Re: [PATCH 0/11] Per-bdi writeback flusher threads v8

[Jens Axboe <jens.axboe@oracle.com>]

On Thu, May 28 2009, Jan Kara wrote:
> On Wed 27-05-09 20:49:59, Theodore Tso wrote:
> > On Wed, May 27, 2009 at 09:45:43PM +0200, Jens Axboe wrote:
> > > 
> > > This one has been tested good, where good means that it boots and
> > > functions normally at least. Whether it fixes your issue, that would be
> > > interesting to know :-)
> > > 
> > 
> > Unfortunately, it doesn't seem to have.  Here's a dmesg with the
> > softlockup report and the sysrq-t output.  Unfortunately the dmesg
> > file is too big for LKML, so I've compressed it so you can get the
> > whole thing.
>   Everybody waits for sys_sync() to complete and they never seem to be
> woken up. Jens, wb_work_complete() seems a bit fishy - who does
> wb_clear_work() in sync_mode == WB_SYNC_ALL which is on stack?

It is tricky, I looked through that several times. I think
wb_work_complete() should do:

        if (!bdi_work_on_stack(work))
                bdi_work_clear(work);
        if (sync_mode == WB_SYNC_NONE || bdi_work_on_stack(work))
                call_rcu(&work->rcu_head, bdi_work_free);

to have bdi_work_clear() called for the on-stack work. I'll double check
this and roll out a new release, once tested. I haven't been able to
reproduce Ted's problem yet, but perhaps my bdi_alloc() test failures
haven't triggered for WB_SYNC_ALL yet.

> > There's also a lockdep warning which fsx triggered.
>   The lockdep warning is definitely unrelated. It's really a possible
> deadlock, although not quite probable. IMHO the problem is that
> sysfs_mutex gets above mmap_sem due to code in sysfs_readdir which calls
> filldir() which may cause page fault. At the same time it gets quite low
> on the lock stack because filesystems call sysfs functions from their
> internal functions (in this case ext4_put_super) holding quite some locks.
> Adding a few CC's for this.

Thanks, I didn't see any bdi related stuff in there either.

-- 
Jens Axboe