Messages in this thread | | | Date | Thu, 20 Mar 2014 04:21:55 +0000 | From | Al Viro <> | Subject | Re: [PATCH] fs/namespace: don't clobber mnt_hash.next while umounting [v2] |
| |
On Wed, Mar 19, 2014 at 09:02:33PM -0700, Linus Torvalds wrote: > Quite frankly, if that's the main issue, then may I suggest aiming to > use a 'hlist' instead of a doubly-linked list? Those have the > advantage that they are NULL-terminated. > > Yeah, hlists have some disadvantages too, which might not make them > work in this case, but really, for mnt_hash? hlists are generally > *exactly* what you want for hash lists, because the head is smaller. > And because of the NULL termination rather than having the head used > in the middle of a circular list, you don't get the termination > problems when moving entries across chains. > > I did not look whether there was some reason a hlist isn't appropriate > here. Maybe you can tell me.
Er... I have, actually, right in the part you've snipped ;-)
<unsnip> I would prefer to deal with (1) by turning mnt_hash into hlist; the problem with that is __lookup_mnt_last(). That sucker is only called under mount_lock, so RCU issues do not play there, but it's there and it complicates things. There might be a way to get rid of that thing for good, but that's more invasive than what I'd be happy with for backports. </unsnip>
hlist _is_ better, no questions there, but surgery required to deal with __lookup_mnt_last()[1] is too invasive for backports and even more so - for -final. I would prefer to have the merge window happen after LSF/MM, obviously, but I thought you wanted to open it this Sunday?
[1] that is, with cases like "/tmp/b is a slave of /tmp/a, bind foo on /tmp/b/c, then bind bar on /tmp/a/c, then umount /tmp/a/c". The only kinda-sorta sane semantics we'd been able to come up with is what we do right now and that's where __lookup_mnt_last() has come from.
| |