[Intel-gfx] [PATCH 4/4] dma-buf: nuke reservation_object seq number

[Chris Wilson]

Quoting Koenig, Christian (2019-08-14 18:58:32)
> Am 14.08.19 um 19:48 schrieb Chris Wilson:
> > Quoting Chris Wilson (2019-08-14 18:38:20)
> >> Quoting Chris Wilson (2019-08-14 18:22:53)
> >>> Quoting Chris Wilson (2019-08-14 18:06:18)
> >>>> Quoting Chris Wilson (2019-08-14 17:42:48)
> >>>>> Quoting Daniel Vetter (2019-08-14 16:39:08)
> >>>>>>>>> +       } while (rcu_access_pointer(obj->fence_excl) != *excl);
> >>>>>> What if someone is real fast (like really real fast) and recycles the
> >>>>>> exclusive fence so you read the same pointer twice, but everything else
> >>>>>> changed? reused fence pointer is a lot more likely than seqlock wrapping
> >>>>>> around.
> >>>>> It's an exclusive fence. If it is replaced, it must be later than all
> >>>>> the shared fences (and dependent on them directly or indirectly), and
> >>>>> so still a consistent snapshot.
> >>>> An extension of that argument says we don't even need to loop,
> >>>>
> >>>> *list = rcu_dereference(obj->fence);
> >>>> *shared_count = *list ? (*list)->shared_count : 0;
> >>>> smp_rmb();
> >>>> *excl = rcu_dereference(obj->fence_excl);
> >>>>
> >>>> Gives a consistent snapshot. It doesn't matter if the fence_excl is
> >>>> before or after the shared_list -- if it's after, it's a superset of all
> >>>> fences, if it's before, we have a correct list of shared fences the
> >>>> earlier fence_excl.
> >>> The problem is that the point of the loop is that we do need a check on
> >>> the fences after the full memory barrier.
> >>>
> >>> Thinking of the rationale beaten out for dma_fence_get_excl_rcu_safe()
> >>>
> >>> We don't have a full memory barrier here, so this cannot be used safely
> >>> in light of fence reallocation.
> >> i.e. we need to restore the loops in the callers,
> >>
> >> diff --git a/drivers/gpu/drm/i915/gem/i915_gem_busy.c b/drivers/gpu/drm/i915/gem/i915_gem_busy.c
> >> index a2aff1d8290e..f019062c8cd7 100644
> >> --- a/drivers/gpu/drm/i915/gem/i915_gem_busy.c
> >> +++ b/drivers/gpu/drm/i915/gem/i915_gem_busy.c
> >> @@ -110,6 +110,7 @@ i915_gem_busy_ioctl(struct drm_device *dev, void *data,
> >>           * to report the overall busyness. This is what the wait-ioctl does.
> >>           *
> >>           */
> >> +retry:
> >>          dma_resv_fences(obj->base.resv, &excl, &list, &shared_count);
> >>
> >>          /* Translate the exclusive fence to the READ *and* WRITE engine */
> >> @@ -122,6 +123,10 @@ i915_gem_busy_ioctl(struct drm_device *dev, void *data,
> >>                  args->busy |= busy_check_reader(fence);
> >>          }
> >>
> >> +       smp_rmb();
> >> +       if (excl != rcu_access_pointer(obj->base.resv->fence_excl))
> >> +               goto retry;
> >> +
> >>
> >> wrap that up as
> >>
> >> static inline bool
> >> dma_resv_fences_retry(struct dma_resv *resv, struct dma_fence *excl)
> >> {
> >>          smp_rmb();
> >>          return excl != rcu_access_pointer(resv->fence_excl);
> >> }
> > I give up. It's not just the fence_excl that's an issue here.
> >
> > Any of the shared fences may be replaced after dma_resv_fences()
> > and so the original shared fence pointer may be reassigned (even under
> > RCU).
> 
> Yeah, but this should be harmless. See fences are always replaced either 
> when they are signaled anyway or by later fences from the same context.
> 
> And existing fences shouldn't be re-used while under RCU, or is anybody 
> still using SLAB_TYPESAFE_BY_RCU?

Yes. We go through enough fences that the freelist is a noticeable
improvement.
-Chris