md: wake raid456 reshape waiters before suspend

[ Upstream commit cf86bb5 ]

During raid456 reshape, direct IO across the reshape position can sleep
in raid5_make_request() waiting for reshape progress while still
holding an active_io reference. If userspace then freezes reshape and
writes md/suspend_lo or md/suspend_hi, mddev_suspend() kills active_io
and waits for all in-flight IO to drain.

This can deadlock: the IO needs reshape progress to continue, but the
reshape thread is already frozen, so the active_io reference is never
dropped and suspend never completes.

raid5_prepare_suspend() already wakes wait_for_reshape for dm-raid. Do
the same for normal md suspend when reshape is already interrupted, so
waiting raid456 IO can abort, drop its reference, and let suspend
finish.

The mdadm test tests/25raid456-reshape-deadlock reproduces the hang.

Fixes: 714d201 ("md: add new helpers to suspend/resume array")
Link: https://lore.kernel.org/linux-raid/20260327140729.2030564-1-yukuai@fnnas.com/
Signed-off-by: Yu Kuai <yukuai@fnnas.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>

Author: hailan94

drivers/md/md.c

@@ -486,6 +486,17 @@ int mddev_suspend(struct mddev *mddev, bool interruptible)
 	}
 
 	percpu_ref_kill(&mddev->active_io);
+
+	/*
+	 * RAID456 IO can sleep in wait_for_reshape while still holding an
+	 * active_io reference. If reshape is already interrupted or frozen,
+	 * wake those waiters so they can abort and drop the reference instead
+	 * of deadlocking suspend.
+	 */
+	if (mddev->pers && mddev->pers->prepare_suspend &&
+	    reshape_interrupted(mddev))
+		mddev->pers->prepare_suspend(mddev);
+
 	if (interruptible)
 		err = wait_event_interruptible(mddev->sb_wait,
 				percpu_ref_is_zero(&mddev->active_io));