Re: [PATCH v10 2/5] KVM: arm64: Use per guest ID register for ID_AA64PFR0_EL1.[CSV2|CSV3]

[Marc Zyngier <maz@kernel.org>]

On Mon, 22 May 2023 23:18:32 +0100,
Jing Zhang <jingzhangos@google.com> wrote:
> 
> With per guest ID registers, ID_AA64PFR0_EL1.[CSV2|CSV3] settings from
> userspace can be stored in its corresponding ID register.
> 
> The setting of CSV bits for protected VMs are removed according to the
> discussion from Fuad below:
> https://lore.kernel.org/all/CA+EHjTwXA9TprX4jeG+-D+c8v9XG+oFdU1o6TSkvVye145_OvA@mail.gmail.com
> 
> Besides the removal of CSV bits setting for protected VMs, No other
> functional change intended.

One thing that you don't mention is the addition of some locking,
which is a pretty significant change.

> 
> Signed-off-by: Jing Zhang <jingzhangos@google.com>
> ---
>  arch/arm64/include/asm/kvm_host.h |  2 --
>  arch/arm64/kvm/arm.c              | 17 ---------
>  arch/arm64/kvm/sys_regs.c         | 58 +++++++++++++++++++++++++------
>  3 files changed, 47 insertions(+), 30 deletions(-)
> 
> diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/include/asm/kvm_host.h
> index 069606170c82..8a2fde6c04c4 100644
> --- a/arch/arm64/include/asm/kvm_host.h
> +++ b/arch/arm64/include/asm/kvm_host.h
> @@ -257,8 +257,6 @@ struct kvm_arch {
>  
>  	cpumask_var_t supported_cpus;
>  
> -	u8 pfr0_csv2;
> -	u8 pfr0_csv3;
>  	struct {
>  		u8 imp:4;
>  		u8 unimp:4;
> diff --git a/arch/arm64/kvm/arm.c b/arch/arm64/kvm/arm.c
> index 774656a0718d..5114521ace60 100644
> --- a/arch/arm64/kvm/arm.c
> +++ b/arch/arm64/kvm/arm.c
> @@ -102,22 +102,6 @@ static int kvm_arm_default_max_vcpus(void)
>  	return vgic_present ? kvm_vgic_get_max_vcpus() : KVM_MAX_VCPUS;
>  }
>  
> -static void set_default_spectre(struct kvm *kvm)
> -{
> -	/*
> -	 * The default is to expose CSV2 == 1 if the HW isn't affected.
> -	 * Although this is a per-CPU feature, we make it global because
> -	 * asymmetric systems are just a nuisance.
> -	 *
> -	 * Userspace can override this as long as it doesn't promise
> -	 * the impossible.
> -	 */
> -	if (arm64_get_spectre_v2_state() == SPECTRE_UNAFFECTED)
> -		kvm->arch.pfr0_csv2 = 1;
> -	if (arm64_get_meltdown_state() == SPECTRE_UNAFFECTED)
> -		kvm->arch.pfr0_csv3 = 1;
> -}
> -
>  /**
>   * kvm_arch_init_vm - initializes a VM data structure
>   * @kvm:	pointer to the KVM struct
> @@ -161,7 +145,6 @@ int kvm_arch_init_vm(struct kvm *kvm, unsigned long type)
>  	/* The maximum number of VCPUs is limited by the host's GIC model */
>  	kvm->max_vcpus = kvm_arm_default_max_vcpus();
>  
> -	set_default_spectre(kvm);
>  	kvm_arm_init_hypercalls(kvm);
>  	kvm_arm_init_id_regs(kvm);
>  
> diff --git a/arch/arm64/kvm/sys_regs.c b/arch/arm64/kvm/sys_regs.c
> index d2ee3a1c7f03..9fb1c2f8f5a5 100644
> --- a/arch/arm64/kvm/sys_regs.c
> +++ b/arch/arm64/kvm/sys_regs.c
> @@ -1218,10 +1218,6 @@ static u64 kvm_arm_read_id_reg(const struct kvm_vcpu *vcpu, u32 id)
>  		if (!vcpu_has_sve(vcpu))
>  			val &= ~ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_SVE);
>  		val &= ~ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_AMU);
> -		val &= ~ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV2);
> -		val |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV2), (u64)vcpu->kvm->arch.pfr0_csv2);
> -		val &= ~ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV3);
> -		val |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV3), (u64)vcpu->kvm->arch.pfr0_csv3);
>  		if (kvm_vgic_global_state.type == VGIC_V3) {
>  			val &= ~ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_GIC);
>  			val |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_GIC), 1);
> @@ -1359,7 +1355,11 @@ static int set_id_aa64pfr0_el1(struct kvm_vcpu *vcpu,
>  			       const struct sys_reg_desc *rd,
>  			       u64 val)
>  {
> +	struct kvm_arch *arch = &vcpu->kvm->arch;

The use of kvm_arch as an anchor is very non-idiomatic. Use the kvm
pointer for this if you must, but I'd rather you spell the whole thing
out.

> +	u64 old_val = read_id_reg(vcpu, rd);
> +	u64 new_val = val;
>  	u8 csv2, csv3;
> +	int ret = 0;
>  
>  	/*
>  	 * Allow AA64PFR0_EL1.CSV2 to be set from userspace as long as
> @@ -1377,17 +1377,26 @@ static int set_id_aa64pfr0_el1(struct kvm_vcpu *vcpu,
>  	    (csv3 && arm64_get_meltdown_state() != SPECTRE_UNAFFECTED))
>  		return -EINVAL;
>  
> +	mutex_lock(&arch->config_lock);
>  	/* We can only differ with CSV[23], and anything else is an error */
> -	val ^= read_id_reg(vcpu, rd);
> +	val ^= old_val;
>  	val &= ~(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV2) |
>  		 ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV3));
> -	if (val)
> -		return -EINVAL;
> -
> -	vcpu->kvm->arch.pfr0_csv2 = csv2;
> -	vcpu->kvm->arch.pfr0_csv3 = csv3;
> +	if (val) {
> +		ret = -EINVAL;
> +		goto out;
> +	}
>  
> -	return 0;
> +	/* Only allow userspace to change the idregs before VM running */
> +	if (kvm_vm_has_ran_once(vcpu->kvm)) {
> +		if (new_val != old_val)
> +			ret = -EBUSY;

This sort of check should be done exactly once in a central spot. For
similar reasons, the config_lock should be take in a unique location
so that we can actually reason about this globally rather than at a
microscopic level.

Something like this (which applies to the full series):

diff --git a/arch/arm64/kvm/sys_regs.c b/arch/arm64/kvm/sys_regs.c
index b3eacfc592eb..e184b9350166 100644
--- a/arch/arm64/kvm/sys_regs.c
+++ b/arch/arm64/kvm/sys_regs.c
@@ -1534,7 +1534,6 @@ static int set_id_aa64dfr0_el1(struct kvm_vcpu *vcpu,
 			       const struct sys_reg_desc *rd,
 			       u64 val)
 {
-	struct kvm_arch *arch = &vcpu->kvm->arch;
 	u8 pmuver, host_pmuver;
 	bool valid_pmu;
 	int ret = 0;
@@ -1557,14 +1556,6 @@ static int set_id_aa64dfr0_el1(struct kvm_vcpu *vcpu,
 	if (kvm_vcpu_has_pmu(vcpu) != valid_pmu)
 		return -EINVAL;
 
-	mutex_lock(&arch->config_lock);
-	/* Only allow userspace to change the idregs before VM running */
-	if (kvm_vm_has_ran_once(vcpu->kvm)) {
-		if (val != read_id_reg(vcpu, rd))
-			ret = -EBUSY;
-		goto out;
-	}
-
 	if (!valid_pmu) {
 		/*
 		 * Ignore the PMUVer field in @val. The PMUVer would be determined
@@ -1592,7 +1583,6 @@ static int set_id_aa64dfr0_el1(struct kvm_vcpu *vcpu,
 			   pmuver == ID_AA64DFR0_EL1_PMUVer_IMP_DEF);
 
 out:
-	mutex_unlock(&arch->config_lock);
 	return ret;
 }
 
@@ -1617,7 +1607,6 @@ static int set_id_dfr0_el1(struct kvm_vcpu *vcpu,
 			   const struct sys_reg_desc *rd,
 			   u64 val)
 {
-	struct kvm_arch *arch = &vcpu->kvm->arch;
 	u8 perfmon, host_perfmon;
 	bool valid_pmu;
 	int ret = 0;
@@ -1641,14 +1630,6 @@ static int set_id_dfr0_el1(struct kvm_vcpu *vcpu,
 	if (kvm_vcpu_has_pmu(vcpu) != valid_pmu)
 		return -EINVAL;
 
-	mutex_lock(&arch->config_lock);
-	/* Only allow userspace to change the idregs before VM running */
-	if (kvm_vm_has_ran_once(vcpu->kvm)) {
-		if (val != read_id_reg(vcpu, rd))
-			ret = -EBUSY;
-		goto out;
-	}
-
 	if (!valid_pmu) {
 		/*
 		 * Ignore the PerfMon field in @val. The PerfMon would be determined
@@ -1676,7 +1657,6 @@ static int set_id_dfr0_el1(struct kvm_vcpu *vcpu,
 			   perfmon == ID_DFR0_EL1_PerfMon_IMPDEF);
 
 out:
-	mutex_unlock(&arch->config_lock);
 	return ret;
 }
 
@@ -1690,11 +1670,7 @@ static int set_id_dfr0_el1(struct kvm_vcpu *vcpu,
 static int get_id_reg(struct kvm_vcpu *vcpu, const struct sys_reg_desc *rd,
 		      u64 *val)
 {
-	struct kvm_arch *arch = &vcpu->kvm->arch;
-
-	mutex_lock(&arch->config_lock);
 	*val = read_id_reg(vcpu, rd);
-	mutex_unlock(&arch->config_lock);
 
 	return 0;
 }
@@ -1702,21 +1678,12 @@ static int get_id_reg(struct kvm_vcpu *vcpu, const struct sys_reg_desc *rd,
 static int set_id_reg(struct kvm_vcpu *vcpu, const struct sys_reg_desc *rd,
 		      u64 val)
 {
-	struct kvm_arch *arch = &vcpu->kvm->arch;
 	u32 id = reg_to_encoding(rd);
 	int ret = 0;
 
-	mutex_lock(&arch->config_lock);
-	/* Only allow userspace to change the idregs before VM running */
-	if (kvm_vm_has_ran_once(vcpu->kvm)) {
-		if (val != read_id_reg(vcpu, rd))
-			ret = -EBUSY;
-	} else {
-		ret = arm64_check_features(vcpu, rd, val);
-		if (!ret)
-			IDREG(vcpu->kvm, id) = val;
-	}
-	mutex_unlock(&arch->config_lock);
+	ret = arm64_check_features(vcpu, rd, val);
+	if (!ret)
+		IDREG(vcpu->kvm, id) = val;
 
 	return ret;
 }
@@ -3438,6 +3405,9 @@ int kvm_sys_reg_get_user(struct kvm_vcpu *vcpu, const struct kvm_one_reg *reg,
 	if (!r || sysreg_hidden_user(vcpu, r))
 		return -ENOENT;
 
+	if (is_id_reg(reg_to_encoding(r)))
+		mutex_lock(&vcpu->kvm->arch.config_lock);
+
 	if (r->get_user) {
 		ret = (r->get_user)(vcpu, r, &val);
 	} else {
@@ -3445,6 +3415,9 @@ int kvm_sys_reg_get_user(struct kvm_vcpu *vcpu, const struct kvm_one_reg *reg,
 		ret = 0;
 	}
 
+	if (is_id_reg(reg_to_encoding(r)))
+		mutex_unlock(&vcpu->kvm->arch.config_lock);
+
 	if (!ret)
 		ret = put_user(val, uaddr);
 
@@ -3482,9 +3455,21 @@ int kvm_sys_reg_set_user(struct kvm_vcpu *vcpu, const struct kvm_one_reg *reg,
 	if (!r || sysreg_hidden_user(vcpu, r))
 		return -ENOENT;
 
+	/* Only allow userspace to change the idregs before VM running */
+	if (is_id_reg(reg_to_encoding(r)) &&
+	    kvm_vm_has_ran_once(vcpu->kvm)) {
+		if (val == read_id_reg(vcpu, r))
+			return 0;
+		return -EBUSY;
+	}
+
 	if (sysreg_user_write_ignore(vcpu, r))
 		return 0;
 
+	/* ID regs are global to the VM and cannot be updated concurrently */
+	if (is_id_reg(reg_to_encoding(r)))
+		mutex_lock(&vcpu->kvm->arch.config_lock);
+
 	if (r->set_user) {
 		ret = (r->set_user)(vcpu, r, val);
 	} else {
@@ -3492,6 +3477,9 @@ int kvm_sys_reg_set_user(struct kvm_vcpu *vcpu, const struct kvm_one_reg *reg,
 		ret = 0;
 	}
 
+	if (is_id_reg(reg_to_encoding(r)))
+		mutex_unlock(&vcpu->kvm->arch.config_lock);
+
 	return ret;
 }

and you can then restore the code to its original shape, as there is
no need to change the control flow anymore.

> +	} else {
> +		IDREG(vcpu->kvm, reg_to_encoding(rd)) = new_val;
> +	}
> +out:
> +	mutex_unlock(&arch->config_lock);
> +	return ret;
>  }
>  
>  static int set_id_aa64dfr0_el1(struct kvm_vcpu *vcpu,
> @@ -1479,7 +1488,12 @@ static int set_id_dfr0_el1(struct kvm_vcpu *vcpu,
>  static int get_id_reg(struct kvm_vcpu *vcpu, const struct sys_reg_desc *rd,
>  		      u64 *val)
>  {

Right above this function is a comment that says the idreg are
immutable. Time to revisit it?

> +	struct kvm_arch *arch = &vcpu->kvm->arch;
> +
> +	mutex_lock(&arch->config_lock);
>  	*val = read_id_reg(vcpu, rd);
> +	mutex_unlock(&arch->config_lock);
> +
>  	return 0;
>  }
>  
> @@ -3364,6 +3378,7 @@ void kvm_arm_init_id_regs(struct kvm *kvm)
>  {
>  	const struct sys_reg_desc *idreg;
>  	struct sys_reg_params params;
> +	u64 val;
>  	u32 id;
>  
>  	/* Find the first idreg (SYS_ID_PFR0_EL1) in sys_reg_descs. */
> @@ -3386,6 +3401,27 @@ void kvm_arm_init_id_regs(struct kvm *kvm)
>  		idreg++;
>  		id = reg_to_encoding(idreg);
>  	}
> +
> +	/*
> +	 * The default is to expose CSV2 == 1 if the HW isn't affected.
> +	 * Although this is a per-CPU feature, we make it global because
> +	 * asymmetric systems are just a nuisance.
> +	 *
> +	 * Userspace can override this as long as it doesn't promise
> +	 * the impossible.
> +	 */
> +	val = IDREG(kvm, SYS_ID_AA64PFR0_EL1);
> +
> +	if (arm64_get_spectre_v2_state() == SPECTRE_UNAFFECTED) {
> +		val &= ~ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV2);
> +		val |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV2), 1);
> +	}
> +	if (arm64_get_meltdown_state() == SPECTRE_UNAFFECTED) {
> +		val &= ~ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV3);
> +		val |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV3), 1);
> +	}
> +
> +	IDREG(kvm, SYS_ID_AA64PFR0_EL1) = val;
>  }
>  
>  int __init kvm_sys_reg_table_init(void)

Thanks,

	M.

-- 
Without deviation from the norm, progress is not possible.