Re: [Xen-devel] [PATCH linux v2 0/9] xen: pvhvm: support bootup on secondary vCPUs

[Vitaly Kuznetsov <vkuznets@redhat.com>]

Julien Grall <julien.grall@arm.com> writes:

> Hello,
>
> On 25/07/16 14:39, Vitaly Kuznetsov wrote:
>> Julien Grall <julien.grall@arm.com> writes:
>>
>>> Hi David,
>>>
>>> On 25/07/16 13:38, David Vrabel wrote:
>>>> On 30/06/16 16:56, Vitaly Kuznetsov wrote:
>>>>> It may happen that Xen's and Linux's ideas of vCPU id diverge. In
>>>>> particular, when we crash on a secondary vCPU we may want to do kdump
>>>>> and unlike plain kexec where we do migrate_to_reboot_cpu() we try booting
>>>>> on the vCPU which crashed. This doesn't work very well for PVHVM guests as
>>>>> we have a number of hypercalls where we pass vCPU id as a parameter. These
>>>>> hypercalls either fail or do something unexpected. To solve the issue we
>>>>> need to have a mapping between Linux's and Xen's vCPU ids.
>>>>>
>>>>> This series solves the issue for x86 PVHVM guests. PV guests don't (and
>>>>> probably won't) support kdump so I always assume Xen's vCPU id == Linux's
>>>>> vCPU id. ARM guests will probably need to get proper mapping once we start
>>>>> supporting kexec/kdump there.
>>>>
>>>> Applied to for-linus-4.8, thanks.
>>>
>>> It would have been nice to send a ping before applying. This patch
>>> series is containing Xen ARM code which has not been acked by Stefano,
>>> nor had feedback from ARM side.
>>>
>>> For instance given that all the hypercalls are representing a "vcpu
>>> id" using "uint32_t" it is a bit weird to use "int" to define
>>> xen_vcpu_id (see patch #3).
>>
>> CPU id is usually 'int' in linux and now we pass it to all
>> hypercalls as it is.
>
> Well, we need to differentiate between the internal representation of
> the CPU which is based on the boot order and the logical CPU ID. For
> instance on ARM, the logical CPU ID may not be contiguous nor 0 for
> the first CPU.
>
> From my understanding, the macros in patch #3 will be used at the last
> minute when prepare the hypercall data. IHMO this is very similar to a
> logical ID and defined as uint32_t by the hypercall ABI.
>
> Although, I agree that currently we use the internal CPU id on ARM
> which is very unfortunate because this value is based on the order of
> the nodes in the device tree.
>
> One way to abolish it on ARM would be to use the MPIDR (or at least a
> part) for the VCPU ID.
>

I probably know too little about ARM but it seems that unlike x86 we
don't need the knowledge of _other_ vCPU ids before we start them so
MPIDR looks very suitable.

>> It is a bit more convenient in the mapping I
>> introduce as we can set it to a negative value to indicate there is no
>> mapping available. I can definitely change that and use something like
>> U32_MAX-1 to instead but I'm not sure it is worth it...
>
> I looked at the definition of cpu_acpi_id on x86 which return
> x86_cpu_to_acpiid that has been defined to an uint32_t.
>
> So you are assuming that it will never be possible to have an ID >
> 0x80000000.
>
> Also, this may not be true on ARM depending how we define the VCPU
> mapping. We could decide to use the MPIDR which is in this case may be
> considered as "negative".

While we're not obliged to have the same type for xen_vcpu_id on all
arches I see no point in diverging without a reason. I can do v3 making
the mapping uint32 and indicating the missing value as U32_MAX-1 if
nobody is against the idea.

-- 
  Vitaly