Re: [PATCH 0/8] Set bi_rw when alloc bio before call bio_add_page.

Re: [PATCH 0/8] Set bi_rw when alloc bio before call bio_add_page.

[Konrad Rzeszutek Wilk]

On Mon, Jul 30, 2012 at 03:14:28PM +0800, majianpeng wrote:
> When exec bio_alloc, the bi_rw is zero.But after calling bio_add_page,

What about submit_bio? That sets the bi_rw as well?
> it will use bi_rw.
> Fox example, in functiion __bio_add_page,it will call merge_bvec_fn().

function.
> The merge_bvec_fn of raid456 will use the bi_rw to judge the merge.
> >> if ((bvm->bi_rw & 1) == WRITE)
> >> return biovec->bv_len; /* always allow writes to be mergeable */

So what does that mean? Without this patch what happens?

> 
> 
> Jianpeng Ma (8):

Can you collapse all of this in just one patch?

>   Evalue bio->bi_rw after calling bio_alloc() and before calling

It is 'evaluate'
>     bio_add_page().
>   Evalue bio->bi_rw after calling bio_alloc() and before calling    
>     bio_add_page().
>   Evalue bio->bi_rw after calling bio_alloc() and before calling    
>     bio_add_page().
>   Evalue bio->bi_rw after calling bio_alloc() and before calling    
>     bio_add_page().
>   Evalue bio->bi_rw after calling bio_alloc() and before calling    
>     bio_add_page().
>   Evalue bio->bi_rw after calling bio_alloc() and before calling    
>     bio_add_page().
>   Evalue bio->bi_rw after calling bio_alloc() and before calling    
>     bio_add_page().
>   Evalue bio->bi_rw after calling bio_alloc() and before calling    
>     bio_add_page().
> 
>  block/blk-lib.c                     |    1 +
>  drivers/block/xen-blkback/blkback.c |    1 +

I am not really sure if that is correct. Does 'submit_bio' not do the
job properly?

>  fs/btrfs/check-integrity.c          |    1 +
>  fs/direct-io.c                      |    1 +
>  fs/ext4/page-io.c                   |    1 +
>  fs/jfs/jfs_metapage.c               |    1 +
>  fs/ocfs2/cluster/heartbeat.c        |    8 +++++---
>  fs/xfs/xfs_aops.c                   |    2 ++
>  fs/xfs/xfs_buf.c                    |    1 +
>  9 files changed, 14 insertions(+), 3 deletions(-)
> 
> -- 
> 1.7.9.5

Re: [PATCH 0/8] Set bi_rw when alloc bio before call bio_add_page.

[Dave Chinner]

On Mon, Jul 30, 2012 at 03:14:28PM +0800, majianpeng wrote:
> When exec bio_alloc, the bi_rw is zero.But after calling bio_add_page,
> it will use bi_rw.
> Fox example, in functiion __bio_add_page,it will call merge_bvec_fn().
> The merge_bvec_fn of raid456 will use the bi_rw to judge the merge.
> >> if ((bvm->bi_rw & 1) == WRITE)
> >> return biovec->bv_len; /* always allow writes to be mergeable */

So if bio_add_page() requires bi_rw to be set, then shouldn't it be
set up for every caller? I noticed there are about 50 call sites for
bio_add_page(), and you've only touched about 10 of them. Indeed, I
notice that the RAID0/1 code uses bio_add_page, and as that can be
stacked on top of RAID456, it also needs to set bi_rw correctly.
As a result, your patch set is nowhere near complete, not does it
document that bio_add_page requires that bi_rw be set before calling
(which is the new API requirement, AFAICT).

So, my question is whether the RAID456 code is doing something
valid.  That write optimisation is clearly not enabled for a
significant amount of code and filesystems, so the first thing to do
is quantify the benefit of the optimisation. I can't evalute the
merit of this change without data telling me it is worthwhile, and
it's a lot of code to churn for no benefit....

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

Re: Re: [PATCH 0/8] Set bi_rw when alloc bio before call bio_add_page.

[Dave Chinner]

On Tue, Jul 31, 2012 at 08:55:59AM +0800, majianpeng wrote:
> On 2012-07-31 05:42 Dave Chinner <david@fromorbit.com> Wrote:
> >On Mon, Jul 30, 2012 at 03:14:28PM +0800, majianpeng wrote:
> >> When exec bio_alloc, the bi_rw is zero.But after calling bio_add_page,
> >> it will use bi_rw.
> >> Fox example, in functiion __bio_add_page,it will call merge_bvec_fn().
> >> The merge_bvec_fn of raid456 will use the bi_rw to judge the merge.
> >> >> if ((bvm->bi_rw & 1) == WRITE)
> >> >> return biovec->bv_len; /* always allow writes to be mergeable */
> >
> >So if bio_add_page() requires bi_rw to be set, then shouldn't it be
> >set up for every caller? I noticed there are about 50 call sites for
> >bio_add_page(), and you've only touched about 10 of them. Indeed, I
> >notice that the RAID0/1 code uses bio_add_page, and as that can be
> >stacked on top of RAID456, it also needs to set bi_rw correctly.
> >As a result, your patch set is nowhere near complete, not does it
> >document that bio_add_page requires that bi_rw be set before calling
> >(which is the new API requirement, AFAICT).
> There are many place call bio_add_page and I send some of those. Because my abilty, so I only send 
> some patchs which i understand clearly.

Sure, but my point is that there is no point changing only a few and
ignoring the great majority of callers. Either fix them all, fix it
some other way (e.g. API change), or remove the code from the RAID5
function that requires it.

> In __bio_add_page:
> >>if (q->merge_bvec_fn) {
> >>				struct bvec_merge_data bvm = {
> >>					/* prev_bvec is already charged in
> >>					   bi_size, discharge it in order to
> >>					   simulate merging updated prev_bvec
> >>					   as new bvec. */
> >>					.bi_bdev = bio->bi_bdev,
> >>					.bi_sector = bio->bi_sector,
> >>					.bi_size = bio->bi_size - prev_bv_len,
> >>					.bi_rw = bio->bi_rw,
> >>				};
> it used bio->bi_rw.
> Before raid5_mergeable_bvec appearing, in kernel 'merge_bvec_fn' did not use bio->bi_rw.

Right, but as things stand right now, the RAID5 code is a no-op
because nobody is setting bio->bi_rw correctly. it is effectively
dead code.

> But i think we shold not suppose bi_rw not meanless.

To decide whether we should take it to have meaning, data is
required to show that the RAID5 optimisation it enables is
worthwhile.  If the optimisation is not worthwhile, then the correct
thing to do is remove the optimisation in the RAID5 code and remove
the bi_rw field from the struct bvec_merge_data.

> >So, my question is whether the RAID456 code is doing something
> >valid.  That write optimisation is clearly not enabled for a
> >significant amount of code and filesystems, so the first thing to do
> >is quantify the benefit of the optimisation. I can't evalute the
> >merit of this change without data telling me it is worthwhile, and
> >it's a lot of code to churn for no benefit....
> >
> Sorry, we do not think the 'merge_bvec_fn' did not use bi_rw.

It's entirely possible that when bi_rw was added to struct
bvec_merge_data, the person who added it was mistaken that bi_rw was
set at this point in time when in fact it never has been. Hence it's
presence and reliance on it would be a bug.

That's what I'm asking - is this actually beneificial, or should it
simply be removed from struct bvec_merge_data? Data is needed to
answer that question....

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

Re: Re: [PATCH 0/8] Set bi_rw when alloc bio before call bio_add_page.

[Muthu Kumar]

[ Resending in plain text... sorry for the duplicate ]

Hi,

On Mon, Jul 30, 2012 at 6:14 PM, Dave Chinner <david@fromorbit.com> wrote:
>
> On Tue, Jul 31, 2012 at 08:55:59AM +0800, majianpeng wrote:
> > On 2012-07-31 05:42 Dave Chinner <david@fromorbit.com> Wrote:
> > >On Mon, Jul 30, 2012 at 03:14:28PM +0800, majianpeng wrote:
> > >> When exec bio_alloc, the bi_rw is zero.But after calling
> > >> bio_add_page,
> > >> it will use bi_rw.
> > >> Fox example, in functiion __bio_add_page,it will call
> > >> merge_bvec_fn().
> > >> The merge_bvec_fn of raid456 will use the bi_rw to judge the merge.
> > >> >> if ((bvm->bi_rw & 1) == WRITE)
> > >> >> return biovec->bv_len; /* always allow writes to be mergeable */
> > >
> > >So if bio_add_page() requires bi_rw to be set, then shouldn't it be
> > >set up for every caller? I noticed there are about 50 call sites for
> > >bio_add_page(), and you've only touched about 10 of them. Indeed, I
> > >notice that the RAID0/1 code uses bio_add_page, and as that can be
> > >stacked on top of RAID456, it also needs to set bi_rw correctly.
> > >As a result, your patch set is nowhere near complete, not does it
> > >document that bio_add_page requires that bi_rw be set before calling
> > >(which is the new API requirement, AFAICT).
> > There are many place call bio_add_page and I send some of those. Because
> > my abilty, so I only send
> > some patchs which i understand clearly.
>
> Sure, but my point is that there is no point changing only a few and
> ignoring the great majority of callers. Either fix them all, fix it
> some other way (e.g. API change), or remove the code from the RAID5
> function that requires it.

A while back, we tried to address this by changing the alloc functions to
take rw argument and set it (as per Jens suggestion). I guess the patch did
not make it in. Please check:

https://lkml.org/lkml/2011/7/11/275

and the follow ups. If needed, I can dust up that patch and resend it.

>

<snip>

>
> It's entirely possible that when bi_rw was added to struct
> bvec_merge_data, the person who added it was mistaken that bi_rw was
> set at this point in time when in fact it never has been. Hence it's
> presence and reliance on it would be a bug.
>
> That's what I'm asking - is this actually beneificial, or should it
> simply be removed from struct bvec_merge_data? Data is needed to
> answer that question....

There are cases where we found it really beneficial to know the rw
field to decide if the can be really merged or not.


Regards,
Muthu


>
> Cheers,
>
> Dave.
> --
> Dave Chinner
> david@fromorbit.com
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/