Re: [PATCH] xfs: don't sleep in xlog_cil_force_lsn on shutdown

[Dave Chinner <david@fromorbit.com>]

On Tue, May 06, 2014 at 09:27:46AM -0400, Brian Foster wrote:
> On Tue, May 06, 2014 at 11:04:40AM +1000, Dave Chinner wrote:
> > From: Dave Chinner <dchinner@redhat.com>
> > 
> > Reports of a shutdown hang when fsyncing a directory have surfaced,
> > such as this:
> > 
> > [ 3663.394472] Call Trace:
> > [ 3663.397199]  [<ffffffff815f1889>] schedule+0x29/0x70
> > [ 3663.402743]  [<ffffffffa01feda5>] xlog_cil_force_lsn+0x185/0x1a0 [xfs]
> > [ 3663.416249]  [<ffffffffa01fd3af>] _xfs_log_force_lsn+0x6f/0x2f0 [xfs]
> > [ 3663.429271]  [<ffffffffa01a339d>] xfs_dir_fsync+0x7d/0xe0 [xfs]
> > [ 3663.435873]  [<ffffffff811df8c5>] do_fsync+0x65/0xa0
> > [ 3663.441408]  [<ffffffff811dfbc0>] SyS_fsync+0x10/0x20
> > [ 3663.447043]  [<ffffffff815fc7d9>] system_call_fastpath+0x16/0x1b
> > 
> > If we trigger a shutdown in xlog_cil_push() from xlog_write(), we
> > will never wake waiters on the current push sequence number, so
> > anything waiting in xlog_cil_force_lsn() for that push sequence
> > number to come up will not get woken and hence stall the shutdown.
> > 
> > Fix this by ensuring we call wake_up_all(&cil->xc_commit_wait) in
> > the push abort handling, in the log shutdown code when waking all
> > waiters, and adding a shutdown check in the sequence completion wait
> > loops to ensure they abort when a wakeup due to a shutdown occurs.
> > 
> > Reported-by: Boris Ranto <branto@redhat.com>
> > Reported-by: Eric Sandeen <esandeen@redhat.com>
> > Signed-off-by: Dave Chinner <dchinner@redhat.com>
> > ---
> 
> Previously posted here, for reference:
> 
> http://oss.sgi.com/archives/xfs/2014-04/msg00801.html
> 
> >  fs/xfs/xfs_log.c     |  7 +++++--
> >  fs/xfs/xfs_log_cil.c | 50 ++++++++++++++++++++++++++++++++++++++++++--------
> >  2 files changed, 47 insertions(+), 10 deletions(-)
> > 
> > diff --git a/fs/xfs/xfs_log.c b/fs/xfs/xfs_log.c
> > index a5f8bd9..dbba2d7 100644
> > --- a/fs/xfs/xfs_log.c
> > +++ b/fs/xfs/xfs_log.c
> > @@ -3952,11 +3952,14 @@ xfs_log_force_umount(
> >  		retval = xlog_state_ioerror(log);
> >  		spin_unlock(&log->l_icloglock);
> >  	}
> > +
> >  	/*
> > -	 * Wake up everybody waiting on xfs_log_force.
> > -	 * Callback all log item committed functions as if the
> > +	 * Wake up everybody waiting on xfs_log_force. This needs to wake anyone
> > +	 * waiting on a CIL push that is issued as part of a log force first
> > +	 * before running the log item committed callback functions as if the
> >  	 * log writes were completed.
> >  	 */
> > +	wake_up_all(&log->l_cilp->xc_commit_wait);
> >  	xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
> >  
> 
> This looks fine to me with the defensive reasoning described in the
> aforementioned link, but it also looks like it could race with a force
> and sleep because we don't take xc_push_lock. We take the lock for the
> same wake up down in xlog_cil_committed(), so a hang seems unlikely at
> this point.

We can't really race in any meaningful way- the filesystem and log
are already marked as shut down. Hence any new sleeper at this point
will detect a shutdown before trying to sleep. Even if we do race,
the xlog_cil_committed() will catch any stragglers...

> Given that the comment is kind of wordy (and unless we want to do the
> locking here as well), could we update the comment to reflect this?
> E.g., something like:
> 
> 	/*
> 	 * Wake up everybody waiting on a CIL push and/or log force. Wake the
> 	 * CIL push first as if the log writes were completed. The abort
> 	 * handling in the log item committed callback functions will do this
> 	 * again under lock to avoid races.
> 	 */
> 
> Thoughts?

Makes sense. I'll change it.

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs