[Cluster-devel] cluster/gfs-kernel/src/gfs glock.c

[bmarzins at sourceware.org]

CVSROOT:	/cvs/cluster
Module name:	cluster
Branch: 	RHEL51
Changes by:	bmarzins at sourceware.org	2008-01-24 20:08:43

Modified files:
	gfs-kernel/src/gfs: glock.c 

Log message:
	Fix for bz #426291.  gfs_glock_dq was traversing the gl_holders list without
	holding the gl_spin spinlock, this was causing a problem when the list item
	it was currently looking at got removed from the list.  The solution is to
	not traverse the list, because it is unncessary. Unfortunately, there is also
	a bug in this section of code, where you can't guarantee that you will not
	cache a glock held with GL_NOCACHE.  Fixing this issue requires significantly
	more work.

Patches:
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/gfs-kernel/src/gfs/glock.c.diff?cvsroot=cluster&only_with_tag=RHEL51&r1=1.29.2.5&r2=1.29.2.5.2.1

--- cluster/gfs-kernel/src/gfs/glock.c	2007/06/26 19:42:31	1.29.2.5
+++ cluster/gfs-kernel/src/gfs/glock.c	2008/01/24 20:08:43	1.29.2.5.2.1
@@ -1618,8 +1618,6 @@
 	struct gfs_sbd *sdp = gl->gl_sbd;
 	struct gfs_glock_operations *glops = gl->gl_ops;
 	struct list_head *pos;
-	struct gfs_holder *tmp_gh = NULL;
-	int count = 0;
 
 	atomic_inc(&gl->gl_sbd->sd_glock_dq_calls);
 
@@ -1630,14 +1628,13 @@
 		set_bit(GLF_SYNC, &gl->gl_flags);
 
 	/* Don't cache glock; request demote to unlock at inter-node scope */
-	if (gh->gh_flags & GL_NOCACHE) {
-		list_for_each(pos, &gl->gl_holders) {
-			tmp_gh = list_entry(pos, struct gfs_holder, gh_list);
-			++count;
-		}
-		if (tmp_gh == gh && count == 1)
-			handle_callback(gl, LM_ST_UNLOCKED);
-	}
+	if (gh->gh_flags & GL_NOCACHE && gl->gl_holders.next == &gh->gh_list &&
+	    gl->gl_holders.prev == &gh->gh_list)
+		/* There's a race here.  If there are two holders, and both
+ 		 * are dq'ed at almost the same time, you can't guarantee that
+ 		 * you will call handle_callback. Fixing this will require
+ 		 * some refactoring */
+		handle_callback(gl, LM_ST_UNLOCKED);
 
 	lock_on_glock(gl);