Re: [Qemu-devel] [PATCH 1/7] block: only force IO completion in .bdrv_truncate if we are shrinking

[Peter Lieven]

Am 08.03.2013 12:56, schrieb Kevin Wolf:
> Am 08.03.2013 um 12:46 hat Peter Lieven geschrieben:
>>
>> Am 08.03.2013 um 10:35 schrieb Kevin Wolf <address@hidden>:
>>
>>> Am 08.03.2013 um 10:23 hat Paolo Bonzini geschrieben:
>>>> Il 08/03/2013 08:53, Peter Lieven ha scritto:
>>>>>>
>>>>>> I think the fix is to only call it for the monitor command.  Optionally,
>>>>>> when shrinking, assert that there are no requests in flight.
>>>>>
>>>>> Okay.
>>>>>
>>>>> What is the plan? just fix this or fix the whole thing (which seems to
>>>>> be some
>>>>> work).
>>>>>
>>>>> The suggested patch from Jeff will break iscsi_truncate as bs->growable
>>>>> is 1 currently.
>>>>
>>>> I guess it's up to Kevin and Jeff.
>>>
>>> Someone needs to send a fix for the qcow1 (and probably vmdk)
>>> regression, otherwise I'll have to revert the patch.
>>
>> What about Paolos suggestion to call bdrv_drain_all() from the block_resize
>> command and not in bdrv_truncate? It is not a the complete solution, but it
>> will fix the regression. However, what happens if someone resizes a qcow2
>> device?
> 
> I suppose you mean qcow1? It doesn't support resizing images, but even
> if it did, this shouldn't be a problem: The problematic case is the call
> of bdrv_drain_all() during a read/write function, which would have to
> wait for itself to complete. As soon as you restrict the
> bdrv_drain_all() to the monitor, waiting for in-flight I/Os should just
> work.

Ok, then please ignore / revert the patch that added bdrv_drain_all() in 
bdrv_truncate() and
I will sent a new one that adds bdrv_drain_all() to qmp_block_resize().

Peter


> 
> Kevin
>