Re: lockdep trace with xfs + mm in it from 5.7.0-rc5

From: Dave Chinner <david@fromorbit.com>
To: "Darrick J. Wong" <darrick.wong@oracle.com>
Cc: Dave Airlie <airlied@gmail.com>,
	LKML <linux-kernel@vger.kernel.org>,
	Linus Torvalds <torvalds@linux-foundation.org>,
	xfs <linux-xfs@vger.kernel.org>
Subject: Re: lockdep trace with xfs + mm in it from 5.7.0-rc5
Date: Sun, 24 May 2020 09:35:25 +1000	[thread overview]
Message-ID: <20200523233525.GO2040@dread.disaster.area> (raw)
In-Reply-To: <20200522204308.GC8230@magnolia>

On Fri, May 22, 2020 at 01:43:08PM -0700, Darrick J. Wong wrote:
> On Fri, May 22, 2020 at 10:30:27AM +1000, Dave Chinner wrote:
> > On Thu, May 21, 2020 at 04:13:12PM -0700, Darrick J. Wong wrote:
> > > [cc linux-xfs]
> > > 
> > > On Fri, May 22, 2020 at 08:21:50AM +1000, Dave Airlie wrote:
> > > > Hi,
> > > > 
> > > > Just updated a rawhide VM to the Fedora 5.7.0-rc5 kernel, did some
> > > > package building,
> > > > 
> > > > got the below trace, not sure if it's known and fixed or unknown.
> > > 
> > > It's a known false-positive.  An inode can't simultaneously be getting
> > > reclaimed due to zero refcount /and/ be the target of a getxattr call.
> > > Unfortunately, lockdep can't tell the difference, and it seems a little
> > > strange to set NOFS on the allocation (which increases the chances of a
> > > runtime error) just to quiet that down.
> > 
> > __GFP_NOLOCKDEP is the intended flag to telling memory allocation
> > that lockdep is stupid.
> > 
> > However, it seems that the patches that were in progress some months
> > ago to convert XFS to kmalloc interfaces and using GFP flags
> > directly stalled - being able to mark locations like this with
> > __GFP_NOLOCKDEP was one of the main reasons for getting rid of all
> > the internal XFS memory allocation wrappers...
> 
> Question is, should I spend time adding a GFP_NOLOCKDEP bandaid to XFS
> or would my time be better spent reviewing your async inode reclaim
> series to make this go away for real?

Heh. I started to write that async reclaim would make this go away,
but then I realised it won't because we still do an XFS_ILOCK_EXCL
call in xfs_inode_reclaim() right at the end to synchronise with
anything that was blocked in the ILOCK during a lockless lookup
waiting for reclaim to drop the lock after setting ip->i_ino = 0.

So that patchset doesn't make the lockdep issues go away. I still
need to work out if we can get rid of that ILOCK cycling in
xfs_reclaim_inode() by changing the lockless lookup code, but that's
a separate problem...

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com