Re: [Qemu-devel] [PATCH v2 00/12] block: qiov_offset parameter for io

[Vladimir Sementsov-Ogievskiy]

28.06.2019 11:43, Stefan Hajnoczi wrote:
> On Tue, Jun 04, 2019 at 07:15:02PM +0300, Vladimir Sementsov-Ogievskiy wrote:
>> Hi all!
>>
>> Here is new parameter qiov_offset for io path, to avoid
>> a lot of places with same pattern of creating local_qiov or hd_qiov
>> variables.
>>
>> These series also includes my
>> "[Qemu-devel] [PATCH 0/2] block/io: refactor padding"
>> with some changes [described in 01 and 03 emails]
>>
>> Vladimir Sementsov-Ogievskiy (12):
>>    util/iov: introduce qemu_iovec_init_extended
>>    util/iov: improve qemu_iovec_is_zero
>>    block/io: refactor padding
>>    block: define .*_part io handlers in BlockDriver
>>    block/io: bdrv_co_do_copy_on_readv: use and support qiov_offset
>>    block/io: bdrv_co_do_copy_on_readv: lazy allocation
>>    block/io: bdrv_aligned_preadv: use and support qiov_offset
>>    block/io: bdrv_aligned_pwritev: use and support qiov_offset
>>    block/io: introduce bdrv_co_p{read,write}v_part
>>    block/qcow2: refactor qcow2_co_preadv to use buffer-based io
>>    block/qcow2: implement .bdrv_co_preadv_part
>>    block/qcow2: implement .bdrv_co_pwritev(_compressed)_part
>>
>>   block/qcow2.h             |   1 +
>>   include/block/block_int.h |  21 ++
>>   include/qemu/iov.h        |  10 +-
>>   block/backup.c            |   2 +-
>>   block/io.c                | 532 ++++++++++++++++++++++----------------
>>   block/qcow2-cluster.c     |  14 +-
>>   block/qcow2.c             | 131 +++++-----
>>   qemu-img.c                |   4 +-
>>   util/iov.c                | 153 +++++++++--
>>   9 files changed, 559 insertions(+), 309 deletions(-)
>>
>> -- 
>> 2.18.0
>>
>>
> 
> I don't see a significant advantage after taking into account more
> complex code (e.g. additional block driver interfaces) and the risk of
> introducing new bugs.  A measurable performance improvement would make
> this refactoring more attractive.  Still:
> 
> Acked-by: Stefan Hajnoczi <address@hidden>
> 

Hmm, I understand your doubt. And I doubt that there will be some significant 
performance
gain.

What will you say if instead of adding new interfaces I just add qiov_offset 
parameter to
old ones, and add one boolean field to BlockDriver like .supports_qiov_offset, 
so, generic
code will use non-zero qiov_offset only for drivers supporting it? And, maybe, 
add corresponding
asserts to all not-supporting driver handlers?

-- 
Best regards,
Vladimir