[libvirt] [PATCH v4 5/8] qemu_monitor: implement query-cpu-model-comparison

Collin Walling walling at linux.ibm.com
Thu Jul 25 18:26:25 UTC 2019


On 7/24/19 12:18 PM, Boris Fiuczynski wrote:
> On 7/17/19 4:03 PM, Collin Walling wrote:

[...]

>> diff --git a/src/qemu/qemu_monitor_json.c b/src/qemu/qemu_monitor_json.c
>> index b599ee6..3c33c63 100644
>> --- a/src/qemu/qemu_monitor_json.c
>> +++ b/src/qemu/qemu_monitor_json.c
>> @@ -5875,6 +5875,101 @@ 
>> qemuMonitorJSONGetCPUModelBaseline(qemuMonitorPtr mon,
>>   }
>> +static int
>> +qemuMonitorJSONParseCPUModelPropName(size_t pos ATTRIBUTE_UNUSED,
>> +                                     virJSONValuePtr item,
>> +                                     void *opaque)
>> +{
>> +    if 
>> (!qemuMonitorJSONParseCPUModelProperty(virJSONValueGetString(item),
>> +                                              item, opaque))
> 
> This need to be changed into
> 
>      if (qemuMonitorJSONParseCPUModelProperty(virJSONValueGetString(item),
>                                               item, opaque) < 0)
> 

Yes, thank you for catching this one.

>> +        return -1;
>> +
>> +    virJSONValueFree(item);
>> +    return 0;
>> +}
>> +
>> +
>> +int
>> +qemuMonitorJSONGetCPUModelComparison(qemuMonitorPtr mon,
>> +                                     const char *model_a_name,
>> +                                     int model_a_nprops,
>> +                                     virCPUFeatureDefPtr model_a_props,
>> +                                     const char *model_b_name,
>> +                                     int model_b_nprops,
>> +                                     virCPUFeatureDefPtr model_b_props,
>> +                                     qemuMonitorCPUModelInfoPtr 
>> *model_result)
>> +{
>> +    int ret = -1;
>> +    virJSONValuePtr model_a;
>> +    virJSONValuePtr model_b = NULL;
>> +    virJSONValuePtr cmd = NULL;
>> +    virJSONValuePtr reply = NULL;
>> +    virJSONValuePtr data;
>> +    const char *result_name;
>> +    virJSONValuePtr props;
>> +    qemuMonitorCPUModelInfoPtr compare = NULL;
>> +
>> +    if (!(model_a = qemuMonitorJSONMakeCPUModel(model_a_name, 
>> model_a_nprops,
>> +                                                model_a_props, true)) ||
>> +        !(model_b = qemuMonitorJSONMakeCPUModel(model_b_name, 
>> model_b_nprops,
>> +                                                model_b_props, true)))
>> +        goto cleanup;
>> +
>> +    if (!(cmd = qemuMonitorJSONMakeCommand("query-cpu-model-comparison",
>> +                                           "a:modela", &model_a,
>> +                                           "a:modelb", &model_b,
>> +                                           NULL)))
>> +        goto cleanup;
>> +
>> +    if (qemuMonitorJSONCommand(mon, cmd, &reply) < 0)
>> +        goto cleanup;
>> +
>> +    if (qemuMonitorJSONCheckError(cmd, reply) < 0)
> 
> Did you test scenarios which contain unknown cpu model name or unknown 
> features?
> This caused errors for me during testing that resulted in e.g.
>   error : qemuMonitorJSONCheckError:415 : internal error: unable to 
> execute QEMU command 'query-cpu-model-comparison': Property '.boris' not 
> found
>   warning : virQEMUCapsLogProbeFailure:4586 : Failed to probe 
> capabilities for /usr/bin/qemu-system-s390x: internal error: unable to 
> execute QEMU command 'query-cpu-model-comparison': Property '.boris' not 
> found
> 
> If I think about a scenario like: I run "virsh domcapabilties" on a new 
> machine with new cpu model and new cpu features, new libvirt, new qemu 
> and use "virsh hypervisor-cpu-compare" with the output xml on my old 
> machine with old cpu &feature, older libvirt and older qemu
> I would expect to get as an answer: Incompatible
> 
> If my expectation to get incompatible is wrong please correct me but an 
> "internal error" is not what should occur.
> What is the qemu specification for this?
> 
> 

Two routes:

1) Fix up the logic in virQEMUCapsCPUModelComparison

Please see my comments on patch 8.

2) Change the error class in the QEMU QMP code

The QMP code in QEMU currently reports a "GenericError" whenever
something goes wrong (such as an unknown CPU model or feature). A
possible solution to this would be to alter that code to set a specific
error class, have libvirt check for that specific error type in the
JSON, then handle it appropriately to print a cleaner message.

There is one drawback: we would technically be reporting this error
twice -- once in virsh, and another in the libvirt daemon (the internal
error) unless we wanted to suppress it.

This is one solution without too much messy code.

>> +        goto cleanup;
>> +
>> +    data = virJSONValueObjectGetObject(reply, "return");
>> +
>> +    if (!(result_name = virJSONValueObjectGetString(data, "result"))) {
>> +        virReportError(VIR_ERR_INTERNAL_ERROR, "%s",
>> +                       _("query-cpu-model-comparison reply data was 
>> missing "
>> +                         "'result'"));
>> +        goto cleanup;
>> +    }
>> +
>> +    if (!(props = virJSONValueObjectGetArray(data, 
>> "responsible-properties"))) {
>> +        virReportError(VIR_ERR_INTERNAL_ERROR, "%s",
>> +                       _("query-cpu-model-comparison reply data was 
>> missing "
>> +                         "'responsible-properties'"));
>> +        goto cleanup;
>> +    }
>> +
>> +    if (VIR_ALLOC(compare) < 0)
>> +        goto cleanup;
>> +
>> +    if (VIR_STRDUP(compare->name, result_name) < 0)
>> +        goto cleanup;
>> +
>> +    if (VIR_ALLOC_N(compare->props, virJSONValueArraySize(props)) < 0)
>> +        goto cleanup;
>> +
>> +    if (virJSONValueArrayForeachSteal(props,
>> +                                      
>> qemuMonitorJSONParseCPUModelPropName,
>> +                                      compare) < 0)
>> +        goto cleanup;
>> +
>> +    VIR_STEAL_PTR(*model_result, compare);
>> +    ret = 0;
>> +
>> + cleanup:
>> +    qemuMonitorCPUModelInfoFree(compare);
>> +    virJSONValueFree(cmd);
>> +    virJSONValueFree(reply);
>> +    virJSONValueFree(model_a);
>> +    virJSONValueFree(model_b);
>> +    return ret;
>> +}
>> +
>> +

[...]




More information about the libvir-list mailing list