From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: with ECARTIS (v1.0.0; list xfs); Sun, 24 Jun 2007 22:45:47 -0700 (PDT) Received: from larry.melbourne.sgi.com (larry.melbourne.sgi.com [134.14.52.130]) by oss.sgi.com (8.12.10/8.12.10/SuSE Linux 0.7) with SMTP id l5P5jZds017242 for ; Sun, 24 Jun 2007 22:45:44 -0700 Date: Mon, 25 Jun 2007 09:22:46 +1000 From: David Chinner Subject: Re: [REGRESSION 2.6-git] possible circular locking dependency detected with XFS Message-ID: <20070624232246.GF86004887@sgi.com> References: <6101e8c40706221340k65f15957k39a04193cb6e7c01@mail.gmail.com> <6bffcb0e0706221553s3a74ef58hcadc69bfa252283@mail.gmail.com> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <6bffcb0e0706221553s3a74ef58hcadc69bfa252283@mail.gmail.com> Sender: xfs-bounce@oss.sgi.com Errors-to: xfs-bounce@oss.sgi.com List-Id: xfs To: Michal Piotrowski Cc: Oliver Pinter , linux-kernel , xfs@oss.sgi.com On Sat, Jun 23, 2007 at 12:53:11AM +0200, Michal Piotrowski wrote: > Hi Oliver, > > On 22/06/07, Oliver Pinter wrote: > >Hi all! > > > >I found this info: > > > >======================================================= [ INFO: possible > >circular locking dependency detected ] 2.6.22-rc5-wifi1 #2 > >------------------------------------------------------- mount/2209 is > >trying to acquire lock: (&(&ip->i_lock)->mr_lock/1){--..}, at: [] > >xfs_ilock+0x66/0x90 > > > >but task is already holding lock: (&(&ip->i_lock)->mr_lock){----}, at: > >[] xfs_ilock+0x66/0x90 > > > > AFAIR it is not a regression. It is a known bug (harmless). FWIW, it's not even a bug. The bug (if any) is due to the fact we can't properly express the XFS locking rules with lockdep. We recently added a bunch of notations that fixed the common false positives we were seeing, but as a result, it appears we now have a whole new set of false positive reports coming in that are even harder to fix. As Christoph Hellwig has previously noted, the correct way to fix this in XFS is to completely change the locking within XFS directory operations to do strict parent/child locking like the VFS does. Unfortunately, that's not as simple as it sounds, because inode flushing and log tail pushing rely on inodes being locked in ascending inode order to prevent deadlocks within XFS. That means when we lock multiple inodes in link, rename, etc, we have to lock them in ascending order. The exception to this is create, mkdir, mknod because the newly created inode will not be locked by definition so it is always safe to lock it. Hence if the new inode's number is less than the parent inode's number we can get lockdep warning about circular locking dependencies which don't actually exist. That is where this warning is coming from.... Cheers, Dave. -- Dave Chinner Principal Engineer SGI Australian Software Group