[Differential] [Commented On] D1833: Add memory barriers to buf_ring

[kmacy (Kip Macy)]

kmacy added a comment.

I've been spending more time looking at this than I care to admit. I don't understand why we need the load_acq_32 on prod_tail. The atomic_store_rel_int in enqueue should guarantee that the store to the ring happens before the update to prod_tail. So the consumer should _never_ see a prod_tail that is newer than the ring[prod_next] value. I suspect that something else is going on that we don't understand.

REVISION DETAIL
  https://reviews.freebsd.org/D1833

To: zbb, kmacy, rpaulo, imp
Cc: meloun-miracle-cz, onwahe-gmail-com, andrew, ian, adrian, freebsd-arm