RE: [PATCH v2] drm/amdgpu/userq: fix lock missing for userq fence error set

AMD General

Regards,
      Prike

> -----Original Message-----
> From: Alex Deucher <[email protected]>
> Sent: Wednesday, August 19, 2026 11:52 PM
> To: Liang, Prike <[email protected]>
> Cc: [email protected]; Deucher, Alexander
> <[email protected]>; Koenig, Christian <[email protected]>
> Subject: Re: [PATCH v2] drm/amdgpu/userq: fix lock missing for userq fence error
> set
>
> On Thu, Aug 13, 2026 at 11:50 PM Prike Liang <[email protected]> wrote:
> >
> > amdgpu_userq_fence_driver() and amdgpu_userq_fence_driver_destroy()
> > don't acquire the dma_fence spinlock, so locking the dma_fence lock
> > before test the signaled state and set error state.
> >
> > Signed-off-by: Prike Liang <[email protected]>
> > ---
> >  .../gpu/drm/amd/amdgpu/amdgpu_userq_fence.c   | 20 ++++++++++++-------
> >  1 file changed, 13 insertions(+), 7 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq_fence.c
> > b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq_fence.c
> > index 19af1adc2921..09e2079cfdbd 100644
> > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq_fence.c
> > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq_fence.c
> > @@ -191,14 +191,15 @@ void amdgpu_userq_fence_driver_destroy(struct kref
> *ref)
> >         struct dma_fence *f;
> >
> >         spin_lock_irqsave(&fence_drv->fence_list_lock, flags);
> > +       lockdep_assert_held(&fence_drv->fence_list_lock);
> >         list_for_each_entry_safe(fence, tmp, &fence_drv->fences, link) {
> >                 f = &fence->base;
> > -
> > -               if (!dma_fence_is_signaled(f)) {
> > +               spin_lock(dma_fence_spinlock(f));
> > +               if (!dma_fence_is_signaled_locked(f)) {
> >                         dma_fence_set_error(f, -ECANCELED);
> > -                       dma_fence_signal(f);
> > +                       dma_fence_signal_locked(f);
> >                 }
> > -
> > +               spin_unlock(dma_fence_spinlock(f));
> >                 list_del(&fence->link);
> >                 dma_fence_put(f);
> >         }
> > @@ -423,11 +424,16 @@ amdgpu_userq_fence_driver_set_error(struct
> amdgpu_userq_fence *fence,
> >         struct dma_fence *f;
> >
> >         spin_lock_irqsave(&fence_drv->fence_list_lock, flags);
> > -
> > +       lockdep_assert_held(&fence_drv->fence_list_lock);
> >         f = rcu_dereference_protected(&fence->base,
> >                                       lockdep_is_held(&fence_drv->fence_list_lock));
> > -       if (f && !dma_fence_is_signaled_locked(f))
> > -               dma_fence_set_error(f, error);
> > +       if (f) {
> > +               /* nest f->lock inside fence_list_lock */
> > +               spin_lock(dma_fence_spinlock(f));
>
> Why do we need to take the fence spinlock for userqs, but not for kernel queues?
> See amdgpu_fence.c

For the kernel queue ring fence, we also need to hold the dma_fence lock when calling dma_fence_is_signaled_locked(). The reason that does not assert the lock missing for the kernel ring fence is the .signaled() callback is not implemented, so dma_fence_signal_locked() will never be invoked.

Thanks,
Prike

> Alex
>
> > +               if (!dma_fence_is_signaled_locked(f))
> > +                       dma_fence_set_error(f, error);
> > +               spin_unlock(dma_fence_spinlock(f));
> > +       }
> >         spin_unlock_irqrestore(&fence_drv->fence_list_lock, flags);  }
> >
> > --
> > 2.34.1
> >