On Fri, Jan 12, 2024 at 02:19:00PM -0800, Erik Kurzinger wrote: > When waiting for a syncobj timeline point whose fence has not yet been > submitted with the WAIT_FOR_SUBMIT flag, a callback is registered using > drm_syncobj_fence_add_wait and the thread is put to sleep until the > timeout expires. If the fence is submitted before then, > drm_syncobj_add_point will wake up the sleeping thread immediately which > will proceed to wait for the fence to be signaled. > > However, if the WAIT_AVAILABLE flag is used instead, > drm_syncobj_fence_add_wait won't get called, meaning the waiting thread > will always sleep for the full timeout duration, even if the fence gets > submitted earlier. If it turns out that the fence *has* been submitted > by the time it eventually wakes up, it will still indicate to userspace > that the wait completed successfully (it won't return -ETIME), but it > will have taken much longer than it should have. > > To fix this, we must call drm_syncobj_fence_add_wait if *either* the > WAIT_FOR_SUBMIT flag or the WAIT_AVAILABLE flag is set. The only > difference being that with WAIT_FOR_SUBMIT we will also wait for the > fence to be signaled after it has been submitted while with > WAIT_AVAILABLE we will return immediately. > > IGT test patch: https://lists.freedesktop.org/archives/igt-dev/2024-January/067282.html > > Fixes: 01d6c3578379 ("drm/syncobj: add support for timeline point wait v8") > Signed-off-by: Erik Kurzinger <ekurzinger@xxxxxxxxxx> > --- > drivers/gpu/drm/drm_syncobj.c | 3 ++- > 1 file changed, 2 insertions(+), 1 deletion(-) > > diff --git a/drivers/gpu/drm/drm_syncobj.c b/drivers/gpu/drm/drm_syncobj.c > index a8e6b61a188c..a1443c673f30 100644 > --- a/drivers/gpu/drm/drm_syncobj.c > +++ b/drivers/gpu/drm/drm_syncobj.c > @@ -1121,7 +1121,8 @@ static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs, > * fallthough and try a 0 timeout wait! > */ > > - if (flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT) { > + if (flags & (DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT | You also need to adjust the condition earlier in the function that guards the lockdep_assert_none call. I also noticed that unliked drm_syncobj_find_fence() we don't have the might_sleep check in there, would be good to add that (in an additional patch). Also I think it'd be good to reject the DRM_SYNCOBJ_WAIT_FLAGS_WAIT_AVAILABLE flag in drm_syncobj_find_fence() since that's a clear misuses of that. Found that since I checked whether there's more conditions that should be patched. Cheers, Sima > + DRM_SYNCOBJ_WAIT_FLAGS_WAIT_AVAILABLE)) { > for (i = 0; i < count; ++i) > drm_syncobj_fence_add_wait(syncobjs[i], &entries[i]); > } > -- > 2.43.0 > > -- Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch
