This is the mail archive of the
cluster-cvs@sourceware.org
mailing list for the cluster.
cluster/gfs2-kernel/src/gfs2 glock.c incore.h ...
- From: teigland at sourceware dot org
- To: cluster-cvs at sources dot redhat dot com
- Date: 15 Nov 2005 20:55:06 -0000
- Subject: cluster/gfs2-kernel/src/gfs2 glock.c incore.h ...
CVSROOT: /cvs/cluster
Module name: cluster
Changes by: teigland@sourceware.org 2005-11-15 20:55:06
Modified files:
gfs2-kernel/src/gfs2: glock.c incore.h ops_fstype.c
Log message:
When using lock_dlm, all gfs2 unmounts would panic in
invalidate_inode_buffers (<- invalidate_list <- invalidate_inodes).
Invalidate_inodes requires that the sb inodes list not change while
it's running, but async unlock completion callbacks from lock_dlm
are scheduled during invalidate_inodes. These callbacks do glock_put()
which does a final iput(), clearing the inode and causing the panic.
The fix is a new semaphore (sd_invalidate_inodes_mutex) which blocks
glock_put's during invalidate_inodes.
Patches:
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/gfs2-kernel/src/gfs2/glock.c.diff?cvsroot=cluster&r1=1.41&r2=1.42
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/gfs2-kernel/src/gfs2/incore.h.diff?cvsroot=cluster&r1=1.21&r2=1.22
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/gfs2-kernel/src/gfs2/ops_fstype.c.diff?cvsroot=cluster&r1=1.34&r2=1.35