Re: 3.4.4-rt13: btrfs + xfstests 006 = BOOM.. and a bonus rt_mutex deadlock report for absolutely free!

[Chris Mason <chris.mason@fusionio.com>]

On Sat, Jul 14, 2012 at 04:14:43AM -0600, Mike Galbraith wrote:
> On Fri, 2012-07-13 at 08:50 -0400, Chris Mason wrote: 
> > On Wed, Jul 11, 2012 at 11:47:40PM -0600, Mike Galbraith wrote:
> > > Greetings,
> > 
> > [ deadlocks with btrfs and the recent RT kernels ]
> > 
> > I talked with Thomas about this and I think the problem is the
> > single-reader nature of the RW rwlocks.  The lockdep report below
> > mentions that btrfs is calling:
> > 
> > > [  692.963099]  [<ffffffff811fabd2>] btrfs_clear_path_blocking+0x32/0x70
> > 
> > In this case, the task has a number of blocking read locks on the btrfs buffers,
> > and we're trying to turn them back into spinning read locks.  Even
> > though btrfs is taking the read rwlock, it doesn't think of this as a new
> > lock operation because we were blocking out new writers.
> > 
> > If the second task has taken the spinning read lock, it is going to
> > prevent that clear_path_blocking operation from progressing, even though
> > it would have worked on a non-RT kernel.
> > 
> > The solution should be to make the blocking read locks in btrfs honor the
> > single-reader semantics.  This means not allowing more than one blocking
> > reader and not allowing a spinning reader when there is a blocking
> > reader.  Strictly speaking btrfs shouldn't need recursive readers on a
> > single lock, so I wouldn't worry about that part.
> > 
> > There is also a chunk of code in btrfs_clear_path_blocking that makes
> > sure to strictly honor top down locking order during the conversion.  It
> > only does this when lockdep is enabled because in non-RT kernels we
> > don't need to worry about it.  For RT we'll want to enable that as well.
> > 
> > I'll give this a shot later today.
> 
> I took a poke at it.  Did I do something similar to what you had in
> mind, or just hide behind performance stealing paranoid trylock loops?
> Box survived 1000 x xfstests 006 and dbench [-s] massive right off the
> bat, so it gets posted despite skepticism.

Great, thanks!  I got stuck in bug land on Friday.  You mentioned
performance problems earlier on Saturday, did this improve performance?

One other question:

>  again:
> +#ifdef CONFIG_PREEMPT_RT_BASE
> +	while (atomic_read(&eb->blocking_readers))
> +		cpu_chill();
> +	while(!read_trylock(&eb->lock))
> +		cpu_chill();
> +	if (atomic_read(&eb->blocking_readers)) {
> +		read_unlock(&eb->lock);
> +		goto again;
> +	}

Why use read_trylock() in a loop instead of just trying to take the
lock?  Is this an RTism or are there other reasons?  

-chris