From: Jan Kara Subject: Re: [PATCH RFC] ext3 data=guarded v5 Date: Thu, 30 Apr 2009 13:38:23 +0200 Message-ID: <20090430113823.GC29220@duck.suse.cz> References: <1240941840.15136.44.camel@think.oraclecorp.com> <20090429085632.GA18273@duck.suse.cz> <1241014093.20099.28.camel@think.oraclecorp.com> <20090429191533.GB22936@duck.suse.cz> <1241034089.20099.60.camel@think.oraclecorp.com> <20090429200412.GA27924@duck.suse.cz> <1241037421.20099.70.camel@think.oraclecorp.com> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: Linus Torvalds , Theodore Ts'o , Linux Kernel Developers List , Ext4 Developers List , Mike Galbraith To: Chris Mason Return-path: Received: from cantor.suse.de ([195.135.220.2]:34016 "EHLO mx1.suse.de" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1760762AbZD3Li1 (ORCPT ); Thu, 30 Apr 2009 07:38:27 -0400 Content-Disposition: inline In-Reply-To: <1241037421.20099.70.camel@think.oraclecorp.com> Sender: linux-ext4-owner@vger.kernel.org List-ID: On Wed 29-04-09 16:37:01, Chris Mason wrote: > On Wed, 2009-04-29 at 22:04 +0200, Jan Kara wrote: > > > > What we don't want to do is have a call to write() over existing blocks > > > in the file add new things to the data=ordered list. I don't see how we > > > can avoid that without datanew. > > Yes, what I suggest would do exactly that: > > In ordered_writepage() in the beginning we do: > > page_bufs = page_buffers(page); > > if (!walk_page_buffers(NULL, page_bufs, 0, PAGE_CACHE_SIZE, > > NULL, buffer_unmapped)) { > > return block_write_full_page(page, NULL, wbc); > > } > > So we only get to starting a transaction and file some buffers if some buffer > > in the page is unmapped. Write() maps / allocates all buffers in write_begin() > > so they are never added to ordered lists in writepage(). > > Right, writepage doesn't really need datanew. > > > We rely on write_end > > to do it. So the only case where not all buffers in the page are mapped is > > when we have to allocate in writepage() (mmaped write) or the two cases I > > describe above. > > But I still think write_end does need datanew. That's where 99% of the > ordered buffers are going to come from when we overwrite the contents of > an existing file. Ah, true, buffer_new() can be cleared in __block_prepare_write() in some cases. Frankly, I don't see a reason why that happens but that's another story. Honza -- Jan Kara SUSE Labs, CR