[Qemu-devel] Re: [PATCH 1/3] expose paravirt feature list to cpuid

[Anthony Liguori <aliguori@us.ibm.com>]

Glauber Costa wrote:
> Expose the paravirt features provided by the host to the guest.
> The features exposed uses qemu just as a medium, since they only
> depend on a negotiation between the host and guest directly.
>
> As a direct result, we're now able to run the kvm pvclock in qemu
> based kvm guests.
>
> Signed-off-by: Glauber Costa <glommer@redhat.com>
> ---
>  kvm.h             |    1 +
>  target-i386/kvm.c |   49 +++++++++++++++++++++++++++++++++++++++++++++++--
>  2 files changed, 48 insertions(+), 2 deletions(-)
>
> diff --git a/kvm.h b/kvm.h
> index efce145..db58baa 100644
> --- a/kvm.h
> +++ b/kvm.h
> @@ -17,6 +17,7 @@
>  #include "config.h"
>
>  #ifdef CONFIG_KVM
> +#include <linux/kvm_para.h>
>  extern int kvm_allowed;
>
>  #define kvm_enabled() (kvm_allowed)
> diff --git a/target-i386/kvm.c b/target-i386/kvm.c
> index 2412ae4..7f4e5aa 100644
> --- a/target-i386/kvm.c
> +++ b/target-i386/kvm.c
> @@ -33,16 +33,61 @@
>      do { } while (0)
>  #endif
>
> +struct kvm_para_features {
> +    int cap;
> +    int feature;
> +} para_features[] = {
> +#ifdef KVM_CAP_CLOCKSOURCE
> +	{ KVM_CAP_CLOCKSOURCE, KVM_FEATURE_CLOCKSOURCE },
> +#endif
> +#ifdef KVM_CAP_NOP_IO_DELAY
> +	{ KVM_CAP_NOP_IO_DELAY, KVM_FEATURE_NOP_IO_DELAY },
> +#endif
> +#ifdef KVM_CAP_PV_MMU
> +	{ KVM_CAP_PV_MMU, KVM_FEATURE_MMU_OP },
> +#endif
> +#ifdef KVM_CAP_CR3_CACHE
> +	{ KVM_CAP_CR3_CACHE, KVM_FEATURE_CR3_CACHE },
> +#endif
> +};
> +
> +static uint32_t get_para_features(CPUState *env)
> +{
> +    uint32_t i, features = 0;
> +
> +    for (i = 0; i < ARRAY_SIZE(para_features); i++) {
> +        if (kvm_ioctl(env->kvm_state, KVM_CHECK_EXTENSION, para_features[i].cap))
> +            features |= (1 << para_features[i].feature);
> +    }
> +
> +    return features;
> +}
>  int kvm_arch_init_vcpu(CPUState *env)
>  {
>      struct {
>          struct kvm_cpuid cpuid;
>          struct kvm_cpuid_entry entries[100];
>      } __attribute__((packed)) cpuid_data;
> -    uint32_t limit, i, cpuid_i;
> +    uint32_t limit, i, cpuid_i = 0;
>      uint32_t eax, ebx, ecx, edx;
> +#ifdef KVM_CPUID_SIGNATURE
> +    struct kvm_cpuid_entry *pv_ent;
> +
> +    /* Paravirtualization CPUIDs */
> +    pv_ent = &cpuid_data.entries[cpuid_i++];
> +    memset(pv_ent, 0, sizeof(*pv_ent));
> +    pv_ent->function = KVM_CPUID_SIGNATURE;
> +    pv_ent->eax = 0;
> +    pv_ent->ebx = 0x4b4d564b; /* KVMKVMK */
> +    pv_ent->ecx = 0x564b4d56; /* VMKV */
> +    pv_ent->edx = 0x0000004d; /* M */
> +
> +    pv_ent = &cpuid_data.entries[cpuid_i++];
> +    memset(pv_ent, 0, sizeof(*pv_ent));
> +    pv_ent->function = KVM_CPUID_FEATURES;
> +    pv_ent->eax = get_para_features(env);
>   

I think this would be better suited for target-i386/helper.c in 
cpu_x86_cpuid().  It can be in an if (kvm_enabled()) for now but there 
should be a comment mentioning that we could enable some of these 
features in normal QEMU.

It can call a kvm helper function from target-i386/kvm.c to determine 
which features are supported (that could be later hooked by TCG).

Regards,

Anthony Liguori

> +#endif
>
> -    cpuid_i = 0;
>
>      cpu_x86_cpuid(env, 0, &eax, &ebx, &ecx, &edx);
>      limit = eax;
>