LU-7973 llite: lock i_lock/dcache_lock before __d_drop()

There has been several Lustre Client crashes reported by sites
running with Lustre versions 2.1/2.5, all showing the same
dentry->d_hash->next corrupted pointer cause.

This patch fixes a regression that has been introduced since a
long time by commit :
d10c9e05ab593f361fdfd27842766a1924e63e58
(LU-506 kernel: FC15 - support dcache scalability changes.)

where i_lock/dcache_lock protection usage has been removed and
that is likely to cause racy condition during dentry [un]hashing
and to be the root cause of these crashes.

Signed-off-by: Bruno Faccini <bruno.faccini@intel.com>
Change-Id: Ie1239058222f15f7e094d8fae4d37e4a36449707
Reviewed-on: http://review.whamcloud.com/19287
Tested-by: Jenkins
Tested-by: Maloo <hpdd-maloo@intel.com>
Reviewed-by: Lai Siyao <lai.siyao@intel.com>
Reviewed-by: Yang Sheng <yang.sheng@intel.com>
Reviewed-by: Oleg Drokin <oleg.drokin@intel.com>

diff --git a/lustre/llite/file.c b/lustre/llite/file.c
index 2eb5b75..bbcbd29 100644 (file)
--- a/lustre/llite/file.c
+++ b/lustre/llite/file.c
@@ -3372,8 +3372,11 @@ static int __ll_inode_revalidate(struct dentry *dentry, __u64 ibits)
                    do_lookup() -> ll_revalidate_it(). We cannot use d_drop
                    here to preserve get_cwd functionality on 2.6.
                    Bug 10503 */
-               if (!dentry->d_inode->i_nlink)
+               if (!dentry->d_inode->i_nlink) {
+                       ll_lock_dcache(inode);
                        d_lustre_invalidate(dentry, 0);
+                       ll_unlock_dcache(inode);
+               }
 
                 ll_lookup_finish_locks(&oit, dentry);
         } else if (!ll_have_md_lock(dentry->d_inode, &ibits, LCK_MINMODE)) {