From: Abhijith Das <adas@redhat.com>
Date: Thu, 16 Apr 2009 10:07:48 -0500
Subject: [gfs2] blocked after recovery
Message-id: 49E749C4.1010809@redhat.com
O-Subject: [RHEL5.4] [PATCH] GFS2 - bz 483541 - gfs2 blocked after recovery
Bugzilla: 483541
RH-Acked-by: Steven Whitehouse <swhiteho@redhat.com>
RH-Acked-by: Bob Peterson <rpeterso@redhat.com>
RH-Acked-by: Josef Bacik <josef@redhat.com>

The problem is that the glock workqueue is getting stuck
while trying to get the transaction lock. Its trying to
get that lock since it needs to start a transaction in
order to release a glock (due to GFS2's journalling
implementation). Thats then blocking incoming requests
from being processed.

This patch changes the log code so that we don't need to
get a lock at this particular point in time no longer do
we have to wonder about what to do if the transaction
cannot be allocated, since we can put that on the stack.
Also it should be slightly faster, since we don't need to
grab the transaction glock in this case.

The patch relies on the fact that we cache the transaction
glock forever, unless:
a) There is a demotion request (which syncs the filesystem
as part of its glops code) or
b) Upon umount

We can thus assume that if there is something in the ail
list, it must be because we have the transaction glock
cached. If we didn't hold it, then that would mean that the
filesystem has been synced and further transactions would
have been blocked in gfs2_trans_begin() so that the ail
list must be empty.

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

diff --git a/fs/gfs2/glops.c b/fs/gfs2/glops.c
index 7ce0016..3cf72b3 100644
--- a/fs/gfs2/glops.c
+++ b/fs/gfs2/glops.c
@@ -38,20 +38,25 @@
 static void gfs2_ail_empty_gl(struct gfs2_glock *gl)
 {
 	struct gfs2_sbd *sdp = gl->gl_sbd;
-	unsigned int blocks;
 	struct list_head *head = &gl->gl_ail_list;
 	struct gfs2_bufdata *bd;
 	struct buffer_head *bh;
-	int error;
+	struct gfs2_trans tr;
 
-	blocks = atomic_read(&gl->gl_ail_count);
-	if (!blocks)
-		return;
+	memset(&tr, 0, sizeof(tr));
+	tr.tr_revokes = atomic_read(&gl->gl_ail_count);
 
-	error = gfs2_trans_begin(sdp, 0, blocks);
-	if (gfs2_assert_withdraw(sdp, !error))
+	if (!tr.tr_revokes)
 		return;
 
+	/* A shortened, inline version of gfs2_trans_begin() */
+	tr.tr_reserved = 1 + gfs2_struct2blk(sdp, tr.tr_revokes, sizeof(u64));
+	tr.tr_ip = (unsigned long)__builtin_return_address(0);
+	INIT_LIST_HEAD(&tr.tr_list_buf);
+	gfs2_log_reserve(sdp, tr.tr_reserved, 1);
+	BUG_ON(current->journal_info);
+	current->journal_info = &tr;
+
 	gfs2_log_lock(sdp);
 	while (!list_empty(head)) {
 		bd = list_entry(head->next, struct gfs2_bufdata,
diff --git a/fs/gfs2/trans.c b/fs/gfs2/trans.c
index c224ac8..0bcbfc1 100644
--- a/fs/gfs2/trans.c
+++ b/fs/gfs2/trans.c
@@ -88,9 +88,11 @@ void gfs2_trans_end(struct gfs2_sbd *sdp)
 
 	if (!tr->tr_touched) {
 		gfs2_log_release(sdp, tr->tr_reserved);
-		gfs2_glock_dq(&tr->tr_t_gh);
-		gfs2_holder_uninit(&tr->tr_t_gh);
-		kfree(tr);
+		if (tr->tr_t_gh.gh_gl) {
+			gfs2_glock_dq(&tr->tr_t_gh);
+			gfs2_holder_uninit(&tr->tr_t_gh);
+			kfree(tr);
+		}
 		return;
 	}
 
@@ -106,9 +108,11 @@ void gfs2_trans_end(struct gfs2_sbd *sdp)
 	}
 
 	gfs2_log_commit(sdp, tr);
-        gfs2_glock_dq(&tr->tr_t_gh);
-        gfs2_holder_uninit(&tr->tr_t_gh);
-        kfree(tr);
+	if (tr->tr_t_gh.gh_gl) {
+		gfs2_glock_dq(&tr->tr_t_gh);
+		gfs2_holder_uninit(&tr->tr_t_gh);
+		kfree(tr);
+	}
 
 	if (sdp->sd_vfs->s_flags & MS_SYNCHRONOUS)
 		gfs2_log_flush(sdp, NULL);