linux/drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c, branch v6.16

drm/amdgpu: Fail DMABUF map of XGMI-accessible memory

2025-05-01T15:01:46Z

If peer memory is XGMI-accessible, we should never access it through PCIe P2P DMA mappings. PCIe P2P is slower, has different coherence behaviour, limited or no support for atomics, or may not work at all. Fail with a warning if DMABUF mappings of such memory are attempted. Signed-off-by: Felix Kuehling Reviewed-by: Christian König Signed-off-by: Alex Deucher (cherry picked from commit dbe4c63689bc6b5fd3ab72650ea4b6a667e96a68)

drm/amdgpu: Allow P2P access through XGMI

2025-04-22T20:47:12Z

If peer memory is accessible through XGMI, allow leaving it in VRAM rather than forcing its migration to GTT on DMABuf attachment. Signed-off-by: Felix Kuehling Tested-by: Hao (Claire) Zhou Reviewed-by: Christian König Signed-off-by: Alex Deucher (cherry picked from commit 372c8d72c3680fdea3fbb2d6b089f76b4a6d596a)

drm/amdgpu: Don't pin VRAM without DMABUF_MOVE_NOTIFY

2025-04-22T20:44:28Z

Pinning of VRAM is for peer devices that don't support dynamic attachment and move notifiers. But it requires that all such peer devices are able to access VRAM via PCIe P2P. Any device without P2P access requires migration to GTT, which fails if the memory is already pinned for another peer device. Sharing between GPUs should not require pinning in VRAM. However, if DMABUF_MOVE_NOTIFY is disabled in the kernel build, even DMABufs shared between GPUs must be pinned, which can lead to failures and functional regressions on systems where some peer GPUs are not P2P accessible. Disable VRAM pinning if move notifiers are disabled in the kernel build to fix regressions when sharing BOs between GPUs. Signed-off-by: Felix Kuehling Tested-by: Hao (Claire) Zhou Reviewed-by: Christian König Signed-off-by: Alex Deucher (cherry picked from commit 05185812ae3695fe049c14847ce3cbeccff1bf2e)

drm/amdgpu: Use allowed_domains for pinning dmabufs

2025-04-22T20:44:02Z

When determining the domains for pinning DMABufs, filter allowed_domains and fail with a warning if VRAM is forbidden and GTT is not an allowed domain. Fixes: f5e7fabd1f5c ("drm/amdgpu: allow pinning DMA-bufs into VRAM if all importers can do P2P") Suggested-by: Christian König Signed-off-by: Felix Kuehling Reviewed-by: Christian König Signed-off-by: Alex Deucher (cherry picked from commit 3940796a6eefa555fec688a4adee5659ef9fa431)

drm/amdgpu: allow pinning DMA-bufs into VRAM if all importers can do P2P

2025-04-07T19:18:30Z

Try pinning into VRAM to allow P2P with RDMA NICs without ODP support if all attachments can do P2P. If any attachment can't do P2P just pin into GTT instead. Acked-by: Simona Vetter Signed-off-by: Christian König Signed-off-by: Felix Kuehling Reviewed-by: Felix Kuehling Tested-by: Pak Nin Lui Cc: Simona Vetter Signed-off-by: Alex Deucher

drm/amdgpu/dma_buf: fix page_link check

2025-04-07T18:31:45Z

The page_link lower bits of the first sg could contain something like SG_END, if we are mapping a single VRAM page or contiguous blob which fits into one sg entry. Rather pull out the struct page, and use that in our check to know if we mapped struct pages vs VRAM. Fixes: f44ffd677fb3 ("drm/amdgpu: add support for exporting VRAM using DMA-buf v3") Signed-off-by: Matthew Auld Cc: Christian König Cc: amd-gfx@lists.freedesktop.org Cc: # v5.8+ Reviewed-by: Christian König Signed-off-by: Alex Deucher

drm/amdgpu: track bo memory stats at runtime

2024-12-19T15:56:28Z

Before, every time fdinfo is queried we try to lock all the BOs in the VM and calculate memory usage from scratch. This works okay if the fdinfo is rarely read and the VMs don't have a ton of BOs. If either of these conditions is not true, we get a massive performance hit. In this new revision, we track the BOs as they change states. This way when the fdinfo is queried we only need to take the status lock and copy out the usage stats with minimal impact to the runtime performance. With this new approach however, we would no longer be able to track active buffers. Signed-off-by: Yunxiang Li Reviewed-by: Christian König Link: https://patchwork.freedesktop.org/patch/msgid/20241219151411.1150-6-Yunxiang.Li@amd.com Signed-off-by: Christian König

drm/amdgpu: remove unused function parameter

2024-12-19T15:56:25Z

amdgpu_vm_bo_invalidate doesn't use the adev parameter and not all callers have a reference to adev handy, so remove it for cleanliness. Signed-off-by: Yunxiang Li Reviewed-by: Christian König Link: https://patchwork.freedesktop.org/patch/msgid/20241219151411.1150-5-Yunxiang.Li@amd.com Signed-off-by: Christian König

drm/amdgpu: revert "take runtime pm reference when we attach a buffer" v2

2024-06-14T20:17:13Z

This reverts commit b8c415e3bf98 ("drm/amdgpu: take runtime pm reference when we attach a buffer") and commit 425285d39afd ("drm/amdgpu: add amdgpu runpm usage trace for separate funcs"). Taking a runtime pm reference for DMA-buf is actually completely unnecessary and even dangerous. The problem is that calling pm_runtime_get_sync() from the DMA-buf callbacks is illegal because we have the reservation locked here which is also taken during resume. So this would deadlock. When the buffer is in GTT it is still accessible even when the GPU is powered down and when it is in VRAM the buffer gets migrated to GTT before powering down. The only use case which would make it mandatory to keep the runtime pm reference would be if we pin the buffer into VRAM, and that's not something we currently do. v2: improve the commit message Signed-off-by: Christian König Reviewed-by: Alex Deucher Signed-off-by: Alex Deucher CC: stable@vger.kernel.org

drm/amdgpu: Reduce mem_type to domain double indirection

2024-05-17T21:40:38Z

All apart from AMDGPU_GEM_DOMAIN_GTT memory domains map 1:1 to TTM placements. And the former be either AMDGPU_PL_PREEMPT or TTM_PL_TT, depending on AMDGPU_GEM_CREATE_PREEMPTIBLE. Simplify a few places in the code which convert the TTM placement into a domain by checking against the current placement directly. In the conversion AMDGPU_PL_PREEMPT either does not have to be handled because amdgpu_mem_type_to_domain() cannot return that value anyway. v2: * Remove AMDGPU_PL_PREEMPT handling. v3: * Rebase. Signed-off-by: Tvrtko Ursulin Reviewed-by: Christian König # v1 Reviewed-by: Felix Kuehling # v2 Signed-off-by: Alex Deucher