Re: [RFC PATCH 09/10] target/riscv: Restrict KVM-specific fields from ArchCPU

[Daniel Henrique Barboza]

On 4/5/23 13:04, Philippe Mathieu-Daudé wrote:
> These fields shouldn't be accessed when KVM is not available.
>
> Signed-off-by: Philippe Mathieu-Daudé <philmd@linaro.org>
> ---
> RFC: The migration part is likely invalid...
>
> kvmtimer_needed() is defined in target/riscv/machine.c as
>
>    static bool kvmtimer_needed(void *opaque)
>    {
>        return kvm_enabled();
>    }
>
> which depends on a host feature.


kvm_enabled() can be false even when CONFIG_KVM is true when a KVM capable host
is running a TCG guest, for example. In that case env->kvm_timer_* states exist
but aren't initialized, and shouldn't be migrated.

Thus it's not just a host feature, but a host feature + accel option. I think
this is fine.

> ---
>   target/riscv/cpu.h     | 2 ++
>   target/riscv/machine.c | 4 ++++
>   2 files changed, 6 insertions(+)
>
> diff --git a/target/riscv/cpu.h b/target/riscv/cpu.h
> index 638e47c75a..82939235ab 100644
> --- a/target/riscv/cpu.h
> +++ b/target/riscv/cpu.h
> @@ -377,12 +377,14 @@ struct CPUArchState {
>       hwaddr kernel_addr;
>       hwaddr fdt_addr;
>   
> +#ifdef CONFIG_KVM
>       /* kvm timer */
>       bool kvm_timer_dirty;
>       uint64_t kvm_timer_time;
>       uint64_t kvm_timer_compare;
>       uint64_t kvm_timer_state;
>       uint64_t kvm_timer_frequency;
> +#endif /* CONFIG_KVM */
>   };
>   
>   OBJECT_DECLARE_CPU_TYPE(RISCVCPU, RISCVCPUClass, RISCV_CPU)
> diff --git a/target/riscv/machine.c b/target/riscv/machine.c
> index 9c455931d8..e45d564ec3 100644
> --- a/target/riscv/machine.c
> +++ b/target/riscv/machine.c
> @@ -201,10 +201,12 @@ static bool kvmtimer_needed(void *opaque)
>   
>   static int cpu_post_load(void *opaque, int version_id)
>   {
> +#ifdef CONFIG_KVM
>       RISCVCPU *cpu = opaque;
>       CPURISCVState *env = &cpu->env;
>   
>       env->kvm_timer_dirty = true;
> +#endif
>       return 0;
>   }
>   
> @@ -215,9 +217,11 @@ static const VMStateDescription vmstate_kvmtimer = {
>       .needed = kvmtimer_needed,
>       .post_load = cpu_post_load,
>       .fields = (VMStateField[]) {
> +#ifdef CONFIG_KVM
>           VMSTATE_UINT64(env.kvm_timer_time, RISCVCPU),
>           VMSTATE_UINT64(env.kvm_timer_compare, RISCVCPU),
>           VMSTATE_UINT64(env.kvm_timer_state, RISCVCPU),
> +#endif

Here you're creating an empty 'cpu/kvmtimer' vmstate that won't be migrated 
anyway
because kvmtimer_needed (== kvm_enabled()) will be always false if CONFIG_KVM=n.

I'd say it's better to just get rid of the whole vmstate in this case, but I 
don't
like the precedence of having vmstates being gated by build flags.


Reviewed-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com>



>           VMSTATE_END_OF_LIST()
>       }
>   };