Re: [PATCH v2 3/3] target/arm: Restrict v7A TCG cpus to TCG accel

["Philippe Mathieu-Daudé" <philmd@redhat.com>]

On 3/18/21 11:47 AM, Philippe Mathieu-Daudé wrote:
> On 3/18/21 10:56 AM, Claudio Fontana wrote:
>> On 3/18/21 10:47 AM, Philippe Mathieu-Daudé wrote:
>>> Hi Peter,
>>>
>>> On 3/11/21 11:43 AM, Claudio Fontana wrote:
>>>> On 2/21/21 11:26 PM, Philippe Mathieu-Daudé wrote:
>>>>> KVM requires the target cpu to be at least ARMv8 architecture
>>>>> (support on ARMv7 has been dropped in commit 82bf7ae84ce:
>>>>> "target/arm: Remove KVM support for 32-bit Arm hosts").
>>>>>
>>>>> A KVM-only build won't be able to run TCG cpus, move the
>>>>> v7A CPU definitions to cpu_tcg.c.
>>>>>
>>>>> Reported-by: Peter Maydell <peter.maydell@linaro.org>
>>>>> Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
>>>>> Signed-off-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
>>>>
>>>>
>>>> Here I think it's better to keep the "a15" cpu around,
>>>> until we fix the board configuration situation.
>>>>
>>>> I added a patch that does that into my KVM-only build series, to avoid the resulting breakages.
>>>
>>> Actually I got a downstream report that this break migration from QEMU
>>> 5.2 to QEMU 6.0.
>>>
>>> I first thought it was on an old kernel (with 32-bit KVM enabled),
>>> but not, it is for Aarch64 VMs on recent KVM (without 32-bit support)
>>> but the 'virt' machines were started with default Cortex-A15 CPU...
>>>
>>>     mc->default_cpu_type = ARM_CPU_TYPE_NAME("cortex-a15");
>>>
>>> I'm not sure upstream should care about this case, but I though
>>> maybe you could give me hints about the best way to keep old VMs
>>> working, as this likely affects any distribution. Obviously not
>>> upgrading QEMU is not a solution :)
>>>
>>> The simplest seems to revert 82bf7ae84ce and this patch, but I
>>> doubt this will be enough.
>>>
>>> Maybe there is some clever thing to do before reverting 82bf7ae84ce,
>>> that could also benefit upstream, by doing something with versioned
>>> machines? I have no idea (yet) how that work and if it could work.
>>
>> Does just applying my series fix it?
> 
> But we are past soft-freeze so I'm looking for a surgical fix.
> 
> I'll send a partial revert for now.

Still, I'm not sure it makes sense. If you want to migrate a such
machine, KVM can not virtualize it, so you'll be forced to use TCG
right? In that case cpu_tcg is built in and you have the A15.

IOW the problem is not this patch, it is that since 82bf7ae84ce we
can not migrate A15.

So we need both 1/ revert 82bf7ae84ce and 2/ be sure the kernel
support 32-bit host... Am I missing something?