From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mx2.netapp.com ([216.240.18.37]:12353 "EHLO mx2.netapp.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1754253Ab0GNNSB convert rfc822-to-8bit (ORCPT ); Wed, 14 Jul 2010 09:18:01 -0400 Subject: Re: [PATCH 1/3] NFS: Fix up the fsync code From: Trond Myklebust To: Christoph Hellwig Cc: linux-nfs@vger.kernel.org In-Reply-To: <20100714070008.GA22884@infradead.org> References: <20100712220932.20019.21170.stgit@heimdal.trondhjem.org> <20100713011306.GA364@infradead.org> <1279050636.17993.14.camel@heimdal.trondhjem.org> <20100714070008.GA22884@infradead.org> Content-Type: text/plain; charset="UTF-8" Date: Wed, 14 Jul 2010 09:17:23 -0400 Message-ID: <1279113443.3103.10.camel@heimdal.trondhjem.org> Sender: linux-nfs-owner@vger.kernel.org List-ID: MIME-Version: 1.0 On Wed, 2010-07-14 at 03:00 -0400, Christoph Hellwig wrote: > On Tue, Jul 13, 2010 at 03:50:36PM -0400, Trond Myklebust wrote: > > NFS can't distinguish between a datasync and a full sync: a successful > > COMMIT operation guarantees that both data+metadata updates are on disk. > > For this reason we ignore the 'datasync' parameter in our fsync > > implementation. > > That's not what the datasync parameter means. > > Both fsync and fdatasync will commit data and metadata to disk, the > questions is how much metadata we need to commit. For fdatasync it's > only the metadata requires to locate the file data on disk, an > fsync requires everything (which is the above + timestamps basically). > > I suspect for NFS the difference still doesn't matter, I'd just try > to make it clear. Right. My point was that in NFSv3 and NFSv4, COMMIT always acts like fsync(): it commits data + _all_ metadata (i.e. including size +timestamps) to disk. The exception is NFSv4.1 with pNFS, where you have an extra 'LAYOUTCOMMIT' operation that tells the metadata server when to write the size+timestamps (because the WRITE and COMMIT operations are sent to the data servers). So we might want to distinguish between fsync() and fdatasync() when we merge that code. > > Would it perhaps help if I added a comment to that effect in > > nfs_file_fsync() itself? > > Yes, comments explaining such higher level concepts are always good. OK. I'll add something to that effect. Thanks for the review! Trond