Return-Path: Received: from mail-gx0-f174.google.com ([209.85.161.174]:37405 "EHLO mail-gx0-f174.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1754511Ab1FAQBP convert rfc822-to-8bit (ORCPT ); Wed, 1 Jun 2011 12:01:15 -0400 Received: by gxk21 with SMTP id 21so2263523gxk.19 for ; Wed, 01 Jun 2011 09:01:14 -0700 (PDT) In-Reply-To: <4DE65202.2010502@panasas.com> References: <1306898310-9229-1-git-send-email-dros@netapp.com> <4DE5D287.603@panasas.com> <09816808-BA7B-4EAF-A18A-866B5A98BF25@netapp.com> <4DE65202.2010502@panasas.com> Date: Wed, 1 Jun 2011 12:01:13 -0400 Message-ID: Subject: Re: [PATCH] NFS: filelayout should use nfs_generic_pg_test From: Fred Isaman To: Benny Halevy Cc: Weston Andros Adamson , Boaz Harrosh , trond@netapp.com, linux-nfs@vger.kernel.org Content-Type: text/plain; charset=ISO-8859-1 Sender: linux-nfs-owner@vger.kernel.org List-ID: MIME-Version: 1.0 On Wed, Jun 1, 2011 at 10:51 AM, Benny Halevy wrote: > On 2011-06-01 17:44, Weston Andros Adamson wrote: >> >> On Jun 1, 2011, at 1:47 AM, Boaz Harrosh wrote: >> >>> On 06/01/2011 06:18 AM, Weston Andros Adamson wrote: >>>> Use nfs_generic_pg_test instead of pnfs_generic_pg_test. >>>> >>>> This fixes the BUG at fs/nfs/write.c:941 introduced by >>>> 89a58e32d9105c01022a757fb32ddc3b51bf0025. >>>> >>>> I was able to trigger this BUG reliably using pynfs in pnfs mode, >>>> by using dd(1) to write many small blocks. >>>> >>>> Signed-off-by: Weston Andros Adamson >>>> --- >>>> Fix proposed by Trond. >>>> >>>> Benny- Does this make sense? >>>> >>>> fs/nfs/nfs4filelayout.c ?| ? ?2 +- >>>> fs/nfs/pagelist.c ? ? ? ?| ? ?5 ++++- >>>> include/linux/nfs_page.h | ? ?3 ++- >>>> 3 files changed, 7 insertions(+), 3 deletions(-) >>>> >>>> diff --git a/fs/nfs/nfs4filelayout.c b/fs/nfs/nfs4filelayout.c >>>> index 4269088..1c3bb72 100644 >>>> --- a/fs/nfs/nfs4filelayout.c >>>> +++ b/fs/nfs/nfs4filelayout.c >>>> @@ -661,7 +661,7 @@ filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev, >>>> ? ? u64 p_stripe, r_stripe; >>>> ? ? u32 stripe_unit; >>>> >>>> - ? if (!pnfs_generic_pg_test(pgio, prev, req)) >>>> + ? if (!nfs_generic_pg_test(pgio, prev, req)) >>>> ? ? ? ? ? ? return 0; >>>> >>> >>> pnfs_generic_pg_test is the one that gets the layout. >>> >>> What you've done is revert to MDS IO >>> >>> Boaz >> >> Ah, you're right - I didn't even notice that! ?I usually confirm client -> DS communication with tcpdump. ?I was working for too long yesterday :) >> >> Patch: recalled. ?Discussion about a real fix: started. >> >> -dros > > I think the following should work: > > Benny > > git diff --stat -p -M > ?fs/nfs/nfs4filelayout.c | ? 10 ++++++++++ > ?1 files changed, 10 insertions(+), 0 deletions(-) > > diff --git a/fs/nfs/nfs4filelayout.c b/fs/nfs/nfs4filelayout.c > index 4269088..9f1d445 100644 > --- a/fs/nfs/nfs4filelayout.c > +++ b/fs/nfs/nfs4filelayout.c > @@ -661,6 +661,16 @@ filelayout_pg_test(struct nfs_pageio_descriptor > *pgio, struct nfs_page *prev, > ? ? ? ?u64 p_stripe, r_stripe; > ? ? ? ?u32 stripe_unit; > > + ? ? ? /* > + ? ? ? ?* FIXME: ideally we should be able to coalesce all requests > + ? ? ? ?* that are not block boundary aligned, but currently this > + ? ? ? ?* is problematic for the case of bsize < PAGE_CACHE_SIZE, > + ? ? ? ?* since nfs_flush_multi and nfs_pagein_multi assume you > + ? ? ? ?* can have only one struct nfs_page. > + ? ? ? ?*/ > + ? ? ? if (desc->pg_bsize < PAGE_SIZE) > + ? ? ? ? ? ? ? return 0; > + > ? ? ? ?if (!pnfs_generic_pg_test(pgio, prev, req)) > ? ? ? ? ? ? ? ?return 0; > Note this moves a test that was once part of the plain nfs code into the file layout driver. Why don't other drivers need this test? Fred