Re: Re: [RFC PATCH] raid5: Add R5_ReadNoMerge flag which prevent bio from merging at block layer

[majianpeng <majianpeng@gmail.com>]

On 2012-07-04 10:30 NeilBrown <neilb@suse.de> Wrote:
>On Tue, 3 Jul 2012 16:52:49 +0800 majianpeng <majianpeng@gmail.com> wrote:
>
>> On 2012-07-03 14:38 NeilBrown <neilb@suse.de> Wrote:
>> >On Tue, 3 Jul 2012 14:23:41 +0800 majianpeng <majianpeng@gmail.com> wrote:
>> >
>> >> Because bios will merge at block-layer,so bios-error may caused by other
>> >> bio which be merged into to the same request.
>> >> Using this flag,it will find exactly error-sector and not do redundant
>> >> operation like re-write and re-read.
>> >> 
>> >> Signed-off-by: majianpeng <majianpeng@gmail.com>
>> >
>> >Hi,
>> > I think this patch needs a more detailed explanation.
>> >
>> >What exactly is the situation that causes a problem, and what exactly is the
>> >problem that it causes?
>> >
>> >Pretend that I don't know anything about what happens below the md level..
>> >
>> >Thanks,
>> >NeilBrown
>> >
>> >
....
>> >>  
>> >
>> >
>> How about the below explanation:
>> 
>> Because bio will be merged at block-layer,so bios-error may caused by other
>> bio which be merged into to the same request.
>> For example: if chunk_aligned_read failed, it will add bio to some stipe.
>> But because the bio-merge function,those bios at most be merged.It will like
>> the chunk_aligned_read and returen error.Then it will re-write and re-read.
>> Suppose RAID5 created by n disk and chunk-size is 512K.
>> If read 512k chunk_aligned and met error(sector 0 is media error), 
>> then add 512/4=128 stipe.If those bios merged and must be error.The rewrite operation
>> will read (n-1) * 128 and computer 128 stripe.But using this flag,we only exec one rewrite.
>> This occur in resync/repair situation and chunk-aligned-read.
>> 
>> May be using this flag,we can find exact bad-sector.Not recorded the whole request bios 
>> as bad-sector
>
>Thanks.  That make some sense.
>
>I wonder if we can just use REQ_FUA to stop requests being merged, rather
>than create a new flag.  In general we should probably be using REQ_FUA in
>all the cases where we are checking for, and trying to fix, read errors.
>
>Thanks,
>NeilBrown
>
>
Using REQ_FUA/REQ_FLUSH can have same effect.
But i think REQ_FUA/REQ_FLUSH operation do more work for example, pre_flush/post_pflush and the harddisk will do more
work.But our purpose only exec bio slonely.