]> git.ipfire.org Git - thirdparty/kernel/stable.git/commitdiff
drm/amdgpu: csa unmap use uninterruptible lock
authorPhilip Yang <Philip.Yang@amd.com>
Wed, 28 Jan 2026 03:22:38 +0000 (11:22 +0800)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Fri, 30 Jan 2026 09:27:43 +0000 (10:27 +0100)
[ Upstream commit a0fa7873f2f869087b1e7793f7fac3713a1e3afe ]

After process exit to unmap csa and free GPU vm, if signal is accepted
and then waiting to take vm lock is interrupted and return, it causes
memory leaking and below warning backtrace.

Change to use uninterruptible wait lock fix the issue.

WARNING: CPU: 69 PID: 167800 at amd/amdgpu/amdgpu_kms.c:1525
 amdgpu_driver_postclose_kms+0x294/0x2a0 [amdgpu]
 Call Trace:
  <TASK>
  drm_file_free.part.0+0x1da/0x230 [drm]
  drm_close_helper.isra.0+0x65/0x70 [drm]
  drm_release+0x6a/0x120 [drm]
  amdgpu_drm_release+0x51/0x60 [amdgpu]
  __fput+0x9f/0x280
  ____fput+0xe/0x20
  task_work_run+0x67/0xa0
  do_exit+0x217/0x3c0
  do_group_exit+0x3b/0xb0
  get_signal+0x14a/0x8d0
  arch_do_signal_or_restart+0xde/0x100
  exit_to_user_mode_loop+0xc1/0x1a0
  exit_to_user_mode_prepare+0xf4/0x100
  syscall_exit_to_user_mode+0x17/0x40
  do_syscall_64+0x69/0xc0

Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
(cherry picked from commit 7dbbfb3c171a6f63b01165958629c9c26abf38ab)
Cc: stable@vger.kernel.org
[The third parameter of drm_exec_init() was introduced by commit
 05d249352f1a ("drm/exec: Pass in initial # of objects") after Linux 6.8.
 This code targets linux 6.6, so the current implementation is used
 and the third parameter is not needed.]
Signed-off-by: Li hongliang <1468888505@139.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c

index 7200110197415f530243907691526b14084086cf..bff44fc21e8bf5d93eab23463abaf03dd53c975d 100644 (file)
@@ -110,7 +110,7 @@ int amdgpu_unmap_static_csa(struct amdgpu_device *adev, struct amdgpu_vm *vm,
        struct drm_exec exec;
        int r;
 
-       drm_exec_init(&exec, DRM_EXEC_INTERRUPTIBLE_WAIT);
+       drm_exec_init(&exec, 0);
        drm_exec_until_all_locked(&exec) {
                r = amdgpu_vm_lock_pd(vm, &exec, 0);
                if (likely(!r))