locks: perform inode_ref of pl_update_refkeeper outside locked region

Performing inode_ref inside the critical section of pl_update_refkeeper
(around pl_inode_t->mutex) causes a deadlock with inode_table_t->lock.
The other thread (process state dump) holds the inode_table_t->lock first
to loop over inodes and then while dumping an inode's lock context, tries
to acquire pl_inode_t->mutex thereby completing the deadlock.

The fix is to perform inode_ref outside the critical section in
pl_update_refkeeper

Change-Id: I689ab4a9b46b36287740279fdec6159182c86119
BUG: 3753
Reviewed-on: http://review.gluster.com/629
Tested-by: Gluster Build System <jenkins@build.gluster.com>
Reviewed-by: Vijay Bellur <vijay@gluster.com>

1 files changed, 6 insertions, 1 deletions

diff --git a/xlators/features/locks/src/common.c b/xlators/features/locks/src/common.c
index 4e089a0ee36..5e80c16d71a 100644
--- a/xlators/features/locks/src/common.c
+++ b/xlators/features/locks/src/common.c
@@ -400,6 +400,7 @@ pl_update_refkeeper (xlator_t *this, inode_t *inode)
         pl_inode_t *pl_inode  = NULL;
         int         is_empty  = 0;
         int         need_unref = 0;
+        int         need_ref = 0;
 
         pl_inode = pl_inode_get (this, inode);
 
@@ -413,13 +414,17 @@ pl_update_refkeeper (xlator_t *this, inode_t *inode)
                 }
 
                 if (!is_empty && !pl_inode->refkeeper) {
-                        pl_inode->refkeeper = inode_ref (inode);
+                        need_ref = 1;
+                        pl_inode->refkeeper = inode;
                 }
         }
         pthread_mutex_unlock (&pl_inode->mutex);
 
         if (need_unref)
                 inode_unref (inode);
+
+        if (need_ref)
+                inode_ref (inode);
 }