On Fri, Nov 29, 2013 at 04:14:16AM +0000, Al Viro wrote:
> And yes, it has fixed the problem with generic/234. I'll do full xfstests
> run to see if there's anything else, but this one is obviously needed.
> I'll send it with sane commit message (along with follow_dotdot_rcu()
> fix) later tonight. path_init() race is a separate story - that one should
> probably go separately, since we'll want it in all branches starting with
> early 2011 or so.
OK, it survives. However, looking a bit more at follow_dotdot_rcu()...
AFAICS, we have a narrow oopsable race, from 2.6.38 and to 3.12 - think what
happens if we are walking through /tmp/foo/bar/../baz in RCU mode and we'd just
reached /tmp/foo/bar. handle_dotdot() is called, calling follow_dotdot_rcu().
OK, we are not about to cross a mountpoint. Read ->d_seq of /tmp/foo into
seq, check that nd->seq matches /tmp/foo/bar (it does, everything's fine)
and set nd->path.dentry to /tmp/foo, with nd->seq set to seq. Then
we check if the /tmp/foo is overmounted by something; it isn't and now we set
Sure, it's _very_ hard to get into trouble here - we need somebody to remove
/tmp/foo/bar *and* /tmp/foo while we'd been walking vfsmount hash,
but in theory it is not impossible to get NULL nd->inode. Then
link_path_walk() gets to checking that we have a directory and we get
an oops on checking inode flags.
I really don't like the way we have nd->inode updates scattered all over
the place in fs/namei.c ;-/ I'm looking into possible ways to deal
with it sanely, but that'll have to wait for tomorrow...
Anyway, I've pushed the minimal regression fix to vfs.git; please, pull
Al Viro (1):
fix bogus path_put() of nd->root after some unlazy_walk() failures
fs/namei.c | 3 +--
1 files changed, 1 insertions(+), 2 deletions(-)