Skip to contentSkip to conversation

Enable shared device assignment

PATCH Superseded Chenyi Qiang 5 participants 31 messages Mar 10, 2025 – Mar 31, 2025

Subsystems (2)

v3 of 7 Superseded Latest is v7 →

lore
Share & tools
lore
#0 lore raw

v3 cover Enable shared device assignment

This is the v3 series of the shared device assignment support.

The overview of this series:
- Patch 1-3: preparation patches. The function exposure and some defintion change to return values.
- Patch 4-5: Introduce a new object to implement RamDiscardManager interface and a callback to notify the shared/private state change.
- Patch 6: Store the new object including guest_memfd information in RAMBlock. Register the RamDiscardManager instance to the target RAMBlock's MemoryRegion so that the object can notify the page conversion events to other systems.
- Patch 7: Unlock the coordinate discard so that the shared device assignment (VFIO) can work with guest_memfd.

Compared with v2 series, the main changes are:

- Introduce the new patch 04 to unify the definiton of ReplayRamPopulate() and ReplayRamDiscard().
- In state_change() callback, fallback to a "1 block at a time" handling in case of mixed state. In addition, change the bitmap before calling the
listener so that the listener can do the change according to the
latest status.
- In kvm_convert_memory(), use ReplayPopulated() and ReplayDiscard() interface to trigger set_attribute() and add the undo operation if state_change() failed.
- v2: https://lore.kernel.org/qemu-devel/[email protected]/

More small changes or details can be found in the individual patches.

---
Original cover letter:

Background
==========
Confidential VMs have two classes of memory: shared and private memory. Shared memory is accessible from the host/VMM while private memory is not. Confidential VMs can decide which memory is shared/private and convert memory between shared/private at runtime.

"guest_memfd" is a new kind of fd whose primary goal is to serve guest private memory. In current implementation, shared memory is allocated with normal methods (e.g. mmap or fallocate) while private memory is allocated from guest_memfd. When a VM performs memory conversions, QEMU frees pages via madvise or via PUNCH_HOLE on memfd or guest_memfd from one side, and allocates new pages from the other side. This will cause a stale IOMMU mapping issue mentioned in [1] when we try to enable shared device assignment in confidential VMs.

Solution
========
The key to enable shared device assignment is to update the IOMMU mappings on page conversion. RamDiscardManager, an existing interface currently utilized by virtio-mem, offers a means to modify IOMMU mappings in accordance with VM page assignment. Page conversion is similar to hot-removing a page in one mode and adding it back in the other.

This series implements a RamDiscardManager for confidential VMs and utilizes its infrastructure to notify VFIO of page conversions.

Relationship with in-place page conversion
==========================================
To support 1G page support for guest_memfd [2], the current direction is to allow mmap() of guest_memfd to userspace so that both private and shared memory can use the same physical pages as the backend. This in-place page conversion design eliminates the need to discard pages during shared/private conversions. However, device assignment will still be blocked because the in-place page conversion will reject the conversion when the page is pinned by VFIO.

To address this, the key difference lies in the sequence of VFIO map/unmap operations and the page conversion. It can be adjusted to achieve unmap-before-conversion-to-private and map-after-conversion-to-shared, ensuring compatibility with guest_memfd.

Limitation
==========
One limitation is that VFIO expects the DMA mapping for a specific IOVA to be mapped and unmapped with the same granularity. The guest may perform partial conversions, such as converting a small region within a larger region. To prevent such invalid cases, all operations are performed with 4K granularity. This could be optimized after the cut_mapping operation [3] is introduced in future. We can alway perform a split-before-unmap if partial conversions happen. If the split succeeds, the unmap will succeed and be atomic. If the split fails, the unmap process fails.

Testing
=======
This patch series is tested based on TDX patches available at:
KVM: https://github.com/intel/tdx/tree/kvm-coco-queue-snapshot/kvm-coco-queue-snapshot-20250308
QEMU: https://github.com/intel-staging/qemu-tdx/tree/tdx-upstream-snapshot-2025-03-10

To facilitate shared device assignment with the NIC, employ the legacy type1 VFIO with the QEMU command:

qemu-system-x86_64 [...]
-device vfio-pci,host=XX:XX.X

The parameter of dma_entry_limit needs to be adjusted. For example, a 16GB guest needs to adjust the parameter like vfio_iommu_type1.dma_entry_limit=4194304.

If use the iommufd-backed VFIO with the qemu command:

qemu-system-x86_64 [...]

    -object iommufd,id=iommufd0 \
    -device vfio-pci,host=XX:XX.X,iommufd=iommufd0

No additional adjustment required.

Following the bootup of the TD guest, the guest's IP address becomes visible, and iperf is able to successfully send and receive data.

Related link
============
[1] https://lore.kernel.org/qemu-devel/[email protected]/
[2] https://lore.kernel.org/lkml/[email protected]/
[3] https://lore.kernel.org/linux-iommu/[email protected]/

Chenyi Qiang (7):
memory: Export a helper to get intersection of a MemoryRegionSection
with a given range
memory: Change memory_region_set_ram_discard_manager() to return the
result
memory: Unify the definiton of ReplayRamPopulate() and
ReplayRamDiscard()
memory-attribute-manager: Introduce MemoryAttributeManager to manage
RAMBLock with guest_memfd
memory-attribute-manager: Introduce a callback to notify the
shared/private state change
memory: Attach MemoryAttributeManager to guest_memfd-backed RAMBlocks
RAMBlock: Make guest_memfd require coordinate discard

 accel/kvm/kvm-all.c                       |  50 ++-
 hw/virtio/virtio-mem.c                    |  81 ++--
 include/exec/memory.h                     |  64 ++-
 include/exec/ramblock.h                   |   2 +
 include/system/memory-attribute-manager.h |  60 +++
 migration/ram.c                           |   5 +-
 system/memory-attribute-manager.c         | 471 ++++++++++++++++++++++
 system/memory.c                           |  22 +-
 system/meson.build                        |   1 +
 system/physmem.c                          |  17 +-
 10 files changed, 690 insertions(+), 83 deletions(-)
 create mode 100644 include/system/memory-attribute-manager.h
 create mode 100644 system/memory-attribute-manager.c
Signature · 1 line
2.43.5
#1 lore raw

PATCH v3 1/7 memory: Export a helper to get intersection of a MemoryRegionSection with a given range

Rename the helper to memory_region_section_intersect_range() to make it more generic. Meanwhile, define the @end as Int128 and replace the related operations with Int128_* format since the helper is exported as a wider API.

---
Changes in v3:
- No change

Changes in v2:
- Make memory_region_section_intersect_range() an inline function.
- Add Reviewed-by from David
- Define the @end as Int128 and use the related Int128_* ops as a wilder API (Alexey)

---
 hw/virtio/virtio-mem.c | 32 +++++---------------------------
 include/exec/memory.h  | 27 +++++++++++++++++++++++++++
 2 files changed, 32 insertions(+), 27 deletions(-)
hw/virtio/virtio-mem.c +5 -27
index b1a003736b..21f16e4912 100644
@@ -244,28 +244,6 @@ static int virtio_mem_for_each_plugged_range(VirtIOMEM *vmem, void *arg,
     return ret;
 }
 
-/*
- * Adjust the memory section to cover the intersection with the given range.
- *
- * Returns false if the intersection is empty, otherwise returns true.
- */
-static bool virtio_mem_intersect_memory_section(MemoryRegionSection *s,
-                                                uint64_t offset, uint64_t size)
-{
-    uint64_t start = MAX(s->offset_within_region, offset);
-    uint64_t end = MIN(s->offset_within_region + int128_get64(s->size),
-                       offset + size);
-
-    if (end <= start) {
-        return false;
-    }
-
-    s->offset_within_address_space += start - s->offset_within_region;
-    s->offset_within_region = start;
-    s->size = int128_make64(end - start);
-    return true;
-}
-
 typedef int (*virtio_mem_section_cb)(MemoryRegionSection *s, void *arg);
 
 static int virtio_mem_for_each_plugged_section(const VirtIOMEM *vmem,
@@ -287,7 +265,7 @@ static int virtio_mem_for_each_plugged_section(const VirtIOMEM *vmem,
                                       first_bit + 1) - 1;
         size = (last_bit - first_bit + 1) * vmem->block_size;
 
-        if (!virtio_mem_intersect_memory_section(&tmp, offset, size)) {
+        if (!memory_region_section_intersect_range(&tmp, offset, size)) {
             break;
         }
         ret = cb(&tmp, arg);
@@ -319,7 +297,7 @@ static int virtio_mem_for_each_unplugged_section(const VirtIOMEM *vmem,
                                  first_bit + 1) - 1;
         size = (last_bit - first_bit + 1) * vmem->block_size;
 
-        if (!virtio_mem_intersect_memory_section(&tmp, offset, size)) {
+        if (!memory_region_section_intersect_range(&tmp, offset, size)) {
             break;
         }
         ret = cb(&tmp, arg);
@@ -355,7 +333,7 @@ static void virtio_mem_notify_unplug(VirtIOMEM *vmem, uint64_t offset,
     QLIST_FOREACH(rdl, &vmem->rdl_list, next) {
         MemoryRegionSection tmp = *rdl->section;
 
-        if (!virtio_mem_intersect_memory_section(&tmp, offset, size)) {
+        if (!memory_region_section_intersect_range(&tmp, offset, size)) {
             continue;
         }
         rdl->notify_discard(rdl, &tmp);
@@ -371,7 +349,7 @@ static int virtio_mem_notify_plug(VirtIOMEM *vmem, uint64_t offset,
     QLIST_FOREACH(rdl, &vmem->rdl_list, next) {
         MemoryRegionSection tmp = *rdl->section;
 
-        if (!virtio_mem_intersect_memory_section(&tmp, offset, size)) {
+        if (!memory_region_section_intersect_range(&tmp, offset, size)) {
             continue;
         }
         ret = rdl->notify_populate(rdl, &tmp);
@@ -388,7 +366,7 @@ static int virtio_mem_notify_plug(VirtIOMEM *vmem, uint64_t offset,
             if (rdl2 == rdl) {
                 break;
             }
-            if (!virtio_mem_intersect_memory_section(&tmp, offset, size)) {
+            if (!memory_region_section_intersect_range(&tmp, offset, size)) {
                 continue;
             }
             rdl2->notify_discard(rdl2, &tmp);
include/exec/memory.h +27 -0
index 3ee1901b52..3bebc43d59 100644
@@ -1202,6 +1202,33 @@ MemoryRegionSection *memory_region_section_new_copy(MemoryRegionSection *s);
  */
 void memory_region_section_free_copy(MemoryRegionSection *s);
 
+/**
+ * memory_region_section_intersect_range: Adjust the memory section to cover
+ * the intersection with the given range.
+ *
+ * @s: the #MemoryRegionSection to be adjusted
+ * @offset: the offset of the given range in the memory region
+ * @size: the size of the given range
+ *
+ * Returns false if the intersection is empty, otherwise returns true.
+ */
+static inline bool memory_region_section_intersect_range(MemoryRegionSection *s,
+                                                         uint64_t offset, uint64_t size)
+{
+    uint64_t start = MAX(s->offset_within_region, offset);
+    Int128 end = int128_min(int128_add(int128_make64(s->offset_within_region), s->size),
+                            int128_add(int128_make64(offset), int128_make64(size)));
+
+    if (int128_le(end, int128_make64(start))) {
+        return false;
+    }
+
+    s->offset_within_address_space += start - s->offset_within_region;
+    s->offset_within_region = start;
+    s->size = int128_sub(end, int128_make64(start));
+    return true;
+}
+
 /**
  * memory_region_init: Initialize a memory region
  *
Signature · 1 line
2.43.5
#2 lore raw

PATCH v3 2/7 memory: Change memory_region_set_ram_discard_manager() to return the result

In reply to Chenyi Qiang

Modify memory_region_set_ram_discard_manager() to return false if a RamDiscardManager is already set in the MemoryRegion. The caller must handle this failure, such as having virtio-mem undo its actions and fail the realize() process. Opportunistically move the call earlier to avoid complex error handling.

This change is beneficial when introducing a new RamDiscardManager instance besides virtio-mem. After ram_block_coordinated_discard_require(true) unlocks all RamDiscardManager instances, only one instance is allowed to be set for a MemoryRegion at present.

---
Changes in v3:
- Move set_ram_discard_manager() up to avoid a g_free()
- Clean up set_ram_discard_manager() definition

Changes in v2:
- newly added.

---
 hw/virtio/virtio-mem.c | 29 ++++++++++++++++-------------
 include/exec/memory.h  |  6 +++---
 system/memory.c        | 10 +++++++---
 3 files changed, 26 insertions(+), 19 deletions(-)
hw/virtio/virtio-mem.c +16 -13
index 21f16e4912..d0d3a0240f 100644
@@ -1049,6 +1049,17 @@ static void virtio_mem_device_realize(DeviceState *dev, Error **errp)
         return;
     }
 
+    /*
+     * Set ourselves as RamDiscardManager before the plug handler maps the
+     * memory region and exposes it via an address space.
+     */
+    if (memory_region_set_ram_discard_manager(&vmem->memdev->mr,
+                                              RAM_DISCARD_MANAGER(vmem))) {
+        error_setg(errp, "Failed to set RamDiscardManager");
+        ram_block_coordinated_discard_require(false);
+        return;
+    }
+
     /*
      * We don't know at this point whether shared RAM is migrated using
      * QEMU or migrated using the file content. "x-ignore-shared" will be
@@ -1124,13 +1135,6 @@ static void virtio_mem_device_realize(DeviceState *dev, Error **errp)
     vmem->system_reset = VIRTIO_MEM_SYSTEM_RESET(obj);
     vmem->system_reset->vmem = vmem;
     qemu_register_resettable(obj);
-
-    /*
-     * Set ourselves as RamDiscardManager before the plug handler maps the
-     * memory region and exposes it via an address space.
-     */
-    memory_region_set_ram_discard_manager(&vmem->memdev->mr,
-                                          RAM_DISCARD_MANAGER(vmem));
 }
 
 static void virtio_mem_device_unrealize(DeviceState *dev)
@@ -1138,12 +1142,6 @@ static void virtio_mem_device_unrealize(DeviceState *dev)
     VirtIODevice *vdev = VIRTIO_DEVICE(dev);
     VirtIOMEM *vmem = VIRTIO_MEM(dev);
 
-    /*
-     * The unplug handler unmapped the memory region, it cannot be
-     * found via an address space anymore. Unset ourselves.
-     */
-    memory_region_set_ram_discard_manager(&vmem->memdev->mr, NULL);
-
     qemu_unregister_resettable(OBJECT(vmem->system_reset));
     object_unref(OBJECT(vmem->system_reset));
 
@@ -1156,6 +1154,11 @@ static void virtio_mem_device_unrealize(DeviceState *dev)
     virtio_del_queue(vdev, 0);
     virtio_cleanup(vdev);
     g_free(vmem->bitmap);
+    /*
+     * The unplug handler unmapped the memory region, it cannot be
+     * found via an address space anymore. Unset ourselves.
+     */
+    memory_region_set_ram_discard_manager(&vmem->memdev->mr, NULL);
     ram_block_coordinated_discard_require(false);
 }
 
include/exec/memory.h +3 -3
index 3bebc43d59..390477b588 100644
@@ -2487,13 +2487,13 @@ static inline bool memory_region_has_ram_discard_manager(MemoryRegion *mr)
  *
  * This function must not be called for a mapped #MemoryRegion, a #MemoryRegion
  * that does not cover RAM, or a #MemoryRegion that already has a
- * #RamDiscardManager assigned.
+ * #RamDiscardManager assigned. Return 0 if the rdm is set successfully.
  *
  * @mr: the #MemoryRegion
  * @rdm: #RamDiscardManager to set
  */
-void memory_region_set_ram_discard_manager(MemoryRegion *mr,
-                                           RamDiscardManager *rdm);
+int memory_region_set_ram_discard_manager(MemoryRegion *mr,
+                                          RamDiscardManager *rdm);
 
 /**
  * memory_region_find: translate an address/size relative to a
system/memory.c +7 -3
index b17b5538ff..62d6b410f0 100644
@@ -2115,12 +2115,16 @@ RamDiscardManager *memory_region_get_ram_discard_manager(MemoryRegion *mr)
     return mr->rdm;
 }
 
-void memory_region_set_ram_discard_manager(MemoryRegion *mr,
-                                           RamDiscardManager *rdm)
+int memory_region_set_ram_discard_manager(MemoryRegion *mr,
+                                          RamDiscardManager *rdm)
 {
     g_assert(memory_region_is_ram(mr));
-    g_assert(!rdm || !mr->rdm);
+    if (mr->rdm && rdm) {
+        return -EBUSY;
+    }
+
     mr->rdm = rdm;
+    return 0;
 }
 
 uint64_t ram_discard_manager_get_min_granularity(const RamDiscardManager *rdm,
Signature · 1 line
2.43.5
#3 lore raw

PATCH v3 3/7 memory: Unify the definiton of ReplayRamPopulate() and ReplayRamDiscard()

In reply to Chenyi Qiang

Update ReplayRamDiscard() function to return the result and unify the ReplayRamPopulate() and ReplayRamDiscard() to ReplayRamStateChange() at the same time due to their identical definitions. This unification simplifies related structures, such as VirtIOMEMReplayData, which makes it more cleaner and maintainable. It also paves the way for future extension when introducing new ReplayRamDiscard() functions to return the results.

---
Changes in v3:
- Newly added.

---
 hw/virtio/virtio-mem.c | 20 ++++++++++----------
 include/exec/memory.h  | 31 ++++++++++++++++---------------
 migration/ram.c        |  5 +++--
 system/memory.c        | 12 ++++++------
 4 files changed, 35 insertions(+), 33 deletions(-)
hw/virtio/virtio-mem.c +10 -10
index d0d3a0240f..816ae8abdd 100644
@@ -1733,7 +1733,7 @@ static bool virtio_mem_rdm_is_populated(const RamDiscardManager *rdm,
 }
 
 struct VirtIOMEMReplayData {
-    void *fn;
+    ReplayRamStateChange fn;
     void *opaque;
 };
 
@@ -1741,12 +1741,12 @@ static int virtio_mem_rdm_replay_populated_cb(MemoryRegionSection *s, void *arg)
 {
     struct VirtIOMEMReplayData *data = arg;
 
-    return ((ReplayRamPopulate)data->fn)(s, data->opaque);
+    return data->fn(s, data->opaque);
 }
 
 static int virtio_mem_rdm_replay_populated(const RamDiscardManager *rdm,
                                            MemoryRegionSection *s,
-                                           ReplayRamPopulate replay_fn,
+                                           ReplayRamStateChange replay_fn,
                                            void *opaque)
 {
     const VirtIOMEM *vmem = VIRTIO_MEM(rdm);
@@ -1765,14 +1765,14 @@ static int virtio_mem_rdm_replay_discarded_cb(MemoryRegionSection *s,
 {
     struct VirtIOMEMReplayData *data = arg;
 
-    ((ReplayRamDiscard)data->fn)(s, data->opaque);
+    data->fn(s, data->opaque);
     return 0;
 }
 
-static void virtio_mem_rdm_replay_discarded(const RamDiscardManager *rdm,
-                                            MemoryRegionSection *s,
-                                            ReplayRamDiscard replay_fn,
-                                            void *opaque)
+static int virtio_mem_rdm_replay_discarded(const RamDiscardManager *rdm,
+                                           MemoryRegionSection *s,
+                                           ReplayRamStateChange replay_fn,
+                                           void *opaque)
 {
     const VirtIOMEM *vmem = VIRTIO_MEM(rdm);
     struct VirtIOMEMReplayData data = {
@@ -1781,8 +1781,8 @@ static void virtio_mem_rdm_replay_discarded(const RamDiscardManager *rdm,
     };
 
     g_assert(s->mr == &vmem->memdev->mr);
-    virtio_mem_for_each_unplugged_section(vmem, s, &data,
-                                          virtio_mem_rdm_replay_discarded_cb);
+    return virtio_mem_for_each_unplugged_section(vmem, s, &data,
+                                                 virtio_mem_rdm_replay_discarded_cb);
 }
 
 static void virtio_mem_rdm_register_listener(RamDiscardManager *rdm,
include/exec/memory.h +16 -15
index 390477b588..aa67d84329 100644
@@ -566,8 +566,7 @@ static inline void ram_discard_listener_init(RamDiscardListener *rdl,
     rdl->double_discard_supported = double_discard_supported;
 }
 
-typedef int (*ReplayRamPopulate)(MemoryRegionSection *section, void *opaque);
-typedef void (*ReplayRamDiscard)(MemoryRegionSection *section, void *opaque);
+typedef int (*ReplayRamStateChange)(MemoryRegionSection *section, void *opaque);
 
 /*
  * RamDiscardManagerClass:
@@ -641,36 +640,38 @@ struct RamDiscardManagerClass {
     /**
      * @replay_populated:
      *
-     * Call the #ReplayRamPopulate callback for all populated parts within the
+     * Call the #ReplayRamStateChange callback for all populated parts within the
      * #MemoryRegionSection via the #RamDiscardManager.
      *
      * In case any call fails, no further calls are made.
      *
      * @rdm: the #RamDiscardManager
      * @section: the #MemoryRegionSection
-     * @replay_fn: the #ReplayRamPopulate callback
+     * @replay_fn: the #ReplayRamStateChange callback
      * @opaque: pointer to forward to the callback
      *
      * Returns 0 on success, or a negative error if any notification failed.
      */
     int (*replay_populated)(const RamDiscardManager *rdm,
                             MemoryRegionSection *section,
-                            ReplayRamPopulate replay_fn, void *opaque);
+                            ReplayRamStateChange replay_fn, void *opaque);
 
     /**
      * @replay_discarded:
      *
-     * Call the #ReplayRamDiscard callback for all discarded parts within the
+     * Call the #ReplayRamStateChange callback for all discarded parts within the
      * #MemoryRegionSection via the #RamDiscardManager.
      *
      * @rdm: the #RamDiscardManager
      * @section: the #MemoryRegionSection
-     * @replay_fn: the #ReplayRamDiscard callback
+     * @replay_fn: the #ReplayRamStateChange callback
      * @opaque: pointer to forward to the callback
+     *
+     * Returns 0 on success, or a negative error if any notification failed.
      */
-    void (*replay_discarded)(const RamDiscardManager *rdm,
-                             MemoryRegionSection *section,
-                             ReplayRamDiscard replay_fn, void *opaque);
+    int (*replay_discarded)(const RamDiscardManager *rdm,
+                            MemoryRegionSection *section,
+                            ReplayRamStateChange replay_fn, void *opaque);
 
     /**
      * @register_listener:
@@ -713,13 +714,13 @@ bool ram_discard_manager_is_populated(const RamDiscardManager *rdm,
 
 int ram_discard_manager_replay_populated(const RamDiscardManager *rdm,
                                          MemoryRegionSection *section,
-                                         ReplayRamPopulate replay_fn,
+                                         ReplayRamStateChange replay_fn,
                                          void *opaque);
 
-void ram_discard_manager_replay_discarded(const RamDiscardManager *rdm,
-                                          MemoryRegionSection *section,
-                                          ReplayRamDiscard replay_fn,
-                                          void *opaque);
+int ram_discard_manager_replay_discarded(const RamDiscardManager *rdm,
+                                         MemoryRegionSection *section,
+                                         ReplayRamStateChange replay_fn,
+                                         void *opaque);
 
 void ram_discard_manager_register_listener(RamDiscardManager *rdm,
                                            RamDiscardListener *rdl,
migration/ram.c +3 -2
index ce28328141..053730367b 100644
@@ -816,8 +816,8 @@ static inline bool migration_bitmap_clear_dirty(RAMState *rs,
     return ret;
 }
 
-static void dirty_bitmap_clear_section(MemoryRegionSection *section,
-                                       void *opaque)
+static int dirty_bitmap_clear_section(MemoryRegionSection *section,
+                                      void *opaque)
 {
     const hwaddr offset = section->offset_within_region;
     const hwaddr size = int128_get64(section->size);
@@ -836,6 +836,7 @@ static void dirty_bitmap_clear_section(MemoryRegionSection *section,
     }
     *cleared_bits += bitmap_count_one_with_offset(rb->bmap, start, npages);
     bitmap_clear(rb->bmap, start, npages);
+    return 0;
 }
 
 /*
system/memory.c +6 -6
index 62d6b410f0..8622d17133 100644
@@ -2147,7 +2147,7 @@ bool ram_discard_manager_is_populated(const RamDiscardManager *rdm,
 
 int ram_discard_manager_replay_populated(const RamDiscardManager *rdm,
                                          MemoryRegionSection *section,
-                                         ReplayRamPopulate replay_fn,
+                                         ReplayRamStateChange replay_fn,
                                          void *opaque)
 {
     RamDiscardManagerClass *rdmc = RAM_DISCARD_MANAGER_GET_CLASS(rdm);
@@ -2156,15 +2156,15 @@ int ram_discard_manager_replay_populated(const RamDiscardManager *rdm,
     return rdmc->replay_populated(rdm, section, replay_fn, opaque);
 }
 
-void ram_discard_manager_replay_discarded(const RamDiscardManager *rdm,
-                                          MemoryRegionSection *section,
-                                          ReplayRamDiscard replay_fn,
-                                          void *opaque)
+int ram_discard_manager_replay_discarded(const RamDiscardManager *rdm,
+                                         MemoryRegionSection *section,
+                                         ReplayRamStateChange replay_fn,
+                                         void *opaque)
 {
     RamDiscardManagerClass *rdmc = RAM_DISCARD_MANAGER_GET_CLASS(rdm);
 
     g_assert(rdmc->replay_discarded);
-    rdmc->replay_discarded(rdm, section, replay_fn, opaque);
+    return rdmc->replay_discarded(rdm, section, replay_fn, opaque);
 }
 
 void ram_discard_manager_register_listener(RamDiscardManager *rdm,
Signature · 1 line
2.43.5
#4 lore raw

Re: [PATCH v3 3/7] memory: Unify the definiton of ReplayRamPopulate() and ReplayRamDiscard()

Quoted 17 lines · Chenyi Qiang

...

migration/ram.c +2 -2
index ce28328141..053730367b 100644
@@ -816,8 +816,8 @@ static inline bool migration_bitmap_clear_dirty(RAMState *rs,
      return ret;
  }
 
-static void dirty_bitmap_clear_section(MemoryRegionSection *section,
-                                       void *opaque)
+static int dirty_bitmap_clear_section(MemoryRegionSection *section,
+                                      void *opaque)
  {
      const hwaddr offset = section->offset_within_region;
      const hwaddr size = int128_get64(section->size);

migration/ram.c static void dirty_bitmap_clear_section(MemoryRegionSection *section, · line 838

Quoted 3 lines
migration/ram.c +0 -0
@@ -836,6 +836,7 @@ static void dirty_bitmap_clear_section(MemoryRegionSection *section,
      }
      *cleared_bits += bitmap_count_one_with_offset(rb->bmap, start, npages);
      bitmap_clear(rb->bmap, start, npages);

It appears that the ram_discard_manager_replay_discarded() path would clear all private pages from the dirty bitmap over here, and thus break migration.

Perhaps the MemoryAttributeManager should be excluded from this flow, something like below?

@@ -910,6 +909,9 @@ static uint64_t ramblock_dirty_bitmap_clear_discarded_pages(RAMBlock *rb)

             .size = int128_make64(qemu_ram_get_used_length(rb)),
         };

+ if (object_dynamic_cast(OBJECT(rdm), TYPE_MEMORY_ATTRIBUTE_MANAGER))
+ return 0;
+

         ram_discard_manager_replay_discarded(rdm, &section,
                                              dirty_bitmap_clear_section,
                                              &cleared_bits);
#5 lore raw

Re: [PATCH v3 3/7] memory: Unify the definiton of ReplayRamPopulate() and ReplayRamDiscard()

Quoted 19 lines · Kishen Maloor
Quoted 18 lines · Chenyi Qiang

...

migration/ram.c +2 -1
index ce28328141..053730367b 100644
@@ -816,8 +816,8 @@ static inline bool
 migration_bitmap_clear_dirty(RAMState *rs,
       return ret;
   }
   -static void dirty_bitmap_clear_section(MemoryRegionSection *section,
-                                       void *opaque)
+static int dirty_bitmap_clear_section(MemoryRegionSection *section,
+                                      void *opaque)
   {
       const hwaddr offset = section->offset_within_region;

      const hwaddr size = int128_get64(section->size);

migration/ram.c +0 -0
@@ -836,6 +836,7 @@ static void
 dirty_bitmap_clear_section(MemoryRegionSection *section,
Quoted 8 lines

migration/ram.c static void dirty_bitmap_clear_section(MemoryRegionSection *section, · line 838

Quoted 4 lines
migration/ram.c +0 -0
       }
       *cleared_bits += bitmap_count_one_with_offset(rb->bmap, start,
 npages);
       bitmap_clear(rb->bmap, start, npages);

It appears that the ram_discard_manager_replay_discarded() path would clear all private pages
from the dirty bitmap over here, and thus break migration.

Exactly, but migration for confidential VM support with guest_memfd is not supported yet.

Quoted 3 lines

Perhaps the MemoryAttributeManager should be excluded from this flow, something like below?

I think it would already fail before reaching this path if initiating confidential VMs migration. But adding such check or assert for some hints is also OK for me. No strong preference.

Quoted 14 lines
(quoted diff) +0 -0
@@ -910,6 +909,9 @@ static uint64_t
 ramblock_dirty_bitmap_clear_discarded_pages(RAMBlock *rb)
              .size = int128_make64(qemu_ram_get_used_length(rb)),
          };


+        if (object_dynamic_cast(OBJECT(rdm), TYPE_MEMORY_ATTRIBUTE_MANAGER))
+            return 0;
+
         ram_discard_manager_replay_discarded(rdm, &section, dirty_bitmap_clear_section, &cleared_bits);

#6 lore raw

PATCH v3 4/7 memory-attribute-manager: Introduce MemoryAttributeManager to manage RAMBLock with guest_memfd

In reply to Chenyi Qiang

As the commit 852f0048f3 ("RAMBlock: make guest_memfd require uncoordinated discard") highlighted, some subsystems like VFIO may disable ram block discard. However, guest_memfd relies on the discard operation to perform page conversion between private and shared memory. This can lead to stale IOMMU mapping issue when assigning a hardware device to a confidential VM via shared memory. To address this, it is crucial to ensure systems like VFIO refresh its IOMMU mappings.

RamDiscardManager is an existing concept (used by virtio-mem) to adjust VFIO mappings in relation to VM page assignment. Effectively page conversion is similar to hot-removing a page in one mode and adding it back in the other. Therefore, similar actions are required for page conversion events. Introduce the RamDiscardManager to guest_memfd to facilitate this process.

Since guest_memfd is not an object, it cannot directly implement the RamDiscardManager interface. One potential attempt is to implement it in HostMemoryBackend. This is not appropriate because guest_memfd is per RAMBlock. Some RAMBlocks have a memory backend but others do not. In particular, the ones like virtual BIOS calling memory_region_init_ram_guest_memfd() do not.

To manage the RAMBlocks with guest_memfd, define a new object named MemoryAttributeManager to implement the RamDiscardManager interface. The object stores guest_memfd information such as shared_bitmap, and handles page conversion notification. Using the name of MemoryAttributeManager is aimed to make it more generic. The term "Memory" emcompasses not only RAM but also private MMIO in TEE I/O, which might rely on this object/interface to handle page conversion events in the future. The term "Attribute" allows for the management of various attributes beyond shared and private. For instance, it could support scenarios where discard vs. populated and shared vs. private states co-exists, such as supporting virtio-mem or something similar in the future.

In the current context, MemoryAttributeManager signifies discarded state as private and populated state as shared. Memory state is tracked at the host page size granularity, as the minimum memory conversion size can be one page per request. Additionally, VFIO expects the DMA mapping for a specific iova to be mapped and unmapped with the same granularity. Confidential VMs may perform partial conversions, e.g. conversion happens on a small region within a large region. To prevent such invalid cases and until cut_mapping operation support is introduced, all operations are performed with 4K granularity.

---
Changes in v3:
- Some rename (bitmap_size->shared_bitmap_size, first_one/zero_bit->first_bit, etc.)
- Change shared_bitmap_size from uint32_t to unsigned
- Return mgr->mr->ram_block->page_size in get_block_size()
- Move set_ram_discard_manager() up to avoid a g_free() in failure case.
- Add const for the memory_attribute_manager_get_block_size()
- Unify the ReplayRamPopulate and ReplayRamDiscard and related callback.

Changes in v2:
- Rename the object name to MemoryAttributeManager
- Rename the bitmap to shared_bitmap to make it more clear.
- Remove block_size field and get it from a helper. In future, we can get the page_size from RAMBlock if necessary.
- Remove the unncessary "struct" before GuestMemfdReplayData
- Remove the unncessary g_free() for the bitmap
- Add some error report when the callback failure for populated/discarded section.
- Move the realize()/unrealize() definition to this patch.

---
 include/system/memory-attribute-manager.h |  42 ++++
 system/memory-attribute-manager.c         | 283 ++++++++++++++++++++++
 system/meson.build                        |   1 +
 3 files changed, 326 insertions(+)
 create mode 100644 include/system/memory-attribute-manager.h
 create mode 100644 system/memory-attribute-manager.c
include/system/memory-attribute-manager.h new file +42 -0
new file mode 100644
index 0000000000..23375a14b8
@@ -0,0 +1,42 @@
+/*
+ * QEMU memory attribute manager
+ *
+ * Copyright Intel
+ *
+ * Author:
+ *      Chenyi Qiang <[email protected]>
+ *
+ * This work is licensed under the terms of the GNU GPL, version 2 or later.
+ * See the COPYING file in the top-level directory
+ *
+ */
+
+#ifndef SYSTEM_MEMORY_ATTRIBUTE_MANAGER_H
+#define SYSTEM_MEMORY_ATTRIBUTE_MANAGER_H
+
+#include "system/hostmem.h"
+
+#define TYPE_MEMORY_ATTRIBUTE_MANAGER "memory-attribute-manager"
+
+OBJECT_DECLARE_TYPE(MemoryAttributeManager, MemoryAttributeManagerClass, MEMORY_ATTRIBUTE_MANAGER)
+
+struct MemoryAttributeManager {
+    Object parent;
+
+    MemoryRegion *mr;
+
+    /* 1-setting of the bit represents the memory is populated (shared) */
+    unsigned shared_bitmap_size;
+    unsigned long *shared_bitmap;
+
+    QLIST_HEAD(, RamDiscardListener) rdl_list;
+};
+
+struct MemoryAttributeManagerClass {
+    ObjectClass parent_class;
+};
+
+int memory_attribute_manager_realize(MemoryAttributeManager *mgr, MemoryRegion *mr);
+void memory_attribute_manager_unrealize(MemoryAttributeManager *mgr);
+
+#endif
system/memory-attribute-manager.c new file +283 -0

283 lines not on this page · raw message

system/meson.build +1 -0
index 4952f4b2c7..ab07ff1442 100644
@@ -15,6 +15,7 @@ system_ss.add(files(
   'dirtylimit.c',
   'dma-helpers.c',
   'globals.c',
+  'memory-attribute-manager.c',
   'memory_mapping.c',
   'qdev-monitor.c',
   'qtest.c',
Signature · 1 line
2.43.5
#7 lore raw

Re: [PATCH v3 4/7] memory-attribute-manager: Introduce MemoryAttributeManager to manage RAMBLock with guest_memfd

Quoted 22 lines · Chenyi Qiang

As the commit 852f0048f3 ("RAMBlock: make guest_memfd require uncoordinated discard") highlighted, some subsystems like VFIO may disable ram block discard. However, guest_memfd relies on the discard operation to perform page conversion between private and shared memory. This can lead to stale IOMMU mapping issue when assigning a hardware device to a confidential VM via shared memory. To address this, it is crucial to ensure systems like VFIO refresh its IOMMU mappings.

RamDiscardManager is an existing concept (used by virtio-mem) to adjust VFIO mappings in relation to VM page assignment. Effectively page conversion is similar to hot-removing a page in one mode and adding it back in the other. Therefore, similar actions are required for page conversion events. Introduce the RamDiscardManager to guest_memfd to facilitate this process.

Since guest_memfd is not an object, it cannot directly implement the RamDiscardManager interface. One potential attempt is to implement it in HostMemoryBackend. This is not appropriate because guest_memfd is per RAMBlock. Some RAMBlocks have a memory backend but others do not. In particular, the ones like virtual BIOS calling memory_region_init_ram_guest_memfd() do not.

Quoted 2 lines

To manage the RAMBlocks with guest_memfd, define a new object named MemoryAttributeManager to implement the RamDiscardManager interface. The

Isn't this should be the other way around. 'MemoryAttributeManager' should be an interface and RamDiscardManager a type of it, an implementation?

MemoryAttributeManager have the data like 'shared_bitmap' etc that information can also be used by the other implementation types?

Or maybe I am getting it entirely wrong.

Thanks,
Pankaj

Quoted 378 lines

object stores guest_memfd information such as shared_bitmap, and handles page conversion notification. Using the name of MemoryAttributeManager is aimed to make it more generic. The term "Memory" emcompasses not only RAM but also private MMIO in TEE I/O, which might rely on this object/interface to handle page conversion events in the future. The term "Attribute" allows for the management of various attributes beyond shared and private. For instance, it could support scenarios where discard vs. populated and shared vs. private states co-exists, such as supporting virtio-mem or something similar in the future.

In the current context, MemoryAttributeManager signifies discarded state as private and populated state as shared. Memory state is tracked at the host page size granularity, as the minimum memory conversion size can be one page per request. Additionally, VFIO expects the DMA mapping for a specific iova to be mapped and unmapped with the same granularity. Confidential VMs may perform partial conversions, e.g. conversion happens on a small region within a large region. To prevent such invalid cases and until cut_mapping operation support is introduced, all operations are performed with 4K granularity.

---
Changes in v3:
- Some rename (bitmap_size->shared_bitmap_size, first_one/zero_bit->first_bit, etc.)
- Change shared_bitmap_size from uint32_t to unsigned
- Return mgr->mr->ram_block->page_size in get_block_size()
- Move set_ram_discard_manager() up to avoid a g_free() in failure case.
- Add const for the memory_attribute_manager_get_block_size()
- Unify the ReplayRamPopulate and ReplayRamDiscard and related callback.

Changes in v2:
- Rename the object name to MemoryAttributeManager
- Rename the bitmap to shared_bitmap to make it more clear.
- Remove block_size field and get it from a helper. In future, we can get the page_size from RAMBlock if necessary.
- Remove the unncessary "struct" before GuestMemfdReplayData
- Remove the unncessary g_free() for the bitmap
- Add some error report when the callback failure for populated/discarded section.
- Move the realize()/unrealize() definition to this patch.

---
  include/system/memory-attribute-manager.h |  42 ++++
  system/memory-attribute-manager.c         | 283 ++++++++++++++++++++++
  system/meson.build                        |   1 +
  3 files changed, 326 insertions(+)

create mode 100644 include/system/memory-attribute-manager.h create mode 100644 system/memory-attribute-manager.c

include/system/memory-attribute-manager.h new file +42 -0
new file mode 100644
index 0000000000..23375a14b8
@@ -0,0 +1,42 @@
+/*
+ * QEMU memory attribute manager
+ *
+ * Copyright Intel
+ *
+ * Author:
+ *      Chenyi Qiang <[email protected]>
+ *
+ * This work is licensed under the terms of the GNU GPL, version 2 or later.
+ * See the COPYING file in the top-level directory
+ *
+ */
+
+#ifndef SYSTEM_MEMORY_ATTRIBUTE_MANAGER_H
+#define SYSTEM_MEMORY_ATTRIBUTE_MANAGER_H
+
+#include "system/hostmem.h"
+
+#define TYPE_MEMORY_ATTRIBUTE_MANAGER "memory-attribute-manager"
+
+OBJECT_DECLARE_TYPE(MemoryAttributeManager, MemoryAttributeManagerClass, MEMORY_ATTRIBUTE_MANAGER)
+
+struct MemoryAttributeManager {
+    Object parent;
+
+    MemoryRegion *mr;
+
+    /* 1-setting of the bit represents the memory is populated (shared) */
+    unsigned shared_bitmap_size;
+    unsigned long *shared_bitmap;
+
+    QLIST_HEAD(, RamDiscardListener) rdl_list;
+};
+
+struct MemoryAttributeManagerClass {
+    ObjectClass parent_class;
+};
+
+int memory_attribute_manager_realize(MemoryAttributeManager *mgr, MemoryRegion *mr);
+void memory_attribute_manager_unrealize(MemoryAttributeManager *mgr);
+
+#endif
system/memory-attribute-manager.c new file +273 -0
new file mode 100644
index 0000000000..7c3789cf49
@@ -0,0 +1,283 @@
+/*
+ * QEMU memory attribute manager
+ *
+ * Copyright Intel
+ *
+ * Author:
+ *      Chenyi Qiang <[email protected]>
+ *
+ * This work is licensed under the terms of the GNU GPL, version 2 or later.
+ * See the COPYING file in the top-level directory
+ *
+ */
+
+#include "qemu/osdep.h"
+#include "qemu/error-report.h"
+#include "exec/ramblock.h"
+#include "system/memory-attribute-manager.h"
+
+OBJECT_DEFINE_TYPE_WITH_INTERFACES(MemoryAttributeManager,
+                                   memory_attribute_manager,
+                                   MEMORY_ATTRIBUTE_MANAGER,
+                                   OBJECT,
+                                   { TYPE_RAM_DISCARD_MANAGER },
+                                   { })
+
+static size_t memory_attribute_manager_get_block_size(const MemoryAttributeManager *mgr)
+{
+    /*
+     * Because page conversion could be manipulated in the size of at least 4K or 4K aligned,
+     * Use the host page size as the granularity to track the memory attribute.
+     */
+    g_assert(mgr && mgr->mr && mgr->mr->ram_block);
+    g_assert(mgr->mr->ram_block->page_size == qemu_real_host_page_size());
+    return mgr->mr->ram_block->page_size;
+}
+
+
+static bool memory_attribute_rdm_is_populated(const RamDiscardManager *rdm,
+                                              const MemoryRegionSection *section)
+{
+    const MemoryAttributeManager *mgr = MEMORY_ATTRIBUTE_MANAGER(rdm);
+    const int block_size = memory_attribute_manager_get_block_size(mgr);
+    uint64_t first_bit = section->offset_within_region / block_size;
+    uint64_t last_bit = first_bit + int128_get64(section->size) / block_size - 1;
+    unsigned long first_discard_bit;
+
+    first_discard_bit = find_next_zero_bit(mgr->shared_bitmap, last_bit + 1, first_bit);
+    return first_discard_bit > last_bit;
+}
+
+typedef int (*memory_attribute_section_cb)(MemoryRegionSection *s, void *arg);
+
+static int memory_attribute_notify_populate_cb(MemoryRegionSection *section, void *arg)
+{
+    RamDiscardListener *rdl = arg;
+
+    return rdl->notify_populate(rdl, section);
+}
+
+static int memory_attribute_notify_discard_cb(MemoryRegionSection *section, void *arg)
+{
+    RamDiscardListener *rdl = arg;
+
+    rdl->notify_discard(rdl, section);
+
+    return 0;
+}
+
+static int memory_attribute_for_each_populated_section(const MemoryAttributeManager *mgr,
+                                                       MemoryRegionSection *section,
+                                                       void *arg,
+                                                       memory_attribute_section_cb cb)
+{
+    unsigned long first_bit, last_bit;
+    uint64_t offset, size;
+    const int block_size = memory_attribute_manager_get_block_size(mgr);
+    int ret = 0;
+
+    first_bit = section->offset_within_region / block_size;
+    first_bit = find_next_bit(mgr->shared_bitmap, mgr->shared_bitmap_size, first_bit);
+
+    while (first_bit < mgr->shared_bitmap_size) {
+        MemoryRegionSection tmp = *section;
+
+        offset = first_bit * block_size;
+        last_bit = find_next_zero_bit(mgr->shared_bitmap, mgr->shared_bitmap_size,
+                                      first_bit + 1) - 1;
+        size = (last_bit - first_bit + 1) * block_size;
+
+        if (!memory_region_section_intersect_range(&tmp, offset, size)) {
+            break;
+        }
+
+        ret = cb(&tmp, arg);
+        if (ret) {
+            error_report("%s: Failed to notify RAM discard listener: %s", __func__,
+                         strerror(-ret));
+            break;
+        }
+
+        first_bit = find_next_bit(mgr->shared_bitmap, mgr->shared_bitmap_size,
+                                  last_bit + 2);
+    }
+
+    return ret;
+}
+
+static int memory_attribute_for_each_discarded_section(const MemoryAttributeManager *mgr,
+                                                       MemoryRegionSection *section,
+                                                       void *arg,
+                                                       memory_attribute_section_cb cb)
+{
+    unsigned long first_bit, last_bit;
+    uint64_t offset, size;
+    const int block_size = memory_attribute_manager_get_block_size(mgr);
+    int ret = 0;
+
+    first_bit = section->offset_within_region / block_size;
+    first_bit = find_next_zero_bit(mgr->shared_bitmap, mgr->shared_bitmap_size,
+                                   first_bit);
+
+    while (first_bit < mgr->shared_bitmap_size) {
+        MemoryRegionSection tmp = *section;
+
+        offset = first_bit * block_size;
+        last_bit = find_next_bit(mgr->shared_bitmap, mgr->shared_bitmap_size,
+                                      first_bit + 1) - 1;
+        size = (last_bit - first_bit + 1) * block_size;
+
+        if (!memory_region_section_intersect_range(&tmp, offset, size)) {
+            break;
+        }
+
+        ret = cb(&tmp, arg);
+        if (ret) {
+            error_report("%s: Failed to notify RAM discard listener: %s", __func__,
+                         strerror(-ret));
+            break;
+        }
+
+        first_bit = find_next_zero_bit(mgr->shared_bitmap, mgr->shared_bitmap_size,
+                                       last_bit + 2);
+    }
+
+    return ret;
+}
+
+static uint64_t memory_attribute_rdm_get_min_granularity(const RamDiscardManager *rdm,
+                                                         const MemoryRegion *mr)
+{
+    MemoryAttributeManager *mgr = MEMORY_ATTRIBUTE_MANAGER(rdm);
+
+    g_assert(mr == mgr->mr);
+    return memory_attribute_manager_get_block_size(mgr);
+}
+
+static void memory_attribute_rdm_register_listener(RamDiscardManager *rdm,
+                                                   RamDiscardListener *rdl,
+                                                   MemoryRegionSection *section)
+{
+    MemoryAttributeManager *mgr = MEMORY_ATTRIBUTE_MANAGER(rdm);
+    int ret;
+
+    g_assert(section->mr == mgr->mr);
+    rdl->section = memory_region_section_new_copy(section);
+
+    QLIST_INSERT_HEAD(&mgr->rdl_list, rdl, next);
+
+    ret = memory_attribute_for_each_populated_section(mgr, section, rdl,
+                                                      memory_attribute_notify_populate_cb);
+    if (ret) {
+        error_report("%s: Failed to register RAM discard listener: %s", __func__,
+                     strerror(-ret));
+    }
+}
+
+static void memory_attribute_rdm_unregister_listener(RamDiscardManager *rdm,
+                                                     RamDiscardListener *rdl)
+{
+    MemoryAttributeManager *mgr = MEMORY_ATTRIBUTE_MANAGER(rdm);
+    int ret;
+
+    g_assert(rdl->section);
+    g_assert(rdl->section->mr == mgr->mr);
+
+    ret = memory_attribute_for_each_populated_section(mgr, rdl->section, rdl,
+                                                      memory_attribute_notify_discard_cb);
+    if (ret) {
+        error_report("%s: Failed to unregister RAM discard listener: %s", __func__,
+                     strerror(-ret));
+    }
+
+    memory_region_section_free_copy(rdl->section);
+    rdl->section = NULL;
+    QLIST_REMOVE(rdl, next);
+
+}
+
+typedef struct MemoryAttributeReplayData {
+    ReplayRamStateChange fn;
+    void *opaque;
+} MemoryAttributeReplayData;
+
+static int memory_attribute_rdm_replay_cb(MemoryRegionSection *section, void *arg)
+{
+    MemoryAttributeReplayData *data = arg;
+
+    return data->fn(section, data->opaque);
+}
+
+static int memory_attribute_rdm_replay_populated(const RamDiscardManager *rdm,
+                                                 MemoryRegionSection *section,
+                                                 ReplayRamStateChange replay_fn,
+                                                 void *opaque)
+{
+    MemoryAttributeManager *mgr = MEMORY_ATTRIBUTE_MANAGER(rdm);
+    MemoryAttributeReplayData data = { .fn = replay_fn, .opaque = opaque };
+
+    g_assert(section->mr == mgr->mr);
+    return memory_attribute_for_each_populated_section(mgr, section, &data,
+                                                       memory_attribute_rdm_replay_cb);
+}
+
+static int memory_attribute_rdm_replay_discarded(const RamDiscardManager *rdm,
+                                                 MemoryRegionSection *section,
+                                                 ReplayRamStateChange replay_fn,
+                                                 void *opaque)
+{
+    MemoryAttributeManager *mgr = MEMORY_ATTRIBUTE_MANAGER(rdm);
+    MemoryAttributeReplayData data = { .fn = replay_fn, .opaque = opaque };
+
+    g_assert(section->mr == mgr->mr);
+    return memory_attribute_for_each_discarded_section(mgr, section, &data,
+                                                       memory_attribute_rdm_replay_cb);
+}
+
+int memory_attribute_manager_realize(MemoryAttributeManager *mgr, MemoryRegion *mr)
+{
+    uint64_t shared_bitmap_size;
+    const int block_size  = qemu_real_host_page_size();
+    int ret;
+
+    shared_bitmap_size = ROUND_UP(mr->size, block_size) / block_size;
+
+    mgr->mr = mr;
+    ret = memory_region_set_ram_discard_manager(mgr->mr, RAM_DISCARD_MANAGER(mgr));
+    if (ret) {
+        return ret;
+    }
+    mgr->shared_bitmap_size = shared_bitmap_size;
+    mgr->shared_bitmap = bitmap_new(shared_bitmap_size);
+
+    return ret;
+}
+
+void memory_attribute_manager_unrealize(MemoryAttributeManager *mgr)
+{
+    g_free(mgr->shared_bitmap);
+    memory_region_set_ram_discard_manager(mgr->mr, NULL);
+}
+
+static void memory_attribute_manager_init(Object *obj)
+{
+    MemoryAttributeManager *mgr = MEMORY_ATTRIBUTE_MANAGER(obj);
+
+    QLIST_INIT(&mgr->rdl_list);
+}
+
+static void memory_attribute_manager_finalize(Object *obj)
+{
+}
+
+static void memory_attribute_manager_class_init(ObjectClass *oc, void *data)

system/memory-attribute-manager.c line 283

Quoted 10 lines
system/memory-attribute-manager.c new file +10 -0
+{
+    RamDiscardManagerClass *rdmc = RAM_DISCARD_MANAGER_CLASS(oc);
+
+    rdmc->get_min_granularity = memory_attribute_rdm_get_min_granularity;
+    rdmc->register_listener = memory_attribute_rdm_register_listener;
+    rdmc->unregister_listener = memory_attribute_rdm_unregister_listener;
+    rdmc->is_populated = memory_attribute_rdm_is_populated;
+    rdmc->replay_populated = memory_attribute_rdm_replay_populated;
+    rdmc->replay_discarded = memory_attribute_rdm_replay_discarded;
+}

Would this initialization be for

Quoted 12 lines
system/meson.build +1 -0
index 4952f4b2c7..ab07ff1442 100644
@@ -15,6 +15,7 @@ system_ss.add(files(
    'dirtylimit.c',
    'dma-helpers.c',
    'globals.c',
+  'memory-attribute-manager.c',
    'memory_mapping.c',
    'qdev-monitor.c',
    'qtest.c',
#8 lore raw

Re: [PATCH v3 4/7] memory-attribute-manager: Introduce MemoryAttributeManager to manage RAMBLock with guest_memfd

Quoted 23 lines · Pankaj
Quoted 22 lines · Chenyi Qiang

As the commit 852f0048f3 ("RAMBlock: make guest_memfd require uncoordinated discard") highlighted, some subsystems like VFIO may disable ram block discard. However, guest_memfd relies on the discard operation to perform page conversion between private and shared memory. This can lead to stale IOMMU mapping issue when assigning a hardware device to a confidential VM via shared memory. To address this, it is crucial to ensure systems like VFIO refresh its IOMMU mappings.

RamDiscardManager is an existing concept (used by virtio-mem) to adjust VFIO mappings in relation to VM page assignment. Effectively page conversion is similar to hot-removing a page in one mode and adding it back in the other. Therefore, similar actions are required for page conversion events. Introduce the RamDiscardManager to guest_memfd to facilitate this process.

Since guest_memfd is not an object, it cannot directly implement the RamDiscardManager interface. One potential attempt is to implement it in HostMemoryBackend. This is not appropriate because guest_memfd is per RAMBlock. Some RAMBlocks have a memory backend but others do not. In particular, the ones like virtual BIOS calling memory_region_init_ram_guest_memfd() do not.

Quoted 6 lines
Quoted 2 lines

To manage the RAMBlocks with guest_memfd, define a new object named MemoryAttributeManager to implement the RamDiscardManager interface. The

Isn't this should be the other way around. 'MemoryAttributeManager' should be an interface and RamDiscardManager a type of it, an implementation?

We want to use 'MemoryAttributeManager' to represent RAMBlock to implement the RamDiscardManager interface callbacks because RAMBlock is not an object. It includes some metadata of guest_memfd like shared_bitmap at the same time.

I can't get it that make 'MemoryAttributeManager' an interface and RamDiscardManager a type of it. Can you elaborate it a little bit? I think at least we need someone to implement the RamDiscardManager interface.

Quoted 3 lines

MemoryAttributeManager have the data like 'shared_bitmap' etc that information can also be used by the other implementation types?

Shared_bitmap is just the metadata of guest_memfd. Other subsystems may consult its status by RamDiscardManager interface like ram_discard_manager_is_populated().

Quoted 474 lines

Or maybe I am getting it entirely wrong.

Thanks,
Pankaj

Quoted 445 lines

object stores guest_memfd information such as shared_bitmap, and handles page conversion notification. Using the name of MemoryAttributeManager is aimed to make it more generic. The term "Memory" emcompasses not only RAM but also private MMIO in TEE I/O, which might rely on this object/interface to handle page conversion events in the future. The term "Attribute" allows for the management of various attributes beyond shared and private. For instance, it could support scenarios where discard vs. populated and shared vs. private states co-exists, such as supporting virtio-mem or something similar in the future.

In the current context, MemoryAttributeManager signifies discarded state as private and populated state as shared. Memory state is tracked at the host page size granularity, as the minimum memory conversion size can be one
page per request. Additionally, VFIO expects the DMA mapping for a specific iova to be mapped and unmapped with the same granularity. Confidential VMs may perform  partial conversions, e.g. conversion happens on a small region within a large region. To prevent such invalid cases and until cut_mapping operation support is introduced, all operations are performed with 4K granularity.

---
Changes in v3:
     - Some rename (bitmap_size->shared_bitmap_size, first_one/zero_bit->first_bit, etc.)
     - Change shared_bitmap_size from uint32_t to unsigned
     - Return mgr->mr->ram_block->page_size in get_block_size()
     - Move set_ram_discard_manager() up to avoid a g_free() in failure case.
     - Add const for the memory_attribute_manager_get_block_size()
     - Unify the ReplayRamPopulate and ReplayRamDiscard and related callback.

Changes in v2:
     - Rename the object name to MemoryAttributeManager
     - Rename the bitmap to shared_bitmap to make it more clear.
     - Remove block_size field and get it from a helper. In future, we can get the page_size from RAMBlock if necessary.
     - Remove the unncessary "struct" before GuestMemfdReplayData
     - Remove the unncessary g_free() for the bitmap
     - Add some error report when the callback failure for populated/discarded section.
     - Move the realize()/unrealize() definition to this patch.
---
  include/system/memory-attribute-manager.h |  42 ++++ system/memory-attribute-manager.c         | 283 ++++++++++++++++++++++ system/meson.build                        |   1 +

  3 files changed, 326 insertions(+)

  create mode 100644 include/system/memory-attribute-manager.h create mode 100644 system/memory-attribute-manager.c

include/ renamed from include/system/memory-attribute-manager.h +0 -0

system/memory-attribute-manager.h
new file mode 100644
index 0000000000..23375a14b8

include/system/memory-attribute-manager.h new file +9 -0
@@ -0,0 +1,42 @@
+/*
+ * QEMU memory attribute manager
+ *
+ * Copyright Intel
+ *
+ * Author:
+ *      Chenyi Qiang <[email protected]>
+ *
+ * This work is licensed under the terms of the GNU GPL, version 2 or

later.
+ * See the COPYING file in the top-level directory
+ *
+ */

include/system/memory-attribute-manager.h new file +9 -0
+
+#ifndef SYSTEM_MEMORY_ATTRIBUTE_MANAGER_H
+#define SYSTEM_MEMORY_ATTRIBUTE_MANAGER_H
+
+#include "system/hostmem.h"
+
+#define TYPE_MEMORY_ATTRIBUTE_MANAGER "memory-attribute-manager"
+
+OBJECT_DECLARE_TYPE(MemoryAttributeManager,

MemoryAttributeManagerClass, MEMORY_ATTRIBUTE_MANAGER)

include/system/memory-attribute-manager.h new file +7 -0
+
+struct MemoryAttributeManager {
+    Object parent;
+
+    MemoryRegion *mr;
+
+    /* 1-setting of the bit represents the memory is populated

(shared) */
+    unsigned shared_bitmap_size;
+    unsigned long *shared_bitmap;

include/system/memory-attribute-manager.h new file +9 -0
+
+    QLIST_HEAD(, RamDiscardListener) rdl_list;
+};
+
+struct MemoryAttributeManagerClass {
+    ObjectClass parent_class;
+};
+
+int memory_attribute_manager_realize(MemoryAttributeManager *mgr,

MemoryRegion *mr);

include/system/memory-attribute-manager.h new file +3 -0
+void memory_attribute_manager_unrealize(MemoryAttributeManager *mgr);
+
+#endif
system/memory- renamed from system/memory-attribute-manager.c +0 -0

attribute-manager.c
new file mode 100644
index 0000000000..7c3789cf49

system/memory-attribute-manager.c new file +9 -0
@@ -0,0 +1,283 @@
+/*
+ * QEMU memory attribute manager
+ *
+ * Copyright Intel
+ *
+ * Author:
+ *      Chenyi Qiang <[email protected]>
+ *
+ * This work is licensed under the terms of the GNU GPL, version 2 or

later.
+ * See the COPYING file in the top-level directory
+ *
+ */

system/memory-attribute-manager.c new file +14 -0
+
+#include "qemu/osdep.h"
+#include "qemu/error-report.h"
+#include "exec/ramblock.h"
+#include "system/memory-attribute-manager.h"
+
+OBJECT_DEFINE_TYPE_WITH_INTERFACES(MemoryAttributeManager,
+                                   memory_attribute_manager,
+                                   MEMORY_ATTRIBUTE_MANAGER,
+                                   OBJECT,
+                                   { TYPE_RAM_DISCARD_MANAGER },
+                                   { })
+
+static size_t memory_attribute_manager_get_block_size(const

MemoryAttributeManager *mgr)
+{
+    /*
+     * Because page conversion could be manipulated in the size of at least 4K or 4K aligned,
+     * Use the host page size as the granularity to track the memory attribute.
+     */
+    g_assert(mgr && mgr->mr && mgr->mr->ram_block);
+    g_assert(mgr->mr->ram_block->page_size == qemu_real_host_page_size());
+    return mgr->mr->ram_block->page_size;

system/memory-attribute-manager.c new file +4 -0
+}
+
+
+static bool memory_attribute_rdm_is_populated(const RamDiscardManager

*rdm,
+                                              const MemoryRegionSection *section)
+{
+    const MemoryAttributeManager *mgr = MEMORY_ATTRIBUTE_MANAGER(rdm); +    const int block_size = memory_attribute_manager_get_block_size(mgr); +    uint64_t first_bit = section->offset_within_region / block_size; +    uint64_t last_bit = first_bit + int128_get64(section->size) / block_size - 1;
+    unsigned long first_discard_bit;
+
+    first_discard_bit = find_next_zero_bit(mgr->shared_bitmap, last_bit + 1, first_bit);
+    return first_discard_bit > last_bit;

system/memory-attribute-manager.c new file +3 -0
+}
+
+typedef int (*memory_attribute_section_cb)(MemoryRegionSection *s,

void *arg);

system/memory-attribute-manager.c new file +2 -0
+
+static int memory_attribute_notify_populate_cb(MemoryRegionSection

*section, void *arg)

system/memory-attribute-manager.c new file +7 -0
+{
+    RamDiscardListener *rdl = arg;
+
+    return rdl->notify_populate(rdl, section);
+}
+
+static int memory_attribute_notify_discard_cb(MemoryRegionSection

*section, void *arg)

system/memory-attribute-manager.c new file +9 -0
+{
+    RamDiscardListener *rdl = arg;
+
+    rdl->notify_discard(rdl, section);
+
+    return 0;
+}
+
+static int memory_attribute_for_each_populated_section(const

MemoryAttributeManager *mgr,
+
MemoryRegionSection *section,
+                                                       void *arg, +
memory_attribute_section_cb cb)

system/memory-attribute-manager.c new file +8 -0
+{
+    unsigned long first_bit, last_bit;
+    uint64_t offset, size;
+    const int block_size = memory_attribute_manager_get_block_size(mgr);
+    int ret = 0;
+
+    first_bit = section->offset_within_region / block_size;
+    first_bit = find_next_bit(mgr->shared_bitmap, mgr-
Quoted 1 line

shared_bitmap_size, first_bit);

system/memory-attribute-manager.c new file +6 -0
+
+    while (first_bit < mgr->shared_bitmap_size) {
+        MemoryRegionSection tmp = *section;
+
+        offset = first_bit * block_size;
+        last_bit = find_next_zero_bit(mgr->shared_bitmap, mgr-
Quoted 1 line

shared_bitmap_size,

+                                      first_bit + 1) - 1; +        size = (last_bit - first_bit + 1) * block_size;
+
+        if (!memory_region_section_intersect_range(&tmp, offset, size)) {
+            break;
+        }
+
+        ret = cb(&tmp, arg);
+        if (ret) {
+            error_report("%s: Failed to notify RAM discard listener:
%s", __func__,
+                         strerror(-ret));
+            break;
+        }
+
+        first_bit = find_next_bit(mgr->shared_bitmap, mgr-

Quoted 1 line

shared_bitmap_size,

+                                  last_bit + 2);
+    }

system/memory-attribute-manager.c new file +5 -0
+
+    return ret;
+}
+
+static int memory_attribute_for_each_discarded_section(const

MemoryAttributeManager *mgr,
+
MemoryRegionSection *section,
+                                                       void *arg, +
memory_attribute_section_cb cb)

system/memory-attribute-manager.c new file +8 -0
+{
+    unsigned long first_bit, last_bit;
+    uint64_t offset, size;
+    const int block_size = memory_attribute_manager_get_block_size(mgr);
+    int ret = 0;
+
+    first_bit = section->offset_within_region / block_size;
+    first_bit = find_next_zero_bit(mgr->shared_bitmap, mgr-
Quoted 1 line

shared_bitmap_size,

+                                   first_bit);

system/memory-attribute-manager.c new file +6 -0
+
+    while (first_bit < mgr->shared_bitmap_size) {
+        MemoryRegionSection tmp = *section;
+
+        offset = first_bit * block_size;
+        last_bit = find_next_bit(mgr->shared_bitmap, mgr-
Quoted 1 line

shared_bitmap_size,

+                                      first_bit + 1) - 1; +        size = (last_bit - first_bit + 1) * block_size;
+
+        if (!memory_region_section_intersect_range(&tmp, offset, size)) {
+            break;
+        }
+
+        ret = cb(&tmp, arg);
+        if (ret) {
+            error_report("%s: Failed to notify RAM discard listener:
%s", __func__,
+                         strerror(-ret));
+            break;
+        }
+
+        first_bit = find_next_zero_bit(mgr->shared_bitmap, mgr-

Quoted 1 line

shared_bitmap_size,

+                                       last_bit + 2); +    }

system/memory-attribute-manager.c new file +5 -0
+
+    return ret;
+}
+
+static uint64_t memory_attribute_rdm_get_min_granularity(const

RamDiscardManager *rdm,
+                                                         const MemoryRegion *mr)

system/memory-attribute-manager.c new file +8 -0
+{
+    MemoryAttributeManager *mgr = MEMORY_ATTRIBUTE_MANAGER(rdm);
+
+    g_assert(mr == mgr->mr);
+    return memory_attribute_manager_get_block_size(mgr);
+}
+
+static void memory_attribute_rdm_register_listener(RamDiscardManager

*rdm,
+                                                   RamDiscardListener *rdl,
+
MemoryRegionSection *section)

system/memory-attribute-manager.c new file +11 -0
+{
+    MemoryAttributeManager *mgr = MEMORY_ATTRIBUTE_MANAGER(rdm);
+    int ret;
+
+    g_assert(section->mr == mgr->mr);
+    rdl->section = memory_region_section_new_copy(section);
+
+    QLIST_INSERT_HEAD(&mgr->rdl_list, rdl, next);
+
+    ret = memory_attribute_for_each_populated_section(mgr, section, rdl,
+                                                     

memory_attribute_notify_populate_cb);
+    if (ret) {
+        error_report("%s: Failed to register RAM discard listener:
%s", __func__,
+                     strerror(-ret));
+    }

system/memory-attribute-manager.c new file +3 -0
+}
+
+static void

memory_attribute_rdm_unregister_listener(RamDiscardManager *rdm, +
RamDiscardListener *rdl)

system/memory-attribute-manager.c new file +8 -0
+{
+    MemoryAttributeManager *mgr = MEMORY_ATTRIBUTE_MANAGER(rdm);
+    int ret;
+
+    g_assert(rdl->section);
+    g_assert(rdl->section->mr == mgr->mr);
+
+    ret = memory_attribute_for_each_populated_section(mgr, rdl-
Quoted 1 line

section, rdl,

+
memory_attribute_notify_discard_cb);
+    if (ret) {
+        error_report("%s: Failed to unregister RAM discard listener:
%s", __func__,
+                     strerror(-ret));
+    }

system/memory-attribute-manager.c new file +13 -0
+
+    memory_region_section_free_copy(rdl->section);
+    rdl->section = NULL;
+    QLIST_REMOVE(rdl, next);
+
+}
+
+typedef struct MemoryAttributeReplayData {
+    ReplayRamStateChange fn;
+    void *opaque;
+} MemoryAttributeReplayData;
+
+static int memory_attribute_rdm_replay_cb(MemoryRegionSection

*section, void *arg)

system/memory-attribute-manager.c new file +7 -0
+{
+    MemoryAttributeReplayData *data = arg;
+
+    return data->fn(section, data->opaque);
+}
+
+static int memory_attribute_rdm_replay_populated(const

RamDiscardManager *rdm,
+                                                 MemoryRegionSection *section,
+                                                 ReplayRamStateChange replay_fn,
+                                                 void *opaque) +{
+    MemoryAttributeManager *mgr = MEMORY_ATTRIBUTE_MANAGER(rdm); +    MemoryAttributeReplayData data = { .fn = replay_fn, .opaque = opaque };
+
+    g_assert(section->mr == mgr->mr);
+    return memory_attribute_for_each_populated_section(mgr, section, &data,
+
memory_attribute_rdm_replay_cb);

system/memory-attribute-manager.c new file +3 -0
+}
+
+static int memory_attribute_rdm_replay_discarded(const

RamDiscardManager *rdm,
+                                                 MemoryRegionSection *section,
+                                                 ReplayRamStateChange replay_fn,
+                                                 void *opaque) +{
+    MemoryAttributeManager *mgr = MEMORY_ATTRIBUTE_MANAGER(rdm); +    MemoryAttributeReplayData data = { .fn = replay_fn, .opaque = opaque };
+
+    g_assert(section->mr == mgr->mr);
+    return memory_attribute_for_each_discarded_section(mgr, section, &data,
+
memory_attribute_rdm_replay_cb);

system/memory-attribute-manager.c new file +3 -0
+}
+
+int memory_attribute_manager_realize(MemoryAttributeManager *mgr,

MemoryRegion *mr)

system/memory-attribute-manager.c new file +9 -0
+{
+    uint64_t shared_bitmap_size;
+    const int block_size  = qemu_real_host_page_size();
+    int ret;
+
+    shared_bitmap_size = ROUND_UP(mr->size, block_size) / block_size;
+
+    mgr->mr = mr;
+    ret = memory_region_set_ram_discard_manager(mgr->mr,

RAM_DISCARD_MANAGER(mgr));
+    if (ret) {
+        return ret;
+    }
+    mgr->shared_bitmap_size = shared_bitmap_size; +    mgr->shared_bitmap = bitmap_new(shared_bitmap_size);

system/memory-attribute-manager.c new file +22 -0
+
+    return ret;
+}
+
+void memory_attribute_manager_unrealize(MemoryAttributeManager *mgr)
+{
+    g_free(mgr->shared_bitmap);
+    memory_region_set_ram_discard_manager(mgr->mr, NULL);
+}
+
+static void memory_attribute_manager_init(Object *obj)
+{
+    MemoryAttributeManager *mgr = MEMORY_ATTRIBUTE_MANAGER(obj);
+
+    QLIST_INIT(&mgr->rdl_list);
+}
+
+static void memory_attribute_manager_finalize(Object *obj)
+{
+}
+
+static void memory_attribute_manager_class_init(ObjectClass *oc, void

*data)

system/memory-attribute-manager.c new file +4 -0
+{
+    RamDiscardManagerClass *rdmc = RAM_DISCARD_MANAGER_CLASS(oc);
+
+    rdmc->get_min_granularity =
Quoted 8 lines

memory_attribute_rdm_get_min_granularity;
+    rdmc->register_listener = memory_attribute_rdm_register_listener; +    rdmc->unregister_listener = memory_attribute_rdm_unregister_listener;
+    rdmc->is_populated = memory_attribute_rdm_is_populated; +    rdmc->replay_populated = memory_attribute_rdm_replay_populated; +    rdmc->replay_discarded = memory_attribute_rdm_replay_discarded; +}

Would this initialization be for

Quoted 12 lines
system/meson.build +1 -0
index 4952f4b2c7..ab07ff1442 100644
@@ -15,6 +15,7 @@ system_ss.add(files(
     'dirtylimit.c',
     'dma-helpers.c',
     'globals.c',
+  'memory-attribute-manager.c',
     'memory_mapping.c',
     'qdev-monitor.c',
     'qtest.c',
#9 lore raw

Re: [PATCH v3 4/7] memory-attribute-manager: Introduce MemoryAttributeManager to manage RAMBLock with guest_memfd

Quoted 33 lines · Chenyi Qiang
Quoted 23 lines · Pankaj
Quoted 22 lines · Chenyi Qiang

As the commit 852f0048f3 ("RAMBlock: make guest_memfd require uncoordinated discard") highlighted, some subsystems like VFIO may disable ram block discard. However, guest_memfd relies on the discard operation to perform page conversion between private and shared memory. This can lead to stale IOMMU mapping issue when assigning a hardware device to a confidential VM via shared memory. To address this, it is crucial to ensure systems like VFIO refresh its IOMMU mappings.

RamDiscardManager is an existing concept (used by virtio-mem) to adjust VFIO mappings in relation to VM page assignment. Effectively page conversion is similar to hot-removing a page in one mode and adding it back in the other. Therefore, similar actions are required for page conversion events. Introduce the RamDiscardManager to guest_memfd to facilitate this process.

Since guest_memfd is not an object, it cannot directly implement the RamDiscardManager interface. One potential attempt is to implement it in HostMemoryBackend. This is not appropriate because guest_memfd is per RAMBlock. Some RAMBlocks have a memory backend but others do not. In particular, the ones like virtual BIOS calling memory_region_init_ram_guest_memfd() do not.

Quoted 6 lines
Quoted 2 lines

To manage the RAMBlocks with guest_memfd, define a new object named MemoryAttributeManager to implement the RamDiscardManager interface. The

Isn't this should be the other way around. 'MemoryAttributeManager' should be an interface and RamDiscardManager a type of it, an implementation?

Quoted 8 lines

We want to use 'MemoryAttributeManager' to represent RAMBlock to implement the RamDiscardManager interface callbacks because RAMBlock is not an object. It includes some metadata of guest_memfd like shared_bitmap at the same time.

I can't get it that make 'MemoryAttributeManager' an interface and RamDiscardManager a type of it. Can you elaborate it a little bit? I think at least we need someone to implement the RamDiscardManager interface.

shared <-> private is translated (abstracted) to "populated <-> discarded", which makes sense. The other way around would be wrong.

It's going to be interesting once we have more logical states, for example supporting virtio-mem for confidential VMs.

Then we'd have "shared+populated, private+populated, shared+discard, private+discarded". Not sure if this could simply be achieved by allowing multiple RamDiscardManager that are effectively chained, or if we'd want a different interface.

Signature · 3 lines
Cheers,

David / dhildenb
#10 lore raw

Re: [PATCH v3 4/7] memory-attribute-manager: Introduce MemoryAttributeManager to manage RAMBLock with guest_memfd

Quoted 43 lines · David Hildenbrand
Quoted 33 lines · Chenyi Qiang
Quoted 23 lines · Pankaj
Quoted 22 lines · Chenyi Qiang

As the commit 852f0048f3 ("RAMBlock: make guest_memfd require uncoordinated discard") highlighted, some subsystems like VFIO may disable ram block discard. However, guest_memfd relies on the discard operation to perform page conversion between private and shared memory. This can lead to stale IOMMU mapping issue when assigning a hardware device to a confidential VM via shared memory. To address this, it is crucial to ensure systems like VFIO refresh its IOMMU mappings.

RamDiscardManager is an existing concept (used by virtio-mem) to adjust VFIO mappings in relation to VM page assignment. Effectively page conversion is similar to hot-removing a page in one mode and adding it back in the other. Therefore, similar actions are required for page conversion events. Introduce the RamDiscardManager to guest_memfd to facilitate this process.

Since guest_memfd is not an object, it cannot directly implement the RamDiscardManager interface. One potential attempt is to implement it in HostMemoryBackend. This is not appropriate because guest_memfd is per RAMBlock. Some RAMBlocks have a memory backend but others do not. In particular, the ones like virtual BIOS calling memory_region_init_ram_guest_memfd() do not.

Quoted 6 lines
Quoted 2 lines

To manage the RAMBlocks with guest_memfd, define a new object named MemoryAttributeManager to implement the RamDiscardManager interface. The

Isn't this should be the other way around. 'MemoryAttributeManager' should be an interface and RamDiscardManager a type of it, an implementation?

Quoted 8 lines

We want to use 'MemoryAttributeManager' to represent RAMBlock to implement the RamDiscardManager interface callbacks because RAMBlock is not an object. It includes some metadata of guest_memfd like shared_bitmap at the same time.

I can't get it that make 'MemoryAttributeManager' an interface and RamDiscardManager a type of it. Can you elaborate it a little bit? I think at least we need someone to implement the RamDiscardManager interface.

Quoted 5 lines

shared <-> private is translated (abstracted) to "populated <-> discarded", which makes sense. The other way around would be wrong.

It's going to be interesting once we have more logical states, for example supporting virtio-mem for confidential VMs.

Then we'd have "shared+populated, private+populated, shared+discard, private+discarded". Not sure if this could simply be achieved by allowing multiple RamDiscardManager that are effectively chained, or if we'd want a different interface.

Exactly! In any case generic manager (parent class) would make more sense that can work on different operations/states implemented in child classes (can be chained as well).

Best regards,
Pankaj

Quoted 1 line
#11 lore raw

Re: [PATCH v3 4/7] memory-attribute-manager: Introduce MemoryAttributeManager to manage RAMBLock with guest_memfd

Quoted 55 lines · Pankaj
Quoted 54 lines · David Hildenbrand
Quoted 42 lines · Chenyi Qiang
Quoted 26 lines · Pankaj
Quoted 25 lines · Chenyi Qiang

As the commit 852f0048f3 ("RAMBlock: make guest_memfd require uncoordinated discard") highlighted, some subsystems like VFIO may disable ram block discard. However, guest_memfd relies on the discard operation to perform page conversion between private and shared memory.
This can lead to stale IOMMU mapping issue when assigning a hardware device to a confidential VM via shared memory. To address this, it is crucial to ensure systems like VFIO refresh its IOMMU mappings.

RamDiscardManager is an existing concept (used by virtio-mem) to adjust
VFIO mappings in relation to VM page assignment. Effectively page conversion is similar to hot-removing a page in one mode and adding it back in the other. Therefore, similar actions are required for page conversion events. Introduce the RamDiscardManager to guest_memfd to facilitate this process.

Since guest_memfd is not an object, it cannot directly implement the RamDiscardManager interface. One potential attempt is to implement it in
HostMemoryBackend. This is not appropriate because guest_memfd is per RAMBlock. Some RAMBlocks have a memory backend but others do not. In particular, the ones like virtual BIOS calling memory_region_init_ram_guest_memfd() do not.

Quoted 7 lines
Quoted 3 lines

To manage the RAMBlocks with guest_memfd, define a new object named MemoryAttributeManager to implement the RamDiscardManager interface. The

Isn't this should be the other way around. 'MemoryAttributeManager' should be an interface and RamDiscardManager a type of it, an implementation?

We want to use 'MemoryAttributeManager' to represent RAMBlock to implement the RamDiscardManager interface callbacks because RAMBlock is not an object. It includes some metadata of guest_memfd like shared_bitmap at the same time.

Quoted 4 lines

I can't get it that make 'MemoryAttributeManager' an interface and RamDiscardManager a type of it. Can you elaborate it a little bit? I think at least we need someone to implement the RamDiscardManager interface.

shared <-> private is translated (abstracted) to "populated <-> discarded", which makes sense. The other way around would be wrong.

It's going to be interesting once we have more logical states, for example supporting virtio-mem for confidential VMs.

Quoted 8 lines
Quoted 4 lines

Then we'd have "shared+populated, private+populated, shared+discard, private+discarded". Not sure if this could simply be achieved by allowing multiple RamDiscardManager that are effectively chained, or if we'd want a different interface.

Exactly! In any case generic manager (parent class) would make more sense that can work on different operations/states implemented in child classes (can be chained as well).

Ah, we are talking about the generic state management. Sorry for my slow reaction.

So we need to
1. Define a generic manager Interface, e.g. MemoryStateManager/GenericStateManager.
2. Make RamDiscardManager the child of MemoryStateManager which manages the state of populated and discarded.
3. Define a new child manager Interface PrivateSharedManager which manages the state of private and shared.
4. Define a new object ConfidentialMemoryAttribute to implement the PrivateSharedManager interface.
(Welcome to rename the above Interface/Object)

Is my understanding correct?

Quoted 5 lines

Best regards,
Pankaj

Quoted 1 line
#12 lore raw

Re: [PATCH v3 4/7] memory-attribute-manager: Introduce MemoryAttributeManager to manage RAMBLock with guest_memfd

Quoted 73 lines
Quoted 50 lines
Quoted 50 lines
Quoted 39 lines
Quoted 26 lines
Quoted 26 lines

As the commit 852f0048f3 ("RAMBlock: make guest_memfd require uncoordinated discard") highlighted, some subsystems like VFIO may disable ram block discard. However, guest_memfd relies on the discard operation to perform page conversion between private and shared memory.
This can lead to stale IOMMU mapping issue when assigning a hardware device to a confidential VM via shared memory. To address this, it is crucial to ensure systems like VFIO refresh its IOMMU mappings.

RamDiscardManager is an existing concept (used by virtio-mem) to adjust
VFIO mappings in relation to VM page assignment. Effectively page conversion is similar to hot-removing a page in one mode and adding it back in the other. Therefore, similar actions are required for page conversion events. Introduce the RamDiscardManager to guest_memfd to facilitate this process.

Since guest_memfd is not an object, it cannot directly implement the RamDiscardManager interface. One potential attempt is to implement it in
HostMemoryBackend. This is not appropriate because guest_memfd is per RAMBlock. Some RAMBlocks have a memory backend but others do not. In particular, the ones like virtual BIOS calling memory_region_init_ram_guest_memfd() do not.

Quoted 7 lines
Quoted 3 lines

To manage the RAMBlocks with guest_memfd, define a new object named MemoryAttributeManager to implement the RamDiscardManager interface. The

Isn't this should be the other way around. 'MemoryAttributeManager' should be an interface and RamDiscardManager a type of it, an implementation?

We want to use 'MemoryAttributeManager' to represent RAMBlock to implement the RamDiscardManager interface callbacks because RAMBlock is not an object. It includes some metadata of guest_memfd like shared_bitmap at the same time.

Quoted 4 lines

I can't get it that make 'MemoryAttributeManager' an interface and RamDiscardManager a type of it. Can you elaborate it a little bit? I think at least we need someone to implement the RamDiscardManager interface.

shared <-> private is translated (abstracted) to "populated <-> discarded", which makes sense. The other way around would be wrong.

It's going to be interesting once we have more logical states, for example supporting virtio-mem for confidential VMs.

Quoted 8 lines
Quoted 4 lines

Then we'd have "shared+populated, private+populated, shared+discard, private+discarded". Not sure if this could simply be achieved by allowing multiple RamDiscardManager that are effectively chained, or if we'd want a different interface.

Exactly! In any case generic manager (parent class) would make more sense that can work on different operations/states implemented in child classes (can be chained as well).

Ah, we are talking about the generic state management. Sorry for my slow reaction.

So we need to
1. Define a generic manager Interface, e.g. MemoryStateManager/GenericStateManager.
2. Make RamDiscardManager the child of MemoryStateManager which manages the state of populated and discarded.
3. Define a new child manager Interface PrivateSharedManager which manages the state of private and shared.
4. Define a new object ConfidentialMemoryAttribute to implement the PrivateSharedManager interface.
(Welcome to rename the above Interface/Object)

Quoted 1 line

Is my understanding correct?

Yes, in that direction. Where 'RamDiscardManager' & 'PrivateSharedManager' are both child of 'GenericStateManager'.

Depending on listeners registered, corresponding handlers can be called.

Best regards,
Pankaj

#13 lore raw

Re: [PATCH v3 4/7] memory-attribute-manager: Introduce MemoryAttributeManager to manage RAMBLock with guest_memfd

Quoted 78 lines · Pankaj
Quoted 78 lines
Quoted 55 lines
Quoted 55 lines
Quoted 44 lines
Quoted 30 lines
Quoted 30 lines

As the commit 852f0048f3 ("RAMBlock: make guest_memfd require uncoordinated discard") highlighted, some subsystems like VFIO may disable ram block discard. However, guest_memfd relies on the discard
operation to perform page conversion between private and shared memory.
This can lead to stale IOMMU mapping issue when assigning a hardware device to a confidential VM via shared memory. To address this, it is
crucial to ensure systems like VFIO refresh its IOMMU mappings.

RamDiscardManager is an existing concept (used by virtio-mem) to adjust
VFIO mappings in relation to VM page assignment. Effectively page conversion is similar to hot-removing a page in one mode and adding it
back in the other. Therefore, similar actions are required for page conversion events. Introduce the RamDiscardManager to guest_memfd to facilitate this process.

Since guest_memfd is not an object, it cannot directly implement the RamDiscardManager interface. One potential attempt is to implement it in
HostMemoryBackend. This is not appropriate because guest_memfd is per
RAMBlock. Some RAMBlocks have a memory backend but others do not. In particular, the ones like virtual BIOS calling memory_region_init_ram_guest_memfd() do not.

Quoted 7 lines
Quoted 3 lines

To manage the RAMBlocks with guest_memfd, define a new object named MemoryAttributeManager to implement the RamDiscardManager interface. The

Isn't this should be the other way around. 'MemoryAttributeManager' should be an interface and RamDiscardManager a type of it, an implementation?

We want to use 'MemoryAttributeManager' to represent RAMBlock to implement the RamDiscardManager interface callbacks because RAMBlock is
not an object. It includes some metadata of guest_memfd like shared_bitmap at the same time.

Quoted 4 lines

I can't get it that make 'MemoryAttributeManager' an interface and RamDiscardManager a type of it. Can you elaborate it a little bit? I think at least we need someone to implement the RamDiscardManager interface.

shared <-> private is translated (abstracted) to "populated <-> discarded", which makes sense. The other way around would be wrong.

It's going to be interesting once we have more logical states, for example supporting virtio-mem for confidential VMs.

Quoted 8 lines
Quoted 4 lines

Then we'd have "shared+populated, private+populated, shared+discard, private+discarded". Not sure if this could simply be achieved by allowing multiple RamDiscardManager that are effectively chained, or if we'd want a different interface.

Exactly! In any case generic manager (parent class) would make more sense that can work on different operations/states implemented in child classes (can be chained as well).

Ah, we are talking about the generic state management. Sorry for my slow reaction.

So we need to
1. Define a generic manager Interface, e.g. MemoryStateManager/GenericStateManager.
2. Make RamDiscardManager the child of MemoryStateManager which manages the state of populated and discarded.
3. Define a new child manager Interface PrivateSharedManager which manages the state of private and shared.
4. Define a new object ConfidentialMemoryAttribute to implement the PrivateSharedManager interface.
(Welcome to rename the above Interface/Object)

Quoted 6 lines
Quoted 1 line

Is my understanding correct?

Yes, in that direction. Where 'RamDiscardManager' & 'PrivateSharedManager' are both child of 'GenericStateManager'.

Depending on listeners registered, corresponding handlers can be called.

Yes, it would be more generic and future extensive.

Do we need to add this framework change directly? Or keep the current structure (abstract private/shared as discard/populated) and add the generic manager until the real case like virtio-mem for confidential VMs.

Quoted 4 lines

Best regards,
Pankaj

#14 lore raw

Re: [PATCH v3 4/7] memory-attribute-manager: Introduce MemoryAttributeManager to manage RAMBLock with guest_memfd

Quoted 86 lines · Chenyi Qiang
Quoted 83 lines · Pankaj
Quoted 78 lines
Quoted 55 lines
Quoted 55 lines
Quoted 44 lines
Quoted 30 lines
Quoted 30 lines

As the commit 852f0048f3 ("RAMBlock: make guest_memfd require uncoordinated discard") highlighted, some subsystems like VFIO may disable ram block discard. However, guest_memfd relies on the discard
operation to perform page conversion between private and shared memory.
This can lead to stale IOMMU mapping issue when assigning a hardware device to a confidential VM via shared memory. To address this, it is
crucial to ensure systems like VFIO refresh its IOMMU mappings.

RamDiscardManager is an existing concept (used by virtio-mem) to adjust
VFIO mappings in relation to VM page assignment. Effectively page conversion is similar to hot-removing a page in one mode and adding it
back in the other. Therefore, similar actions are required for page conversion events. Introduce the RamDiscardManager to guest_memfd to facilitate this process.

Since guest_memfd is not an object, it cannot directly implement the RamDiscardManager interface. One potential attempt is to implement it in
HostMemoryBackend. This is not appropriate because guest_memfd is per
RAMBlock. Some RAMBlocks have a memory backend but others do not. In particular, the ones like virtual BIOS calling memory_region_init_ram_guest_memfd() do not.

Quoted 7 lines
Quoted 3 lines

To manage the RAMBlocks with guest_memfd, define a new object named MemoryAttributeManager to implement the RamDiscardManager interface. The

Isn't this should be the other way around. 'MemoryAttributeManager' should be an interface and RamDiscardManager a type of it, an implementation?

We want to use 'MemoryAttributeManager' to represent RAMBlock to implement the RamDiscardManager interface callbacks because RAMBlock is
not an object. It includes some metadata of guest_memfd like shared_bitmap at the same time.

Quoted 4 lines

I can't get it that make 'MemoryAttributeManager' an interface and RamDiscardManager a type of it. Can you elaborate it a little bit? I think at least we need someone to implement the RamDiscardManager interface.

shared <-> private is translated (abstracted) to "populated <-> discarded", which makes sense. The other way around would be wrong.

It's going to be interesting once we have more logical states, for example supporting virtio-mem for confidential VMs.

Quoted 8 lines
Quoted 4 lines

Then we'd have "shared+populated, private+populated, shared+discard, private+discarded". Not sure if this could simply be achieved by allowing multiple RamDiscardManager that are effectively chained, or if we'd want a different interface.

Exactly! In any case generic manager (parent class) would make more sense that can work on different operations/states implemented in child classes (can be chained as well).

Ah, we are talking about the generic state management. Sorry for my slow reaction.

So we need to
1. Define a generic manager Interface, e.g. MemoryStateManager/GenericStateManager.
2. Make RamDiscardManager the child of MemoryStateManager which manages the state of populated and discarded.
3. Define a new child manager Interface PrivateSharedManager which manages the state of private and shared.
4. Define a new object ConfidentialMemoryAttribute to implement the PrivateSharedManager interface.
(Welcome to rename the above Interface/Object)

Quoted 1 line

Is my understanding correct?

Yes, in that direction. Where 'RamDiscardManager' & 'PrivateSharedManager' are both child of 'GenericStateManager'.

Quoted 8 lines
Quoted 1 line

Depending on listeners registered, corresponding handlers can be called.

Yes, it would be more generic and future extensive.

Do we need to add this framework change directly? Or keep the current structure (abstract private/shared as discard/populated) and add the generic manager until the real case like virtio-mem for confidential VMs.

Yes, maybe to start with we should add new (discard/populated) changes with the new framework.

In future the current framework can be extended for in-place conversion for private-shared conversion (if require userspace help) and virtio-mem like interfaces. Important is to have proper hierarchy with base bits there.

Thanks,
Pankaj

#15 lore raw

Re: [PATCH v3 4/7] memory-attribute-manager: Introduce MemoryAttributeManager to manage RAMBLock with guest_memfd

Quoted 104 lines · Pankaj
Quoted 94 lines · Chenyi Qiang
Quoted 91 lines · Pankaj
Quoted 86 lines
Quoted 66 lines
Quoted 58 lines
Quoted 50 lines
Quoted 35 lines
Quoted 35 lines

As the commit 852f0048f3 ("RAMBlock: make guest_memfd require uncoordinated discard") highlighted, some subsystems like VFIO may disable ram block discard. However, guest_memfd relies on the discard
operation to perform page conversion between private and shared memory.
This can lead to stale IOMMU mapping issue when assigning a hardware
device to a confidential VM via shared memory. To address this, it is
crucial to ensure systems like VFIO refresh its IOMMU mappings.

RamDiscardManager is an existing concept (used by virtio-mem) to adjust
VFIO mappings in relation to VM page assignment. Effectively page conversion is similar to hot-removing a page in one mode and adding it
back in the other. Therefore, similar actions are required for page
conversion events. Introduce the RamDiscardManager to guest_memfd to
facilitate this process.

Since guest_memfd is not an object, it cannot directly implement the
RamDiscardManager interface. One potential attempt is to implement it in
HostMemoryBackend. This is not appropriate because guest_memfd is per
RAMBlock. Some RAMBlocks have a memory backend but others do not. In
particular, the ones like virtual BIOS calling memory_region_init_ram_guest_memfd() do not.

Quoted 8 lines
Quoted 4 lines

To manage the RAMBlocks with guest_memfd, define a new object named
MemoryAttributeManager to implement the RamDiscardManager interface. The

Isn't this should be the other way around. 'MemoryAttributeManager' should be an interface and RamDiscardManager a type of it, an implementation?

We want to use 'MemoryAttributeManager' to represent RAMBlock to implement the RamDiscardManager interface callbacks because RAMBlock is
not an object. It includes some metadata of guest_memfd like shared_bitmap at the same time.

Quoted 4 lines

I can't get it that make 'MemoryAttributeManager' an interface and RamDiscardManager a type of it. Can you elaborate it a little bit? I think at least we need someone to implement the RamDiscardManager interface.

shared <-> private is translated (abstracted) to "populated <-> discarded", which makes sense. The other way around would be wrong.

Quoted 7 lines

It's going to be interesting once we have more logical states, for example supporting virtio-mem for confidential VMs.

Then we'd have "shared+populated, private+populated, shared+discard, private+discarded". Not sure if this could simply be achieved by allowing multiple RamDiscardManager that are effectively chained, or if we'd want a different interface.

Quoted 4 lines

Exactly! In any case generic manager (parent class) would make more sense that can work on different operations/states implemented in child
classes (can be chained as well).

Ah, we are talking about the generic state management. Sorry for my slow
reaction.

So we need to
1. Define a generic manager Interface, e.g. MemoryStateManager/GenericStateManager.
2. Make RamDiscardManager the child of MemoryStateManager which manages the state of populated and discarded.
3. Define a new child manager Interface PrivateSharedManager which manages the state of private and shared.
4. Define a new object ConfidentialMemoryAttribute to implement the PrivateSharedManager interface.
(Welcome to rename the above Interface/Object)

Quoted 1 line

Is my understanding correct?

Yes, in that direction. Where 'RamDiscardManager' & 'PrivateSharedManager' are both child of 'GenericStateManager'.

Quoted 8 lines
Quoted 1 line

Depending on listeners registered, corresponding handlers can be called.

Yes, it would be more generic and future extensive.

Do we need to add this framework change directly? Or keep the current structure (abstract private/shared as discard/populated) and add the generic manager until the real case like virtio-mem for confidential VMs.

Quoted 7 lines

Yes, maybe to start with we should add new (discard/populated) changes with the new framework.

In future the current framework can be extended for in-place conversion for private-shared conversion (if require userspace help) and virtio-mem like interfaces. Important is to have proper hierarchy with base bits there.

Thanks. Then I will follow this direction.

To abstract the common parent class, what I can think of is to abstract it to manage a pair of opposite states (state set and state clear, like populate and discard) and define some similar common callbacks like notify_set() and notify_clear(), as long as we don't use it to manage more than two states in the future. Otherwise I may define a stub parent class.

Quoted 3 lines

Thanks,
Pankaj

#16 lore raw

Re: [PATCH v3 4/7] memory-attribute-manager: Introduce MemoryAttributeManager to manage RAMBLock with guest_memfd

Quoted 115 lines · Chenyi Qiang
Quoted 104 lines · Pankaj
Quoted 94 lines · Chenyi Qiang
Quoted 91 lines · Pankaj
Quoted 86 lines
Quoted 66 lines
Quoted 58 lines
Quoted 50 lines
Quoted 35 lines
Quoted 35 lines

As the commit 852f0048f3 ("RAMBlock: make guest_memfd require uncoordinated discard") highlighted, some subsystems like VFIO may disable ram block discard. However, guest_memfd relies on the discard
operation to perform page conversion between private and shared memory.
This can lead to stale IOMMU mapping issue when assigning a hardware
device to a confidential VM via shared memory. To address this, it is
crucial to ensure systems like VFIO refresh its IOMMU mappings.

RamDiscardManager is an existing concept (used by virtio-mem) to adjust
VFIO mappings in relation to VM page assignment. Effectively page conversion is similar to hot-removing a page in one mode and adding it
back in the other. Therefore, similar actions are required for page
conversion events. Introduce the RamDiscardManager to guest_memfd to
facilitate this process.

Since guest_memfd is not an object, it cannot directly implement the
RamDiscardManager interface. One potential attempt is to implement it in
HostMemoryBackend. This is not appropriate because guest_memfd is per
RAMBlock. Some RAMBlocks have a memory backend but others do not. In
particular, the ones like virtual BIOS calling memory_region_init_ram_guest_memfd() do not.

Quoted 8 lines
Quoted 4 lines

To manage the RAMBlocks with guest_memfd, define a new object named
MemoryAttributeManager to implement the RamDiscardManager interface. The

Isn't this should be the other way around. 'MemoryAttributeManager' should be an interface and RamDiscardManager a type of it, an implementation?

We want to use 'MemoryAttributeManager' to represent RAMBlock to implement the RamDiscardManager interface callbacks because RAMBlock is
not an object. It includes some metadata of guest_memfd like shared_bitmap at the same time.

Quoted 4 lines

I can't get it that make 'MemoryAttributeManager' an interface and RamDiscardManager a type of it. Can you elaborate it a little bit? I think at least we need someone to implement the RamDiscardManager interface.

shared <-> private is translated (abstracted) to "populated <-> discarded", which makes sense. The other way around would be wrong.

Quoted 7 lines

It's going to be interesting once we have more logical states, for example supporting virtio-mem for confidential VMs.

Then we'd have "shared+populated, private+populated, shared+discard, private+discarded". Not sure if this could simply be achieved by allowing multiple RamDiscardManager that are effectively chained, or if we'd want a different interface.

Quoted 4 lines

Exactly! In any case generic manager (parent class) would make more sense that can work on different operations/states implemented in child
classes (can be chained as well).

Ah, we are talking about the generic state management. Sorry for my slow
reaction.

So we need to
1. Define a generic manager Interface, e.g. MemoryStateManager/GenericStateManager.
2. Make RamDiscardManager the child of MemoryStateManager which manages the state of populated and discarded.
3. Define a new child manager Interface PrivateSharedManager which manages the state of private and shared.
4. Define a new object ConfidentialMemoryAttribute to implement the PrivateSharedManager interface.
(Welcome to rename the above Interface/Object)

Quoted 1 line

Is my understanding correct?

Yes, in that direction. Where 'RamDiscardManager' & 'PrivateSharedManager' are both child of 'GenericStateManager'.

Quoted 8 lines
Quoted 1 line

Depending on listeners registered, corresponding handlers can be called.

Yes, it would be more generic and future extensive.

Do we need to add this framework change directly? Or keep the current structure (abstract private/shared as discard/populated) and add the generic manager until the real case like virtio-mem for confidential VMs.

Quoted 7 lines

Yes, maybe to start with we should add new (discard/populated) changes with the new framework.

In future the current framework can be extended for in-place conversion for private-shared conversion (if require userspace help) and virtio-mem like interfaces. Important is to have proper hierarchy with base bits there.

Quoted 8 lines

Thanks. Then I will follow this direction.

To abstract the common parent class, what I can think of is to abstract it to manage a pair of opposite states (state set and state clear, like populate and discard) and define some similar common callbacks like notify_set() and notify_clear(), as long as we don't use it to manage more than two states in the future. Otherwise I may define a stub parent class.

Hi Pankaj & David,

How about defining them like:

---
For class definition: extract all the callbacks into parent class:

typedef int (*ReplayStateChange)(MemoryRegionSection *section, void *opaque);

struct GenericStateManagerClass {
    InterfaceClass parent_class;
    uint64_t (*get_min_granularity)(const GenericStateManager *gsm,
                                    const MemoryRegion *mr);
    bool (*is_state_set)(const GenericStateManager *gsm,
                         const MemoryRegionSection *section);
    int (*replay_state_set)(const GenericStateManager *gsm,
                            MemoryRegionSection *section,
                            ReplayStateChange replay_fn, void *opaque);
    int (*replay_state_clear)(const GenericStateManager *gsm,
                              MemoryRegionSection *section,
                              ReplayStateChange replay_fn, void *opaque);
    void (*register_listener)(GenericStateManager *gsm,
                              void* listener,
                              MemoryRegionSection *section);
    void (*unregister_listener)(GenericStateManager *gsm,
                                void *listener);
}
struct RamDiscardManagerClass {
    /* private */
    GenericStateManagerClass parent_class;
};
struct PrivateSharedManagerClass {
    /* private */
    GenericStateManagerClass parent_class;
}

---

For listeners, embed the generic listener into specific listeners:

typedef int (*NotifyStateSet)(void *listener, MemoryRegionSection *section); typedef int (*NotifyStateClear)(void *listener, MemoryRegionSection *section);

struct StateChangeListener {
    NotifyStateSet notify_state_set;
    NotifyStateClear notify_state_clear;

    MemoryRegionSection *section;
}
struct RamDiscardListener {
    struct StateChangeListener scl;
    bool double_discard_supported;

    QLIST_ENTRY(RamDiscardListener) next;
};
struct PrivateSharedListener {
    struct StateChangeListener scl;

    QLIST_ENTRY(PrivateSharedListener) next;
}

---

For helpers,

- define the generic helpers:

void generic_state_manager_register_listener(GenericStateManager *gsm,
void *listener, MemoryRegionSection *s)

{
    GenericStateManagerClass gsmc = GENERATE_STATE_MANAGER_GET_CLASS(rdm);

    g_assert(gsmc->register_listener);
    gsmc->register_listener(gsm, listener, s);
}

- Keep RamDiscardManager specific helpers for compatibility if necessary:

void ram_discard_manager_register_listener(RamDiscardManager *rdm,
RamDiscardListener *rdl, MemoryRegionSection *section)

{
    GenericStateManagerClass gsmc = GENERATE_STATE_MANAGER_GET_CLASS(rdm);
    GenericStateManager *gsm = GENERIC_STATE_MANAGER(rdm);

    g_assert(gsmc->register_listener);
    gsmc->register_listener(gsm, (void*)rdl, section);
}
Quoted 5 lines
Quoted 4 lines

Thanks,
Pankaj

#17 lore raw

PATCH v3 5/7 memory-attribute-manager: Introduce a callback to notify the shared/private state change

In reply to Chenyi Qiang

Introduce a new state_change() callback in MemoryAttributeManagerClass to efficiently notify all registered RamDiscardListeners, including VFIO listeners about the memory conversion events in guest_memfd. The existing VFIO listener can dynamically DMA map/unmap the shared pages based on conversion types:
- For conversions from shared to private, the VFIO system ensures the discarding of shared mapping from the IOMMU.
- For conversions from private to shared, it triggers the population of the shared mapping into the IOMMU.

Additionally, there could be some special conversion requests:
- When a conversion request is made for a page already in the desired state, the helper simply returns success.
- For requests involving a range partially in the desired state, only the necessary segments are converted, ensuring the entire range complies with the request efficiently. In this case, fallback to a "1 block at a time" handling.
- In scenarios where a conversion request is declined by other systems, such as a failure from VFIO during notify_populate(), the helper will roll back the request, maintaining consistency.

Note that the bitmap status is updated before the notifier callbacks so that the listener can handle the memory based on the latest status.

Opportunistically introduce a helper to trigger the state_change() callback of the class.

---
Changes in v3:
- Move the bitmap update before notifier callbacks.
- Call the notifier callbacks directly in notify_discard/populate() with the expectation that the request memory range is in the desired attribute.
- For the case that only partial range in the desire status, handle the range with block_size granularity for ease of rollback
(https://lore.kernel.org/qemu-devel/[email protected]/)

Changes in v2:
- Do the alignment changes due to the rename to MemoryAttributeManager
- Move the state_change() helper definition in this patch.

---
 include/system/memory-attribute-manager.h |  18 +++
 system/memory-attribute-manager.c         | 188 ++++++++++++++++++++++
 2 files changed, 206 insertions(+)
include/system/memory-attribute-manager.h +18 -0
index 23375a14b8..3d9227d62a 100644
@@ -34,8 +34,26 @@ struct MemoryAttributeManager {
 
 struct MemoryAttributeManagerClass {
     ObjectClass parent_class;
+
+    int (*state_change)(MemoryAttributeManager *mgr, uint64_t offset, uint64_t size,
+                        bool to_private);
 };
 
+static inline int memory_attribute_manager_state_change(MemoryAttributeManager *mgr, uint64_t offset,
+                                                        uint64_t size, bool to_private)
+{
+    MemoryAttributeManagerClass *klass;
+
+    if (mgr == NULL) {
+        return 0;
+    }
+
+    klass = MEMORY_ATTRIBUTE_MANAGER_GET_CLASS(mgr);
+
+    g_assert(klass->state_change);
+    return klass->state_change(mgr, offset, size, to_private);
+}
+
 int memory_attribute_manager_realize(MemoryAttributeManager *mgr, MemoryRegion *mr);
 void memory_attribute_manager_unrealize(MemoryAttributeManager *mgr);
 
system/memory-attribute-manager.c +188 -0

202 lines not on this page · raw message

Signature · 1 line
2.43.5
#18 lore raw

PATCH v3 6/7 memory: Attach MemoryAttributeManager to guest_memfd-backed RAMBlocks

In reply to Chenyi Qiang

Introduce a new field, memory_attribute_manager, in RAMBlock to link to an MemoryAttributeManager object. This change centralizes all guest_memfd state information (like fd and shared_bitmap) within a RAMBlock, making it easier to manage.

Use the realize()/unrealize() helpers to initialize/uninitialize the MemoryAttributeManager object. Register/unregister the object in the target RAMBlock's MemoryRegion when creating guest_memfd.

In the kvm_convert_memory() function, manage memory state changes by using the shared_bitmap to call set_attribute() only on the specific memory range. Additionally, use the memory_attribute_manager_state_change() helper to notify the reigstered RamDiscardListener of these changes.

---
Changes in v3:
- Use ram_discard_manager_reply_populated/discarded() to set the memory attribute and add the undo support if state_change() failed.
- Didn't add Reviewed-by from Alexey due to the new changes in this commit.

Changes in v2:
- Introduce a new field memory_attribute_manager in RAMBlock.
- Move the state_change() handling during page conversion in this patch.
- Undo what we did if it fails to set.
- Change the order of close(guest_memfd) and memory_attribute_manager cleanup.

---
 accel/kvm/kvm-all.c     | 50 +++++++++++++++++++++++++++++++++++++++--
 include/exec/ramblock.h |  2 ++
 system/physmem.c        | 13 +++++++++++
 3 files changed, 63 insertions(+), 2 deletions(-)
accel/kvm/kvm-all.c +48 -2
index c1fea69d58..a89c5655e8 100644
@@ -48,6 +48,7 @@
 #include "kvm-cpus.h"
 #include "system/dirtylimit.h"
 #include "qemu/range.h"
+#include "system/memory-attribute-manager.h"
 
 #include "hw/boards.h"
 #include "system/stats.h"
@@ -3018,6 +3019,25 @@ static void kvm_eat_signals(CPUState *cpu)
     } while (sigismember(&chkset, SIG_IPI));
 }
 
+typedef struct SetMemoryAttribute {
+    bool to_private;
+} SetMemoryAttribute;
+
+static int kvm_set_memory_attributes_cb(MemoryRegionSection *section,
+                                        void *opaque)
+{
+    hwaddr start = section->offset_within_address_space;
+    hwaddr size = section->size;
+    SetMemoryAttribute *args = opaque;
+    bool to_private = args->to_private;
+
+    if (to_private) {
+        return kvm_set_memory_attributes_private(start, size);
+    } else {
+        return kvm_set_memory_attributes_shared(start, size);
+    }
+}
+
 int kvm_convert_memory(hwaddr start, hwaddr size, bool to_private)
 {
     MemoryRegionSection section;
@@ -3026,6 +3046,7 @@ int kvm_convert_memory(hwaddr start, hwaddr size, bool to_private)
     RAMBlock *rb;
     void *addr;
     int ret = -EINVAL;
+    SetMemoryAttribute args = { .to_private = to_private };
 
     trace_kvm_convert_memory(start, size, to_private ? "shared_to_private" : "private_to_shared");
 
@@ -3077,9 +3098,13 @@ int kvm_convert_memory(hwaddr start, hwaddr size, bool to_private)
     }
 
     if (to_private) {
-        ret = kvm_set_memory_attributes_private(start, size);
+        ret = ram_discard_manager_replay_populated(mr->rdm, &section,
+                                                   kvm_set_memory_attributes_cb,
+                                                   &args);
     } else {
-        ret = kvm_set_memory_attributes_shared(start, size);
+        ret = ram_discard_manager_replay_discarded(mr->rdm, &section,
+                                                   kvm_set_memory_attributes_cb,
+                                                   &args);
     }
     if (ret) {
         goto out_unref;
@@ -3088,6 +3113,27 @@ int kvm_convert_memory(hwaddr start, hwaddr size, bool to_private)
     addr = memory_region_get_ram_ptr(mr) + section.offset_within_region;
     rb = qemu_ram_block_from_host(addr, false, &offset);
 
+    ret = memory_attribute_manager_state_change(MEMORY_ATTRIBUTE_MANAGER(mr->rdm),
+                                                offset, size, to_private);
+    if (ret) {
+        warn_report("Failed to notify the listener the state change of "
+                    "(0x%"HWADDR_PRIx" + 0x%"HWADDR_PRIx") to %s",
+                    start, size, to_private ? "private" : "shared");
+        args.to_private = !to_private;
+        if (to_private) {
+            ret = ram_discard_manager_replay_populated(mr->rdm, &section,
+                                                       kvm_set_memory_attributes_cb,
+                                                       &args);
+        } else {
+            ret = ram_discard_manager_replay_discarded(mr->rdm, &section,
+                                                       kvm_set_memory_attributes_cb,
+                                                       &args);
+        }
+        if (ret) {
+            goto out_unref;
+        }
+    }
+
     if (to_private) {
         if (rb->page_size != qemu_real_host_page_size()) {
             /*
include/exec/ramblock.h +2 -0
index 0babd105c0..06fd365326 100644
@@ -23,6 +23,7 @@
 #include "cpu-common.h"
 #include "qemu/rcu.h"
 #include "exec/ramlist.h"
+#include "system/memory-attribute-manager.h"
 
 struct RAMBlock {
     struct rcu_head rcu;
@@ -42,6 +43,7 @@ struct RAMBlock {
     int fd;
     uint64_t fd_offset;
     int guest_memfd;
+    MemoryAttributeManager *memory_attribute_manager;
     size_t page_size;
     /* dirty bitmap used during migration */
     unsigned long *bmap;
system/physmem.c +13 -0
index c76503aea8..0ed394c5d2 100644
@@ -54,6 +54,7 @@
 #include "system/hostmem.h"
 #include "system/hw_accel.h"
 #include "system/xen-mapcache.h"
+#include "system/memory-attribute-manager.h"
 #include "trace.h"
 
 #ifdef CONFIG_FALLOCATE_PUNCH_HOLE
@@ -1885,6 +1886,16 @@ static void ram_block_add(RAMBlock *new_block, Error **errp)
             qemu_mutex_unlock_ramlist();
             goto out_free;
         }
+
+        new_block->memory_attribute_manager = MEMORY_ATTRIBUTE_MANAGER(object_new(TYPE_MEMORY_ATTRIBUTE_MANAGER));
+        if (memory_attribute_manager_realize(new_block->memory_attribute_manager, new_block->mr)) {
+            error_setg(errp, "Failed to realize memory attribute manager");
+            object_unref(OBJECT(new_block->memory_attribute_manager));
+            close(new_block->guest_memfd);
+            ram_block_discard_require(false);
+            qemu_mutex_unlock_ramlist();
+            goto out_free;
+        }
     }
 
     ram_size = (new_block->offset + new_block->max_length) >> TARGET_PAGE_BITS;
@@ -2138,6 +2149,8 @@ static void reclaim_ramblock(RAMBlock *block)
     }
 
     if (block->guest_memfd >= 0) {
+        memory_attribute_manager_unrealize(block->memory_attribute_manager);
+        object_unref(OBJECT(block->memory_attribute_manager));
         close(block->guest_memfd);
         ram_block_discard_require(false);
     }
Signature · 1 line
2.43.5
#19 lore raw

Re: [PATCH v3 6/7] memory: Attach MemoryAttributeManager to guest_memfd-backed RAMBlocks

Hi David & Alexey,

To keep the bitmap aligned, I add the undo operation for set_memory_attributes() and use the bitmap + replay callback to do set_memory_attributes(). Does this change make sense?

Alexey, I didn't add your Reivewed-by since this patch introduced some new changes.

Quoted 183 lines · Chenyi Qiang

Introduce a new field, memory_attribute_manager, in RAMBlock to link to an MemoryAttributeManager object. This change centralizes all guest_memfd state information (like fd and shared_bitmap) within a RAMBlock, making it easier to manage.

Use the realize()/unrealize() helpers to initialize/uninitialize the MemoryAttributeManager object. Register/unregister the object in the target RAMBlock's MemoryRegion when creating guest_memfd.

In the kvm_convert_memory() function, manage memory state changes by using the shared_bitmap to call set_attribute() only on the specific memory range. Additionally, use the memory_attribute_manager_state_change() helper to notify the reigstered RamDiscardListener of these changes.

---
Changes in v3:
- Use ram_discard_manager_reply_populated/discarded() to set the memory attribute and add the undo support if state_change() failed.
- Didn't add Reviewed-by from Alexey due to the new changes in this commit.

Changes in v2:
- Introduce a new field memory_attribute_manager in RAMBlock.
- Move the state_change() handling during page conversion in this patch.
- Undo what we did if it fails to set.
- Change the order of close(guest_memfd) and memory_attribute_manager cleanup.

---
 accel/kvm/kvm-all.c     | 50 +++++++++++++++++++++++++++++++++++++++--
 include/exec/ramblock.h |  2 ++
 system/physmem.c        | 13 +++++++++++
 3 files changed, 63 insertions(+), 2 deletions(-)
accel/kvm/kvm-all.c +48 -2
index c1fea69d58..a89c5655e8 100644
@@ -48,6 +48,7 @@
 #include "kvm-cpus.h"
 #include "system/dirtylimit.h"
 #include "qemu/range.h"
+#include "system/memory-attribute-manager.h"
 
 #include "hw/boards.h"
 #include "system/stats.h"
@@ -3018,6 +3019,25 @@ static void kvm_eat_signals(CPUState *cpu)
     } while (sigismember(&chkset, SIG_IPI));
 }
 
+typedef struct SetMemoryAttribute {
+    bool to_private;
+} SetMemoryAttribute;
+
+static int kvm_set_memory_attributes_cb(MemoryRegionSection *section,
+                                        void *opaque)
+{
+    hwaddr start = section->offset_within_address_space;
+    hwaddr size = section->size;
+    SetMemoryAttribute *args = opaque;
+    bool to_private = args->to_private;
+
+    if (to_private) {
+        return kvm_set_memory_attributes_private(start, size);
+    } else {
+        return kvm_set_memory_attributes_shared(start, size);
+    }
+}
+
 int kvm_convert_memory(hwaddr start, hwaddr size, bool to_private)
 {
     MemoryRegionSection section;
@@ -3026,6 +3046,7 @@ int kvm_convert_memory(hwaddr start, hwaddr size, bool to_private)
     RAMBlock *rb;
     void *addr;
     int ret = -EINVAL;
+    SetMemoryAttribute args = { .to_private = to_private };
 
     trace_kvm_convert_memory(start, size, to_private ? "shared_to_private" : "private_to_shared");
 
@@ -3077,9 +3098,13 @@ int kvm_convert_memory(hwaddr start, hwaddr size, bool to_private)
     }
 
     if (to_private) {
-        ret = kvm_set_memory_attributes_private(start, size);
+        ret = ram_discard_manager_replay_populated(mr->rdm, &section,
+                                                   kvm_set_memory_attributes_cb,
+                                                   &args);
     } else {
-        ret = kvm_set_memory_attributes_shared(start, size);
+        ret = ram_discard_manager_replay_discarded(mr->rdm, &section,
+                                                   kvm_set_memory_attributes_cb,
+                                                   &args);
     }
     if (ret) {
         goto out_unref;
@@ -3088,6 +3113,27 @@ int kvm_convert_memory(hwaddr start, hwaddr size, bool to_private)
     addr = memory_region_get_ram_ptr(mr) + section.offset_within_region;
     rb = qemu_ram_block_from_host(addr, false, &offset);
 
+    ret = memory_attribute_manager_state_change(MEMORY_ATTRIBUTE_MANAGER(mr->rdm),
+                                                offset, size, to_private);
+    if (ret) {
+        warn_report("Failed to notify the listener the state change of "
+                    "(0x%"HWADDR_PRIx" + 0x%"HWADDR_PRIx") to %s",
+                    start, size, to_private ? "private" : "shared");
+        args.to_private = !to_private;
+        if (to_private) {
+            ret = ram_discard_manager_replay_populated(mr->rdm, &section,
+                                                       kvm_set_memory_attributes_cb,
+                                                       &args);
+        } else {
+            ret = ram_discard_manager_replay_discarded(mr->rdm, &section,
+                                                       kvm_set_memory_attributes_cb,
+                                                       &args);
+        }
+        if (ret) {
+            goto out_unref;
+        }
+    }
+
     if (to_private) {
         if (rb->page_size != qemu_real_host_page_size()) {
             /*
include/exec/ramblock.h +2 -0
index 0babd105c0..06fd365326 100644
@@ -23,6 +23,7 @@
 #include "cpu-common.h"
 #include "qemu/rcu.h"
 #include "exec/ramlist.h"
+#include "system/memory-attribute-manager.h"
 
 struct RAMBlock {
     struct rcu_head rcu;
@@ -42,6 +43,7 @@ struct RAMBlock {
     int fd;
     uint64_t fd_offset;
     int guest_memfd;
+    MemoryAttributeManager *memory_attribute_manager;
     size_t page_size;
     /* dirty bitmap used during migration */
     unsigned long *bmap;
system/physmem.c +13 -0
index c76503aea8..0ed394c5d2 100644
@@ -54,6 +54,7 @@
 #include "system/hostmem.h"
 #include "system/hw_accel.h"
 #include "system/xen-mapcache.h"
+#include "system/memory-attribute-manager.h"
 #include "trace.h"
 
 #ifdef CONFIG_FALLOCATE_PUNCH_HOLE
@@ -1885,6 +1886,16 @@ static void ram_block_add(RAMBlock *new_block, Error **errp)
             qemu_mutex_unlock_ramlist();
             goto out_free;
         }
+
+        new_block->memory_attribute_manager = MEMORY_ATTRIBUTE_MANAGER(object_new(TYPE_MEMORY_ATTRIBUTE_MANAGER));
+        if (memory_attribute_manager_realize(new_block->memory_attribute_manager, new_block->mr)) {
+            error_setg(errp, "Failed to realize memory attribute manager");
+            object_unref(OBJECT(new_block->memory_attribute_manager));
+            close(new_block->guest_memfd);
+            ram_block_discard_require(false);
+            qemu_mutex_unlock_ramlist();
+            goto out_free;
+        }
     }
 
     ram_size = (new_block->offset + new_block->max_length) >> TARGET_PAGE_BITS;
@@ -2138,6 +2149,8 @@ static void reclaim_ramblock(RAMBlock *block)
     }
 
     if (block->guest_memfd >= 0) {
+        memory_attribute_manager_unrealize(block->memory_attribute_manager);
+        object_unref(OBJECT(block->memory_attribute_manager));
         close(block->guest_memfd);
         ram_block_discard_require(false);
     }
#20 lore raw

Re: [PATCH v3 6/7] memory: Attach MemoryAttributeManager to guest_memfd-backed RAMBlocks

Quoted 1 line · Chenyi Qiang

Hi David & Alexey,

Hi,

Quoted 4 lines

To keep the bitmap aligned, I add the undo operation for set_memory_attributes() and use the bitmap + replay callback to do set_memory_attributes(). Does this change make sense?

I assume you mean this hunk:

+ ret = memory_attribute_manager_state_change(MEMORY_ATTRIBUTE_MANAGER(mr->rdm),
+ offset, size, to_private);
+ if (ret) {
+ warn_report("Failed to notify the listener the state change of "
+ "(0x%"HWADDR_PRIx" + 0x%"HWADDR_PRIx") to %s",
+ start, size, to_private ? "private" : "shared");
+ args.to_private = !to_private;
+ if (to_private) {
+ ret = ram_discard_manager_replay_populated(mr->rdm, &section,
+ kvm_set_memory_attributes_cb,
+ &args);
+ } else {
+ ret = ram_discard_manager_replay_discarded(mr->rdm, &section,
+ kvm_set_memory_attributes_cb,
+ &args);
+ }
+ if (ret) {
+ goto out_unref;
+ }

Why is that undo necessary? The bitmap + listeners should be held in sync inside of memory_attribute_manager_state_change(). Handling this in the caller looks wrong.

I thought the current implementation properly handles that internally? In which scenario is that not the case?

Signature · 3 lines
Cheers,

David / dhildenb
#21 lore raw

Re: [PATCH v3 6/7] memory: Attach MemoryAttributeManager to guest_memfd-backed RAMBlocks

Quoted 36 lines · David Hildenbrand
Quoted 1 line · Chenyi Qiang

Hi David & Alexey,

Hi,

Quoted 4 lines

To keep the bitmap aligned, I add the undo operation for set_memory_attributes() and use the bitmap + replay callback to do set_memory_attributes(). Does this change make sense?

I assume you mean this hunk:

+    ret = memory_attribute_manager_state_change(MEMORY_ATTRIBUTE_MANAGER(mr->rdm), +                                                offset, size, to_private); +    if (ret) {
+        warn_report("Failed to notify the listener the state change of " +                    "(0x%"HWADDR_PRIx" + 0x%"HWADDR_PRIx") to %s", +                    start, size, to_private ? "private" : "shared"); +        args.to_private = !to_private;
+        if (to_private) {
+            ret = ram_discard_manager_replay_populated(mr->rdm, &section, +
kvm_set_memory_attributes_cb,
+                                                       &args); +        } else {
+            ret = ram_discard_manager_replay_discarded(mr->rdm, &section, +
kvm_set_memory_attributes_cb,
+                                                       &args); +        }
+        if (ret) {
+            goto out_unref;
+        }

Quoted 4 lines

Why is that undo necessary? The bitmap + listeners should be held in sync inside of
memory_attribute_manager_state_change(). Handling this in the caller looks wrong.

state_change() handles the listener, i.e. VFIO/IOMMU. And the caller handles the core mm side (guest_memfd set_attribute()) undo if state_change() failed. Just want to keep the attribute consistent with the bitmap on both side. Do we need this? If not, the bitmap can only represent the status of listeners.

Quoted 3 lines

I thought the current implementation properly handles that internally? In which scenario is that not the case?

As mentioned above, e.g. if private_to_shared:
1. set_attribute_shared() convert to shared, but bitmap doesn't change at that stage and is still private.
2. state_change() failed, it undo the operation and bitmap status to keep unchanged.
3. core mm side status is inconsistent with bitmap.

Quoted 1 line
#22 lore raw

Re: [PATCH v3 6/7] memory: Attach MemoryAttributeManager to guest_memfd-backed RAMBlocks

Quoted 30 lines · Chenyi Qiang
Quoted 27 lines · David Hildenbrand
Quoted 1 line · Chenyi Qiang

Hi David & Alexey,

Hi,

Quoted 4 lines

To keep the bitmap aligned, I add the undo operation for set_memory_attributes() and use the bitmap + replay callback to do set_memory_attributes(). Does this change make sense?

I assume you mean this hunk:

+    ret = memory_attribute_manager_state_change(MEMORY_ATTRIBUTE_MANAGER(mr->rdm), +                                                offset, size, to_private); +    if (ret) {
+        warn_report("Failed to notify the listener the state change of " +                    "(0x%"HWADDR_PRIx" + 0x%"HWADDR_PRIx") to %s", +                    start, size, to_private ? "private" : "shared"); +        args.to_private = !to_private;
+        if (to_private) {
+            ret = ram_discard_manager_replay_populated(mr->rdm, &section, +
kvm_set_memory_attributes_cb,
+                                                       &args); +        } else {
+            ret = ram_discard_manager_replay_discarded(mr->rdm, &section,

Quoted 8 lines
Quoted 8 lines

+
kvm_set_memory_attributes_cb,
+                                                       &args); +        }
+        if (ret) {
+            goto out_unref;
+        }

We should probably document that memory_attribute_state_change() cannot fail with "to_private", so you can simplify it to only handle the "to shared" case.

Quoted 11 lines
Quoted 5 lines

Why is that undo necessary? The bitmap + listeners should be held in sync inside of
memory_attribute_manager_state_change(). Handling this in the caller looks wrong.

state_change() handles the listener, i.e. VFIO/IOMMU. And the caller handles the core mm side (guest_memfd set_attribute()) undo if state_change() failed. Just want to keep the attribute consistent with the bitmap on both side. Do we need this? If not, the bitmap can only represent the status of listeners.

Ah, so you meant that you effectively want to undo the attribute change, because the state effectively cannot change, and we want to revert the attribute change.

That makes sense when we are converting private->shared.

BTW, I'm thinking if the orders should be the following (with in-place conversion in mind where we would mmap guest_memfd for the shared memory parts).

On private -> shared conversion:

(1) change_attribute()
(2) state_change(): IOMMU pins shared memory
(3) restore_attribute() if it failed

On shared -> private conversion
(1) state_change(): IOMMU unpins shared memory
(2) change_attribute(): can convert in-place because there are not pins

I'm wondering if the whole attribute change could actually be a listener, invoked by the memory_attribute_manager_state_change() call itself in the right order.

We'd probably need listener priorities, and invoke them in reverse order on shared -> private conversion. Just an idea to get rid of the manual ram_discard_manager_replay_discarded() call in your code.

Signature · 3 lines
Cheers,

David / dhildenbh
#23 lore raw

Re: [PATCH v3 6/7] memory: Attach MemoryAttributeManager to guest_memfd-backed RAMBlocks

Quoted 39 lines · David Hildenbrand
Quoted 38 lines · Chenyi Qiang
Quoted 35 lines · David Hildenbrand
Quoted 1 line · Chenyi Qiang

Hi David & Alexey,

Hi,

Quoted 4 lines

To keep the bitmap aligned, I add the undo operation for set_memory_attributes() and use the bitmap + replay callback to do set_memory_attributes(). Does this change make sense?

I assume you mean this hunk:

+    ret = memory_attribute_manager_state_change(MEMORY_ATTRIBUTE_MANAGER(mr->rdm), +                                                offset, size, to_private);
+    if (ret) {
+        warn_report("Failed to notify the listener the state change of "
+                    "(0x%"HWADDR_PRIx" + 0x%"HWADDR_PRIx") to %s", +                    start, size, to_private ? "private" : "shared"); +        args.to_private = !to_private;
+        if (to_private) {
+            ret = ram_discard_manager_replay_populated(mr->rdm, &section,
+
kvm_set_memory_attributes_cb,
+                                                       &args); +        } else {
+            ret = ram_discard_manager_replay_discarded(mr->rdm, &section,
+
kvm_set_memory_attributes_cb,
+                                                       &args); +        }

Quoted 8 lines
Quoted 4 lines
Quoted 4 lines

+        if (ret) {
+            goto out_unref;
+        }

We should probably document that memory_attribute_state_change() cannot fail with "to_private", so you can simplify it to only handle the "to shared" case.

Yes, "to_private" branch is unnecessary.

Quoted 34 lines
Quoted 7 lines
Quoted 6 lines

Why is that undo necessary? The bitmap + listeners should be held in sync inside of
memory_attribute_manager_state_change(). Handling this in the caller looks wrong.

Quoted 5 lines

state_change() handles the listener, i.e. VFIO/IOMMU. And the caller handles the core mm side (guest_memfd set_attribute()) undo if state_change() failed. Just want to keep the attribute consistent with the bitmap on both side. Do we need this? If not, the bitmap can only represent the status of listeners.

Ah, so you meant that you effectively want to undo the attribute change, because the state effectively cannot change, and we want to revert the attribute change.

That makes sense when we are converting private->shared.

BTW, I'm thinking if the orders should be the following (with in-place conversion in mind where we would mmap guest_memfd for the shared memory parts).

On private -> shared conversion:

(1) change_attribute()
(2) state_change(): IOMMU pins shared memory
(3) restore_attribute() if it failed

On shared -> private conversion
(1) state_change(): IOMMU unpins shared memory
(2) change_attribute(): can convert in-place because there are not pins

Quoted 7 lines

I'm wondering if the whole attribute change could actually be a listener, invoked by the memory_attribute_manager_state_change() call itself in the right order.

We'd probably need listener priorities, and invoke them in reverse order on shared -> private conversion. Just an idea to get rid of the manual ram_discard_manager_replay_discarded() call in your code.

Good idea. I think listener priorities can make it more elegant with in-place conversion. And the change_attribute() listener can be given a highest or lowest priority. Maybe we can add this change in advance before in-place conversion available.

Quoted 1 line
#24 lore raw

Re: [PATCH v3 6/7] memory: Attach MemoryAttributeManager to guest_memfd-backed RAMBlocks

Quoted 91 lines · Chenyi Qiang
Quoted 39 lines · David Hildenbrand
Quoted 38 lines · Chenyi Qiang
Quoted 35 lines · David Hildenbrand
Quoted 1 line · Chenyi Qiang

Hi David & Alexey,

Hi,

Quoted 4 lines

To keep the bitmap aligned, I add the undo operation for set_memory_attributes() and use the bitmap + replay callback to do set_memory_attributes(). Does this change make sense?

I assume you mean this hunk:

+    ret = memory_attribute_manager_state_change(MEMORY_ATTRIBUTE_MANAGER(mr->rdm), +                                                offset, size, to_private);
+    if (ret) {
+        warn_report("Failed to notify the listener the state change of "
+                    "(0x%"HWADDR_PRIx" + 0x%"HWADDR_PRIx") to %s", +                    start, size, to_private ? "private" : "shared"); +        args.to_private = !to_private;
+        if (to_private) {
+            ret = ram_discard_manager_replay_populated(mr->rdm, &section,
+
kvm_set_memory_attributes_cb,
+                                                       &args); +        } else {
+            ret = ram_discard_manager_replay_discarded(mr->rdm, &section,
+
kvm_set_memory_attributes_cb,
+                                                       &args); +        }

Quoted 8 lines
Quoted 4 lines
Quoted 4 lines

+        if (ret) {
+            goto out_unref;
+        }

We should probably document that memory_attribute_state_change() cannot fail with "to_private", so you can simplify it to only handle the "to shared" case.

Yes, "to_private" branch is unnecessary.

Quoted 38 lines
Quoted 7 lines
Quoted 6 lines

Why is that undo necessary? The bitmap + listeners should be held in sync inside of
memory_attribute_manager_state_change(). Handling this in the caller looks wrong.

Quoted 5 lines

state_change() handles the listener, i.e. VFIO/IOMMU. And the caller handles the core mm side (guest_memfd set_attribute()) undo if state_change() failed. Just want to keep the attribute consistent with the bitmap on both side. Do we need this? If not, the bitmap can only represent the status of listeners.

Ah, so you meant that you effectively want to undo the attribute change, because the state effectively cannot change, and we want to revert the attribute change.

That makes sense when we are converting private->shared.

BTW, I'm thinking if the orders should be the following (with in-place conversion in mind where we would mmap guest_memfd for the shared memory parts).

On private -> shared conversion:

(1) change_attribute()
(2) state_change(): IOMMU pins shared memory
(3) restore_attribute() if it failed

On shared -> private conversion
(1) state_change(): IOMMU unpins shared memory
(2) change_attribute(): can convert in-place because there are not pins

I'm wondering if the whole attribute change could actually be a listener, invoked by the memory_attribute_manager_state_change() call itself in the right order.

Quoted 8 lines
Quoted 3 lines

We'd probably need listener priorities, and invoke them in reverse order on shared -> private conversion. Just an idea to get rid of the manual ram_discard_manager_replay_discarded() call in your code.

Good idea. I think listener priorities can make it more elegant with in-place conversion. And the change_attribute() listener can be given a highest or lowest priority. Maybe we can add this change in advance before in-place conversion available.

Hi David,

To add the change_attribute() listener priorities changes, I can think of several steps:

1) change the *NotifyRamDiscard() to return the result, because change attribute to private could return failure.
2) Add a list in confidential_guest_support structure (or some other place) to save the wrapped the listener like VFIORamDiscardListener. And add related listener_register/unregister() in kvm_region_add/del()
3) Add priority in listener and related handling to follow the listener expected sequence.

Regarding the step 1), with change_attribute() listener, the to_private path could also fail. The tricky part is still the error handling. If we simply assume it couldn't fail, maybe add a g_assert() to avoid expected case. Or we also need to add the undo operation for to_private case.

Quoted 3 lines
Quoted 2 lines
#25 lore raw

Re: [PATCH v3 6/7] memory: Attach MemoryAttributeManager to guest_memfd-backed RAMBlocks

In reply to Chenyi Qiang

Hi,

Quoted 9 lines · Chenyi Qiang
system/physmem.c +3 -0
@@ -1885,6 +1886,16 @@ static void ram_block_add(RAMBlock *new_block, Error **errp)
             qemu_mutex_unlock_ramlist();
             goto out_free;
         }
+
+        new_block->memory_attribute_manager = MEMORY_ATTRIBUTE_MANAGER(object_new(TYPE_MEMORY_ATTRIBUTE_MANAGER));
+        if (memory_attribute_manager_realize(new_block->memory_attribute_manager, new_block->mr)) {

system/physmem.c static void ram_block_add(RAMBlock *new_block, Error **errp) · line 1901

Quoted 10 lines
system/physmem.c +7 -0
+            error_setg(errp, "Failed to realize memory attribute manager");
+            object_unref(OBJECT(new_block->memory_attribute_manager));
+            close(new_block->guest_memfd);
+            ram_block_discard_require(false);
+            qemu_mutex_unlock_ramlist();
+            goto out_free;
+        }
     }
 
     ram_size = (new_block->offset + new_block->max_length) >> TARGET_PAGE_BITS;

Might as well put the above into a separate memory manager init function to start with. It keeps the goto out_free error path unified, and makes things more future proof if the rest of ram_block_add() ever develops a need to check for errors.

Regards,

Tony

#26 lore raw

Re: [PATCH v3 6/7] memory: Attach MemoryAttributeManager to guest_memfd-backed RAMBlocks

Quoted 21 lines · Tony Lindgren

Hi,

Quoted 18 lines · Chenyi Qiang
system/physmem.c +10 -0
@@ -1885,6 +1886,16 @@ static void ram_block_add(RAMBlock *new_block, Error **errp)
             qemu_mutex_unlock_ramlist();
             goto out_free;
         }
+
+        new_block->memory_attribute_manager = MEMORY_ATTRIBUTE_MANAGER(object_new(TYPE_MEMORY_ATTRIBUTE_MANAGER));
+        if (memory_attribute_manager_realize(new_block->memory_attribute_manager, new_block->mr)) {
+            error_setg(errp, "Failed to realize memory attribute manager");
+            object_unref(OBJECT(new_block->memory_attribute_manager));
+            close(new_block->guest_memfd);
+            ram_block_discard_require(false);
+            qemu_mutex_unlock_ramlist();
+            goto out_free;
+        }
     }
 
Quoted 6 lines

system/physmem.c static void ram_block_add(RAMBlock *new_block, Error **errp) · line 1901

Quoted 1 line
system/physmem.c +0 -0
     ram_size = (new_block->offset + new_block->max_length) >> TARGET_PAGE_BITS;

Might as well put the above into a separate memory manager init function to start with. It keeps the goto out_free error path unified, and makes things more future proof if the rest of ram_block_add() ever develops a need to check for errors.

Which part to be defined in a separate function? The init function of object_new() + realize(), or the error handling operation (object_unref() + close() + ram_block_discard_require(false))?

If need to check for errors in the rest of ram_block_add() in future, how about adding a new label before out_free and move the error handling there?

Quoted 4 lines

Regards,

Tony

#27 lore raw

Re: [PATCH v3 6/7] memory: Attach MemoryAttributeManager to guest_memfd-backed RAMBlocks

Quoted 31 lines · Chenyi Qiang
Quoted 21 lines · Tony Lindgren

Hi,

Quoted 18 lines · Chenyi Qiang
system/physmem.c +10 -0
@@ -1885,6 +1886,16 @@ static void ram_block_add(RAMBlock *new_block, Error **errp)
             qemu_mutex_unlock_ramlist();
             goto out_free;
         }
+
+        new_block->memory_attribute_manager = MEMORY_ATTRIBUTE_MANAGER(object_new(TYPE_MEMORY_ATTRIBUTE_MANAGER));
+        if (memory_attribute_manager_realize(new_block->memory_attribute_manager, new_block->mr)) {
+            error_setg(errp, "Failed to realize memory attribute manager");
+            object_unref(OBJECT(new_block->memory_attribute_manager));
+            close(new_block->guest_memfd);
+            ram_block_discard_require(false);
+            qemu_mutex_unlock_ramlist();
+            goto out_free;
+        }
     }
 
Quoted 6 lines

system/physmem.c static void ram_block_add(RAMBlock *new_block, Error **errp) · line 1901

Quoted 1 line
system/physmem.c +0 -0
     ram_size = (new_block->offset + new_block->max_length) >> TARGET_PAGE_BITS;

Might as well put the above into a separate memory manager init function to start with. It keeps the goto out_free error path unified, and makes things more future proof if the rest of ram_block_add() ever develops a need to check for errors.

Quoted 3 lines

Which part to be defined in a separate function? The init function of object_new() + realize(), or the error handling operation (object_unref() + close() + ram_block_discard_require(false))?

I was thinking the whole thing, including freeing :) But maybe there's something more to consider to keep calls paired.

Quoted 3 lines

If need to check for errors in the rest of ram_block_add() in future, how about adding a new label before out_free and move the error handling there?

Yeah that would work too.

Regards,

Tony

#28 lore raw

Re: [PATCH v3 6/7] memory: Attach MemoryAttributeManager to guest_memfd-backed RAMBlocks

Quoted 32 lines · Tony Lindgren
Quoted 31 lines · Chenyi Qiang
Quoted 27 lines · Tony Lindgren

Hi,

Quoted 9 lines · Chenyi Qiang
system/physmem.c +3 -0
@@ -1885,6 +1886,16 @@ static void ram_block_add(RAMBlock *new_block, Error **errp)
             qemu_mutex_unlock_ramlist();
             goto out_free;
         }
+
+        new_block->memory_attribute_manager = MEMORY_ATTRIBUTE_MANAGER(object_new(TYPE_MEMORY_ATTRIBUTE_MANAGER));
+        if (memory_attribute_manager_realize(new_block->memory_attribute_manager, new_block->mr)) {

system/physmem.c static void ram_block_add(RAMBlock *new_block, Error **errp) · line 1901

Quoted 10 lines
system/physmem.c +7 -0
+            error_setg(errp, "Failed to realize memory attribute manager");
+            object_unref(OBJECT(new_block->memory_attribute_manager));
+            close(new_block->guest_memfd);
+            ram_block_discard_require(false);
+            qemu_mutex_unlock_ramlist();
+            goto out_free;
+        }
     }
 
     ram_size = (new_block->offset + new_block->max_length) >> TARGET_PAGE_BITS;

Might as well put the above into a separate memory manager init function to start with. It keeps the goto out_free error path unified, and makes things more future proof if the rest of ram_block_add() ever develops a need to check for errors.

Quoted 6 lines
Quoted 3 lines

Which part to be defined in a separate function? The init function of object_new() + realize(), or the error handling operation (object_unref() + close() + ram_block_discard_require(false))?

I was thinking the whole thing, including freeing :) But maybe there's something more to consider to keep calls paired.

If putting the whole thing separately, I think the rest part to do error handling still needs to add the same operation. Or I misunderstand something?

Quoted 6 lines
Quoted 4 lines

If need to check for errors in the rest of ram_block_add() in future, how about adding a new label before out_free and move the error handling there?

Yeah that would work too.

I'm not sure if we should add such change directly, or we can wait for the real error check introduced in future.

Quoted 4 lines

Regards,

Tony

#29 lore raw

Re: [PATCH v3 6/7] memory: Attach MemoryAttributeManager to guest_memfd-backed RAMBlocks

Quoted 39 lines · Chenyi Qiang
Quoted 36 lines · Tony Lindgren
Quoted 34 lines · Chenyi Qiang
Quoted 21 lines · Tony Lindgren

Hi,

Quoted 18 lines · Chenyi Qiang
system/physmem.c +10 -0
@@ -1885,6 +1886,16 @@ static void ram_block_add(RAMBlock *new_block, Error **errp)
             qemu_mutex_unlock_ramlist();
             goto out_free;
         }
+
+        new_block->memory_attribute_manager = MEMORY_ATTRIBUTE_MANAGER(object_new(TYPE_MEMORY_ATTRIBUTE_MANAGER));
+        if (memory_attribute_manager_realize(new_block->memory_attribute_manager, new_block->mr)) {
+            error_setg(errp, "Failed to realize memory attribute manager");
+            object_unref(OBJECT(new_block->memory_attribute_manager));
+            close(new_block->guest_memfd);
+            ram_block_discard_require(false);
+            qemu_mutex_unlock_ramlist();
+            goto out_free;
+        }
     }
 
Quoted 6 lines

system/physmem.c static void ram_block_add(RAMBlock *new_block, Error **errp) · line 1901

Quoted 1 line
system/physmem.c +0 -0
     ram_size = (new_block->offset + new_block->max_length) >> TARGET_PAGE_BITS;

Might as well put the above into a separate memory manager init function to start with. It keeps the goto out_free error path unified, and makes things more future proof if the rest of ram_block_add() ever develops a need to check for errors.

Which part to be defined in a separate function? The init function of object_new() + realize(), or the error handling operation (object_unref() + close() + ram_block_discard_require(false))?

Quoted 6 lines
Quoted 2 lines

I was thinking the whole thing, including freeing :) But maybe there's something more to consider to keep calls paired.

If putting the whole thing separately, I think the rest part to do error handling still needs to add the same operation. Or I misunderstand something?

So maybe you suggestion of just a separate clean-up function would work:

new_block->memory_attribute_manager =

    MEMORY_ATTRIBUTE_MANAGER(object_new(TYPE_MEMORY_ATTRIBUTE_MANAGER));
if (memory_attribute_manager_realize(new_block->memory_attribute_manager,
    new_block->mr)) {
    memory_attribute_manager_cleanup(...);
    goto out_free;
}
Quoted 8 lines
Quoted 5 lines
Quoted 3 lines

If need to check for errors in the rest of ram_block_add() in future, how about adding a new label before out_free and move the error handling there?

Yeah that would work too.

I'm not sure if we should add such change directly, or we can wait for the real error check introduced in future.

Right, not sure either.

Regards,

Tony

#30 lore raw

PATCH v3 7/7 RAMBlock: Make guest_memfd require coordinate discard

In reply to Chenyi Qiang

As guest_memfd is now managed by memory_attribute_manager with RamDiscardManager, only block uncoordinated discard.

---
Changes in v3:
- No change.

Changes in v2:
- Change the ram_block_discard_require(false) to ram_block_coordinated_discard_require(false).

---
 system/physmem.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
system/physmem.c +3 -3
index 0ed394c5d2..a30cdd43ee 100644
@@ -1872,7 +1872,7 @@ static void ram_block_add(RAMBlock *new_block, Error **errp)
         assert(kvm_enabled());
         assert(new_block->guest_memfd < 0);
 
-        ret = ram_block_discard_require(true);
+        ret = ram_block_coordinated_discard_require(true);
         if (ret < 0) {
             error_setg_errno(errp, -ret,
                              "cannot set up private guest memory: discard currently blocked");
@@ -1892,7 +1892,7 @@ static void ram_block_add(RAMBlock *new_block, Error **errp)
             error_setg(errp, "Failed to realize memory attribute manager");
             object_unref(OBJECT(new_block->memory_attribute_manager));
             close(new_block->guest_memfd);
-            ram_block_discard_require(false);
+            ram_block_coordinated_discard_require(false);
             qemu_mutex_unlock_ramlist();
             goto out_free;
         }
@@ -2152,7 +2152,7 @@ static void reclaim_ramblock(RAMBlock *block)
         memory_attribute_manager_unrealize(block->memory_attribute_manager);
         object_unref(OBJECT(block->memory_attribute_manager));
         close(block->guest_memfd);
-        ram_block_discard_require(false);
+        ram_block_coordinated_discard_require(false);
     }
 
     g_free(block);
Signature · 1 line
2.43.5