Re: reiserfs locking (v2)

[Al Viro <viro@ZenIV.linux.org.uk>]

On Sat, Jul 03, 2010 at 10:43:23AM +0100, Al Viro wrote:
> On Sat, Jul 03, 2010 at 10:24:42AM +0100, Al Viro wrote:
> 
> > Gyah...  For the 1001st time: readdir() is far from being the only thing that
> > nests mmap_sem inside i_mutex.  In particular, write() does the same thing.
> > 
> > So yes, it *is* a real deadlock, TYVM, with no directories involved.  Open the
> > same file twice, mmap one fd, close it, then have munmap() hitting i_mutex
> > in reiserfs_file_release() race with write() through another fd.
> > 
> > Incidentally, reiserfs_file_release() checks in the fastpath look completely
> > bogus.  Checking i_count?  What the hell is that one about?  And no, these
> > checks won't stop open() coming between them and grabbing i_mutex, so they
> > couldn't prevent the deadlock in question anyway.
> 
> ... and unfortunately it's been that way since the the initial merge in 2.4.early.
> FWIW, it seems that i_count check was a misguided attempt to check that no other
> opened struct file are there, but it's
> 	a) wrong, since way, _way_ back - open() affects d_count, not i_count
> 	b) wrong even with such modification (consider hardlinks)
> 	c) wrong for even more reasons since forever - i_count and d_count could
> be bumped by many things at any time
> 	d) hopelessly racy anyway, since another open() could very well have
> happened just as we'd finished these checks.

OK...  See 22093b8f3d387f77 in vfs-2.6.git for-next (should propagate to
git.kernel.org shortly).  That ought to deal with this crap, assuming I hadn't
fucked up somewhere...