arm_cpu_post_init (Was: Re: arm: "max" CPU class hierarchy changes possible?)

[Claudio Fontana <cfontana@suse.de>]

On 3/11/21 8:10 PM, Andrew Jones wrote:
> On Thu, Mar 11, 2021 at 06:33:15PM +0000, Peter Maydell wrote:
>> On Thu, 11 Mar 2021 at 17:16, Claudio Fontana <cfontana@suse.de> wrote:
>>> Maybe Peter you could clarify similarly what the intended meaning of "max" is on ARM?
>>
>> "max" is "best we can do, whatever that is". (On KVM this is "same as
>> the host".)
>> "host" is "whatever the host is (KVM only)".
>>
>>> KVM: (aarch64-only): aarch64_max_initfn():
>>>
>>> The following comment in the code seems wrong to me:
>>>
>>> /* -cpu max: if KVM is enabled, like -cpu host (best possible with this host); */
>>>
>>> This is not exactly true:
>>>
>>> "-cpu max" calls kvm_arm_set_cpu_features_from_host(), (which checks "dtb_compatible", and if not set gets the features from the host, if set ...?)
>>> After that, calls aarch64_add_sve_properties() and then adds also "svw-max-vq". This code is common with TCG.


As part of this research I noticed that arm_cpu_post_init() is quite confusing, seems really inconsistent to me.

Apparently the intention was to call it from the leaf classes:

commit 51e5ef459eca045d7e8afe880ee60190f0b75b26
Author: Marc-André Lureau <marcandre.lureau@redhat.com>
Date:   Tue Nov 27 12:55:59 2018 +0400

    arm: replace instance_post_init()
    
    Replace arm_cpu_post_init() instance callback by calling it from leaf
    classes, to avoid potential ordering issue with other post_init callbacks.
    
    Signed-off-by: Marc-André Lureau <marcandre.lureau@redhat.com>
    Suggested-by: Igor Mammedov <imammedo@redhat.com>
    Reviewed-by: Igor Mammedov <imammedo@redhat.com>
    Acked-by: Eduardo Habkost <ehabkost@redhat.com>


but then we end up calling it multiple times in the class hierarch, which is a recipe for bugs, and makes it difficult to understand what arm_cpu_post_init()
even means, what calling this function is supposed to do.

For a "max" or "host" cpu on AArch64, this function is called:

for the ARM CPU base class, TYPE_ARM_CPU, in

cpu.c::arm_cpu_instance_init,

then later again for the TYPE_AARCH64_CPU class, child of TYPE_ARM_CPU, in

cpu64.c::aarch64_cpu_instance_init,

then later again for the TYPE_ARM_HOST_CPU class, child of TYPE_AARCH64_CPU, in

cpu.c::arm_host_initfn.

Same for "max".

When looking at 32bit CPUs instead, only the ARM CPU base class ends up calling arm_cpu_post_init.
"Leaf" classes do not do it (see cpu_tcg.c).

What is then arm_cpu_post_init even supposed to mean?

Thanks,

Claudio


>>>
>>> In the case of cpu host instead,
>>>
>>> "-cpu host" calls kvm_arm_set_cpu_features_from_host(), same as max, then calls aarch64_add_sve_properties() but does NOT add "svw-max-vq".
>>>
>>> Is this a bug?
> 
> It was left out intentionally. More below.
> 
>>
>> Maybe; that's a question for Richard or Drew...
>>
>>> Are "max" and "host" for KVM supposed to be the same like with x86?
> 
> Yes, but my understanding of "max" == "host" for KVM is that that only
> applies to the perspective of the guest. What CPU and what CPU features
> the guest can see should be exactly the same with either "max" or "host",
> depending on the enabling/disabling of any optional CPU properties.
> 
> The question here seems to be that, if one has a CPU property, does that
> imply the other should have the same? Which would effectively allow the
> two to be aliases (when KVM is enabled). I don't know, does x86 ensure
> 100% property compatibility?
> 
> I opted not to support sve-max-vq for "host" because I consider it a
> legacy CPU property, one I didn't want to propagate. Indeed it may
> make more sense to depreciate sve-max-vq than to "fix" this issue
> by adding it to "host". Note, we can already create equivalent SVE
> CPUs. The following are the same from the perspective of the guest
> 
>  -accel kvm -cpu host,sve512=on
>  -accel kvm -cpu max,sve512=on
> 
> And, for TCG, these are the same from the perspective of the guest
>  
>  -accel tcg -cpu max,sve512=on
>  -accel tcg -cpu max,sve-max-vq=4
> 
> So we already don't need sve-max-vq.
> 
> Thanks,
> drew
>