xfs
[Top] [All Lists]

Re: XFS locking issues in 3.18-rc3+

To: Andy Lutomirski <luto@xxxxxxxxxxxxxx>
Subject: Re: XFS locking issues in 3.18-rc3+
From: Dave Chinner <david@xxxxxxxxxxxxx>
Date: Sat, 15 Nov 2014 10:02:42 +1100
Cc: xfs@xxxxxxxxxxx
Delivered-to: xfs@xxxxxxxxxxx
In-reply-to: <CALCETrU+D70HzUFh4qfnKqit7vA0TkZCus1+GNBCC4R=_KXDVg@xxxxxxxxxxxxxx>
References: <CALCETrWN71T1JsSNix52+ez_Sxt4i+vL-fX_2RTNbt8Ex1p=vw@xxxxxxxxxxxxxx> <20141114052523.GB29950@dastard> <CALCETrVcp1BY3cO4Y2sdgGC+eQnyOF3xUCK6qng_Z_Y7dNi2yg@xxxxxxxxxxxxxx> <20141114215007.GC29950@dastard> <CALCETrU+D70HzUFh4qfnKqit7vA0TkZCus1+GNBCC4R=_KXDVg@xxxxxxxxxxxxxx>
User-agent: Mutt/1.5.21 (2010-09-15)
On Fri, Nov 14, 2014 at 02:05:43PM -0800, Andy Lutomirski wrote:
> On Fri, Nov 14, 2014 at 1:50 PM, Dave Chinner <david@xxxxxxxxxxxxx> wrote:
> > On Fri, Nov 14, 2014 at 01:02:59PM -0800, Andy Lutomirski wrote:
> >> On Thu, Nov 13, 2014 at 9:25 PM, Dave Chinner <david@xxxxxxxxxxxxx> wrote:
> >> > On Thu, Nov 13, 2014 at 04:57:11PM -0800, Andy Lutomirski wrote:
> >> >> I'm running a kernel somewhere between 3.18-rc3 and 3.18-rc4.  I got
> >> >> the warnings below.  Is this a known issue?
> >> >
> >> > Yup, false positive. shmem is instantiating an inode under the
> >> > mmap_sem and in doing so taking inode locks under the mmap_sem,
> >> > which triggers other filesystems to issue false warnings about
> >> > mmap_sem <-> inode lock order.
> >>
> >> I still have an enormous flood of warnings even if
> >> CONFIG_PROVE_LOCKING=n.  This one is repeated at high frequency:
> >>
> >> [   70.969118] ------------[ cut here ]------------
> >> [   70.969121] WARNING: CPU: 8 PID: 653 at mm/truncate.c:758
> >> pagecache_isize_extended+0x105/0x110()
> >
> > Fixed in 3.18-rc4.
> 
> Yeah, I just confirmed that, too.  Thanks.  (The lockdep splat is
> still there, but that's much less annoying.)

Fixing the lockdep split is, unfortunately, a lot more work. I've
basically got to redesign the way readdir does readahead to hold
locks in different ways and so we don't need to hold various locks
of the filldir callout where the might_fault() triggers the
warnings.

Cheers,

Dave.
-- 
Dave Chinner
david@xxxxxxxxxxxxx

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