linux.git/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c, branch v7.3-rc2

drm/amdkfd: Remove svm_bo eviction fence

2026-07-29T21:52:20+00:00

SVM BOs are now migrated back to system memory synchronously from the TTM
eviction path (svm_range_evict_svm_bo), so the per-svm_bo eviction fence
is no longer used.

Remove the eviction fence from svm_range_bo, drop the
amdgpu_amdkfd_fence->svm_bo back pointer and the
amdgpu_amdkfd_evict_svm_bo() helper, and stop special-casing svm_bo
fences in the KFD fence enable_signaling and check_mm paths. Embed struct
amdgpu_bo directly in svm_range_bo with a dedicated svm_range_bo_destroy()
callback, and keep the owning mm via mmgrab()/mmdrop() instead of through
the fence.

Signed-off-by: Philip Yang 
Reviewed-by: Felix Kuehling 
Signed-off-by: Alex Deucher

drm/amdkfd: Evict SVM BOs synchronously from TTM eviction

2026-07-28T23:17:34+00:00

svm_range_evict_svm_bo_worker() migrated an SVM BO's pages back to system
memory from a work item that took mmap_read_lock. When an mmap writer was
pending, that read lock blocked behind the writer while the thread
allocating a new migration VRAM BO waited on this BO's eviction fence - a
circular wait that hung the SVM workers.

Evict the SVM BO synchronously from the TTM eviction path
(amdgpu_ttm_bo_eviction_valuable) instead of deferring to a work item.
The BO is already reserved and the lock order is mmap_lock -> BO
reservation, so only trylock the owning process's mmap lock; on
contention return -EBUSY so TTM skips this BO. This removes the eviction
work item and the enable_signaling path, so no worker can block on
mmap_read_lock.

The SVM BO uses AMDGPU_GEM_CREATE_DISCARDABLE, so ttm_bo_evict takes the
pipeline_gutting path and skips allocating a system memory placement.
That would be wasted work, since svm_migrate_vram_to_ram allocates the
system pages and copies the data back itself.

Eviction now migrates ranges directly, so it must serialize with the
owning process: it trylocks migrate_mutex under svm_bo->list_lock before
unlinking the range, and svm_range_free() unlinks the range then waits on
migrate_mutex, so a concurrent eviction cannot free a range under it.

Drop the mm reference with mmput_async so exit_mmap() does not run under
the BO reservation.

Signed-off-by: Philip Yang 
Reviewed-by: Felix Kuehling 
Signed-off-by: Alex Deucher

drm/amdgpu: use correct gfp_t for job allocation

2026-07-16T22:26:18+00:00

For job allocation in GPU reset and page fault handling we must use
GFP_ATOMIC to guarantee that we don't cycle back and depend on a
dma_fence submission for the memory allocation.

Add gfp_flags argument to amdgpu_job_alloc() and expose the
gfp_flags of IB pools with amdgpu_ib_pool_gfp_flags() so that
we can use different flags when allocating jobs.

Assisted-by: Claude:Sonnet 4
Signed-off-by: Christian König 
Reviewed-by: Timur Kristóf 
Signed-off-by: Alex Deucher

drm/amdgpu: move job parameter to the end in amdgpu_job_alloc() and *_with_ib()

2026-07-16T22:26:04+00:00

Move the job output parameter to be the last parameter in the
amdgpu_job_alloc() and amdgpu_job_alloc_with_ib() function signature. This
aligns with the common kernel coding convention where output parameters
typically come last.

Assisted-by: Claude:Sonnet 4
Signed-off-by: Christian König 
Reviewed-by: Timur Kristóf 
Signed-off-by: Alex Deucher

drm/amdkfd: Let driver decide buffer size at AMDKFD_IOC_GET_DMABUF_INFO ioctl

2026-06-17T20:30:25+00:00

amdkfd driver needs allocate buffer to return bo metadata to user space. The
buffer size is controlled by user currently. It is a potential security issue
that hostile value (e.g. 2 GiB) lets any render-group user trigger order-MAX
allocation/OOM in kernel context.

This patch first finds bo metadata size. If the size is smaller than user
provided value drive can safely allocate buffer in kernel space and copy to
user space buffer. If not, driver will let user know, not allocate and copy.
User will redo with new buffer in user space.

This patch lets driver decide buffer allocation size to avoid potential hostile
size from user space.

Signed-off-by: Xiaogang Chen 
Reviewed-by: Alex Deucher 
Signed-off-by: Alex Deucher

drm/amdkfd: plumb a helper to reset a KFD user queue

2026-06-17T19:51:35+00:00

Can be called from KGD.

Reviewed-by: Amber Lin 
Signed-off-by: Alex Deucher

drm/amdgpu: unmap all user mappings of framebuffer and doorbell before mode1 reset

2026-05-19T15:45:41+00:00

During Mode 1 reset, the ASIC undergoes a reset cycle and becomes temporarily
inaccessible via PCIe. Any attempt to access framebuffer or MMIO registers during
this window can result in uncompleted PCIe transactions, leading to NMI panics or
system hangs.

To prevent this, Unmap all of the applications mappings of the framebuffer
and doorbell BARs before mode1 reset. Also prevent new mappings from coming in
during the reset process.

v2: remove inode in kfd_dev (Christian)
v3: correct unmap offset (Felix), remove prevent new mappings part
to avoid deadlock (Christian)

Reviewed-by: Felix Kuehling 
Signed-off-by: Yifan Zhang 
Signed-off-by: Alex Deucher

drm/amdgpu: add an option to allow gpu partition allocate all available memory

2026-04-03T17:50:29+00:00

Current driver reports and limits memory allocation for each partition equally
among partitions using same memory partition. Application may not be able to
use all available memory when run on a partitioned gpu though system still has
enough free memory.

Add an option that app can use to have gpu partition allocate all available
memory.

Signed-off-by: Xiaogang Chen 
Reviewed-by: Philip Yang 
Signed-off-by: Alex Deucher

drm/amdgpu: Fix fence put before wait in amdgpu_amdkfd_submit_ib

2026-03-24T17:27:35+00:00

amdgpu_amdkfd_submit_ib() submits a GPU job and gets a fence
from amdgpu_ib_schedule(). This fence is used to wait for job
completion.

Currently, the code drops the fence reference using dma_fence_put()
before calling dma_fence_wait().

If dma_fence_put() releases the last reference, the fence may be
freed before dma_fence_wait() is called. This can lead to a
use-after-free.

Fix this by waiting on the fence first and releasing the reference
only after dma_fence_wait() completes.

Fixes the below:
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c:697 amdgpu_amdkfd_submit_ib() warn: passing freed memory 'f' (line 696)

Fixes: 9ae55f030dc5 ("drm/amdgpu: Follow up change to previous drm scheduler change.")
Cc: Felix Kuehling 
Cc: Dan Carpenter 
Cc: Christian König 
Cc: Alex Deucher 
Signed-off-by: Srinivasan Shanmugam 
Reviewed-by: Christian König 
Signed-off-by: Alex Deucher

Convert 'alloc_obj' family to use the new default GFP_KERNEL argument

2026-02-22T01:09:51+00:00

This was done entirely with mindless brute force, using

    git grep -l '\