Re: 2.6.35-rc4-git3: Reported regressions from 2.6.34

[Frederic Weisbecker <fweisbec@gmail.com>]

On Thu, Jul 08, 2010 at 06:34:25PM -0700, Linus Torvalds wrote:
> On Thu, Jul 8, 2010 at 4:33 PM, Rafael J. Wysocki <rjw@sisk.pl> wrote:
> > Bug-Entry       : http://bugzilla.kernel.org/show_bug.cgi?id=16334
> > Subject         : reiserfs locking (v2)
> > Submitter       : Sergey Senozhatsky <sergey.senozhatsky@gmail.com>
> > Date            : 2010-07-02 9:34 (7 days old)
> > Message-ID      : <20100702093451.GA3973@swordfish.minsk.epam.com>
> > References      : http://marc.info/?l=linux-kernel&m=127806306303590&w=2
> 
> Frederic? Al? I assume this is some late fallout from the BKL removal
> ages ago.. It's the old filldir-vs-mmap crud, but normally it should
> be impossible to trigger because the inode for a directory should
> never be mmap'able, so we should never have the same i_mutex lock used
> for both mmap and for filldir protection.
> 
> We saw some of that oddity long ago, I wonder if it's lockdep being
> confused about some inodes.



I think it has been there from the beginning. At least it was there before
the reiserfs bkl removal in .32.


Indeed the readdir <-> unmap/release inversion problem can not happen.
But Al said that can happen between write and release. (Although I don't see
where write takes the inode mutex).

He also highlighted the fact that reiserfs refcounting based on i_count
was totally broken.

He has a fix the whole in the vfs tree, in the for-next branch on commit
6c2bdaf089a3876226893fab00dd83596c465ad2
"Fix reiserfs_file_release()"

No more uses of the i_mutex on release after that, nor i_count, but a private
openers refcount and a tailpack mutex per reiserfs inode.