Re: [RFC PATCH 2/8] xen/arm: add sve_vl_bits field to domain

[Luca Fancellu <Luca.Fancellu@xxxxxxx>]

> On 11 Jan 2023, at 17:27, Julien Grall <julien@xxxxxxx> wrote:
> 
> Hi Luca,
> 
> On 11/01/2023 14:38, Luca Fancellu wrote:
>> Add sve_vl_bits field to arch_domain and xen_arch_domainconfig
>> structure, to allow the domain to have an information about the
>> SVE feature and the number of SVE register bits that are allowed
>> for this domain.
>> The field is used also to allow or forbid a domain to use SVE,
>> because a value equal to zero means the guest is not allowed to
>> use the feature.
>> When the guest is allowed to use SVE, the zcr_el2 register is
>> updated on context switch to restict the domain on the allowed
>> number of bits chosen, this value is the minimum among the chosen
>> value and the platform supported value.
>> Signed-off-by: Luca Fancellu <luca.fancellu@xxxxxxx>
>> ---
>>  xen/arch/arm/arm64/sve.c             |  9 ++++++
>>  xen/arch/arm/domain.c                | 45 ++++++++++++++++++++++++++++
>>  xen/arch/arm/include/asm/arm64/sve.h | 12 ++++++++
>>  xen/arch/arm/include/asm/domain.h    |  6 ++++
>>  xen/include/public/arch-arm.h        |  2 ++
>>  xen/include/public/domctl.h          |  2 +-
>>  6 files changed, 75 insertions(+), 1 deletion(-)
>> diff --git a/xen/arch/arm/arm64/sve.c b/xen/arch/arm/arm64/sve.c
>> index 326389278292..b7695834f4ba 100644
>> --- a/xen/arch/arm/arm64/sve.c
>> +++ b/xen/arch/arm/arm64/sve.c
>> @@ -6,6 +6,7 @@
>>   */
>>    #include <xen/types.h>
>> +#include <asm/cpufeature.h>
>>  #include <asm/arm64/sve.h>
>>  #include <asm/arm64/sysregs.h>
>>  @@ -36,3 +37,11 @@ register_t vl_to_zcr(uint16_t vl)
>>  {
>>      return ((vl / SVE_VL_MULTIPLE_VAL) - 1U) & ZCR_ELx_LEN_MASK;
>>  }
>> +
>> +/* Get the system sanitized value for VL in bits */
>> +uint16_t get_sys_vl_len(void)
>> +{
>> +    /* ZCR_ELx len field is ((len+1) * 128) = vector bits length */
>> +    return ((system_cpuinfo.zcr64.bits[0] & ZCR_ELx_LEN_MASK) + 1U) *
>> +            SVE_VL_MULTIPLE_VAL;
>> +}
>> diff --git a/xen/arch/arm/domain.c b/xen/arch/arm/domain.c
>> index 8ea3843ea8e8..27f38729302b 100644
>> --- a/xen/arch/arm/domain.c
>> +++ b/xen/arch/arm/domain.c
>> @@ -13,6 +13,7 @@
>>  #include <xen/wait.h>
>>    #include <asm/alternative.h>
>> +#include <asm/arm64/sve.h>
>>  #include <asm/cpuerrata.h>
>>  #include <asm/cpufeature.h>
>>  #include <asm/current.h>
>> @@ -183,6 +184,11 @@ static void ctxt_switch_to(struct vcpu *n)
>>        WRITE_SYSREG(n->arch.cptr_el2, CPTR_EL2);
>>  +#ifdef CONFIG_ARM64_SVE
>> +    if ( is_sve_domain(n->domain) )
>> +        WRITE_SYSREG(n->arch.zcr_el2, ZCR_EL2);
>> +#endif
> 
> I would actually expect that is_sve_domain() returns false when the SVE is 
> not enabled. So can we simply remove the #ifdef?

I was tricked by it too, the arm32 build will fail because it doesn’t know 
ZCR_EL2

> 
>> +
>>      /* VFP */
>>      vfp_restore_state(n);
>>  @@ -551,6 +557,11 @@ int arch_vcpu_create(struct vcpu *v)
>>      v->arch.vmpidr = MPIDR_SMP | vcpuid_to_vaffinity(v->vcpu_id);
>>        v->arch.cptr_el2 = get_default_cptr_flags();
>> +    if ( is_sve_domain(v->domain) )
>> +    {
>> +        v->arch.cptr_el2 &= ~HCPTR_CP(8);
>> +        v->arch.zcr_el2 = vl_to_zcr(v->domain->arch.sve_vl_bits);
>> +    }
>>        v->arch.hcr_el2 = get_default_hcr_flags();
>>  @@ -595,6 +606,7 @@ int arch_sanitise_domain_config(struct 
>> xen_domctl_createdomain *config)
>>      unsigned int max_vcpus;
>>      unsigned int flags_required = (XEN_DOMCTL_CDF_hvm | XEN_DOMCTL_CDF_hap);
>>      unsigned int flags_optional = (XEN_DOMCTL_CDF_iommu | 
>> XEN_DOMCTL_CDF_vpmu);
>> +    unsigned int sve_vl_bits = config->arch.sve_vl_bits;
>>        if ( (config->flags & ~flags_optional) != flags_required )
>>      {
>> @@ -603,6 +615,36 @@ int arch_sanitise_domain_config(struct 
>> xen_domctl_createdomain *config)
>>          return -EINVAL;
>>      }
>>  +    /* Check feature flags */
>> +    if ( sve_vl_bits > 0 ) {
>> +        unsigned int zcr_max_bits;
>> +
>> +        if ( !cpu_has_sve )
>> +        {
>> +            dprintk(XENLOG_INFO, "SVE is unsupported on this machine.\n");
>> +            return -EINVAL;
>> +        }
>> +        else if ( !is_vl_valid(sve_vl_bits) )
>> +        {
>> +            dprintk(XENLOG_INFO, "Unsupported SVE vector length (%u)\n",
>> +                    sve_vl_bits);
>> +            return -EINVAL;
>> +        }
>> +        /*
>> +         * get_sys_vl_len() is the common safe value among all cpus, so if 
>> the
>> +         * value specified by the user is above that value, use the safe 
>> value
>> +         * instead.
>> +         */
>> +        zcr_max_bits = get_sys_vl_len();
>> +        if ( sve_vl_bits > zcr_max_bits )
>> +        {
>> +            config->arch.sve_vl_bits = zcr_max_bits;
>> +            dprintk(XENLOG_INFO,
>> +                    "SVE vector length lowered to %u, safe value among 
>> CPUs\n",
>> +                    zcr_max_bits);
>> +        }
> 
> I don't think Xen should "downgrade" the value. Instead, this should be the 
> decision from the tools. So we want to find a different way to reproduce the 
> value (Andrew may have some ideas here as he was looking at it).

Can you explain this in more details? By the tools you mean xl? This would be 
ok for DomUs, but how to do it for Dom0?
I’ll wait suggestions also from Andrew.

> 
>> +    }
>> +
>>      /* The P2M table must always be shared between the CPU and the IOMMU */
>>      if ( config->iommu_opts & XEN_DOMCTL_IOMMU_no_sharept )
>>      {
>> @@ -745,6 +787,9 @@ int arch_domain_create(struct domain *d,
>>      if ( (rc = domain_vpci_init(d)) != 0 )
>>          goto fail;
>>  +    /* Copy sve_vl_bits to the domain configuration */
>> +    d->arch.sve_vl_bits = config->arch.sve_vl_bits;
>> +
>>      return 0;
>>    fail:
>> diff --git a/xen/arch/arm/include/asm/arm64/sve.h 
>> b/xen/arch/arm/include/asm/arm64/sve.h
>> index bd56e2f24230..f4a660e402ca 100644
>> --- a/xen/arch/arm/include/asm/arm64/sve.h
>> +++ b/xen/arch/arm/include/asm/arm64/sve.h
>> @@ -13,10 +13,17 @@
>>  /* Vector length must be multiple of 128 */
>>  #define SVE_VL_MULTIPLE_VAL (128U)
>>  +static inline bool is_vl_valid(uint16_t vl)
>> +{
>> +    /* SVE vector length is multiple of 128 and maximum 2048 */
>> +    return ((vl % SVE_VL_MULTIPLE_VAL) == 0) && (vl <= SVE_VL_MAX_BITS);
>> +}
>> +
>>  #ifdef CONFIG_ARM64_SVE
>>    register_t compute_max_zcr(void);
>>  register_t vl_to_zcr(uint16_t vl);
>> +uint16_t get_sys_vl_len(void);
>>    #else /* !CONFIG_ARM64_SVE */
>>  @@ -30,6 +37,11 @@ static inline register_t vl_to_zcr(uint16_t vl)
>>      return 0;
>>  }
>>  +static inline uint16_t get_sys_vl_len(void)
>> +{
>> +    return 0;
>> +}
>> +
>>  #endif
>>    #endif /* _ARM_ARM64_SVE_H */
>> diff --git a/xen/arch/arm/include/asm/domain.h 
>> b/xen/arch/arm/include/asm/domain.h
>> index 42eb5df320a7..e4794a9fd2ab 100644
>> --- a/xen/arch/arm/include/asm/domain.h
>> +++ b/xen/arch/arm/include/asm/domain.h
>> @@ -31,6 +31,8 @@ enum domain_type {
>>    #define is_domain_direct_mapped(d) ((d)->cdf & CDF_directmap)
>>  +#define is_sve_domain(d) ((d)->arch.sve_vl_bits > 0)
>> +
>>  /*
>>   * Is the domain using the host memory layout?
>>   *
>> @@ -114,6 +116,9 @@ struct arch_domain
>>      void *tee;
>>  #endif
>>  +    /* max SVE vector length in bits */
>> +    uint16_t sve_vl_bits;
>> +
>>  }  __cacheline_aligned;
>>    struct arch_vcpu
>> @@ -190,6 +195,7 @@ struct arch_vcpu
>>      register_t tpidrro_el0;
>>        /* HYP configuration */
>> +    register_t zcr_el2;
>>      register_t cptr_el2;
>>      register_t hcr_el2;
>>      register_t mdcr_el2;
>> diff --git a/xen/include/public/arch-arm.h b/xen/include/public/arch-arm.h
>> index 1528ced5097a..e18a075105f0 100644
>> --- a/xen/include/public/arch-arm.h
>> +++ b/xen/include/public/arch-arm.h
>> @@ -304,6 +304,8 @@ struct xen_arch_domainconfig {
>>      uint16_t tee_type;
>>      /* IN */
>>      uint32_t nr_spis;
>> +    /* IN */
>> +    uint16_t sve_vl_bits;
> 
> Please spell out the padding clearly (even though I know there is one in this 
> structure that is not mentioned).

Ok I will

> 
>>      /*
>>       * OUT
>>       * Based on the property clock-frequency in the DT timer node.
>> diff --git a/xen/include/public/domctl.h b/xen/include/public/domctl.h
>> index 51be28c3de7c..616d7a1c070d 100644
>> --- a/xen/include/public/domctl.h
>> +++ b/xen/include/public/domctl.h
>> @@ -21,7 +21,7 @@
>>  #include "hvm/save.h"
>>  #include "memory.h"
>>  -#define XEN_DOMCTL_INTERFACE_VERSION 0x00000015
>> +#define XEN_DOMCTL_INTERFACE_VERSION 0x00000016
>>    /*
>>   * NB. xen_domctl.domain is an IN/OUT parameter for this operation.
> 
> Cheers,
> 
> -- 
> Julien Grall