Re: [RFC][PATCH] reiserfs vs BKL

[Peter Zijlstra]

On Sat, 2007-04-21 at 12:14 -0400, Jeff Mahoney wrote:
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
> 
> Peter Zijlstra wrote:
> > Replace all the lock_kernel() instances with reiserfs_write_lock(sb),
> > and make that use an actual per super-block mutex instead of
> > lock_kernel().
> > 
> > This should make reiserfs safe from PREEMPT_BKL=n, since it seems to
> > rely on being able to schedule. Also, it removes the dependency on the
> > BKL, and thereby is not prone to cause prio inversion with remaining BKL
> > users (notably tty).
> > 
> > Compile tested only, since I didn't dare boot it.
> 
> NACK.

darn, I suspected it would be wrong :-/, it was too easy to be right.

> Believe me, I would *love* to nuke the BKL from reiserfs, but a search
> and replace of this nature is just wrong. reiserfs_write_lock() using
> the BKL isn't an accident - it depends on its nesting properties. If you
> did try to boot this kernel, you'd deadlock pretty quickly.

Right, I see. 

> This one has been on my TODO list for a long time. Interestingly, I've
> been doing reiserfs xattr development recently using 2.6.21-rc7-git2,
> and I'm not seeing any of these messages.

Yeah, that would be pretty close to what I ran.

> # CONFIG_PREEMPT_NONE is not set
> CONFIG_PREEMPT_VOLUNTARY=y
> # CONFIG_PREEMPT is not set
> # CONFIG_PREEMPT_BKL is not set

I have CONFIG_PREEMPT=y, but seeing how one of those points came from a
cond_resched() within a lock_kernel() section, I'm not seeing how you
don't get these.

A well, I really do hope you come up with something some day, for me its
time to go change filesystems - bah, the pain...

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/