Re: [Qemu-devel] Failing iotests in v2.3.0-rc2 / master

[Andreas Färber]

Am 15.04.2015 um 11:26 schrieb Kevin Wolf:
> Am 15.04.2015 um 06:53 hat Jeff Cody geschrieben:
>> On Tue, Apr 14, 2015 at 11:57:35AM +0200, Kevin Wolf wrote:
>>> Am 11.04.2015 um 05:41 hat Andreas Färber geschrieben:
>>>> Hi,
>>>>
>>>> 001 seems to hang for -qcow (or is not reasonably "quick": >5 min).
>>>>
>>>> 033 is failing for -vhdx.
>>>>
>>>> (Note that `make check-block` only tests -qcow2, so didn't uncover
>>>> either of them.)
>>>>
>>>> Given a failing test, am I seeing correctly that there is no command
>>>> line option to skip this one failing test? -x seems to be for groups only.
>>>>
>>>> Regards,
>>>> Andreas
>>>>
>>>> $ ./check -v -T -qcow -g quick
>>>> [...]
>>>> 001 6s ...        [05:12:39]
>>>
>>> qcow1 is just really slow. 001 passes for me, after 202 seconds (that's
>>> on my SSD, YMMV).
>>>
>>>> $ ./check -v -T -vhdx -g quick
>>>> [...]
>>>> 033 1s ...        [04:06:09] [04:06:11] - output mismatch (see 033.out.bad)
>>>
>>> This seems to be because blkdebug doesn't implement .bdrv_truncate.
>>> Currently the test case isn't suitable for VHDX, which uses explicit
>>> bdrv_truncate() calls to grow the image file. I'll send a patch for
>>> blkdebug to allow this.
>>>
>>> However, it seems that there is another problem which causes assertion
>>> failures when using VHDX over blkdebug. Jeff, does the following fix
>>> make sense to you? (I think it does, but I don't understand yet why the
>>> assertion failure is only triggered with blkdebug - or in other words:
>>> "how could this ever work?")
>>>
>>> Kevin
>>
>> Kevin,
>>
>> Yes, looking at that fix it makes sense - we are wanting to pad the
>> back part of the block after the actual data with zeros. That back
>> length should be (block size - (bytes avail + block offset)), which is
>> iov2.iov_len.
>>
>> There are two reasons I think we haven't seen this issue (it has been
>> hidden):
>>
>> 1.) If bs->file supports zero init, we don't do any of this
> 
> I see. file does and blkdebug doesn't, so that's the crucial difference.
> 
>> 2.) This is done for the case when the existing BAT state is
>> PAYLOAD_BLOCK_ZERO.  Until recently (commit 30af51c), we didn't
>> create VHDX files with blocks in the PAYLOAD_BLOCK_ZERO state.
> 
> Right, I wasn't aware of this either any more.
> 
>> So it has been a latent bug in a hitherto rarely (if ever) exercised
>> path.
> 
> Thanks for your explanation, it's clear to me now what's going on. I'll
> send out the patches (for both blkdebug and vhdx) right away. You can
> either pick up the vhdx one, or just give your Acked-by and then I'll
> merge it through my block tree.

Might 059 (?) failure for -vmdk be another symptom of the same issue?

Thanks,
Andreas

-- 
SUSE Linux GmbH, Maxfeldstr. 5, 90409 Nürnberg, Germany
GF: Felix Imendörffer, Jane Smithard, Jennifer Guild, Dilip Upmanyu,
Graham Norton; HRB 21284 (AG Nürnberg)