io_uring_poll: kill the no longer necessary barrier after poll_wait()

Now that poll_wait() provides a full barrier we can remove smp_rmb() from
io_uring_poll().

In fact I don't think smp_rmb() was correct, it can't serialize LOADs and
STOREs.

Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Link: https://lore.kernel.org/r/20250107162730.GA18940@redhat.com
Signed-off-by: Christian Brauner <brauner@kernel.org>

diff --git a/io_uring/io_uring.c b/io_uring/io_uring.c
index 06ff41484e29c6e7d8779bd7ff8317ebae003a8d..a64a82b93b865d0c0a30dfebb5bd42ee154ae572 100644 (file)
--- a/io_uring/io_uring.c
+++ b/io_uring/io_uring.c
@@ -2809,13 +2809,12 @@ static __poll_t io_uring_poll(struct file *file, poll_table *wait)
 
        if (unlikely(!ctx->poll_activated))
                io_activate_pollwq(ctx);
-
-       poll_wait(file, &ctx->poll_wq, wait);
        /*
-        * synchronizes with barrier from wq_has_sleeper call in
-        * io_commit_cqring
+        * provides mb() which pairs with barrier from wq_has_sleeper
+        * call in io_commit_cqring
         */
-       smp_rmb();
+       poll_wait(file, &ctx->poll_wq, wait);
+
        if (!io_sqring_full(ctx))
                mask |= EPOLLOUT | EPOLLWRNORM;