From mboxrd@z Thu Jan 1 00:00:00 1970 From: Tao Ma Subject: Re: [PATCH 1/2] ext4: let mpage_submit_io works well when blocksize < pagesize Date: Sun, 27 Nov 2011 16:59:14 +0800 Message-ID: <4ED1FBE2.8020001@tao.ma> References: <1322039726-3587-1-git-send-email-xiaoqiangnk@gmail.com> <4ED100E8.109@tao.ma> <4ED10448.1040204@tao.ma> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1 Content-Transfer-Encoding: 7bit Cc: "linux-ext4@vger.kernel.org" , "achender@linux.vnet.ibm.com" , "tytso@mit.edu" To: Yongqiang Yang Return-path: Received: from oproxy7-pub.bluehost.com ([67.222.55.9]:35409 "HELO oproxy7-pub.bluehost.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with SMTP id S1750987Ab1K0I7k (ORCPT ); Sun, 27 Nov 2011 03:59:40 -0500 In-Reply-To: Sender: linux-ext4-owner@vger.kernel.org List-ID: On 11/27/2011 09:29 AM, Yongqiang Yang wrote: > > > On Saturday, November 26, 2011, Tao Ma > wrote: >> On 11/26/2011 11:13 PM, Yongqiang Yang wrote: >>> Hi, >>> >>> It can be tested by xfstests 61 or 91, I did not remember which one. >> ok, I will try it later. But I still wonder how it happens. See my >> comments below. >>> >>> It can be reproduced by reading on a fallocted block and write the block >>> after the fallocted block. Then the written block can not be written >>> out by da_writepages. >> Why? in read no bh will be created, and in write only the written bh >> will be mapped and set unwritten. How could that happen? Sorry, but this >> explanation doesn't convince me. > Ok! The bug was found with punch hole enabled, and punch hole creates > empty buffers sometimes. > Actually, punching hole can read the whole page like read operation > does, but it do not. May be this is the source of the problem. > The problemastic function is discard_partial_buffers. I post a patch > fixing another bug in it in this series. If we read whole page in this > function, the problem can also be solved. Got it and thanks for the explanation. So we introduced a bug and then fix it in another way. ;) Thanks Tao > > Yongqiang. >>> >>> Yongqiang. >>> >>> On Saturday, November 26, 2011, Tao Ma > >> wrote: >>>> Hi Yongqiang, >>>> On 11/23/2011 05:15 PM, Yongqiang Yang wrote: >>>>> If there is a unwritten but clean buffer in a page and there is a >>> dirty buffer >>>>> after the buffer, then mpage_submit_io does not write the dirty >>> buffer out. >>>>> As a result, da_writepages loops forever. >>>> Did you ever meet with this bug or just find it to be a possible bug by >>>> skimming the code? Actually, I can't find a proper way to get a buffer >>>> which can satisfy the check. >>>> >>>> Thanks >>>> Tao >>>>> >>>>> This patch fixes the problem by checking dirty flag. >>>>> >>>>> Signed-off-by: Yongqiang Yang >>> >> >>>>> --- >>>>> fs/ext4/inode.c | 7 +++++-- >>>>> 1 files changed, 5 insertions(+), 2 deletions(-) >>>>> >>>>> diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c >>>>> index 755f6c7..20a1d17 100644 >>>>> --- a/fs/ext4/inode.c >>>>> +++ b/fs/ext4/inode.c >>>>> @@ -1339,8 +1339,11 @@ static int mpage_da_submit_io(struct >>> mpage_da_data *mpd, >>>>> clear_buffer_unwritten(bh); >>>>> } >>>>> >>>>> - /* skip page if block allocation > undone */ >>>>> - if (buffer_delay(bh) || >>> buffer_unwritten(bh)) >>>>> + /* >>>>> + * skip page if block allocation > undone and >>>>> + * block is dirty >>>>> + */ >>>>> + if (ext4_bh_delay_or_unwritten(NULL, bh)) >>>>> skip_page = 1; >>>>> bh = bh->b_this_page; >>>>> block_start += bh->b_size; >>>> >>>> >>>> >>> >>> -- >>> Best Wishes >>> Yongqiang Yang >>> >> >> > > -- > Best Wishes > Yongqiang Yang >