kernel/linux.git/drivers/gpu/drm/amd/amdkfd, branch linux-6.9.y

drm/amdkfd: Let VRAM allocations go to GTT domain on small APUs

2024-07-11T10:51:21+00:00

[ Upstream commit eb853413d02c8d9b27942429b261a9eef228f005 ] Small APUs(i.e., consumer, embedded products) usually have a small carveout device memory which can't satisfy most compute workloads memory allocation requirements. We can't even run a Basic MNIST Example with a default 512MB carveout. https://github.com/pytorch/examples/tree/main/mnist. Error Log: "torch.cuda.OutOfMemoryError: HIP out of memory. Tried to allocate 84.00 MiB. GPU 0 has a total capacity of 512.00 MiB of which 0 bytes is free. Of the allocated memory 103.83 MiB is allocated by PyTorch, and 22.17 MiB is reserved by PyTorch but unallocated" Though we can change BIOS settings to enlarge carveout size, which is inflexible and may bring complaint. On the other hand, the memory resource can't be effectively used between host and device. The solution is MI300A approach, i.e., let VRAM allocations go to GTT. Then device and host can flexibly and effectively share memory resource. v2: Report local_mem_size_private as 0. (Felix) Signed-off-by: Lang Yu Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher Signed-off-by: Sasha Levin

Revert "drm/amdkfd: fix gfx_target_version for certain 11.0.3 devices"

2024-06-16T11:51:01+00:00

commit dd2b75fd9a79bf418e088656822af06fc253dbe3 upstream. This reverts commit 28ebbb4981cb1fad12e0b1227dbecc88810b1ee8. Revert this commit as apparently the LLVM code to take advantage of this never landed. Reviewed-by: Feifei Xu Signed-off-by: Alex Deucher Cc: Feifei Xu Signed-off-by: Greg Kroah-Hartman

drm/amdkfd: don't allow mapping the MMIO HDP page with large pages

2024-05-10T17:05:13+00:00

We don't get the right offset in that case. The GPU has an unused 4K area of the register BAR space into which you can remap registers. We remap the HDP flush registers into this space to allow userspace (CPU or GPU) to flush the HDP when it updates VRAM. However, on systems with >4K pages, we end up exposing PAGE_SIZE of MMIO space. Fixes: d8e408a82704 ("drm/amdkfd: Expose HDP registers to user space") Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher Cc: stable@vger.kernel.org

Revert "drm/amdkfd: Add partition id field to location_id"

2024-05-08T19:51:18+00:00

This reverts commit c37ce764cd492f044dcdbb39616298f02b0dbc7f. RCCL library is currently not treating spatial partitions differently, hence this change is causing issues. Revert temporarily till RCCL implementation is ready for spatial partitions. Signed-off-by: Lijo Lazar Reviewed-by: Jonathan Kim Signed-off-by: Alex Deucher

drm/amdkfd: Flush the process wq before creating a kfd_process

2024-05-01T01:59:16+00:00

There is a race condition when re-creating a kfd_process for a process. This has been observed when a process under the debugger executes exec(3). In this scenario: - The process executes exec. - This will eventually release the process's mm, which will cause the kfd_process object associated with the process to be freed (kfd_process_free_notifier decrements the reference count to the kfd_process to 0). This causes kfd_process_ref_release to enqueue kfd_process_wq_release to the kfd_process_wq. - The debugger receives the PTRACE_EVENT_EXEC notification, and tries to re-enable AMDGPU traps (KFD_IOC_DBG_TRAP_ENABLE). - When handling this request, KFD tries to re-create a kfd_process. This eventually calls kfd_create_process and kobject_init_and_add. At this point the call to kobject_init_and_add can fail because the old kfd_process.kobj has not been freed yet by kfd_process_wq_release. This patch proposes to avoid this race by making sure to drain kfd_process_wq before creating a new kfd_process object. This way, we know that any cleanup task is done executing when we reach kobject_init_and_add. Signed-off-by: Lancelot SIX Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher

drm/amdkfd: Add VRAM accounting for SVM migration

2024-04-24T03:23:45+00:00

Do VRAM accounting when doing migrations to vram to make sure there is enough available VRAM and migrating to VRAM doesn't evict other possible non-unified memory BOs. If migrating to VRAM fails, driver can fall back to using system memory seamlessly. Signed-off-by: Mukul Joshi Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher

drm/amdkfd: Fix rescheduling of restore worker

2024-04-24T03:23:28+00:00

Handle the case that the restore worker was already scheduled by another eviction while the restore was in progress. Fixes: 9a1c1339abf9 ("drm/amdkfd: Run restore_workers on freezable WQs") Signed-off-by: Felix Kuehling Reviewed-by: Philip Yang Tested-by: Yunxiang Li Signed-off-by: Alex Deucher Cc: stable@vger.kernel.org

drm/amdkfd: Fix eviction fence handling

2024-04-24T03:17:30+00:00

Handle case that dma_fence_get_rcu_safe returns NULL. If restore work is already scheduled, only update its timer. The same work item cannot be queued twice, so undo the extra queue eviction. Fixes: 9a1c1339abf9 ("drm/amdkfd: Run restore_workers on freezable WQs") Signed-off-by: Felix Kuehling Reviewed-by: Philip Yang Tested-by: Gang BA Reviewed-by: Gang BA Signed-off-by: Alex Deucher Cc: stable@vger.kernel.org

drm/amdkfd: Fix memory leak in create_process failure

2024-04-17T15:05:09+00:00

Fix memory leak due to a leaked mmget reference on an error handling code path that is triggered when attempting to create KFD processes while a GPU reset is in progress. Fixes: 0ab2d7532b05 ("drm/amdkfd: prepare per-process debug enable and disable") CC: Xiaogang Chen Signed-off-by: Felix Kuehling Tested-by: Harish Kasiviswanthan Reviewed-by: Mukul Joshi Signed-off-by: Alex Deucher Cc: stable@vger.kernel.org

amdkfd: use calloc instead of kzalloc to avoid integer overflow

2024-04-12T01:11:59+00:00

This uses calloc instead of doing the multiplication which might overflow. Cc: stable@vger.kernel.org Signed-off-by: Dave Airlie