Skip to contentSkip to conversation

DRM GPUVA Manager & Nouveau VM_BIND UAPI

RFC PATCH Superseded Danilo Krummrich 9 participants 43 messages Jun 20, 2023 – Jun 26, 2023

Subsystems (6)

v5 of 10 Superseded Latest is v10 →

lore
Share & tools
lore
#0 lore raw

v5 cover DRM GPUVA Manager & Nouveau VM_BIND UAPI

This patch series provides a new UAPI for the Nouveau driver in order to support Vulkan features, such as sparse bindings and sparse residency.

Furthermore, with the DRM GPUVA manager it provides a new DRM core feature to keep track of GPU virtual address (VA) mappings in a more generic way.

The DRM GPUVA manager is indented to help drivers implement userspace-manageable GPU VA spaces in reference to the Vulkan API. In order to achieve this goal it serves the following purposes in this context.

1) Provide infrastructure to track GPU VA allocations and mappings, making use of the maple_tree.

2) Generically connect GPU VA mappings to their backing buffers, in particular DRM GEM objects.

3) Provide a common implementation to perform more complex mapping operations on the GPU VA space. In particular splitting and merging of GPU VA mappings, e.g. for intersecting mapping requests or partial unmap requests.

The new VM_BIND Nouveau UAPI build on top of the DRM GPUVA manager, itself providing the following new interfaces.

1) Initialize a GPU VA space via the new DRM_IOCTL_NOUVEAU_VM_INIT ioctl for UMDs to specify the portion of VA space managed by the kernel and userspace, respectively.

2) Allocate and free a VA space region as well as bind and unbind memory to the GPUs VA space via the new DRM_IOCTL_NOUVEAU_VM_BIND ioctl.

3) Execute push buffers with the new DRM_IOCTL_NOUVEAU_EXEC ioctl.

Both, DRM_IOCTL_NOUVEAU_VM_BIND and DRM_IOCTL_NOUVEAU_EXEC, make use of the DRM scheduler to queue jobs and support asynchronous processing with DRM syncobjs as synchronization mechanism.

By default DRM_IOCTL_NOUVEAU_VM_BIND does synchronous processing, DRM_IOCTL_NOUVEAU_EXEC supports asynchronous processing only.

The new VM_BIND UAPI for Nouveau makes also use of drm_exec (execution context for GEM buffers) by Christian König. Since the patch implementing drm_exec was not yet merged into drm-next it is part of this series, as well as a small fix for this patch, which was found while testing this series.

This patch series is also available at [1].

There is a Mesa NVK merge request by Dave Airlie [2] implementing the corresponding userspace parts for this series.

The Vulkan CTS test suite passes the sparse binding and sparse residency test cases for the new UAPI together with Dave's Mesa work.

There are also some test cases in the igt-gpu-tools project [3] for the new UAPI and hence the DRM GPU VA manager. However, most of them are testing the DRM GPU VA manager's logic through Nouveau's new UAPI and should be considered just as helper for implementation.

However, I absolutely intend to change those test cases to proper kunit test cases for the DRM GPUVA manager, once and if we agree on it's usefulness and design.

[1] https://gitlab.freedesktop.org/nouvelles/kernel/-/tree/new-uapi-drm-next /
https://gitlab.freedesktop.org/nouvelles/kernel/-/merge_requests/1
[2] https://gitlab.freedesktop.org/nouveau/mesa/-/merge_requests/150/
[3] https://gitlab.freedesktop.org/dakr/igt-gpu-tools/-/tree/wip_nouveau_vm_bind

Changes in V2:
==============
Nouveau:
- Reworked the Nouveau VM_BIND UAPI to avoid memory allocations in fence signalling critical sections. Updates to the VA space are split up in three separate stages, where only the 2. stage executes in a fence signalling critical section:

1. update the VA space, allocate new structures and page tables
2. (un-)map the requested memory bindings
3. free structures and page tables

- Separated generic job scheduler code from specific job implementations.
- Separated the EXEC and VM_BIND implementation of the UAPI.
- Reworked the locking parts of the nvkm/vmm RAW interface, such that (un-)map operations can be executed in fence signalling critical sections.

GPUVA Manager:
- made drm_gpuva_regions optional for users of the GPUVA manager
- allow NULL GEMs for drm_gpuva entries
- swichted from drm_mm to maple_tree for track drm_gpuva / drm_gpuva_region entries
- provide callbacks for users to allocate custom drm_gpuva_op structures to allow inheritance
- added user bits to drm_gpuva_flags
- added a prefetch operation type in order to support generating prefetch operations in the same way other operations generated
- hand the responsibility for mutual exclusion for a GEM's drm_gpuva list to the user; simplified corresponding (un-)link functions

Maple Tree:
- I added two maple tree patches to the series, one to support custom tree walk macros and one to hand the locking responsibility to the user of the GPUVA manager without pre-defined lockdep checks.

Changes in V3:
==============
Nouveau:
- Reworked the Nouveau VM_BIND UAPI to do the job cleanup (including page table cleanup) within a workqueue rather than the job_free() callback of the scheduler itself. A job_free() callback can stall the execution (run() callback) of the next job in the queue. Since the page table cleanup requires to take the same locks as need to be taken for page table allocation, doing it directly in the job_free() callback would still violate the fence signalling critical path.
- Separated Nouveau fence allocation and emit, such that we do not violate the fence signalling critical path in EXEC jobs.
- Implement "regions" (for handling sparse mappings through PDEs and dual page tables) within Nouveau.
- Drop the requirement for every mapping to be contained within a region.
- Add necassary synchronization of VM_BIND job operation sequences in order to work around limitations in page table handling. This will be addressed in a future re-work of Nouveau's page table handling.
- Fixed a couple of race conditions found through more testing. Thanks to Dave for consitently trying to break it. :-)

GPUVA Manager:
- Implement pre-allocation capabilities for tree modifications within fence signalling critical sections.
- Implement accessors to to apply tree modification while walking the GPUVA tree in order to actually support processing of drm_gpuva_ops through callbacks in fence signalling critical sections rather than through pre-allocated operation lists.
- Remove merging of GPUVAs; the kernel has limited to none knowlege about the semantics of mapping sequences. Hence, merging is purely speculative. It seems that gaining a significant (or at least a measurable) performance increase through merging is way more likely to happen when userspace is responsible for merging mappings up to the next larger page size if possible.
- Since merging was removed, regions pretty much loose their right to exist. They might still be useful for handling dual page tables or similar mechanisms, but since Nouveau seems to be the only driver having a need for this for now, regions were removed from the GPUVA manager.
- Fixed a couple of maple_tree related issues; thanks to Liam for helping me out.

Changes in V4:
==============
Nouveau:
- Refactored how specific VM_BIND and EXEC jobs are created and how their arguments are passed to the generic job implementation.
- Fixed a UAF race condition where bind job ops could have been freed already while still waiting for a job cleanup to finish. This is due to in certain cases we need to wait for mappings actually being unmapped before creating sparse regions in the same area.
- Re-based the code onto drm_exec v4 patch.

GPUVA Manager:
- Fixed a maple tree related bug when pre-allocating MA states. (Boris Brezillion)
- Made struct drm_gpuva_fn_ops a const object in all occurrences. (Boris Brezillion)

Changes in V5:
==============
Nouveau:
- Link and unlink GPUVAs outside the fence signalling critical path in nouveau_uvmm_bind_job_submit() holding the dma-resv lock. Mutual exclusion of BO evicts causing mapping invalidation and regular mapping operations is ensured with dma-fences.

GPUVA Manager:
- Removed the separate GEMs GPUVA list lock. Link and unlink as well as iterating the GEM's GPUVA list should be protected with the GEM's dma-resv lock instead.
- Renamed DRM_GPUVA_EVICTED flag to DRM_GPUVA_INVALIDATED. Mappings do not get eviced, they might get invalidated due to eviction.
- Maple tree uses the 'unsinged long' type for node entries. While this works for GPU VA spaces larger than 32-bit on 64-bit kernel, the GPU VA space is limited to 32-bit on 32-bit kernels as well.
As long as we do not have a 64-bit capable maple tree for 32-bit kernels, the GPU VA manager contains checks to throw warnings when GPU VA entries exceed the maple tree's storage capabilities.
- Extended the Documentation and added example code as requested by Donald Robson.

Christian König (1):
drm: execution context for GEM buffers v4

Danilo Krummrich (13):
maple_tree: split up MA_STATE() macro
drm: manager to keep track of GPUs VA mappings
drm: debugfs: provide infrastructure to dump a DRM GPU VA space drm/nouveau: new VM_BIND uapi interfaces
drm/nouveau: get vmm via nouveau_cli_vmm() drm/nouveau: bo: initialize GEM GPU VA interface drm/nouveau: move usercopy helpers to nouveau_drv.h drm/nouveau: fence: separate fence alloc and emit drm/nouveau: fence: fail to emit when fence context is killed drm/nouveau: chan: provide nouveau_channel_kill() drm/nouveau: nvkm/vmm: implement raw ops to manage uvmm drm/nouveau: implement new VM_BIND uAPI
drm/nouveau: debugfs: implement DRM GPU VA debugfs

 Documentation/gpu/driver-uapi.rst             |   11 +
 Documentation/gpu/drm-mm.rst                  |   54 +
 drivers/gpu/drm/Kconfig                       |    6 +
 drivers/gpu/drm/Makefile                      |    3 +
 drivers/gpu/drm/drm_debugfs.c                 |   41 +
 drivers/gpu/drm/drm_exec.c                    |  278 +++
 drivers/gpu/drm/drm_gem.c                     |    3 +
 drivers/gpu/drm/drm_gpuva_mgr.c               | 1971 ++++++++++++++++
 drivers/gpu/drm/nouveau/Kbuild                |    3 +
 drivers/gpu/drm/nouveau/Kconfig               |    2 +
 drivers/gpu/drm/nouveau/dispnv04/crtc.c       |    9 +-
 drivers/gpu/drm/nouveau/include/nvif/if000c.h |   26 +-
 drivers/gpu/drm/nouveau/include/nvif/vmm.h    |   19 +-
 .../gpu/drm/nouveau/include/nvkm/subdev/mmu.h |   20 +-
 drivers/gpu/drm/nouveau/nouveau_abi16.c       |   24 +
 drivers/gpu/drm/nouveau/nouveau_abi16.h       |    1 +
 drivers/gpu/drm/nouveau/nouveau_bo.c          |  204 +-
 drivers/gpu/drm/nouveau/nouveau_bo.h          |    2 +-
 drivers/gpu/drm/nouveau/nouveau_chan.c        |   22 +-
 drivers/gpu/drm/nouveau/nouveau_chan.h        |    1 +
 drivers/gpu/drm/nouveau/nouveau_debugfs.c     |   39 +
 drivers/gpu/drm/nouveau/nouveau_dmem.c        |    9 +-
 drivers/gpu/drm/nouveau/nouveau_drm.c         |   27 +-
 drivers/gpu/drm/nouveau/nouveau_drv.h         |   94 +-
 drivers/gpu/drm/nouveau/nouveau_exec.c        |  418 ++++
 drivers/gpu/drm/nouveau/nouveau_exec.h        |   54 +
 drivers/gpu/drm/nouveau/nouveau_fence.c       |   23 +-
 drivers/gpu/drm/nouveau/nouveau_fence.h       |    5 +-
 drivers/gpu/drm/nouveau/nouveau_gem.c         |   62 +-
 drivers/gpu/drm/nouveau/nouveau_mem.h         |    5 +
 drivers/gpu/drm/nouveau/nouveau_prime.c       |    2 +-
 drivers/gpu/drm/nouveau/nouveau_sched.c       |  461 ++++
 drivers/gpu/drm/nouveau/nouveau_sched.h       |  123 +
 drivers/gpu/drm/nouveau/nouveau_svm.c         |    2 +-
 drivers/gpu/drm/nouveau/nouveau_uvmm.c        | 1979 +++++++++++++++++
 drivers/gpu/drm/nouveau/nouveau_uvmm.h        |  107 +
 drivers/gpu/drm/nouveau/nouveau_vmm.c         |    4 +-
 drivers/gpu/drm/nouveau/nvif/vmm.c            |  100 +-
 .../gpu/drm/nouveau/nvkm/subdev/mmu/uvmm.c    |  213 +-
 drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c |  197 +-
 drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h |   25 +
 .../drm/nouveau/nvkm/subdev/mmu/vmmgf100.c    |   16 +-
 .../drm/nouveau/nvkm/subdev/mmu/vmmgp100.c    |   16 +-
 .../gpu/drm/nouveau/nvkm/subdev/mmu/vmmnv50.c |   27 +-
 include/drm/drm_debugfs.h                     |   25 +
 include/drm/drm_drv.h                         |    6 +
 include/drm/drm_exec.h                        |  119 +
 include/drm/drm_gem.h                         |   52 +
 include/drm/drm_gpuva_mgr.h                   |  682 ++++++
 include/linux/maple_tree.h                    |    7 +-
 include/uapi/drm/nouveau_drm.h                |  209 ++
 51 files changed, 7566 insertions(+), 242 deletions(-)
 create mode 100644 drivers/gpu/drm/drm_exec.c
 create mode 100644 drivers/gpu/drm/drm_gpuva_mgr.c
 create mode 100644 drivers/gpu/drm/nouveau/nouveau_exec.c
 create mode 100644 drivers/gpu/drm/nouveau/nouveau_exec.h
 create mode 100644 drivers/gpu/drm/nouveau/nouveau_sched.c
 create mode 100644 drivers/gpu/drm/nouveau/nouveau_sched.h
 create mode 100644 drivers/gpu/drm/nouveau/nouveau_uvmm.c
 create mode 100644 drivers/gpu/drm/nouveau/nouveau_uvmm.h
 create mode 100644 include/drm/drm_exec.h
 create mode 100644 include/drm/drm_gpuva_mgr.h

base-commit: 2222dcb0775d36de28992f56455ab3967b30d380

Signature · 1 line
2.40.1
#1 lore raw

PATCH v5 1/14 drm: execution context for GEM buffers v4

From: Christian König <[email protected]>

This adds the infrastructure for an execution context for GEM buffers which is similar to the existing TTMs execbuf util and intended to replace it in the long term.

The basic functionality is that we abstracts the necessary loop to lock many different GEM buffers with automated deadlock and duplicate handling.

v2: drop xarray and use dynamic resized array instead, the locking
overhead is unecessary and measurable.
v3: drop duplicate tracking, radeon is really the only one needing that.
v4: fixes issues pointed out by Danilo, some typos in comments and a
helper for lock arrays of GEM objects.

---
 Documentation/gpu/drm-mm.rst |  12 ++
 drivers/gpu/drm/Kconfig      |   6 +
 drivers/gpu/drm/Makefile     |   2 +
 drivers/gpu/drm/drm_exec.c   | 278 +++++++++++++++++++++++++++++++++++
 include/drm/drm_exec.h       | 119 +++++++++++++++
 5 files changed, 417 insertions(+)
 create mode 100644 drivers/gpu/drm/drm_exec.c
 create mode 100644 include/drm/drm_exec.h
Documentation/gpu/drm-mm.rst +12 -0
index a79fd3549ff8..a52e6f4117d6 100644
@@ -493,6 +493,18 @@ DRM Sync Objects
 .. kernel-doc:: drivers/gpu/drm/drm_syncobj.c
    :export:
 
+DRM Execution context
+=====================
+
+.. kernel-doc:: drivers/gpu/drm/drm_exec.c
+   :doc: Overview
+
+.. kernel-doc:: include/drm/drm_exec.h
+   :internal:
+
+.. kernel-doc:: drivers/gpu/drm/drm_exec.c
+   :export:
+
 GPU Scheduler
 =============
 
drivers/gpu/drm/Kconfig +6 -0
index afb3b2f5f425..c2f3d234c89e 100644
@@ -194,6 +194,12 @@ config DRM_TTM
 	  GPU memory types. Will be enabled automatically if a device driver
 	  uses it.
 
+config DRM_EXEC
+	tristate
+	depends on DRM
+	help
+	  Execution context for command submissions
+
 config DRM_BUDDY
 	tristate
 	depends on DRM
drivers/gpu/drm/Makefile +2 -0
index 7a09a89b493b..414855e2a463 100644
@@ -78,6 +78,8 @@ obj-$(CONFIG_DRM_PANEL_ORIENTATION_QUIRKS) += drm_panel_orientation_quirks.o
 #
 # Memory-management helpers
 #
+#
+obj-$(CONFIG_DRM_EXEC) += drm_exec.o
 
 obj-$(CONFIG_DRM_BUDDY) += drm_buddy.o
 
drivers/gpu/drm/drm_exec.c new file +278 -0

278 lines not on this page · raw message

include/drm/drm_exec.h new file +119 -0

119 lines not on this page · raw message

Signature · 1 line
2.40.1
#2 lore raw

PATCH v5 2/14 maple_tree: split up MA_STATE() macro

In reply to Danilo Krummrich

Split up the MA_STATE() macro such that components using the maple tree can easily inherit from struct ma_state and build custom tree walk macros to hide their internals from users.

Example:

struct sample_iterator {
	struct ma_state mas;
	struct sample_mgr *mgr;
};
\#define SAMPLE_ITERATOR(name, __mgr, start)			\
	struct sample_iterator name = {				\
		.mas = MA_STATE_INIT(&(__mgr)->mt, start, 0),	\
		.mgr = __mgr,					\
	}
\#define sample_iter_for_each_range(it__, entry__, end__) \
	mas_for_each(&(it__).mas, entry__, end__)

--

struct sample *sample;
SAMPLE_ITERATOR(si, min);

sample_iter_for_each_range(&si, sample, max) {
	frob(mgr, sample);
}
---
 include/linux/maple_tree.h | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)
include/linux/maple_tree.h +5 -2
index 1fadb5f5978b..87d55334f1c2 100644
@@ -423,8 +423,8 @@ struct ma_wr_state {
 #define MA_ERROR(err) \
 		((struct maple_enode *)(((unsigned long)err << 2) | 2UL))
 
-#define MA_STATE(name, mt, first, end)					\
-	struct ma_state name = {					\
+#define MA_STATE_INIT(mt, first, end)					\
+	{								\
 		.tree = mt,						\
 		.index = first,						\
 		.last = end,						\
@@ -435,6 +435,9 @@ struct ma_wr_state {
 		.mas_flags = 0,						\
 	}
 
+#define MA_STATE(name, mt, first, end)					\
+	struct ma_state name = MA_STATE_INIT(mt, first, end)
+
 #define MA_WR_STATE(name, ma_state, wr_entry)				\
 	struct ma_wr_state name = {					\
 		.mas = ma_state,					\
Signature · 1 line
2.40.1
#3 lore raw

PATCH v5 3/14 drm: manager to keep track of GPUs VA mappings

In reply to Danilo Krummrich

Add infrastructure to keep track of GPU virtual address (VA) mappings with a decicated VA space manager implementation.

New UAPIs, motivated by Vulkan sparse memory bindings graphics drivers start implementing, allow userspace applications to request multiple and arbitrary GPU VA mappings of buffer objects. The DRM GPU VA manager is intended to serve the following purposes in this context.

1) Provide infrastructure to track GPU VA allocations and mappings, making use of the maple_tree.

2) Generically connect GPU VA mappings to their backing buffers, in particular DRM GEM objects.

3) Provide a common implementation to perform more complex mapping operations on the GPU VA space. In particular splitting and merging of GPU VA mappings, e.g. for intersecting mapping requests or partial unmap requests.

---
 Documentation/gpu/drm-mm.rst    |   42 +
 drivers/gpu/drm/Makefile        |    1 +
 drivers/gpu/drm/drm_gem.c       |    3 +
 drivers/gpu/drm/drm_gpuva_mgr.c | 1971 +++++++++++++++++++++++++++++++
 include/drm/drm_drv.h           |    6 +
 include/drm/drm_gem.h           |   52 +
 include/drm/drm_gpuva_mgr.h     |  682 +++++++++++
 7 files changed, 2757 insertions(+)
 create mode 100644 drivers/gpu/drm/drm_gpuva_mgr.c
 create mode 100644 include/drm/drm_gpuva_mgr.h
Documentation/gpu/drm-mm.rst +42 -0
index a52e6f4117d6..0a9d54e723a8 100644
@@ -466,6 +466,48 @@ DRM MM Range Allocator Function References
 .. kernel-doc:: drivers/gpu/drm/drm_mm.c
    :export:
 
+DRM GPU VA Manager
+==================
+
+Overview
+--------
+
+.. kernel-doc:: drivers/gpu/drm/drm_gpuva_mgr.c
+   :doc: Overview
+
+Split and Merge
+---------------
+
+.. kernel-doc:: drivers/gpu/drm/drm_gpuva_mgr.c
+   :doc: Split and Merge
+
+Locking
+-------
+
+.. kernel-doc:: drivers/gpu/drm/drm_gpuva_mgr.c
+   :doc: Locking
+
+Examples
+--------
+
+.. kernel-doc:: drivers/gpu/drm/drm_gpuva_mgr.c
+   :doc: Examples
+
+Quirks
+------
+
+.. kernel-doc:: drivers/gpu/drm/drm_gpuva_mgr.c
+   :doc: Quirks
+
+DRM GPU VA Manager Function References
+--------------------------------------
+
+.. kernel-doc:: include/drm/drm_gpuva_mgr.h
+   :internal:
+
+.. kernel-doc:: drivers/gpu/drm/drm_gpuva_mgr.c
+   :export:
+
 DRM Buddy Allocator
 ===================
 
drivers/gpu/drm/Makefile +1 -0
index 414855e2a463..6d6c9dec66e8 100644
@@ -45,6 +45,7 @@ drm-y := \
 	drm_vblank.o \
 	drm_vblank_work.o \
 	drm_vma_manager.o \
+	drm_gpuva_mgr.o \
 	drm_writeback.o
 drm-$(CONFIG_DRM_LEGACY) += \
 	drm_agpsupport.o \
drivers/gpu/drm/drm_gem.c +3 -0
index 1a5a2cd0d4ec..cd878ebddbd0 100644
@@ -164,6 +164,9 @@ void drm_gem_private_object_init(struct drm_device *dev,
 	if (!obj->resv)
 		obj->resv = &obj->_resv;
 
+	if (drm_core_check_feature(dev, DRIVER_GEM_GPUVA))
+		drm_gem_gpuva_init(obj);
+
 	drm_vma_node_reset(&obj->vma_node);
 	INIT_LIST_HEAD(&obj->lru_node);
 }
drivers/gpu/drm/drm_gpuva_mgr.c new file +1971 -0

1,971 lines not on this page · raw message

include/drm/drm_drv.h +6 -0
index 89e2706cac56..04dbe223b1a5 100644
@@ -104,6 +104,12 @@ enum drm_driver_feature {
 	 * acceleration should be handled by two drivers that are connected using auxiliary bus.
 	 */
 	DRIVER_COMPUTE_ACCEL            = BIT(7),
+	/**
+	 * @DRIVER_GEM_GPUVA:
+	 *
+	 * Driver supports user defined GPU VA bindings for GEM objects.
+	 */
+	DRIVER_GEM_GPUVA		= BIT(8),
 
 	/* IMPORTANT: Below are all the legacy flags, add new ones above. */
 
include/drm/drm_gem.h +52 -0

68 lines not on this page · raw message

include/drm/drm_gpuva_mgr.h new file +682 -0

682 lines not on this page · raw message

Signature · 1 line
2.40.1
#4 lore raw

Re: [PATCH drm-next v5 03/14] drm: manager to keep track of GPUs VA mappings

kernel test robot noticed the following build warnings:

Hi Danilo,

kernel test robot noticed the following build warnings:

[auto build test WARNING on 2222dcb0775d36de28992f56455ab3967b30d380]

url:    https://github.com/intel-lab-lkp/linux/commits/Danilo-Krummrich/drm-execution-context-for-GEM-buffers-v4/20230620-084448
base:   2222dcb0775d36de28992f56455ab3967b30d380
patch link:    https://lore.kernel.org/r/20230620004217.4700-4-dakr%40redhat.com
patch subject: [PATCH drm-next v5 03/14] drm: manager to keep track of GPUs VA mappings
config: m68k-allyesconfig (https://download.01.org/0day-ci/archive/20230620/[email protected]/config)
compiler: m68k-linux-gcc (GCC) 12.3.0
reproduce: (https://download.01.org/0day-ci/archive/20230620/[email protected]/reproduce)

If you fix the issue in a separate patch/commit (i.e. not just a new version of the same patch/commit), kindly add following tags
| Reported-by: kernel test robot <[email protected]>
| Closes: https://lore.kernel.org/oe-kbuild-all/[email protected]/

All warnings (new ones prefixed by >>):

In file included from arch/m68k/include/asm/bug.h:32,

                    from include/linux/bug.h:5,
                    from include/linux/thread_info.h:13,
                    from include/asm-generic/preempt.h:5,
                    from ./arch/m68k/include/generated/asm/preempt.h:1,
                    from include/linux/preempt.h:78,
                    from include/linux/spinlock.h:56,
                    from include/linux/kref.h:16,
                    from include/drm/drm_gem.h:37,
                    from drivers/gpu/drm/drm_gpuva_mgr.c:28:
   drivers/gpu/drm/drm_gpuva_mgr.c: In function 'drm_gpuva_check_overflow':
>> drivers/gpu/drm/drm_gpuva_mgr.c:675:21: warning: format '%lu' expects argument of type 'long unsigned int', but argument 5 has type 'unsigned int' [-Wformat=]
     675 |                     "GPUVA address limited to %lu bytes, see Documentation.\n",
         |                     ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
   include/asm-generic/bug.h:97:62: note: in definition of macro '__WARN_printf'
      97 |                 warn_slowpath_fmt(__FILE__, __LINE__, taint, arg);      \
         |                                                              ^~~
   drivers/gpu/drm/drm_gpuva_mgr.c:674:16: note: in expansion of macro 'WARN'
     674 |         return WARN(check_add_overflow(addr, range, &end),
         |                ^~~~
   drivers/gpu/drm/drm_gpuva_mgr.c:675:49: note: format string is defined here
     675 |                     "GPUVA address limited to %lu bytes, see Documentation.\n",
         |                                               ~~^
         |                                                 |
         |                                                 long unsigned int
         |                                               %u
   drivers/gpu/drm/drm_gpuva_mgr.c: In function '__drm_gpuva_sm_map':
   drivers/gpu/drm/drm_gpuva_mgr.c:1314:32: warning: variable 'prev' set but not used [-Wunused-but-set-variable]
    1314 |         struct drm_gpuva *va, *prev = NULL;
         |                                ^~~~

vim +675 drivers/gpu/drm/drm_gpuva_mgr.c

668
669 static inline bool
670 drm_gpuva_check_overflow(u64 addr, u64 range)
671 {
672 MTREE_INDEX_TYPE end;
673
674 return WARN(check_add_overflow(addr, range, &end),

 > 675			    "GPUVA address limited to %lu bytes, see Documentation.\n",

676 MTREE_INDEX_SIZE);
677 }
678

Signature · 2 lines
0-DAY CI Kernel Test Service
https://github.com/intel/lkp-tests/wiki
#5 lore raw

Re: [PATCH drm-next v5 03/14] drm: manager to keep track of GPUs VA mappings

In reply to Danilo Krummrich
kernel test robot noticed the following build warnings:

Hi Danilo,

kernel test robot noticed the following build warnings:

[auto build test WARNING on 2222dcb0775d36de28992f56455ab3967b30d380]

url:    https://github.com/intel-lab-lkp/linux/commits/Danilo-Krummrich/drm-execution-context-for-GEM-buffers-v4/20230620-084448
base:   2222dcb0775d36de28992f56455ab3967b30d380
patch link:    https://lore.kernel.org/r/20230620004217.4700-4-dakr%40redhat.com
patch subject: [PATCH drm-next v5 03/14] drm: manager to keep track of GPUs VA mappings
config: hexagon-randconfig-r041-20230620 (https://download.01.org/0day-ci/archive/20230620/[email protected]/config)
compiler: clang version 17.0.0 (https://github.com/llvm/llvm-project.git 4a5ac14ee968ff0ad5d2cc1ffa0299048db4c88a)
reproduce: (https://download.01.org/0day-ci/archive/20230620/[email protected]/reproduce)

If you fix the issue in a separate patch/commit (i.e. not just a new version of the same patch/commit), kindly add following tags
| Reported-by: kernel test robot <[email protected]>
| Closes: https://lore.kernel.org/oe-kbuild-all/[email protected]/

All warnings (new ones prefixed by >>):

>> drivers/gpu/drm/drm_gpuva_mgr.c:676:7: warning: format specifies type 'unsigned long' but the argument has type 'unsigned int' [-Wformat]
     676 |         return WARN(check_add_overflow(addr, range, &end),
         |                ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
     677 |                     "GPUVA address limited to %lu bytes, see Documentation.\n",
         |                     ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
         |                                               %u
     678 |                     MTREE_INDEX_SIZE);
         |                     ^~~~~~~~~~~~~~~~~
   drivers/gpu/drm/drm_gpuva_mgr.c:663:26: note: expanded from macro 'MTREE_INDEX_SIZE'
     663 | #define MTREE_INDEX_SIZE sizeof(MTREE_INDEX_TYPE)
         |                          ^
   include/asm-generic/bug.h:133:29: note: expanded from macro 'WARN'
     133 |                 __WARN_printf(TAINT_WARN, format);                      \
         |                 ~~~~~~~~~~~~~~~~~~~~~~~~~~^~~~~~~
   include/asm-generic/bug.h:97:48: note: expanded from macro '__WARN_printf'
      97 |                 warn_slowpath_fmt(__FILE__, __LINE__, taint, arg);      \
         |                                                              ^~~
   drivers/gpu/drm/drm_gpuva_mgr.c:1314:25: warning: variable 'prev' set but not used [-Wunused-but-set-variable]
    1314 |         struct drm_gpuva *va, *prev = NULL;
         |                                ^

2 warnings generated.

vim +676 drivers/gpu/drm/drm_gpuva_mgr.c

668
669 static inline bool
670 drm_gpuva_check_overflow(u64 addr, u64 range)
671 {
672 MTREE_INDEX_TYPE end;
673
674 return WARN(check_add_overflow(addr, range, &end),
675 "GPUVA address limited to %lu bytes, see Documentation.\n",

 > 676			    MTREE_INDEX_SIZE);

677 }
678

Signature · 2 lines
0-DAY CI Kernel Test Service
https://github.com/intel/lkp-tests/wiki
#7 lore raw

Re: [PATCH drm-next v5 03/14] drm: manager to keep track of GPUs VA mappings

In reply to Danilo Krummrich

Hi Danilo,

sorry for the delayed reply. I've trying to dig myself out of a hole at the moment.

Am 20.06.23 um 02:42 schrieb Danilo Krummrich:

Quoted 23 lines

[SNIP]

include/drm/drm_gem.h +7 -0
index bbc721870c13..5ec8148a30ee 100644
@@ -36,6 +36,8 @@
 
  #include <linux/kref.h>
  #include <linux/dma-resv.h>
+#include <linux/list.h>
+#include <linux/mutex.h>
 
  #include <drm/drm_vma_manager.h>
 
@@ -379,6 +381,18 @@ struct drm_gem_object {
  	 */
  	struct dma_resv _resv;
 
+	/**
+	 * @gpuva:
+	 *
+	 * Provides the list of GPU VAs attached to this GEM object.
+	 *

include/drm/drm_gem.h struct drm_gem_object { · line 398

Quoted 10 lines
include/drm/drm_gem.h +7 -0
+	 * Drivers should lock list accesses with the GEMs &dma_resv lock
+	 * (&drm_gem_object.resv).
+	 */
+	struct {
+		struct list_head list;
+	} gpuva;
+
  	/**
  	 * @funcs:
  	 *

I'm pretty sure that it's not a good idea to attach this directly to the GEM object.

As you wrote in the commit message it's highly driver specific what to map and where to map it.

Instead I suggest to have a separate structure for mappings in a VA space which driver can then add to their GEM objects or whatever they want to map into their VMs.

Regards,
Christian.

#8 lore raw

Re: [PATCH drm-next v5 03/14] drm: manager to keep track of GPUs VA mappings

Hi Christian,

Quoted 3 lines · Christian König

Hi Danilo,

sorry for the delayed reply. I've trying to dig myself out of a hole at the moment.

No worries, thank you for taking a look anyway!

Quoted 25 lines

Am 20.06.23 um 02:42 schrieb Danilo Krummrich:

Quoted 23 lines

[SNIP]

include/drm/drm_gem.h +11 -0
index bbc721870c13..5ec8148a30ee 100644
@@ -36,6 +36,8 @@
   #include <linux/kref.h>
   #include <linux/dma-resv.h>
+#include <linux/list.h>
+#include <linux/mutex.h>
   #include <drm/drm_vma_manager.h>
@@ -379,6 +381,18 @@ struct drm_gem_object {
        */
       struct dma_resv _resv;
+    /**
+     * @gpuva:
+     *
+     * Provides the list of GPU VAs attached to this GEM object.
+     *
+     * Drivers should lock list accesses with the GEMs &dma_resv lock
+     * (&drm_gem_object.resv).
+     */
+    struct {
Quoted 8 lines

include/drm/drm_gem.h struct drm_gem_object { · line 398

Quoted 6 lines
include/drm/drm_gem.h +3 -0
+        struct list_head list;
+    } gpuva;
+
       /**
        * @funcs:
        *

I'm pretty sure that it's not a good idea to attach this directly to the GEM object.

Why do you think so? IMHO having a common way to connect mappings to their backing buffers is a good thing, since every driver needs this connection anyway.

E.g. when a BO gets evicted, drivers can just iterate the list of mappings and, as the circumstances require, invalidate the corresponding mappings or to unmap all existing mappings of a given buffer.

What would be the advantage to let every driver implement a driver specific way of keeping this connection? Do you see cases where this kind of connection between mappings and backing buffers wouldn't be good enough? If so, which cases do you have in mind? Maybe we can cover them in a common way as well?

Quoted 2 lines

As you wrote in the commit message it's highly driver specific what to map and where to map it.

In the end the common case should be that in a VA space at least every mapping being backed by a BO is represented by a struct drm_gpuva.

Quoted 2 lines

Instead I suggest to have a separate structure for mappings in a VA space which driver can then add to their GEM objects or whatever they want to map into their VMs.

Which kind of separate structure for mappings? Another one analogous to struct drm_gpuva?

- Danilo

Quoted 5 lines

Regards,
Christian.

#9 lore raw

Re: [PATCH drm-next v5 03/14] drm: manager to keep track of GPUs VA mappings

Am 20.06.23 um 14:23 schrieb Danilo Krummrich:

Quoted 43 lines

Hi Christian,

Quoted 3 lines · Christian König

Hi Danilo,

sorry for the delayed reply. I've trying to dig myself out of a hole at the moment.

No worries, thank you for taking a look anyway!

Quoted 25 lines

Am 20.06.23 um 02:42 schrieb Danilo Krummrich:

Quoted 23 lines

[SNIP]

include/drm/drm_gem.h +11 -0
index bbc721870c13..5ec8148a30ee 100644
@@ -36,6 +36,8 @@
   #include <linux/kref.h>
   #include <linux/dma-resv.h>
+#include <linux/list.h>
+#include <linux/mutex.h>
   #include <drm/drm_vma_manager.h>
@@ -379,6 +381,18 @@ struct drm_gem_object {
        */
       struct dma_resv _resv;
+    /**
+     * @gpuva:
+     *
+     * Provides the list of GPU VAs attached to this GEM object.
+     *
+     * Drivers should lock list accesses with the GEMs &dma_resv lock
+     * (&drm_gem_object.resv).
+     */
+    struct {
Quoted 8 lines

include/drm/drm_gem.h struct drm_gem_object { · line 398

Quoted 6 lines
include/drm/drm_gem.h +3 -0
+        struct list_head list;
+    } gpuva;
+
       /**
        * @funcs:
        *

I'm pretty sure that it's not a good idea to attach this directly to the GEM object.

Quoted 5 lines

Why do you think so? IMHO having a common way to connect mappings to their backing buffers is a good thing, since every driver needs this connection anyway.

E.g. when a BO gets evicted, drivers can just iterate the list of mappings and, as the circumstances require, invalidate the corresponding mappings or to unmap all existing mappings of a given buffer.

What would be the advantage to let every driver implement a driver specific way of keeping this connection?

Flexibility. For example on amdgpu the mappings of a BO are groups by VM address spaces.

E.g. the BO points to multiple bo_vm structures which in turn have lists of their mappings.

Additional to that there is a state maschine associated with the mappings, e.g. if the page tables are up to date or need to be updated etc....

Quoted 1 line

Do you see cases where this kind of connection between mappings and backing buffers wouldn't be good enough? If so, which cases do you have in mind? Maybe we can cover them in a common way as well?

Yeah, we have tons of cases like that. But I have no idea how to generalize them.

Quoted 5 lines
Quoted 3 lines

As you wrote in the commit message it's highly driver specific what to map and where to map it.

In the end the common case should be that in a VA space at least every mapping being backed by a BO is represented by a struct drm_gpuva.

Oh, no! We also have mappings not backed by a BO at all! For example for partial resident textures or data routing to internal hw etc...

You can't be sure that a mapping is backed by a BO at all.

Quoted 5 lines
Quoted 3 lines

Instead I suggest to have a separate structure for mappings in a VA space which driver can then add to their GEM objects or whatever they want to map into their VMs.

Which kind of separate structure for mappings? Another one analogous to struct drm_gpuva?

Well similar to what amdgpu uses BO -> one structure for each combination of BO and VM -> mappings inside that VM

Regards,
Christian.

Quoted 9 lines

- Danilo

Quoted 5 lines

Regards,
Christian.

#10 lore raw

Re: [PATCH drm-next v5 03/14] drm: manager to keep track of GPUs VA mappings

Quoted 46 lines · Christian König

Am 20.06.23 um 14:23 schrieb Danilo Krummrich:

Quoted 45 lines

Hi Christian,

Quoted 3 lines · Christian König

Hi Danilo,

sorry for the delayed reply. I've trying to dig myself out of a hole at the moment.

No worries, thank you for taking a look anyway!

Quoted 25 lines

Am 20.06.23 um 02:42 schrieb Danilo Krummrich:

Quoted 23 lines

[SNIP]

include/drm/drm_gem.h +11 -0
index bbc721870c13..5ec8148a30ee 100644
@@ -36,6 +36,8 @@
   #include <linux/kref.h>
   #include <linux/dma-resv.h>
+#include <linux/list.h>
+#include <linux/mutex.h>
   #include <drm/drm_vma_manager.h>
@@ -379,6 +381,18 @@ struct drm_gem_object {
        */
       struct dma_resv _resv;
+    /**
+     * @gpuva:
+     *
+     * Provides the list of GPU VAs attached to this GEM object.
+     *
+     * Drivers should lock list accesses with the GEMs &dma_resv lock
+     * (&drm_gem_object.resv).
+     */
+    struct {
Quoted 8 lines

include/drm/drm_gem.h struct drm_gem_object { · line 398

Quoted 6 lines
include/drm/drm_gem.h +3 -0
+        struct list_head list;
+    } gpuva;
+
       /**
        * @funcs:
        *

I'm pretty sure that it's not a good idea to attach this directly to the GEM object.

Why do you think so? IMHO having a common way to connect mappings to their backing buffers is a good thing, since every driver needs this connection anyway.

Quoted 7 lines
Quoted 3 lines

E.g. when a BO gets evicted, drivers can just iterate the list of mappings and, as the circumstances require, invalidate the corresponding mappings or to unmap all existing mappings of a given buffer.

What would be the advantage to let every driver implement a driver specific way of keeping this connection?

Flexibility. For example on amdgpu the mappings of a BO are groups by VM address spaces.

E.g. the BO points to multiple bo_vm structures which in turn have lists of their mappings.

Isn't this (almost) the same relationship I introduce with the GPUVA manager?

If you would switch over to the GPUVA manager right now, it would be that every GEM has a list of it's mappings (the gpuva list). The mapping is represented by struct drm_gpuva (of course embedded in driver specific structure(s)) which has a pointer to the VM address space it is part of, namely the GPUVA manager instance. And the GPUVA manager keeps a maple tree of it's mappings as well.

If you still would like to *directly* (indirectly you already have that relationship) keep a list of GPUVA managers (VM address spaces) per GEM, you could still do that in a driver specific way.

Do I miss something?

Quoted 6 lines

Additional to that there is a state maschine associated with the mappings, e.g. if the page tables are up to date or need to be updated etc....

Quoted 1 line

Do you see cases where this kind of connection between mappings and backing buffers wouldn't be good enough? If so, which cases do you have in mind? Maybe we can cover them in a common way as well?

Yeah, we have tons of cases like that. But I have no idea how to generalize them.

They could still remain to be driver specific then, right?

Quoted 10 lines
Quoted 6 lines
Quoted 4 lines

As you wrote in the commit message it's highly driver specific what to map and where to map it.

In the end the common case should be that in a VA space at least every mapping being backed by a BO is represented by a struct drm_gpuva.

Oh, no! We also have mappings not backed by a BO at all! For example for partial resident textures or data routing to internal hw etc...

You can't be sure that a mapping is backed by a BO at all.

I fully agree, that's why I wrote "the common case should be that in a VA space at least every mapping *being backed by a BO* is represented by a struct drm_gpuva".

Mappings not being backed by an actual BO would not be linked to a GEM of course.

Quoted 8 lines
Quoted 6 lines
Quoted 4 lines

Instead I suggest to have a separate structure for mappings in a VA space which driver can then add to their GEM objects or whatever they want to map into their VMs.

Which kind of separate structure for mappings? Another one analogous to struct drm_gpuva?

Well similar to what amdgpu uses BO -> one structure for each combination of BO and VM -> mappings inside that VM

As explained above, I think that's exactly what the GPUVA manager does, just in another order:

BO has list of mappings, mappings have pointer to VM, VM has list (or actually a maple tree) of mappings.

You see any advantages or disadvantages of either order of relationships? For me it looks like it doesn't really matter which one to pick.

- Danilo

Quoted 14 lines

Regards,
Christian.

Quoted 9 lines

- Danilo

Quoted 5 lines

Regards,
Christian.

#11 lore raw

Re: [PATCH drm-next v5 03/14] drm: manager to keep track of GPUs VA mappings

Am 22.06.23 um 16:22 schrieb Danilo Krummrich:

Quoted 56 lines · Christian König
Quoted 47 lines

Am 20.06.23 um 14:23 schrieb Danilo Krummrich:

Quoted 46 lines

Hi Christian,

Quoted 3 lines · Christian König

Hi Danilo,

sorry for the delayed reply. I've trying to dig myself out of a hole at the moment.

No worries, thank you for taking a look anyway!

Quoted 26 lines

Am 20.06.23 um 02:42 schrieb Danilo Krummrich:

Quoted 24 lines

[SNIP]

include/drm/drm_gem.h +8 -0
index bbc721870c13..5ec8148a30ee 100644
@@ -36,6 +36,8 @@
   #include <linux/kref.h>
   #include <linux/dma-resv.h>
+#include <linux/list.h>
+#include <linux/mutex.h>
   #include <drm/drm_vma_manager.h>
@@ -379,6 +381,18 @@ struct drm_gem_object {
        */
       struct dma_resv _resv;
+    /**
+     * @gpuva:
+     *
+     * Provides the list of GPU VAs attached to this GEM object.
+     *
+     * Drivers should lock list accesses with the GEMs &dma_resv 

lock
+     * (&drm_gem_object.resv).
+     */
+    struct {

Quoted 8 lines
Quoted 6 lines

+        struct list_head list;
+    } gpuva;
+
      /**
       * @funcs:
       *

I'm pretty sure that it's not a good idea to attach this directly to the GEM object.

Why do you think so? IMHO having a common way to connect mappings to their backing buffers is a good thing, since every driver needs this connection anyway.

Quoted 7 lines
Quoted 3 lines

E.g. when a BO gets evicted, drivers can just iterate the list of mappings and, as the circumstances require, invalidate the corresponding mappings or to unmap all existing mappings of a given buffer.

What would be the advantage to let every driver implement a driver specific way of keeping this connection?

Flexibility. For example on amdgpu the mappings of a BO are groups by VM address spaces.

E.g. the BO points to multiple bo_vm structures which in turn have lists of their mappings.

Quoted 7 lines

Isn't this (almost) the same relationship I introduce with the GPUVA manager?

If you would switch over to the GPUVA manager right now, it would be that every GEM has a list of it's mappings (the gpuva list). The mapping is represented by struct drm_gpuva (of course embedded in driver specific structure(s)) which has a pointer to the VM address space it is part of, namely the GPUVA manager instance. And the GPUVA manager keeps a maple tree of it's mappings as well.

If you still would like to *directly* (indirectly you already have that relationship) keep a list of GPUVA managers (VM address spaces) per GEM, you could still do that in a driver specific way.

Do I miss something?

How do you efficiently find only the mappings of a BO in one VM?

Keep in mind that we have cases where one BO is shared with hundreds of different VMs as well as potentially the number of mappings can be >10k.

Quoted 9 lines
Quoted 7 lines

Additional to that there is a state maschine associated with the mappings, e.g. if the page tables are up to date or need to be updated etc....

Quoted 1 line

Do you see cases where this kind of connection between mappings and backing buffers wouldn't be good enough? If so, which cases do you have in mind? Maybe we can cover them in a common way as well?

Yeah, we have tons of cases like that. But I have no idea how to generalize them.

They could still remain to be driver specific then, right?

Well does the mapping has a back pointer to the BO? And can that be optional NULL if there is no BO?

Regards,
Christian.

Quoted 48 lines
Quoted 11 lines
Quoted 7 lines
Quoted 5 lines

As you wrote in the commit message it's highly driver specific what to map and where to map it.

In the end the common case should be that in a VA space at least every mapping being backed by a BO is represented by a struct drm_gpuva.

Oh, no! We also have mappings not backed by a BO at all! For example for partial resident textures or data routing to internal hw etc...

You can't be sure that a mapping is backed by a BO at all.

I fully agree, that's why I wrote "the common case should be that in a VA space at least every mapping *being backed by a BO* is represented by a struct drm_gpuva".

Mappings not being backed by an actual BO would not be linked to a GEM of course.

Quoted 8 lines
Quoted 6 lines
Quoted 4 lines

Instead I suggest to have a separate structure for mappings in a VA space which driver can then add to their GEM objects or whatever they want to map into their VMs.

Which kind of separate structure for mappings? Another one analogous to struct drm_gpuva?

Well similar to what amdgpu uses BO -> one structure for each combination of BO and VM -> mappings inside that VM

As explained above, I think that's exactly what the GPUVA manager does, just in another order:

BO has list of mappings, mappings have pointer to VM, VM has list (or actually a maple tree) of mappings.

You see any advantages or disadvantages of either order of relationships? For me it looks like it doesn't really matter which one to pick.

- Danilo

Quoted 14 lines

Regards,
Christian.

Quoted 9 lines

- Danilo

Quoted 5 lines

Regards,
Christian.

#12 lore raw

Re: [PATCH drm-next v5 03/14] drm: manager to keep track of GPUs VA mappings

Quoted 59 lines · Christian König

Am 22.06.23 um 16:22 schrieb Danilo Krummrich:

Quoted 58 lines · Christian König
Quoted 47 lines

Am 20.06.23 um 14:23 schrieb Danilo Krummrich:

Quoted 46 lines

Hi Christian,

Quoted 3 lines · Christian König

Hi Danilo,

sorry for the delayed reply. I've trying to dig myself out of a hole at the moment.

No worries, thank you for taking a look anyway!

Quoted 26 lines

Am 20.06.23 um 02:42 schrieb Danilo Krummrich:

Quoted 24 lines

[SNIP]

include/drm/drm_gem.h +8 -0
index bbc721870c13..5ec8148a30ee 100644
@@ -36,6 +36,8 @@
   #include <linux/kref.h>
   #include <linux/dma-resv.h>
+#include <linux/list.h>
+#include <linux/mutex.h>
   #include <drm/drm_vma_manager.h>
@@ -379,6 +381,18 @@ struct drm_gem_object {
        */
       struct dma_resv _resv;
+    /**
+     * @gpuva:
+     *
+     * Provides the list of GPU VAs attached to this GEM object.
+     *
+     * Drivers should lock list accesses with the GEMs &dma_resv 

lock
+     * (&drm_gem_object.resv).
+     */
+    struct {

Quoted 8 lines
Quoted 6 lines

+        struct list_head list;
+    } gpuva;
+
      /**
       * @funcs:
       *

I'm pretty sure that it's not a good idea to attach this directly to the GEM object.

Why do you think so? IMHO having a common way to connect mappings to their backing buffers is a good thing, since every driver needs this connection anyway.

Quoted 7 lines
Quoted 3 lines

E.g. when a BO gets evicted, drivers can just iterate the list of mappings and, as the circumstances require, invalidate the corresponding mappings or to unmap all existing mappings of a given buffer.

What would be the advantage to let every driver implement a driver specific way of keeping this connection?

Flexibility. For example on amdgpu the mappings of a BO are groups by VM address spaces.

E.g. the BO points to multiple bo_vm structures which in turn have lists of their mappings.

Isn't this (almost) the same relationship I introduce with the GPUVA manager?

Quoted 7 lines
Quoted 5 lines

If you would switch over to the GPUVA manager right now, it would be that every GEM has a list of it's mappings (the gpuva list). The mapping is represented by struct drm_gpuva (of course embedded in driver specific structure(s)) which has a pointer to the VM address space it is part of, namely the GPUVA manager instance. And the GPUVA manager keeps a maple tree of it's mappings as well.

If you still would like to *directly* (indirectly you already have that relationship) keep a list of GPUVA managers (VM address spaces) per GEM, you could still do that in a driver specific way.

Do I miss something?

How do you efficiently find only the mappings of a BO in one VM?

Actually, I think this case should even be more efficient than with a BO having a list of GPUVAs (or mappings):

Having a list of GPUVAs per GEM, each GPUVA has a pointer to it's VM. Hence, you'd only need to iterate the list of mappings for a given BO and check the mappings VM pointer.

Having a list of VMs per BO, you'd have to iterate the whole VM to find the mappings having a pointer to the given BO, right?

I'd think that a single VM potentially has more mapping entries than a single BO was mapped in multiple VMs.

Another case to consider is the case I originally had in mind choosing this relationship: finding all mappings for a given BO, which I guess all drivers need to do in order to invalidate mappings on BO eviction.

Having a list of VMs per BO, wouldn't you need to iterate all of the VMs entirely?

Quoted 7 lines

Keep in mind that we have cases where one BO is shared with hundreds of different VMs as well as potentially the number of mappings can be >10k.

Quoted 4 lines
Quoted 4 lines

Additional to that there is a state maschine associated with the mappings, e.g. if the page tables are up to date or need to be updated etc....

Quoted 7 lines
Quoted 5 lines
Quoted 3 lines
Quoted 1 line

Do you see cases where this kind of connection between mappings and backing buffers wouldn't be good enough? If so, which cases do you have in mind? Maybe we can cover them in a common way as well?

Yeah, we have tons of cases like that. But I have no idea how to generalize them.

They could still remain to be driver specific then, right?

Well does the mapping has a back pointer to the BO? And can that be optional NULL if there is no BO?

Yes to both.

- Danilo

Quoted 53 lines

Regards,
Christian.

Quoted 48 lines
Quoted 11 lines
Quoted 7 lines
Quoted 5 lines

As you wrote in the commit message it's highly driver specific what to map and where to map it.

In the end the common case should be that in a VA space at least every mapping being backed by a BO is represented by a struct drm_gpuva.

Oh, no! We also have mappings not backed by a BO at all! For example for partial resident textures or data routing to internal hw etc...

You can't be sure that a mapping is backed by a BO at all.

I fully agree, that's why I wrote "the common case should be that in a VA space at least every mapping *being backed by a BO* is represented by a struct drm_gpuva".

Mappings not being backed by an actual BO would not be linked to a GEM of course.

Quoted 8 lines
Quoted 6 lines
Quoted 4 lines

Instead I suggest to have a separate structure for mappings in a VA space which driver can then add to their GEM objects or whatever they want to map into their VMs.

Which kind of separate structure for mappings? Another one analogous to struct drm_gpuva?

Well similar to what amdgpu uses BO -> one structure for each combination of BO and VM -> mappings inside that VM

As explained above, I think that's exactly what the GPUVA manager does, just in another order:

BO has list of mappings, mappings have pointer to VM, VM has list (or actually a maple tree) of mappings.

You see any advantages or disadvantages of either order of relationships? For me it looks like it doesn't really matter which one to pick.

- Danilo

Quoted 14 lines

Regards,
Christian.

Quoted 9 lines

- Danilo

Quoted 5 lines

Regards,
Christian.

#13 lore raw

Re: [PATCH drm-next v5 03/14] drm: manager to keep track of GPUs VA mappings

Quoted 62 lines · Danilo Krummrich
Quoted 61 lines · Christian König

Am 22.06.23 um 16:22 schrieb Danilo Krummrich:

Quoted 60 lines · Christian König
Quoted 47 lines

Am 20.06.23 um 14:23 schrieb Danilo Krummrich:

Quoted 46 lines

Hi Christian,

Quoted 3 lines · Christian König

Hi Danilo,

sorry for the delayed reply. I've trying to dig myself out of a hole at the moment.

No worries, thank you for taking a look anyway!

Quoted 26 lines

Am 20.06.23 um 02:42 schrieb Danilo Krummrich:

Quoted 24 lines

[SNIP]

include/drm/drm_gem.h +8 -0
index bbc721870c13..5ec8148a30ee 100644
@@ -36,6 +36,8 @@
   #include <linux/kref.h>
   #include <linux/dma-resv.h>
+#include <linux/list.h>
+#include <linux/mutex.h>
   #include <drm/drm_vma_manager.h>
@@ -379,6 +381,18 @@ struct drm_gem_object {
        */
       struct dma_resv _resv;
+    /**
+     * @gpuva:
+     *
+     * Provides the list of GPU VAs attached to this GEM object.
+     *
+     * Drivers should lock list accesses with the GEMs &dma_resv 

lock
+     * (&drm_gem_object.resv).
+     */
+    struct {

Quoted 8 lines
Quoted 6 lines

+        struct list_head list;
+    } gpuva;
+
      /**
       * @funcs:
       *

I'm pretty sure that it's not a good idea to attach this directly to the GEM object.

Why do you think so? IMHO having a common way to connect mappings to their backing buffers is a good thing, since every driver needs this connection anyway.

Quoted 7 lines
Quoted 3 lines

E.g. when a BO gets evicted, drivers can just iterate the list of mappings and, as the circumstances require, invalidate the corresponding mappings or to unmap all existing mappings of a given buffer.

What would be the advantage to let every driver implement a driver specific way of keeping this connection?

Flexibility. For example on amdgpu the mappings of a BO are groups by VM address spaces.

E.g. the BO points to multiple bo_vm structures which in turn have lists of their mappings.

Isn't this (almost) the same relationship I introduce with the GPUVA manager?

If you would switch over to the GPUVA manager right now, it would be that every GEM has a list of it's mappings (the gpuva list). The mapping is represented by struct drm_gpuva (of course embedded in driver specific structure(s)) which has a pointer to the VM address space it is part of, namely the GPUVA manager instance. And the GPUVA manager keeps a maple tree of it's mappings as well.

Quoted 7 lines
Quoted 5 lines
Quoted 3 lines

If you still would like to *directly* (indirectly you already have that relationship) keep a list of GPUVA managers (VM address spaces) per GEM, you could still do that in a driver specific way.

Do I miss something?

How do you efficiently find only the mappings of a BO in one VM?

Actually, I think this case should even be more efficient than with a BO having a list of GPUVAs (or mappings):

*than with a BO having a list of VMs:

Quoted 84 lines

Having a list of GPUVAs per GEM, each GPUVA has a pointer to it's VM. Hence, you'd only need to iterate the list of mappings for a given BO and check the mappings VM pointer.

Having a list of VMs per BO, you'd have to iterate the whole VM to find the mappings having a pointer to the given BO, right?

I'd think that a single VM potentially has more mapping entries than a single BO was mapped in multiple VMs.

Another case to consider is the case I originally had in mind choosing this relationship: finding all mappings for a given BO, which I guess all drivers need to do in order to invalidate mappings on BO eviction.

Having a list of VMs per BO, wouldn't you need to iterate all of the VMs entirely?

Quoted 8 lines

Keep in mind that we have cases where one BO is shared with hundreds of different VMs as well as potentially the number of mappings can be

Quoted 6 lines

10k.

Quoted 3 lines

Additional to that there is a state maschine associated with the mappings, e.g. if the page tables are up to date or need to be updated etc....

Quoted 7 lines
Quoted 5 lines
Quoted 3 lines
Quoted 1 line

Do you see cases where this kind of connection between mappings and backing buffers wouldn't be good enough? If so, which cases do you have in mind? Maybe we can cover them in a common way as well?

Yeah, we have tons of cases like that. But I have no idea how to generalize them.

They could still remain to be driver specific then, right?

Well does the mapping has a back pointer to the BO? And can that be optional NULL if there is no BO?

Yes to both.

- Danilo

Quoted 53 lines

Regards,
Christian.

Quoted 48 lines
Quoted 11 lines
Quoted 7 lines
Quoted 5 lines

As you wrote in the commit message it's highly driver specific what to map and where to map it.

In the end the common case should be that in a VA space at least every mapping being backed by a BO is represented by a struct drm_gpuva.

Oh, no! We also have mappings not backed by a BO at all! For example for partial resident textures or data routing to internal hw etc...

You can't be sure that a mapping is backed by a BO at all.

I fully agree, that's why I wrote "the common case should be that in a VA space at least every mapping *being backed by a BO* is represented by a struct drm_gpuva".

Mappings not being backed by an actual BO would not be linked to a GEM of course.

Quoted 8 lines
Quoted 6 lines
Quoted 4 lines

Instead I suggest to have a separate structure for mappings in a VA space which driver can then add to their GEM objects or whatever they want to map into their VMs.

Which kind of separate structure for mappings? Another one analogous to struct drm_gpuva?

Well similar to what amdgpu uses BO -> one structure for each combination of BO and VM -> mappings inside that VM

As explained above, I think that's exactly what the GPUVA manager does, just in another order:

BO has list of mappings, mappings have pointer to VM, VM has list (or actually a maple tree) of mappings.

You see any advantages or disadvantages of either order of relationships? For me it looks like it doesn't really matter which one to pick.

- Danilo

Quoted 14 lines

Regards,
Christian.

Quoted 9 lines

- Danilo

Quoted 5 lines

Regards,
Christian.

#14 lore raw

Re: [PATCH drm-next v5 03/14] drm: manager to keep track of GPUs VA mappings

Quoted 86 lines · Danilo Krummrich
Quoted 85 lines · Danilo Krummrich
Quoted 84 lines · Christian König

Am 22.06.23 um 16:22 schrieb Danilo Krummrich:

Quoted 83 lines · Christian König
Quoted 60 lines

Am 20.06.23 um 14:23 schrieb Danilo Krummrich:

Quoted 50 lines

Hi Christian,

Quoted 4 lines · Christian König

Hi Danilo,

sorry for the delayed reply. I've trying to dig myself out of a hole at the moment.

No worries, thank you for taking a look anyway!

Quoted 27 lines

Am 20.06.23 um 02:42 schrieb Danilo Krummrich:

Quoted 25 lines

[SNIP]

include/drm/drm_gem.h +8 -0
index bbc721870c13..5ec8148a30ee 100644
@@ -36,6 +36,8 @@
   #include <linux/kref.h>
   #include <linux/dma-resv.h>
+#include <linux/list.h>
+#include <linux/mutex.h>
   #include <drm/drm_vma_manager.h>
@@ -379,6 +381,18 @@ struct drm_gem_object {
        */
       struct dma_resv _resv;
+    /**
+     * @gpuva:
+     *
+     * Provides the list of GPU VAs attached to this GEM object.
+     *
+     * Drivers should lock list accesses with

the GEMs &dma_resv lock
+     * (&drm_gem_object.resv).
+     */
+    struct {
+        struct list_head list;

Quoted 8 lines
Quoted 5 lines

+    } gpuva;
+
      /**
       * @funcs:
       *

I'm pretty sure that it's not a good idea to attach this directly to the GEM object.

Why do you think so? IMHO having a common way to connect mappings to their backing buffers is a good thing, since every driver needs this connection anyway.

Quoted 8 lines

E.g. when a BO gets evicted, drivers can just iterate the list of mappings and, as the circumstances require, invalidate the corresponding mappings or to unmap all existing mappings of a given buffer.

What would be the advantage to let every driver implement a driver specific way of keeping this connection?

Quoted 5 lines

Flexibility. For example on amdgpu the mappings of a BO are groups by VM address spaces.

E.g. the BO points to multiple bo_vm structures which in turn have lists of their mappings.

Isn't this (almost) the same relationship I introduce with the GPUVA manager?

If you would switch over to the GPUVA manager right now, it would be that every GEM has a list of it's mappings (the gpuva list). The mapping is represented by struct drm_gpuva (of course embedded in driver specific structure(s)) which has a pointer to the VM address space it is part of, namely the GPUVA manager instance. And the GPUVA manager keeps a maple tree of it's mappings as well.

If you still would like to *directly* (indirectly you already have that relationship) keep a list of GPUVA managers (VM address spaces) per GEM, you could still do that in a driver specific way.

Quoted 27 lines
Quoted 6 lines
Quoted 3 lines
Quoted 1 line

Do I miss something?

How do you efficiently find only the mappings of a BO in one VM?

Actually, I think this case should even be more efficient than with a BO having a list of GPUVAs (or mappings):

*than with a BO having a list of VMs:

Quoted 18 lines

Having a list of GPUVAs per GEM, each GPUVA has a pointer to it's VM. Hence, you'd only need to iterate the list of mappings for a given BO and check the mappings VM pointer.

Having a list of VMs per BO, you'd have to iterate the whole VM to find the mappings having a pointer to the given BO, right?

I'd think that a single VM potentially has more mapping entries than a single BO was mapped in multiple VMs.

Another case to consider is the case I originally had in mind choosing this relationship: finding all mappings for a given BO, which I guess all drivers need to do in order to invalidate mappings on BO eviction.

Having a list of VMs per BO, wouldn't you need to iterate all of the VMs entirely?

FWIW I agree with Danilo here on basically all points.

- Having GPUVA list per GEM makes a ton of sense wrt eviction and invalidation. Xe had this list prior to GPUVA, after GPUVA it is just in a common place.
- From a GPUVA to you can resolve a GEM
- GPUVA can have NULL GEM (userptr or sparse binding in Xe)
- From a VM all GPUVA can be resolved via the maple tree

This seems to do everything any reasonable driver would need.

Quoted 5 lines
Quoted 5 lines
Quoted 5 lines

Keep in mind that we have cases where one BO is shared with hundreds of different VMs as well as potentially the number of mappings can be >10k.

Christian - Can you explain this use case? It seems like an odd one but regardless the GPUVA design still works we'd just have 10k GPUVA in the BO's list.

Quoted 14 lines
Quoted 14 lines
Quoted 14 lines
Quoted 14 lines
Quoted 13 lines

Additional to that there is a state maschine associated with the mappings, e.g. if the page tables are up to date or need to be updated etc....

Quoted 4 lines

Do you see cases where this kind of connection between mappings and backing buffers wouldn't be good enough? If so, which cases do you have in mind? Maybe we can cover them in a common way as well?

Yeah, we have tons of cases like that. But I have no idea how to generalize them.

Quoted 7 lines
Quoted 7 lines
Quoted 4 lines
Quoted 1 line

They could still remain to be driver specific then, right?

Well does the mapping has a back pointer to the BO? And can that be optional NULL if there is no BO?

Yes to both.

In Xe we have GPUVA with NULL BOs for userptrs and for sparse bindings. It seems to work rather well.

Matt

Quoted 73 lines
Quoted 72 lines

- Danilo

Quoted 70 lines

Regards,
Christian.

Quoted 65 lines
Quoted 11 lines
Quoted 10 lines
Quoted 6 lines

As you wrote in the commit message it's highly driver specific what to map and where to map it.

In the end the common case should be that in a VA space at least every mapping being backed by a BO is represented by a struct drm_gpuva.

Quoted 5 lines

Oh, no! We also have mappings not backed by a BO at all! For example for partial resident textures or data routing to internal hw etc...

You can't be sure that a mapping is backed by a BO at all.

I fully agree, that's why I wrote "the common case should be that in a VA space at least every mapping *being backed by a BO* is represented by a struct drm_gpuva".

Mappings not being backed by an actual BO would not be linked to a GEM of course.

Quoted 8 lines
Quoted 8 lines
Quoted 7 lines

Instead I suggest to have a separate structure for mappings in a VA space which driver can then add to their GEM objects or whatever they want to map into their VMs.

Quoted 5 lines
Quoted 2 lines

Which kind of separate structure for mappings? Another one analogous to struct drm_gpuva?

Well similar to what amdgpu uses BO -> one structure for each combination of BO and VM -> mappings inside that VM

As explained above, I think that's exactly what the GPUVA manager does, just in another order:

BO has list of mappings, mappings have pointer to VM, VM has list (or actually a maple tree) of mappings.

You see any advantages or disadvantages of either order of relationships? For me it looks like it doesn't really matter which one to pick.

- Danilo

Quoted 14 lines

Regards,
Christian.

Quoted 9 lines

- Danilo

Quoted 5 lines

Regards,
Christian.

#15 lore raw

Re: [PATCH drm-next v5 03/14] drm: manager to keep track of GPUs VA mappings

In reply to Danilo Krummrich

Am 22.06.23 um 17:07 schrieb Danilo Krummrich:

Quoted 65 lines · Danilo Krummrich
Quoted 64 lines · Christian König
Quoted 63 lines

Am 22.06.23 um 16:22 schrieb Danilo Krummrich:

Quoted 62 lines · Christian König
Quoted 47 lines

Am 20.06.23 um 14:23 schrieb Danilo Krummrich:

Quoted 46 lines

Hi Christian,

Quoted 3 lines · Christian König

Hi Danilo,

sorry for the delayed reply. I've trying to dig myself out of a hole at the moment.

No worries, thank you for taking a look anyway!

Quoted 26 lines

Am 20.06.23 um 02:42 schrieb Danilo Krummrich:

Quoted 24 lines

[SNIP]

include/drm/drm_gem.h +8 -0
index bbc721870c13..5ec8148a30ee 100644
@@ -36,6 +36,8 @@
   #include <linux/kref.h>
   #include <linux/dma-resv.h>
+#include <linux/list.h>
+#include <linux/mutex.h>
   #include <drm/drm_vma_manager.h>
@@ -379,6 +381,18 @@ struct drm_gem_object {
        */
       struct dma_resv _resv;
+    /**
+     * @gpuva:
+     *
+     * Provides the list of GPU VAs attached to this GEM object.
+     *
+     * Drivers should lock list accesses with the GEMs 

&dma_resv lock
+     * (&drm_gem_object.resv).
+     */
+    struct {

Quoted 8 lines
Quoted 6 lines

+        struct list_head list;
+    } gpuva;
+
      /**
       * @funcs:
       *

I'm pretty sure that it's not a good idea to attach this directly to the GEM object.

Why do you think so? IMHO having a common way to connect mappings to their backing buffers is a good thing, since every driver needs this connection anyway.

Quoted 7 lines
Quoted 3 lines

E.g. when a BO gets evicted, drivers can just iterate the list of mappings and, as the circumstances require, invalidate the corresponding mappings or to unmap all existing mappings of a given buffer.

What would be the advantage to let every driver implement a driver specific way of keeping this connection?

Flexibility. For example on amdgpu the mappings of a BO are groups by VM address spaces.

E.g. the BO points to multiple bo_vm structures which in turn have lists of their mappings.

Isn't this (almost) the same relationship I introduce with the GPUVA manager?

If you would switch over to the GPUVA manager right now, it would be that every GEM has a list of it's mappings (the gpuva list). The mapping is represented by struct drm_gpuva (of course embedded in driver specific structure(s)) which has a pointer to the VM address space it is part of, namely the GPUVA manager instance. And the GPUVA manager keeps a maple tree of it's mappings as well.

If you still would like to *directly* (indirectly you already have that relationship) keep a list of GPUVA managers (VM address spaces) per GEM, you could still do that in a driver specific way.

Quoted 10 lines
Quoted 5 lines
Quoted 3 lines
Quoted 1 line

Do I miss something?

How do you efficiently find only the mappings of a BO in one VM?

Actually, I think this case should even be more efficient than with a BO having a list of GPUVAs (or mappings):

*than with a BO having a list of VMs:

Quoted 2 lines

Having a list of GPUVAs per GEM, each GPUVA has a pointer to it's VM. Hence, you'd only need to iterate the list of mappings for a given BO and check the mappings VM pointer.

Yeah, and that is extremely time consuming if you have tons of mappings in different VMs.

Quoted 2 lines
Quoted 2 lines

Having a list of VMs per BO, you'd have to iterate the whole VM to find the mappings having a pointer to the given BO, right?

No, you don't seem to understand what I'm suggesting.

Currently you have a list of mappings attached to the BO, so when you need to make sure that a specific BO is up to date in a specific VM you either need to iterate over the VM or the BO. Neither of that is a good idea.

What you need is a representation of the data used for each BO+VM combination. In other words another indirection which allows you to handle all the mappings of a BO inside a VM at once.

Quoted 6 lines
Quoted 6 lines

I'd think that a single VM potentially has more mapping entries than a single BO was mapped in multiple VMs.

Another case to consider is the case I originally had in mind choosing this relationship: finding all mappings for a given BO, which I guess all drivers need to do in order to invalidate mappings on BO eviction.

Having a list of VMs per BO, wouldn't you need to iterate all of the VMs entirely?

No, see how amdgpu works.

Regards,
Christian.

#16 lore raw

Re: [PATCH drm-next v5 03/14] drm: manager to keep track of GPUs VA mappings

Quoted 80 lines · Christian König

Am 22.06.23 um 17:07 schrieb Danilo Krummrich:

Quoted 65 lines · Danilo Krummrich
Quoted 64 lines · Christian König
Quoted 63 lines

Am 22.06.23 um 16:22 schrieb Danilo Krummrich:

Quoted 62 lines · Christian König
Quoted 47 lines

Am 20.06.23 um 14:23 schrieb Danilo Krummrich:

Quoted 46 lines

Hi Christian,

Quoted 3 lines · Christian König

Hi Danilo,

sorry for the delayed reply. I've trying to dig myself out of a hole at the moment.

No worries, thank you for taking a look anyway!

Quoted 26 lines

Am 20.06.23 um 02:42 schrieb Danilo Krummrich:

Quoted 24 lines

[SNIP]

include/drm/drm_gem.h +8 -0
index bbc721870c13..5ec8148a30ee 100644
@@ -36,6 +36,8 @@
   #include <linux/kref.h>
   #include <linux/dma-resv.h>
+#include <linux/list.h>
+#include <linux/mutex.h>
   #include <drm/drm_vma_manager.h>
@@ -379,6 +381,18 @@ struct drm_gem_object {
        */
       struct dma_resv _resv;
+    /**
+     * @gpuva:
+     *
+     * Provides the list of GPU VAs attached to this GEM object.
+     *
+     * Drivers should lock list accesses with the GEMs 

&dma_resv lock
+     * (&drm_gem_object.resv).
+     */
+    struct {

Quoted 8 lines
Quoted 6 lines

+        struct list_head list;
+    } gpuva;
+
      /**
       * @funcs:
       *

I'm pretty sure that it's not a good idea to attach this directly to the GEM object.

Why do you think so? IMHO having a common way to connect mappings to their backing buffers is a good thing, since every driver needs this connection anyway.

Quoted 7 lines
Quoted 3 lines

E.g. when a BO gets evicted, drivers can just iterate the list of mappings and, as the circumstances require, invalidate the corresponding mappings or to unmap all existing mappings of a given buffer.

What would be the advantage to let every driver implement a driver specific way of keeping this connection?

Flexibility. For example on amdgpu the mappings of a BO are groups by VM address spaces.

E.g. the BO points to multiple bo_vm structures which in turn have lists of their mappings.

Isn't this (almost) the same relationship I introduce with the GPUVA manager?

If you would switch over to the GPUVA manager right now, it would be that every GEM has a list of it's mappings (the gpuva list). The mapping is represented by struct drm_gpuva (of course embedded in driver specific structure(s)) which has a pointer to the VM address space it is part of, namely the GPUVA manager instance. And the GPUVA manager keeps a maple tree of it's mappings as well.

If you still would like to *directly* (indirectly you already have that relationship) keep a list of GPUVA managers (VM address spaces) per GEM, you could still do that in a driver specific way.

Quoted 10 lines
Quoted 5 lines
Quoted 3 lines
Quoted 1 line

Do I miss something?

How do you efficiently find only the mappings of a BO in one VM?

Actually, I think this case should even be more efficient than with a BO having a list of GPUVAs (or mappings):

*than with a BO having a list of VMs:

Quoted 2 lines

Having a list of GPUVAs per GEM, each GPUVA has a pointer to it's VM. Hence, you'd only need to iterate the list of mappings for a given BO and check the mappings VM pointer.

Yeah, and that is extremely time consuming if you have tons of mappings in different VMs.

Quoted 1 line
Quoted 1 line
Quoted 7 lines
Quoted 1 line
Quoted 1 line

Having a list of VMs per BO, you'd have to iterate the whole VM to find the mappings having a pointer to the given BO, right?

No, you don't seem to understand what I'm suggesting.

Currently you have a list of mappings attached to the BO, so when you need to make sure that a specific BO is up to date in a specific VM you either need to iterate over the VM or the BO. Neither of that is a good idea.

What you need is a representation of the data used for each BO+VM combination. In other words another indirection which allows you to handle all the mappings of a BO inside a VM at once.

Ok, after having a quick look at amdgpu, I can see what you mean.

The missing piece for me was that the BO+VM abstraction itself keeps a list of mappings for this specific BO and VM.

Just to make it obvious for other people following the discussion, let me quickly sketch up how this approach would look like for the GPUVA manager:

1. We would need a new structure to represent the BO+VM combination, something like:

	struct drm_gpuva_mgr_gem {
    		struct drm_gpuva_manager *mgr;
		struct drm_gem_object *obj;
		struct list_head gpuva_list;
	};

with a less horrible name, hopefully.

2. Create an instance of struct drm_gpuva_mgr_gem once a GEM becomes associated with a GPUVA manager (VM) and attach it to the GEMs, as by now, "gpuva" list.

In amdgpu, for example, this seems to be the case once a GEM object is opened, since there is one VM per file_priv.

However, for other drivers this could be different, hence drivers would need to take care about this.

3. Attach GPUVAs to the new gpuva_list of the corresponding instance of struct drm_gpuva_mgr_gem.

4. Drivers would need to clean up the instance of struct drm_gpuva_mgr_gem, once the GEM is not associated with the GPUVA manager anymore.

As pointed out by Christian, this would optimize the "get all mappings backed by a specific BO from a given VM" use case.

The question for me is, do other drivers than amdgpu commonly need this?

And what does amdgpu need this for? Maybe amdgpu does something we're not doing (yet)?

Christian - I know you didn't ask for "do it the way amdgpu does", instead you voted for keeping it entirely driver specific. But I think everyone is pretty close and I'm still optimistic that we could just generalize this.

- Danilo

Quoted 13 lines
Quoted 7 lines
Quoted 7 lines

I'd think that a single VM potentially has more mapping entries than a single BO was mapped in multiple VMs.

Another case to consider is the case I originally had in mind choosing this relationship: finding all mappings for a given BO, which I guess all drivers need to do in order to invalidate mappings on BO eviction.

Having a list of VMs per BO, wouldn't you need to iterate all of the VMs entirely?

No, see how amdgpu works.

Regards,
Christian.

#17 lore raw

Re: [PATCH drm-next v5 03/14] drm: manager to keep track of GPUs VA mappings

Am 23.06.23 um 15:55 schrieb Danilo Krummrich:

Quoted 35 lines

[SNIP]

Quoted 12 lines
Quoted 8 lines
Quoted 3 lines
Quoted 1 line

How do you efficiently find only the mappings of a BO in one VM?

Actually, I think this case should even be more efficient than with a BO having a list of GPUVAs (or mappings):

*than with a BO having a list of VMs:

Quoted 2 lines

Having a list of GPUVAs per GEM, each GPUVA has a pointer to it's VM. Hence, you'd only need to iterate the list of mappings for a given BO and check the mappings VM pointer.

Yeah, and that is extremely time consuming if you have tons of mappings in different VMs.

Quoted 1 line
Quoted 1 line
Quoted 7 lines
Quoted 1 line
Quoted 1 line

Having a list of VMs per BO, you'd have to iterate the whole VM to find the mappings having a pointer to the given BO, right?

No, you don't seem to understand what I'm suggesting.

Currently you have a list of mappings attached to the BO, so when you need to make sure that a specific BO is up to date in a specific VM you either need to iterate over the VM or the BO. Neither of that is a good idea.

What you need is a representation of the data used for each BO+VM combination. In other words another indirection which allows you to handle all the mappings of a BO inside a VM at once.

Ok, after having a quick look at amdgpu, I can see what you mean.

The missing piece for me was that the BO+VM abstraction itself keeps a list of mappings for this specific BO and VM.

Just to make it obvious for other people following the discussion, let me quickly sketch up how this approach would look like for the GPUVA manager:

1. We would need a new structure to represent the BO+VM combination, something like:

    struct drm_gpuva_mgr_gem {
            struct drm_gpuva_manager *mgr;
        struct drm_gem_object *obj;
        struct list_head gpuva_list;
    };

Quoted 7 lines

with a less horrible name, hopefully.

2. Create an instance of struct drm_gpuva_mgr_gem once a GEM becomes associated with a GPUVA manager (VM) and attach it to the GEMs, as by now, "gpuva" list.

In amdgpu, for example, this seems to be the case once a GEM object is opened, since there is one VM per file_priv.

However, for other drivers this could be different, hence drivers would need to take care about this.

Yes, exactly that.

Quoted 10 lines

3. Attach GPUVAs to the new gpuva_list of the corresponding instance of struct drm_gpuva_mgr_gem.

4. Drivers would need to clean up the instance of struct drm_gpuva_mgr_gem, once the GEM is not associated with the GPUVA manager anymore.

As pointed out by Christian, this would optimize the "get all mappings backed by a specific BO from a given VM" use case.

The question for me is, do other drivers than amdgpu commonly need this?

I have no idea.

Quoted 2 lines

And what does amdgpu need this for? Maybe amdgpu does something we're not doing (yet)?

Basically when we do a CS we need to make sure that the VM used by this CS is up to date. For this we walk over the relocation list of BOs and check the status of each BO+VM structure.

This is done because we don't want to update all VMs at the same time, but rather just those who needs the update.

Quoted 2 lines

Christian - I know you didn't ask for "do it the way amdgpu does", instead you voted for keeping it entirely driver specific. But I think everyone is pretty close and I'm still optimistic that we could just generalize this.

Well, you should *not* necessarily do it like amdgpu does! Basically the implementation in amdgpu was driven by requirements, e.g. we need that, let's do it like this.

It's perfectly possible that other requirements (e.g. focus on Vulkan) lead to a completely different implementation.

It's just that ideally I would like to have an implementation where I can apply at least the basics to amdgpu as well.

Regards,
Christian.

Quoted 17 lines

- Danilo

Quoted 13 lines
Quoted 7 lines
Quoted 7 lines

I'd think that a single VM potentially has more mapping entries than a single BO was mapped in multiple VMs.

Another case to consider is the case I originally had in mind choosing this relationship: finding all mappings for a given BO, which I guess all drivers need to do in order to invalidate mappings on BO eviction.

Having a list of VMs per BO, wouldn't you need to iterate all of the VMs entirely?

No, see how amdgpu works.

Regards,
Christian.

maintainer
#18 lore raw

Re: [PATCH drm-next v5 03/14] drm: manager to keep track of GPUs VA mappings

Quoted 11 lines
Quoted 4 lines

As pointed out by Christian, this would optimize the "get all mappings backed by a specific BO from a given VM" use case.

The question for me is, do other drivers than amdgpu commonly need this?

I have no idea.

Quoted 3 lines

And what does amdgpu need this for? Maybe amdgpu does something we're not doing (yet)?

Quoted 6 lines

Basically when we do a CS we need to make sure that the VM used by this CS is up to date. For this we walk over the relocation list of BOs and check the status of each BO+VM structure.

This is done because we don't want to update all VMs at the same time, but rather just those who needs the update.

This seems like a legacy from GL and possibly older vulkan, going forward vulkan can't rely on passing lists of objects into the kernel due to things like buffer_device_address, I'm not sure we should optimise for this situation, and moving the tracking list into the drivers is going to mean having a bunch of drivers all having the same boilerplate, to do the same thing just so amdgpu can't avoid doing it.

Now there might be some benchmark somewhere we can produce a benefit in this, and if there is then we should consider going this way for all drivers and not just allowing drivers to choose their own path here.

Quoted 10 lines
Quoted 5 lines

Christian - I know you didn't ask for "do it the way amdgpu does", instead you voted for keeping it entirely driver specific. But I think everyone is pretty close and I'm still optimistic that we could just generalize this.

Well, you should *not* necessarily do it like amdgpu does! Basically the implementation in amdgpu was driven by requirements, e.g. we need that, let's do it like this.

Quoted 6 lines

It's perfectly possible that other requirements (e.g. focus on Vulkan) lead to a completely different implementation.

It's just that ideally I would like to have an implementation where I can apply at least the basics to amdgpu as well.

I think we can still do that just either have an option to disable using the list internally in the gpuva or have the driver keep it's own tracking alongside, there may still be use cases where it can use the gpuva tracking instead of it's own.

I don't think we should forklift what is pretty likely to be common code across every driver that uses this going forward just to benefit an amdgpu design decision for older stacks.

Dave.

#19 lore raw

Re: [PATCH drm-next v5 03/14] drm: manager to keep track of GPUs VA mappings

In reply to Danilo Krummrich

Hi Danilo,
One comment below, but otherwise it looks great. Thanks for adding the example! Thanks,
Donald

Quoted 26 lines · Danilo Krummrich

+/**
+ * DOC: Overview
+ *
+ * The DRM GPU VA Manager, represented by struct drm_gpuva_manager keeps track
+ * of a GPU's virtual address (VA) space and manages the corresponding virtual
+ * mappings represented by &drm_gpuva objects. It also keeps track of the
+ * mapping's backing &drm_gem_object buffers.
+ *
+ * &drm_gem_object buffers maintain a list of &drm_gpuva objects representing
+ * all existent GPU VA mappings using this &drm_gem_object as backing buffer.
+ *
+ * GPU VAs can be flagged as sparse, such that drivers may use GPU VAs to also
+ * keep track of sparse PTEs in order to support Vulkan 'Sparse Resources'.
+ *
+ * The GPU VA manager internally uses a &maple_tree to manage the
+ * &drm_gpuva mappings within a GPU's virtual address space.
+ *
+ * The &drm_gpuva_manager contains a special &drm_gpuva representing the
+ * portion of VA space reserved by the kernel. This node is initialized together
+ * with the GPU VA manager instance and removed when the GPU VA manager is
+ * destroyed.
+ *
+ * In a typical application drivers would embed struct drm_gpuva_manager and
+ * struct drm_gpuva within their own driver specific structures, there won't be
+ * any memory allocations of it's own nor memory allocations of &drm_gpuva

Quoted 8 lines

+ * entries.
+ *
+ * However, the &drm_gpuva_manager needs to allocate nodes for it's internal
+ * tree structures when &drm_gpuva entries are inserted. In order to support
+ * inserting &drm_gpuva entries from dma-fence signalling critical sections the
+ * &drm_gpuva_manager provides struct drm_gpuva_prealloc. Drivers may create
+ * pre-allocated nodes which drm_gpuva_prealloc_create() and subsequently insert
+ * a new &drm_gpuva entry with drm_gpuva_insert_prealloc().

I think it might be worth moving or repeating this paragraph to 'Split and Merge' where I've added the other comment below. I think these functions are only used to set up for drm_gpuva_sm_map(). Please ignore me if I'm wrong.

Quoted 10 lines

+ */

drivers/gpu/drm/drm_gpuva_mgr.c new file +9 -0
+
+/**
+ * DOC: Split and Merge
+ *
+ * Besides it's capability to manage and represent a GPU VA space, the
+ * &drm_gpuva_manager also provides functions to let the &drm_gpuva_manager
+ * calculate a sequence of operations to satisfy a given map or unmap request.
+ *
+ * Therefore the DRM GPU VA manager provides an algorithm implementing splitting

drivers/gpu/drm/drm_gpuva_mgr.c line 83

Quoted 10 lines
drivers/gpu/drm/drm_gpuva_mgr.c new file +10 -0
+ * and merging of existent GPU VA mappings with the ones that are requested to
+ * be mapped or unmapped. This feature is required by the Vulkan API to
+ * implement Vulkan 'Sparse Memory Bindings' - drivers UAPIs often refer to this
+ * as VM BIND.
+ *
+ * Drivers can call drm_gpuva_sm_map() to receive a sequence of callbacks
+ * containing map, unmap and remap operations for a given newly requested
+ * mapping. The sequence of callbacks represents the set of operations to
+ * execute in order to integrate the new mapping cleanly into the current state
+ * of the GPU VA space.

Here

Quoted 7 lines

+ *
+ * Depending on how the new GPU VA mapping intersects with the existent mappings
+ * of the GPU VA space the &drm_gpuva_fn_ops callbacks contain an arbitrary
+ * amount of unmap operations, a maximum of two remap operations and a single
+ * map operation. The caller might receive no callback at all if no operation is
+ * required, e.g. if the requested mapping already exists in the exact same way.

#20 lore raw

PATCH v5 4/14 drm: debugfs: provide infrastructure to dump a DRM GPU VA space

In reply to Danilo Krummrich

This commit adds a function to dump a DRM GPU VA space and a macro for drivers to register the struct drm_info_list 'gpuvas' entry.

Most likely, most drivers might maintain one DRM GPU VA space per struct drm_file, but there might also be drivers not having a fixed relation between DRM GPU VA spaces and a DRM core infrastructure, hence we need the indirection via the driver iterating it's maintained DRM GPU VA spaces.

---
 drivers/gpu/drm/drm_debugfs.c | 41 +++++++++++++++++++++++++++++++++++
 include/drm/drm_debugfs.h     | 25 +++++++++++++++++++++
 2 files changed, 66 insertions(+)
drivers/gpu/drm/drm_debugfs.c +41 -0
index 4855230ba2c6..82180fb1c200 100644
@@ -39,6 +39,7 @@
 #include <drm/drm_file.h>
 #include <drm/drm_gem.h>
 #include <drm/drm_managed.h>
+#include <drm/drm_gpuva_mgr.h>
 
 #include "drm_crtc_internal.h"
 #include "drm_internal.h"
@@ -175,6 +176,46 @@ static const struct file_operations drm_debugfs_fops = {
 	.release = single_release,
 };
 
+/**
+ * drm_debugfs_gpuva_info - dump the given DRM GPU VA space
+ * @m: pointer to the &seq_file to write
+ * @mgr: the &drm_gpuva_manager representing the GPU VA space
+ *
+ * Dumps the GPU VA mappings of a given DRM GPU VA manager.
+ *
+ * For each DRM GPU VA space drivers should call this function from their
+ * &drm_info_list's show callback.
+ *
+ * Returns: 0 on success, -ENODEV if the &mgr is not initialized
+ */
+int drm_debugfs_gpuva_info(struct seq_file *m,
+			   struct drm_gpuva_manager *mgr)
+{
+	DRM_GPUVA_ITER(it, mgr, 0);
+	struct drm_gpuva *va, *kva = &mgr->kernel_alloc_node;
+
+	if (!mgr->name)
+		return -ENODEV;
+
+	seq_printf(m, "DRM GPU VA space (%s) [0x%016llx;0x%016llx]\n",
+		   mgr->name, mgr->mm_start, mgr->mm_start + mgr->mm_range);
+	seq_printf(m, "Kernel reserved node [0x%016llx;0x%016llx]\n",
+		   kva->va.addr, kva->va.addr + kva->va.range);
+	seq_puts(m, "\n");
+	seq_puts(m, " VAs | start              | range              | end                | object             | object offset\n");
+	seq_puts(m, "-------------------------------------------------------------------------------------------------------------\n");
+	drm_gpuva_iter_for_each(va, it) {
+		if (unlikely(va == &mgr->kernel_alloc_node))
+			continue;
+
+		seq_printf(m, "     | 0x%016llx | 0x%016llx | 0x%016llx | 0x%016llx | 0x%016llx\n",
+			   va->va.addr, va->va.range, va->va.addr + va->va.range,
+			   (u64)va->gem.obj, va->gem.offset);
+	}
+
+	return 0;
+}
+EXPORT_SYMBOL(drm_debugfs_gpuva_info);
 
 /**
  * drm_debugfs_create_files - Initialize a given set of debugfs files for DRM
include/drm/drm_debugfs.h +25 -0
index 7616f457ce70..cb2c1956a214 100644
@@ -34,6 +34,22 @@
 
 #include <linux/types.h>
 #include <linux/seq_file.h>
+
+#include <drm/drm_gpuva_mgr.h>
+
+/**
+ * DRM_DEBUGFS_GPUVA_INFO - &drm_info_list entry to dump a GPU VA space
+ * @show: the &drm_info_list's show callback
+ * @data: driver private data
+ *
+ * Drivers should use this macro to define a &drm_info_list entry to provide a
+ * debugfs file for dumping the GPU VA space regions and mappings.
+ *
+ * For each DRM GPU VA space drivers should call drm_debugfs_gpuva_info() from
+ * their @show callback.
+ */
+#define DRM_DEBUGFS_GPUVA_INFO(show, data) {"gpuvas", show, DRIVER_GEM_GPUVA, data}
+
 /**
  * struct drm_info_list - debugfs info list entry
  *
@@ -134,6 +150,9 @@ void drm_debugfs_add_file(struct drm_device *dev, const char *name,
 
 void drm_debugfs_add_files(struct drm_device *dev,
 			   const struct drm_debugfs_info *files, int count);
+
+int drm_debugfs_gpuva_info(struct seq_file *m,
+			   struct drm_gpuva_manager *mgr);
 #else
 static inline void drm_debugfs_create_files(const struct drm_info_list *files,
 					    int count, struct dentry *root,
@@ -155,6 +174,12 @@ static inline void drm_debugfs_add_files(struct drm_device *dev,
 					 const struct drm_debugfs_info *files,
 					 int count)
 {}
+
+static inline int drm_debugfs_gpuva_info(struct seq_file *m,
+					 struct drm_gpuva_manager *mgr)
+{
+	return 0;
+}
 #endif
 
 #endif /* _DRM_DEBUGFS_H_ */
Signature · 1 line
2.40.1
#21 lore raw

PATCH v5 5/14 drm/nouveau: new VM_BIND uapi interfaces

In reply to Danilo Krummrich

This commit provides the interfaces for the new UAPI motivated by the Vulkan API. It allows user mode drivers (UMDs) to:

1) Initialize a GPU virtual address (VA) space via the new DRM_IOCTL_NOUVEAU_VM_INIT ioctl. UMDs can provide a kernel reserved VA area.

2) Bind and unbind GPU VA space mappings via the new DRM_IOCTL_NOUVEAU_VM_BIND ioctl.

3) Execute push buffers with the new DRM_IOCTL_NOUVEAU_EXEC ioctl.

Both, DRM_IOCTL_NOUVEAU_VM_BIND and DRM_IOCTL_NOUVEAU_EXEC support asynchronous processing with DRM syncobjs as synchronization mechanism.

The default DRM_IOCTL_NOUVEAU_VM_BIND is synchronous processing, DRM_IOCTL_NOUVEAU_EXEC supports asynchronous processing only.

---
 Documentation/gpu/driver-uapi.rst |   8 ++
 include/uapi/drm/nouveau_drm.h    | 209 ++++++++++++++++++++++++++++++
 2 files changed, 217 insertions(+)
Documentation/gpu/driver-uapi.rst +8 -0
index 4411e6919a3d..9c7ca6e33a68 100644
@@ -6,3 +6,11 @@ drm/i915 uAPI
 =============
 
 .. kernel-doc:: include/uapi/drm/i915_drm.h
+
+drm/nouveau uAPI
+================
+
+VM_BIND / EXEC uAPI
+-------------------
+
+.. kernel-doc:: include/uapi/drm/nouveau_drm.h
include/uapi/drm/nouveau_drm.h +209 -0

227 lines not on this page · raw message

Signature · 1 line
2.40.1
#22 lore raw

PATCH v5 6/14 drm/nouveau: get vmm via nouveau_cli_vmm()

In reply to Danilo Krummrich

Provide a getter function for the client's current vmm context. Since we'll add a new (u)vmm context for UMD bindings in subsequent commits, this will keep the code clean.

---
 drivers/gpu/drm/nouveau/nouveau_bo.c   | 2 +-
 drivers/gpu/drm/nouveau/nouveau_chan.c | 2 +-
 drivers/gpu/drm/nouveau/nouveau_drv.h  | 9 +++++++++
 drivers/gpu/drm/nouveau/nouveau_gem.c  | 6 +++---
 4 files changed, 14 insertions(+), 5 deletions(-)
drivers/gpu/drm/nouveau/nouveau_bo.c +1 -1
index c2ec91cc845d..7724fe63067d 100644
@@ -204,7 +204,7 @@ nouveau_bo_alloc(struct nouveau_cli *cli, u64 *size, int *align, u32 domain,
 	struct nouveau_drm *drm = cli->drm;
 	struct nouveau_bo *nvbo;
 	struct nvif_mmu *mmu = &cli->mmu;
-	struct nvif_vmm *vmm = cli->svm.cli ? &cli->svm.vmm : &cli->vmm.vmm;
+	struct nvif_vmm *vmm = &nouveau_cli_vmm(cli)->vmm;
 	int i, pi = -1;
 
 	if (!*size) {
drivers/gpu/drm/nouveau/nouveau_chan.c +1 -1
index e648ecd0c1a0..1068abe41024 100644
@@ -148,7 +148,7 @@ nouveau_channel_prep(struct nouveau_drm *drm, struct nvif_device *device,
 
 	chan->device = device;
 	chan->drm = drm;
-	chan->vmm = cli->svm.cli ? &cli->svm : &cli->vmm;
+	chan->vmm = nouveau_cli_vmm(cli);
 	atomic_set(&chan->killed, 0);
 
 	/* allocate memory for dma push buffer */
drivers/gpu/drm/nouveau/nouveau_drv.h +9 -0
index b5de312a523f..81350e685b50 100644
@@ -112,6 +112,15 @@ struct nouveau_cli_work {
 	struct dma_fence_cb cb;
 };
 
+static inline struct nouveau_vmm *
+nouveau_cli_vmm(struct nouveau_cli *cli)
+{
+	if (cli->svm.cli)
+		return &cli->svm;
+
+	return &cli->vmm;
+}
+
 void nouveau_cli_work_queue(struct nouveau_cli *, struct dma_fence *,
 			    struct nouveau_cli_work *);
 
drivers/gpu/drm/nouveau/nouveau_gem.c +3 -3
index ab9062e50977..45ca4eb98f54 100644
@@ -103,7 +103,7 @@ nouveau_gem_object_open(struct drm_gem_object *gem, struct drm_file *file_priv)
 	struct nouveau_bo *nvbo = nouveau_gem_object(gem);
 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
 	struct device *dev = drm->dev->dev;
-	struct nouveau_vmm *vmm = cli->svm.cli ? &cli->svm : &cli->vmm;
+	struct nouveau_vmm *vmm = nouveau_cli_vmm(cli);
 	struct nouveau_vma *vma;
 	int ret;
 
@@ -180,7 +180,7 @@ nouveau_gem_object_close(struct drm_gem_object *gem, struct drm_file *file_priv)
 	struct nouveau_bo *nvbo = nouveau_gem_object(gem);
 	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
 	struct device *dev = drm->dev->dev;
-	struct nouveau_vmm *vmm = cli->svm.cli ? &cli->svm : & cli->vmm;
+	struct nouveau_vmm *vmm = nouveau_cli_vmm(cli);
 	struct nouveau_vma *vma;
 	int ret;
 
@@ -269,7 +269,7 @@ nouveau_gem_info(struct drm_file *file_priv, struct drm_gem_object *gem,
 {
 	struct nouveau_cli *cli = nouveau_cli(file_priv);
 	struct nouveau_bo *nvbo = nouveau_gem_object(gem);
-	struct nouveau_vmm *vmm = cli->svm.cli ? &cli->svm : &cli->vmm;
+	struct nouveau_vmm *vmm = nouveau_cli_vmm(cli);
 	struct nouveau_vma *vma;
 
 	if (is_power_of_2(nvbo->valid_domains))
Signature · 1 line
2.40.1
#23 lore raw

PATCH v5 7/14 drm/nouveau: bo: initialize GEM GPU VA interface

In reply to Danilo Krummrich

Initialize the GEM's DRM GPU VA manager interface in preparation for the (u)vmm implementation, provided by subsequent commits, to make use of it.

---
 drivers/gpu/drm/nouveau/nouveau_bo.c | 3 +++
 1 file changed, 3 insertions(+)
drivers/gpu/drm/nouveau/nouveau_bo.c +3 -0
index 7724fe63067d..057bc995f19b 100644
@@ -215,11 +215,14 @@ nouveau_bo_alloc(struct nouveau_cli *cli, u64 *size, int *align, u32 domain,
 	nvbo = kzalloc(sizeof(struct nouveau_bo), GFP_KERNEL);
 	if (!nvbo)
 		return ERR_PTR(-ENOMEM);
+
 	INIT_LIST_HEAD(&nvbo->head);
 	INIT_LIST_HEAD(&nvbo->entry);
 	INIT_LIST_HEAD(&nvbo->vma_list);
 	nvbo->bo.bdev = &drm->ttm.bdev;
 
+	drm_gem_gpuva_init(&nvbo->bo.base);
+
 	/* This is confusing, and doesn't actually mean we want an uncached
 	 * mapping, but is what NOUVEAU_GEM_DOMAIN_COHERENT gets translated
 	 * into in nouveau_gem_new().
Signature · 1 line
2.40.1
#24 lore raw

PATCH v5 8/14 drm/nouveau: move usercopy helpers to nouveau_drv.h

In reply to Danilo Krummrich

Move the usercopy helpers to a common driver header file to make it usable for the new API added in subsequent commits.

---
 drivers/gpu/drm/nouveau/nouveau_drv.h | 26 ++++++++++++++++++++++++++
 drivers/gpu/drm/nouveau/nouveau_gem.c | 26 --------------------------
 2 files changed, 26 insertions(+), 26 deletions(-)
drivers/gpu/drm/nouveau/nouveau_drv.h +26 -0
index 81350e685b50..20a7f31b9082 100644
@@ -130,6 +130,32 @@ nouveau_cli(struct drm_file *fpriv)
 	return fpriv ? fpriv->driver_priv : NULL;
 }
 
+static inline void
+u_free(void *addr)
+{
+	kvfree(addr);
+}
+
+static inline void *
+u_memcpya(uint64_t user, unsigned nmemb, unsigned size)
+{
+	void *mem;
+	void __user *userptr = (void __force __user *)(uintptr_t)user;
+
+	size *= nmemb;
+
+	mem = kvmalloc(size, GFP_KERNEL);
+	if (!mem)
+		return ERR_PTR(-ENOMEM);
+
+	if (copy_from_user(mem, userptr, size)) {
+		u_free(mem);
+		return ERR_PTR(-EFAULT);
+	}
+
+	return mem;
+}
+
 #include <nvif/object.h>
 #include <nvif/parent.h>
 
drivers/gpu/drm/nouveau/nouveau_gem.c +0 -26
index 45ca4eb98f54..a48f42aaeab9 100644
@@ -613,32 +613,6 @@ nouveau_gem_pushbuf_validate(struct nouveau_channel *chan,
 	return 0;
 }
 
-static inline void
-u_free(void *addr)
-{
-	kvfree(addr);
-}
-
-static inline void *
-u_memcpya(uint64_t user, unsigned nmemb, unsigned size)
-{
-	void *mem;
-	void __user *userptr = (void __force __user *)(uintptr_t)user;
-
-	size *= nmemb;
-
-	mem = kvmalloc(size, GFP_KERNEL);
-	if (!mem)
-		return ERR_PTR(-ENOMEM);
-
-	if (copy_from_user(mem, userptr, size)) {
-		u_free(mem);
-		return ERR_PTR(-EFAULT);
-	}
-
-	return mem;
-}
-
 static int
 nouveau_gem_pushbuf_reloc_apply(struct nouveau_cli *cli,
 				struct drm_nouveau_gem_pushbuf *req,
Signature · 1 line
2.40.1
#25 lore raw

PATCH v5 9/14 drm/nouveau: fence: separate fence alloc and emit

In reply to Danilo Krummrich

The new (VM_BIND) UAPI exports DMA fences through DRM syncobjs. Hence, in order to emit fences within DMA fence signalling critical sections (e.g. as typically done in the DRM GPU schedulers run_job() callback) we need to separate fence allocation and fence emitting.

---
 drivers/gpu/drm/nouveau/dispnv04/crtc.c |  9 ++++-
 drivers/gpu/drm/nouveau/nouveau_bo.c    | 52 +++++++++++++++----------
 drivers/gpu/drm/nouveau/nouveau_chan.c  |  6 ++-
 drivers/gpu/drm/nouveau/nouveau_dmem.c  |  9 +++--
 drivers/gpu/drm/nouveau/nouveau_fence.c | 16 +++-----
 drivers/gpu/drm/nouveau/nouveau_fence.h |  3 +-
 drivers/gpu/drm/nouveau/nouveau_gem.c   |  5 ++-
 7 files changed, 59 insertions(+), 41 deletions(-)
drivers/gpu/drm/nouveau/dispnv04/crtc.c +8 -1
index a6f2e681bde9..a34924523133 100644
@@ -1122,11 +1122,18 @@ nv04_page_flip_emit(struct nouveau_channel *chan,
 	PUSH_NVSQ(push, NV_SW, NV_SW_PAGE_FLIP, 0x00000000);
 	PUSH_KICK(push);
 
-	ret = nouveau_fence_new(chan, false, pfence);
+	ret = nouveau_fence_new(pfence);
 	if (ret)
 		goto fail;
 
+	ret = nouveau_fence_emit(*pfence, chan);
+	if (ret)
+		goto fail_fence_unref;
+
 	return 0;
+
+fail_fence_unref:
+	nouveau_fence_unref(pfence);
 fail:
 	spin_lock_irqsave(&dev->event_lock, flags);
 	list_del(&s->head);
drivers/gpu/drm/nouveau/nouveau_bo.c +31 -21

60 lines not on this page · raw message

drivers/gpu/drm/nouveau/nouveau_chan.c +4 -2
index 1068abe41024..f47c0363683c 100644
@@ -62,9 +62,11 @@ nouveau_channel_idle(struct nouveau_channel *chan)
 		struct nouveau_fence *fence = NULL;
 		int ret;
 
-		ret = nouveau_fence_new(chan, false, &fence);
+		ret = nouveau_fence_new(&fence);
 		if (!ret) {
-			ret = nouveau_fence_wait(fence, false, false);
+			ret = nouveau_fence_emit(fence, chan);
+			if (!ret)
+				ret = nouveau_fence_wait(fence, false, false);
 			nouveau_fence_unref(&fence);
 		}
 
drivers/gpu/drm/nouveau/nouveau_dmem.c +6 -3
index 789857faa048..4ad40e42cae1 100644
@@ -209,7 +209,8 @@ static vm_fault_t nouveau_dmem_migrate_to_ram(struct vm_fault *vmf)
 		goto done;
 	}
 
-	nouveau_fence_new(dmem->migrate.chan, false, &fence);
+	if (!nouveau_fence_new(&fence))
+		nouveau_fence_emit(fence, dmem->migrate.chan);
 	migrate_vma_pages(&args);
 	nouveau_dmem_fence_done(&fence);
 	dma_unmap_page(drm->dev->dev, dma_addr, PAGE_SIZE, DMA_BIDIRECTIONAL);
@@ -402,7 +403,8 @@ nouveau_dmem_evict_chunk(struct nouveau_dmem_chunk *chunk)
 		}
 	}
 
-	nouveau_fence_new(chunk->drm->dmem->migrate.chan, false, &fence);
+	if (!nouveau_fence_new(&fence))
+		nouveau_fence_emit(fence, chunk->drm->dmem->migrate.chan);
 	migrate_device_pages(src_pfns, dst_pfns, npages);
 	nouveau_dmem_fence_done(&fence);
 	migrate_device_finalize(src_pfns, dst_pfns, npages);
@@ -675,7 +677,8 @@ static void nouveau_dmem_migrate_chunk(struct nouveau_drm *drm,
 		addr += PAGE_SIZE;
 	}
 
-	nouveau_fence_new(drm->dmem->migrate.chan, false, &fence);
+	if (!nouveau_fence_new(&fence))
+		nouveau_fence_emit(fence, chunk->drm->dmem->migrate.chan);
 	migrate_vma_pages(args);
 	nouveau_dmem_fence_done(&fence);
 	nouveau_pfns_map(svmm, args->vma->vm_mm, args->start, pfns, i);
drivers/gpu/drm/nouveau/nouveau_fence.c +5 -11
index ee5e9d40c166..e946408f945b 100644
@@ -210,6 +210,9 @@ nouveau_fence_emit(struct nouveau_fence *fence, struct nouveau_channel *chan)
 	struct nouveau_fence_priv *priv = (void*)chan->drm->fence;
 	int ret;
 
+	if (unlikely(!chan->fence))
+		return -ENODEV;
+
 	fence->channel  = chan;
 	fence->timeout  = jiffies + (15 * HZ);
 
@@ -396,25 +399,16 @@ nouveau_fence_unref(struct nouveau_fence **pfence)
 }
 
 int
-nouveau_fence_new(struct nouveau_channel *chan, bool sysmem,
-		  struct nouveau_fence **pfence)
+nouveau_fence_new(struct nouveau_fence **pfence)
 {
 	struct nouveau_fence *fence;
-	int ret = 0;
-
-	if (unlikely(!chan->fence))
-		return -ENODEV;
 
 	fence = kzalloc(sizeof(*fence), GFP_KERNEL);
 	if (!fence)
 		return -ENOMEM;
 
-	ret = nouveau_fence_emit(fence, chan);
-	if (ret)
-		nouveau_fence_unref(&fence);
-
 	*pfence = fence;
-	return ret;
+	return 0;
 }
 
 static const char *nouveau_fence_get_get_driver_name(struct dma_fence *fence)
drivers/gpu/drm/nouveau/nouveau_fence.h +1 -2
index 0ca2bc85adf6..7c73c7c9834a 100644
@@ -17,8 +17,7 @@ struct nouveau_fence {
 	unsigned long timeout;
 };
 
-int  nouveau_fence_new(struct nouveau_channel *, bool sysmem,
-		       struct nouveau_fence **);
+int  nouveau_fence_new(struct nouveau_fence **);
 void nouveau_fence_unref(struct nouveau_fence **);
 
 int  nouveau_fence_emit(struct nouveau_fence *, struct nouveau_channel *);
drivers/gpu/drm/nouveau/nouveau_gem.c +4 -1
index a48f42aaeab9..9c8d1b911a01 100644
@@ -873,8 +873,11 @@ nouveau_gem_ioctl_pushbuf(struct drm_device *dev, void *data,
 		}
 	}
 
-	ret = nouveau_fence_new(chan, false, &fence);
+	ret = nouveau_fence_new(&fence);
+	if (!ret)
+		ret = nouveau_fence_emit(fence, chan);
 	if (ret) {
+		nouveau_fence_unref(&fence);
 		NV_PRINTK(err, cli, "error fencing pushbuf: %d\n", ret);
 		WIND_RING(chan);
 		goto out;
Signature · 1 line
2.40.1
#26 lore raw

Re: [PATCH drm-next v5 09/14] drm/nouveau: fence: separate fence alloc and emit

kernel test robot noticed the following build errors:

Hi Danilo,

kernel test robot noticed the following build errors:

[auto build test ERROR on 2222dcb0775d36de28992f56455ab3967b30d380]

url:    https://github.com/intel-lab-lkp/linux/commits/Danilo-Krummrich/drm-execution-context-for-GEM-buffers-v4/20230620-084448
base:   2222dcb0775d36de28992f56455ab3967b30d380
patch link:    https://lore.kernel.org/r/20230620004217.4700-10-dakr%40redhat.com
patch subject: [PATCH drm-next v5 09/14] drm/nouveau: fence: separate fence alloc and emit
config: arm64-allyesconfig (https://download.01.org/0day-ci/archive/20230621/[email protected]/config)
compiler: aarch64-linux-gcc (GCC) 12.3.0
reproduce: (https://download.01.org/0day-ci/archive/20230621/[email protected]/reproduce)

If you fix the issue in a separate patch/commit (i.e. not just a new version of the same patch/commit), kindly add following tags
| Reported-by: kernel test robot <[email protected]>
| Closes: https://lore.kernel.org/oe-kbuild-all/[email protected]/

All errors (new ones prefixed by >>):

drivers/gpu/drm/nouveau/nouveau_dmem.c: In function 'nouveau_dmem_migrate_chunk':

>> drivers/gpu/drm/nouveau/nouveau_dmem.c:681:43: error: 'chunk' undeclared (first use in this function)
     681 |                 nouveau_fence_emit(fence, chunk->drm->dmem->migrate.chan);
         |                                           ^~~~~

drivers/gpu/drm/nouveau/nouveau_dmem.c:681:43: note: each undeclared identifier is reported only once for each function it appears in

vim +/chunk +681 drivers/gpu/drm/nouveau/nouveau_dmem.c

664
665 static void nouveau_dmem_migrate_chunk(struct nouveau_drm *drm,
666 struct nouveau_svmm *svmm, struct migrate_vma *args,
667 dma_addr_t *dma_addrs, u64 *pfns)
668 {
669 struct nouveau_fence *fence;
670 unsigned long addr = args->start, nr_dma = 0, i;
671
672 for (i = 0; addr < args->end; i++) {
673 args->dst[i] = nouveau_dmem_migrate_copy_one(drm, svmm,
674 args->src[i], dma_addrs + nr_dma, pfns + i);
675 if (!dma_mapping_error(drm->dev->dev, dma_addrs[nr_dma]))
676 nr_dma++;
677 addr += PAGE_SIZE;
678 }
679
680 if (!nouveau_fence_new(&fence))

 > 681			nouveau_fence_emit(fence, chunk->drm->dmem->migrate.chan);

682 migrate_vma_pages(args);
683 nouveau_dmem_fence_done(&fence);
684 nouveau_pfns_map(svmm, args->vma->vm_mm, args->start, pfns, i);
685
686 while (nr_dma--) {
687 dma_unmap_page(drm->dev->dev, dma_addrs[nr_dma], PAGE_SIZE,
688 DMA_BIDIRECTIONAL);
689 }
690 migrate_vma_finalize(args);
691 }
692

Signature · 2 lines
0-DAY CI Kernel Test Service
https://github.com/intel/lkp-tests/wiki
#27 lore raw

PATCH v5 10/14 drm/nouveau: fence: fail to emit when fence context is killed

In reply to Danilo Krummrich

The new VM_BIND UAPI implementation introduced in subsequent commits will allow asynchronous jobs processing push buffers and emitting fences.

If a fence context is killed, e.g. due to a channel fault, jobs which are already queued for execution might still emit new fences. In such a case a job would hang forever.

To fix that, fail to emit a new fence on a killed fence context with -ENODEV to unblock the job.

---
 drivers/gpu/drm/nouveau/nouveau_fence.c | 7 +++++++
 drivers/gpu/drm/nouveau/nouveau_fence.h | 2 +-
 2 files changed, 8 insertions(+), 1 deletion(-)
drivers/gpu/drm/nouveau/nouveau_fence.c +7 -0
index e946408f945b..77c739a55b19 100644
@@ -96,6 +96,7 @@ nouveau_fence_context_kill(struct nouveau_fence_chan *fctx, int error)
 		if (nouveau_fence_signal(fence))
 			nvif_event_block(&fctx->event);
 	}
+	fctx->killed = 1;
 	spin_unlock_irqrestore(&fctx->lock, flags);
 }
 
@@ -229,6 +230,12 @@ nouveau_fence_emit(struct nouveau_fence *fence, struct nouveau_channel *chan)
 		dma_fence_get(&fence->base);
 		spin_lock_irq(&fctx->lock);
 
+		if (unlikely(fctx->killed)) {
+			spin_unlock_irq(&fctx->lock);
+			dma_fence_put(&fence->base);
+			return -ENODEV;
+		}
+
 		if (nouveau_fence_update(chan, fctx))
 			nvif_event_block(&fctx->event);
 
drivers/gpu/drm/nouveau/nouveau_fence.h +1 -1
index 7c73c7c9834a..2c72d96ef17d 100644
@@ -44,7 +44,7 @@ struct nouveau_fence_chan {
 	char name[32];
 
 	struct nvif_event event;
-	int notify_ref, dead;
+	int notify_ref, dead, killed;
 };
 
 struct nouveau_fence_priv {
Signature · 1 line
2.40.1
#28 lore raw

PATCH v5 11/14 drm/nouveau: chan: provide nouveau_channel_kill()

In reply to Danilo Krummrich

The new VM_BIND UAPI implementation introduced in subsequent commits will allow asynchronous jobs processing push buffers and emitting fences.

If a job times out, we need a way to recover from this situation. For now, simply kill the channel to unblock all hung up jobs and signal userspace that the device is dead on the next EXEC or VM_BIND ioctl.

---
 drivers/gpu/drm/nouveau/nouveau_chan.c | 14 +++++++++++---
 drivers/gpu/drm/nouveau/nouveau_chan.h |  1 +
 2 files changed, 12 insertions(+), 3 deletions(-)
drivers/gpu/drm/nouveau/nouveau_chan.c +11 -3
index f47c0363683c..a975f8b0e0e5 100644
@@ -40,6 +40,14 @@ MODULE_PARM_DESC(vram_pushbuf, "Create DMA push buffers in VRAM");
 int nouveau_vram_pushbuf;
 module_param_named(vram_pushbuf, nouveau_vram_pushbuf, int, 0400);
 
+void
+nouveau_channel_kill(struct nouveau_channel *chan)
+{
+	atomic_set(&chan->killed, 1);
+	if (chan->fence)
+		nouveau_fence_context_kill(chan->fence, -ENODEV);
+}
+
 static int
 nouveau_channel_killed(struct nvif_event *event, void *repv, u32 repc)
 {
@@ -47,9 +55,9 @@ nouveau_channel_killed(struct nvif_event *event, void *repv, u32 repc)
 	struct nouveau_cli *cli = (void *)chan->user.client;
 
 	NV_PRINTK(warn, cli, "channel %d killed!\n", chan->chid);
-	atomic_set(&chan->killed, 1);
-	if (chan->fence)
-		nouveau_fence_context_kill(chan->fence, -ENODEV);
+
+	if (unlikely(!atomic_read(&chan->killed)))
+		nouveau_channel_kill(chan);
 
 	return NVIF_EVENT_DROP;
 }
drivers/gpu/drm/nouveau/nouveau_chan.h +1 -0
index e06a8ffed31a..e483f4a254da 100644
@@ -65,6 +65,7 @@ int  nouveau_channel_new(struct nouveau_drm *, struct nvif_device *, bool priv,
 			 u32 vram, u32 gart, struct nouveau_channel **);
 void nouveau_channel_del(struct nouveau_channel **);
 int  nouveau_channel_idle(struct nouveau_channel *);
+void nouveau_channel_kill(struct nouveau_channel *);
 
 extern int nouveau_vram_pushbuf;
 
Signature · 1 line
2.40.1
#29 lore raw

PATCH v5 12/14 drm/nouveau: nvkm/vmm: implement raw ops to manage uvmm

In reply to Danilo Krummrich

The new VM_BIND UAPI uses the DRM GPU VA manager to manage the VA space. Hence, we a need a way to manipulate the MMUs page tables without going through the internal range allocator implemented by nvkm/vmm.

This patch adds a raw interface for nvkm/vmm to pass the resposibility for managing the address space and the corresponding map/unmap/sparse operations to the upper layers.

---
 drivers/gpu/drm/nouveau/include/nvif/if000c.h |  26 ++-
 drivers/gpu/drm/nouveau/include/nvif/vmm.h    |  19 +-
 .../gpu/drm/nouveau/include/nvkm/subdev/mmu.h |  20 +-
 drivers/gpu/drm/nouveau/nouveau_svm.c         |   2 +-
 drivers/gpu/drm/nouveau/nouveau_vmm.c         |   4 +-
 drivers/gpu/drm/nouveau/nvif/vmm.c            | 100 +++++++-
 .../gpu/drm/nouveau/nvkm/subdev/mmu/uvmm.c    | 213 ++++++++++++++++--
 drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c | 197 ++++++++++++----
 drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h |  25 ++
 .../drm/nouveau/nvkm/subdev/mmu/vmmgf100.c    |  16 +-
 .../drm/nouveau/nvkm/subdev/mmu/vmmgp100.c    |  16 +-
 .../gpu/drm/nouveau/nvkm/subdev/mmu/vmmnv50.c |  27 ++-
 12 files changed, 566 insertions(+), 99 deletions(-)
drivers/gpu/drm/nouveau/include/nvif/if000c.h +25 -1
index 9c7ff56831c5..a5a182b3c28d 100644
@@ -3,7 +3,10 @@
 struct nvif_vmm_v0 {
 	__u8  version;
 	__u8  page_nr;
-	__u8  managed;
+#define NVIF_VMM_V0_TYPE_UNMANAGED                                         0x00
+#define NVIF_VMM_V0_TYPE_MANAGED                                           0x01
+#define NVIF_VMM_V0_TYPE_RAW                                               0x02
+	__u8  type;
 	__u8  pad03[5];
 	__u64 addr;
 	__u64 size;
@@ -17,6 +20,7 @@ struct nvif_vmm_v0 {
 #define NVIF_VMM_V0_UNMAP                                                  0x04
 #define NVIF_VMM_V0_PFNMAP                                                 0x05
 #define NVIF_VMM_V0_PFNCLR                                                 0x06
+#define NVIF_VMM_V0_RAW                                                    0x07
 #define NVIF_VMM_V0_MTHD(i)                                         ((i) + 0x80)
 
 struct nvif_vmm_page_v0 {
@@ -66,6 +70,26 @@ struct nvif_vmm_unmap_v0 {
 	__u64 addr;
 };
 
+struct nvif_vmm_raw_v0 {
+	__u8 version;
+#define NVIF_VMM_RAW_V0_GET	0x0
+#define NVIF_VMM_RAW_V0_PUT	0x1
+#define NVIF_VMM_RAW_V0_MAP	0x2
+#define NVIF_VMM_RAW_V0_UNMAP	0x3
+#define NVIF_VMM_RAW_V0_SPARSE	0x4
+	__u8  op;
+	__u8  sparse;
+	__u8  ref;
+	__u8  shift;
+	__u32 argc;
+	__u8  pad01[7];
+	__u64 addr;
+	__u64 size;
+	__u64 offset;
+	__u64 memory;
+	__u64 argv;
+};
+
 struct nvif_vmm_pfnmap_v0 {
 	__u8  version;
 	__u8  page;
drivers/gpu/drm/nouveau/include/nvif/vmm.h +17 -2
index a2ee92201ace..0ecedd0ee0a5 100644
@@ -4,6 +4,12 @@
 struct nvif_mem;
 struct nvif_mmu;
 
+enum nvif_vmm_type {
+	UNMANAGED,
+	MANAGED,
+	RAW,
+};
+
 enum nvif_vmm_get {
 	ADDR,
 	PTES,
@@ -30,8 +36,9 @@ struct nvif_vmm {
 	int page_nr;
 };
 
-int nvif_vmm_ctor(struct nvif_mmu *, const char *name, s32 oclass, bool managed,
-		  u64 addr, u64 size, void *argv, u32 argc, struct nvif_vmm *);
+int nvif_vmm_ctor(struct nvif_mmu *, const char *name, s32 oclass,
+		  enum nvif_vmm_type, u64 addr, u64 size, void *argv, u32 argc,
+		  struct nvif_vmm *);
 void nvif_vmm_dtor(struct nvif_vmm *);
 int nvif_vmm_get(struct nvif_vmm *, enum nvif_vmm_get, bool sparse,
 		 u8 page, u8 align, u64 size, struct nvif_vma *);
@@ -39,4 +46,12 @@ void nvif_vmm_put(struct nvif_vmm *, struct nvif_vma *);
 int nvif_vmm_map(struct nvif_vmm *, u64 addr, u64 size, void *argv, u32 argc,
 		 struct nvif_mem *, u64 offset);
 int nvif_vmm_unmap(struct nvif_vmm *, u64);
+
+int nvif_vmm_raw_get(struct nvif_vmm *vmm, u64 addr, u64 size, u8 shift);
+int nvif_vmm_raw_put(struct nvif_vmm *vmm, u64 addr, u64 size, u8 shift);
+int nvif_vmm_raw_map(struct nvif_vmm *vmm, u64 addr, u64 size, u8 shift,
+		     void *argv, u32 argc, struct nvif_mem *mem, u64 offset);
+int nvif_vmm_raw_unmap(struct nvif_vmm *vmm, u64 addr, u64 size,
+		       u8 shift, bool sparse);
+int nvif_vmm_raw_sparse(struct nvif_vmm *vmm, u64 addr, u64 size, bool ref);
 #endif
drivers/gpu/drm/nouveau/include/nvkm/subdev/mmu.h +19 -1
index 70e7887ef4b4..2fd2f2433fc7 100644
@@ -17,6 +17,7 @@ struct nvkm_vma {
 	bool part:1; /* Region was split from an allocated region by map(). */
 	bool busy:1; /* Region busy (for temporarily preventing user access). */
 	bool mapped:1; /* Region contains valid pages. */
+	bool no_comp:1; /* Force no memory compression. */
 	struct nvkm_memory *memory; /* Memory currently mapped into VMA. */
 	struct nvkm_tags *tags; /* Compression tag reference. */
 };
@@ -27,10 +28,26 @@ struct nvkm_vmm {
 	const char *name;
 	u32 debug;
 	struct kref kref;
-	struct mutex mutex;
+
+	struct {
+		struct mutex vmm;
+		struct mutex ref;
+		struct mutex map;
+	} mutex;
 
 	u64 start;
 	u64 limit;
+	struct {
+		struct {
+			u64 addr;
+			u64 size;
+		} p;
+		struct {
+			u64 addr;
+			u64 size;
+		} n;
+		bool raw;
+	} managed;
 
 	struct nvkm_vmm_pt *pd;
 	struct list_head join;
@@ -70,6 +87,7 @@ struct nvkm_vmm_map {
 
 	const struct nvkm_vmm_page *page;
 
+	bool no_comp;
 	struct nvkm_tags *tags;
 	u64 next;
 	u64 type;
drivers/gpu/drm/nouveau/nouveau_svm.c +1 -1
index a74ba8d84ba7..186351ecf72f 100644
@@ -350,7 +350,7 @@ nouveau_svmm_init(struct drm_device *dev, void *data,
 	 * VMM instead of the standard one.
 	 */
 	ret = nvif_vmm_ctor(&cli->mmu, "svmVmm",
-			    cli->vmm.vmm.object.oclass, true,
+			    cli->vmm.vmm.object.oclass, MANAGED,
 			    args->unmanaged_addr, args->unmanaged_size,
 			    &(struct gp100_vmm_v0) {
 				.fault_replay = true,
drivers/gpu/drm/nouveau/nouveau_vmm.c +2 -2
index 67d6619fcd5e..a6602c012671 100644
@@ -128,8 +128,8 @@ nouveau_vmm_fini(struct nouveau_vmm *vmm)
 int
 nouveau_vmm_init(struct nouveau_cli *cli, s32 oclass, struct nouveau_vmm *vmm)
 {
-	int ret = nvif_vmm_ctor(&cli->mmu, "drmVmm", oclass, false, PAGE_SIZE,
-				0, NULL, 0, &vmm->vmm);
+	int ret = nvif_vmm_ctor(&cli->mmu, "drmVmm", oclass, UNMANAGED,
+				PAGE_SIZE, 0, NULL, 0, &vmm->vmm);
 	if (ret)
 		return ret;
 
drivers/gpu/drm/nouveau/nvif/vmm.c +97 -3

120 lines not on this page · raw message

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/uvmm.c +195 -18

324 lines not on this page · raw message

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c +152 -45

306 lines not on this page · raw message

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h +25 -0
index f6188aa9171c..f9bc30cdb2b3 100644
@@ -163,6 +163,7 @@ int nvkm_vmm_new_(const struct nvkm_vmm_func *, struct nvkm_mmu *,
 		  u32 pd_header, bool managed, u64 addr, u64 size,
 		  struct lock_class_key *, const char *name,
 		  struct nvkm_vmm **);
+struct nvkm_vma *nvkm_vma_new(u64 addr, u64 size);
 struct nvkm_vma *nvkm_vmm_node_search(struct nvkm_vmm *, u64 addr);
 struct nvkm_vma *nvkm_vmm_node_split(struct nvkm_vmm *, struct nvkm_vma *,
 				     u64 addr, u64 size);
@@ -173,6 +174,30 @@ void nvkm_vmm_put_locked(struct nvkm_vmm *, struct nvkm_vma *);
 void nvkm_vmm_unmap_locked(struct nvkm_vmm *, struct nvkm_vma *, bool pfn);
 void nvkm_vmm_unmap_region(struct nvkm_vmm *, struct nvkm_vma *);
 
+int nvkm_vmm_raw_get(struct nvkm_vmm *vmm, u64 addr, u64 size, u8 refd);
+void nvkm_vmm_raw_put(struct nvkm_vmm *vmm, u64 addr, u64 size, u8 refd);
+void nvkm_vmm_raw_unmap(struct nvkm_vmm *vmm, u64 addr, u64 size,
+			bool sparse, u8 refd);
+int nvkm_vmm_raw_sparse(struct nvkm_vmm *, u64 addr, u64 size, bool ref);
+
+static inline bool
+nvkm_vmm_in_managed_range(struct nvkm_vmm *vmm, u64 start, u64 size)
+{
+	u64 p_start = vmm->managed.p.addr;
+	u64 p_end = p_start + vmm->managed.p.size;
+	u64 n_start = vmm->managed.n.addr;
+	u64 n_end = n_start + vmm->managed.n.size;
+	u64 end = start + size;
+
+	if (start >= p_start && end <= p_end)
+		return true;
+
+	if (start >= n_start && end <= n_end)
+		return true;
+
+	return false;
+}
+
 #define NVKM_VMM_PFN_ADDR                                 0xfffffffffffff000ULL
 #define NVKM_VMM_PFN_ADDR_SHIFT                                              12
 #define NVKM_VMM_PFN_APER                                 0x00000000000000f0ULL
drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmgf100.c +9 -7
index 5438384d9a67..5e857c02e9aa 100644
@@ -287,15 +287,17 @@ gf100_vmm_valid(struct nvkm_vmm *vmm, void *argv, u32 argc,
 			return -EINVAL;
 		}
 
-		ret = nvkm_memory_tags_get(memory, device, tags,
-					   nvkm_ltc_tags_clear,
-					   &map->tags);
-		if (ret) {
-			VMM_DEBUG(vmm, "comp %d", ret);
-			return ret;
+		if (!map->no_comp) {
+			ret = nvkm_memory_tags_get(memory, device, tags,
+						   nvkm_ltc_tags_clear,
+						   &map->tags);
+			if (ret) {
+				VMM_DEBUG(vmm, "comp %d", ret);
+				return ret;
+			}
 		}
 
-		if (map->tags->mn) {
+		if (!map->no_comp && map->tags->mn) {
 			u64 tags = map->tags->mn->offset + (map->offset >> 17);
 			if (page->shift == 17 || !gm20x) {
 				map->type |= tags << 44;
drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmgp100.c +9 -7
index 17899fc95b2d..f3630d0e0d55 100644
@@ -453,15 +453,17 @@ gp100_vmm_valid(struct nvkm_vmm *vmm, void *argv, u32 argc,
 			return -EINVAL;
 		}
 
-		ret = nvkm_memory_tags_get(memory, device, tags,
-					   nvkm_ltc_tags_clear,
-					   &map->tags);
-		if (ret) {
-			VMM_DEBUG(vmm, "comp %d", ret);
-			return ret;
+		if (!map->no_comp) {
+			ret = nvkm_memory_tags_get(memory, device, tags,
+						   nvkm_ltc_tags_clear,
+						   &map->tags);
+			if (ret) {
+				VMM_DEBUG(vmm, "comp %d", ret);
+				return ret;
+			}
 		}
 
-		if (map->tags->mn) {
+		if (!map->no_comp && map->tags->mn) {
 			tags = map->tags->mn->offset + (map->offset >> 16);
 			map->ctag |= ((1ULL << page->shift) >> 16) << 36;
 			map->type |= tags << 36;
drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmnv50.c +15 -12
index b7548dcd72c7..ff08ad5005a9 100644
@@ -296,19 +296,22 @@ nv50_vmm_valid(struct nvkm_vmm *vmm, void *argv, u32 argc,
 			return -EINVAL;
 		}
 
-		ret = nvkm_memory_tags_get(memory, device, tags, NULL,
-					   &map->tags);
-		if (ret) {
-			VMM_DEBUG(vmm, "comp %d", ret);
-			return ret;
-		}
+		if (!map->no_comp) {
+			ret = nvkm_memory_tags_get(memory, device, tags, NULL,
+						   &map->tags);
+			if (ret) {
+				VMM_DEBUG(vmm, "comp %d", ret);
+				return ret;
+			}
 
-		if (map->tags->mn) {
-			u32 tags = map->tags->mn->offset + (map->offset >> 16);
-			map->ctag |= (u64)comp << 49;
-			map->type |= (u64)comp << 47;
-			map->type |= (u64)tags << 49;
-			map->next |= map->ctag;
+			if (map->tags->mn) {
+				u32 tags = map->tags->mn->offset +
+					   (map->offset >> 16);
+				map->ctag |= (u64)comp << 49;
+				map->type |= (u64)comp << 47;
+				map->type |= (u64)tags << 49;
+				map->next |= map->ctag;
+			}
 		}
 	}
 
Signature · 1 line
2.40.1
#30 lore raw

PATCH v5 13/14 drm/nouveau: implement new VM_BIND uAPI

In reply to Danilo Krummrich

This commit provides the implementation for the new uapi motivated by the Vulkan API. It allows user mode drivers (UMDs) to:

1) Initialize a GPU virtual address (VA) space via the new DRM_IOCTL_NOUVEAU_VM_INIT ioctl for UMDs to specify the portion of VA space managed by the kernel and userspace, respectively.

2) Allocate and free a VA space region as well as bind and unbind memory to the GPUs VA space via the new DRM_IOCTL_NOUVEAU_VM_BIND ioctl. UMDs can request the named operations to be processed either synchronously or asynchronously. It supports DRM syncobjs (incl. timelines) as synchronization mechanism. The management of the GPU VA mappings is implemented with the DRM GPU VA manager.

3) Execute push buffers with the new DRM_IOCTL_NOUVEAU_EXEC ioctl. The execution happens asynchronously. It supports DRM syncobj (incl. timelines) as synchronization mechanism. DRM GEM object locking is handled with drm_exec.

Both, DRM_IOCTL_NOUVEAU_VM_BIND and DRM_IOCTL_NOUVEAU_EXEC, use the DRM GPU scheduler for the asynchronous paths.

---
 Documentation/gpu/driver-uapi.rst       |    3 +
 drivers/gpu/drm/nouveau/Kbuild          |    3 +
 drivers/gpu/drm/nouveau/Kconfig         |    2 +
 drivers/gpu/drm/nouveau/nouveau_abi16.c |   24 +
 drivers/gpu/drm/nouveau/nouveau_abi16.h |    1 +
 drivers/gpu/drm/nouveau/nouveau_bo.c    |  147 +-
 drivers/gpu/drm/nouveau/nouveau_bo.h    |    2 +-
 drivers/gpu/drm/nouveau/nouveau_drm.c   |   27 +-
 drivers/gpu/drm/nouveau/nouveau_drv.h   |   59 +-
 drivers/gpu/drm/nouveau/nouveau_exec.c  |  418 +++++
 drivers/gpu/drm/nouveau/nouveau_exec.h  |   54 +
 drivers/gpu/drm/nouveau/nouveau_gem.c   |   25 +-
 drivers/gpu/drm/nouveau/nouveau_mem.h   |    5 +
 drivers/gpu/drm/nouveau/nouveau_prime.c |    2 +-
 drivers/gpu/drm/nouveau/nouveau_sched.c |  461 ++++++
 drivers/gpu/drm/nouveau/nouveau_sched.h |  123 ++
 drivers/gpu/drm/nouveau/nouveau_uvmm.c  | 1979 +++++++++++++++++++++++
 drivers/gpu/drm/nouveau/nouveau_uvmm.h  |  107 ++
 18 files changed, 3377 insertions(+), 65 deletions(-)
 create mode 100644 drivers/gpu/drm/nouveau/nouveau_exec.c
 create mode 100644 drivers/gpu/drm/nouveau/nouveau_exec.h
 create mode 100644 drivers/gpu/drm/nouveau/nouveau_sched.c
 create mode 100644 drivers/gpu/drm/nouveau/nouveau_sched.h
 create mode 100644 drivers/gpu/drm/nouveau/nouveau_uvmm.c
 create mode 100644 drivers/gpu/drm/nouveau/nouveau_uvmm.h
Documentation/gpu/driver-uapi.rst +3 -0
index 9c7ca6e33a68..c08bcbb95fb3 100644
@@ -13,4 +13,7 @@ drm/nouveau uAPI
 VM_BIND / EXEC uAPI
 -------------------
 
+.. kernel-doc:: drivers/gpu/drm/nouveau/nouveau_exec.c
+    :doc: Overview
+
 .. kernel-doc:: include/uapi/drm/nouveau_drm.h
drivers/gpu/drm/nouveau/Kbuild +3 -0
index 5e5617006da5..cf6b3a80c0c8 100644
@@ -47,6 +47,9 @@ nouveau-y += nouveau_prime.o
 nouveau-y += nouveau_sgdma.o
 nouveau-y += nouveau_ttm.o
 nouveau-y += nouveau_vmm.o
+nouveau-y += nouveau_exec.o
+nouveau-y += nouveau_sched.o
+nouveau-y += nouveau_uvmm.o
 
 # DRM - modesetting
 nouveau-$(CONFIG_DRM_NOUVEAU_BACKLIGHT) += nouveau_backlight.o
drivers/gpu/drm/nouveau/Kconfig +2 -0
index a70bd65e1400..c52e8096cca4 100644
@@ -10,6 +10,8 @@ config DRM_NOUVEAU
 	select DRM_KMS_HELPER
 	select DRM_TTM
 	select DRM_TTM_HELPER
+	select DRM_EXEC
+	select DRM_SCHED
 	select I2C
 	select I2C_ALGOBIT
 	select BACKLIGHT_CLASS_DEVICE if DRM_NOUVEAU_BACKLIGHT
drivers/gpu/drm/nouveau/nouveau_abi16.c +24 -0
index 82dab51d8aeb..a112f28681d3 100644
@@ -35,6 +35,7 @@
 #include "nouveau_chan.h"
 #include "nouveau_abi16.h"
 #include "nouveau_vmm.h"
+#include "nouveau_sched.h"
 
 static struct nouveau_abi16 *
 nouveau_abi16(struct drm_file *file_priv)
@@ -125,6 +126,17 @@ nouveau_abi16_chan_fini(struct nouveau_abi16 *abi16,
 {
 	struct nouveau_abi16_ntfy *ntfy, *temp;
 
+	/* When a client exits without waiting for it's queued up jobs to
+	 * finish it might happen that we fault the channel. This is due to
+	 * drm_file_free() calling drm_gem_release() before the postclose()
+	 * callback. Hence, we can't tear down this scheduler entity before
+	 * uvmm mappings are unmapped. Currently, we can't detect this case.
+	 *
+	 * However, this should be rare and harmless, since the channel isn't
+	 * needed anymore.
+	 */
+	nouveau_sched_entity_fini(&chan->sched_entity);
+
 	/* wait for all activity to stop before cleaning up */
 	if (chan->chan)
 		nouveau_channel_idle(chan->chan);
@@ -261,6 +273,13 @@ nouveau_abi16_ioctl_channel_alloc(ABI16_IOCTL_ARGS)
 	if (!drm->channel)
 		return nouveau_abi16_put(abi16, -ENODEV);
 
+	/* If uvmm wasn't initialized until now disable it completely to prevent
+	 * userspace from mixing up UAPIs.
+	 *
+	 * The client lock is already acquired by nouveau_abi16_get().
+	 */
+	__nouveau_cli_uvmm_disable(cli);
+
 	device = &abi16->device;
 	engine = NV_DEVICE_HOST_RUNLIST_ENGINES_GR;
 
@@ -304,6 +323,11 @@ nouveau_abi16_ioctl_channel_alloc(ABI16_IOCTL_ARGS)
 	if (ret)
 		goto done;
 
+	ret = nouveau_sched_entity_init(&chan->sched_entity, &drm->sched,
+					drm->sched_wq);
+	if (ret)
+		goto done;
+
 	init->channel = chan->chan->chid;
 
 	if (device->info.family >= NV_DEVICE_INFO_V0_TESLA)
drivers/gpu/drm/nouveau/nouveau_abi16.h +1 -0
index 27eae85f33e6..8209eb28feaf 100644
@@ -26,6 +26,7 @@ struct nouveau_abi16_chan {
 	struct nouveau_bo *ntfy;
 	struct nouveau_vma *ntfy_vma;
 	struct nvkm_mm  heap;
+	struct nouveau_sched_entity sched_entity;
 };
 
 struct nouveau_abi16 {
drivers/gpu/drm/nouveau/nouveau_bo.c +92 -55

186 lines not on this page · raw message

drivers/gpu/drm/nouveau/nouveau_bo.h +1 -1
index 774dd93ca76b..cb85207d9e8f 100644
@@ -73,7 +73,7 @@ extern struct ttm_device_funcs nouveau_bo_driver;
 
 void nouveau_bo_move_init(struct nouveau_drm *);
 struct nouveau_bo *nouveau_bo_alloc(struct nouveau_cli *, u64 *size, int *align,
-				    u32 domain, u32 tile_mode, u32 tile_flags);
+				    u32 domain, u32 tile_mode, u32 tile_flags, bool internal);
 int  nouveau_bo_init(struct nouveau_bo *, u64 size, int align, u32 domain,
 		     struct sg_table *sg, struct dma_resv *robj);
 int  nouveau_bo_new(struct nouveau_cli *, u64 size, int align, u32 domain,
drivers/gpu/drm/nouveau/nouveau_drm.c +25 -2
index 7aac9384600e..3b364a38c5be 100644
@@ -68,6 +68,9 @@
 #include "nouveau_platform.h"
 #include "nouveau_svm.h"
 #include "nouveau_dmem.h"
+#include "nouveau_exec.h"
+#include "nouveau_uvmm.h"
+#include "nouveau_sched.h"
 
 DECLARE_DYNDBG_CLASSMAP(drm_debug_classes, DD_CLASS_TYPE_DISJOINT_BITS, 0,
 			"DRM_UT_CORE",
@@ -196,6 +199,8 @@ nouveau_cli_fini(struct nouveau_cli *cli)
 	WARN_ON(!list_empty(&cli->worker));
 
 	usif_client_fini(cli);
+	nouveau_uvmm_fini(&cli->uvmm);
+	nouveau_sched_entity_fini(&cli->sched_entity);
 	nouveau_vmm_fini(&cli->svm);
 	nouveau_vmm_fini(&cli->vmm);
 	nvif_mmu_dtor(&cli->mmu);
@@ -301,6 +306,12 @@ nouveau_cli_init(struct nouveau_drm *drm, const char *sname,
 	}
 
 	cli->mem = &mems[ret];
+
+	ret = nouveau_sched_entity_init(&cli->sched_entity, &drm->sched,
+					drm->sched_wq);
+	if (ret)
+		goto done;
+
 	return 0;
 done:
 	if (ret)
@@ -554,10 +565,14 @@ nouveau_drm_device_init(struct drm_device *dev)
 	nvif_parent_ctor(&nouveau_parent, &drm->parent);
 	drm->master.base.object.parent = &drm->parent;
 
-	ret = nouveau_cli_init(drm, "DRM-master", &drm->master);
+	ret = nouveau_sched_init(drm);
 	if (ret)
 		goto fail_alloc;
 
+	ret = nouveau_cli_init(drm, "DRM-master", &drm->master);
+	if (ret)
+		goto fail_sched;
+
 	ret = nouveau_cli_init(drm, "DRM", &drm->client);
 	if (ret)
 		goto fail_master;
@@ -614,7 +629,6 @@ nouveau_drm_device_init(struct drm_device *dev)
 	}
 
 	return 0;
-
 fail_dispinit:
 	nouveau_display_destroy(dev);
 fail_dispctor:
@@ -627,6 +641,8 @@ nouveau_drm_device_init(struct drm_device *dev)
 	nouveau_cli_fini(&drm->client);
 fail_master:
 	nouveau_cli_fini(&drm->master);
+fail_sched:
+	nouveau_sched_fini(drm);
 fail_alloc:
 	nvif_parent_dtor(&drm->parent);
 	kfree(drm);
@@ -678,6 +694,8 @@ nouveau_drm_device_fini(struct drm_device *dev)
 	}
 	mutex_unlock(&drm->clients_lock);
 
+	nouveau_sched_fini(drm);
+
 	nouveau_cli_fini(&drm->client);
 	nouveau_cli_fini(&drm->master);
 	nvif_parent_dtor(&drm->parent);
@@ -1179,6 +1197,9 @@ nouveau_ioctls[] = {
 	DRM_IOCTL_DEF_DRV(NOUVEAU_GEM_CPU_PREP, nouveau_gem_ioctl_cpu_prep, DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(NOUVEAU_GEM_CPU_FINI, nouveau_gem_ioctl_cpu_fini, DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(NOUVEAU_GEM_INFO, nouveau_gem_ioctl_info, DRM_RENDER_ALLOW),
+	DRM_IOCTL_DEF_DRV(NOUVEAU_VM_INIT, nouveau_uvmm_ioctl_vm_init, DRM_RENDER_ALLOW),
+	DRM_IOCTL_DEF_DRV(NOUVEAU_VM_BIND, nouveau_uvmm_ioctl_vm_bind, DRM_RENDER_ALLOW),
+	DRM_IOCTL_DEF_DRV(NOUVEAU_EXEC, nouveau_exec_ioctl_exec, DRM_RENDER_ALLOW),
 };
 
 long
@@ -1226,6 +1247,8 @@ nouveau_driver_fops = {
 static struct drm_driver
 driver_stub = {
 	.driver_features = DRIVER_GEM |
+			   DRIVER_SYNCOBJ | DRIVER_SYNCOBJ_TIMELINE |
+			   DRIVER_GEM_GPUVA |
 			   DRIVER_MODESET |
 			   DRIVER_RENDER,
 	.open = nouveau_drm_open,
drivers/gpu/drm/nouveau/nouveau_drv.h +57 -2

99 lines not on this page · raw message

drivers/gpu/drm/nouveau/nouveau_exec.c new file +418 -0

418 lines not on this page · raw message

drivers/gpu/drm/nouveau/nouveau_exec.h new file +54 -0

54 lines not on this page · raw message

drivers/gpu/drm/nouveau/nouveau_gem.c +21 -4
index 9c8d1b911a01..3b0fbaedfb57 100644
@@ -120,7 +120,11 @@ nouveau_gem_object_open(struct drm_gem_object *gem, struct drm_file *file_priv)
 		goto out;
 	}
 
-	ret = nouveau_vma_new(nvbo, vmm, &vma);
+	/* only create a VMA on binding */
+	if (!nouveau_cli_uvmm(cli))
+		ret = nouveau_vma_new(nvbo, vmm, &vma);
+	else
+		ret = 0;
 	pm_runtime_mark_last_busy(dev);
 	pm_runtime_put_autosuspend(dev);
 out:
@@ -187,6 +191,9 @@ nouveau_gem_object_close(struct drm_gem_object *gem, struct drm_file *file_priv)
 	if (vmm->vmm.object.oclass < NVIF_CLASS_VMM_NV50)
 		return;
 
+	if (nouveau_cli_uvmm(cli))
+		return;
+
 	ret = ttm_bo_reserve(&nvbo->bo, false, false, NULL);
 	if (ret)
 		return;
@@ -231,7 +238,7 @@ nouveau_gem_new(struct nouveau_cli *cli, u64 size, int align, uint32_t domain,
 		domain |= NOUVEAU_GEM_DOMAIN_CPU;
 
 	nvbo = nouveau_bo_alloc(cli, &size, &align, domain, tile_mode,
-				tile_flags);
+				tile_flags, false);
 	if (IS_ERR(nvbo))
 		return PTR_ERR(nvbo);
 
@@ -279,13 +286,15 @@ nouveau_gem_info(struct drm_file *file_priv, struct drm_gem_object *gem,
 	else
 		rep->domain = NOUVEAU_GEM_DOMAIN_VRAM;
 	rep->offset = nvbo->offset;
-	if (vmm->vmm.object.oclass >= NVIF_CLASS_VMM_NV50) {
+	if (vmm->vmm.object.oclass >= NVIF_CLASS_VMM_NV50 &&
+	    !nouveau_cli_uvmm(cli)) {
 		vma = nouveau_vma_find(nvbo, vmm);
 		if (!vma)
 			return -EINVAL;
 
 		rep->offset = vma->addr;
-	}
+	} else
+		rep->offset = 0;
 
 	rep->size = nvbo->bo.base.size;
 	rep->map_handle = drm_vma_node_offset_addr(&nvbo->bo.base.vma_node);
@@ -310,6 +319,11 @@ nouveau_gem_ioctl_new(struct drm_device *dev, void *data,
 	struct nouveau_bo *nvbo = NULL;
 	int ret = 0;
 
+	/* If uvmm wasn't initialized until now disable it completely to prevent
+	 * userspace from mixing up UAPIs.
+	 */
+	nouveau_cli_uvmm_disable(cli);
+
 	ret = nouveau_gem_new(cli, req->info.size, req->align,
 			      req->info.domain, req->info.tile_mode,
 			      req->info.tile_flags, &nvbo);
@@ -721,6 +735,9 @@ nouveau_gem_ioctl_pushbuf(struct drm_device *dev, void *data,
 	if (unlikely(!abi16))
 		return -ENOMEM;
 
+	if (unlikely(nouveau_cli_uvmm(cli)))
+		return -ENOSYS;
+
 	list_for_each_entry(temp, &abi16->channels, head) {
 		if (temp->chan->chid == req->channel) {
 			chan = temp->chan;
drivers/gpu/drm/nouveau/nouveau_mem.h +5 -0
index 76c86d8bb01e..5365a3d3a17f 100644
@@ -35,4 +35,9 @@ int nouveau_mem_vram(struct ttm_resource *, bool contig, u8 page);
 int nouveau_mem_host(struct ttm_resource *, struct ttm_tt *);
 void nouveau_mem_fini(struct nouveau_mem *);
 int nouveau_mem_map(struct nouveau_mem *, struct nvif_vmm *, struct nvif_vma *);
+int
+nouveau_mem_map_fixed(struct nouveau_mem *mem,
+		      struct nvif_vmm *vmm,
+		      u8 kind, u64 addr,
+		      u64 offset, u64 range);
 #endif
drivers/gpu/drm/nouveau/nouveau_prime.c +1 -1
index f42c2b1b0363..6a883b9a799a 100644
@@ -50,7 +50,7 @@ struct drm_gem_object *nouveau_gem_prime_import_sg_table(struct drm_device *dev,
 
 	dma_resv_lock(robj, NULL);
 	nvbo = nouveau_bo_alloc(&drm->client, &size, &align,
-				NOUVEAU_GEM_DOMAIN_GART, 0, 0);
+				NOUVEAU_GEM_DOMAIN_GART, 0, 0, true);
 	if (IS_ERR(nvbo)) {
 		obj = ERR_CAST(nvbo);
 		goto unlock;
drivers/gpu/drm/nouveau/nouveau_sched.c new file +461 -0

461 lines not on this page · raw message

drivers/gpu/drm/nouveau/nouveau_sched.h new file +123 -0

123 lines not on this page · raw message

drivers/gpu/drm/nouveau/nouveau_uvmm.c new file +1979 -0

1,405 lines not on this page · raw message

drivers/gpu/drm/nouveau/nouveau_uvmm.h new file +107 -0

107 lines not on this page · raw message

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

Signature · 1 line
2.40.1

Part of the diff is not shown here. View the raw message.

#31 lore raw

PATCH v5 14/14 drm/nouveau: debugfs: implement DRM GPU VA debugfs

In reply to Danilo Krummrich

Provide the driver indirection iterating over all DRM GPU VA spaces to enable the common 'gpuvas' debugfs file for dumping DRM GPU VA spaces.

---
 drivers/gpu/drm/nouveau/nouveau_debugfs.c | 39 +++++++++++++++++++++++
 1 file changed, 39 insertions(+)
drivers/gpu/drm/nouveau/nouveau_debugfs.c +39 -0
index 99d022a91afc..053f703f2f68 100644
@@ -203,6 +203,44 @@ nouveau_debugfs_pstate_open(struct inode *inode, struct file *file)
 	return single_open(file, nouveau_debugfs_pstate_get, inode->i_private);
 }
 
+static void
+nouveau_debugfs_gpuva_regions(struct seq_file *m, struct nouveau_uvmm *uvmm)
+{
+	MA_STATE(mas, &uvmm->region_mt, 0, 0);
+	struct nouveau_uvma_region *reg;
+
+	seq_puts  (m, " VA regions  | start              | range              | end                \n");
+	seq_puts  (m, "----------------------------------------------------------------------------\n");
+	mas_for_each(&mas, reg, ULONG_MAX)
+		seq_printf(m, "             | 0x%016llx | 0x%016llx | 0x%016llx\n",
+			   reg->va.addr, reg->va.range, reg->va.addr + reg->va.range);
+}
+
+static int
+nouveau_debugfs_gpuva(struct seq_file *m, void *data)
+{
+	struct drm_info_node *node = (struct drm_info_node *) m->private;
+	struct nouveau_drm *drm = nouveau_drm(node->minor->dev);
+	struct nouveau_cli *cli;
+
+	mutex_lock(&drm->clients_lock);
+	list_for_each_entry(cli, &drm->clients, head) {
+		struct nouveau_uvmm *uvmm = nouveau_cli_uvmm(cli);
+
+		if (!uvmm)
+			continue;
+
+		nouveau_uvmm_lock(uvmm);
+		drm_debugfs_gpuva_info(m, &uvmm->umgr);
+		seq_puts(m, "\n");
+		nouveau_debugfs_gpuva_regions(m, uvmm);
+		nouveau_uvmm_unlock(uvmm);
+	}
+	mutex_unlock(&drm->clients_lock);
+
+	return 0;
+}
+
 static const struct file_operations nouveau_pstate_fops = {
 	.owner = THIS_MODULE,
 	.open = nouveau_debugfs_pstate_open,
@@ -214,6 +252,7 @@ static const struct file_operations nouveau_pstate_fops = {
 static struct drm_info_list nouveau_debugfs_list[] = {
 	{ "vbios.rom",  nouveau_debugfs_vbios_image, 0, NULL },
 	{ "strap_peek", nouveau_debugfs_strap_peek, 0, NULL },
+	DRM_DEBUGFS_GPUVA_INFO(nouveau_debugfs_gpuva, NULL),
 };
 #define NOUVEAU_DEBUGFS_ENTRIES ARRAY_SIZE(nouveau_debugfs_list)
 
Signature · 1 line
2.40.1
#32 lore raw
In reply to Danilo Krummrich

Hi Donald,

forgot to add your email address to the patch series - sorry about that.

This series (v5) contains the Documentation changes you requested.

- Danilo

Quoted 266 lines · Danilo Krummrich

This patch series provides a new UAPI for the Nouveau driver in order to support Vulkan features, such as sparse bindings and sparse residency.

Furthermore, with the DRM GPUVA manager it provides a new DRM core feature to keep track of GPU virtual address (VA) mappings in a more generic way.

The DRM GPUVA manager is indented to help drivers implement userspace-manageable GPU VA spaces in reference to the Vulkan API. In order to achieve this goal it serves the following purposes in this context.

1) Provide infrastructure to track GPU VA allocations and mappings, making use of the maple_tree.

2) Generically connect GPU VA mappings to their backing buffers, in particular DRM GEM objects.

3) Provide a common implementation to perform more complex mapping operations on the GPU VA space. In particular splitting and merging of GPU VA mappings, e.g. for intersecting mapping requests or partial unmap requests.

The new VM_BIND Nouveau UAPI build on top of the DRM GPUVA manager, itself providing the following new interfaces.

1) Initialize a GPU VA space via the new DRM_IOCTL_NOUVEAU_VM_INIT ioctl for UMDs to specify the portion of VA space managed by the kernel and userspace, respectively.

2) Allocate and free a VA space region as well as bind and unbind memory to the GPUs VA space via the new DRM_IOCTL_NOUVEAU_VM_BIND ioctl.

3) Execute push buffers with the new DRM_IOCTL_NOUVEAU_EXEC ioctl.

Both, DRM_IOCTL_NOUVEAU_VM_BIND and DRM_IOCTL_NOUVEAU_EXEC, make use of the DRM scheduler to queue jobs and support asynchronous processing with DRM syncobjs as synchronization mechanism.

By default DRM_IOCTL_NOUVEAU_VM_BIND does synchronous processing, DRM_IOCTL_NOUVEAU_EXEC supports asynchronous processing only.

The new VM_BIND UAPI for Nouveau makes also use of drm_exec (execution context for GEM buffers) by Christian König. Since the patch implementing drm_exec was not yet merged into drm-next it is part of this series, as well as a small fix for this patch, which was found while testing this series.

This patch series is also available at [1].

There is a Mesa NVK merge request by Dave Airlie [2] implementing the corresponding userspace parts for this series.

The Vulkan CTS test suite passes the sparse binding and sparse residency test cases for the new UAPI together with Dave's Mesa work.

There are also some test cases in the igt-gpu-tools project [3] for the new UAPI and hence the DRM GPU VA manager. However, most of them are testing the DRM GPU VA manager's logic through Nouveau's new UAPI and should be considered just as helper for implementation.

However, I absolutely intend to change those test cases to proper kunit test cases for the DRM GPUVA manager, once and if we agree on it's usefulness and design.

[1] https://gitlab.freedesktop.org/nouvelles/kernel/-/tree/new-uapi-drm-next /
https://gitlab.freedesktop.org/nouvelles/kernel/-/merge_requests/1
[2] https://gitlab.freedesktop.org/nouveau/mesa/-/merge_requests/150/
[3] https://gitlab.freedesktop.org/dakr/igt-gpu-tools/-/tree/wip_nouveau_vm_bind

Changes in V2:
==============
Nouveau:
- Reworked the Nouveau VM_BIND UAPI to avoid memory allocations in fence signalling critical sections. Updates to the VA space are split up in three separate stages, where only the 2. stage executes in a fence signalling critical section:

1. update the VA space, allocate new structures and page tables
2. (un-)map the requested memory bindings
3. free structures and page tables

- Separated generic job scheduler code from specific job implementations.
- Separated the EXEC and VM_BIND implementation of the UAPI.
- Reworked the locking parts of the nvkm/vmm RAW interface, such that (un-)map operations can be executed in fence signalling critical sections.

GPUVA Manager:
- made drm_gpuva_regions optional for users of the GPUVA manager
- allow NULL GEMs for drm_gpuva entries
- swichted from drm_mm to maple_tree for track drm_gpuva / drm_gpuva_region entries
- provide callbacks for users to allocate custom drm_gpuva_op structures to allow inheritance
- added user bits to drm_gpuva_flags
- added a prefetch operation type in order to support generating prefetch operations in the same way other operations generated
- hand the responsibility for mutual exclusion for a GEM's drm_gpuva list to the user; simplified corresponding (un-)link functions

Maple Tree:
- I added two maple tree patches to the series, one to support custom tree walk macros and one to hand the locking responsibility to the user of the GPUVA manager without pre-defined lockdep checks.

Changes in V3:
==============
Nouveau:
- Reworked the Nouveau VM_BIND UAPI to do the job cleanup (including page table cleanup) within a workqueue rather than the job_free() callback of the scheduler itself. A job_free() callback can stall the execution (run() callback) of the next job in the queue. Since the page table cleanup requires to take the same locks as need to be taken for page table allocation, doing it directly in the job_free() callback would still violate the fence signalling critical path.
- Separated Nouveau fence allocation and emit, such that we do not violate the fence signalling critical path in EXEC jobs.
- Implement "regions" (for handling sparse mappings through PDEs and dual page tables) within Nouveau.
- Drop the requirement for every mapping to be contained within a region.
- Add necassary synchronization of VM_BIND job operation sequences in order to work around limitations in page table handling. This will be addressed in a future re-work of Nouveau's page table handling.
- Fixed a couple of race conditions found through more testing. Thanks to Dave for consitently trying to break it. :-)

GPUVA Manager:
- Implement pre-allocation capabilities for tree modifications within fence signalling critical sections.
- Implement accessors to to apply tree modification while walking the GPUVA tree in order to actually support processing of drm_gpuva_ops through callbacks in fence signalling critical sections rather than through pre-allocated operation lists.
- Remove merging of GPUVAs; the kernel has limited to none knowlege about the semantics of mapping sequences. Hence, merging is purely speculative. It seems that gaining a significant (or at least a measurable) performance increase through merging is way more likely to happen when userspace is responsible for merging mappings up to the next larger page size if possible.
- Since merging was removed, regions pretty much loose their right to exist. They might still be useful for handling dual page tables or similar mechanisms, but since Nouveau seems to be the only driver having a need for this for now, regions were removed from the GPUVA manager.
- Fixed a couple of maple_tree related issues; thanks to Liam for helping me out.

Changes in V4:
==============
Nouveau:
- Refactored how specific VM_BIND and EXEC jobs are created and how their arguments are passed to the generic job implementation.
- Fixed a UAF race condition where bind job ops could have been freed already while still waiting for a job cleanup to finish. This is due to in certain cases we need to wait for mappings actually being unmapped before creating sparse regions in the same area.
- Re-based the code onto drm_exec v4 patch.

GPUVA Manager:
- Fixed a maple tree related bug when pre-allocating MA states. (Boris Brezillion)
- Made struct drm_gpuva_fn_ops a const object in all occurrences. (Boris Brezillion)

Changes in V5:
==============
Nouveau:
- Link and unlink GPUVAs outside the fence signalling critical path in nouveau_uvmm_bind_job_submit() holding the dma-resv lock. Mutual exclusion of BO evicts causing mapping invalidation and regular mapping operations is ensured with dma-fences.

GPUVA Manager:
- Removed the separate GEMs GPUVA list lock. Link and unlink as well as iterating the GEM's GPUVA list should be protected with the GEM's dma-resv lock instead.
- Renamed DRM_GPUVA_EVICTED flag to DRM_GPUVA_INVALIDATED. Mappings do not get eviced, they might get invalidated due to eviction.
- Maple tree uses the 'unsinged long' type for node entries. While this works for GPU VA spaces larger than 32-bit on 64-bit kernel, the GPU VA space is limited to 32-bit on 32-bit kernels as well.
As long as we do not have a 64-bit capable maple tree for 32-bit kernels, the GPU VA manager contains checks to throw warnings when GPU VA entries exceed the maple tree's storage capabilities.
- Extended the Documentation and added example code as requested by Donald Robson.

Christian König (1):
drm: execution context for GEM buffers v4

Danilo Krummrich (13):
maple_tree: split up MA_STATE() macro
drm: manager to keep track of GPUs VA mappings
drm: debugfs: provide infrastructure to dump a DRM GPU VA space drm/nouveau: new VM_BIND uapi interfaces drm/nouveau: get vmm via nouveau_cli_vmm() drm/nouveau: bo: initialize GEM GPU VA interface drm/nouveau: move usercopy helpers to nouveau_drv.h drm/nouveau: fence: separate fence alloc and emit drm/nouveau: fence: fail to emit when fence context is killed drm/nouveau: chan: provide nouveau_channel_kill() drm/nouveau: nvkm/vmm: implement raw ops to manage uvmm drm/nouveau: implement new VM_BIND uAPI
drm/nouveau: debugfs: implement DRM GPU VA debugfs

  Documentation/gpu/driver-uapi.rst             |   11 +
  Documentation/gpu/drm-mm.rst                  |   54 +
  drivers/gpu/drm/Kconfig                       |    6 +
  drivers/gpu/drm/Makefile                      |    3 +
  drivers/gpu/drm/drm_debugfs.c                 |   41 +
  drivers/gpu/drm/drm_exec.c                    |  278 +++
  drivers/gpu/drm/drm_gem.c                     |    3 +
  drivers/gpu/drm/drm_gpuva_mgr.c               | 1971 ++++++++++++++++
  drivers/gpu/drm/nouveau/Kbuild                |    3 +
  drivers/gpu/drm/nouveau/Kconfig               |    2 +
  drivers/gpu/drm/nouveau/dispnv04/crtc.c       |    9 +-
  drivers/gpu/drm/nouveau/include/nvif/if000c.h |   26 +-
  drivers/gpu/drm/nouveau/include/nvif/vmm.h    |   19 +-
  .../gpu/drm/nouveau/include/nvkm/subdev/mmu.h |   20 +-
  drivers/gpu/drm/nouveau/nouveau_abi16.c       |   24 +
  drivers/gpu/drm/nouveau/nouveau_abi16.h       |    1 +
  drivers/gpu/drm/nouveau/nouveau_bo.c          |  204 +-
  drivers/gpu/drm/nouveau/nouveau_bo.h          |    2 +-
  drivers/gpu/drm/nouveau/nouveau_chan.c        |   22 +-
  drivers/gpu/drm/nouveau/nouveau_chan.h        |    1 +
  drivers/gpu/drm/nouveau/nouveau_debugfs.c     |   39 +
  drivers/gpu/drm/nouveau/nouveau_dmem.c        |    9 +-
  drivers/gpu/drm/nouveau/nouveau_drm.c         |   27 +-
  drivers/gpu/drm/nouveau/nouveau_drv.h         |   94 +-
  drivers/gpu/drm/nouveau/nouveau_exec.c        |  418 ++++
  drivers/gpu/drm/nouveau/nouveau_exec.h        |   54 +
  drivers/gpu/drm/nouveau/nouveau_fence.c       |   23 +-
  drivers/gpu/drm/nouveau/nouveau_fence.h       |    5 +-
  drivers/gpu/drm/nouveau/nouveau_gem.c         |   62 +-
  drivers/gpu/drm/nouveau/nouveau_mem.h         |    5 +
  drivers/gpu/drm/nouveau/nouveau_prime.c       |    2 +-
  drivers/gpu/drm/nouveau/nouveau_sched.c       |  461 ++++
  drivers/gpu/drm/nouveau/nouveau_sched.h       |  123 +
  drivers/gpu/drm/nouveau/nouveau_svm.c         |    2 +-
  drivers/gpu/drm/nouveau/nouveau_uvmm.c        | 1979 +++++++++++++++++
  drivers/gpu/drm/nouveau/nouveau_uvmm.h        |  107 +
  drivers/gpu/drm/nouveau/nouveau_vmm.c         |    4 +-
  drivers/gpu/drm/nouveau/nvif/vmm.c            |  100 +-
  .../gpu/drm/nouveau/nvkm/subdev/mmu/uvmm.c    |  213 +-
  drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c |  197 +-
  drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h |   25 +
  .../drm/nouveau/nvkm/subdev/mmu/vmmgf100.c    |   16 +-
  .../drm/nouveau/nvkm/subdev/mmu/vmmgp100.c    |   16 +-
  .../gpu/drm/nouveau/nvkm/subdev/mmu/vmmnv50.c |   27 +-
  include/drm/drm_debugfs.h                     |   25 +
  include/drm/drm_drv.h                         |    6 +
  include/drm/drm_exec.h                        |  119 +
  include/drm/drm_gem.h                         |   52 +
  include/drm/drm_gpuva_mgr.h                   |  682 ++++++
  include/linux/maple_tree.h                    |    7 +-
  include/uapi/drm/nouveau_drm.h                |  209 ++
  51 files changed, 7566 insertions(+), 242 deletions(-)

create mode 100644 drivers/gpu/drm/drm_exec.c create mode 100644 drivers/gpu/drm/drm_gpuva_mgr.c create mode 100644 drivers/gpu/drm/nouveau/nouveau_exec.c create mode 100644 drivers/gpu/drm/nouveau/nouveau_exec.h create mode 100644 drivers/gpu/drm/nouveau/nouveau_sched.c create mode 100644 drivers/gpu/drm/nouveau/nouveau_sched.h create mode 100644 drivers/gpu/drm/nouveau/nouveau_uvmm.c create mode 100644 drivers/gpu/drm/nouveau/nouveau_uvmm.h create mode 100644 include/drm/drm_exec.h
create mode 100644 include/drm/drm_gpuva_mgr.h

base-commit: 2222dcb0775d36de28992f56455ab3967b30d380

maintainer
#33 lore raw

Since this is feature is nouveau only currently and doesn't disturb the current nouveau code paths, I'd like to try and get this work in tree so other drivers can work from it.

If there are any major objections to this, I'm happy to pull it back out again, but I'd like to get some acks/rb in the next couple of days in order to land some of it.

Dave.

Quoted 275 lines

forgot to add your email address to the patch series - sorry about that.

This series (v5) contains the Documentation changes you requested.

- Danilo

Quoted 266 lines · Danilo Krummrich

This patch series provides a new UAPI for the Nouveau driver in order to support Vulkan features, such as sparse bindings and sparse residency.

Furthermore, with the DRM GPUVA manager it provides a new DRM core feature to keep track of GPU virtual address (VA) mappings in a more generic way.

The DRM GPUVA manager is indented to help drivers implement userspace-manageable GPU VA spaces in reference to the Vulkan API. In order to achieve this goal it serves the following purposes in this context.

1) Provide infrastructure to track GPU VA allocations and mappings, making use of the maple_tree.

2) Generically connect GPU VA mappings to their backing buffers, in particular DRM GEM objects.

3) Provide a common implementation to perform more complex mapping operations on the GPU VA space. In particular splitting and merging of GPU VA mappings, e.g. for intersecting mapping requests or partial unmap requests.

The new VM_BIND Nouveau UAPI build on top of the DRM GPUVA manager, itself providing the following new interfaces.

1) Initialize a GPU VA space via the new DRM_IOCTL_NOUVEAU_VM_INIT ioctl for UMDs to specify the portion of VA space managed by the kernel and userspace, respectively.

2) Allocate and free a VA space region as well as bind and unbind memory to the GPUs VA space via the new DRM_IOCTL_NOUVEAU_VM_BIND ioctl.

3) Execute push buffers with the new DRM_IOCTL_NOUVEAU_EXEC ioctl.

Both, DRM_IOCTL_NOUVEAU_VM_BIND and DRM_IOCTL_NOUVEAU_EXEC, make use of the DRM scheduler to queue jobs and support asynchronous processing with DRM syncobjs as synchronization mechanism.

By default DRM_IOCTL_NOUVEAU_VM_BIND does synchronous processing, DRM_IOCTL_NOUVEAU_EXEC supports asynchronous processing only.

The new VM_BIND UAPI for Nouveau makes also use of drm_exec (execution context for GEM buffers) by Christian König. Since the patch implementing drm_exec was not yet merged into drm-next it is part of this series, as well as a small fix for this patch, which was found while testing this series.

This patch series is also available at [1].

There is a Mesa NVK merge request by Dave Airlie [2] implementing the corresponding userspace parts for this series.

The Vulkan CTS test suite passes the sparse binding and sparse residency test cases for the new UAPI together with Dave's Mesa work.

There are also some test cases in the igt-gpu-tools project [3] for the new UAPI and hence the DRM GPU VA manager. However, most of them are testing the DRM GPU VA manager's logic through Nouveau's new UAPI and should be considered just as helper for implementation.

However, I absolutely intend to change those test cases to proper kunit test cases for the DRM GPUVA manager, once and if we agree on it's usefulness and design.

[1] https://gitlab.freedesktop.org/nouvelles/kernel/-/tree/new-uapi-drm-next /
https://gitlab.freedesktop.org/nouvelles/kernel/-/merge_requests/1
[2] https://gitlab.freedesktop.org/nouveau/mesa/-/merge_requests/150/
[3] https://gitlab.freedesktop.org/dakr/igt-gpu-tools/-/tree/wip_nouveau_vm_bind

Changes in V2:
==============
Nouveau:
- Reworked the Nouveau VM_BIND UAPI to avoid memory allocations in fence signalling critical sections. Updates to the VA space are split up in three separate stages, where only the 2. stage executes in a fence signalling critical section:

1. update the VA space, allocate new structures and page tables
2. (un-)map the requested memory bindings
3. free structures and page tables

- Separated generic job scheduler code from specific job implementations.
- Separated the EXEC and VM_BIND implementation of the UAPI.
- Reworked the locking parts of the nvkm/vmm RAW interface, such that (un-)map operations can be executed in fence signalling critical sections.

GPUVA Manager:
- made drm_gpuva_regions optional for users of the GPUVA manager
- allow NULL GEMs for drm_gpuva entries
- swichted from drm_mm to maple_tree for track drm_gpuva / drm_gpuva_region entries
- provide callbacks for users to allocate custom drm_gpuva_op structures to allow inheritance
- added user bits to drm_gpuva_flags
- added a prefetch operation type in order to support generating prefetch operations in the same way other operations generated
- hand the responsibility for mutual exclusion for a GEM's drm_gpuva list to the user; simplified corresponding (un-)link functions

Maple Tree:
- I added two maple tree patches to the series, one to support custom tree walk macros and one to hand the locking responsibility to the user of the GPUVA manager without pre-defined lockdep checks.

Changes in V3:
==============
Nouveau:
- Reworked the Nouveau VM_BIND UAPI to do the job cleanup (including page table cleanup) within a workqueue rather than the job_free() callback of the scheduler itself. A job_free() callback can stall the execution (run() callback) of the next job in the queue. Since the page table cleanup requires to take the same locks as need to be taken for page table allocation, doing it directly in the job_free() callback would still violate the fence signalling critical path.
- Separated Nouveau fence allocation and emit, such that we do not violate the fence signalling critical path in EXEC jobs.
- Implement "regions" (for handling sparse mappings through PDEs and dual page tables) within Nouveau.
- Drop the requirement for every mapping to be contained within a region.
- Add necassary synchronization of VM_BIND job operation sequences in order to work around limitations in page table handling. This will be addressed in a future re-work of Nouveau's page table handling.
- Fixed a couple of race conditions found through more testing. Thanks to Dave for consitently trying to break it. :-)

GPUVA Manager:
- Implement pre-allocation capabilities for tree modifications within fence signalling critical sections.
- Implement accessors to to apply tree modification while walking the GPUVA tree in order to actually support processing of drm_gpuva_ops through callbacks in fence signalling critical sections rather than through pre-allocated operation lists.
- Remove merging of GPUVAs; the kernel has limited to none knowlege about the semantics of mapping sequences. Hence, merging is purely speculative. It seems that gaining a significant (or at least a measurable) performance increase through merging is way more likely to happen when userspace is responsible for merging mappings up to the next larger page size if possible.
- Since merging was removed, regions pretty much loose their right to exist. They might still be useful for handling dual page tables or similar mechanisms, but since Nouveau seems to be the only driver having a need for this for now, regions were removed from the GPUVA manager.
- Fixed a couple of maple_tree related issues; thanks to Liam for helping me out.

Changes in V4:
==============
Nouveau:
- Refactored how specific VM_BIND and EXEC jobs are created and how their arguments are passed to the generic job implementation.
- Fixed a UAF race condition where bind job ops could have been freed already while still waiting for a job cleanup to finish. This is due to in certain cases we need to wait for mappings actually being unmapped before creating sparse regions in the same area.
- Re-based the code onto drm_exec v4 patch.

GPUVA Manager:
- Fixed a maple tree related bug when pre-allocating MA states. (Boris Brezillion)
- Made struct drm_gpuva_fn_ops a const object in all occurrences. (Boris Brezillion)

Changes in V5:
==============
Nouveau:
- Link and unlink GPUVAs outside the fence signalling critical path in nouveau_uvmm_bind_job_submit() holding the dma-resv lock. Mutual exclusion of BO evicts causing mapping invalidation and regular mapping operations is ensured with dma-fences.

GPUVA Manager:
- Removed the separate GEMs GPUVA list lock. Link and unlink as well as iterating the GEM's GPUVA list should be protected with the GEM's dma-resv lock instead.
- Renamed DRM_GPUVA_EVICTED flag to DRM_GPUVA_INVALIDATED. Mappings do not get eviced, they might get invalidated due to eviction.
- Maple tree uses the 'unsinged long' type for node entries. While this works for GPU VA spaces larger than 32-bit on 64-bit kernel, the GPU VA space is limited to 32-bit on 32-bit kernels as well.
As long as we do not have a 64-bit capable maple tree for 32-bit kernels, the GPU VA manager contains checks to throw warnings when GPU VA entries exceed the maple tree's storage capabilities.
- Extended the Documentation and added example code as requested by Donald Robson.

Christian König (1):
drm: execution context for GEM buffers v4

Danilo Krummrich (13):
maple_tree: split up MA_STATE() macro
drm: manager to keep track of GPUs VA mappings
drm: debugfs: provide infrastructure to dump a DRM GPU VA space drm/nouveau: new VM_BIND uapi interfaces drm/nouveau: get vmm via nouveau_cli_vmm() drm/nouveau: bo: initialize GEM GPU VA interface drm/nouveau: move usercopy helpers to nouveau_drv.h drm/nouveau: fence: separate fence alloc and emit drm/nouveau: fence: fail to emit when fence context is killed drm/nouveau: chan: provide nouveau_channel_kill() drm/nouveau: nvkm/vmm: implement raw ops to manage uvmm drm/nouveau: implement new VM_BIND uAPI
drm/nouveau: debugfs: implement DRM GPU VA debugfs

  Documentation/gpu/driver-uapi.rst             |   11 +
  Documentation/gpu/drm-mm.rst                  |   54 +
  drivers/gpu/drm/Kconfig                       |    6 +
  drivers/gpu/drm/Makefile                      |    3 +
  drivers/gpu/drm/drm_debugfs.c                 |   41 +
  drivers/gpu/drm/drm_exec.c                    |  278 +++
  drivers/gpu/drm/drm_gem.c                     |    3 +
  drivers/gpu/drm/drm_gpuva_mgr.c               | 1971 ++++++++++++++++
  drivers/gpu/drm/nouveau/Kbuild                |    3 +
  drivers/gpu/drm/nouveau/Kconfig               |    2 +
  drivers/gpu/drm/nouveau/dispnv04/crtc.c       |    9 +-
  drivers/gpu/drm/nouveau/include/nvif/if000c.h |   26 +-
  drivers/gpu/drm/nouveau/include/nvif/vmm.h    |   19 +-
  .../gpu/drm/nouveau/include/nvkm/subdev/mmu.h |   20 +-
  drivers/gpu/drm/nouveau/nouveau_abi16.c       |   24 +
  drivers/gpu/drm/nouveau/nouveau_abi16.h       |    1 +
  drivers/gpu/drm/nouveau/nouveau_bo.c          |  204 +-
  drivers/gpu/drm/nouveau/nouveau_bo.h          |    2 +-
  drivers/gpu/drm/nouveau/nouveau_chan.c        |   22 +-
  drivers/gpu/drm/nouveau/nouveau_chan.h        |    1 +
  drivers/gpu/drm/nouveau/nouveau_debugfs.c     |   39 +
  drivers/gpu/drm/nouveau/nouveau_dmem.c        |    9 +-
  drivers/gpu/drm/nouveau/nouveau_drm.c         |   27 +-
  drivers/gpu/drm/nouveau/nouveau_drv.h         |   94 +-
  drivers/gpu/drm/nouveau/nouveau_exec.c        |  418 ++++
  drivers/gpu/drm/nouveau/nouveau_exec.h        |   54 +
  drivers/gpu/drm/nouveau/nouveau_fence.c       |   23 +-
  drivers/gpu/drm/nouveau/nouveau_fence.h       |    5 +-
  drivers/gpu/drm/nouveau/nouveau_gem.c         |   62 +-
  drivers/gpu/drm/nouveau/nouveau_mem.h         |    5 +
  drivers/gpu/drm/nouveau/nouveau_prime.c       |    2 +-
  drivers/gpu/drm/nouveau/nouveau_sched.c       |  461 ++++
  drivers/gpu/drm/nouveau/nouveau_sched.h       |  123 +
  drivers/gpu/drm/nouveau/nouveau_svm.c         |    2 +-
  drivers/gpu/drm/nouveau/nouveau_uvmm.c        | 1979 +++++++++++++++++
  drivers/gpu/drm/nouveau/nouveau_uvmm.h        |  107 +
  drivers/gpu/drm/nouveau/nouveau_vmm.c         |    4 +-
  drivers/gpu/drm/nouveau/nvif/vmm.c            |  100 +-
  .../gpu/drm/nouveau/nvkm/subdev/mmu/uvmm.c    |  213 +-
  drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c |  197 +-
  drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h |   25 +
  .../drm/nouveau/nvkm/subdev/mmu/vmmgf100.c    |   16 +-
  .../drm/nouveau/nvkm/subdev/mmu/vmmgp100.c    |   16 +-
  .../gpu/drm/nouveau/nvkm/subdev/mmu/vmmnv50.c |   27 +-
  include/drm/drm_debugfs.h                     |   25 +
  include/drm/drm_drv.h                         |    6 +
  include/drm/drm_exec.h                        |  119 +
  include/drm/drm_gem.h                         |   52 +
  include/drm/drm_gpuva_mgr.h                   |  682 ++++++
  include/linux/maple_tree.h                    |    7 +-
  include/uapi/drm/nouveau_drm.h                |  209 ++
  51 files changed, 7566 insertions(+), 242 deletions(-)

create mode 100644 drivers/gpu/drm/drm_exec.c create mode 100644 drivers/gpu/drm/drm_gpuva_mgr.c create mode 100644 drivers/gpu/drm/nouveau/nouveau_exec.c create mode 100644 drivers/gpu/drm/nouveau/nouveau_exec.h create mode 100644 drivers/gpu/drm/nouveau/nouveau_sched.c create mode 100644 drivers/gpu/drm/nouveau/nouveau_sched.h create mode 100644 drivers/gpu/drm/nouveau/nouveau_uvmm.c create mode 100644 drivers/gpu/drm/nouveau/nouveau_uvmm.h create mode 100644 include/drm/drm_exec.h
create mode 100644 include/drm/drm_gpuva_mgr.h

base-commit: 2222dcb0775d36de28992f56455ab3967b30d380

#34 lore raw
Quoted 5 lines · Dave Airlie

Since this is feature is nouveau only currently and doesn't disturb the current nouveau code paths, I'd like to try and get this work in tree so other drivers can work from it.

Quoted 30 lines

If there are any major objections to this, I'm happy to pull it back out again, but I'd like to get some acks/rb in the next couple of days in order to land some of it.

Dave.

Quoted 23 lines

forgot to add your email address to the patch series - sorry about that.

This series (v5) contains the Documentation changes you requested.

- Danilo

Quoted 15 lines · Danilo Krummrich

This patch series provides a new UAPI for the Nouveau driver in order to support Vulkan features, such as sparse bindings and sparse residency.

Furthermore, with the DRM GPUVA manager it provides a new DRM core feature to keep track of GPU virtual address (VA) mappings in a more generic way.

The DRM GPUVA manager is indented to help drivers implement userspace-manageable GPU VA spaces in reference to the Vulkan API. In order to achieve this goal it serves the following purposes in this context.

1) Provide infrastructure to track GPU VA allocations and mappings, making use of the maple_tree.

2) Generically connect GPU VA mappings to their backing buffers, in particular DRM GEM objects.

Will this manager be able to connect GPU VA mappings to host memory allocations (aka user pointers) ?

I only skimmed over the uapi definitions, but from that quick glance I saw you can only pass a (gem) handle to the vm bind uapi.

I think it is an important feature because you don't want to have two GPU VA managers running in your driver (if that's even possible). Maybe we should at least try to make sure the uapi is/will be compatible with such an extension.

Oded

Oded

Quoted 252 lines
Quoted 252 lines
Quoted 251 lines

3) Provide a common implementation to perform more complex mapping operations on the GPU VA space. In particular splitting and merging of GPU VA mappings, e.g. for intersecting mapping requests or partial unmap requests.

The new VM_BIND Nouveau UAPI build on top of the DRM GPUVA manager, itself providing the following new interfaces.

1) Initialize a GPU VA space via the new DRM_IOCTL_NOUVEAU_VM_INIT ioctl for UMDs to specify the portion of VA space managed by the kernel and userspace, respectively.

2) Allocate and free a VA space region as well as bind and unbind memory to the GPUs VA space via the new DRM_IOCTL_NOUVEAU_VM_BIND ioctl.

3) Execute push buffers with the new DRM_IOCTL_NOUVEAU_EXEC ioctl.

Both, DRM_IOCTL_NOUVEAU_VM_BIND and DRM_IOCTL_NOUVEAU_EXEC, make use of the DRM scheduler to queue jobs and support asynchronous processing with DRM syncobjs as synchronization mechanism.

By default DRM_IOCTL_NOUVEAU_VM_BIND does synchronous processing, DRM_IOCTL_NOUVEAU_EXEC supports asynchronous processing only.

The new VM_BIND UAPI for Nouveau makes also use of drm_exec (execution context for GEM buffers) by Christian König. Since the patch implementing drm_exec was not yet merged into drm-next it is part of this series, as well as a small fix for this patch, which was found while testing this series.

This patch series is also available at [1].

There is a Mesa NVK merge request by Dave Airlie [2] implementing the corresponding userspace parts for this series.

The Vulkan CTS test suite passes the sparse binding and sparse residency test cases for the new UAPI together with Dave's Mesa work.

There are also some test cases in the igt-gpu-tools project [3] for the new UAPI and hence the DRM GPU VA manager. However, most of them are testing the DRM GPU VA manager's logic through Nouveau's new UAPI and should be considered just as helper for implementation.

However, I absolutely intend to change those test cases to proper kunit test cases for the DRM GPUVA manager, once and if we agree on it's usefulness and design.

[1] https://gitlab.freedesktop.org/nouvelles/kernel/-/tree/new-uapi-drm-next /
https://gitlab.freedesktop.org/nouvelles/kernel/-/merge_requests/1
[2] https://gitlab.freedesktop.org/nouveau/mesa/-/merge_requests/150/
[3] https://gitlab.freedesktop.org/dakr/igt-gpu-tools/-/tree/wip_nouveau_vm_bind

Changes in V2:
==============
Nouveau:
- Reworked the Nouveau VM_BIND UAPI to avoid memory allocations in fence signalling critical sections. Updates to the VA space are split up in three separate stages, where only the 2. stage executes in a fence signalling critical section:

1. update the VA space, allocate new structures and page tables
2. (un-)map the requested memory bindings
3. free structures and page tables

- Separated generic job scheduler code from specific job implementations.
- Separated the EXEC and VM_BIND implementation of the UAPI.
- Reworked the locking parts of the nvkm/vmm RAW interface, such that (un-)map operations can be executed in fence signalling critical sections.

GPUVA Manager:
- made drm_gpuva_regions optional for users of the GPUVA manager
- allow NULL GEMs for drm_gpuva entries
- swichted from drm_mm to maple_tree for track drm_gpuva / drm_gpuva_region entries
- provide callbacks for users to allocate custom drm_gpuva_op structures to allow inheritance
- added user bits to drm_gpuva_flags
- added a prefetch operation type in order to support generating prefetch operations in the same way other operations generated
- hand the responsibility for mutual exclusion for a GEM's drm_gpuva list to the user; simplified corresponding (un-)link functions

Maple Tree:
- I added two maple tree patches to the series, one to support custom tree walk macros and one to hand the locking responsibility to the user of the GPUVA manager without pre-defined lockdep checks.

Changes in V3:
==============
Nouveau:
- Reworked the Nouveau VM_BIND UAPI to do the job cleanup (including page table cleanup) within a workqueue rather than the job_free() callback of the scheduler itself. A job_free() callback can stall the execution (run() callback) of the next job in the queue. Since the page table cleanup requires to take the same locks as need to be taken for page table allocation, doing it directly in the job_free() callback would still violate the fence signalling critical path.
- Separated Nouveau fence allocation and emit, such that we do not violate the fence signalling critical path in EXEC jobs.
- Implement "regions" (for handling sparse mappings through PDEs and dual page tables) within Nouveau.
- Drop the requirement for every mapping to be contained within a region.
- Add necassary synchronization of VM_BIND job operation sequences in order to work around limitations in page table handling. This will be addressed in a future re-work of Nouveau's page table handling.
- Fixed a couple of race conditions found through more testing. Thanks to Dave for consitently trying to break it. :-)

GPUVA Manager:
- Implement pre-allocation capabilities for tree modifications within fence signalling critical sections.
- Implement accessors to to apply tree modification while walking the GPUVA tree in order to actually support processing of drm_gpuva_ops through callbacks in fence signalling critical sections rather than through pre-allocated operation lists.
- Remove merging of GPUVAs; the kernel has limited to none knowlege about the semantics of mapping sequences. Hence, merging is purely speculative. It seems that gaining a significant (or at least a measurable) performance increase through merging is way more likely to happen when userspace is responsible for merging mappings up to the next larger page size if possible.
- Since merging was removed, regions pretty much loose their right to exist. They might still be useful for handling dual page tables or similar mechanisms, but since Nouveau seems to be the only driver having a need for this for now, regions were removed from the GPUVA manager.
- Fixed a couple of maple_tree related issues; thanks to Liam for helping me out.

Changes in V4:
==============
Nouveau:
- Refactored how specific VM_BIND and EXEC jobs are created and how their arguments are passed to the generic job implementation.
- Fixed a UAF race condition where bind job ops could have been freed already while still waiting for a job cleanup to finish. This is due to in certain cases we need to wait for mappings actually being unmapped before creating sparse regions in the same area.
- Re-based the code onto drm_exec v4 patch.

GPUVA Manager:
- Fixed a maple tree related bug when pre-allocating MA states. (Boris Brezillion)
- Made struct drm_gpuva_fn_ops a const object in all occurrences. (Boris Brezillion)

Changes in V5:
==============
Nouveau:
- Link and unlink GPUVAs outside the fence signalling critical path in nouveau_uvmm_bind_job_submit() holding the dma-resv lock. Mutual exclusion of BO evicts causing mapping invalidation and regular mapping operations is ensured with dma-fences.

GPUVA Manager:
- Removed the separate GEMs GPUVA list lock. Link and unlink as well as iterating the GEM's GPUVA list should be protected with the GEM's dma-resv lock instead.
- Renamed DRM_GPUVA_EVICTED flag to DRM_GPUVA_INVALIDATED. Mappings do not get eviced, they might get invalidated due to eviction.
- Maple tree uses the 'unsinged long' type for node entries. While this works for GPU VA spaces larger than 32-bit on 64-bit kernel, the GPU VA space is limited to 32-bit on 32-bit kernels as well.
As long as we do not have a 64-bit capable maple tree for 32-bit kernels, the GPU VA manager contains checks to throw warnings when GPU VA entries exceed the maple tree's storage capabilities.
- Extended the Documentation and added example code as requested by Donald Robson.

Christian König (1):
drm: execution context for GEM buffers v4

Danilo Krummrich (13):
maple_tree: split up MA_STATE() macro
drm: manager to keep track of GPUs VA mappings
drm: debugfs: provide infrastructure to dump a DRM GPU VA space drm/nouveau: new VM_BIND uapi interfaces drm/nouveau: get vmm via nouveau_cli_vmm() drm/nouveau: bo: initialize GEM GPU VA interface drm/nouveau: move usercopy helpers to nouveau_drv.h drm/nouveau: fence: separate fence alloc and emit drm/nouveau: fence: fail to emit when fence context is killed drm/nouveau: chan: provide nouveau_channel_kill() drm/nouveau: nvkm/vmm: implement raw ops to manage uvmm drm/nouveau: implement new VM_BIND uAPI
drm/nouveau: debugfs: implement DRM GPU VA debugfs

  Documentation/gpu/driver-uapi.rst             |   11 +
  Documentation/gpu/drm-mm.rst                  |   54 +
  drivers/gpu/drm/Kconfig                       |    6 +
  drivers/gpu/drm/Makefile                      |    3 +
  drivers/gpu/drm/drm_debugfs.c                 |   41 +
  drivers/gpu/drm/drm_exec.c                    |  278 +++
  drivers/gpu/drm/drm_gem.c                     |    3 +
  drivers/gpu/drm/drm_gpuva_mgr.c               | 1971 ++++++++++++++++
  drivers/gpu/drm/nouveau/Kbuild                |    3 +
  drivers/gpu/drm/nouveau/Kconfig               |    2 +
  drivers/gpu/drm/nouveau/dispnv04/crtc.c       |    9 +-
  drivers/gpu/drm/nouveau/include/nvif/if000c.h |   26 +-
  drivers/gpu/drm/nouveau/include/nvif/vmm.h    |   19 +-
  .../gpu/drm/nouveau/include/nvkm/subdev/mmu.h |   20 +-
  drivers/gpu/drm/nouveau/nouveau_abi16.c       |   24 +
  drivers/gpu/drm/nouveau/nouveau_abi16.h       |    1 +
  drivers/gpu/drm/nouveau/nouveau_bo.c          |  204 +-
  drivers/gpu/drm/nouveau/nouveau_bo.h          |    2 +-
  drivers/gpu/drm/nouveau/nouveau_chan.c        |   22 +-
  drivers/gpu/drm/nouveau/nouveau_chan.h        |    1 +
  drivers/gpu/drm/nouveau/nouveau_debugfs.c     |   39 +
  drivers/gpu/drm/nouveau/nouveau_dmem.c        |    9 +-
  drivers/gpu/drm/nouveau/nouveau_drm.c         |   27 +-
  drivers/gpu/drm/nouveau/nouveau_drv.h         |   94 +-
  drivers/gpu/drm/nouveau/nouveau_exec.c        |  418 ++++
  drivers/gpu/drm/nouveau/nouveau_exec.h        |   54 +
  drivers/gpu/drm/nouveau/nouveau_fence.c       |   23 +-
  drivers/gpu/drm/nouveau/nouveau_fence.h       |    5 +-
  drivers/gpu/drm/nouveau/nouveau_gem.c         |   62 +-
  drivers/gpu/drm/nouveau/nouveau_mem.h         |    5 +
  drivers/gpu/drm/nouveau/nouveau_prime.c       |    2 +-
  drivers/gpu/drm/nouveau/nouveau_sched.c       |  461 ++++
  drivers/gpu/drm/nouveau/nouveau_sched.h       |  123 +
  drivers/gpu/drm/nouveau/nouveau_svm.c         |    2 +-
  drivers/gpu/drm/nouveau/nouveau_uvmm.c        | 1979 +++++++++++++++++
  drivers/gpu/drm/nouveau/nouveau_uvmm.h        |  107 +
  drivers/gpu/drm/nouveau/nouveau_vmm.c         |    4 +-
  drivers/gpu/drm/nouveau/nvif/vmm.c            |  100 +-
  .../gpu/drm/nouveau/nvkm/subdev/mmu/uvmm.c    |  213 +-
  drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c |  197 +-
  drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h |   25 +
  .../drm/nouveau/nvkm/subdev/mmu/vmmgf100.c    |   16 +-
  .../drm/nouveau/nvkm/subdev/mmu/vmmgp100.c    |   16 +-
  .../gpu/drm/nouveau/nvkm/subdev/mmu/vmmnv50.c |   27 +-
  include/drm/drm_debugfs.h                     |   25 +
  include/drm/drm_drv.h                         |    6 +
  include/drm/drm_exec.h                        |  119 +
  include/drm/drm_gem.h                         |   52 +
  include/drm/drm_gpuva_mgr.h                   |  682 ++++++
  include/linux/maple_tree.h                    |    7 +-
  include/uapi/drm/nouveau_drm.h                |  209 ++
  51 files changed, 7566 insertions(+), 242 deletions(-)

create mode 100644 drivers/gpu/drm/drm_exec.c create mode 100644 drivers/gpu/drm/drm_gpuva_mgr.c create mode 100644 drivers/gpu/drm/nouveau/nouveau_exec.c create mode 100644 drivers/gpu/drm/nouveau/nouveau_exec.h create mode 100644 drivers/gpu/drm/nouveau/nouveau_sched.c create mode 100644 drivers/gpu/drm/nouveau/nouveau_sched.h create mode 100644 drivers/gpu/drm/nouveau/nouveau_uvmm.c create mode 100644 drivers/gpu/drm/nouveau/nouveau_uvmm.h create mode 100644 include/drm/drm_exec.h
create mode 100644 include/drm/drm_gpuva_mgr.h

base-commit: 2222dcb0775d36de28992f56455ab3967b30d380

maintainer
#35 lore raw
Quoted 40 lines · Oded Gabbay
Quoted 5 lines · Dave Airlie

Since this is feature is nouveau only currently and doesn't disturb the current nouveau code paths, I'd like to try and get this work in tree so other drivers can work from it.

Quoted 30 lines

If there are any major objections to this, I'm happy to pull it back out again, but I'd like to get some acks/rb in the next couple of days in order to land some of it.

Dave.

Quoted 23 lines

forgot to add your email address to the patch series - sorry about that.

This series (v5) contains the Documentation changes you requested.

- Danilo

Quoted 15 lines · Danilo Krummrich

This patch series provides a new UAPI for the Nouveau driver in order to support Vulkan features, such as sparse bindings and sparse residency.

Furthermore, with the DRM GPUVA manager it provides a new DRM core feature to keep track of GPU virtual address (VA) mappings in a more generic way.

The DRM GPUVA manager is indented to help drivers implement userspace-manageable GPU VA spaces in reference to the Vulkan API. In order to achieve this goal it serves the following purposes in this context.

1) Provide infrastructure to track GPU VA allocations and mappings, making use of the maple_tree.

2) Generically connect GPU VA mappings to their backing buffers, in particular DRM GEM objects.

Will this manager be able to connect GPU VA mappings to host memory allocations (aka user pointers) ?

Quoted 8 lines

I only skimmed over the uapi definitions, but from that quick glance I saw you can only pass a (gem) handle to the vm bind uapi.

I think it is an important feature because you don't want to have two GPU VA managers running in your driver (if that's even possible). Maybe we should at least try to make sure the uapi is/will be compatible with such an extension.

I think that would have to be a new uAPI entry point anyways, since managing user ptrs is extra, but the uAPI is nouveau specific and nouveau has no hostptr support as of now.

The gpuva manager is kernel internal, I think adding host ptr tracking is useful, but I don't think it's a blocker right now.

One of the reasons I'd like to get this in the tree is to add things like that instead of overloading this initial patchset with feature creep.

Dave.

#36 lore raw
Quoted 55 lines · Dave Airlie
Quoted 40 lines · Oded Gabbay
Quoted 5 lines · Dave Airlie

Since this is feature is nouveau only currently and doesn't disturb the current nouveau code paths, I'd like to try and get this work in tree so other drivers can work from it.

Quoted 30 lines

If there are any major objections to this, I'm happy to pull it back out again, but I'd like to get some acks/rb in the next couple of days in order to land some of it.

Dave.

Quoted 23 lines

forgot to add your email address to the patch series - sorry about that.

This series (v5) contains the Documentation changes you requested.

- Danilo

Quoted 15 lines · Danilo Krummrich

This patch series provides a new UAPI for the Nouveau driver in order to support Vulkan features, such as sparse bindings and sparse residency.

Furthermore, with the DRM GPUVA manager it provides a new DRM core feature to keep track of GPU virtual address (VA) mappings in a more generic way.

The DRM GPUVA manager is indented to help drivers implement userspace-manageable GPU VA spaces in reference to the Vulkan API. In order to achieve this goal it serves the following purposes in this context.

1) Provide infrastructure to track GPU VA allocations and mappings, making use of the maple_tree.

2) Generically connect GPU VA mappings to their backing buffers, in particular DRM GEM objects.

Will this manager be able to connect GPU VA mappings to host memory allocations (aka user pointers) ?

Quoted 8 lines

I only skimmed over the uapi definitions, but from that quick glance I saw you can only pass a (gem) handle to the vm bind uapi.

I think it is an important feature because you don't want to have two GPU VA managers running in your driver (if that's even possible). Maybe we should at least try to make sure the uapi is/will be compatible with such an extension.

I think that would have to be a new uAPI entry point anyways, since managing user ptrs is extra, but the uAPI is nouveau specific and nouveau has no hostptr support as of now.

Quoted 8 lines

The gpuva manager is kernel internal, I think adding host ptr tracking is useful, but I don't think it's a blocker right now.

One of the reasons I'd like to get this in the tree is to add things like that instead of overloading this initial patchset with feature creep.

Dave.

ok, that makes sense.
Thanks,
Oded

Reviewed-by on the v5 series
#37 lore raw
In reply to Danilo Krummrich

Hi Danilo,

Quoted 183 lines · On Tue, 20 Jun 2023 02:42:03 +0200 Danilo Krummrich <[email protected]> wrote:

This patch series provides a new UAPI for the Nouveau driver in order to support Vulkan features, such as sparse bindings and sparse residency.

Furthermore, with the DRM GPUVA manager it provides a new DRM core feature to keep track of GPU virtual address (VA) mappings in a more generic way.

The DRM GPUVA manager is indented to help drivers implement userspace-manageable GPU VA spaces in reference to the Vulkan API. In order to achieve this goal it serves the following purposes in this context.

1) Provide infrastructure to track GPU VA allocations and mappings, making use of the maple_tree.

2) Generically connect GPU VA mappings to their backing buffers, in particular DRM GEM objects.

3) Provide a common implementation to perform more complex mapping operations on the GPU VA space. In particular splitting and merging of GPU VA mappings, e.g. for intersecting mapping requests or partial unmap requests.

The new VM_BIND Nouveau UAPI build on top of the DRM GPUVA manager, itself providing the following new interfaces.

1) Initialize a GPU VA space via the new DRM_IOCTL_NOUVEAU_VM_INIT ioctl for UMDs to specify the portion of VA space managed by the kernel and userspace, respectively.

2) Allocate and free a VA space region as well as bind and unbind memory to the GPUs VA space via the new DRM_IOCTL_NOUVEAU_VM_BIND ioctl.

3) Execute push buffers with the new DRM_IOCTL_NOUVEAU_EXEC ioctl.

Both, DRM_IOCTL_NOUVEAU_VM_BIND and DRM_IOCTL_NOUVEAU_EXEC, make use of the DRM scheduler to queue jobs and support asynchronous processing with DRM syncobjs as synchronization mechanism.

By default DRM_IOCTL_NOUVEAU_VM_BIND does synchronous processing, DRM_IOCTL_NOUVEAU_EXEC supports asynchronous processing only.

The new VM_BIND UAPI for Nouveau makes also use of drm_exec (execution context for GEM buffers) by Christian König. Since the patch implementing drm_exec was not yet merged into drm-next it is part of this series, as well as a small fix for this patch, which was found while testing this series.

This patch series is also available at [1].

There is a Mesa NVK merge request by Dave Airlie [2] implementing the corresponding userspace parts for this series.

The Vulkan CTS test suite passes the sparse binding and sparse residency test cases for the new UAPI together with Dave's Mesa work.

There are also some test cases in the igt-gpu-tools project [3] for the new UAPI and hence the DRM GPU VA manager. However, most of them are testing the DRM GPU VA manager's logic through Nouveau's new UAPI and should be considered just as helper for implementation.

However, I absolutely intend to change those test cases to proper kunit test cases for the DRM GPUVA manager, once and if we agree on it's usefulness and design.

[1] https://gitlab.freedesktop.org/nouvelles/kernel/-/tree/new-uapi-drm-next /
https://gitlab.freedesktop.org/nouvelles/kernel/-/merge_requests/1
[2] https://gitlab.freedesktop.org/nouveau/mesa/-/merge_requests/150/
[3] https://gitlab.freedesktop.org/dakr/igt-gpu-tools/-/tree/wip_nouveau_vm_bind

Changes in V2:
==============
Nouveau:
- Reworked the Nouveau VM_BIND UAPI to avoid memory allocations in fence signalling critical sections. Updates to the VA space are split up in three separate stages, where only the 2. stage executes in a fence signalling critical section:

1. update the VA space, allocate new structures and page tables
2. (un-)map the requested memory bindings
3. free structures and page tables

- Separated generic job scheduler code from specific job implementations.
- Separated the EXEC and VM_BIND implementation of the UAPI.
- Reworked the locking parts of the nvkm/vmm RAW interface, such that (un-)map operations can be executed in fence signalling critical sections.

GPUVA Manager:
- made drm_gpuva_regions optional for users of the GPUVA manager
- allow NULL GEMs for drm_gpuva entries
- swichted from drm_mm to maple_tree for track drm_gpuva / drm_gpuva_region entries
- provide callbacks for users to allocate custom drm_gpuva_op structures to allow inheritance
- added user bits to drm_gpuva_flags
- added a prefetch operation type in order to support generating prefetch operations in the same way other operations generated
- hand the responsibility for mutual exclusion for a GEM's drm_gpuva list to the user; simplified corresponding (un-)link functions

Maple Tree:
- I added two maple tree patches to the series, one to support custom tree walk macros and one to hand the locking responsibility to the user of the GPUVA manager without pre-defined lockdep checks.

Changes in V3:
==============
Nouveau:
- Reworked the Nouveau VM_BIND UAPI to do the job cleanup (including page table cleanup) within a workqueue rather than the job_free() callback of the scheduler itself. A job_free() callback can stall the execution (run() callback) of the next job in the queue. Since the page table cleanup requires to take the same locks as need to be taken for page table allocation, doing it directly in the job_free() callback would still violate the fence signalling critical path.
- Separated Nouveau fence allocation and emit, such that we do not violate the fence signalling critical path in EXEC jobs.
- Implement "regions" (for handling sparse mappings through PDEs and dual page tables) within Nouveau.
- Drop the requirement for every mapping to be contained within a region.
- Add necassary synchronization of VM_BIND job operation sequences in order to work around limitations in page table handling. This will be addressed in a future re-work of Nouveau's page table handling.
- Fixed a couple of race conditions found through more testing. Thanks to Dave for consitently trying to break it. :-)

GPUVA Manager:
- Implement pre-allocation capabilities for tree modifications within fence signalling critical sections.
- Implement accessors to to apply tree modification while walking the GPUVA tree in order to actually support processing of drm_gpuva_ops through callbacks in fence signalling critical sections rather than through pre-allocated operation lists.
- Remove merging of GPUVAs; the kernel has limited to none knowlege about the semantics of mapping sequences. Hence, merging is purely speculative. It seems that gaining a significant (or at least a measurable) performance increase through merging is way more likely to happen when userspace is responsible for merging mappings up to the next larger page size if possible.
- Since merging was removed, regions pretty much loose their right to exist. They might still be useful for handling dual page tables or similar mechanisms, but since Nouveau seems to be the only driver having a need for this for now, regions were removed from the GPUVA manager.
- Fixed a couple of maple_tree related issues; thanks to Liam for helping me out.

Changes in V4:
==============
Nouveau:
- Refactored how specific VM_BIND and EXEC jobs are created and how their arguments are passed to the generic job implementation.
- Fixed a UAF race condition where bind job ops could have been freed already while still waiting for a job cleanup to finish. This is due to in certain cases we need to wait for mappings actually being unmapped before creating sparse regions in the same area.
- Re-based the code onto drm_exec v4 patch.

GPUVA Manager:
- Fixed a maple tree related bug when pre-allocating MA states. (Boris Brezillion)
- Made struct drm_gpuva_fn_ops a const object in all occurrences. (Boris Brezillion)

Changes in V5:
==============
Nouveau:
- Link and unlink GPUVAs outside the fence signalling critical path in nouveau_uvmm_bind_job_submit() holding the dma-resv lock. Mutual exclusion of BO evicts causing mapping invalidation and regular mapping operations is ensured with dma-fences.

GPUVA Manager:
- Removed the separate GEMs GPUVA list lock. Link and unlink as well as iterating the GEM's GPUVA list should be protected with the GEM's dma-resv lock instead.
- Renamed DRM_GPUVA_EVICTED flag to DRM_GPUVA_INVALIDATED. Mappings do not get eviced, they might get invalidated due to eviction.
- Maple tree uses the 'unsinged long' type for node entries. While this works for GPU VA spaces larger than 32-bit on 64-bit kernel, the GPU VA space is limited to 32-bit on 32-bit kernels as well.
As long as we do not have a 64-bit capable maple tree for 32-bit kernels, the GPU VA manager contains checks to throw warnings when GPU VA entries exceed the maple tree's storage capabilities.
- Extended the Documentation and added example code as requested by Donald Robson.

Quoted 7 lines

Christian König (1):
drm: execution context for GEM buffers v4

Danilo Krummrich (13):
maple_tree: split up MA_STATE() macro
drm: manager to keep track of GPUs VA mappings
drm: debugfs: provide infrastructure to dump a DRM GPU VA space

Core drm patches are

The only thing I'm worried about is the 'sync mapping requests have to go through the async path and wait for all previous async requests to be processed' problem I mentioned in one of your previous submission, but I'm happy leave that for later.

Quoted 76 lines

drm/nouveau: new VM_BIND uapi interfaces drm/nouveau: get vmm via nouveau_cli_vmm() drm/nouveau: bo: initialize GEM GPU VA interface drm/nouveau: move usercopy helpers to nouveau_drv.h drm/nouveau: fence: separate fence alloc and emit drm/nouveau: fence: fail to emit when fence context is killed drm/nouveau: chan: provide nouveau_channel_kill() drm/nouveau: nvkm/vmm: implement raw ops to manage uvmm drm/nouveau: implement new VM_BIND uAPI drm/nouveau: debugfs: implement DRM GPU VA debugfs

 Documentation/gpu/driver-uapi.rst             |   11 +
 Documentation/gpu/drm-mm.rst                  |   54 +
 drivers/gpu/drm/Kconfig                       |    6 +
 drivers/gpu/drm/Makefile                      |    3 +
 drivers/gpu/drm/drm_debugfs.c                 |   41 +
 drivers/gpu/drm/drm_exec.c                    |  278 +++
 drivers/gpu/drm/drm_gem.c                     |    3 +
 drivers/gpu/drm/drm_gpuva_mgr.c               | 1971 ++++++++++++++++
 drivers/gpu/drm/nouveau/Kbuild                |    3 +
 drivers/gpu/drm/nouveau/Kconfig               |    2 +
 drivers/gpu/drm/nouveau/dispnv04/crtc.c       |    9 +-
 drivers/gpu/drm/nouveau/include/nvif/if000c.h |   26 +-
 drivers/gpu/drm/nouveau/include/nvif/vmm.h    |   19 +-
 .../gpu/drm/nouveau/include/nvkm/subdev/mmu.h |   20 +-
 drivers/gpu/drm/nouveau/nouveau_abi16.c       |   24 +
 drivers/gpu/drm/nouveau/nouveau_abi16.h       |    1 +
 drivers/gpu/drm/nouveau/nouveau_bo.c          |  204 +-
 drivers/gpu/drm/nouveau/nouveau_bo.h          |    2 +-
 drivers/gpu/drm/nouveau/nouveau_chan.c        |   22 +-
 drivers/gpu/drm/nouveau/nouveau_chan.h        |    1 +
 drivers/gpu/drm/nouveau/nouveau_debugfs.c     |   39 +
 drivers/gpu/drm/nouveau/nouveau_dmem.c        |    9 +-
 drivers/gpu/drm/nouveau/nouveau_drm.c         |   27 +-
 drivers/gpu/drm/nouveau/nouveau_drv.h         |   94 +-
 drivers/gpu/drm/nouveau/nouveau_exec.c        |  418 ++++
 drivers/gpu/drm/nouveau/nouveau_exec.h        |   54 +
 drivers/gpu/drm/nouveau/nouveau_fence.c       |   23 +-
 drivers/gpu/drm/nouveau/nouveau_fence.h       |    5 +-
 drivers/gpu/drm/nouveau/nouveau_gem.c         |   62 +-
 drivers/gpu/drm/nouveau/nouveau_mem.h         |    5 +
 drivers/gpu/drm/nouveau/nouveau_prime.c       |    2 +-
 drivers/gpu/drm/nouveau/nouveau_sched.c       |  461 ++++
 drivers/gpu/drm/nouveau/nouveau_sched.h       |  123 +
 drivers/gpu/drm/nouveau/nouveau_svm.c         |    2 +-
 drivers/gpu/drm/nouveau/nouveau_uvmm.c        | 1979 +++++++++++++++++
 drivers/gpu/drm/nouveau/nouveau_uvmm.h        |  107 +
 drivers/gpu/drm/nouveau/nouveau_vmm.c         |    4 +-
 drivers/gpu/drm/nouveau/nvif/vmm.c            |  100 +-
 .../gpu/drm/nouveau/nvkm/subdev/mmu/uvmm.c    |  213 +-
 drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c |  197 +-
 drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h |   25 +
 .../drm/nouveau/nvkm/subdev/mmu/vmmgf100.c    |   16 +-
 .../drm/nouveau/nvkm/subdev/mmu/vmmgp100.c    |   16 +-
 .../gpu/drm/nouveau/nvkm/subdev/mmu/vmmnv50.c |   27 +-
 include/drm/drm_debugfs.h                     |   25 +
 include/drm/drm_drv.h                         |    6 +
 include/drm/drm_exec.h                        |  119 +
 include/drm/drm_gem.h                         |   52 +
 include/drm/drm_gpuva_mgr.h                   |  682 ++++++
 include/linux/maple_tree.h                    |    7 +-
 include/uapi/drm/nouveau_drm.h                |  209 ++
 51 files changed, 7566 insertions(+), 242 deletions(-)
 create mode 100644 drivers/gpu/drm/drm_exec.c
 create mode 100644 drivers/gpu/drm/drm_gpuva_mgr.c
 create mode 100644 drivers/gpu/drm/nouveau/nouveau_exec.c
 create mode 100644 drivers/gpu/drm/nouveau/nouveau_exec.h
 create mode 100644 drivers/gpu/drm/nouveau/nouveau_sched.c
 create mode 100644 drivers/gpu/drm/nouveau/nouveau_sched.h
 create mode 100644 drivers/gpu/drm/nouveau/nouveau_uvmm.c
 create mode 100644 drivers/gpu/drm/nouveau/nouveau_uvmm.h
 create mode 100644 include/drm/drm_exec.h
 create mode 100644 include/drm/drm_gpuva_mgr.h

base-commit: 2222dcb0775d36de28992f56455ab3967b30d380

#38 lore raw

Hi Boris,

Quoted 196 lines · Boris Brezillon

Hi Danilo,

Quoted 183 lines · On Tue, 20 Jun 2023 02:42:03 +0200 Danilo Krummrich <[email protected]> wrote:

This patch series provides a new UAPI for the Nouveau driver in order to support Vulkan features, such as sparse bindings and sparse residency.

Furthermore, with the DRM GPUVA manager it provides a new DRM core feature to keep track of GPU virtual address (VA) mappings in a more generic way.

The DRM GPUVA manager is indented to help drivers implement userspace-manageable GPU VA spaces in reference to the Vulkan API. In order to achieve this goal it serves the following purposes in this context.

1) Provide infrastructure to track GPU VA allocations and mappings, making use of the maple_tree.

2) Generically connect GPU VA mappings to their backing buffers, in particular DRM GEM objects.

3) Provide a common implementation to perform more complex mapping operations on the GPU VA space. In particular splitting and merging of GPU VA mappings, e.g. for intersecting mapping requests or partial unmap requests.

The new VM_BIND Nouveau UAPI build on top of the DRM GPUVA manager, itself providing the following new interfaces.

1) Initialize a GPU VA space via the new DRM_IOCTL_NOUVEAU_VM_INIT ioctl for UMDs to specify the portion of VA space managed by the kernel and userspace, respectively.

2) Allocate and free a VA space region as well as bind and unbind memory to the GPUs VA space via the new DRM_IOCTL_NOUVEAU_VM_BIND ioctl.

3) Execute push buffers with the new DRM_IOCTL_NOUVEAU_EXEC ioctl.

Both, DRM_IOCTL_NOUVEAU_VM_BIND and DRM_IOCTL_NOUVEAU_EXEC, make use of the DRM scheduler to queue jobs and support asynchronous processing with DRM syncobjs as synchronization mechanism.

By default DRM_IOCTL_NOUVEAU_VM_BIND does synchronous processing, DRM_IOCTL_NOUVEAU_EXEC supports asynchronous processing only.

The new VM_BIND UAPI for Nouveau makes also use of drm_exec (execution context for GEM buffers) by Christian König. Since the patch implementing drm_exec was not yet merged into drm-next it is part of this series, as well as a small fix for this patch, which was found while testing this series.

This patch series is also available at [1].

There is a Mesa NVK merge request by Dave Airlie [2] implementing the corresponding userspace parts for this series.

The Vulkan CTS test suite passes the sparse binding and sparse residency test cases for the new UAPI together with Dave's Mesa work.

There are also some test cases in the igt-gpu-tools project [3] for the new UAPI and hence the DRM GPU VA manager. However, most of them are testing the DRM GPU VA manager's logic through Nouveau's new UAPI and should be considered just as helper for implementation.

However, I absolutely intend to change those test cases to proper kunit test cases for the DRM GPUVA manager, once and if we agree on it's usefulness and design.

[1] https://gitlab.freedesktop.org/nouvelles/kernel/-/tree/new-uapi-drm-next /
https://gitlab.freedesktop.org/nouvelles/kernel/-/merge_requests/1
[2] https://gitlab.freedesktop.org/nouveau/mesa/-/merge_requests/150/
[3] https://gitlab.freedesktop.org/dakr/igt-gpu-tools/-/tree/wip_nouveau_vm_bind

Changes in V2:
==============
Nouveau:
- Reworked the Nouveau VM_BIND UAPI to avoid memory allocations in fence signalling critical sections. Updates to the VA space are split up in three separate stages, where only the 2. stage executes in a fence signalling critical section:

1. update the VA space, allocate new structures and page tables
2. (un-)map the requested memory bindings
3. free structures and page tables

- Separated generic job scheduler code from specific job implementations.
- Separated the EXEC and VM_BIND implementation of the UAPI.
- Reworked the locking parts of the nvkm/vmm RAW interface, such that (un-)map operations can be executed in fence signalling critical sections.

GPUVA Manager:
- made drm_gpuva_regions optional for users of the GPUVA manager
- allow NULL GEMs for drm_gpuva entries
- swichted from drm_mm to maple_tree for track drm_gpuva / drm_gpuva_region entries
- provide callbacks for users to allocate custom drm_gpuva_op structures to allow inheritance
- added user bits to drm_gpuva_flags
- added a prefetch operation type in order to support generating prefetch operations in the same way other operations generated
- hand the responsibility for mutual exclusion for a GEM's drm_gpuva list to the user; simplified corresponding (un-)link functions

Maple Tree:
- I added two maple tree patches to the series, one to support custom tree walk macros and one to hand the locking responsibility to the user of the GPUVA manager without pre-defined lockdep checks.

Changes in V3:
==============
Nouveau:
- Reworked the Nouveau VM_BIND UAPI to do the job cleanup (including page table cleanup) within a workqueue rather than the job_free() callback of the scheduler itself. A job_free() callback can stall the execution (run() callback) of the next job in the queue. Since the page table cleanup requires to take the same locks as need to be taken for page table allocation, doing it directly in the job_free() callback would still violate the fence signalling critical path.
- Separated Nouveau fence allocation and emit, such that we do not violate the fence signalling critical path in EXEC jobs.
- Implement "regions" (for handling sparse mappings through PDEs and dual page tables) within Nouveau.
- Drop the requirement for every mapping to be contained within a region.
- Add necassary synchronization of VM_BIND job operation sequences in order to work around limitations in page table handling. This will be addressed in a future re-work of Nouveau's page table handling.
- Fixed a couple of race conditions found through more testing. Thanks to Dave for consitently trying to break it. :-)

GPUVA Manager:
- Implement pre-allocation capabilities for tree modifications within fence signalling critical sections.
- Implement accessors to to apply tree modification while walking the GPUVA tree in order to actually support processing of drm_gpuva_ops through callbacks in fence signalling critical sections rather than through pre-allocated operation lists.
- Remove merging of GPUVAs; the kernel has limited to none knowlege about the semantics of mapping sequences. Hence, merging is purely speculative. It seems that gaining a significant (or at least a measurable) performance increase through merging is way more likely to happen when userspace is responsible for merging mappings up to the next larger page size if possible.
- Since merging was removed, regions pretty much loose their right to exist. They might still be useful for handling dual page tables or similar mechanisms, but since Nouveau seems to be the only driver having a need for this for now, regions were removed from the GPUVA manager.
- Fixed a couple of maple_tree related issues; thanks to Liam for helping me out.

Changes in V4:
==============
Nouveau:
- Refactored how specific VM_BIND and EXEC jobs are created and how their arguments are passed to the generic job implementation.
- Fixed a UAF race condition where bind job ops could have been freed already while still waiting for a job cleanup to finish. This is due to in certain cases we need to wait for mappings actually being unmapped before creating sparse regions in the same area.
- Re-based the code onto drm_exec v4 patch.

GPUVA Manager:
- Fixed a maple tree related bug when pre-allocating MA states. (Boris Brezillion)
- Made struct drm_gpuva_fn_ops a const object in all occurrences. (Boris Brezillion)

Changes in V5:
==============
Nouveau:
- Link and unlink GPUVAs outside the fence signalling critical path in nouveau_uvmm_bind_job_submit() holding the dma-resv lock. Mutual exclusion of BO evicts causing mapping invalidation and regular mapping operations is ensured with dma-fences.

GPUVA Manager:
- Removed the separate GEMs GPUVA list lock. Link and unlink as well as iterating the GEM's GPUVA list should be protected with the GEM's dma-resv lock instead.
- Renamed DRM_GPUVA_EVICTED flag to DRM_GPUVA_INVALIDATED. Mappings do not get eviced, they might get invalidated due to eviction.
- Maple tree uses the 'unsinged long' type for node entries. While this works for GPU VA spaces larger than 32-bit on 64-bit kernel, the GPU VA space is limited to 32-bit on 32-bit kernels as well.
As long as we do not have a 64-bit capable maple tree for 32-bit kernels, the GPU VA manager contains checks to throw warnings when GPU VA entries exceed the maple tree's storage capabilities.
- Extended the Documentation and added example code as requested by Donald Robson.

Quoted 7 lines

Christian König (1):
drm: execution context for GEM buffers v4

Danilo Krummrich (13):
maple_tree: split up MA_STATE() macro
drm: manager to keep track of GPUs VA mappings
drm: debugfs: provide infrastructure to dump a DRM GPU VA space

Quoted 8 lines

Core drm patches are

The only thing I'm worried about is the 'sync mapping requests have to go through the async path and wait for all previous async requests to be processed' problem I mentioned in one of your previous submission, but I'm happy leave that for later.

Yes, I'm aware of this limitation.

Let me quickly try to explain where this limitation comes from and how I intend to address it.

In order to be able to allocate the required page tables for a mapping request and in order to free corresponding page tables once the (async) job finished I need to know the corresponding sequence of operations (drm_gpuva_ops) to fulfill the mapping request.

This requires me to update the GPUVA space in the ioctl() rather than in the async stage, because otherwise I would need to wait for previous jobs to finish before being able to submit subsequent jobs to the job queue, since I need an up to date view of the GPUVA space in order to calculate the sequence of operations to fulfill a mapping request.

As a consequence all jobs need to be processed in the order they were submitted, including synchronous jobs.

@Matt: I think you will have the same limitation with synchronous jobs as your implementation in XE should be similar?

In order to address it I want to switch to using callbacks rather than 'pre-allocated' drm_gpuva_ops and update the GPUVA space within the asynchronous stage. This would allow me to 'fit' synchronous jobs between jobs waiting in the async job queue. However, to do this I have to re-work how the page table handling in Nouveau is implemented, since this would require me to be able to manage the page tables without knowing the exact sequence of operations to fulfill a mapping request.

- Danilo

Quoted 78 lines
Quoted 77 lines

drm/nouveau: new VM_BIND uapi interfaces drm/nouveau: get vmm via nouveau_cli_vmm() drm/nouveau: bo: initialize GEM GPU VA interface drm/nouveau: move usercopy helpers to nouveau_drv.h drm/nouveau: fence: separate fence alloc and emit drm/nouveau: fence: fail to emit when fence context is killed drm/nouveau: chan: provide nouveau_channel_kill() drm/nouveau: nvkm/vmm: implement raw ops to manage uvmm drm/nouveau: implement new VM_BIND uAPI drm/nouveau: debugfs: implement DRM GPU VA debugfs

  Documentation/gpu/driver-uapi.rst             |   11 +
  Documentation/gpu/drm-mm.rst                  |   54 +
  drivers/gpu/drm/Kconfig                       |    6 +
  drivers/gpu/drm/Makefile                      |    3 +
  drivers/gpu/drm/drm_debugfs.c                 |   41 +
  drivers/gpu/drm/drm_exec.c                    |  278 +++
  drivers/gpu/drm/drm_gem.c                     |    3 +
  drivers/gpu/drm/drm_gpuva_mgr.c               | 1971 ++++++++++++++++
  drivers/gpu/drm/nouveau/Kbuild                |    3 +
  drivers/gpu/drm/nouveau/Kconfig               |    2 +
  drivers/gpu/drm/nouveau/dispnv04/crtc.c       |    9 +-
  drivers/gpu/drm/nouveau/include/nvif/if000c.h |   26 +-
  drivers/gpu/drm/nouveau/include/nvif/vmm.h    |   19 +-
  .../gpu/drm/nouveau/include/nvkm/subdev/mmu.h |   20 +-
  drivers/gpu/drm/nouveau/nouveau_abi16.c       |   24 +
  drivers/gpu/drm/nouveau/nouveau_abi16.h       |    1 +
  drivers/gpu/drm/nouveau/nouveau_bo.c          |  204 +-
  drivers/gpu/drm/nouveau/nouveau_bo.h          |    2 +-
  drivers/gpu/drm/nouveau/nouveau_chan.c        |   22 +-
  drivers/gpu/drm/nouveau/nouveau_chan.h        |    1 +
  drivers/gpu/drm/nouveau/nouveau_debugfs.c     |   39 +
  drivers/gpu/drm/nouveau/nouveau_dmem.c        |    9 +-
  drivers/gpu/drm/nouveau/nouveau_drm.c         |   27 +-
  drivers/gpu/drm/nouveau/nouveau_drv.h         |   94 +-
  drivers/gpu/drm/nouveau/nouveau_exec.c        |  418 ++++
  drivers/gpu/drm/nouveau/nouveau_exec.h        |   54 +
  drivers/gpu/drm/nouveau/nouveau_fence.c       |   23 +-
  drivers/gpu/drm/nouveau/nouveau_fence.h       |    5 +-
  drivers/gpu/drm/nouveau/nouveau_gem.c         |   62 +-
  drivers/gpu/drm/nouveau/nouveau_mem.h         |    5 +
  drivers/gpu/drm/nouveau/nouveau_prime.c       |    2 +-
  drivers/gpu/drm/nouveau/nouveau_sched.c       |  461 ++++
  drivers/gpu/drm/nouveau/nouveau_sched.h       |  123 +
  drivers/gpu/drm/nouveau/nouveau_svm.c         |    2 +-
  drivers/gpu/drm/nouveau/nouveau_uvmm.c        | 1979 +++++++++++++++++
  drivers/gpu/drm/nouveau/nouveau_uvmm.h        |  107 +
  drivers/gpu/drm/nouveau/nouveau_vmm.c         |    4 +-
  drivers/gpu/drm/nouveau/nvif/vmm.c            |  100 +-
  .../gpu/drm/nouveau/nvkm/subdev/mmu/uvmm.c    |  213 +-
  drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c |  197 +-
  drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h |   25 +
  .../drm/nouveau/nvkm/subdev/mmu/vmmgf100.c    |   16 +-
  .../drm/nouveau/nvkm/subdev/mmu/vmmgp100.c    |   16 +-
  .../gpu/drm/nouveau/nvkm/subdev/mmu/vmmnv50.c |   27 +-
  include/drm/drm_debugfs.h                     |   25 +
  include/drm/drm_drv.h                         |    6 +
  include/drm/drm_exec.h                        |  119 +
  include/drm/drm_gem.h                         |   52 +
  include/drm/drm_gpuva_mgr.h                   |  682 ++++++
  include/linux/maple_tree.h                    |    7 +-
  include/uapi/drm/nouveau_drm.h                |  209 ++
  51 files changed, 7566 insertions(+), 242 deletions(-)

create mode 100644 drivers/gpu/drm/drm_exec.c create mode 100644 drivers/gpu/drm/drm_gpuva_mgr.c create mode 100644 drivers/gpu/drm/nouveau/nouveau_exec.c create mode 100644 drivers/gpu/drm/nouveau/nouveau_exec.h create mode 100644 drivers/gpu/drm/nouveau/nouveau_sched.c create mode 100644 drivers/gpu/drm/nouveau/nouveau_sched.h create mode 100644 drivers/gpu/drm/nouveau/nouveau_uvmm.c create mode 100644 drivers/gpu/drm/nouveau/nouveau_uvmm.h create mode 100644 include/drm/drm_exec.h
create mode 100644 include/drm/drm_gpuva_mgr.h

base-commit: 2222dcb0775d36de28992f56455ab3967b30d380

#39 lore raw

Hi Danilo,

Quoted 27 lines · On Tue, 20 Jun 2023 14:46:07 +0200 Danilo Krummrich <[email protected]> wrote:
Quoted 4 lines

The only thing I'm worried about is the 'sync mapping requests have to go through the async path and wait for all previous async requests to be processed' problem I mentioned in one of your previous submission, but I'm happy leave that for later.

Yes, I'm aware of this limitation.

Let me quickly try to explain where this limitation comes from and how I intend to address it.

In order to be able to allocate the required page tables for a mapping request and in order to free corresponding page tables once the (async) job finished I need to know the corresponding sequence of operations (drm_gpuva_ops) to fulfill the mapping request.

This requires me to update the GPUVA space in the ioctl() rather than in the async stage, because otherwise I would need to wait for previous jobs to finish before being able to submit subsequent jobs to the job queue, since I need an up to date view of the GPUVA space in order to calculate the sequence of operations to fulfill a mapping request.

As a consequence all jobs need to be processed in the order they were submitted, including synchronous jobs.

@Matt: I think you will have the same limitation with synchronous jobs as your implementation in XE should be similar?

Quoted 8 lines

In order to address it I want to switch to using callbacks rather than 'pre-allocated' drm_gpuva_ops and update the GPUVA space within the asynchronous stage.
This would allow me to 'fit' synchronous jobs
between jobs waiting in the async job queue. However, to do this I have to re-work how the page table handling in Nouveau is implemented, since this would require me to be able to manage the page tables without knowing the exact sequence of operations to fulfill a mapping request.

Ok, so I think that's more or less what we're trying to do right now in PowerVR.

- First, we make sure we reserve enough MMU page tables for a given map operation to succeed no matter the VM state in the VM_BIND job submission path (our VM_BIND ioctl). That means we're always over-provisioning and returning unused memory back when the operation is done if we end up using less memory.
- We pre-allocate for the mapple-tree insertions.
- Then we map using drm_gpuva_sm_map() and the callbacks we provided in the drm_sched::run_job() path. We guarantee that no memory is allocated in that path thanks to the pre-allocation/reservation we've done at VM_BIND job submission time.

The problem I see with this v5 is that:

1/ We now have a dma_resv_lock_held() in drm_gpuva_{link,unlink}(),
which, in our case, is called in the async drm_sched::run_job() path, and we don't hold the lock in that path (it's been released just after the job submission).
2/ I'm worried that Liam's plan to only reserve what's actually needed
based on the mapple tree state is going to play against us, because the mapple-tree is only modified at job exec time, and we might have several unmaps happening between the moment we created and queued the jobs, and the moment they actually get executed, meaning the mapple-tree reservation might no longer fit the bill.

For issue #1, it shouldn't be to problematic if we use a regular lock to insert to/remove from the GEM gpuva list.

For issue #2, I can see a way out if, instead of freeing gpuva nodes, we flag those as unused when we see that something happening later in the queue is going to map a section being unmapped. All of this implies keeping access to already queued VM_BIND jobs (using the spsc queue at the entity level is not practical), and iterating over them every time a new sync or async job is queued to flag what needs to be retained. It would obviously be easier if we could tell the mapple-tree API 'provision as if the tree was empty', so all we have to do is just over-provision for both the page tables and mapple-tree insertion, and free the unused mem when the operation is done.

Don't know if you already thought about that and/or have solutions to solve these issues.

Regards,

Boris

#40 lore raw

Hi Boris,

Quoted 55 lines · Boris Brezillon

Hi Danilo,

Quoted 27 lines · On Tue, 20 Jun 2023 14:46:07 +0200 Danilo Krummrich <[email protected]> wrote:
Quoted 4 lines

The only thing I'm worried about is the 'sync mapping requests have to go through the async path and wait for all previous async requests to be processed' problem I mentioned in one of your previous submission, but I'm happy leave that for later.

Yes, I'm aware of this limitation.

Let me quickly try to explain where this limitation comes from and how I intend to address it.

In order to be able to allocate the required page tables for a mapping request and in order to free corresponding page tables once the (async) job finished I need to know the corresponding sequence of operations (drm_gpuva_ops) to fulfill the mapping request.

This requires me to update the GPUVA space in the ioctl() rather than in the async stage, because otherwise I would need to wait for previous jobs to finish before being able to submit subsequent jobs to the job queue, since I need an up to date view of the GPUVA space in order to calculate the sequence of operations to fulfill a mapping request.

As a consequence all jobs need to be processed in the order they were submitted, including synchronous jobs.

@Matt: I think you will have the same limitation with synchronous jobs as your implementation in XE should be similar?

Quoted 8 lines

In order to address it I want to switch to using callbacks rather than 'pre-allocated' drm_gpuva_ops and update the GPUVA space within the asynchronous stage.
This would allow me to 'fit' synchronous jobs
between jobs waiting in the async job queue. However, to do this I have to re-work how the page table handling in Nouveau is implemented, since this would require me to be able to manage the page tables without knowing the exact sequence of operations to fulfill a mapping request.

Ok, so I think that's more or less what we're trying to do right now in PowerVR.

- First, we make sure we reserve enough MMU page tables for a given map operation to succeed no matter the VM state in the VM_BIND job submission path (our VM_BIND ioctl). That means we're always over-provisioning and returning unused memory back when the operation is done if we end up using less memory.
- We pre-allocate for the mapple-tree insertions.
- Then we map using drm_gpuva_sm_map() and the callbacks we provided in the drm_sched::run_job() path. We guarantee that no memory is allocated in that path thanks to the pre-allocation/reservation we've done at VM_BIND job submission time.

Quoted 6 lines

The problem I see with this v5 is that:

1/ We now have a dma_resv_lock_held() in drm_gpuva_{link,unlink}(),
which, in our case, is called in the async drm_sched::run_job() path, and we don't hold the lock in that path (it's been released just after the job submission).

My solution to this, as by now, is to - in the same way we pre-allocate - to just pre-link and pre-unlink. And then fix things up in the cleanup path.

However, depending on the driver, this might require you to set a flag in the driver specific structure (embedding struct drm_gpuva) whether the gpuva is actually mapped (as in has active page table entries). Maybe we could also just add such a flag to struct drm_gpuva. But yeah, doesn't sound too nice to be honest...

Quoted 6 lines

2/ I'm worried that Liam's plan to only reserve what's actually needed
based on the mapple tree state is going to play against us, because the mapple-tree is only modified at job exec time, and we might have several unmaps happening between the moment we created and queued the jobs, and the moment they actually get executed, meaning the mapple-tree reservation might no longer fit the bill.

Yes, I'm aware and I explained to Liam in detail why we need the mas_preallocate_worst_case() way of doing it.

See this mail: https://lore.kernel.org/nouveau/[email protected]/T/#ma326e200b1de1e3c9df4e9fcb3bf243061fee8b5

He hasn't answered yet, but I hope we can just get (or actually keep) such a function (hopefully with better naming), since it shouldn't interfere with anything else.

Quoted 3 lines

For issue #1, it shouldn't be to problematic if we use a regular lock to insert to/remove from the GEM gpuva list.

Yes, that's why I had a separate GEM gpuva list lock in the first place. However, this doesn't really work when generating ops rather than using the callback interface.

Have a look at drm_gpuva_gem_unmap_ops_create() requested by Matt for XE. This function generates drm_gpuva_ops to unmap all mappings of a given GEM. In order to do that the function must iterate the GEM's gpuva list and allocate operations for each mapping. As a consequence the gpuva list lock wouldn't be allowed to be taken in the fence signalling path (run_job()) any longer. Hence, we can just protect the list with the GEM's dma-resv lock.

However, I can understand that it might be inconvenient for the callback interface and admittedly my solution to that isn't that nice as well. Hence the following idea:

For drivers to be able to use their own lock for that it would be enough to get rid of the lockdep checks. We could just add a flag to the GPUVA manager to let the driver indicate it wants to do it's own locking for the GPUVA list and skip the lockdep checks for the dma-resv lock in that case.

Quoted 12 lines

For issue #2, I can see a way out if, instead of freeing gpuva nodes, we flag those as unused when we see that something happening later in the queue is going to map a section being unmapped. All of this implies keeping access to already queued VM_BIND jobs (using the spsc queue at the entity level is not practical), and iterating over them every time a new sync or async job is queued to flag what needs to be retained. It would obviously be easier if we could tell the mapple-tree API 'provision as if the tree was empty', so all we have to do is just over-provision for both the page tables and mapple-tree insertion, and free the unused mem when the operation is done.

Quoted 2 lines

Don't know if you already thought about that and/or have solutions to solve these issues.

As already mentioned above, I'd just expect we can keep it the over-provision way, as you say. I think it's a legit use case to not know the state of the maple tree at the time the pre-allocated nodes will be used and keeping that should not interfere with Liams plan to (hopefully separately) optimize for the pre-allocation use case they have within -mm.

But let's wait for his take on that.

- Danilo

Quoted 5 lines

Regards,

Boris

#41 lore raw

Hi Danilo,

Quoted 85 lines · On Thu, 22 Jun 2023 15:58:23 +0200 Danilo Krummrich <[email protected]> wrote:

Hi Boris,

Quoted 55 lines · Boris Brezillon

Hi Danilo,

Quoted 27 lines · On Tue, 20 Jun 2023 14:46:07 +0200 Danilo Krummrich <[email protected]> wrote:
Quoted 4 lines

The only thing I'm worried about is the 'sync mapping requests have to go through the async path and wait for all previous async requests to be processed' problem I mentioned in one of your previous submission, but I'm happy leave that for later.

Yes, I'm aware of this limitation.

Let me quickly try to explain where this limitation comes from and how I intend to address it.

In order to be able to allocate the required page tables for a mapping request and in order to free corresponding page tables once the (async) job finished I need to know the corresponding sequence of operations (drm_gpuva_ops) to fulfill the mapping request.

This requires me to update the GPUVA space in the ioctl() rather than in the async stage, because otherwise I would need to wait for previous jobs to finish before being able to submit subsequent jobs to the job queue, since I need an up to date view of the GPUVA space in order to calculate the sequence of operations to fulfill a mapping request.

As a consequence all jobs need to be processed in the order they were submitted, including synchronous jobs.

@Matt: I think you will have the same limitation with synchronous jobs as your implementation in XE should be similar?

Quoted 8 lines

In order to address it I want to switch to using callbacks rather than 'pre-allocated' drm_gpuva_ops and update the GPUVA space within the asynchronous stage.
This would allow me to 'fit' synchronous jobs
between jobs waiting in the async job queue. However, to do this I have to re-work how the page table handling in Nouveau is implemented, since this would require me to be able to manage the page tables without knowing the exact sequence of operations to fulfill a mapping request.

Ok, so I think that's more or less what we're trying to do right now in PowerVR.

- First, we make sure we reserve enough MMU page tables for a given map operation to succeed no matter the VM state in the VM_BIND job submission path (our VM_BIND ioctl). That means we're always over-provisioning and returning unused memory back when the operation is done if we end up using less memory.
- We pre-allocate for the mapple-tree insertions.
- Then we map using drm_gpuva_sm_map() and the callbacks we provided in the drm_sched::run_job() path. We guarantee that no memory is allocated in that path thanks to the pre-allocation/reservation we've done at VM_BIND job submission time.

Quoted 6 lines

The problem I see with this v5 is that:

1/ We now have a dma_resv_lock_held() in drm_gpuva_{link,unlink}(),
which, in our case, is called in the async drm_sched::run_job() path, and we don't hold the lock in that path (it's been released just after the job submission).

My solution to this, as by now, is to - in the same way we pre-allocate
- to just pre-link and pre-unlink. And then fix things up in the cleanup path.

However, depending on the driver, this might require you to set a flag in the driver specific structure (embedding struct drm_gpuva) whether the gpuva is actually mapped (as in has active page table entries). Maybe we could also just add such a flag to struct drm_gpuva. But yeah, doesn't sound too nice to be honest...

Quoted 6 lines

2/ I'm worried that Liam's plan to only reserve what's actually needed
based on the mapple tree state is going to play against us, because the mapple-tree is only modified at job exec time, and we might have several unmaps happening between the moment we created and queued the jobs, and the moment they actually get executed, meaning the mapple-tree reservation might no longer fit the bill.

Yes, I'm aware and I explained to Liam in detail why we need the mas_preallocate_worst_case() way of doing it.

Quoted 6 lines

See this mail:
https://lore.kernel.org/nouveau/[email protected]/T/#ma326e200b1de1e3c9df4e9fcb3bf243061fee8b5

He hasn't answered yet, but I hope we can just get (or actually keep) such a function (hopefully with better naming), since it shouldn't interfere with anything else.

My bad, I started reading your reply and got interrupted. Never got back to it, which I should definitely have done before posting my questions. Anyway, glad to hear we're on the same page regarding the mas_preallocate_worst_case() thing.

Quoted 9 lines
Quoted 4 lines

For issue #1, it shouldn't be to problematic if we use a regular lock to insert to/remove from the GEM gpuva list.

Yes, that's why I had a separate GEM gpuva list lock in the first place. However, this doesn't really work when generating ops rather than using the callback interface.

Quoted 7 lines

Have a look at drm_gpuva_gem_unmap_ops_create() requested by Matt for XE. This function generates drm_gpuva_ops to unmap all mappings of a given GEM. In order to do that the function must iterate the GEM's gpuva list and allocate operations for each mapping. As a consequence the gpuva list lock wouldn't be allowed to be taken in the fence signalling path (run_job()) any longer. Hence, we can just protect the list with the GEM's dma-resv lock.

Yeah, I see why using dma_resv when pre-inserting the mapping is useful, it just didn't really work with late mapping insertion.

Quoted 10 lines

However, I can understand that it might be inconvenient for the callback interface and admittedly my solution to that isn't that nice as well. Hence the following idea:

For drivers to be able to use their own lock for that it would be enough to get rid of the lockdep checks. We could just add a flag to the GPUVA manager to let the driver indicate it wants to do it's own locking for the GPUVA list and skip the lockdep checks for the dma-resv lock in that case.

Sounds good to me.

Quoted 16 lines
Quoted 13 lines

For issue #2, I can see a way out if, instead of freeing gpuva nodes, we flag those as unused when we see that something happening later in the queue is going to map a section being unmapped. All of this implies keeping access to already queued VM_BIND jobs (using the spsc queue at the entity level is not practical), and iterating over them every time a new sync or async job is queued to flag what needs to be retained. It would obviously be easier if we could tell the mapple-tree API 'provision as if the tree was empty', so all we have to do is just over-provision for both the page tables and mapple-tree insertion, and free the unused mem when the operation is done.

Quoted 2 lines

Don't know if you already thought about that and/or have solutions to solve these issues.

Quoted 8 lines

As already mentioned above, I'd just expect we can keep it the over-provision way, as you say. I think it's a legit use case to not know the state of the maple tree at the time the pre-allocated nodes will be used and keeping that should not interfere with Liams plan to (hopefully separately) optimize for the pre-allocation use case they have within -mm.

But let's wait for his take on that.

Sure. As I said, I'm fine getting this version merged, we can sort out the changes needed for PowerVR later. Just thought I'd mention those issues early, so you're not surprised when we come back with crazy requests (which apparently are not that crazy ;-)).

Regards,

Boris

#42 lore raw
Quoted 93 lines · Boris Brezillon

Hi Danilo,

Quoted 88 lines · On Thu, 22 Jun 2023 15:58:23 +0200 Danilo Krummrich <[email protected]> wrote:

Hi Boris,

Quoted 55 lines · Boris Brezillon

Hi Danilo,

Quoted 27 lines · On Tue, 20 Jun 2023 14:46:07 +0200 Danilo Krummrich <[email protected]> wrote:
Quoted 4 lines

The only thing I'm worried about is the 'sync mapping requests have to go through the async path and wait for all previous async requests to be processed' problem I mentioned in one of your previous submission, but I'm happy leave that for later.

Yes, I'm aware of this limitation.

Let me quickly try to explain where this limitation comes from and how I intend to address it.

In order to be able to allocate the required page tables for a mapping request and in order to free corresponding page tables once the (async) job finished I need to know the corresponding sequence of operations (drm_gpuva_ops) to fulfill the mapping request.

This requires me to update the GPUVA space in the ioctl() rather than in the async stage, because otherwise I would need to wait for previous jobs to finish before being able to submit subsequent jobs to the job queue, since I need an up to date view of the GPUVA space in order to calculate the sequence of operations to fulfill a mapping request.

As a consequence all jobs need to be processed in the order they were submitted, including synchronous jobs.

@Matt: I think you will have the same limitation with synchronous jobs as your implementation in XE should be similar?

Quoted 8 lines

In order to address it I want to switch to using callbacks rather than 'pre-allocated' drm_gpuva_ops and update the GPUVA space within the asynchronous stage.
This would allow me to 'fit' synchronous jobs
between jobs waiting in the async job queue. However, to do this I have to re-work how the page table handling in Nouveau is implemented, since this would require me to be able to manage the page tables without knowing the exact sequence of operations to fulfill a mapping request.

Ok, so I think that's more or less what we're trying to do right now in PowerVR.

- First, we make sure we reserve enough MMU page tables for a given map operation to succeed no matter the VM state in the VM_BIND job submission path (our VM_BIND ioctl). That means we're always over-provisioning and returning unused memory back when the operation is done if we end up using less memory.
- We pre-allocate for the mapple-tree insertions.
- Then we map using drm_gpuva_sm_map() and the callbacks we provided in the drm_sched::run_job() path. We guarantee that no memory is allocated in that path thanks to the pre-allocation/reservation we've done at VM_BIND job submission time.

Quoted 6 lines

The problem I see with this v5 is that:

1/ We now have a dma_resv_lock_held() in drm_gpuva_{link,unlink}(),
which, in our case, is called in the async drm_sched::run_job() path, and we don't hold the lock in that path (it's been released just after the job submission).

My solution to this, as by now, is to - in the same way we pre-allocate
- to just pre-link and pre-unlink. And then fix things up in the cleanup path.

However, depending on the driver, this might require you to set a flag in the driver specific structure (embedding struct drm_gpuva) whether the gpuva is actually mapped (as in has active page table entries). Maybe we could also just add such a flag to struct drm_gpuva. But yeah, doesn't sound too nice to be honest...

Quoted 6 lines

2/ I'm worried that Liam's plan to only reserve what's actually needed
based on the mapple tree state is going to play against us, because the mapple-tree is only modified at job exec time, and we might have several unmaps happening between the moment we created and queued the jobs, and the moment they actually get executed, meaning the mapple-tree reservation might no longer fit the bill.

Yes, I'm aware and I explained to Liam in detail why we need the mas_preallocate_worst_case() way of doing it.

See this mail:
https://lore.kernel.org/nouveau/[email protected]/T/#ma326e200b1de1e3c9df4e9fcb3bf243061fee8b5

Quoted 8 lines
Quoted 3 lines

He hasn't answered yet, but I hope we can just get (or actually keep) such a function (hopefully with better naming), since it shouldn't interfere with anything else.

My bad, I started reading your reply and got interrupted. Never got back to it, which I should definitely have done before posting my questions. Anyway, glad to hear we're on the same page regarding the mas_preallocate_worst_case() thing.

No worries, I should probably also reply to Liams patch introducing the change. I will do that in a minute.

Quoted 26 lines
Quoted 10 lines
Quoted 5 lines

For issue #1, it shouldn't be to problematic if we use a regular lock to insert to/remove from the GEM gpuva list.

Yes, that's why I had a separate GEM gpuva list lock in the first place. However, this doesn't really work when generating ops rather than using the callback interface.

Quoted 7 lines

Have a look at drm_gpuva_gem_unmap_ops_create() requested by Matt for XE. This function generates drm_gpuva_ops to unmap all mappings of a given GEM. In order to do that the function must iterate the GEM's gpuva list and allocate operations for each mapping. As a consequence the gpuva list lock wouldn't be allowed to be taken in the fence signalling path (run_job()) any longer. Hence, we can just protect the list with the GEM's dma-resv lock.

Yeah, I see why using dma_resv when pre-inserting the mapping is useful, it just didn't really work with late mapping insertion.

Quoted 5 lines

However, I can understand that it might be inconvenient for the callback interface and admittedly my solution to that isn't that nice as well. Hence the following idea:

Quoted 7 lines
Quoted 5 lines

For drivers to be able to use their own lock for that it would be enough to get rid of the lockdep checks. We could just add a flag to the GPUVA manager to let the driver indicate it wants to do it's own locking for the GPUVA list and skip the lockdep checks for the dma-resv lock in that case.

Sounds good to me.

I think it's way better than the pre-link / pre-unlink mess. I will add this to v6.

Quoted 24 lines
Quoted 24 lines
Quoted 14 lines

For issue #2, I can see a way out if, instead of freeing gpuva nodes, we flag those as unused when we see that something happening later in the queue is going to map a section being unmapped. All of this implies keeping access to already queued VM_BIND jobs (using the spsc queue at the entity level is not practical), and iterating over them every time a new sync or async job is queued to flag what needs to be retained. It would obviously be easier if we could tell the mapple-tree API 'provision as if the tree was empty', so all we have to do is just over-provision for both the page tables and mapple-tree insertion, and free the unused mem when the operation is done.

Quoted 2 lines

Don't know if you already thought about that and/or have solutions to solve these issues.

As already mentioned above, I'd just expect we can keep it the over-provision way, as you say. I think it's a legit use case to not know the state of the maple tree at the time the pre-allocated nodes will be used and keeping that should not interfere with Liams plan to (hopefully separately) optimize for the pre-allocation use case they have within -mm.

Quoted 6 lines
Quoted 1 line

But let's wait for his take on that.

Sure. As I said, I'm fine getting this version merged, we can sort out the changes needed for PowerVR later. Just thought I'd mention those issues early, so you're not surprised when we come back with crazy requests (which apparently are not that crazy ;-)).

They're not crazy at all, in fact they entirely represent what the callback interface was designed for. :-)

- Danilo

Quoted 5 lines

Regards,

Boris