Re: [PATCH 3/4] qemu: kvm: Enable XSAVE live migration support

[Jan Kiszka <jan.kiszka@siemens.com>]

Sheng Yang wrote:
> Signed-off-by: Sheng Yang <sheng@linux.intel.com>
> ---
>  target-i386/cpu.h     |    7 ++-
>  target-i386/kvm.c     |  139 ++++++++++++++++++++++++++++++++++++++++++++++++-
>  target-i386/machine.c |   20 +++++++
>  3 files changed, 163 insertions(+), 3 deletions(-)
> 
> diff --git a/target-i386/cpu.h b/target-i386/cpu.h
> index 548ab80..680eed1 100644
> --- a/target-i386/cpu.h
> +++ b/target-i386/cpu.h
> @@ -718,6 +718,11 @@ typedef struct CPUX86State {
>      uint16_t fpus_vmstate;
>      uint16_t fptag_vmstate;
>      uint16_t fpregs_format_vmstate;
> +
> +    uint64_t xstate_bv;
> +    XMMReg ymmh_regs[CPU_NB_REGS];
> +
> +    uint64_t xcr0;
>  } CPUX86State;
>  
>  CPUX86State *cpu_x86_init(const char *cpu_model);
> @@ -895,7 +900,7 @@ uint64_t cpu_get_tsc(CPUX86State *env);
>  #define cpu_list_id x86_cpu_list
>  #define cpudef_setup	x86_cpudef_setup
>  
> -#define CPU_SAVE_VERSION 11
> +#define CPU_SAVE_VERSION 12
>  
>  /* MMU modes definitions */
>  #define MMU_MODE0_SUFFIX _kernel
> diff --git a/target-i386/kvm.c b/target-i386/kvm.c
> index bb6a12f..e490c0a 100644
> --- a/target-i386/kvm.c
> +++ b/target-i386/kvm.c
> @@ -289,6 +289,8 @@ void kvm_arch_reset_vcpu(CPUState *env)
>      } else {
>          env->mp_state = KVM_MP_STATE_RUNNABLE;
>      }
> +    /* Legal xcr0 for loading */
> +    env->xcr0 = 1;
>  }
>  
>  static int kvm_has_msr_star(CPUState *env)
> @@ -504,6 +506,68 @@ static int kvm_put_fpu(CPUState *env)
>      return kvm_vcpu_ioctl(env, KVM_SET_FPU, &fpu);
>  }
>  
> +#ifdef KVM_CAP_XSAVE
> +#define XSAVE_CWD_RIP     2
> +#define XSAVE_CWD_RDP     4
> +#define XSAVE_MXCSR       6
> +#define XSAVE_ST_SPACE    8
> +#define XSAVE_XMM_SPACE   40
> +#define XSAVE_XSTATE_BV   128
> +#define XSAVE_YMMH_SPACE  144
> +#endif
> +
> +static int kvm_put_xsave(CPUState *env)
> +{
> +#ifdef KVM_CAP_XSAVE
> +    int i;
> +    struct kvm_xsave* xsave;
> +    uint16_t cwd, swd, twd, fop;
> +
> +    if (kvm_check_extension(env->kvm_state, KVM_CAP_XSAVE))

That's still one syscall too much for this path (which will be a
fast-path for Kemari). Get that value during arch_init.

> +        return kvm_put_fpu(env);
> +
> +    xsave = qemu_memalign(4096, sizeof(struct kvm_xsave));
> +    memset(xsave, 0, sizeof(struct kvm_xsave));
> +    cwd = swd = twd = fop = 0;
> +    swd = env->fpus & ~(7 << 11);
> +    swd |= (env->fpstt & 7) << 11;
> +    cwd = env->fpuc;
> +    for (i = 0; i < 8; ++i)
> +        twd |= (!env->fptags[i]) << i;
> +    xsave->region[0] = (uint32_t)(swd << 16) + cwd;
> +    xsave->region[1] = (uint32_t)(fop << 16) + twd;
> +    memcpy(&xsave->region[XSAVE_ST_SPACE], env->fpregs,
> +            sizeof env->fpregs);
> +    memcpy(&xsave->region[XSAVE_XMM_SPACE], env->xmm_regs,
> +            sizeof env->xmm_regs);
> +    xsave->region[XSAVE_MXCSR] = env->mxcsr;
> +    *(uint64_t *)&xsave->region[XSAVE_XSTATE_BV] = env->xstate_bv;
> +    memcpy(&xsave->region[XSAVE_YMMH_SPACE], env->ymmh_regs,
> +            sizeof env->ymmh_regs);
> +    return kvm_vcpu_ioctl(env, KVM_SET_XSAVE, xsave);
> +#else
> +    return kvm_put_fpu(env);
> +#endif
> +}
> +
> +static int kvm_put_xcrs(CPUState *env)
> +{
> +#ifdef KVM_CAP_XCRS
> +    struct kvm_xcrs xcrs;
> +
> +    if (!kvm_check_extension(env->kvm_state, KVM_CAP_XCRS))
> +        return 0;
> +
> +    xcrs.nr_xcrs = 1;
> +    xcrs.flags = 0;
> +    xcrs.xcrs[0].xcr = 0;
> +    xcrs.xcrs[0].value = env->xcr0;
> +    return kvm_vcpu_ioctl(env, KVM_SET_XCRS, &xcrs);
> +#else
> +    return 0;
> +#endif
> +}
> +
>  static int kvm_put_sregs(CPUState *env)
>  {
>      struct kvm_sregs sregs;
> @@ -621,6 +685,69 @@ static int kvm_get_fpu(CPUState *env)
>      return 0;
>  }
>  
> +static int kvm_get_xsave(CPUState *env)
> +{
> +#ifdef KVM_CAP_XSAVE
> +    struct kvm_xsave* xsave;
> +    int ret, i;
> +    uint16_t cwd, swd, twd, fop;
> +
> +    if (!kvm_check_extension(env->kvm_state, KVM_CAP_XSAVE))
> +        return kvm_get_fpu(env);
> +
> +    xsave = qemu_memalign(4096, sizeof(struct kvm_xsave));
> +    ret = kvm_vcpu_ioctl(env, KVM_GET_XSAVE, xsave);
> +    if (ret < 0)
> +        return ret;
> +
> +    cwd = (uint16_t)xsave->region[0];
> +    swd = (uint16_t)(xsave->region[0] >> 16);
> +    twd = (uint16_t)xsave->region[1];
> +    fop = (uint16_t)(xsave->region[1] >> 16);
> +    env->fpstt = (swd >> 11) & 7;
> +    env->fpus = swd;
> +    env->fpuc = cwd;
> +    for (i = 0; i < 8; ++i)
> +        env->fptags[i] = !((twd >> i) & 1);
> +    env->mxcsr = xsave->region[XSAVE_MXCSR];
> +    memcpy(env->fpregs, &xsave->region[XSAVE_ST_SPACE],
> +            sizeof env->fpregs);
> +    memcpy(env->xmm_regs, &xsave->region[XSAVE_XMM_SPACE],
> +            sizeof env->xmm_regs);
> +    env->xstate_bv = *(uint64_t *)&xsave->region[XSAVE_XSTATE_BV];
> +    memcpy(env->ymmh_regs, &xsave->region[XSAVE_YMMH_SPACE],
> +            sizeof env->ymmh_regs);
> +    return 0;
> +#else
> +    return kvm_get_fpu(env);
> +#endif
> +}
> +
> +static int kvm_get_xcrs(CPUState *env)
> +{
> +#ifdef KVM_CAP_XCRS
> +    int i, ret;
> +    struct kvm_xcrs xcrs;
> +
> +    if (!kvm_check_extension(env->kvm_state, KVM_CAP_XCRS))
> +        return 0;
> +
> +    ret = kvm_vcpu_ioctl(env, KVM_GET_XCRS, &xcrs);
> +    if (ret < 0)
> +        return ret;
> +
> +    for (i = 0; i < xcrs.nr_xcrs; i++)
> +        /* Only support xcr0 now */
> +        if (xcrs.xcrs[0].xcr == 0) {
> +            env->xcr0 = xcrs.xcrs[0].value;
> +            break;
> +        }
> +    return 0;
> +#else
> +    return 0;
> +#endif
> +}
> +
>  static int kvm_get_sregs(CPUState *env)
>  {
>      struct kvm_sregs sregs;
> @@ -965,7 +1092,11 @@ int kvm_arch_put_registers(CPUState *env, int level)
>      if (ret < 0)
>          return ret;
>  
> -    ret = kvm_put_fpu(env);
> +    ret = kvm_put_xsave(env);
> +    if (ret < 0)
> +        return ret;
> +
> +    ret = kvm_put_xcrs(env);
>      if (ret < 0)
>          return ret;
>  
> @@ -1009,7 +1140,11 @@ int kvm_arch_get_registers(CPUState *env)
>      if (ret < 0)
>          return ret;
>  
> -    ret = kvm_get_fpu(env);
> +    ret = kvm_get_xsave(env);
> +    if (ret < 0)
> +        return ret;
> +
> +    ret = kvm_get_xcrs(env);
>      if (ret < 0)
>          return ret;
>  
> diff --git a/target-i386/machine.c b/target-i386/machine.c
> index b547e2a..5f8376c 100644
> --- a/target-i386/machine.c
> +++ b/target-i386/machine.c
> @@ -47,6 +47,22 @@ static const VMStateDescription vmstate_xmm_reg = {
>  #define VMSTATE_XMM_REGS(_field, _state, _n)                         \
>      VMSTATE_STRUCT_ARRAY(_field, _state, _n, 0, vmstate_xmm_reg, XMMReg)
>  
> +/* YMMH format is the same as XMM */
> +static const VMStateDescription vmstate_ymmh_reg = {
> +    .name = "ymmh_reg",
> +    .version_id = 1,
> +    .minimum_version_id = 1,
> +    .minimum_version_id_old = 1,
> +    .fields      = (VMStateField []) {
> +        VMSTATE_UINT64(XMM_Q(0), XMMReg),
> +        VMSTATE_UINT64(XMM_Q(1), XMMReg),
> +        VMSTATE_END_OF_LIST()
> +    }
> +};
> +
> +#define VMSTATE_YMMH_REGS_VARS(_field, _state, _n, _v)                         \
> +    VMSTATE_STRUCT_ARRAY(_field, _state, _n, _v, vmstate_ymmh_reg, XMMReg)
> +
>  static const VMStateDescription vmstate_mtrr_var = {
>      .name = "mtrr_var",
>      .version_id = 1,
> @@ -453,6 +469,10 @@ static const VMStateDescription vmstate_cpu = {
>          /* KVM pvclock msr */
>          VMSTATE_UINT64_V(system_time_msr, CPUState, 11),
>          VMSTATE_UINT64_V(wall_clock_msr, CPUState, 11),
> +        /* XSAVE related fields */
> +        VMSTATE_UINT64_V(xcr0, CPUState, 12),
> +        VMSTATE_UINT64_V(xstate_bv, CPUState, 12),
> +        VMSTATE_YMMH_REGS_VARS(ymmh_regs, CPUState, CPU_NB_REGS, 12),
>          VMSTATE_END_OF_LIST()
>          /* The above list is not sorted /wrt version numbers, watch out! */
>      }

Looks good otherwise.

Thanks,
Jan

-- 
Siemens AG, Corporate Technology, CT T DE IT 1
Corporate Competence Center Embedded Linux