Re: [Qemu-devel] [PATCH] blockdev-backup: don't check aio_context too early

[John Snow <jsnow@redhat.com>]

On 5/9/19 4:00 AM, Kevin Wolf wrote:
> Am 09.05.2019 um 00:55 hat John Snow geschrieben:
>>
>> On 5/7/19 4:50 AM, Kevin Wolf wrote:
>>> Am 06.05.2019 um 22:33 hat John Snow geschrieben:
>>>> in blockdev_backup_prepare, we check to make sure that the target is
>>>> associated with a compatible aio context. However, do_blockdev_backup is
>>>> called later and has some logic to move the target to a compatible
>>>> aio_context. The transaction version will fail certain commands
>>>> needlessly early as a result.
>>>>
>>>> Allow blockdev_backup_prepare to simply call do_blockdev_backup, which
>>>> will ultimately decide if the contexts are compatible or not.
>>>>
>>>> Note: the transaction version has always disallowed this operation since
>>>> its initial commit bd8baecd (2014), whereas the version of
>>>> qmp_blockdev_backup at the time, from commit c29c1dd312f, tried to
>>>> enforce the aio_context switch instead. It's not clear, and I can't see
>>>> from the mailing list archives at the time, why the two functions take a
>>>> different approach. It wasn't until later in efd7556708b (2016) that the
>>>> standalone version tried to determine if it could set the context or
>>>> not.
>>>>
>>>> Reported-by: aihua liang <aliang@redhat.com>
>>>> Fixes: https://bugzilla.redhat.com/show_bug.cgi?id=1683498
>>>
>>> Signed-off-by is missing, and a testcase, too. :-)
>>>
>>>> diff --git a/blockdev.c b/blockdev.c
>>>> index 79fbac8450..a81d88980c 100644
>>>> --- a/blockdev.c
>>>> +++ b/blockdev.c
>>>> @@ -1872,10 +1872,6 @@ static void blockdev_backup_prepare(BlkActionState *common, Error **errp)
>>>>      }
>>>>  
>>>>      aio_context = bdrv_get_aio_context(bs);
>>>> -    if (aio_context != bdrv_get_aio_context(target)) {
>>>> -        error_setg(errp, "Backup between two IO threads is not implemented");
>>>> -        return;
>>>> -    }
>>>>      aio_context_acquire(aio_context);
>>>>      state->bs = bs;
>>>
>>> The actual change looks good to me.
>>>
>>> Kevin
>>>
>>
>> (See the Red Hat BZ for details on the test I am more or less trying to
>> replicate in iotests -- but the jist of it is using an iothread on one
>> device and not specifying one for the other, then backing up both to two
>> blockdev-add created nodes not attached to any device.)
>>
>> Is there some trick to getting this to fail with accel=qtest? I'm
>> noticing that if the VM is paused or if I use accel=qtest, the iothread
>> contexts for the drives appear to go ... unresolved? and the test passes.
>>
>> I've only had luck (so far) reproducing this with accel=kvm on a running
>> VM (the drives can be empty) and I don't actually know why that is just
>> yet -- I guess these aio_context objects get resolved at runtime?
>>
>> Anyway, this seems to be a little tricky so far, maybe you have some
>> advice for me?
> 
> Are you using virtio-blk? I think it only actually enables dataplane
> (and moves its node to an iothread) during initialisation of the guest
> driver. This makes it quite annoying for testing, I've had the same
> problem before.
> 
> However, virtio-scsi does what you would expect and puts the node into
> the right AioContext directly on creation.
> 
> So maybe just switching to virtio-scsi solves you problem?
> 

Aha. I will give that a try, thanks! I was indeed using virtio-blk.

>> (Also note: doing a full backup and an incremental backup for two
>> perfectly empty 64GB qcow2 drives takes over 6 seconds in total. It
>> probably shouldn't, right? There's something about the initial backup
>> that appears to take a pretty long time.)
> 
> Sounds like Someone (TM) should have a closer look, yes.
> 
> Kevin
> 

OK, I'll dig into that after I finish this piece.

Thank you,
--js