Re: linux-next: reiserfs lockout with xattr ?

[Frederic Weisbecker <fweisbec@gmail.com>]

On Mon, Aug 10, 2009 at 10:08:26PM +0200, Laurent Riffard wrote:
> Le 10/08/2009 01:33, Frederic Weisbecker a écrit :
> > 
> > Thanks a lot for this report. Reiserfs xattrs are definetly paths I
> > haven't tested deeply.
> > 
> > I'm trying to reproduce these issues and will let you know once I get
> > a fix.
> > 
> > (Btw, if you're still running this tree, you should update with the
> > latest patch I just pushed, it's a very important fix).
> > 
> > Thanks,
> > 
> > Frederic.
> 
> I tried with your lastest patch (git head is at "kill-the-bkl/reiserfs: 
> fix early readdir offset increment"), and the lockout still happens.
> 
> ~~
> laurent


Ok, it looks like the last two patches in my tree have introduced a locking
inversion dependency between the reiserfs write lock and mm->mmap_sem.

In filldir, we take the reiserfs lock and we might take mm->mmap_sem if
we fault while copying the dir entry to the user.

In mmap, the mm->mmap_sem is taken before the call to the fs mmap callback
and there we take the reiserfs lock.

Locking inversion: AB - BA

What is really sad here is that I have no choice: I can't take the reiserfs lock
while filling a dir entry to the user.
This part, introduced in the last two patches of the tree optimized a lot
the throughput, and now I have to revert them :-(