Re: [PATCH 2/2] dma-buf/dma-fence: Unify return codes for signalled fences

[Philipp Stanner]

On Tue, 2025-11-25 at 13:20 +0000, Tvrtko Ursulin wrote:
> 
> On 25/11/2025 10:44, Philipp Stanner wrote:
> > Almost all users of dma_fence_signal() ignore the return code which
> > would indicate that the fence was already signaled. The same return code
> > by dma_fence_add_callback() cannot be ignored, however, because it's
> > needed to detect races.
> > 
> > For an already signaled fence, dma_fence_signal() returns -EINVAL,
> > whereas dma_fence_add_callback() returns -ENOENT.
> > 
> > Unify the error codes by having dma_fence_signal() return -ENOENT, too.
> > 
> > Signed-off-by: Philipp Stanner <phasta@xxxxxxxxxx>
> > ---
> >   drivers/dma-buf/dma-fence.c | 2 +-
> >   1 file changed, 1 insertion(+), 1 deletion(-)
> > 
> > diff --git a/drivers/dma-buf/dma-fence.c b/drivers/dma-buf/dma-fence.c
> > index 3a48896ded62..09d97624e647 100644
> > --- a/drivers/dma-buf/dma-fence.c
> > +++ b/drivers/dma-buf/dma-fence.c
> > @@ -373,7 +373,7 @@ int dma_fence_signal_timestamp_locked(struct dma_fence *fence,
> >   	lockdep_assert_held(fence->lock);
> >   
> >   	if (unlikely(__dma_fence_is_signaled(fence)))
> > -		return -EINVAL;
> > +		return -ENOENT;
> >   
> >   	/* Stash the cb_list before replacing it with the timestamp */
> >   	list_replace(&fence->cb_list, &cb_list);
> 
> Story checks out AFAICT - only two callers fetch the error, xe and kfd, 
> and neither does anything with it. So I'd say it makes sense to unify
> the errno.
> 
> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxxx>
> 
> Regards,
> 
> Tvrtko
> 
> P.S. Just not sure of the 1/2 on which this one ends depending on. I 
> *think* I suggested the helper in the context of some discussion long
> long time ago but what it was? And what about all the drivers which look 
> at the signaled bit directly?

That was in a discussion where we could have needed such a function to
solve a bug in Nouveau where dma_fence_is_signaled() unexpetedely
signaled a fence. AFAIR that caused locking issues in that context.

The other test_bit() users could be ported, of course.


P.