udf deadlock (was Re: hugetlbfs lockdep spew revisited.)

udf deadlock (was Re: hugetlbfs lockdep spew revisited.)

[Al Viro]

On Fri, Feb 17, 2012 at 12:49:22AM +0000, Al Viro wrote:
> Folks, this is not a false positive and it has nothing to do with misannotation
> for directories.  Deadlock is real; I have no idea WTF do we what ->i_mutex
> held over that area in hugetlbfs ->mmap(), but doing that is really, really
> wrong, whatever the reason.

Arrrrgh...  Some grepping around has uncovered another deadlock on
i_mutex/mmap_sem and this one is not hard to hit at all.

Thread A:
	opens file on UDF (O_RDWR open)
	does big, fat write() to it
Thread B:
	opens the same file (also O_RDWR)
	mmaps it
	closes
	does munmap()

and there we are - munmap() will end up closing the second struct file,
call udf_release_file() and we are hitting ->i_mutex while under
->mmap_sem.  Blocking on it, actually, since generic_file_aio_write()
in the first thread is holding ->i_mutex.  And as soon as thread A gets
around to faulting the next piece of data in, well...  To widen the
window a lot, mmap something large sitting on NFS and do write() from
that mmapped area.  Race window as wide as one could ask for...

What happens there is prealloc discard on close; do we even want ->i_mutex
there these days?  Note that there's also
	down_write(&UDF_I(inode)->i_data_sem);
in udf_release_file()...

Re: udf deadlock (was Re: hugetlbfs lockdep spew revisited.)

[Jan Kara]

On Fri 17-02-12 17:48:18, Al Viro wrote:
> On Fri, Feb 17, 2012 at 12:49:22AM +0000, Al Viro wrote:
> > Folks, this is not a false positive and it has nothing to do with misannotation
> > for directories.  Deadlock is real; I have no idea WTF do we what ->i_mutex
> > held over that area in hugetlbfs ->mmap(), but doing that is really, really
> > wrong, whatever the reason.
> 
> Arrrrgh...  Some grepping around has uncovered another deadlock on
> i_mutex/mmap_sem and this one is not hard to hit at all.
> 
> Thread A:
> 	opens file on UDF (O_RDWR open)
> 	does big, fat write() to it
> Thread B:
> 	opens the same file (also O_RDWR)
> 	mmaps it
> 	closes
> 	does munmap()
> 
> and there we are - munmap() will end up closing the second struct file,
> call udf_release_file() and we are hitting ->i_mutex while under
> ->mmap_sem.  Blocking on it, actually, since generic_file_aio_write()
> in the first thread is holding ->i_mutex.  And as soon as thread A gets
> around to faulting the next piece of data in, well...  To widen the
> window a lot, mmap something large sitting on NFS and do write() from
> that mmapped area.  Race window as wide as one could ask for...
  Right, I didn't realize ->release() may be called with mmap_sem held.
Thanks for spotting this.  BTW: Documentation/filesystems/Locking might
need an update since it states:
locking rules:
        All may block except for ->setlease.
        No VFS locks held on entry except for ->setlease.

> What happens there is prealloc discard on close; do we even want ->i_mutex
> there these days?  Note that there's also
> 	down_write(&UDF_I(inode)->i_data_sem);
> in udf_release_file()...
  I've looked around and it seems we don't need i_mutex for anything.
i_data_sem should be enough. So I'll just remove i_mutex.

								Honza
-- 
Jan Kara <jack@suse.cz>
SUSE Labs, CR