ocfs2: fix a tiny race when running dirop_fileop_racer
authorYiwen Jiang <jiangyiwen@huawei.com>
Mon, 23 Jun 2014 20:22:09 +0000 (13:22 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Mon, 23 Jun 2014 23:47:45 +0000 (16:47 -0700)
When running dirop_fileop_racer we found a dead lock case.

2 nodes, say Node A and Node B, mount the same ocfs2 volume.  Create
/race/16/1 in the filesystem, and let the inode number of dir 16 is less
than the inode number of dir race.

Node A                            Node B
mv /race/16/1 /race/
                                  right after Node A has got the
                                  EX mode of /race/16/, and tries to
                                  get EX mode of /race
                                  ls /race/16/

In this case, Node A has got the EX mode of /race/16/, and wants to get EX
mode of /race/.  Node B has got the PR mode of /race/, and wants to get
the PR mode of /race/16/.  Since EX and PR are mutually exclusive, dead
lock happens.

This patch fixes this case by locking in ancestor order before trying
inode number order.

Signed-off-by: Yiwen Jiang <jiangyiwen@huawei.com>
Signed-off-by: Joseph Qi <joseph.qi@huawei.com>
Cc: Joel Becker <jlbec@evilplan.org>
Reviewed-by: Mark Fasheh <mfasheh@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fs/ocfs2/namei.c
fs/ocfs2/ocfs2_trace.h

index 5819bb5d64839f52dcdfa6254b67bc7bf7603294..879f9f36b6073be477702e5cf66924dd00262e81 100644 (file)
@@ -991,6 +991,65 @@ leave:
        return status;
 }
 
+static int ocfs2_check_if_ancestor(struct ocfs2_super *osb,
+               u64 src_inode_no, u64 dest_inode_no)
+{
+       int ret = 0, i = 0;
+       u64 parent_inode_no = 0;
+       u64 child_inode_no = src_inode_no;
+       struct inode *child_inode;
+
+#define MAX_LOOKUP_TIMES 32
+       while (1) {
+               child_inode = ocfs2_iget(osb, child_inode_no, 0, 0);
+               if (IS_ERR(child_inode)) {
+                       ret = PTR_ERR(child_inode);
+                       break;
+               }
+
+               ret = ocfs2_inode_lock(child_inode, NULL, 0);
+               if (ret < 0) {
+                       iput(child_inode);
+                       if (ret != -ENOENT)
+                               mlog_errno(ret);
+                       break;
+               }
+
+               ret = ocfs2_lookup_ino_from_name(child_inode, "..", 2,
+                               &parent_inode_no);
+               ocfs2_inode_unlock(child_inode, 0);
+               iput(child_inode);
+               if (ret < 0) {
+                       ret = -ENOENT;
+                       break;
+               }
+
+               if (parent_inode_no == dest_inode_no) {
+                       ret = 1;
+                       break;
+               }
+
+               if (parent_inode_no == osb->root_inode->i_ino) {
+                       ret = 0;
+                       break;
+               }
+
+               child_inode_no = parent_inode_no;
+
+               if (++i >= MAX_LOOKUP_TIMES) {
+                       mlog(ML_NOTICE, "max lookup times reached, filesystem "
+                                       "may have nested directories, "
+                                       "src inode: %llu, dest inode: %llu.\n",
+                                       (unsigned long long)src_inode_no,
+                                       (unsigned long long)dest_inode_no);
+                       ret = 0;
+                       break;
+               }
+       }
+
+       return ret;
+}
+
 /*
  * The only place this should be used is rename!
  * if they have the same id, then the 1st one is the only one locked.
@@ -1002,6 +1061,7 @@ static int ocfs2_double_lock(struct ocfs2_super *osb,
                             struct inode *inode2)
 {
        int status;
+       int inode1_is_ancestor, inode2_is_ancestor;
        struct ocfs2_inode_info *oi1 = OCFS2_I(inode1);
        struct ocfs2_inode_info *oi2 = OCFS2_I(inode2);
        struct buffer_head **tmpbh;
@@ -1015,9 +1075,26 @@ static int ocfs2_double_lock(struct ocfs2_super *osb,
        if (*bh2)
                *bh2 = NULL;
 
-       /* we always want to lock the one with the lower lockid first. */
+       /* we always want to lock the one with the lower lockid first.
+        * and if they are nested, we lock ancestor first */
        if (oi1->ip_blkno != oi2->ip_blkno) {
-               if (oi1->ip_blkno < oi2->ip_blkno) {
+               inode1_is_ancestor = ocfs2_check_if_ancestor(osb, oi2->ip_blkno,
+                               oi1->ip_blkno);
+               if (inode1_is_ancestor < 0) {
+                       status = inode1_is_ancestor;
+                       goto bail;
+               }
+
+               inode2_is_ancestor = ocfs2_check_if_ancestor(osb, oi1->ip_blkno,
+                               oi2->ip_blkno);
+               if (inode2_is_ancestor < 0) {
+                       status = inode2_is_ancestor;
+                       goto bail;
+               }
+
+               if ((inode1_is_ancestor == 1) ||
+                               (oi1->ip_blkno < oi2->ip_blkno &&
+                               inode2_is_ancestor == 0)) {
                        /* switch id1 and id2 around */
                        tmpbh = bh2;
                        bh2 = bh1;
@@ -1135,6 +1212,21 @@ static int ocfs2_rename(struct inode *old_dir,
                        goto bail;
                }
                rename_lock = 1;
+
+               /* here we cannot guarantee the inodes haven't just been
+                * changed, so check if they are nested again */
+               status = ocfs2_check_if_ancestor(osb, new_dir->i_ino,
+                               old_inode->i_ino);
+               if (status < 0) {
+                       mlog_errno(status);
+                       goto bail;
+               } else if (status == 1) {
+                       status = -EPERM;
+                       trace_ocfs2_rename_not_permitted(
+                                       (unsigned long long)old_inode->i_ino,
+                                       (unsigned long long)new_dir->i_ino);
+                       goto bail;
+               }
        }
 
        /* if old and new are the same, this'll just do one lock. */
index 1b60c62aa9d6fa62940c0a9b8f9889943604170e..6cb019b7c6a83c4ec449baff12652e84b5fca06a 100644 (file)
@@ -2292,6 +2292,8 @@ TRACE_EVENT(ocfs2_rename,
                  __entry->new_len, __get_str(new_name))
 );
 
+DEFINE_OCFS2_ULL_ULL_EVENT(ocfs2_rename_not_permitted);
+
 TRACE_EVENT(ocfs2_rename_target_exists,
        TP_PROTO(int new_len, const char *new_name),
        TP_ARGS(new_len, new_name),