Re: [Qemu-devel] [RFC 00/13] aio: drop io_flush()

[Stefan Hajnoczi <stefanha@gmail.com>]

On Fri, Apr 12, 2013 at 2:22 PM, Kevin Wolf <kwolf@redhat.com> wrote:
> Am 12.04.2013 um 14:06 hat Stefan Hajnoczi geschrieben:
>> On Fri, Apr 12, 2013 at 12:27 PM, Paolo Bonzini <pbonzini@redhat.com> wrote:
>> > Il 12/04/2013 12:04, Kevin Wolf ha scritto:
>> >> Or actually, if you want to avoid .bdrv_drain for now, the patch that I
>> >> started when I thought a bit about this, had a default .bdrv_drain
>> >> implementation that just forwarded the request to bs->file if it wasn't
>> >> implemented by a block driver. For the QED case, this would work.
>> >
>> > It's similar to bdrv_co_flush.  Drain bs first, then bs->backing_hd
>> > (this is not needed in bdrv_co_flush), then bs->file, then the driver
>> > can do it on other files.
>>
>> Thanks for explaining Kevin.  I didn't check whether bdrv_states
>> included all BlockDriverStates.
>>
>> The simplest solution is to put all BlockDriverStates on a global list.
>>
>> A .bdrv_drain() interface is more flexible but I don't see a need for it yet.
>
> We don't need the .bdrv_drain() callback yet, but please implement it
> without a global list and with a function that forwards requests to
> bs->file and bs->backing_hd like Paolo suggested. This will be the
> obvious place to add the .bdrv_drain() callback later.
>
> The reason why I prefer this approach is that it allows draining a
> single BlockDriverState. After your patches we're really close to
> enabling this and converting some bdrv_drain_all() to more specific
> bdrv_drain(bs) calls.

Fair enough, that approach does make it possible to drain single
devices in the future.

Stefan