Thank you for your contribution! Sashiko AI review found 1 potential
issue to consider:
- [Medium] drm/amdgpu/userq: missing resume_work flush in
  amdgpu_userq_query_status()
--

commit dcee7ab7a079dac06e55c3b0a31589e130baf8e9
Author: Prike Liang <[email protected]>

drm/amdgpu/userq: report invalid VA through the broken status flag

This commit sets AMDGPU_USERQ_QUERY_STATUS_FLAGS_BROKEN in
amdgpu_userq_query_status() when a usermode queue is in
AMDGPU_USERQ_STATE_INVALID_VA, and updates the UAPI comment in
amdgpu_drm.h.

> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c 
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
> index 496137fca1f0c..b48cf2a2118c0 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
> @@ -837,6 +837,7 @@ amdgpu_userq_query_status(struct drm_file *filp, union 
> drm_amdgpu_userq *args)
>  
>       args->out_qs.flags = 0;
>       if (queue->state == AMDGPU_USERQ_STATE_HUNG ||
> +         queue->state == AMDGPU_USERQ_STATE_INVALID_VA ||
>           queue->generation != amdgpu_vm_generation(adev, &fpriv->vm))
>               args->out_qs.flags |= AMDGPU_USERQ_QUERY_STATUS_FLAGS_BROKEN;
>       amdgpu_userq_put(queue);

[Severity: Medium]
Can amdgpu_userq_query_status() miss AMDGPU_USERQ_STATE_INVALID_VA when
called before uq_mgr->resume_work finishes?

The only place that sets queue->state to AMDGPU_USERQ_STATE_INVALID_VA is
amdgpu_userq_restore_all(), which runs asynchronously from
uq_mgr->resume_work:

CPU0                                  kworker
amdgpu_gem_va_ioctl(UNMAP)
  amdgpu_userq_gem_va_unmap_validate()
    dma_resv_wait_timeout()
                                      amdgpu_eviction_fence_suspend_worker()
                                        amdgpu_userq_evict()
                                          queue->state =
                                            AMDGPU_USERQ_STATE_UNMAPPED
                                        dma_fence_signal(ev_fence)
                                        schedule_delayed_work(
                                          &uq_mgr->resume_work, 0)
  drm_exec_fini()
amdgpu_userq_query_status()
  mutex_lock(&uq_mgr->userq_mutex)
  queue->state ==
    AMDGPU_USERQ_STATE_UNMAPPED  <-- BROKEN not set
  mutex_unlock(&uq_mgr->userq_mutex)
                                      amdgpu_userq_restore_worker()
                                        ...
                                        amdgpu_userq_restore_all()
                                          queue->state =
                                            AMDGPU_USERQ_STATE_INVALID_VA

Because amdgpu_gem_va_ioctl() holds the VM reservation lock until
drm_exec_fini() just before returning to userspace, uq_mgr->resume_work
cannot finish validating the VM and calling amdgpu_userq_restore_all()
before the unmap ioctl returns.  Unlike amdgpu_userq_ensure_ev_fence(),
amdgpu_userq_query_status() does not call
flush_delayed_work(&uq_mgr->resume_work) before acquiring
uq_mgr->userq_mutex, so querying queue status after unmapping a queue VA
reports flags as 0 instead of AMDGPU_USERQ_QUERY_STATUS_FLAGS_BROKEN.

Also, if a queue VA mapping is removed via AMDGPU_VA_OP_CLEAR or
AMDGPU_VA_OP_REPLACE (amdgpu_vm_bo_clear_mappings()),
amdgpu_userq_gem_va_unmap_validate() is never called to trigger eviction and
transition the queue to AMDGPU_USERQ_STATE_INVALID_VA.

-- 
Sashiko AI review ยท 
https://sashiko.dev/#/patchset/[email protected]?part=3

Reply via email to