[libvirt] [PATCH 4/4] qemu: hot-unplug of watchdog

[Michal Privoznik]

On 09/12/2017 04:29 PM, John Ferlan wrote:
> 
> 
> On 09/05/2017 07:45 AM, Michal Privoznik wrote:
>> https://bugzilla.redhat.com/show_bug.cgi?id=1447169
>>
>> Signed-off-by: Michal Privoznik <mprivozn at redhat.com>
>> ---
>>  src/qemu/qemu_driver.c                             |  4 +-
>>  src/qemu/qemu_hotplug.c                            | 61 ++++++++++++++++++++++
>>  src/qemu/qemu_hotplug.h                            |  3 ++
>>  tests/qemuhotplugtest.c                            |  7 ++-
>>  .../qemuhotplug-watchdog-full.xml                  |  3 ++
>>  5 files changed, 76 insertions(+), 2 deletions(-)
>>  create mode 100644 tests/qemuhotplugtestdevices/qemuhotplug-watchdog-full.xml
>>
>> diff --git a/src/qemu/qemu_driver.c b/src/qemu/qemu_driver.c
>> index 626148dba..4c636b956 100644
>> --- a/src/qemu/qemu_driver.c
>> +++ b/src/qemu/qemu_driver.c
>> @@ -7692,12 +7692,14 @@ qemuDomainDetachDeviceLive(virDomainObjPtr vm,
>>      case VIR_DOMAIN_DEVICE_SHMEM:
>>          ret = qemuDomainDetachShmemDevice(driver, vm, dev->data.shmem);
>>          break;
>> +    case VIR_DOMAIN_DEVICE_WATCHDOG:
>> +        ret = qemuDomainDetachWatchdog(driver, vm, dev->data.watchdog);
>> +        break;
>>  
>>      case VIR_DOMAIN_DEVICE_FS:
>>      case VIR_DOMAIN_DEVICE_INPUT:
>>      case VIR_DOMAIN_DEVICE_SOUND:
>>      case VIR_DOMAIN_DEVICE_VIDEO:
>> -    case VIR_DOMAIN_DEVICE_WATCHDOG:
>>      case VIR_DOMAIN_DEVICE_GRAPHICS:
>>      case VIR_DOMAIN_DEVICE_HUB:
>>      case VIR_DOMAIN_DEVICE_SMARTCARD:
>> diff --git a/src/qemu/qemu_hotplug.c b/src/qemu/qemu_hotplug.c
>> index 989146eb9..44472ce9f 100644
>> --- a/src/qemu/qemu_hotplug.c
>> +++ b/src/qemu/qemu_hotplug.c
>> @@ -4360,6 +4360,25 @@ qemuDomainRemoveShmemDevice(virQEMUDriverPtr driver,
>>  }
>>  
>>  
>> +static int
>> +qemuDomainRemoveWatchdog(virQEMUDriverPtr driver,
>> +                         virDomainObjPtr vm,
>> +                         virDomainWatchdogDefPtr watchdog)
>> +{
>> +    virObjectEventPtr event = NULL;
>> +
>> +    VIR_DEBUG("Removing watchdog %s from domain %p %s",
> 
> Is "Removing watchdog watchdog0 from ..." a bit superfluous?
> 
> Perhaps just "Removing '%s' from ..."
> 
>> +              watchdog->info.alias, vm, vm->def->name);
>> +
> 
> This would/could be the place for the virDomainAuditWatchdog for
> "detach" too.
> 
>> +    event = virDomainEventDeviceRemovedNewFromObj(vm, watchdog->info.alias);
>> +    qemuDomainEventQueue(driver, event);
>> +    qemuDomainReleaseDeviceAddress(vm, &watchdog->info, NULL);
>> +    virDomainWatchdogDefFree(vm->def->watchdog);
>> +    vm->def->watchdog = NULL;
>> +    return 0;
>> +}
>> +
>> +
>>  int
>>  qemuDomainRemoveDevice(virQEMUDriverPtr driver,
>>                         virDomainObjPtr vm,
>> @@ -5076,6 +5095,48 @@ qemuDomainDetachShmemDevice(virQEMUDriverPtr driver,
>>  }
>>  
>>  
>> +int
>> +qemuDomainDetachWatchdog(virQEMUDriverPtr driver,
>> +                         virDomainObjPtr vm,
>> +                         virDomainWatchdogDefPtr dev)
>> +{
>> +    int ret = -1;
>> +    virDomainWatchdogDefPtr watchdog;
> 
> Why not watchdog = vm->def->watchdog; here?
> 
>> +    qemuDomainObjPrivatePtr priv = vm->privateData;
>> +
>> +    /* While domains can have up to one watchdog, the one supplied by user
> 
> s/by/by the/
> 
>> +     * doesn't necessarily match the one domain has. Refuse to detach in such
>> +     * case. */
>> +    if (!(vm->def->watchdog &&
>> +          STREQ_NULLABLE(dev->info.alias,
>> +                         vm->def->watchdog->info.alias))) {
> 
> if (!watchdog || STRNEQ_NULLABLE(dev->info.alias, watchdog->info.alias))
> 
> Trying to think why NULLABLE would be necessary... So it seems it's
> required that that input XML has the alias - something not quite right
> with that...

Is that so? We match devices based on their alias in a lot of places.
Users are expected to pass the full device XML anyway. So it's up to us
how we pick the right device to be detached. For instance, when
detaching a vNIC we match MAC addresses and ignore the rest, since MAC
identifies the device uniquely. So for instance, if the detach XML
provided by user has <bandwidth/> set but the one in domain doesn't have
it, we detach the device anyway. One can argue this is a buggy
behaviour. But I just don't think we should care. There's a line we have
to draw between protecting users from themselves and too complex and
mostly useless code. So we've documented that users are supposed to pass
the device XML as is in the domain XML.

> 
> I would think for unplug there'd be something that would compare the
> input "model" and "action" values to whatever watchdog is currently
> present and if they don't match, then fail. Not sure why comparing the
> alias is "right". It's not something we require of other unplugs, is it?

Sort of. For instance when detaching disks we only care about the
target. And for a lot of other devices we require alias too.

> We just need to make sure the same "something" is being removed. Since
> the alias is essentially static for the single available watchdog
> device, then as long as the model and action are the same, we can
> remove; otherwise, someone has to fix the input XML to have the matching
> model and action.

Okay, I'll add these two checks.

Michal