io_uring: fix skipping disabling sqo on exec

If there are no requests at the time __io_uring_task_cancel() is called,
tctx_inflight() returns zero and and it terminates not getting a chance
to go through __io_uring_files_cancel() and do
io_disable_sqo_submit(). And we absolutely want them disabled by the
time cancellation ends.

Reported-by: Jens Axboe <[email protected]>
Signed-off-by: Pavel Begunkov <[email protected]>
Fixes: d9d0521 ("io_uring: stop SQPOLL submit on creator's death")
Signed-off-by: Jens Axboe <[email protected]>

Author: isilence

fs/io_uring.c

@@ -9085,6 +9085,10 @@ void __io_uring_task_cancel(void)
 	/* make sure overflow events are dropped */
 	atomic_inc(&tctx->in_idle);
 
+	/* trigger io_disable_sqo_submit() */
+	if (tctx->sqpoll)
+		__io_uring_files_cancel(NULL);
+
 	do {
 		/* read completions before cancelations */
 		inflight = tctx_inflight(tctx);