From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from relay.sgi.com (relay2.corp.sgi.com [137.38.102.29]) by oss.sgi.com (Postfix) with ESMTP id F2D3F7CA1 for ; Wed, 3 Aug 2016 17:34:01 -0500 (CDT) Received: from cuda.sgi.com (cuda1.sgi.com [192.48.157.11]) by relay2.corp.sgi.com (Postfix) with ESMTP id 9869F304048 for ; Wed, 3 Aug 2016 15:34:01 -0700 (PDT) Received: from ipmail06.adl6.internode.on.net (ipmail06.adl6.internode.on.net [150.101.137.145]) by cuda.sgi.com with ESMTP id NeM7PISiLrdvEYf7 for ; Wed, 03 Aug 2016 15:33:58 -0700 (PDT) Date: Thu, 4 Aug 2016 08:32:42 +1000 From: Dave Chinner Subject: Re: [PATCH] xfs: don't invalidate whole file on DAX read/write Message-ID: <20160803223242.GV16044@dastard> References: <1470181226-20935-1-git-send-email-david@fromorbit.com> <20160803153437.GC4576@quack2.suse.cz> MIME-Version: 1.0 Content-Disposition: inline In-Reply-To: <20160803153437.GC4576@quack2.suse.cz> List-Id: XFS Filesystem from SGI List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit Errors-To: xfs-bounces@oss.sgi.com Sender: xfs-bounces@oss.sgi.com To: Jan Kara Cc: xfs@oss.sgi.com On Wed, Aug 03, 2016 at 05:34:37PM +0200, Jan Kara wrote: > On Wed 03-08-16 09:40:26, Dave Chinner wrote: > > From: Dave Chinner > > > > When we do DAX IO, we try to invalidate the entire page cache held > > on the file. This is incorrect as it will trash the entire mapping > > tree that now tracks dirty state in exceptional entries in the radix > > tree slots. > > > > What we are trying to do is remove cached pages (e.g from reads > > into holes) that sit in the radix tree over the range we are about > > to write to. Hence we should just limit the invalidation to the > > range we are about to overwrite. > > The patch looks good. Just one comment below. > > > > > Reported-by: Jan Kara > > Signed-off-by: Dave Chinner > > --- > > fs/xfs/xfs_file.c | 13 ++++++++++++- > > 1 file changed, 12 insertions(+), 1 deletion(-) > > > > diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c > > index ed95e5b..e612a02 100644 > > --- a/fs/xfs/xfs_file.c > > +++ b/fs/xfs/xfs_file.c > > @@ -741,9 +741,20 @@ xfs_file_dax_write( > > * page is inserted into the pagecache when we have to serve a write > > * fault on a hole. It should never be dirtied and can simply be > > * dropped from the pagecache once we get real data for the page. > > + * > > + * XXX: This is racy against mmap, and there's nothing we can do about > > + * it. dax_do_io() should really do this invalidation internally as > > + * it will know if we've allocated over a holei for this specific IO and > > + * if so it needs to update the mapping tree and invalidate existing > > + * PTEs over the newly allocated range. Remove this invalidation when > > + * dax_do_io() is fixed up. > > And would it be OK for XFS if dax_do_io() actually invalidated page cache / > PTEs under just XFS_IOLOCK_SHARED? Because currently you seem to be careful > to call invalidate_inode_pages2() only when holding the lock exclusively > and then demote it to a shared one when calling dax_do_io(). That really only exists to prevent multiple IOs trying to do invalidation at once. In the direct IO code, we don't want multiple page cache flushers running at once - one is enough - so we serialise on that state knowing that once the invalidation is done the remaining EXCL lock waiters will pass straight through. For DAX, I don't think that's a problem - the invalidation is ranged, and it's unlikely there will be overlaps, and mapping/pte invalidation is done under fine grained locks so we don't have to worry about races there, either. So it seems fine to me to do this under a SHARED lock. It will still serialise against truncate and other extent manipulation operations, and that's mainly what we care about here. Cheers, Dave. -- Dave Chinner david@fromorbit.com _______________________________________________ xfs mailing list xfs@oss.sgi.com http://oss.sgi.com/mailman/listinfo/xfs