Re: [Qemu-devel] [PATCH v6 0/3] linux-aio: introduce submit I/O as a batch

[Zhang Haoyu]

>>> Hi,
>>> 
>>> The commit 580b6b2aa2(dataplane: use the QEMU block layer for I/O)
>>> introduces ~40% throughput regression on virtio-blk dataplane, and
>>> one of causes is that submitting I/O as a batch is removed.
>>> 
>>> This patchset trys to introduce this mechanism on block, at least,
>>> linux-aio can benefit from that.
>>> 
>>> With these patches, it is observed that thoughout on virtio-blk
>>> dataplane can be improved a lot, see data in commit log of patch
>>> 3/3.
>>> 
>>> It should be possible to apply the batch mechanism to other devices
>>> (such as virtio-scsi) too.
>>> 
>>> TODO:
>>>     - support queuing I/O to multi files for scsi devies, which
>>>     need some changes to linux-aio
>>> 
>>> V6:
>>>     - fix requests leak if part of them arn't submitted successfully,
>>>     pointed by Stefan
>>>     - linux-aio.c coding style fix
>>> 
>>> V5:
>>>     - rebase on v2.1.0-rc0 of qemu.git/master
>>>     - block/linux-aio.c code style fix
>>>     - don't flush io queue before flush, pointed by Paolo
>>> 
>>> V4:
>>>     - support other non-raw formats with under-optimized performance
>>>     - use reference counter for plug & unplug
>>>     - flush io queue before sending flush command
>>> 
>>> V3:
>>>     - only support submitting I/O as a batch for raw format, pointed by
>>>     Kevin
>>> 
>>> V2:
>>>     - define return value of bdrv_io_unplug as void, suggested by Paolo
>>>     - avoid busy-wait for handling io_submit
>>> V1:
>>>     - move queuing io stuff into linux-aio.c as suggested by Paolo
>>
>>Thanks, applied to my block tree:
>>https://github.com/stefanha/qemu/commits/block
>>
>Can we use the queued io data as caches, 
>io write will directly return and tell the guest the io is completed after the 
>io is enqueued, 
>better user experience for burst io,
>and io-read will firstly search the io queue, if matched data found, directly 
>get the data from the queue, 
>if not, then read the data from the disk or host page cache.

Because host page cache cannot assure the order of io-write, leading to VM 
image data corruption if suddenly poweroff happened,
but we can assure the order of io-write in queue, and use 
cache=none/writethrough/directsync to submit the batched IO to disk directly, 
bypass host page cache.
IO merging also can be performed in the queue.

>Any ideas?
>
>Thanks,
>Zhang Haoyu
>
>>In Patch 2 we should complete requests with -EIO if io_submit() returned
>>0 <= ret < len.  I fixed this up when applying because the patch was
>>completing with a bogus ret value.
>>
>>Stefan