From: Dave Chinner <david@fromorbit.com>
Subject: Re: [PATCH 1/6] fs: add hole punching to fallocate
Date: Wed, 12 Jan 2011 23:44:31 +1100
Message-ID: <20110112124431.GP28803@dastard>
References: <1289248327-16308-1-git-send-email-josef@redhat.com>
 <20101109011222.GD2715@dastard>
 <20101109033038.GF3099@thunk.org>
 <20101109044242.GH2715@dastard>
 <20101109214147.GK3099@thunk.org>
 <20101109234049.GQ2715@dastard>
 <AANLkTimwmJ_ZoE9oAuA1WGhCgK585jDznqnc6k0=9Ntb@mail.gmail.com>
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Cc: Ted Ts'o <tytso@mit.edu>, Josef Bacik <josef@redhat.com>,
	linux-kernel@vger.kernel.org, linux-btrfs@vger.kernel.org,
	linux-ext4@vger.kernel.org, linux-fsdevel@vger.kernel.org,
	xfs@oss.sgi.com, joel.becker@oracle.com, cmm@us.ibm.com,
	cluster-devel@redhat.com
To: Lawrence Greenfield <leg@google.com>
Content-Disposition: inline
In-Reply-To: <AANLkTimwmJ_ZoE9oAuA1WGhCgK585jDznqnc6k0=9Ntb@mail.gmail.com>
Sender: linux-ext4-owner@vger.kernel.org

On Tue, Jan 11, 2011 at 04:13:42PM -0500, Lawrence Greenfield wrote:
> On Tue, Nov 9, 2010 at 6:40 PM, Dave Chinner <david@fromorbit.com> wrote:
> > The historical reason for such behaviour existing in XFS was that in
> > 1997 the CPU and IO latency cost of unwritten extent conversion was
> > significant,

.....

> >> (Take for example a trusted cluster filesystem backend that checks the
> >> object checksum before returning any data to the user; and if the
> >> check fails the cluster file system will try to use some other replica
> >> stored on some other server.)
> >
> > IOWs, all they want to do is avoid the unwritten extent conversion
> > overhead. Time has shown that a bad security/performance tradeoff
> > decision was made 13 years ago in XFS, so I see little reason to
> > repeat it for ext4 today....
> 
> I'd make use of FALLOC_FL_EXPOSE_OLD_DATA. It's not the CPU overhead
> of extent conversion. It's that extent conversion causes more metadata
> operations than what you'd have otherwise,

Yes, that's the "IO latency" part of the cost I mentioned above.

> which means systems that
> want to use O_DIRECT and make sure the data doesn't go away either
> have to write O_DIRECT|O_DSYNC or need to call fdatasync().

Seriously, we tell application writers _all the time_ that they
*must* use fsync/fdatasync to guarantee their data is on stable
storage and that they cannot rely on side-effects of filesystem or
storage specific behaviours (like ext3 ordered mode) to do that job
for them.

You're suggesting that by introducing FALLOC_FL_EXPOSE_OLD_DATA,
applications can rely on filesystem/storage specific behaviour to
guarantee data is on stable storage without the use of
fdatasync/fsync. Wht you describe is definitely storage specific,
because volatile write caches still needs the fdatasync to issue a
cache flush.

Do you see the same conflict here that I do?

> cluster file system implementor

Which one?

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com