On Wed, Sep 4, 2019 at 1:23 PM Aleksa Sarai cyphar@cyphar.com wrote:
This patch allows for LOOKUP_BENEATH and LOOKUP_IN_ROOT to safely permit ".." resolution (in the case of LOOKUP_BENEATH the resolution will still fail if ".." resolution would resolve a path outside of the root -- while LOOKUP_IN_ROOT will chroot(2)-style scope it). Magic-link jumps are still disallowed entirely because now they could result in inconsistent behaviour if resolution encounters a subsequent ".."[*].
This is the only patch in the series that makes me go "umm".
Why is it ok to re-initialize m_seq, which is used by other things too? I think it's because we're out of RCU lookup, but there's no comment about it, and it looks iffy to me. I'd rather have a separate sequence count that doesn't have two users with different lifetime rules.
But even apart from that, I think from a "patch continuity" standpoint it would be better to introduce the sequence counts as just an error condition first - iow, not have the "path_is_under()" check, but just return -EXDEV if the sequence number doesn't match.
So you'd have three stages:
1) ".." always returns -EXDEV
2) ".." returns -EXDEV if there was a concurrent rename/mount
3) ".." returns -EXDEV if there was a concurrent rename/mount and we reset the sequence numbers and check if you escaped.
becasue the sequence number reset really does make me go "hmm", plus I get this nagging little feeling in the back of my head that you can cause nasty O(n^2) lookup cost behavior with deep paths, lots of "..", and repeated path_is_under() calls.
So (1) sounds safe. (2) sounds simple. And (3) is where I think subtle things start happening.
Also, I'm not 100% convinced that (3) is needed at all. I think the retry could be done in user space instead, which needs to have a fallback anyway. Yes? No?
Linus