RE: [PATCH v4 00/21] AMX Support in KVM

[Tian, Kevin]

> From: Zhong, Yang <yang.zhong@xxxxxxxxx>
> Sent: Wednesday, December 29, 2021 9:13 PM
> 
> Highly appreciate for your review. This version mostly addressed the
> comments
> from Sean. Most comments are adopted except three which are not closed
> and
> need more discussions:
> 
>   - Move the entire xfd write emulation code to x86.c. Doing so requires
>     introducing a new kvm_x86_ops callback to disable msr write bitmap.
>     According to Paolo's earlier comment he prefers to handle it in vmx.c.
> 
>   - Directly check msr_bitmap in update_exception_bitmap() (for
>     trapping #NM) and vcpu_enter_guest() (for syncing guest xfd after
>     vm-exit) instead of introducing an extra flag in the last patch. However,
>     doing so requires another new kvm_x86_ops callback for checking
>     msr_bitmap since vcpu_enter_guest() is x86 common code. Having an
>     extra flag sounds simpler here (at least for the initial AMX support).
>     It does penalize nested guest with one xfd sync per exit, but it's not
>     worse than a normal guest which initializes xfd but doesn't run
>     AMX applications at all. Those could be improved afterwards.

Another option is to move xfd sync into vmx_vcpu_run(), given that
disabling xfd interception is vmx specific code thus it makes some
sense to also handle related side-effect in vmx.c. If this can be agreed
then yes there is no need of an extra flag and just checking msr_bitmap
is sufficient (and more accurate).

Paolo, how is your opinion?

> 
>   - Disable #NM trap for nested guest. This version still chooses to always
>     trap #NM (regardless in L1 or L2) as long as xfd write interception is
> disabled.
>     In reality #NM is rare if nested guest doesn't intend to run AMX
> applications
>     and always-trap is safer than dynamic trap for the basic support in case
>     of any oversight here.
> 

Sean just pointed out some potential issues in current logic, about handling
#NM raised in L2 guest (could happen just due to L1 interception).

It's being discussed here:

	https://lore.kernel.org/all/YcyaN7V4wwGI7wDV@xxxxxxxxxx/

Thanks
Kevin