Re: [PATCH v10 0/5] drm/gpusvm: split MM and device state across gpusvm/range/pages

[Matthew Brost <matthew.brost@intel.com>]

On Tue, Jun 30, 2026 at 06:21:22PM +0800, Honglei Huang wrote:
> The intent of this series is to make drm_gpusvm more flexible and give
> drivers more freedom over how they assemble the MM related and device
> side operations. It implements the direction Matt suggested in [1]:
> Mirror MR in gitlab: [4]
> 
>   - Move struct drm_gpusvm_pages out of struct drm_gpusvm_range.
>   - Embed a struct drm_device in struct drm_gpusvm_pages and drive all
>     DMA through it.
>   - Drop struct drm_device from struct drm_gpusvm.
>   - Have the driver's range structure embed one or more struct
>     drm_gpusvm_pages in addition to struct drm_gpusvm_range.
>   - Drop the range-based helpers (drm_gpusvm_range_pages_valid,
>     drm_gpusvm_range_get_pages, drm_gpusvm_range_unmap_pages) and update
>     drivers to use the drm_gpusvm_pages helpers instead.
> 
> In essence the series does only two abstractions, plus the xe
> adaptation that follows from them:
> 
>   - range vs pages: split drm_gpusvm_range (MM / VA range state) from
>     drm_gpusvm_pages (device physical related), so the two sides can
>     have independent lifetimes and ownership.
>   - drm_gpusvm vs drm_device: make drm_gpusvm pure MM level and push
>     the device side down onto drm_gpusvm_pages, which is where DMA
>     actually happens.
>   - xe is updated to fit the modifications, no functional change
>     intended.
> 
> V10:
>   Fix two issues found by the AI review:
>   - patch 1: fix a KCSAN data race in xe_svm_alloc_vram(): read
>     range->base.flags.__flags with READ_ONCE() and assert on a local
>     copy, pairing with the WRITE_ONCE() in
>     drm_gpusvm_range_set_unmapped().
>   - patch 3: remove the DMA unmap in xe_svm_fini(): since
>     drm_gpusvm_range_remove() no longer unmaps synchronously, explicitly
>     drm_gpusvm_unmap_pages() all remaining ranges before
>     drm_gpusvm_fini().
> 
> V9:
>   - patch 3: fix the build with CONFIG_DRM_XE_GPUSVM disabled: move
>     pages out of the nested base struct in the stub xe_svm_range and
>     route has_dma_mapping through range->pages. No functional change.
> 
> V8:
>   - patch 4: add reviewed-by for Matt's review.
> 
> V7:
>   - patch 1: split MM state flags: the AI review found a KCSAN / memory
>     model cleanliness issue. Address it for consistency with
>     drm_gpusvm_pages_flags, set the range flags with WRITE_ONCE() on
>     __flags and read them with READ_ONCE().
> 
> V6:
>   - The AI review flagged a potential DMA free issue: the DMA unmap
>     step was moved into the range_free callback, but on the invalidate
>     path a range can be removed from the MMU interval tree while its DMA
>     mappings are still live, so a concurrent unmap event can miss it.
>   - patch 3: have xe_svm_range embed one drm_gpusvm_pages: explicitly
>     call drm_gpusvm_unmap_pages() before drm_gpusvm_range_remove() in the
>     garbage collector, so a range is never off the tree while still DMA
>     mapped, and document this caller contract in drm_gpusvm_range_remove()
>     kernel-doc.
>   - patch 4: move struct drm_gpusvm_pages out: document the
>     unmap before remove contract in the garbage collector example and
>     note that range_free()'s drm_gpusvm_free_pages() as a final fallback.
>   - patch 1: split MM state flags: return -EACCES directly.
>   - Fold in the pre existing IOVA/DMA unmap fixes the AI review found
>     previously sent separately: the uninitialized dma_addr[0].dir on
>     the get_pages() error path, the whole reservation IOVA free for
>     mixed ranges, and the device mapping leak on the get_pages() error
>     path. [6]
> 
> V5:
>   - add reviewed-by in patches 1, 2, 3, 5 for Matt's review.
> 
> V4:
>   - drm_gpusvm_init_pages(): memset() the pages to zero before recording
>     the owning drm_device.
>   - DOC: overview: recommend a zeroing allocator: kcalloc() for the
>     N:1 pages array.
>   - Rebased onto the latest drm-xe.
>   - The AI review of this series flagged two preexisting issues in the
>     IOVA unmap path that are not introduced by this series; they are
>     fixed in a separate series [5].
> 
> V3:
>   - Fix a kernel-doc/Sphinx warning from the kernel test robot: use
>     ".. code-block:: c" for the drm_gpusvm_pages example in DOC: overview.
>   - drm_gpusvm_range_set_unmapped(): use WRITE_ONCE() on the whole
>     pages[i].flags.__flags word to pair with the lockless READ_ONCE()
>     readers and avoid a data race.
>   - xe_userptr_setup(): call drm_gpusvm_init_pages() before
>     mmu_interval_notifier_insert() to avoid exposing uninitialized
>     pages.drm to invalidation callbacks.
>   - Fix per commit build of the set_unmapped() pages.
> 
> V2:
>   - Followed in Matt's v0 review fixups [2]:
>      - keep unmapped flag in pages structures.
>      - add pages_count to drm_gpusvm_range_set_unmapped() to set the pages
>        unmapped flag, so the framework can check unmapped status in
>        drm_gpusvm_get_pages().
>   - Add drm_gpusvm_init_pages to init the drm_device and sequence number.
>   - Remove drm_device from drm_gpusvm_get_pages() parameters.
>   - Reworked the DOC: overview and usage examples to describe the new
>     model: struct drm_gpusvm_pages, the 1:1 / N:1 driver layouts, and
>     examples that operate on a driver embedded pages object by the
>     drm_gpusvm_pages helpers and etc.
>   - remove WARN_ON_ONCE in __drm_gpusvm_unmap_pages.
>   - Dropped RFC.
> 
> Follow-up (not in this series):
> 
>   - modify drm_gpusvm_get_pages() to support one time hmm range fault
>     and multi drm device dma mapping.
>   - Add no dma device support for drm_gpusvm_get_pages().
> 
> tests:
> AMDGPU:
>   based on amdgpu adaptation patch in [3], but still SVM:DRM = 1:1,
>   1:n is on going needs many modifications and testings.
> 
>   Tested on gfx943 (MI300X) and gfx906 (MI60) with XNACK on/off:
>   - KFD test: 95%+ passed.
>   - ROCR test: all passed.
>   - HIP catch test: gfx943 (MI300X): 99% passed.
>                     gfx906 (MI60): 99% passed.
> INTEL XE:
>   CI tests passed in rev5 tests.[7]


CI full is in flight, but it seems likely to pass. Once it completes, we
can merge this; we just need to decide which branch to use.

The series does not apply cleanly to drm-misc-next because it depends on
changes in drm-xe-next that have not yet been merged into drm-misc-next.
It does apply cleanly to drm-xe-next, though.

If we merge the patches in drm-xe-next and pick them up via a rebase in
the AMD trees, that works. Alternatively, we can ask the drm-misc-next
maintainers to rebase and apply them there. Either option works for me.

Matt

> 
> links:
> [1] https://lore.kernel.org/amd-gfx/acRgr7QwdULsn6G2@gsse-cloud1/#:~:text=I%20think%20roughly,drm_gpusvm_pages%0A%20%20helpers%20instead.
> [2] https://gitlab.freedesktop.org/mbrost/xe-kernel-driver-svn-perf-6-15-2025/-/commit/623f6a50c037d9e44f6c9fbe6859a0ba7ad50177
> [3] https://lore.kernel.org/amd-gfx/20260603065030.2554403-1-honglei1.huang@amd.com/
> [4] https://gitlab.freedesktop.org/drm/xe/kernel/-/merge_requests/360
> [5] https://lore.kernel.org/all/20260627033325.3795298-1-honglei1.huang@amd.com/
> [6] https://lore.kernel.org/all/20260628061757.4093701-1-honglei1.huang@amd.com/
> [7] https://patchwork.freedesktop.org/series/169384/#rev5
> 
> Honglei Huang (5):
>   drm/gpusvm: split MM state flags out of drm_gpusvm_pages_flags
>   drm/gpusvm: embed struct drm_device into drm_gpusvm_pages
>   drm/xe: have xe_svm_range embed one drm_gpusvm_pages
>   drm/gpusvm: move struct drm_gpusvm_pages out of struct
>     drm_gpusvm_range
>   drm/gpusvm: let the drm_gpusvm core context purely MM level
> 
>  drivers/gpu/drm/drm_gpusvm.c    | 243 +++++++++++++++++++-------------
>  drivers/gpu/drm/xe/xe_pt.c      |   2 +-
>  drivers/gpu/drm/xe/xe_svm.c     |  66 +++++++--
>  drivers/gpu/drm/xe/xe_svm.h     |  14 +-
>  drivers/gpu/drm/xe/xe_userptr.c |   5 +-
>  include/drm/drm_gpusvm.h        |  67 ++++++---
>  6 files changed, 255 insertions(+), 142 deletions(-)
> 
> -- 
> 2.34.1
>