[PATCH v6 03/11] drm/amdgpu: rework the userq doorbell object destroy

[Liang, Prike]

[Public]

Regards,
      Prike

> -----Original Message-----
> From: Koenig, Christian <Christian.Koenig at amd.com>
> Sent: Tuesday, July 15, 2025 4:49 PM
> To: Liang, Prike <Prike.Liang at amd.com>; amd-gfx at lists.freedesktop.org
> Cc: Deucher, Alexander <Alexander.Deucher at amd.com>
> Subject: Re: [PATCH v6 03/11] drm/amdgpu: rework the userq doorbell object
> destroy
>
> On 15.07.25 10:07, Liang, Prike wrote:
> > [Public]
> >
> > Regards,
> >       Prike
> >
> >> -----Original Message-----
> >> From: Koenig, Christian <Christian.Koenig at amd.com>
> >> Sent: Friday, July 11, 2025 8:01 PM
> >> To: Liang, Prike <Prike.Liang at amd.com>; amd-gfx at lists.freedesktop.org
> >> Cc: Deucher, Alexander <Alexander.Deucher at amd.com>
> >> Subject: Re: [PATCH v6 03/11] drm/amdgpu: rework the userq doorbell
> >> object destroy
> >>
> >> On 11.07.25 11:39, Prike Liang wrote:
> >>> This patch aims to unify and destroy the userq doorbell objects at
> >>> mes_userq_mqd_destroy(), and this change will also help with
> >>> unpinning and destroying the userq doorbell objects for
> >>> amdgpu_userq_mgr_fini() during releasing the drm files.
> >>>
> >>> Signed-off-by: Prike Liang <Prike.Liang at amd.com>
> >>> Reviewed-by: Alex Deucher <alexander.deucher at amd.com>
> >>> ---
> >>>  drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c  | 6 ------
> >>> drivers/gpu/drm/amd/amdgpu/mes_userqueue.c | 7 +++++++
> >>>  2 files changed, 7 insertions(+), 6 deletions(-)
> >>>
> >>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
> >>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
> >>> index 3d2a7f8946cf..15e833b1b3e3 100644
> >>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
> >>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
> >>> @@ -312,12 +312,6 @@ amdgpu_userq_destroy(struct drm_file *filp, int
> >> queue_id)
> >>>             return -EINVAL;
> >>>     }
> >>>     amdgpu_userq_wait_for_last_fence(uq_mgr, queue);
> >>> -   r = amdgpu_bo_reserve(queue->db_obj.obj, true);
> >>> -   if (!r) {
> >>> -           amdgpu_bo_unpin(queue->db_obj.obj);
> >>> -           amdgpu_bo_unreserve(queue->db_obj.obj);
> >>> -   }
> >>> -   amdgpu_bo_unref(&queue->db_obj.obj);
> >>>     r = amdgpu_userq_unmap_helper(uq_mgr, queue);
> >>>     /*TODO: It requires a reset for userq hw unmap error*/
> >>>     if (unlikely(r != AMDGPU_USERQ_STATE_UNMAPPED)) { diff --git
> >>> a/drivers/gpu/drm/amd/amdgpu/mes_userqueue.c
> >>> b/drivers/gpu/drm/amd/amdgpu/mes_userqueue.c
> >>> index 1457fb49a794..15aa1ca67a11 100644
> >>> --- a/drivers/gpu/drm/amd/amdgpu/mes_userqueue.c
> >>> +++ b/drivers/gpu/drm/amd/amdgpu/mes_userqueue.c
> >>> @@ -336,6 +336,13 @@ mes_userq_mqd_destroy(struct amdgpu_userq_mgr
> >> *uq_mgr,
> >>>                   struct amdgpu_usermode_queue *queue)  {
> >>>     amdgpu_userq_destroy_object(uq_mgr, &queue->fw_obj);
> >>> +
> >>> +   if (!amdgpu_bo_reserve(queue->db_obj.obj, true)) {
> >>> +           amdgpu_bo_unpin(queue->db_obj.obj);
> >>> +           amdgpu_bo_unreserve(queue->db_obj.obj);
> >>> +           amdgpu_userq_destroy_object(uq_mgr, &queue->db_obj);
> >>> +   }
> >>> +
> >>
> >> That makes no sense to do here. The pinning isn't done in
> >> mes_userq_mqd_create() either.
> > Yes, but the doorbell BO is pinned by
> > amdgpu_userq_get_doorbell_index(), which is still Invoked during userq
> > BOs creation phase. This patch wants to free the doorbell object like some other
> userq objects at the unified place of mes_userq_mqd_destroy().
>
> Yeah and exactly that is not a good idea.
>
> The doorbell object is provided by userspace and not allocated by the kernel like the
> MQD.
>
> So destroying it here makes no sense at all. You are most likely messing up the
> doorbell reference count with that.
>
> >> In general we should avoid pinning the MQD in the first place, that
> >> buffer needs to be fences instead.
> > If here not pin the userq doorbell BO, then will the doorbell index be
> > changed when the doorbell BO is moved?
>
> Correct, yes. The doorbell index needs to be updated on each resume of the
> userqueue.
>
> We haven't implemented that yet since we weren't sure if the MES FW could handle
> that (and because the eviction fences wasn't ready at that time).
Thank you for the history lesson. If there's no objections, I plan to rework the userq doorbell object management separately.

> Regards,
> Christian.
>
> >
> >> Regards,
> >> Christian.
> >>
> >>>     kfree(queue->userq_prop);
> >>>     amdgpu_userq_destroy_object(uq_mgr, &queue->mqd);  }
> >