Re: [libvirt] [FOR 1.3.0 PATCH] conf: add net device prefix for Xen

[Jim Fehlig <jfehlig@suse.com>]

On 02/22/2016 04:15 PM, Joao Martins wrote:
>
> On 12/08/2015 04:06 PM, Jim Fehlig wrote:
>> Daniel P. Berrange wrote:
>>> On Mon, Dec 07, 2015 at 10:59:32PM -0700, Jim Fehlig wrote:
>>>> On 12/07/2015 11:52 AM, Daniel P. Berrange wrote:
>>>>> On Mon, Dec 07, 2015 at 09:42:21AM -0700, Jim Fehlig wrote:
>>>>>> In commit d2e5538b1, the libxl driver was changed to copy interface
>>>>>> names autogenerated by libxl to the corresponding network def in the
>>>>>> domain's virDomainDef object. The copied name is freed when the domain
>>>>>> transitions to the shutoff state. But when migrating a domain, the
>>>>>> autogenerated name is included in the XML sent to the destination host.
>>>>>> It is possible an interface with the same name already exists on the
>>>>>> destination host, causing migration to fail. Indeed the Xen project's
>>>>>> OSSTEST CI already encountered such a failure.
>>>>>>
>>>>>> This patch defines another VIR_NET_GENERATED_PREFIX for Xen, allowing
>>>>>> the autogenerated names to be excluded when parsing and formatting
>>>>>> inactive config.
>>>>>>
>>>>>> Signed-off-by: Jim Fehlig <jfehlig@suse.com>
>>>>>> ---
>>>>>>
>>>>>> This is an alternative approach to Joao's fix for this regression
>>>>>>
>>>>>> https://www.redhat.com/archives/libvir-list/2015-December/msg00197.html
>>>>>>
>>>>>> I think it is the same approach used by the qemu driver. My only
>>>>>> reservation is that it expands the potential for clashes with
>>>>>> user-defined names. I.e. with this change both 'vnet' and 'vif' are
>>>>>> reserved prefixes.
>>>>> Hmm, yes, tricky one.
>>>>>
>>>>> If we only care about XML parsing, then you could register a post
>>>>> parse callback instead to do this.
>>>> AFAIK, XML parsing is all that's in play here.
>>>>
>>>>> I'm not clear why we also have it in the virDomainNetDefFormat
>>>>> method - and we can't solve that with a post-parse callback.
>>>>>
>>>>>
>>>>> The other option would be to make the reserved prefix be a
>>>>> capability that the parser/formatter could read.
>>>> This seems like the best option, since a post-parse callback doesn't solve the
>>>> problem in virDomainNetDefFormat. It also has the upshot of making the prefix
>>>> visible and known to users. But I doubt such a change is suitable during 1.3.0
>>>> freeze.  With the freeze in mind, seems the best solution to the libxl migration
>>>> regression is revert d2e5538b1. It can be added again post-1.3.0 release, after
>>>> adding the prefix to capabilities.
>>>>
>>>> DV, since you may be making the release soon, feel free to revert d2e5538b1 if
>>>> you agree.
>>> Yeah, just go ahead & revert it Jim,  DV isn't doing the releae until
>>> tomorrow morning
>> I've pushed the revert.
>>
>> Joao, sorry for yanking this for 1.3.0. We can get it in 1.3.1, after exposing
>> the prefix in capabilities.
>>
> Hey Jim,
>
> I had the impression we had pushed this back in (i.e. cherry-picking d2e5538)
> but I was double checking and that's doesn't seem to be case. In adding support
> for the prefix in capabilities I found out one issue on the migration failure
> path leading to dereferencing a NULL pointer on the destination. If you agree
> could we squash the following chunk in addition to the cherry-pick of d2e5538:
>
> diff --git a/src/libxl/libxl_domain.c b/src/libxl/libxl_domain.c
> index 63c5b24..f73bfb3 100644
> --- a/src/libxl/libxl_domain.c
> +++ b/src/libxl/libxl_domain.c
> @@ -768,7 +768,7 @@ libxlDomainCleanup(libxlDriverPrivatePtr driver,
>          for (i = 0; i < vm->def->nnets; i++) {
>              virDomainNetDefPtr net = vm->def->nets[i];
>
> -            if (STRPREFIX(net->ifname, "vif"))
> +            if (net->ifname && STRPREFIX(net->ifname, "vif"))
>                  VIR_FREE(net->ifname);
>          }
>      }
> diff --git a/src/libxl/libxl_driver.c b/src/libxl/libxl_driver.c
> index 3f4457f..5479441 100644
> --- a/src/libxl/libxl_driver.c
> +++ b/src/libxl/libxl_driver.c
> @@ -5590,7 +5590,7 @@ static virHypervisorDriver libxlHypervisorDriver = {
>      .domainGetJobStats = libxlDomainGetJobStats, /* 1.3.1 */
>      .domainMemoryStats = libxlDomainMemoryStats, /* 1.3.0 */
>      .domainGetCPUStats = libxlDomainGetCPUStats, /* 1.3.0 */
> -    .domainInterfaceStats = libxlDomainInterfaceStats, /* 1.3.0 */
> +    .domainInterfaceStats = libxlDomainInterfaceStats, /* 1.3.2 */
>      .connectDomainEventRegister = libxlConnectDomainEventRegister, /* 0.9.0 */
>      .connectDomainEventDeregister = libxlConnectDomainEventDeregister, /* 0.9.0 */
>      .domainManagedSave = libxlDomainManagedSave, /* 0.9.2 */
>
>
> Or do you think it should be deferred to 1.3.3 ?

I wanted to ask you about this on the recent Xen+libvirt+OpenStack sync. Thanks
for checking. It would be a shame if this patch didn't make 1.3.2 since all of
your prerequisite work is in, but perhaps a new patch is best given the changes?

Regards,
Jim