Re: [Qemu-devel] [PATCH 01/15] qemu coroutine: support bypass mode

[Ming Lei]

On Thu, Jul 31, 2014 at 6:45 PM, Paolo Bonzini <address@hidden> wrote:
> Il 31/07/2014 11:47, Ming Lei ha scritto:
>>> Block mirroring of a device for example is done using coroutines.
>>> As block mirroring can be done on a raw device you need coroutines.
>>
>> If block layer knows the mirroring is in-progress, it still can enable
>> coroutine by ignoring bypass coroutine, or just let device disable
>> bypass coroutine in case of mirroring, and the current APIs are very
>> flexible.
>
> What matters is not whether you're mirroring.  What matters is whether
> you're calling bdrv_aio_readv/writev or bdrv_co_readv/writev.  Under
> some limitation, if you are calling bdrv_aio_readv/writev you can bypass
> coroutines.  (In fact drive mirroring uses those APIs too so it doesn't
> need coroutines and can benefit from the speedup too. :)  But
> drive-backup does need them).
>
> The limitations are essentially "would bdrv_co_do_preadv or
> bdrv_co_do_pwritev do anything special?"  This means for example for
> bdrv_co_do_pwritev:
>
> - bs->drv must be non-NULL
>
> - bs->read_only must be false
>
> - bdrv_check_byte_request(bs, offset, bytes) must return false
>
> - bs->io_limits_enabled must be false
>
> - the request must be aligned
>
> - (in bdrv_aligned_pwritev) the before_write_notifiers must be empty
>
> - (in bdrv_aligned_pwritev) bs->detect_zeroes must be off
>
> - (in bdrv_aligned_pwritev) the BDRV_REQ_ZERO_WRITE flag must be off
>
> - (in bdrv_aligned_pwritev) bs->enable_write_cache must be false
>
> and the hard part is organizing the code so that the code duplication is
> as limited as possible.

In this patchset, only implementation of bypassing coroutine is
introduced, and simply apply it on raw image for dataplane, and it
won't break drive mirror & backup, if I understand correctly. So I
suggest we focus on the implementation of bypassing coroutine.

All the above is mostly about if the bypass can be applied in
other cases, so this patchset shouldn't have caused code duplication.

I appreciate very much the implementation of bypassing coroutine
can be reviewed in the discussion.

Thanks,