xfs
[Top] [All Lists]

Re: Still seeing hangs in xlog_grant_log_space

To: Dave Chinner <david@xxxxxxxxxxxxx>
Subject: Re: Still seeing hangs in xlog_grant_log_space
From: Ben Myers <bpm@xxxxxxx>
Date: Fri, 8 Jun 2012 12:09:43 -0500
Cc: Brian Foster <bfoster@xxxxxxxxxx>, xfs@xxxxxxxxxxx
In-reply-to: <20120608002826.GV4347@dastard>
References: <CAH4wwdGWHSZoveLJMxu5pjr22NEEeW7oG8TS+snoM8RY=ZeRmg@xxxxxxxxxxxxxx> <CADLDEKsGtsw-rrSOE7gY4T81u+p41b34ixv0B7Dh07afJ73n2w@xxxxxxxxxxxxxx> <CAH4wwdFu7DEkHFZ5Bf7_PtLPsG0hUyUDoov03q=82R6t+QkERg@xxxxxxxxxxxxxx> <20120605235447.GF22848@dastard> <4FCF5DB9.2000808@xxxxxxxxxx> <20120607013531.GP22848@dastard> <4FD0B7B6.3000704@xxxxxxxxxx> <20120608002826.GV4347@dastard>
User-agent: Mutt/1.5.20 (2009-06-14)
On Fri, Jun 08, 2012 at 10:28:26AM +1000, Dave Chinner wrote:
...
> And because the inodes pin the tail of the AIL, there's not enough
> space in the log for the xfs_sync_worker to trigger a log force via
> the dummy transaction, and hence we deadlock.
>
> FWIW, there's an argument that can be made here for an unconditional
> log force in xfs_sync_worker() to provide a "get out gaol free" card
> here. 

No kidding!
http://oss.sgi.com/archives/xfs/2012-05/msg00312.html

> The thing is, I would prefer that the filesystems hang so that
> we find out about these issues and have to understand them and fix
> them. IMO, there is nothing harder to detect and debug than short
> duration, temporary stalls of the filesystem...

I agree.. such a patch is not for general consumption.  We want to fix the
actual problem, not work around it with a prod on a timer.  ;)

Regards,
        Ben

<Prev in Thread] Current Thread [Next in Thread>