Check the ipt.error value, it must have been either cleared to zero or
set to another error than the default -EINVAL if we don't go through the
waitqueue proc addition. Just give up on poll at that point and return
failure, this will fallback to async work.
io_poll_add() doesn't suffer from this failure case, as it returns the
error value directly.
Cc: stable@xxxxxxxxxxxxxxx # v5.7+
Reported-by: syzbot+a730016dc0bdce4f6ff5@xxxxxxxxxxxxxxxxxxxxxxxxx
Signed-off-by: Jens Axboe <axboe@xxxxxxxxx>
---
diff --git a/fs/io_uring.c b/fs/io_uring.c
index 99582cf5106b..8a2afd8c33c9 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -4883,7 +4883,7 @@ static bool io_arm_poll_handler(struct io_kiocb *req)
ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
io_async_wake);
- if (ret) {
+ if (ret || ipt.error) {
io_poll_remove_double(req, apoll->double_poll);
spin_unlock_irq(&ctx->completion_lock);
kfree(apoll->double_poll);
--
Jens Axboe
