dm: limit target bio polling to one shot

dm_poll_bio() is the ->poll_bio() callback for a stacked dm device.
The caller only knows about the dm queue, so it may decide to do a
spinning poll if it thinks a single queue is being polled. Passing those
flags unchanged to the mapped clone lets blk_mq_poll() spin on a target
queue from inside dm_poll_bio().

With io_uring IOPOLL on a dm-stripe target this can keep a task in

  dm_poll_bio() -> bio_poll() -> blk_mq_poll()

long enough to trigger an RCU CPU stall, before io_uring gets back to
io_iopoll_check() and its need_resched() check.

Keep dm's ->poll_bio() bounded by forcing one-shot polling for target
bios. The caller can invoke dm_poll_bio() again if it wants to keep
polling, and it also gets a chance to reap completions or reschedule
between passes.

Fixes: f22ecf9 ("blk-mq: delete task running check in blk_hctx_poll()")
Signed-off-by: Fengnan Chang <changfengnan@bytedance.com>
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>

Author: Fengnan Chang

drivers/md/dm.c

@@ -2098,8 +2098,17 @@ static bool dm_poll_dm_io(struct dm_io *io, struct io_comp_batch *iob,
 	WARN_ON_ONCE(!dm_tio_is_normal(&io->tio));
 
 	/* don't poll if the mapped io is done */
-	if (atomic_read(&io->io_count) > 1)
-		bio_poll(&io->tio.clone, iob, flags);
+	if (atomic_read(&io->io_count) > 1) {
+		/*
+		 * DM hides the target queues from the upper poller, which may
+		 * decide it is safe to spin on a single stacked queue.  Do not
+		 * pass that spinning policy down to a target queue: one slow
+		 * clone could keep the task inside dm_poll_bio() for a long
+		 * time.  Poll target bios once and let the caller decide
+		 * whether to keep polling, reap completions or reschedule.
+		 */
+		bio_poll(&io->tio.clone, iob, flags | BLK_POLL_ONESHOT);
+	}
 
 	/* bio_poll holds the last reference */
 	return atomic_read(&io->io_count) == 1;