[PATCH] fix bio_add_page for non trivial merge_bvec_fn case

[PATCH] fix bio_add_page for non trivial merge_bvec_fn case

[Dmitri Monakhov]

We have to properly decrease all related bio's counters, especially bi_size
in order to merge_bvec_fn return right result. Usually this result in
false merge rejects for two absolutely valid bio_vecs. This may cause 
significant performance penalty for example Itanium: page_size == 16k,
fs_block_size == 1k and block device is raid with small chunk_size.

Signed-off-by: Dmitri Monakhov <dmonakhov@openvz.org>
---
 fs/bio.c |   16 ++++++++++++----
 1 files changed, 12 insertions(+), 4 deletions(-)

diff --git a/fs/bio.c b/fs/bio.c
index 7856257..d713074 100644
--- a/fs/bio.c
+++ b/fs/bio.c
@@ -332,14 +332,21 @@ static int __bio_add_page(struct request_queue *q, struct bio *bio, struct page
 
 		if (page == prev->bv_page &&
 		    offset == prev->bv_offset + prev->bv_len) {
+			/* Temprory detacth last bio_vec. */
+			bio->bi_size -= prev->bv_len;
+			bio->bi_vcnt--;
+			bio->bi_phys_segments--;
+			bio->bi_hw_segments--;
+
 			prev->bv_len += len;
 			if (q->merge_bvec_fn &&
 			    q->merge_bvec_fn(q, bio, prev) < len) {
 				prev->bv_len -= len;
-				return 0;
+				len = 0;
 			}
 
-			goto done;
+			bio->bi_size += prev->bv_len;
+			goto out_add_bvec;
 		}
 	}
 
@@ -394,11 +401,12 @@ static int __bio_add_page(struct request_queue *q, struct bio *bio, struct page
 	    BIOVEC_VIRT_MERGEABLE(bvec-1, bvec)))
 		bio->bi_flags &= ~(1 << BIO_SEG_VALID);
 
+	bio->bi_size += len;
+out_add_bvec:
 	bio->bi_vcnt++;
 	bio->bi_phys_segments++;
 	bio->bi_hw_segments++;
- done:
-	bio->bi_size += len;
 	return len;
 }
 
-- 
1.5.4.rc4

Re: [PATCH] fix bio_add_page for non trivial merge_bvec_fn case

[Andrew Morton]

On Sun, 08 Jun 2008 19:45:08 +0400
Dmitri Monakhov <dmonakhov@openvz.org> wrote:

> We have to properly decrease all related bio's counters, especially bi_size
> in order to merge_bvec_fn return right result. Usually this result in
> false merge rejects for two absolutely valid bio_vecs. This may cause 
> significant performance penalty for example Itanium: page_size == 16k,
> fs_block_size == 1k and block device is raid with small chunk_size.
> 

Please cc Jens on BIO changes.

> ---
>  fs/bio.c |   16 ++++++++++++----
>  1 files changed, 12 insertions(+), 4 deletions(-)
> 
> diff --git a/fs/bio.c b/fs/bio.c
> index 7856257..d713074 100644
> --- a/fs/bio.c
> +++ b/fs/bio.c
> @@ -332,14 +332,21 @@ static int __bio_add_page(struct request_queue *q, struct bio *bio, struct page
>  
>  		if (page == prev->bv_page &&
>  		    offset == prev->bv_offset + prev->bv_len) {
> +			/* Temprory detacth last bio_vec. */

whoa, drunken speling.

> +			bio->bi_size -= prev->bv_len;
> +			bio->bi_vcnt--;
> +			bio->bi_phys_segments--;
> +			bio->bi_hw_segments--;
> +
>  			prev->bv_len += len;
>  			if (q->merge_bvec_fn &&
>  			    q->merge_bvec_fn(q, bio, prev) < len) {
>  				prev->bv_len -= len;
> -				return 0;
> +				len = 0;
>  			}
>  
> -			goto done;
> +			bio->bi_size += prev->bv_len;
> +			goto out_add_bvec;
>  		}
>  	}
>  
> @@ -394,11 +401,12 @@ static int __bio_add_page(struct request_queue *q, struct bio *bio, struct page
>  	    BIOVEC_VIRT_MERGEABLE(bvec-1, bvec)))
>  		bio->bi_flags &= ~(1 << BIO_SEG_VALID);
>  
> +	bio->bi_size += len;
> +out_add_bvec:
>  	bio->bi_vcnt++;
>  	bio->bi_phys_segments++;
>  	bio->bi_hw_segments++;
> - done:
> -	bio->bi_size += len;
>  	return len;
>  }

For some reason patch(1) claims this hunk was corrupted.  I typed it in
by hand.

Re: [PATCH] fix bio_add_page for non trivial merge_bvec_fn case

[Jens Axboe]

On Mon, Jun 30 2008, Andrew Morton wrote:
> On Sun, 08 Jun 2008 19:45:08 +0400
> Dmitri Monakhov <dmonakhov@openvz.org> wrote:
> 
> > We have to properly decrease all related bio's counters, especially bi_size
> > in order to merge_bvec_fn return right result. Usually this result in
> > false merge rejects for two absolutely valid bio_vecs. This may cause 
> > significant performance penalty for example Itanium: page_size == 16k,
> > fs_block_size == 1k and block device is raid with small chunk_size.
> > 
> 
> Please cc Jens on BIO changes.
> 
> > ---
> >  fs/bio.c |   16 ++++++++++++----
> >  1 files changed, 12 insertions(+), 4 deletions(-)
> > 
> > diff --git a/fs/bio.c b/fs/bio.c
> > index 7856257..d713074 100644
> > --- a/fs/bio.c
> > +++ b/fs/bio.c
> > @@ -332,14 +332,21 @@ static int __bio_add_page(struct request_queue *q, struct bio *bio, struct page
> >  
> >  		if (page == prev->bv_page &&
> >  		    offset == prev->bv_offset + prev->bv_len) {
> > +			/* Temprory detacth last bio_vec. */
> 
> whoa, drunken speling.
> 
> > +			bio->bi_size -= prev->bv_len;
> > +			bio->bi_vcnt--;
> > +			bio->bi_phys_segments--;
> > +			bio->bi_hw_segments--;
> > +

This logic isn't quite right, the rules for what constitutes a new hw or
phys segment is not a 1:1 mapping with number of pages in the bio. How
about just dropping the segment decrement? The merge_bvec fn should not
care, and we'll retry and coalesce segment count if we get to the limit
anyway.

-- 
Jens Axboe