From: johann <johann>
Date: Thu, 12 Nov 2009 09:57:44 +0000 (+0000)
Subject: Branch b_release_1_8_1
X-Git-Tag: GIT_EPOCH_B_RELEASE_1_8_1~2^2~4
X-Git-Url: https://git.whamcloud.com/?a=commitdiff_plain;h=755867f781a218cbd4d794b2835361db7479516e;p=fs%2Flustre-release.git

Branch b_release_1_8_1
b=20892
i=shadow
i=yangsheng

Use lock_rename instead of the BKL.
---

diff --git a/lustre/ChangeLog b/lustre/ChangeLog
index d1330a8..0f6940a 100644
--- a/lustre/ChangeLog
+++ b/lustre/ChangeLog
@@ -12,6 +12,12 @@ TBD Sun Microsystems, Inc.
 	* ext4 support for RHEL5 is experimental and thus should not be
 	  used in production.
 
+Severity   : major
+Bugzilla   : 20892
+Description: bad entry in directory xxx: inode out of bounds
+Details    : fix locking issue in the rename path which could race with any
+	     other operations updating the same directory.
+
 -------------------------------------------------------------------------------
 
 2009-10-16 Sun Microsystems, Inc.
diff --git a/lustre/mds/mds_fs.c b/lustre/mds/mds_fs.c
index 44f2560..bca2e22 100644
--- a/lustre/mds/mds_fs.c
+++ b/lustre/mds/mds_fs.c
@@ -1001,10 +1001,8 @@ int mds_obd_create(struct obd_export *exp, struct obdo *oa,
         if (IS_ERR(handle))
                 GOTO(out_dput2, rc = PTR_ERR(handle));
 
-        lock_kernel();
         rc = ll_vfs_rename(parent_inode, dchild, mds->mds_vfsmnt,
                            parent_inode, new_child, mds->mds_vfsmnt);
-        unlock_kernel();
         if (rc)
                 CERROR("error renaming new object "LPU64":%u: rc %d\n",
                        oa->o_id, oa->o_generation, rc);
diff --git a/lustre/mds/mds_reint.c b/lustre/mds/mds_reint.c
index e39f313..73ae161 100644
--- a/lustre/mds/mds_reint.c
+++ b/lustre/mds/mds_reint.c
@@ -2490,6 +2490,7 @@ static int mds_reint_rename(struct mds_update_record *rec, int offset,
         struct dentry *de_tgtdir = NULL;
         struct dentry *de_old = NULL;
         struct dentry *de_new = NULL;
+        struct dentry *trap;
         struct inode *old_inode = NULL, *new_inode = NULL;
         struct inode *inodes[PTLRPC_NUM_VERSIONS] = { NULL };
         struct mds_obd *mds = mds_req2mds(req);
@@ -2604,11 +2605,6 @@ no_unlink:
         OBD_FAIL_WRITE(obd, OBD_FAIL_MDS_REINT_RENAME_WRITE,
                        de_srcdir->d_inode->i_sb);
 
-        /* Check if we are moving old entry into its child. 2.6 does not
-           check for this in vfs_rename() anymore */
-        if (is_subdir(de_new, de_old))
-                GOTO(cleanup, rc = -EINVAL);
-
         lmm = lustre_msg_buf(req->rq_repmsg, offset + 1, 0);
         /* check that lmm size is not 0 */
         sz = lustre_msg_buflen(req->rq_repmsg, offset + 1) > 0 ?
@@ -2619,13 +2615,23 @@ no_unlink:
         if (IS_ERR(handle))
                 GOTO(cleanup, rc = PTR_ERR(handle));
 
-        lock_kernel();
+        trap = lock_rename(de_tgtdir, de_srcdir);
+        /* source should not be ancestor of target */
+        if (de_old == trap) {
+                unlock_rename(de_tgtdir, de_srcdir);
+                GOTO(cleanup, rc = -EINVAL);
+        }
+        /* target should not be an ancestor of source */
+        if (de_new == trap) {
+                unlock_rename(de_tgtdir, de_srcdir);
+                GOTO(cleanup, rc = -ENOTEMPTY);
+        }
         de_old->d_fsdata = req;
         de_new->d_fsdata = req;
 
-        rc = ll_vfs_rename(de_srcdir->d_inode, de_old, mds->mds_vfsmnt, 
+        rc = ll_vfs_rename(de_srcdir->d_inode, de_old, mds->mds_vfsmnt,
                            de_tgtdir->d_inode, de_new, mds->mds_vfsmnt);
-        unlock_kernel();
+        unlock_rename(de_tgtdir, de_srcdir);
 
         if (rc == 0) {
                 struct iattr iattr;