io_uring: remove unnecessary barrier before wq_has_sleeper

wq_has_sleeper has a full barrier internally. The smp_rmb barrier in
io_uring_poll synchronizes with it.

Signed-off-by: Stefan Bühler <source@stbuehler.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This commit is contained in:
Stefan Bühler 2019-04-24 23:54:17 +02:00 committed by Jens Axboe
parent 1e84b97b73
commit 4f7067c3fb
1 changed files with 4 additions and 7 deletions

View File

@ -418,12 +418,6 @@ static void io_commit_cqring(struct io_ring_ctx *ctx)
/* order cqe stores with ring update */
smp_store_release(&ring->r.tail, ctx->cached_cq_tail);
/*
* Write sider barrier of tail update, app has read side. See
* comment at the top of this file.
*/
smp_wmb();
if (wq_has_sleeper(&ctx->cq_wait)) {
wake_up_interruptible(&ctx->cq_wait);
kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
@ -2677,7 +2671,10 @@ static __poll_t io_uring_poll(struct file *file, poll_table *wait)
__poll_t mask = 0;
poll_wait(file, &ctx->cq_wait, wait);
/* See comment at the top of this file */
/*
* synchronizes with barrier from wq_has_sleeper call in
* io_commit_cqring
*/
smp_rmb();
if (READ_ONCE(ctx->sq_ring->r.tail) - ctx->cached_sq_head !=
ctx->sq_ring->ring_entries)