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. -Chris