);
+		sdata = ERR_PTR(-ENOSPC);
+		goto exit_unlock;
+	}
+
+	sdata->token = token;
+	sdata->type = type;
+	sdata->quota = quota;
+	sdata->sid = id;
+	refcount_set(&sdata->ref, 1);
+
+	/*
+	 * Per set XA is used to store private IDs within the set,

get ready

drivers/iommu/ioasid.c +1 -0
+	 * for ioasid_set private ID and system-wide IOASID

allocation

drivers/iommu/ioasid.c

Quoted 2 lines
drivers/iommu/ioasid.c +2 -0
+	 * results.
+	 */

I'm not sure that I understood the per-set XA correctly. Is it used to save both private ID and real ID allocated from the system-wide pool? If so, isn't private id might be equal to real ID?

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 557

Quoted 2 lines
drivers/iommu/ioasid.c +2 -0
+	xa_init_flags(&sdata->xa, XA_FLAGS_ALLOC);
+	ioasid_capacity_avail -= quota;

As mentioned above, the ioasid_set consumed one extra ioasid, so

Quoted 5 lines
Quoted 4 lines
	ioasid_capacity_avail -= (quota + 1);

?

Same explanation, make sense?

Sure.

Quoted 39 lines
Quoted 39 lines
Quoted 25 lines
drivers/iommu/ioasid.c +20 -1
+
   exit_unlock:
   	spin_unlock(&ioasid_allocator_lock);
+
+	return sdata;
   }
-EXPORT_SYMBOL_GPL(ioasid_free);
+EXPORT_SYMBOL_GPL(ioasid_alloc_set);
+
+void ioasid_set_get_locked(struct ioasid_set *set)
+{
+	if (WARN_ON(xa_load(&ioasid_sets, set->sid) != set)) {
+		pr_warn("Invalid set data\n");
+		return;
+	}
+
+	refcount_inc(&set->ref);
+}
+EXPORT_SYMBOL_GPL(ioasid_set_get_locked);
+
+void ioasid_set_get(struct ioasid_set *set)
+{
+	spin_lock(&ioasid_allocator_lock);
+	ioasid_set_get_locked(set);

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 590

Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+	spin_unlock(&ioasid_allocator_lock);
+}
+EXPORT_SYMBOL_GPL(ioasid_set_get);
+
+void ioasid_set_put_locked(struct ioasid_set *set)
+{
+	struct ioasid_data *entry;
+	unsigned long index;
+
+	if (WARN_ON(xa_load(&ioasid_sets, set->sid) != set)) {

There are multiple occurences of this line of code, how about defining it as a inline helper?

Quoted 7 lines
Quoted 5 lines

static inline bool ioasid_is_valid(struct ioasid_set *set)

{
	return xa_load(&ioasid_sets, set->sid) == set;
}

Sounds good. perhaps the name should be
ioasid_set_is_valid()?

Looks good to me.

Quoted 303 lines
Quoted 43 lines
Quoted 43 lines
drivers/iommu/ioasid.c +7 -0
+		pr_warn("Invalid set data\n");
+		return;
+	}
+
+	if (!refcount_dec_and_test(&set->ref)) {
+		pr_debug("%s: IOASID set %d has %d users\n",
+			__func__, set->sid,

refcount_read(&set->ref));

drivers/iommu/ioasid.c +9 -0
+		return;
+	}
+
+	/* The set is already empty, we just destroy the set. */
+	if (xa_empty(&set->xa))
+		goto done_destroy;
+
+	/*
+	 * Free all PASIDs from system-wide IOASID pool, all

subscribers gets

drivers/iommu/ioasid.c +2 -0
+	 * notified and do cleanup of their own.
+	 * Note that some references of the IOASIDs within the set

can still

drivers/iommu/ioasid.c +1 -0
+	 * be held after the free call. This is OK in that the

IOASIDs will be

drivers/iommu/ioasid.c +1 -0
+	 * marked inactive, the only operations can be done is

ioasid_put.

drivers/iommu/ioasid.c +1 -0
+	 * No need to track IOASID set states since there is no

reclaim phase.

drivers/iommu/ioasid.c +14 -0
+	 */
+	xa_for_each(&set->xa, index, entry) {
+		ioasid_free_locked(set, index);
+		/* Free from per set private pool */
+		xa_erase(&set->xa, index);
+	}
+
+done_destroy:
+	/* Return the quota back to system pool */
+	ioasid_capacity_avail += set->quota;
+	kfree_rcu(set, rcu);
+
+	/*
+	 * Token got released right away after the ioasid_set is

freed.

Quoted 9 lines

drivers/iommu/ioasid.c

Quoted 6 lines
drivers/iommu/ioasid.c +1 -0
+	 * If a new set is created immediately with the newly

released token,

drivers/iommu/ioasid.c +1 -0
+	 * it will not allocate the same IOASIDs unless they are

reclaimed.

drivers/iommu/ioasid.c +2 -0
+	 */
+	xa_erase(&ioasid_sets, set->sid);

No. pointer is used after free.

Right, will move it before kfree_rcu.

Quoted 51 lines
Quoted 41 lines
drivers/iommu/ioasid.c +5 -0
+}
+EXPORT_SYMBOL_GPL(ioasid_set_put_locked);
+
+/**
+ * ioasid_set_put - Drop a reference to the IOASID set. Free all

IOASIDs within
+ * the set if there are no more users.
+ *
+ * @set: The IOASID set ID to be freed
+ *
+ * If refcount drops to zero, all IOASIDs allocated within the set will be
+ * freed.
+ */

drivers/iommu/ioasid.c +14 -0
+void ioasid_set_put(struct ioasid_set *set)
+{
+	spin_lock(&ioasid_allocator_lock);
+	ioasid_set_put_locked(set);
+	spin_unlock(&ioasid_allocator_lock);
+}
+EXPORT_SYMBOL_GPL(ioasid_set_put);
+
+/**
+ * ioasid_adjust_set - Adjust the quota of an IOASID set
+ * @set:	IOASID set to be assigned
+ * @quota:	Quota allowed in this set
+ *
+ * Return 0 on success. If the new quota is smaller than the

number of
+ * IOASIDs already allocated, -EINVAL will be returned. No change will be
+ * made to the existing quota.
+ */

drivers/iommu/ioasid.c +7 -0
+int ioasid_adjust_set(struct ioasid_set *set, int quota)
+{
+	int ret = 0;
+
+	spin_lock(&ioasid_allocator_lock);
+	if (set->nr_ioasids > quota) {
+		pr_err("New quota %d is smaller than outstanding

IOASIDs %d\n",

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 671

Quoted 6 lines
drivers/iommu/ioasid.c +6 -0
+			quota, set->nr_ioasids);
+		ret = -EINVAL;
+		goto done_unlock;
+	}
+
+	if (quota >= ioasid_capacity_avail) {

This check doesn't make sense since you are updating (not asking for) a quota.

Quoted 3 lines

if ((quota > set->quota) &&

      (quota - set->quota > ioasid_capacity_avail))

Good point, will fix.

Thanks a lot!

Quoted 175 lines

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 677

Quoted 6 lines
drivers/iommu/ioasid.c +6 -0
+		ret = -ENOSPC;
+		goto done_unlock;
+	}
+
+	/* Return the delta back to system pool */
+	ioasid_capacity_avail += set->quota - quota;

ioasid_capacity_avail is defined as a unsigned int, hence this always increase the available capacity value even the caller is asking for a bigger quota?

Quoted 19 lines
drivers/iommu/ioasid.c +3 -0
+
+	/*
+	 * May have a policy to prevent giving all available

IOASIDs

drivers/iommu/ioasid.c +1 -0
+	 * to one set. But we don't enforce here, it should be in

the

drivers/iommu/ioasid.c +13 -0
+	 * upper layers.
+	 */
+	set->quota = quota;
+
+done_unlock:
+	spin_unlock(&ioasid_allocator_lock);
+
+	return ret;
+}
+EXPORT_SYMBOL_GPL(ioasid_adjust_set);
+
+/**
+ * ioasid_set_for_each_ioasid - Iterate over all the IOASIDs
Quoted 4 lines

within the set
+ *
+ * Caller must hold a reference of the set and handles its own locking.

Do you need to hold ioasid_allocator_lock here?

Quoted 138 lines

+ */

drivers/iommu/ioasid.c +20 -0
+int ioasid_set_for_each_ioasid(struct ioasid_set *set,
+			       void (*fn)(ioasid_t id, void *data),
+			       void *data)
+{
+	struct ioasid_data *entry;
+	unsigned long index;
+	int ret = 0;
+
+	if (xa_empty(&set->xa)) {
+		pr_warn("No IOASIDs in the set %d\n", set->sid);
+		return -ENOENT;
+	}
+
+	xa_for_each(&set->xa, index, entry) {
+		fn(index, data);
+	}
+
+	return ret;
+}
+EXPORT_SYMBOL_GPL(ioasid_set_for_each_ioasid);
 
   /**
    * ioasid_find - Find IOASID data
include/linux/ioasid.h +0 -0
index 9c44947a68c8..412d025d440e 100644
@@ -10,8 +10,35 @@ typedef unsigned int ioasid_t;
   typedef ioasid_t (*ioasid_alloc_fn_t)(ioasid_t min, ioasid_t max,

void *data); typedef void (*ioasid_free_fn_t)(ioasid_t ioasid, void *data);

include/linux/ioasid.h +14 -0
+/* IOASID set types */
+enum ioasid_set_type {
+	IOASID_SET_TYPE_NULL = 1, /* Set token is NULL */
+	IOASID_SET_TYPE_MM,	  /* Set token is a mm_struct,
+				   * i.e. associated with a process
+				   */
+	IOASID_SET_TYPE_NR,
+};
+
+/**
+ * struct ioasid_set - Meta data about ioasid_set
+ * @type:	Token types and other features
+ * @token:	Unique to identify an IOASID set
+ * @xa:		XArray to store ioasid_set private IDs, can

be used for
+ * guest-host IOASID mapping, or just a private
IOASID namespace.
+ * @quota: Max number of IOASIDs can be allocated within
the set
+ * @nr_ioasids Number of IOASIDs currently allocated in the
set
+ * @sid: ID of the set
+ * @ref: Reference count of the users
+ */
struct ioasid_set {

include/linux/ioasid.h +11 -3
-	int dummy;
+	void *token;
+	struct xarray xa;
+	int type;
+	int quota;
+	int nr_ioasids;
+	int sid;
+	refcount_t ref;
+	struct rcu_head rcu;
   };
 
   /**
@@ -29,31 +56,64 @@ struct ioasid_allocator_ops {
   	void *pdata;
   };
 
-#define DECLARE_IOASID_SET(name) struct ioasid_set name = { 0 }
-
   #if IS_ENABLED(CONFIG_IOASID)
+void ioasid_install_capacity(ioasid_t total);
+ioasid_t ioasid_get_capacity(void);
+struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota,

int type); +int ioasid_adjust_set(struct ioasid_set *set, int quota); +void ioasid_set_get_locked(struct ioasid_set *set);

include/linux/ioasid.h +3 -0
+void ioasid_set_put_locked(struct ioasid_set *set);
+void ioasid_set_put(struct ioasid_set *set);
+
   ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min,

ioasid_t max, void *private);

include/linux/ioasid.h +5 -3
-void ioasid_free(ioasid_t ioasid);
-void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid,
-		  bool (*getter)(void *));
+void ioasid_free(struct ioasid_set *set, ioasid_t ioasid);
+
+bool ioasid_is_active(ioasid_t ioasid);
+
+void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid, bool

(*getter)(void *)); +int ioasid_attach_data(ioasid_t ioasid, void *data); int ioasid_register_allocator(struct ioasid_allocator_ops *allocator); void ioasid_unregister_allocator(struct ioasid_allocator_ops *allocator); -int ioasid_attach_data(ioasid_t ioasid, void *data); -

include/linux/ioasid.h +13 -0
+void ioasid_is_in_set(struct ioasid_set *set, ioasid_t ioasid);
+int ioasid_set_for_each_ioasid(struct ioasid_set *sdata,
+			       void (*fn)(ioasid_t id, void *data),
+			       void *data);
   #else /* !CONFIG_IOASID */
+static inline void ioasid_install_capacity(ioasid_t total)
+{
+}
+
+static inline ioasid_t ioasid_get_capacity(void)
+{
+	return 0;
+}
+
   static inline ioasid_t ioasid_alloc(struct ioasid_set *set,

ioasid_t min, ioasid_t max, void *private)

   {
   	return INVALID_IOASID;
   }
include/linux/ioasid.h +1 -1
-static inline void ioasid_free(ioasid_t ioasid)
+static inline void ioasid_free(struct ioasid_set *set, ioasid_t

ioasid) +{

include/linux/ioasid.h +8 -0
+}
+
+static inline bool ioasid_is_active(ioasid_t ioasid)
+{
+	return false;
+}
+
+static inline struct ioasid_set *ioasid_alloc_set(void *token,

ioasid_t quota, int type) +{

include/linux/ioasid.h +4 -0
+	return ERR_PTR(-ENOTSUPP);
+}
+
+static inline void ioasid_set_put(struct ioasid_set *set)
   {
   }
Quoted 15 lines
Quoted 8 lines

-static inline void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid,

include/linux/ioasid.h +1 -1
-				bool (*getter)(void *))
+static inline void *ioasid_find(struct ioasid_set *set, ioasid_t
ioasid, bool (*getter)(void *)) {
   	return NULL;
   }

Best regards,
baolu
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

Quoted 2 lines

[Jacob Pan]

Best regards,
baolu
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

#23 lore raw

Re: [PATCH v2 3/9] iommu/ioasid: Introduce ioasid_set APIs

In reply to Jean-Philippe Brucker
Quoted 185 lines · On Mon, 24 Aug 2020 20:28:48 +0200 Jean-Philippe Brucker…
Quoted 184 lines · Jacob Pan

ioasid_set was introduced as an arbitrary token that are shared by a group of IOASIDs. For example, if IOASID #1 and #2 are allocated via the same ioasid_set*, they are viewed as to belong to the same set.

For guest SVA usages, system-wide IOASID resources need to be partitioned such that VMs can have its own quota and being managed separately. ioasid_set is the perfect candidate for meeting such requirements. This patch redefines and extends ioasid_set with the following new fields:
- Quota
- Reference count
- Storage of its namespace
- The token is stored in the new ioasid_set but with optional types

ioasid_set level APIs are introduced that wires up these new data. Existing users of IOASID APIs are converted where a host IOASID set is allocated for bare-metal usage.

---
 drivers/iommu/intel/iommu.c |  27 ++-
 drivers/iommu/intel/pasid.h |   1 +
 drivers/iommu/intel/svm.c   |   8 +-
 drivers/iommu/ioasid.c      | 390

+++++++++++++++++++++++++++++++++++++++++---
include/linux/ioasid.h | 82 ++++++++-- 5 files changed, 465
insertions(+), 43 deletions(-)

(quoted diff) +0 -0

b/drivers/iommu/intel/iommu.c index a3a0b5c8921d..5813eeaa5edb 100644 --- a/drivers/iommu/intel/iommu.c
+++ b/drivers/iommu/intel/iommu.c

drivers/iommu/intel/iommu.c +4 -0
@@ -42,6 +42,7 @@
 #include <linux/crash_dump.h>
 #include <linux/numa.h>
 #include <linux/swiotlb.h>
+#include <linux/ioasid.h>
 #include <asm/irq_remapping.h>
 #include <asm/cacheflush.h>
 #include <asm/iommu.h>
@@ -103,6 +104,9 @@
  */
 #define INTEL_IOMMU_PGSIZES	(~0xFFFUL)
 
+/* PASIDs used by host SVM */
+struct ioasid_set *host_pasid_set;
+
 static inline int agaw_to_level(int agaw)
 {
 	return agaw + 2;
@@ -3103,8 +3107,8 @@ static void intel_vcmd_ioasid_free(ioasid_t
 ioasid, void *data)
 	 * Sanity check the ioasid owner is done at upper layer,

e.g. VFIO
* We can only free the PASID when all the devices are
unbound. */

drivers/iommu/intel/iommu.c +2 -2
-	if (ioasid_find(NULL, ioasid, NULL)) {
-		pr_alert("Cannot free active IOASID %d\n", ioasid);
+	if (IS_ERR(ioasid_find(host_pasid_set, ioasid, NULL))) {
+		pr_err("Cannot free IOASID %d, not in system

set\n", ioasid); return;

drivers/iommu/intel/iommu.c +13 -0
 	}
 	vcmd_free_pasid(iommu, ioasid);
@@ -3288,6 +3292,19 @@ static int __init init_dmars(void)
 	if (ret)
 		goto free_iommu;
 
+	/* PASID is needed for scalable mode irrespective to SVM */
+	if (intel_iommu_sm) {
+		ioasid_install_capacity(intel_pasid_max_id);
+		/* We should not run out of IOASIDs at boot */
+		host_pasid_set = ioasid_alloc_set(NULL,
 PID_MAX_DEFAULT,
+
 IOASID_SET_TYPE_NULL);
+		if (IS_ERR_OR_NULL(host_pasid_set)) {
+			pr_err("Failed to enable host PASID
 allocator %lu\n",
+				PTR_ERR(host_pasid_set));
+			intel_iommu_sm = 0;
+		}
+	}
+
drivers/iommu/intel/iommu.c +3 -4
 	/*
 	 * for each drhd
 	 *   enable fault log
@@ -5149,7 +5166,7 @@ static void auxiliary_unlink_device(struct
 dmar_domain *domain, domain->auxd_refcnt--;
 
 	if (!domain->auxd_refcnt && domain->default_pasid > 0)
-		ioasid_free(domain->default_pasid);
+		ioasid_free(host_pasid_set, domain->default_pasid);
 }
 
 static int aux_domain_add_dev(struct dmar_domain *domain,
@@ -5167,7 +5184,7 @@ static int aux_domain_add_dev(struct
 dmar_domain *domain, int pasid;
 
 		/* No private data needed for the default pasid */
-		pasid = ioasid_alloc(NULL, PASID_MIN,
+		pasid = ioasid_alloc(host_pasid_set, PASID_MIN,
 				     pci_max_pasids(to_pci_dev(dev))
- 1, NULL);
 		if (pasid == INVALID_IOASID) {
@@ -5210,7 +5227,7 @@ static int aux_domain_add_dev(struct
 dmar_domain *domain, spin_unlock(&iommu->lock);
 	spin_unlock_irqrestore(&device_domain_lock, flags);
 	if (!domain->auxd_refcnt && domain->default_pasid > 0)
-		ioasid_free(domain->default_pasid);
+		ioasid_free(host_pasid_set, domain->default_pasid);
 
 	return ret;
 }
(quoted diff) +0 -0

b/drivers/iommu/intel/pasid.h index c9850766c3a9..ccdc23446015 100644 --- a/drivers/iommu/intel/pasid.h
+++ b/drivers/iommu/intel/pasid.h

drivers/iommu/intel/pasid.h +1 -0
@@ -99,6 +99,7 @@ static inline bool pasid_pte_is_present(struct
 pasid_entry *pte) }
 
 extern u32 intel_pasid_max_id;
+extern struct ioasid_set *host_pasid_set;
 int intel_pasid_alloc_id(void *ptr, int start, int end, gfp_t gfp);
 void intel_pasid_free_id(int pasid);
 void *intel_pasid_lookup_id(int pasid);
drivers/iommu/intel/svm.c +3 -3
index 37a9beabc0ca..634e191ca2c3 100644
@@ -551,7 +551,7 @@ intel_svm_bind_mm(struct device *dev, int
 flags, struct svm_dev_ops *ops, pasid_max = intel_pasid_max_id;
 
 		/* Do not use PASID 0, reserved for RID to PASID */
-		svm->pasid = ioasid_alloc(NULL, PASID_MIN,
+		svm->pasid = ioasid_alloc(host_pasid_set,
 PASID_MIN, pasid_max - 1, svm);
 		if (svm->pasid == INVALID_IOASID) {
 			kfree(svm);
@@ -568,7 +568,7 @@ intel_svm_bind_mm(struct device *dev, int
 flags, struct svm_dev_ops *ops, if (mm) {
 			ret =
 mmu_notifier_register(&svm->notifier, mm); if (ret) {
-				ioasid_free(svm->pasid);
+				ioasid_free(host_pasid_set,
 svm->pasid); kfree(svm);
 				kfree(sdev);
 				goto out;
@@ -586,7 +586,7 @@ intel_svm_bind_mm(struct device *dev, int
 flags, struct svm_dev_ops *ops, if (ret) {
 			if (mm)
 				mmu_notifier_unregister(&svm->notifier,
 mm);
-			ioasid_free(svm->pasid);
+			ioasid_free(host_pasid_set, svm->pasid);
 			kfree(svm);
 			kfree(sdev);
drivers/iommu/intel/svm.c +1 -1
 			goto out;
@@ -655,7 +655,7 @@ static int intel_svm_unbind_mm(struct device
 *dev, int pasid) kfree_rcu(sdev, rcu);
 
 			if (list_empty(&svm->devs)) {
-				ioasid_free(svm->pasid);
+				ioasid_free(host_pasid_set,
 svm->pasid); if (svm->mm)
 					mmu_notifier_unregister(&svm->notifier,
 svm->mm); list_del(&svm->list);
drivers/iommu/ioasid.c +0 -0
index 5f63af07acd5..f73b3dbfc37a 100644
@@ -1,22 +1,58 @@
 // SPDX-License-Identifier: GPL-2.0
 /*
  * I/O Address Space ID allocator. There is one global IOASID

space, split into
- * subsets. Users create a subset with DECLARE_IOASID_SET, then allocate and
- * free IOASIDs with ioasid_alloc and ioasid_free.
+ * subsets. Users create a subset with ioasid_alloc_set, then allocate/free IDs
+ * with ioasid_alloc and ioasid_free.
*/

Quoted 9 lines

drivers/iommu/ioasid.c line 11

Quoted 6 lines
drivers/iommu/ioasid.c +1 -1
-#include <linux/ioasid.h>
 #include <linux/module.h>
 #include <linux/slab.h>
 #include <linux/spinlock.h>
 #include <linux/xarray.h>
+#include <linux/ioasid.h>  

Spurious change (best keep the includes in alphabetical order)

I changed order so that I don't have to include more headers in ioasid.h, but it is better to keep the order and add headers to ioasid.h

Quoted 5 lines

drivers/iommu/ioasid.c line 13

Quoted 2 lines
drivers/iommu/ioasid.c +2 -0
+
+static DEFINE_XARRAY_ALLOC(ioasid_sets);  

I'd prefer keeping all static variables together

yes, same here :) will move.

Quoted 6 lines
Quoted 6 lines
drivers/iommu/ioasid.c +5 -0
+enum ioasid_state {
+	IOASID_STATE_INACTIVE,
+	IOASID_STATE_ACTIVE,
+	IOASID_STATE_FREE_PENDING,
+};
Quoted 12 lines
Quoted 9 lines

+/**
+ * struct ioasid_data - Meta data about ioasid
+ *
+ * @id: Unique ID
+ * @users Number of active users
+ * @state Track state of the IOASID
+ * @set Meta data of the set this IOASID belongs to
+ * @private Private data associated with the IOASID
+ * @rcu For free after RCU grace period

nit: it would be nicer to follow the struct order

yes, will do.

Quoted 37 lines
Quoted 37 lines

+ */
struct ioasid_data {

drivers/iommu/ioasid.c +19 -0
 	ioasid_t id;
 	struct ioasid_set *set;
+	refcount_t users;
+	enum ioasid_state state;
 	void *private;
 	struct rcu_head rcu;
 };
 
+/* Default to PCIe standard 20 bit PASID */
+#define PCI_PASID_MAX 0x100000
+static ioasid_t ioasid_capacity = PCI_PASID_MAX;
+static ioasid_t ioasid_capacity_avail = PCI_PASID_MAX;
+
+void ioasid_install_capacity(ioasid_t total)
+{
+	ioasid_capacity = ioasid_capacity_avail = total;
+}
+EXPORT_SYMBOL_GPL(ioasid_install_capacity);
+
+ioasid_t ioasid_get_capacity()
+{
+	return ioasid_capacity;
+}
+EXPORT_SYMBOL_GPL(ioasid_get_capacity);
+
 /*
  * struct ioasid_allocator_data - Internal data structure to hold

information
* about an allocator. There are two types of allocators:

drivers/iommu/ioasid.c +1 -1
@@ -306,11 +342,23 @@ ioasid_t ioasid_alloc(struct ioasid_set *set,
 ioasid_t min, ioasid_t max, {
 	struct ioasid_data *data;
 	void *adata;
-	ioasid_t id;
+	ioasid_t id = INVALID_IOASID;
Quoted 9 lines

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 350

Quoted 6 lines
drivers/iommu/ioasid.c +3 -0
+
+	spin_lock(&ioasid_allocator_lock);
+	/* Check if the IOASID set has been allocated and

initialized */

drivers/iommu/ioasid.c +2 -0
+	if (WARN_ON(xa_load(&ioasid_sets, set->sid) != set)) {
+		pr_warn("Invalid set\n");  

WARN_ON() is sufficient

right, will do

Quoted 4 lines
Quoted 4 lines
drivers/iommu/ioasid.c +4 -0
+		goto done_unlock;
+	}
+
+	if (set->quota <= set->nr_ioasids) {
Quoted 9 lines
Quoted 2 lines
drivers/iommu/ioasid.c +1 -0
+		pr_err("IOASID set %d out of quota %d\n",

set->sid, set->quota);

As this can be called directly by userspace via VFIO, I wonder if we should remove non-bug error messages like this one to avoid leaking internal IDs, or at least rate-limit them. We already have a few, perhaps we should deal with them before the VFIO_IOMMU_ALLOC_PASID patches land?

Good point, will hide internal IDs and add _ratelimited for VFIO possible call path.

Quoted 25 lines
Quoted 25 lines
drivers/iommu/ioasid.c +3 -1
+		goto done_unlock;
+	}
 
 	data = kzalloc(sizeof(*data), GFP_ATOMIC);
 	if (!data)
-		return INVALID_IOASID;
+		goto done_unlock;
 
 	data->set = set;
 	data->private = private;
@@ -319,7 +367,6 @@ ioasid_t ioasid_alloc(struct ioasid_set *set,
 ioasid_t min, ioasid_t max,
 	 * Custom allocator needs allocator data to perform

platform specific

drivers/iommu/ioasid.c +0 -1
 	 * operations.
 	 */
-	spin_lock(&ioasid_allocator_lock);
 	adata = active_allocator->flags &

IOASID_ALLOCATOR_CUSTOM ? active_allocator->ops->pdata : data; id = active_allocator->ops->alloc(min, max, adata); if (id == INVALID_IOASID) { @@ -335,42 +382,339 @@ ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, goto exit_free; }

drivers/iommu/ioasid.c +1 -0
 	data->id = id;
+	data->state = IOASID_STATE_ACTIVE;
Quoted 8 lines
Quoted 6 lines
drivers/iommu/ioasid.c +5 -0
+	refcount_set(&data->users, 1);
+
+	/* Store IOASID in the per set data */
+	if (xa_err(xa_store(&set->xa, id, data, GFP_ATOMIC))) {
+		pr_err("Failed to ioasid %d in set %d\n", id,

set->sid);

"Failed to store"

will fix

Quoted 2 lines

Don't we need to call active_allocator->ops->free()?

yes, also erase if custom allocator is used. How about:

	active_allocator->ops->free(id, active_allocator->ops->pdata);
	if (active_allocator->flags & IOASID_ALLOCATOR_CUSTOM)
		xa_erase(&active_allocator->xa, id);
Quoted 4 lines

And I need to think about this more, but do you see any issue with revoking here the data that we published into the xarray above through alloc()? We might need to free data in an RCU callback.

I don't see we need to do kfree_rcu since there shouldn't be any readers before the alloc function returns. Or I am missing something?

Quoted 33 lines
Quoted 33 lines
drivers/iommu/ioasid.c +13 -4
+		goto exit_free;
+	}
+	set->nr_ioasids++;
+	goto done_unlock;
 
-	spin_unlock(&ioasid_allocator_lock);
-	return id;
 exit_free:
-	spin_unlock(&ioasid_allocator_lock);
 	kfree(data);
-	return INVALID_IOASID;
+done_unlock:
+	spin_unlock(&ioasid_allocator_lock);
+	return id;
 }
 EXPORT_SYMBOL_GPL(ioasid_alloc);
 
+static void ioasid_do_free(struct ioasid_data *data)
+{
+	struct ioasid_data *ioasid_data;
+	struct ioasid_set *sdata;
+
+	active_allocator->ops->free(data->id,

active_allocator->ops->pdata);

drivers/iommu/ioasid.c +1 -0
+	/* Custom allocator needs additional steps to free the xa

element */

drivers/iommu/ioasid.c +2 -0
+	if (active_allocator->flags & IOASID_ALLOCATOR_CUSTOM) {
+		ioasid_data = xa_erase(&active_allocator->xa,

data->id);

drivers/iommu/ioasid.c +4 -0
+		kfree_rcu(ioasid_data, rcu);
+	}
+
+	sdata = xa_load(&ioasid_sets, data->set->sid);
Quoted 9 lines

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 420

Quoted 4 lines
drivers/iommu/ioasid.c +4 -0
+	if (!sdata) {
+		pr_err("No set %d for IOASID %d\n", data->set->sid,
+		       data->id);
+		return;  

I don't think we're allowed to fail at this point. If we need more sanity-check on the parameters, it should be before we start removing from the active_allocator above. Otherwise this should be a WARN

Good point, I will move this earlier as a WARN.

Quoted 7 lines

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 423

Quoted 3 lines
drivers/iommu/ioasid.c +3 -0
+	}
+	xa_erase(&sdata->xa, data->id);
+	sdata->nr_ioasids--;  

Would be nicer to perform the cleanup in the order opposite from ioasid_alloc()

Yes, will switch the order.

Quoted 47 lines
Quoted 47 lines
drivers/iommu/ioasid.c +3 -0
+}
+
+static void ioasid_free_locked(struct ioasid_set *set, ioasid_t

ioasid) +{

drivers/iommu/ioasid.c +5 -0
+	struct ioasid_data *data;
+
+	data = xa_load(&active_allocator->xa, ioasid);
+	if (!data) {
+		pr_err("Trying to free unknown IOASID %u\n",

ioasid);

drivers/iommu/ioasid.c +5 -0
+		return;
+	}
+
+	if (data->set != set) {
+		pr_warn("Cannot free IOASID %u due to set

ownership\n", ioasid);

drivers/iommu/ioasid.c +11 -2
+		return;
+	}
+	data->state = IOASID_STATE_FREE_PENDING;
+
+	if (!refcount_dec_and_test(&data->users))
+		return;
+
+	ioasid_do_free(data);
+}
+
 /**
- * ioasid_free - Free an IOASID
- * @ioasid: the ID to remove
+ * ioasid_free - Drop reference on an IOASID. Free if refcount

drops to 0,
+ * including free from its set and system-wide list.
+ * @set: The ioasid_set to check permission with. If not
NULL, IOASID
+ * free will fail if the set does not match.
+ * @ioasid: The IOASID to remove
*/

drivers/iommu/ioasid.c +6 -2
-void ioasid_free(ioasid_t ioasid)
+void ioasid_free(struct ioasid_set *set, ioasid_t ioasid)
 {
-	struct ioasid_data *ioasid_data;
+	spin_lock(&ioasid_allocator_lock);
+	ioasid_free_locked(set, ioasid);
+	spin_unlock(&ioasid_allocator_lock);
+}
+EXPORT_SYMBOL_GPL(ioasid_free);
Quoted 10 lines
Quoted 7 lines

+/**
+ * ioasid_alloc_set - Allocate a new IOASID set for a given token
+ *
+ * @token: Unique token of the IOASID set, cannot be NULL
+ * @quota: Quota allowed in this set. Only for new set
creation
+ * @flags: Special requirements

There is no @flags, but @type is missing

right, will fix.

Quoted 10 lines
Quoted 5 lines

+ *
+ * IOASID can be limited system-wide resource that requires quota management.
+ * If caller does not wish to enforce quota, use IOASID_SET_NO_QUOTA flag.

The flag isn't introduced in this patch. How about passing @quota == 0 in this case? For now I'm fine with leaving this as TODO and returning -EINVAL.

I will remove the flag for now. We can add @quota == 0 case later.

Quoted 66 lines
Quoted 57 lines

+ *
+ * Token will be stored in the ioasid_set returned. A reference will be taken
+ * upon finding a matching set or newly created set.
+ * IOASID allocation within the set and other per set operations will use
+ * the retured ioasid_set *.
+ *
+ */
+struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota, int type) +{

drivers/iommu/ioasid.c +9 -0
+	struct ioasid_set *sdata;
+	unsigned long index;
+	ioasid_t id;
+
+	if (type >= IOASID_SET_TYPE_NR)
+		return ERR_PTR(-EINVAL);
+
+	/*
+	 * Need to check space available if we share system-wide

quota.

drivers/iommu/ioasid.c +1 -0
+	 * TODO: we may need to support quota free sets in the

future.

drivers/iommu/ioasid.c +1 -3
+	 */
 	spin_lock(&ioasid_allocator_lock);
-	ioasid_data = xa_load(&active_allocator->xa, ioasid);
-	if (!ioasid_data) {
-		pr_err("Trying to free unknown IOASID %u\n",

ioasid);

drivers/iommu/ioasid.c +2 -0
+	if (quota > ioasid_capacity_avail) {
+		pr_warn("Out of IOASID capacity! ask %d, avail

%d\n",

drivers/iommu/ioasid.c +2 -1
+			quota, ioasid_capacity_avail);
+		sdata = ERR_PTR(-ENOSPC);
 		goto exit_unlock;
 	}
 
-	active_allocator->ops->free(ioasid,

active_allocator->ops->pdata);

drivers/iommu/ioasid.c +0 -1
-	/* Custom allocator needs additional steps to free the xa

element */

drivers/iommu/ioasid.c +0 -2
-	if (active_allocator->flags & IOASID_ALLOCATOR_CUSTOM) {
-		ioasid_data = xa_erase(&active_allocator->xa,

ioasid);

drivers/iommu/ioasid.c +2 -1
-		kfree_rcu(ioasid_data, rcu);
+	/*
+	 * Token is only unique within its types but right now we

have only

drivers/iommu/ioasid.c +1 -0
+	 * mm type. If we have more token types, we have to match

type as well.

drivers/iommu/ioasid.c +7 -0
+	 */
+	switch (type) {
+	case IOASID_SET_TYPE_MM:
+		/* Search existing set tokens, reject duplicates */
+		xa_for_each(&ioasid_sets, index, sdata) {
+			if (sdata->token == token &&
+				sdata->type == IOASID_SET_TYPE_MM)
Quoted 1 line

{

Should be aligned at the "if ("

According to the function doc, shouldn't we take a reference to the set in this case, and return it?
"A reference will be taken upon finding a matching set or newly created set."

Quoted 5 lines

However it might be better to separate the two functionalities into ioasid_alloc_set() and ioasid_find_set(). Because two modules can want to work on the same set for an mm, but they won't pass the same @quota, right? So it'd make more sense for one of them (VFIO) to alloc the set and the other to reuse it.

I went back and forth on this but yes, it is better to keep them separate. So far the only use case for find is ioasid_find_mm_set()
Will take reference during set allocation.

Quoted 16 lines
Quoted 16 lines
drivers/iommu/ioasid.c +13 -0
+				sdata = ERR_PTR(-EEXIST);
+				goto exit_unlock;
+			}
+		}
+		break;
+	case IOASID_SET_TYPE_NULL:
+		if (!token)
+			break;
+		fallthrough;
+	default:
+		pr_err("Invalid token and IOASID type\n");
+		sdata = ERR_PTR(-EINVAL);
+		goto exit_unlock;
 	}
Quoted 6 lines

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 526

Quoted 3 lines
drivers/iommu/ioasid.c +1 -0
+	/* REVISIT: may support set w/o quota, use system

available */

drivers/iommu/ioasid.c +1 -0
+	if (!quota) {  

Maybe move this next to the other quota check above

yes, that is better. no need to be under the spinlock either.

Quoted 7 lines
Quoted 7 lines
drivers/iommu/ioasid.c +7 -0
+		sdata = ERR_PTR(-EINVAL);
+		goto exit_unlock;
+	}
+
+	sdata = kzalloc(sizeof(*sdata), GFP_ATOMIC);
+	if (!sdata) {
+		sdata = ERR_PTR(-ENOMEM);
Quoted 9 lines
Quoted 6 lines
drivers/iommu/ioasid.c +5 -0
+		goto exit_unlock;
+	}
+
+	if (xa_alloc(&ioasid_sets, &id, sdata,
+			XA_LIMIT(0, ioasid_capacity_avail -

quota),

Why this limit? sid could just be an arbitrary u32 (xa_limit_32b)

True, but I was thinking the number of sets cannot be more than the number of available IOASIDs. XA allocator should be able to find empty slots within this range. Performance is better if the IDs are densely populated.

Quoted 14 lines
Quoted 14 lines
drivers/iommu/ioasid.c +14 -0
+			GFP_ATOMIC)) {
+		kfree(sdata);
+		sdata = ERR_PTR(-ENOSPC);
+		goto exit_unlock;
+	}
+
+	sdata->token = token;
+	sdata->type = type;
+	sdata->quota = quota;
+	sdata->sid = id;
+	refcount_set(&sdata->ref, 1);
+
+	/*
+	 * Per set XA is used to store private IDs within the set,
Quoted 9 lines

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 556

Quoted 6 lines

get ready

drivers/iommu/ioasid.c +1 -0
+	 * for ioasid_set private ID and system-wide IOASID

allocation

drivers/iommu/ioasid.c +3 -0
+	 * results.
+	 */
+	xa_init_flags(&sdata->xa, XA_FLAGS_ALLOC);  

Since it's only used for storage, you could use xa_init()

good point, will do.

Quoted 8 lines
Quoted 8 lines
drivers/iommu/ioasid.c +4 -1
+	ioasid_capacity_avail -= quota;
+
 exit_unlock:
 	spin_unlock(&ioasid_allocator_lock);
+
+	return sdata;
 }
-EXPORT_SYMBOL_GPL(ioasid_free);
Quoted 9 lines

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 569

Quoted 6 lines
drivers/iommu/ioasid.c +6 -0
+EXPORT_SYMBOL_GPL(ioasid_alloc_set);
+
+void ioasid_set_get_locked(struct ioasid_set *set)
+{
+	if (WARN_ON(xa_load(&ioasid_sets, set->sid) != set)) {
+		pr_warn("Invalid set data\n");  

WARN_ON() is sufficient

yes

Quoted 10 lines

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 575

Quoted 6 lines
drivers/iommu/ioasid.c +6 -0
+		return;
+	}
+
+	refcount_inc(&set->ref);
+}
+EXPORT_SYMBOL_GPL(ioasid_set_get_locked);  

Why is this function public, is it for an iterator? Might want to add a lockdep_assert_held() annotation.

You are right, no need to be public now.

Earlier, we had a design for users to acquire ioasid_set reference in the notifier handler (under lock). Since KVM may not know when or if the ioasid_set for its VM is created, we wanted KVM to listen to ioasid_set allocation event then register notifier block.
In the current design we have a notifier pending list so users such as KVM does not need to be aware of ioasid_set.

Quoted 10 lines
Quoted 10 lines
drivers/iommu/ioasid.c +10 -0
+
+void ioasid_set_get(struct ioasid_set *set)
+{
+	spin_lock(&ioasid_allocator_lock);
+	ioasid_set_get_locked(set);
+	spin_unlock(&ioasid_allocator_lock);
+}
+EXPORT_SYMBOL_GPL(ioasid_set_get);
+
+void ioasid_set_put_locked(struct ioasid_set *set)
Quoted 9 lines

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 591

Quoted 6 lines
drivers/iommu/ioasid.c +6 -0
+{
+	struct ioasid_data *entry;
+	unsigned long index;
+
+	if (WARN_ON(xa_load(&ioasid_sets, set->sid) != set)) {
+		pr_warn("Invalid set data\n");  

WARN_ON() is sufficient

got it

Quoted 43 lines
Quoted 43 lines
drivers/iommu/ioasid.c +6 -0
+		return;
+	}
+
+	if (!refcount_dec_and_test(&set->ref)) {
+		pr_debug("%s: IOASID set %d has %d users\n",
+			__func__, set->sid,

refcount_read(&set->ref));

drivers/iommu/ioasid.c +9 -0
+		return;
+	}
+
+	/* The set is already empty, we just destroy the set. */
+	if (xa_empty(&set->xa))
+		goto done_destroy;
+
+	/*
+	 * Free all PASIDs from system-wide IOASID pool, all

subscribers gets

drivers/iommu/ioasid.c +2 -0
+	 * notified and do cleanup of their own.
+	 * Note that some references of the IOASIDs within the set

can still

drivers/iommu/ioasid.c +1 -0
+	 * be held after the free call. This is OK in that the

IOASIDs will be

drivers/iommu/ioasid.c +1 -0
+	 * marked inactive, the only operations can be done is

ioasid_put.

drivers/iommu/ioasid.c +1 -0
+	 * No need to track IOASID set states since there is no

reclaim phase.

drivers/iommu/ioasid.c +14 -0
+	 */
+	xa_for_each(&set->xa, index, entry) {
+		ioasid_free_locked(set, index);
+		/* Free from per set private pool */
+		xa_erase(&set->xa, index);
+	}
+
+done_destroy:
+	/* Return the quota back to system pool */
+	ioasid_capacity_avail += set->quota;
+	kfree_rcu(set, rcu);
+
+	/*
+	 * Token got released right away after the ioasid_set is

freed.

drivers/iommu/ioasid.c +1 -0
+	 * If a new set is created immediately with the newly

released token,

Quoted 9 lines

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 631

Quoted 6 lines
drivers/iommu/ioasid.c +1 -0
+	 * it will not allocate the same IOASIDs unless they are

reclaimed.

drivers/iommu/ioasid.c +4 -0
+	 */
+	xa_erase(&ioasid_sets, set->sid);
+}
+EXPORT_SYMBOL_GPL(ioasid_set_put_locked);  

Same comment as ioasid_set_get_locked

same, will remove :)

Quoted 46 lines
Quoted 46 lines
drivers/iommu/ioasid.c +3 -0
+
+/**
+ * ioasid_set_put - Drop a reference to the IOASID set. Free all

IOASIDs within
+ * the set if there are no more users.
+ *
+ * @set: The IOASID set ID to be freed
+ *
+ * If refcount drops to zero, all IOASIDs allocated within the set will be
+ * freed.
+ */

drivers/iommu/ioasid.c +14 -0
+void ioasid_set_put(struct ioasid_set *set)
+{
+	spin_lock(&ioasid_allocator_lock);
+	ioasid_set_put_locked(set);
+	spin_unlock(&ioasid_allocator_lock);
+}
+EXPORT_SYMBOL_GPL(ioasid_set_put);
+
+/**
+ * ioasid_adjust_set - Adjust the quota of an IOASID set
+ * @set:	IOASID set to be assigned
+ * @quota:	Quota allowed in this set
+ *
+ * Return 0 on success. If the new quota is smaller than the

number of
+ * IOASIDs already allocated, -EINVAL will be returned. No change will be
+ * made to the existing quota.
+ */

drivers/iommu/ioasid.c +7 -0
+int ioasid_adjust_set(struct ioasid_set *set, int quota)
+{
+	int ret = 0;
+
+	spin_lock(&ioasid_allocator_lock);
+	if (set->nr_ioasids > quota) {
+		pr_err("New quota %d is smaller than outstanding

IOASIDs %d\n",

drivers/iommu/ioasid.c +7 -0
+			quota, set->nr_ioasids);
+		ret = -EINVAL;
+		goto done_unlock;
+	}
+
+	if (quota >= ioasid_capacity_avail) {
+		ret = -ENOSPC;
Quoted 9 lines

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 677

Quoted 5 lines
drivers/iommu/ioasid.c +5 -0
+		goto done_unlock;
+	}
+
+	/* Return the delta back to system pool */
+	ioasid_capacity_avail += set->quota - quota;  

I think this is correct as long as the above check is fixed (as pointed out by Baolu). A check that quota >= 0 could be nice too.

yes, will fix per Baolu.
Will add a check for quota > 0 but 0 is not supported yet.

Quoted 27 lines
Quoted 27 lines
drivers/iommu/ioasid.c +3 -0
+
+	/*
+	 * May have a policy to prevent giving all available

IOASIDs

drivers/iommu/ioasid.c +1 -0
+	 * to one set. But we don't enforce here, it should be in

the

drivers/iommu/ioasid.c +13 -0
+	 * upper layers.
+	 */
+	set->quota = quota;
+
+done_unlock:
+	spin_unlock(&ioasid_allocator_lock);
+
+	return ret;
+}
+EXPORT_SYMBOL_GPL(ioasid_adjust_set);
+
+/**
+ * ioasid_set_for_each_ioasid - Iterate over all the IOASIDs

within the set
+ *
+ * Caller must hold a reference of the set and handles its own locking.
+ */

drivers/iommu/ioasid.c +3 -0
+int ioasid_set_for_each_ioasid(struct ioasid_set *set,
+			       void (*fn)(ioasid_t id, void *data),
+			       void *data)
Quoted 8 lines

drivers/iommu/ioasid.c ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max, · line 705

Quoted 6 lines
drivers/iommu/ioasid.c +6 -0
+{
+	struct ioasid_data *entry;
+	unsigned long index;
+	int ret = 0;
+
[… more lines in the raw message]

Lines of this file are not shown. View the raw message.

The rest of this diff is not shown. View the raw message.

Who calls this function?

This is called by Yi's VFIO PASID code. Not sure when he will send out but should be soon. The use case is when a guest terminates, VFIO can call this function to free all PASIDs for this guest.

Quoted 5 lines

Empty xa may be a normal use-case if the
caller just uses it for sweeping, so pr_warn() could be problematic. The returned value also isn't particularly accurate if concurrent ioasid_alloc/free are allowed, so I'd drop this.

Right, xa could be empty if the guest never allocated PASIDS. Will drop the check and make the function void.

Quoted 28 lines
Quoted 28 lines
drivers/iommu/ioasid.c +11 -0

Lines of this file are not shown. View the raw message.

include/linux/ioasid.h +0 -0
index 9c44947a68c8..412d025d440e 100644

Lines of this file are not shown. View the raw message.

The rest of this diff is not shown. View the raw message.

void *data); typedef void (*ioasid_free_fn_t)(ioasid_t ioasid, void *data);

include/linux/ioasid.h +6 -0

Lines of this file are not shown. View the raw message.

The rest of this diff is not shown. View the raw message.

Quoted 9 lines

include/linux/ioasid.h

Quoted 6 lines
include/linux/ioasid.h +6 -0

Lines of this file are not shown. View the raw message.

The rest of this diff is not shown. View the raw message.

nit: doesn't follow struct order

will fix.

Quoted 29 lines
Quoted 29 lines

13 lines not shown · raw message

include/linux/ioasid.h +8 -1

Lines of this file are not shown. View the raw message.

The rest of this diff is not shown. View the raw message.

Quoted 13 lines

11 lines not shown · raw message

103 lines not shown · raw message

Parts of this message are not shown here. View the raw message.

#20 lore raw

Re: [PATCH v2 3/9] iommu/ioasid: Introduce ioasid_set APIs

In reply to Randy Dunlap
Quoted 10 lines · On Mon, 24 Aug 2020 11:30:47 -0700 Randy Dunlap <[email protected]> wrote:
Quoted 9 lines · Jean-Philippe Brucker
Quoted 9 lines

+/**
+ * struct ioasid_data - Meta data about ioasid
+ *
+ * @id: Unique ID
+ * @users Number of active users
+ * @state Track state of the IOASID
+ * @set Meta data of the set this IOASID belongs to
+ * @private Private data associated with the IOASID
+ * @rcu For free after RCU grace period

Quoted 4 lines
Quoted 1 line

nit: it would be nicer to follow the struct order

and use a ':' after each struct member name, as is done for @id:

Got it, thanks.

#22 lore raw

Re: [PATCH v2 3/9] iommu/ioasid: Introduce ioasid_set APIs

In reply to Randy Dunlap
Quoted 6 lines · On Mon, 24 Aug 2020 11:34:29 -0700 Randy Dunlap <[email protected]> wrote:
Quoted 5 lines · Jean-Philippe Brucker
Quoted 3 lines

+/**
+ * struct ioasid_set - Meta data about ioasid_set
+ * @type: Token types and other features

nit: doesn't follow struct order

Quoted 8 lines
Quoted 5 lines
Quoted 5 lines

+ * @token: Unique to identify an IOASID set
+ * @xa: XArray to store ioasid_set private IDs, can be used for
+ * guest-host IOASID mapping, or just a private IOASID namespace.
+ * @quota: Max number of IOASIDs can be allocated within the set
+ * @nr_ioasids Number of IOASIDs currently allocated in the set

* @nr_ioasids: Number of IOASIDs currently allocated in the set

got it. thanks!

Quoted 6 lines
Quoted 6 lines
Quoted 6 lines

+ * @sid: ID of the set
+ * @ref: Reference count of the users
+ */
struct ioasid_set {

include/linux/ioasid.h +1 -1
-	int dummy;
+	void *token;
Quoted 10 lines
Quoted 8 lines

include/linux/ioasid.h typedef unsigned int ioasid_t; · line 42

Quoted 8 lines
include/linux/ioasid.h +7 -0
+	struct xarray xa;
+	int type;
+	int quota;
+	int nr_ioasids;
+	int sid;
+	refcount_t ref;
+	struct rcu_head rcu;
 };  

[Jacob Pan]
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu