From mboxrd@z Thu Jan  1 00:00:00 1970
From: Christoph Hellwig <hch@infradead.org>
Subject: Re: [patch 8/8] fs: add i_op->sync_inode
Date: Fri, 7 Jan 2011 02:24:30 -0500
Message-ID: <20110107072430.GA32308@infradead.org>
References: <20101218014634.943276411@kernel.dk>
 <20101218015117.759480620@kernel.dk>
 <20101229151246.GA22033@infradead.org>
 <20110104062725.GD3402@amd>
 <20110104065736.GA8013@infradead.org>
 <20110104080323.GC4090@amd>
 <20110104092501.GB2760@infradead.org>
 <20110104094922.GA4812@amd>
 <20110106204510.GA2872@infradead.org>
 <20110107044734.GA4552@amd>
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Cc: Christoph Hellwig <hch@infradead.org>,
	linux-fsdevel@vger.kernel.org, linux-kernel@vger.kernel.org,
	Andrew Morton <akpm@linux-foundation.org>
To: Nick Piggin <npiggin@kernel.dk>
Return-path: <linux-fsdevel-owner@vger.kernel.org>
Received: from bombadil.infradead.org ([18.85.46.34]:49400 "EHLO
	bombadil.infradead.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org
	with ESMTP id S1752541Ab1AGHYb (ORCPT
	<rfc822;linux-fsdevel@vger.kernel.org>);
	Fri, 7 Jan 2011 02:24:31 -0500
Content-Disposition: inline
In-Reply-To: <20110107044734.GA4552@amd>
Sender: linux-fsdevel-owner@vger.kernel.org
List-ID: <linux-fsdevel.vger.kernel.org>

On Fri, Jan 07, 2011 at 03:47:34PM +1100, Nick Piggin wrote:
> No, you misunderstand 1. I am saying they should be treated as
> WB_SYNC_NONE.
> 
> In fact 2 would cause much more IO, because dirty writeout would
> never clean them so it will just keep writing them out. I don't
> know how 2 could be feasible.

WB_SYNC_NONE means ->write_inode behaves non-blocking.  That is
we do not block on memory allocations, and we do not take locks
blocking.  Most journaling filesystems currently take the easy
way out an make it a no-op due to that, but take a look at XFS
how complicated it is to avoid the blocking if you want a non-noop
implementation.

> So, back to my original question: what is the performance problem
> with treating write_inode as WB_SYNC_NONE, and then having .fsync
> and .sync_fs do the integrity?

See above - we'll block in the flusher thread and cause it to stall,
which is really nasty as it does all data I/O writeback.  The salling
may also block sync() although I don't think it's as important there.