Re: [Qemu-devel] [PATCH v3 8/9] qmp: add query-block-jobs

[Kevin Wolf]

Am 15.12.2011 09:27, schrieb Stefan Hajnoczi:
> On Wed, Dec 14, 2011 at 03:54:52PM +0100, Kevin Wolf wrote:
>> Am 13.12.2011 14:52, schrieb Stefan Hajnoczi:
>>> diff --git a/hmp.c b/hmp.c
>>> index 66d9d0f..c16d6a1 100644
>>> --- a/hmp.c
>>> +++ b/hmp.c
>>> @@ -499,6 +499,46 @@ void hmp_info_pci(Monitor *mon)
>>>      qapi_free_PciInfoList(info);
>>>  }
>>>  
>>> +void hmp_info_block_jobs(Monitor *mon)
>>> +{
>>> +    BlockJobInfoList *list;
>>> +    Error *err = NULL;
>>> +
>>> +    list = qmp_query_block_jobs(&err);
>>> +    assert(!err);
>>> +
>>> +    if (!list) {
>>> +        monitor_printf(mon, "No active jobs\n");
>>> +        return;
>>> +    }
>>> +
>>> +    while (list) {
>>> +        /* The HMP output for streaming jobs is special because 
>>> historically it
>>> +         * was different from other job types so applications may depend 
>>> on the
>>> +         * exact string.
>>> +         */
>>
>> Er, what? This is new code. What HMP clients use this string? I know
>> that libvirt already got support for this before we implemented it, but
>> shouldn't that be QMP only?
> 
> Libvirt HMP uses this particular string, which turned out to be
> sub-optimal once I realized we might support other types of block jobs
> in the future.
> 
> You can still build libvirt HMP-only by disabling the yajl library
> dependency.  The approach I've taken is to make the interfaces available
> over both HMP and QMP (and so has the libvirt-side code).
> 
> In any case, we have defined both HMP and QMP.  Libvirt implements both
> and I don't think there's a reason to provide only QMP.
> 
> Luiz: For future features, are we supposed to provide only QMP
> interfaces, not HMP?

Of course, qemu should provide them as HMP command. But libvirt
shouldn't use HMP commands. HMP is intended for human users, not as an
API for management.

And I was pretty sure that we all agreed that HMP should be considered
unstable after a transition period. For new commands there's certainly
no reason to have a transition period, so I would consider them unstable
from the very beginning. After all, there are no qemu versions that
support the feature in question, but don't support QMP.

Kevin