Skip to content

Commit

Permalink
reiserfs: fix deadlock with nfs racing on create/lookup
Browse files Browse the repository at this point in the history
Reiserfs is currently able to be deadlocked by having two NFS clients
where one has removed and recreated a file and another is accessing the
file with an open file handle.

If one client deletes and recreates a file with timing such that the
recreated file obtains the same [dirid, objectid] pair as the original
file while another client accesses the file via file handle, the create
and lookup can race and deadlock if the lookup manages to create the
in-memory inode first.

The create thread, in insert_inode_locked4, will hold the write lock
while waiting on the other inode to be unlocked. The lookup thread,
anywhere in the iget path, will release and reacquire the write lock while
it schedules. If it needs to reacquire the lock while the create thread
has it, it will never be able to make forward progress because it needs
to reacquire the lock before ultimately unlocking the inode.

This patch drops the write lock across the insert_inode_locked4 call so
that the ordering of inode_wait -> write lock is retained. Since this
would have been the case before the BKL push-down, this is safe.

Signed-off-by: Jeff Mahoney <jeffm@suse.com>
Signed-off-by: Jan Kara <jack@suse.cz>
  • Loading branch information
Jeff Mahoney authored and Jan Kara committed May 31, 2013
1 parent 4a85701 commit a1457c0
Showing 1 changed file with 7 additions and 2 deletions.
9 changes: 7 additions & 2 deletions fs/reiserfs/inode.c
Original file line number Diff line number Diff line change
Expand Up @@ -1811,11 +1811,16 @@ int reiserfs_new_inode(struct reiserfs_transaction_handle *th,
TYPE_STAT_DATA, SD_SIZE, MAX_US_INT);
memcpy(INODE_PKEY(inode), &(ih.ih_key), KEY_SIZE);
args.dirid = le32_to_cpu(ih.ih_key.k_dir_id);
if (insert_inode_locked4(inode, args.objectid,
reiserfs_find_actor, &args) < 0) {

reiserfs_write_unlock(inode->i_sb);
err = insert_inode_locked4(inode, args.objectid,
reiserfs_find_actor, &args);
reiserfs_write_lock(inode->i_sb);
if (err) {
err = -EINVAL;
goto out_bad_inode;
}

if (old_format_only(sb))
/* not a perfect generation count, as object ids can be reused, but
** this is as good as reiserfs can do right now.
Expand Down

0 comments on commit a1457c0

Please sign in to comment.