io_uring: spin for sq thread to idle on shutdown

As part of io_uring shutdown, we cancel work that is pending and won't
necessarily complete on its own. That includes requests like poll
commands and timeouts.

If we're using SQPOLL for kernel side submission and we shutdown the
ring immediately after queueing such work, we can race with the sqthread
doing the submission. This means we may miss cancelling some work, which
results in the io_uring shutdown hanging forever.

Cc: stable@vger.kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>

diff --git a/fs/io_uring.c b/fs/io_uring.c
index edb00ae..87f8655 100644 (file)
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -5070,7 +5070,8 @@ static int io_sq_thread(void *data)
                         * reap events and wake us up.
                         */
                        if (inflight ||
-                           (!time_after(jiffies, timeout) && ret != -EBUSY)) {
+                           (!time_after(jiffies, timeout) && ret != -EBUSY &&
+                           !percpu_ref_is_dying(&ctx->refs))) {
                                cond_resched();
                                continue;
                        }
@@ -6324,6 +6325,16 @@ static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
        percpu_ref_kill(&ctx->refs);
        mutex_unlock(&ctx->uring_lock);
 
+       /*
+        * Wait for sq thread to idle, if we have one. It won't spin on new
+        * work after we've killed the ctx ref above. This is important to do
+        * before we cancel existing commands, as the thread could otherwise
+        * be queueing new work post that. If that's work we need to cancel,
+        * it could cause shutdown to hang.
+        */
+       while (ctx->sqo_thread && !wq_has_sleeper(&ctx->sqo_wait))
+               cpu_relax();
+
        io_kill_timeouts(ctx);
        io_poll_remove_all(ctx);