[RFC PATCH v3 03/29] KVM: arm64: Introduce struct id_reg_info

Reiji Watanabe reijiw at google.com
Thu Nov 18 20:47:53 PST 2021


Hi Eric,

On Thu, Nov 18, 2021 at 12:36 PM Eric Auger <eauger at redhat.com> wrote:
>
> Hi Reiji,
>
> On 11/17/21 7:43 AM, Reiji Watanabe wrote:
> > This patch lays the groundwork to make ID registers writable.
> >
> > Introduce struct id_reg_info for an ID register to manage the
> > register specific control of its value for the guest, and provide set
> > of functions commonly used for ID registers to make them writable.
> >
> > The id_reg_info is used to do register specific initialization,
> > validation of the ID register and etc.  Not all ID registers must
> > have the id_reg_info. ID registers that don't have the id_reg_info
> > are handled in a common way that is applied to all ID registers.
> >
> > At present, changing an ID register from userspace is allowed only
> > if the ID register has the id_reg_info, but that will be changed
> > by the following patches.
> >
> > No ID register has the structure yet and the following patches
> > will add the id_reg_info for some ID registers.
> >
> > Signed-off-by: Reiji Watanabe <reijiw at google.com>
> > ---
> >  arch/arm64/include/asm/sysreg.h |   1 +
> >  arch/arm64/kvm/sys_regs.c       | 226 ++++++++++++++++++++++++++++++--
> >  2 files changed, 218 insertions(+), 9 deletions(-)
> >
> > diff --git a/arch/arm64/include/asm/sysreg.h b/arch/arm64/include/asm/sysreg.h
> > index 16b3f1a1d468..597609f26331 100644
> > --- a/arch/arm64/include/asm/sysreg.h
> > +++ b/arch/arm64/include/asm/sysreg.h
> > @@ -1197,6 +1197,7 @@
> >  #define ICH_VTR_TDS_MASK     (1 << ICH_VTR_TDS_SHIFT)
> >
> >  #define ARM64_FEATURE_FIELD_BITS     4
> > +#define ARM64_FEATURE_FIELD_MASK     ((1ull << ARM64_FEATURE_FIELD_BITS) - 1)
> >
> >  /* Create a mask for the feature bits of the specified feature. */
> >  #define ARM64_FEATURE_MASK(x)        (GENMASK_ULL(x##_SHIFT + ARM64_FEATURE_FIELD_BITS - 1, x##_SHIFT))
> > diff --git a/arch/arm64/kvm/sys_regs.c b/arch/arm64/kvm/sys_regs.c
> > index 5608d3410660..1552cd5581b7 100644
> > --- a/arch/arm64/kvm/sys_regs.c
> > +++ b/arch/arm64/kvm/sys_regs.c
> > @@ -265,6 +265,181 @@ static bool trap_raz_wi(struct kvm_vcpu *vcpu,
> >               return read_zero(vcpu, p);
> >  }
> >
> > +/*
> > + * A value for FCT_LOWER_SAFE must be zero and changing that will affect
> > + * ftr_check_types of id_reg_info.
> > + */
> > +enum feature_check_type {
> > +     FCT_LOWER_SAFE = 0,
> > +     FCT_HIGHER_SAFE,
> > +     FCT_HIGHER_OR_ZERO_SAFE,
> > +     FCT_EXACT,
> > +     FCT_EXACT_OR_ZERO_SAFE,
> > +     FCT_IGNORE,     /* Don't check (any value is fine) */
> Maybe you can remove the _SAFE suffix (EXACT does not have it).

I am inclined to keep 'SAFE' (otherwise, I am likely to forget
if lower is safe or not).

> s/EXACT/EQUAL ?

I will fix that FCT_EXACT to FCT_EQUAL_SAFE.

> > +};
> > +
> > +static int arm64_check_feature_one(enum feature_check_type type, int val,
> > +                                int limit)
> > +{
> > +     bool is_safe = false;
> > +
> > +     if (val == limit)
> > +             return 0;
> even if the type is unexpected?

I will remove it.

> > +
> > +     switch (type) {
> > +     case FCT_LOWER_SAFE:
> > +             is_safe = (val <= limit);
> > +             break;
> > +     case FCT_HIGHER_OR_ZERO_SAFE:
> > +             if (val == 0) {
> > +                     is_safe = true;
> > +                     break;
> > +             }
> > +             fallthrough;
> > +     case FCT_HIGHER_SAFE:
> > +             is_safe = (val >= limit);
> > +             break;
> > +     case FCT_EXACT:
> > +             break;
> > +     case FCT_EXACT_OR_ZERO_SAFE:
> > +             is_safe = (val == 0);
> > +             break;
> > +     case FCT_IGNORE:
> > +             is_safe = true;
> > +             break;
> > +     default:
> > +             WARN_ONCE(1, "Unexpected feature_check_type (%d)\n", type);
> > +             break;
> > +     }
> > +
> > +     return is_safe ? 0 : -1;
> > +}
> > +
> > +#define      FCT_TYPE_MASK           0x7
> > +#define      FCT_TYPE_SHIFT          1
> > +#define      FCT_SIGN_MASK           0x1
> > +#define      FCT_SIGN_SHIFT          0
> > +#define      FCT_TYPE(val)   ((val >> FCT_TYPE_SHIFT) & FCT_TYPE_MASK)
> > +#define      FCT_SIGN(val)   ((val >> FCT_SIGN_SHIFT) & FCT_SIGN_MASK)
> > +
> > +#define      MAKE_FCT(shift, type, sign)                             \
> > +     ((u64)((((type) & FCT_TYPE_MASK) << FCT_TYPE_SHIFT) |   \
> > +            (((sign) & FCT_SIGN_MASK) << FCT_SIGN_SHIFT)) << (shift))
> > +
> > +/* For signed field */
> > +#define      S_FCT(shift, type)      MAKE_FCT(shift, type, 1)
> > +/* For unigned field */
> > +#define      U_FCT(shift, type)      MAKE_FCT(shift, type, 0)
> > +
> > +/*
> > + * @val and @lim are both a value of the ID register. The function checks
> > + * if all features indicated in @val can be supported for guests on the host,
> > + * which supports features indicated in @lim. @check_types indicates how> + * features in the ID register needs to be checked.
> > + * See comments for id_reg_info's ftr_check_types field for more detail.
> What about RES0 fields which may exist? add a comment to reassure about
> the fact they are properly handled if there are?

Any fields including RES0 should be checked based on check_types.
I will explicitly state that in the comment.

> > + */
> > +static int arm64_check_features(u64 check_types, u64 val, u64 lim)
> > +{
> > +     int i;
> > +
> > +     for (i = 0; i < 64; i += ARM64_FEATURE_FIELD_BITS) {
> > +             u8 ftr_check = (check_types >> i) & ARM64_FEATURE_FIELD_MASK;
> > +             bool is_sign = FCT_SIGN(ftr_check);
> > +             enum feature_check_type fctype = FCT_TYPE(ftr_check);
> > +             int fval, flim, ret;
> > +
> > +             fval = cpuid_feature_extract_field(val, i, is_sign);
> > +             flim = cpuid_feature_extract_field(lim, i, is_sign);
> > +
> > +             ret = arm64_check_feature_one(fctype, fval, flim);
> > +             if (ret)
> > +                     return -E2BIG;
> > +     }
> > +     return 0;
> > +}
> > +
> > +struct id_reg_info {
> > +     u32     sys_reg;        /* Register ID */
> use struct kernel-doc comments instead?
> > +
> > +     /*
> > +      * Limit value of the register for a vcpu. The value is the sanitized
> > +      * system value with bits cleared for unsupported features for the
> > +      * guest.
> > +      */
> > +     u64     vcpu_limit_val;
> > +
> > +     /*
> > +      * The ftr_check_types is comprised of a set of 4 bits fields.
> nit: s/bits field/bit field here and below

I will fix them.

> > +      * Each 4 bits field is for a feature indicated by the same bits
> > +      * field of the ID register and indicates how the feature support
> > +      * for guests needs to be checked.
> > +      * The bit 0 indicates that the corresponding ID register field
> > +      * is signed(1) or unsigned(0).
> > +      * The bits [3:1] hold feature_check_type for the field.
> > +      * If all zero, all features in the ID register are treated as unsigned
> > +      * fields and checked based on Principles of the ID scheme for fields
> > +      * in ID registers (FCT_LOWER_SAFE of feature_check_type).
> values set by the guest are checked against host ID field values
> according to FCT_LOWER_SAFE test? You do not actually explicitly explain
> what the check is about although this may be obvious for you?

How about this ?

        /*
         * The ftr_check_types is comprised of a set of 4 bit fields.
         * Each 4 bit field is for a feature indicated by the same bit field
         * of the ID register and indicates how the field needs to be checked
         * (by arm64_check_feature_one) against the host's ID field when
         * userspace tries to set the register.
         * The bit 0 indicates that the corresponding ID register field is
         * signed(1) or unsigned(0). The bits [3:1] hold feature_check_type
         * for the field (FCT_LOWER_SAFE == 0, etc).
         * e.g. for ID_AA64PFR0_EL1.SVE(bits [35:32]), bits[35:32] of
         * ftr_check_types for the register should be 0. It means the SVE
         * field is treated as an unsigned field, and userspace can set the
         * field to a equal or lower value than the host's ID field value.
         */

> > +      */
> > +     u64     ftr_check_types;
> > +
> > +     /* Initialization function of the id_reg_info */
> > +     void (*init)(struct id_reg_info *id_reg);
> > +
> > +     /* Register specific validation function */
> validation callback? it does not register anything. We have check
> customization means already in ftr_check_types so it is difficult to
> guess at that point why this cb is needed, all the more so it applies
> after the ftr_checks.

I am going to add the following comment. Does it look clear enough for you ?

        /*
         * This is an optional ID register specific validation function.
         * When userspace tries to set the ID register, arm64_check_features()
         * will check if the requested value indicates any features that cannot
         * be supported by KVM on the host.  But, some ID register fields need
         * a special checking and this function can be used for such fields.
         * e.g. KVM_CREATE_DEVICE must be used to configure GICv3 for a guest.
         * ID_AA64PFR0_EL1.GIC shouldn't be set to 1 unless GICv3 is configured.
         * The validation function for ID_AA64PFR0_EL1 could be used to check
         * the field is consistent with GICv3 configuration.
         */

> > +     int (*validate)(struct kvm_vcpu *vcpu, const struct id_reg_info *id_reg,
> > +                     u64 val);
> > +
> > +     /* Return the reset value of the register for the vCPU */
> > +     u64 (*get_reset_val)(struct kvm_vcpu *vcpu,
> > +                          const struct id_reg_info *id_reg);
> > +};
> > +
> > +static void id_reg_info_init(struct id_reg_info *id_reg)
> > +{
> > +     id_reg->vcpu_limit_val = read_sanitised_ftr_reg(id_reg->sys_reg);
> > +     if (id_reg->init)
> > +             id_reg->init(id_reg);
> > +}
> > +
> > +/*
> > + * An ID register that needs special handling to control the value for the
> > + * guest must have its own id_reg_info in id_reg_info_table.
> > + * (i.e. the reset value is different from the host's sanitized value,
> > + * the value is affected by opt-in features, some fields needs specific
> s/needs/need

I will fix it.

Thank you for your review !

Regards
Reiji



More information about the linux-arm-kernel mailing list