From: Abhijith Das <adas@redhat.com>
Date: Fri, 10 Jul 2009 13:42:03 -0500
Subject: [gfs2] umount.gfs2 hangs eating CPU
Message-id: 4A578B7B.2090105@redhat.com
O-Subject: [RHEL5.4 PATCH][GFS2] bz 508876 - umount.gfs2 hangs eating CPU
Bugzilla: 508876
RH-Acked-by: Steven Whitehouse <swhiteho@redhat.com>

This is the fix from upstream. It moves the sem from around the code which
receives messages from the dlm to being around the code which processes the
particular messages which cause the race during umount. This means that other
messages from the dlm can continue to be processed during this time.

Since the write side of the lock is held for all of the invalidate_inodes()
routine, the read side acts like a barrier during that process.

This is entirely due to the extra set of inodes that we are carrying around,
and once we are rid of those, this need for this rw semaphore will go away.

Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Signed-off-by: Abhijith Das <adas@redhat.com>

diff --git a/fs/gfs2/glock.c b/fs/gfs2/glock.c
index 03409a2..875a3af 100644
--- a/fs/gfs2/glock.c
+++ b/fs/gfs2/glock.c
@@ -695,6 +695,7 @@ static void glock_work_func(void *data)
 
 	if (test_and_clear_bit(GLF_REPLY_PENDING, &gl->gl_flags))
 		finish_xmote(gl, gl->gl_reply);
+	down_read(&gfs2_umount_flush_sem);
 	spin_lock(&gl->gl_spin);
 	if (test_and_clear_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
 	    gl->gl_state != LM_ST_UNLOCKED &&
@@ -707,6 +708,7 @@ static void glock_work_func(void *data)
 	}
 	run_queue(gl, 0);
 	spin_unlock(&gl->gl_spin);
+	up_read(&gfs2_umount_flush_sem);
 	if (!delay ||
 	    queue_delayed_work(glock_workqueue, &gl->gl_work, delay) == 0)
 		gfs2_glock_put(gl);
@@ -1365,7 +1367,6 @@ void gfs2_glock_cb(void *cb_data, unsigned int type, void *data)
 		struct lm_async_cb *async = data;
 		struct gfs2_glock *gl;
 
-		down_read(&gfs2_umount_flush_sem);
 		gl = gfs2_glock_find(sdp, &async->lc_name);
 		if (gfs2_assert_warn(sdp, gl))
 			return;
@@ -1373,7 +1374,6 @@ void gfs2_glock_cb(void *cb_data, unsigned int type, void *data)
 		set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
 		if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
 			gfs2_glock_put(gl);
-		up_read(&gfs2_umount_flush_sem);
 		return;
 	}