Re: VFS deadlock ?

From: Linus Torvalds
Date: Thu Mar 21 2013 - 20:15:25 EST


On Thu, Mar 21, 2013 at 5:08 PM, Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
>
> Not really. Do that and yes, this deadlock goes away. But the locking
> order in general goes to hell - we order directory inodes by "which dentry
> is an ancestor of another?" So we have no warranty that we won't get
> alias1/foo/bar/baz < alias2/foo. Take rename_lock() on those two and
> have it race with rmdir alias2/foo/bar/baz (locks alias2/foo/bar, then
> alias2/foo/bar/baz) and rmdir alias2/foo/bar (locks alias2/foo and
> alias2/foo/bar). Oops - we have a cycle now...

Hmm. But again, that can't actually happen here. We're in /proc. You
can't move the entries around. Also, we only changed the locking order
for the "inode is identical" case where we take only *one* lock, we
didn't change it for the cases where we take multiple locks (and order
them topologically).

So I agree that we need to avoid aliased directories in the *general*
case. I'm just arguing that for the specific case of /proc, we should
be ok. No?

Linus
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/