[Top] [All Lists]

Re: [PATCH] XFS: Use KM_NOFS for incore inode extent tree allocation

To: Christoph Hellwig <hch@xxxxxxxxxxxxx>
Subject: Re: [PATCH] XFS: Use KM_NOFS for incore inode extent tree allocation
From: Dave Chinner <david@xxxxxxxxxxxxx>
Date: Mon, 21 Jul 2008 20:59:15 +1000
Cc: xfs@xxxxxxxxxxx
In-reply-to: <20080721075235.GA6692@xxxxxxxxxxxxx>
Mail-followup-to: Christoph Hellwig <hch@xxxxxxxxxxxxx>, xfs@xxxxxxxxxxx
References: <1216615959-23010-1-git-send-email-david@xxxxxxxxxxxxx> <20080721075235.GA6692@xxxxxxxxxxxxx>
Sender: xfs-bounce@xxxxxxxxxxx
User-agent: Mutt/1.5.18 (2008-05-17)
On Mon, Jul 21, 2008 at 03:52:35AM -0400, Christoph Hellwig wrote:
> On Mon, Jul 21, 2008 at 02:52:39PM +1000, Dave Chinner wrote:
> > If we allow incore extent tree allocations to recurse into the
> > filesystem under memory pressure, new delayed allocations through
> > xfs_iomap_write_delay() can deadlock on themselves if memory reclaim
> > tries to write back dirty pages from that inode.
> > 
> > It will deadlock in xfs_iomap_write_allocate() trying to take the
> > ilock we already hold. This can also show up as complex ABBA
> > deadlocks when multiple threeads are triggering memory reclaim when
> > trying to allocate extents.
> > 
> > The main cause of this is the fact that delayed allocation is
> > not done in a transaction, so KM_NOFS is not automatically
> > added to the allocations to prevent this recursion.
> > 
> > Mark all allocations done for the incore inode extent tree as
> > KM_NOFS to ensure they never recurse back into the filesystem.
> Looks good.  Note that KM_NOFS alone already means a allocation
> that can't fail, so no need to or it to KM_SLEEP.

Right. I'll update the patch and resend it.

> And long term we should try to look into allowing these to fail,
> allocations that aren't allowed to fail but can't recurse back into
> the fs still have a chance to deadlock.

We need dirty transaction rollback capabilities before we can do that.


Dave Chinner

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