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

[Dmitry Monakhov <dmonakhov@openvz.org>]

On Tue, Jan 26, 2010 at 4:29 PM, Jens Axboe <jens.axboe@oracle.com> wrote:
> On Tue, Jan 26 2010, Dmitry Monakhov wrote:
>> Hi, year ago I've sent a patch which fix false bio merge rejects, but
>> seems patch was missed. Currently the issue is still present.
>>
>
>> From 92a97ef181e15caa94bd56a1ade5c337db599b79 Mon Sep 17 00:00:00 2001
>> From: Dmitry Monakhov <dmonakhov@openvz.org>
>> Date: Tue, 26 Jan 2010 16:01:34 +0300
>> Subject: [PATCH] [PATCH] block: fix bio_add_page for non trivial merge_bvec_fn case
>>
>> We have to properly decrease bi_size in order to merge_bvec_fn return
>> right result.  Otherwise 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: Dmitry Monakhov <dmonakhov@openvz.org>
>> ---
>>  fs/bio.c |    3 ++-
>>  1 files changed, 2 insertions(+), 1 deletions(-)
>>
>> diff --git a/fs/bio.c b/fs/bio.c
>> index 76e6713..9f8e517 100644
>> --- a/fs/bio.c
>> +++ b/fs/bio.c
>> @@ -548,7 +548,8 @@ static int __bio_add_page(struct request_queue *q, struct bio *bio, struct page
>>                               struct bvec_merge_data bvm = {
>>                                       .bi_bdev = bio->bi_bdev,
>>                                       .bi_sector = bio->bi_sector,
>> -                                     .bi_size = bio->bi_size,
>> +                                     .bi_size = bio->bi_size -
>> +                                                     (prev->bv_len - len),
>>                                       .bi_rw = bio->bi_rw,
>>                               };
>
> Hmm confused. why isn't this just bio->bi_size - len?
because we have following scenario:
0) old_bv_len = prev->bv_len ;
1) prev->bv_len += len;
2)->merge_bvec_fn()
      usually it looks like follows
      if (bio->bv_len + bvm->bi_size > max_chunk_size)
          goto fail
     So formally we detach last bvec increase it size and try to
attach it to bio.

     but old_bv_len is already accounted in bi_size, so we have to decrease
    bi_size to this value, and old_bv_len = (prev->bv_len - len) at
that moment.
    I've used math manipuation in order to avoid temporary variable
    because stack size is matter on block layer.
>
> --
> Jens Axboe
>
>