md: only call set_in_sync() when it is expected to succeed.

Since commit 4ad23a9 ("MD: use per-cpu counter for
writes_pending"), set_in_sync() is substantially more expensive: it
can wait for a full RCU grace period which can be 10s of milliseconds.

So we should only call it when the cost is justified.

md_check_recovery() currently calls set_in_sync() every time it finds
anything to do (on non-external active arrays).  For an array
performing resync or recovery, this will be quite often.
Each call will introduce a delay to the md thread, which can noticeable
affect IO submission latency.

In md_check_recovery() we only need to call set_in_sync() if
'safemode' was non-zero at entry, meaning that there has been not
recent IO.  So we save this "safemode was nonzero" state, and only
call set_in_sync() if it was non-zero.

This measurably reduces mean and maximum IO submission latency during
resync/recovery.

Reported-and-tested-by: Jack Wang <jinpu.wang@cloud.ionos.com>
Fixes: 4ad23a9 ("MD: use per-cpu counter for writes_pending")
Cc: stable@vger.kernel.org (v4.12+)
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Song Liu <songliubraving@fb.com>

drivers/md/md.c

@@ -8932,6 +8932,7 @@ void md_check_recovery(struct mddev *mddev)
 
 	if (mddev_trylock(mddev)) {
 		int spares = 0;
+		bool try_set_sync = mddev->safemode != 0;
 
 		if (!mddev->external && mddev->safemode == 1)
 			mddev->safemode = 0;
@@ -8977,7 +8978,7 @@ void md_check_recovery(struct mddev *mddev)
 			}
 		}
 
-		if (!mddev->external && !mddev->in_sync) {
+		if (try_set_sync && !mddev->external && !mddev->in_sync) {
 			spin_lock(&mddev->lock);
 			set_in_sync(mddev);
 			spin_unlock(&mddev->lock);