[libvirt] [PATCH v2 6/7] virsh: domstate: report detailed state if available

[Michal Privoznik]

On 4/1/19 10:17 AM, Bjoern Walk wrote:
> Michal Privoznik <mprivozn at redhat.com> [2019-03-27, 02:05PM +0100]:
>> On 3/25/19 9:04 AM, Bjoern Walk wrote:
>>> +    case VIR_DOMAIN_STATE_INFO_TYPE_QEMU_S390: {
>>> +        int core;
>>> +        unsigned long long psw_mask;
>>> +        unsigned long long psw_addr;
>>> +        const char *reason;
>>> +
>>> +        if (virTypedParamsGetInt(params, nparams,
>>> +                                 VIR_DOMAIN_STATE_PARAMS_PANIC_INFO_S390_CORE, &core) < 0 ||
>>> +            virTypedParamsGetULLong(params, nparams,
>>> +                                    VIR_DOMAIN_STATE_PARAMS_PANIC_INFO_S390_PSW_MASK, &psw_mask) < 0 ||
>>> +            virTypedParamsGetULLong(params, nparams,
>>> +                                    VIR_DOMAIN_STATE_PARAMS_PANIC_INFO_S390_PSW_ADDR, &psw_addr) < 0 ||
>>> +            virTypedParamsGetString(params, nparams,
>>> +                                    VIR_DOMAIN_STATE_PARAMS_PANIC_INFO_S390_PSW_REASON, &reason) < 0) {
>>> +            return NULL;
>>> +        }
>>> +
>>> +        ignore_value(virAsprintf(&ret, "s390: core='%d' psw-mask='0x%016llx' "
>>> +                                 "psw-addr='0x%016llx' reason='%s'",
>>> +                                 core, psw_mask, psw_addr, reason));
>>
>> Don't we want to print this in some more machine friendly way? For
>> instance one argument per line? Also, we have a function (sort of)
>> that prints out typed params. You may take the inspiration from
>> virshDomainStatsPrintRecord()
> 
> One argument per line is certainly doable. How to exploit the print
> function I am not so sure, because I want to have special formatting for
> different parameters (i.e., the ULL values for PSW_(MASK|ADDR) are hex).
> I don't see a way to get around this with a generic function.

Ah, okay. Just use whatever you want then. I did not realize this when I 
was making the suggestion.

> 
>>> +        }
>>> +        break;
>>> +    case VIR_DOMAIN_STATE_INFO_TYPE_NONE:
>>> +    case VIR_DOMAIN_STATE_INFO_TYPE_LAST:
>>> +        break;
>>> +    }
>>> +
>>> +    return ret;
>>> +}
>>> +
>>>    static bool
>>>    cmdDomstate(vshControl *ctl, const vshCmd *cmd)
>>>    {
>>>        virDomainPtr dom;
>>>        bool ret = true;
>>>        bool showReason = vshCommandOptBool(cmd, "reason");
>>> +    bool showInfo = vshCommandOptBool(cmd, "info");
>>> +    virTypedParameterPtr params = NULL;
>>> +    int nparams = 0;
>>>        int state, reason;
>>> +    char *info = NULL;
>>>    
>>>        if (!(dom = virshCommandOptDomain(ctl, cmd, NULL)))
>>>            return false;
>>>    
>>> -    if ((state = virshDomainState(ctl, dom, &reason)) < 0) {
>>> -        ret = false;
>>> -        goto cleanup;
>>> +    if (virDomainGetStateParams(dom, &state, &reason, &params, &nparams, 0) < 0) {
>>> +        if ((state = virshDomainState(ctl, dom, &reason)) < 0) {
>>> +            ret = false;
>>> +            goto cleanup;
>>> +        }
>>>        }
>>
>> This looks fishy. Firstly, if --info was requested but we're talking to
>> an older daemon then virDomainGetStateParams() will fail (the API is not
>> there) and virshDomainState() is called instead which doesn't fill out
>> @params and thus we won't print any additional info even though it was
>> requested.
> 
> I intent to have the new API function superseed the existing one, so I
> actually want to call virDomainGetStateParams() unconditionally and only
> fall back to virshDomainState() when the new API is not there (so that
> hopefully at some distant point in the future, we could deprecate the
> old stuff).

I understand that but consider the following scenario: a shell script 
that calls 'virsh domstate $dom' every so often. If we switch to new API 
then this will still fetch expected info for the client but also it will 
produce an error message logged at the server side (if it is old and 
doesn't support the new API).

That's the main reason I think we should use new API only if necessary, 
i.e. if --info was requested. But if you still want to prefer the new 
API we should do it the proper way. That is something among these lines:

if (virDomainGetStateParams() < 0) {
   if (!last_error || last_error->code != VIR_ERR_NO_SUPPORT)
     goto error;

   /* fallback */
   if (virshDomainState() < 0)
     goto error;
}

Note that we have to check for the reason why vurDomainGetStateParams() 
failed and fallback on old API if and only if the new API is not supported.

> 
>> Also, printing out some error message would be nice:
> 
> If the new API is not available, I wanted to just silently have the old
> behaviour. The domain state and reason can always be retrieve (by means
> of the old API) only additional information could be missing. I am not
> sure if we should error out if this is the case.

I think we should. I mean, if I'd run 'virsh domstate --info' that means 
I am expecting some extended info on domain state. If I don't get it, 
isn't that a failure?

> 
>> diff --git i/tools/virsh-domain-monitor.c w/tools/virsh-domain-monitor.c
>> index bf9d4970a7..27c33f354a 100644
>> --- i/tools/virsh-domain-monitor.c
>> +++ w/tools/virsh-domain-monitor.c
>> @@ -1477,11 +1477,13 @@ cmdDomstate(vshControl *ctl, const vshCmd *cmd)
>>       if (!(dom = virshCommandOptDomain(ctl, cmd, NULL)))
>>           return false;
>>   
>> -    if (virDomainGetStateParams(dom, &state, &reason, &params, &nparams, 0) < 0) {
>> -        if ((state = virshDomainState(ctl, dom, &reason)) < 0) {
>> -            ret = false;
>> -            goto cleanup;
>> -        }
>> +    if ((showInfo &&
>> +         virDomainGetStateParams(dom, &state, &reason, &params, &nparams, 0) < 0) ||
>> +        (!showInfo &&
>> +         (state = virshDomainState(ctl, dom, &reason)) < 0)) {
>> +        vshError(ctl, _("Unable to get state of domain %s"), virDomainGetName(dom));
>> +        ret = false;
>> +        goto cleanup;
>>       }
> 

Michal