[PATCH v2 0/8] KVM: arm64: Prefix of fixes from HAFDBS series (for 7.4)

[PATCH v2 0/8] KVM: arm64: Prefix of fixes from HAFDBS series (for 7.4)

[Oliver Upton]

v1: https://lore.kernel.org/kvmarm/20261104192124.3932743-1-oupton@kernel.org/

v1 -> v2:
 - Hide KVM_CAP_ARM_EL2 when provided a VM with MTE enabled (Sashiko)
 - Document MTE and EL2 as mutually exclusive
 - Pick up Leo's R-b (thanks!)

Oliver Upton (8):
  KVM: arm64: nv: Disallow KVM_CAP_ARM_MTE + FEAT_EL2
  KVM: arm64: nv: Add CGT description for FEAT_PoPS instructions
  KVM: arm64: Separate user / guest visibility flags
  KVM: arm64: nv: Treat the VNCR GICv3 registers as hidden from user
  KVM: arm64: Use ->visibility() to filter MTE CMO instructions
  KVM: arm64: Add sys_reg_descs for remaining MTE CMO instructions
  KVM: arm64: nv: Enforce sysreg visibility before evaulating guest
    traps
  KVM: arm64: Stop leaning on __kvm_inject_sea() for abort injection

 Documentation/virt/kvm/api.rst   |   1 +
 arch/arm64/kvm/arm.c             |  16 ++--
 arch/arm64/kvm/emulate-nested.c  |  26 +++++-
 arch/arm64/kvm/inject_fault.c    | 133 ++++++++++++++++++-------------
 arch/arm64/kvm/sys_regs.c        |  84 +++++++++++--------
 arch/arm64/kvm/sys_regs.h        |  27 +++++--
 arch/arm64/kvm/vgic-sys-reg-v3.c |   2 +-
 arch/arm64/kvm/vgic-sys-reg-v5.c |   2 +-
 8 files changed, 183 insertions(+), 108 deletions(-)


base-commit: 6628e7d4c6b25c8d18f1345e3a20d21829aae7e4
-- 
2.47.3

[PATCH v2 1/8] KVM: arm64: nv: Disallow KVM_CAP_ARM_MTE + FEAT_EL2

[Oliver Upton]

KVM's nested implementation already hides FEAT_MTE and its laundry list
of derivatives from an NV guest. While good, this is insonsistent with
the UAPI which does not explicitly prevent the combination of MTE and
NV.

Bite the bullet and refuse to create a nested VM if KVM_CAP_ARM_MTE has
been enabled.

Cc: stable@vger.kernel.org
Fixes: a7484c80e5ca ("KVM: arm64: Allow userspace to request KVM_ARM_VCPU_EL2*")
Signed-off-by: Oliver Upton <oupton@kernel.org>
---
 Documentation/virt/kvm/api.rst |  1 +
 arch/arm64/kvm/arm.c           | 16 +++++++++++-----
 2 files changed, 12 insertions(+), 5 deletions(-)

diff --git a/Documentation/virt/kvm/api.rst b/Documentation/virt/kvm/api.rst
index e95c9cda085d..4ce4166d415f 100644
--- a/Documentation/virt/kvm/api.rst
+++ b/Documentation/virt/kvm/api.rst
@@ -3602,6 +3602,7 @@ Possible features:
 	- KVM_ARM_VCPU_HAS_EL2: Enable Nested Virtualisation support,
 	  booting the guest from EL2 instead of EL1.
 	  Depends on KVM_CAP_ARM_EL2.
+          Cannot be enabled if KVM_CAP_ARM_MTE has been enabled for the VM.
 	  The VM is running with HCR_EL2.E2H being RES1 (VHE) unless
 	  KVM_ARM_VCPU_HAS_EL2_E2H0 is also set.
 
diff --git a/arch/arm64/kvm/arm.c b/arch/arm64/kvm/arm.c
index 157bf4b81202..4108ff30830c 100644
--- a/arch/arm64/kvm/arm.c
+++ b/arch/arm64/kvm/arm.c
@@ -480,7 +480,11 @@ int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext)
 		r = cpus_have_final_cap(ARM64_HAS_32BIT_EL1);
 		break;
 	case KVM_CAP_ARM_EL2:
-		r = cpus_have_final_cap(ARM64_HAS_NESTED_VIRT);
+		/* Nested virt is incompatible with MTE */
+		if (kvm && kvm_has_mte(kvm))
+			r = 0;
+		else
+			r = cpus_have_final_cap(ARM64_HAS_NESTED_VIRT);
 		break;
 	case KVM_CAP_ARM_EL2_E2H0:
 		r = cpus_have_final_cap(ARM64_HAS_HCR_NV1);
@@ -1793,13 +1797,15 @@ static int kvm_vcpu_init_check_features(struct kvm_vcpu *vcpu,
 	    !test_bit(KVM_ARM_VCPU_PMU_V3, &features))
 		return -EINVAL;
 
+	/* MTE is incompatible with AArch32 and nested virt */
+	if (kvm_has_mte(vcpu->kvm) &&
+	    (test_bit(KVM_ARM_VCPU_EL1_32BIT, &features) ||
+	     test_bit(KVM_ARM_VCPU_HAS_EL2, &features)))
+		return -EINVAL;
+
 	if (!test_bit(KVM_ARM_VCPU_EL1_32BIT, &features))
 		return 0;
 
-	/* MTE is incompatible with AArch32 */
-	if (kvm_has_mte(vcpu->kvm))
-		return -EINVAL;
-
 	/* NV is incompatible with AArch32 */
 	if (test_bit(KVM_ARM_VCPU_HAS_EL2, &features))
 		return -EINVAL;
-- 
2.47.3

Re: [PATCH v2 1/8] KVM: arm64: nv: Disallow KVM_CAP_ARM_MTE + FEAT_EL2

[Wei-Lin Chang]

On Fri, Nov 06, 2026 at 10:31:47AM -0800, Oliver Upton wrote:
> KVM's nested implementation already hides FEAT_MTE and its laundry list
> of derivatives from an NV guest. While good, this is insonsistent with
> the UAPI which does not explicitly prevent the combination of MTE and
> NV.
> 
> Bite the bullet and refuse to create a nested VM if KVM_CAP_ARM_MTE has
> been enabled.
> 
> Cc: stable@vger.kernel.org
> Fixes: a7484c80e5ca ("KVM: arm64: Allow userspace to request KVM_ARM_VCPU_EL2*")
> Signed-off-by: Oliver Upton <oupton@kernel.org>
> ---
>  Documentation/virt/kvm/api.rst |  1 +
>  arch/arm64/kvm/arm.c           | 16 +++++++++++-----
>  2 files changed, 12 insertions(+), 5 deletions(-)
> 
> diff --git a/Documentation/virt/kvm/api.rst b/Documentation/virt/kvm/api.rst
> index e95c9cda085d..4ce4166d415f 100644
> --- a/Documentation/virt/kvm/api.rst
> +++ b/Documentation/virt/kvm/api.rst
> @@ -3602,6 +3602,7 @@ Possible features:
>  	- KVM_ARM_VCPU_HAS_EL2: Enable Nested Virtualisation support,
>  	  booting the guest from EL2 instead of EL1.
>  	  Depends on KVM_CAP_ARM_EL2.
> +          Cannot be enabled if KVM_CAP_ARM_MTE has been enabled for the VM.
>  	  The VM is running with HCR_EL2.E2H being RES1 (VHE) unless
>  	  KVM_ARM_VCPU_HAS_EL2_E2H0 is also set.
>  
> diff --git a/arch/arm64/kvm/arm.c b/arch/arm64/kvm/arm.c
> index 157bf4b81202..4108ff30830c 100644
> --- a/arch/arm64/kvm/arm.c
> +++ b/arch/arm64/kvm/arm.c
> @@ -480,7 +480,11 @@ int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext)
>  		r = cpus_have_final_cap(ARM64_HAS_32BIT_EL1);
>  		break;
>  	case KVM_CAP_ARM_EL2:
> -		r = cpus_have_final_cap(ARM64_HAS_NESTED_VIRT);
> +		/* Nested virt is incompatible with MTE */
> +		if (kvm && kvm_has_mte(kvm))
> +			r = 0;
> +		else
> +			r = cpus_have_final_cap(ARM64_HAS_NESTED_VIRT);

More of a question, do we really want to consider VM feature
compatibility during KVM_CHECK_EXTENSION?

This duplicates the checks which is a bit error-prone. For example I
think we need to also update the KVM_CAP_ARM_EL1_32BIT case to check
kvm_has_mte() if we want to do this. Additionally the information
reported can still become stale if userspace changes it afterwards.

OTOH the KVM_CAP_ARM_MTE case becomes awkward, as its contract states it
needs to be enabled before creating vCPUs so it is strange for it to
check for whether there are vCPUs initialized with NV.

Thanks,
Wei-Lin Chang

>  		break;
>  	case KVM_CAP_ARM_EL2_E2H0:
>  		r = cpus_have_final_cap(ARM64_HAS_HCR_NV1);
> @@ -1793,13 +1797,15 @@ static int kvm_vcpu_init_check_features(struct kvm_vcpu *vcpu,
>  	    !test_bit(KVM_ARM_VCPU_PMU_V3, &features))
>  		return -EINVAL;
>  
> +	/* MTE is incompatible with AArch32 and nested virt */
> +	if (kvm_has_mte(vcpu->kvm) &&
> +	    (test_bit(KVM_ARM_VCPU_EL1_32BIT, &features) ||
> +	     test_bit(KVM_ARM_VCPU_HAS_EL2, &features)))
> +		return -EINVAL;
> +
>  	if (!test_bit(KVM_ARM_VCPU_EL1_32BIT, &features))
>  		return 0;
>  
> -	/* MTE is incompatible with AArch32 */
> -	if (kvm_has_mte(vcpu->kvm))
> -		return -EINVAL;
> -
>  	/* NV is incompatible with AArch32 */
>  	if (test_bit(KVM_ARM_VCPU_HAS_EL2, &features))
>  		return -EINVAL;
> -- 
> 2.47.3
> 

[PATCH v2 2/8] KVM: arm64: nv: Add CGT description for FEAT_PoPS instructions

[Oliver Upton]

The DC CIVAPS and DC CIGDVAPS instructions are both affected by
HCR_EL2.TPCP. Add the missing CGT descriptors.

Fixes: a7484c80e5ca ("KVM: arm64: Allow userspace to request KVM_ARM_VCPU_EL2*")
Reviewed-by: Leonardo Bras <leo.bras@arm.com>
Signed-off-by: Oliver Upton <oupton@kernel.org>
---
 arch/arm64/kvm/emulate-nested.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/arch/arm64/kvm/emulate-nested.c b/arch/arm64/kvm/emulate-nested.c
index 2bec81668361..d04ac5203454 100644
--- a/arch/arm64/kvm/emulate-nested.c
+++ b/arch/arm64/kvm/emulate-nested.c
@@ -747,6 +747,8 @@ static const struct encoding_to_trap_config encoding_to_cgt[] __initconst = {
 	SR_TRAP(SYS_DC_CGDVAP,		CGT_HCR_TPC),
 	SR_TRAP(SYS_DC_CGVADP,		CGT_HCR_TPC),
 	SR_TRAP(SYS_DC_CGDVADP,		CGT_HCR_TPC),
+	SR_TRAP(SYS_DC_CIVAPS,		CGT_HCR_TPC),
+	SR_TRAP(SYS_DC_CIGDVAPS,	CGT_HCR_TPC),
 	SR_TRAP(SYS_IC_IVAU,		CGT_HCR_TPU_TOCU),
 	SR_TRAP(SYS_IC_IALLU,		CGT_HCR_TPU_TOCU),
 	SR_TRAP(SYS_IC_IALLUIS,		CGT_HCR_TPU_TICAB),
-- 
2.47.3

[PATCH v2 3/8] KVM: arm64: Separate user / guest visibility flags

[Oliver Upton]

There's need for even more visibility hacks to differentiate between
userspace and guest. Decompose REG_HIDDEN into two parts and test for
the user-specific flag on all the appropriate UAPI surfaces.

Signed-off-by: Oliver Upton <oupton@kernel.org>
---
 arch/arm64/kvm/sys_regs.c        | 10 +++++-----
 arch/arm64/kvm/sys_regs.h        | 23 +++++++++++++++++------
 arch/arm64/kvm/vgic-sys-reg-v3.c |  2 +-
 arch/arm64/kvm/vgic-sys-reg-v5.c |  2 +-
 4 files changed, 24 insertions(+), 13 deletions(-)

diff --git a/arch/arm64/kvm/sys_regs.c b/arch/arm64/kvm/sys_regs.c
index 850eba167aad..870dcb25d0d7 100644
--- a/arch/arm64/kvm/sys_regs.c
+++ b/arch/arm64/kvm/sys_regs.c
@@ -4866,7 +4866,7 @@ static void perform_access(struct kvm_vcpu *vcpu,
 	trace_kvm_sys_access(*vcpu_pc(vcpu), params, r);
 
 	/* Check for regs disabled by runtime config */
-	if (sysreg_hidden(vcpu, r)) {
+	if (sysreg_guest_hidden(vcpu, r)) {
 		kvm_inject_undefined(vcpu);
 		return;
 	}
@@ -5490,7 +5490,7 @@ id_to_sys_reg_desc(struct kvm_vcpu *vcpu, u64 id,
 	r = get_reg_by_id(id, table, num);
 
 	/* Not saved in the sys_reg array and not otherwise accessible? */
-	if (r && (!(r->reg || r->get_user) || sysreg_hidden(vcpu, r)))
+	if (r && (!(r->reg || r->get_user) || sysreg_user_hidden(vcpu, r)))
 		r = NULL;
 
 	return r;
@@ -5571,7 +5571,7 @@ int kvm_sys_reg_get_user(struct kvm_vcpu *vcpu, const struct kvm_one_reg *reg,
 	int ret;
 
 	r = id_to_sys_reg_desc(vcpu, id, table, num);
-	if (!r || sysreg_hidden(vcpu, r))
+	if (!r || sysreg_user_hidden(vcpu, r))
 		return -ENOENT;
 
 	if (r->get_user) {
@@ -5611,7 +5611,7 @@ int kvm_sys_reg_set_user(struct kvm_vcpu *vcpu, const struct kvm_one_reg *reg,
 		return -EFAULT;
 
 	r = id_to_sys_reg_desc(vcpu, id, table, num);
-	if (!r || sysreg_hidden(vcpu, r))
+	if (!r || sysreg_user_hidden(vcpu, r))
 		return -ENOENT;
 
 	if (sysreg_user_write_ignore(vcpu, r))
@@ -5705,7 +5705,7 @@ static int walk_one_sys_reg(const struct kvm_vcpu *vcpu,
 	if (!(rd->reg || rd->get_user))
 		return 0;
 
-	if (sysreg_hidden(vcpu, rd))
+	if (sysreg_user_hidden(vcpu, rd))
 		return 0;
 
 	if (!copy_reg_to_user(rd, uind))
diff --git a/arch/arm64/kvm/sys_regs.h b/arch/arm64/kvm/sys_regs.h
index a3cccad2766f..50214271b24a 100644
--- a/arch/arm64/kvm/sys_regs.h
+++ b/arch/arm64/kvm/sys_regs.h
@@ -104,9 +104,14 @@ struct sys_reg_desc {
 				   const struct sys_reg_desc *rd);
 };
 
-#define REG_HIDDEN		(1 << 0) /* hidden from userspace and guest */
-#define REG_RAZ			(1 << 1) /* RAZ from userspace and guest */
-#define REG_USER_WI		(1 << 2) /* WI from userspace only */
+#define REG_GUEST_HIDDEN	BIT(0)
+#define REG_USER_HIDDEN		BIT(1)
+	/* Hidden from userspace and guest */
+#define REG_HIDDEN		(REG_GUEST_HIDDEN | REG_USER_HIDDEN)
+	/* RAZ from userspace and guest */
+#define REG_RAZ			BIT(2)
+	/* WI from userspace only */
+#define REG_USER_WI		BIT(3)
 
 static __printf(2, 3)
 inline void print_sys_reg_msg(const struct sys_reg_params *p,
@@ -168,10 +173,16 @@ static inline unsigned int sysreg_visibility(const struct kvm_vcpu *vcpu,
 	return r->visibility(vcpu, r);
 }
 
-static inline bool sysreg_hidden(const struct kvm_vcpu *vcpu,
-				 const struct sys_reg_desc *r)
+static inline bool sysreg_user_hidden(const struct kvm_vcpu *vcpu,
+				      const struct sys_reg_desc *r)
+{
+	return sysreg_visibility(vcpu, r) & REG_USER_HIDDEN;
+}
+
+static inline bool sysreg_guest_hidden(const struct kvm_vcpu *vcpu,
+				       const struct sys_reg_desc *r)
 {
-	return sysreg_visibility(vcpu, r) & REG_HIDDEN;
+	return sysreg_visibility(vcpu, r) & REG_GUEST_HIDDEN;
 }
 
 static inline bool sysreg_visible_as_raz(const struct kvm_vcpu *vcpu,
diff --git a/arch/arm64/kvm/vgic-sys-reg-v3.c b/arch/arm64/kvm/vgic-sys-reg-v3.c
index 89315336b9d3..f41c21ee9e89 100644
--- a/arch/arm64/kvm/vgic-sys-reg-v3.c
+++ b/arch/arm64/kvm/vgic-sys-reg-v3.c
@@ -465,7 +465,7 @@ int vgic_v3_has_cpu_sysregs_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *
 	r = get_reg_by_id(attr_to_id(attr->attr), gic_v3_icc_reg_descs,
 			  ARRAY_SIZE(gic_v3_icc_reg_descs));
 
-	if (r && !sysreg_hidden(vcpu, r))
+	if (r && !sysreg_user_hidden(vcpu, r))
 		return 0;
 
 	return -ENXIO;
diff --git a/arch/arm64/kvm/vgic-sys-reg-v5.c b/arch/arm64/kvm/vgic-sys-reg-v5.c
index 6b255eacbf60..e62d5e908f5a 100644
--- a/arch/arm64/kvm/vgic-sys-reg-v5.c
+++ b/arch/arm64/kvm/vgic-sys-reg-v5.c
@@ -495,7 +495,7 @@ int vgic_v5_has_cpu_sysregs_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *
 	r = get_reg_by_id(attr_to_id(attr->attr), gic_v5_icc_reg_descs,
 			  ARRAY_SIZE(gic_v5_icc_reg_descs));
 
-	if (r && !sysreg_hidden(vcpu, r))
+	if (r && !sysreg_user_hidden(vcpu, r))
 		return 0;
 
 	return -ENXIO;
-- 
2.47.3

[PATCH v2 4/8] KVM: arm64: nv: Treat the VNCR GICv3 registers as hidden from user

[Oliver Upton]

KVM's emulation of sysreg traps will soon need to evaluate the
visibility of a sysreg prior to the effects of the L1 hypervisor's traps
to cope with instructions that intersect with a CGT but lack a
corresponding FGU bit (I know, not enough TLAs).

The VNCR GICv3 registers are marked as hidden from both user/guest on
the basis that KVM never handles these traps locally and unconditinally
relies on VNCR redirection. But an L1 hypervisor could trap these.

Remap the visibility as being specific to user accesses, giving the
indication that guest accesses to the sysreg are legal at least from an
architecture POV. Note that the guest side of the visibility was a bit
belt-and-braces as the bad_vncr_trap() accessor will act as a backstop
against any unexpected local trap handling.

Signed-off-by: Oliver Upton <oupton@kernel.org>
---
 arch/arm64/kvm/sys_regs.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/arch/arm64/kvm/sys_regs.c b/arch/arm64/kvm/sys_regs.c
index 870dcb25d0d7..0b21b454242c 100644
--- a/arch/arm64/kvm/sys_regs.c
+++ b/arch/arm64/kvm/sys_regs.c
@@ -1020,10 +1020,10 @@ static u64 reset_mpidr(struct kvm_vcpu *vcpu, const struct sys_reg_desc *r)
 	return mpidr;
 }
 
-static unsigned int hidden_visibility(const struct kvm_vcpu *vcpu,
-				      const struct sys_reg_desc *r)
+static unsigned int hidden_user_visibility(const struct kvm_vcpu *vcpu,
+					   const struct sys_reg_desc *r)
 {
-	return REG_HIDDEN;
+	return REG_USER_HIDDEN;
 }
 
 static unsigned int pmu_visibility(const struct kvm_vcpu *vcpu,
@@ -2757,7 +2757,7 @@ static bool bad_redir_trap(struct kvm_vcpu *vcpu,
 #define EL2_REG_VNCR_FILT(name, vis)			\
 	EL2_REG_FILTERED(name, bad_vncr_trap, reset_val, 0, vis)
 #define EL2_REG_VNCR_GICv3(name)			\
-	EL2_REG_VNCR_FILT(name, hidden_visibility)
+	EL2_REG_VNCR_FILT(name, hidden_user_visibility)
 #define EL2_REG_REDIR(name, rst, v)	EL2_REG(name, bad_redir_trap, rst, v)
 
 #define TIMER_REG(name, vis)					   \
-- 
2.47.3

[PATCH v2 5/8] KVM: arm64: Use ->visibility() to filter MTE CMO instructions

[Oliver Upton]

More descriptors for MTE CMO instructions are on the way; prepare by
using sys_reg_desc::visibility() to hide instructions when MTE isn't
implemented.

Signed-off-by: Oliver Upton <oupton@kernel.org>
---
 arch/arm64/kvm/sys_regs.c | 30 +++++++++++++-----------------
 1 file changed, 13 insertions(+), 17 deletions(-)

diff --git a/arch/arm64/kvm/sys_regs.c b/arch/arm64/kvm/sys_regs.c
index 0b21b454242c..b45635471344 100644
--- a/arch/arm64/kvm/sys_regs.c
+++ b/arch/arm64/kvm/sys_regs.c
@@ -556,17 +556,6 @@ static bool access_dcsw(struct kvm_vcpu *vcpu,
 	return true;
 }
 
-static bool access_dcgsw(struct kvm_vcpu *vcpu,
-			 struct sys_reg_params *p,
-			 const struct sys_reg_desc *r)
-{
-	if (!kvm_has_mte(vcpu->kvm))
-		return undef_access(vcpu, p, r);
-
-	/* Treat MTE S/W ops as we treat the classic ones: with contempt */
-	return access_dcsw(vcpu, p, r);
-}
-
 static void get_access_mask(const struct sys_reg_desc *r, u64 *mask, u64 *shift)
 {
 	switch (r->aarch32_map) {
@@ -4306,10 +4295,17 @@ static bool handle_tlbi_el1(struct kvm_vcpu *vcpu, struct sys_reg_params *p,
 		.access = (access_fn),					\
 	}
 
+#define MTE_CMO_INSN(insn, access_fn)					\
+	{								\
+		SYS_DESC(SYS_##insn),					\
+		.access	= (access_fn),					\
+		.visibility = mte_visibility,				\
+	}
+
 static struct sys_reg_desc sys_insn_descs[] = {
 	{ SYS_DESC(SYS_DC_ISW), access_dcsw },
-	{ SYS_DESC(SYS_DC_IGSW), access_dcgsw },
-	{ SYS_DESC(SYS_DC_IGDSW), access_dcgsw },
+	MTE_CMO_INSN(DC_IGSW, access_dcsw),
+	MTE_CMO_INSN(DC_IGDSW, access_dcsw),
 
 	SYS_INSN(AT_S1E1R, handle_at_s1e01),
 	SYS_INSN(AT_S1E1W, handle_at_s1e01),
@@ -4320,11 +4316,11 @@ static struct sys_reg_desc sys_insn_descs[] = {
 	SYS_INSN(AT_S1E1A, handle_at_s1e01),
 
 	{ SYS_DESC(SYS_DC_CSW), access_dcsw },
-	{ SYS_DESC(SYS_DC_CGSW), access_dcgsw },
-	{ SYS_DESC(SYS_DC_CGDSW), access_dcgsw },
+	MTE_CMO_INSN(DC_CGSW, access_dcsw),
+	MTE_CMO_INSN(DC_CGDSW, access_dcsw),
 	{ SYS_DESC(SYS_DC_CISW), access_dcsw },
-	{ SYS_DESC(SYS_DC_CIGSW), access_dcgsw },
-	{ SYS_DESC(SYS_DC_CIGDSW), access_dcgsw },
+	MTE_CMO_INSN(DC_CIGSW, access_dcsw),
+	MTE_CMO_INSN(DC_CIGDSW, access_dcsw),
 
 	SYS_INSN(TLBI_VMALLE1OS, handle_tlbi_el1),
 	SYS_INSN(TLBI_VAE1OS, handle_tlbi_el1),
-- 
2.47.3

[PATCH v2 6/8] KVM: arm64: Add sys_reg_descs for remaining MTE CMO instructions

[Oliver Upton]

Filtering unsupported cache maintenance instructions from a guest relies
on either having a suitable FGU bit for the feature or a descriptor with
a visibility() hook.

MTE CMO instructions can be trapped by HCR_EL2.TPCP and lack a suitable
FGU. Add descriptors for the CMO instructions that UNDEF if MTE isn't
implemented for the VM.

Signed-off-by: Oliver Upton <oupton@kernel.org>
---
 arch/arm64/kvm/sys_regs.c | 18 ++++++++++++++++++
 1 file changed, 18 insertions(+)

diff --git a/arch/arm64/kvm/sys_regs.c b/arch/arm64/kvm/sys_regs.c
index b45635471344..7f7a197a9ad6 100644
--- a/arch/arm64/kvm/sys_regs.c
+++ b/arch/arm64/kvm/sys_regs.c
@@ -4295,6 +4295,12 @@ static bool handle_tlbi_el1(struct kvm_vcpu *vcpu, struct sys_reg_params *p,
 		.access = (access_fn),					\
 	}
 
+static bool bad_cmo_trap(struct kvm_vcpu *vcpu, struct sys_reg_params *p,
+			 const struct sys_reg_desc *r)
+{
+	return bad_trap(vcpu, p, r, "MTE CMO trap");
+}
+
 #define MTE_CMO_INSN(insn, access_fn)					\
 	{								\
 		SYS_DESC(SYS_##insn),					\
@@ -4304,7 +4310,9 @@ static bool handle_tlbi_el1(struct kvm_vcpu *vcpu, struct sys_reg_params *p,
 
 static struct sys_reg_desc sys_insn_descs[] = {
 	{ SYS_DESC(SYS_DC_ISW), access_dcsw },
+	MTE_CMO_INSN(DC_IGVAC, bad_cmo_trap),
 	MTE_CMO_INSN(DC_IGSW, access_dcsw),
+	MTE_CMO_INSN(DC_IGDVAC, bad_cmo_trap),
 	MTE_CMO_INSN(DC_IGDSW, access_dcsw),
 
 	SYS_INSN(AT_S1E1R, handle_at_s1e01),
@@ -4321,6 +4329,7 @@ static struct sys_reg_desc sys_insn_descs[] = {
 	{ SYS_DESC(SYS_DC_CISW), access_dcsw },
 	MTE_CMO_INSN(DC_CIGSW, access_dcsw),
 	MTE_CMO_INSN(DC_CIGDSW, access_dcsw),
+	MTE_CMO_INSN(DC_CIGDVAPS, bad_cmo_trap),
 
 	SYS_INSN(TLBI_VMALLE1OS, handle_tlbi_el1),
 	SYS_INSN(TLBI_VAE1OS, handle_tlbi_el1),
@@ -4394,6 +4403,15 @@ static struct sys_reg_desc sys_insn_descs[] = {
 	SYS_INSN(TLBI_VALE1NXS, handle_tlbi_el1),
 	SYS_INSN(TLBI_VAALE1NXS, handle_tlbi_el1),
 
+	MTE_CMO_INSN(DC_CGVAC, bad_cmo_trap),
+	MTE_CMO_INSN(DC_CGDVAC, bad_cmo_trap),
+	MTE_CMO_INSN(DC_CGVAP, bad_cmo_trap),
+	MTE_CMO_INSN(DC_CGDVAP, bad_cmo_trap),
+	MTE_CMO_INSN(DC_CGVADP, bad_cmo_trap),
+	MTE_CMO_INSN(DC_CGDVADP, bad_cmo_trap),
+	MTE_CMO_INSN(DC_CIGVAC, bad_cmo_trap),
+	MTE_CMO_INSN(DC_CIGDVAC, bad_cmo_trap),
+
 	SYS_INSN(AT_S1E2R, handle_at_s1e2),
 	SYS_INSN(AT_S1E2W, handle_at_s1e2),
 	SYS_INSN(AT_S12E1R, handle_at_s12),
-- 
2.47.3

Re: [PATCH v2 6/8] KVM: arm64: Add sys_reg_descs for remaining MTE CMO instructions

[sashiko-bot]

Thank you for your contribution! Sashiko AI review found 1 potential
issue to consider:
- [Low] KVM: arm64: possible missing MTE_CMO_INSN entries for DC GVA and
  DC GZVA
--

commit c80fdbddf913183cac8f63dfc60ce6f48445eafa
Author: Oliver Upton <oupton@kernel.org>

KVM: arm64: Add sys_reg_descs for remaining MTE CMO instructions

This commit adds MTE_CMO_INSN descriptors with mte_visibility to
sys_insn_descs[] for MTE cache maintenance instructions trapped by
HCR_EL2.TPCP so that they inject an UNDEFINED exception when MTE is not
enabled for the VM.

> diff --git a/arch/arm64/kvm/sys_regs.c b/arch/arm64/kvm/sys_regs.c
> index b456354713443..7f7a197a9ad6b 100644
> --- a/arch/arm64/kvm/sys_regs.c
> +++ b/arch/arm64/kvm/sys_regs.c
[ ... ]
> @@ -4394,6 +4403,15 @@ static struct sys_reg_desc sys_insn_descs[] = {
>  	SYS_INSN(TLBI_VALE1NXS, handle_tlbi_el1),
>  	SYS_INSN(TLBI_VAALE1NXS, handle_tlbi_el1),
>  
> +	MTE_CMO_INSN(DC_CGVAC, bad_cmo_trap),
> +	MTE_CMO_INSN(DC_CGDVAC, bad_cmo_trap),
> +	MTE_CMO_INSN(DC_CGVAP, bad_cmo_trap),
> +	MTE_CMO_INSN(DC_CGDVAP, bad_cmo_trap),
> +	MTE_CMO_INSN(DC_CGVADP, bad_cmo_trap),
> +	MTE_CMO_INSN(DC_CGDVADP, bad_cmo_trap),
> +	MTE_CMO_INSN(DC_CIGVAC, bad_cmo_trap),
> +	MTE_CMO_INSN(DC_CIGDVAC, bad_cmo_trap),

[Severity: Low]
This question rests on an architecture detail that could not be confirmed
from the kernel source: can SYS_DC_GVA and SYS_DC_GZVA trap to EL2 via
HCR_EL2.TDZ or HFGITR_EL2.DCZVA when MTE is not enabled for the VM?

If so, encoding_to_cgt[] maps both instructions to CGT_HCR_TDZ and
encoding_to_fgt[] maps them to HFGITR_EL2.DCZVA, which is marked NEVER_FGU
in config.c because the trap bit is shared with SYS_DC_ZVA.  Without
MTE_CMO_INSN entries for SYS_DC_GVA and SYS_DC_GZVA in sys_insn_descs[],
triage_sysreg_trap() has neither an FGU bit nor a visibility() callback to
filter them, and will forward traps from a nested guest to L1 when L1 sets
HCR_EL2.TDZ or HFGITR_EL2.DCZVA instead of injecting an UNDEFINED
exception.

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20261106183154.3590825-1-oupton@kernel.org?part=6

Re: [PATCH v2 6/8] KVM: arm64: Add sys_reg_descs for remaining MTE CMO instructions

[Oliver Upton]

On Fri, Nov 06, 2026 at 06:45:39PM +0000, sashiko-bot@kernel.org wrote:
> Thank you for your contribution! Sashiko AI review found 1 potential
> issue to consider:
> - [Low] KVM: arm64: possible missing MTE_CMO_INSN entries for DC GVA and
>   DC GZVA
> --
> 
> commit c80fdbddf913183cac8f63dfc60ce6f48445eafa
> Author: Oliver Upton <oupton@kernel.org>
> 
> KVM: arm64: Add sys_reg_descs for remaining MTE CMO instructions
> 
> This commit adds MTE_CMO_INSN descriptors with mte_visibility to
> sys_insn_descs[] for MTE cache maintenance instructions trapped by
> HCR_EL2.TPCP so that they inject an UNDEFINED exception when MTE is not
> enabled for the VM.
> 
> > diff --git a/arch/arm64/kvm/sys_regs.c b/arch/arm64/kvm/sys_regs.c
> > index b456354713443..7f7a197a9ad6b 100644
> > --- a/arch/arm64/kvm/sys_regs.c
> > +++ b/arch/arm64/kvm/sys_regs.c
> [ ... ]
> > @@ -4394,6 +4403,15 @@ static struct sys_reg_desc sys_insn_descs[] = {
> >  	SYS_INSN(TLBI_VALE1NXS, handle_tlbi_el1),
> >  	SYS_INSN(TLBI_VAALE1NXS, handle_tlbi_el1),
> >  
> > +	MTE_CMO_INSN(DC_CGVAC, bad_cmo_trap),
> > +	MTE_CMO_INSN(DC_CGDVAC, bad_cmo_trap),
> > +	MTE_CMO_INSN(DC_CGVAP, bad_cmo_trap),
> > +	MTE_CMO_INSN(DC_CGDVAP, bad_cmo_trap),
> > +	MTE_CMO_INSN(DC_CGVADP, bad_cmo_trap),
> > +	MTE_CMO_INSN(DC_CGDVADP, bad_cmo_trap),
> > +	MTE_CMO_INSN(DC_CIGVAC, bad_cmo_trap),
> > +	MTE_CMO_INSN(DC_CIGDVAC, bad_cmo_trap),
> 
> [Severity: Low]
> This question rests on an architecture detail that could not be confirmed
> from the kernel source: can SYS_DC_GVA and SYS_DC_GZVA trap to EL2 via
> HCR_EL2.TDZ or HFGITR_EL2.DCZVA when MTE is not enabled for the VM?
> 
> If so, encoding_to_cgt[] maps both instructions to CGT_HCR_TDZ and
> encoding_to_fgt[] maps them to HFGITR_EL2.DCZVA, which is marked NEVER_FGU
> in config.c because the trap bit is shared with SYS_DC_ZVA.  Without
> MTE_CMO_INSN entries for SYS_DC_GVA and SYS_DC_GZVA in sys_insn_descs[],
> triage_sysreg_trap() has neither an FGU bit nor a visibility() callback to
> filter them, and will forward traps from a nested guest to L1 when L1 sets
> HCR_EL2.TDZ or HFGITR_EL2.DCZVA instead of injecting an UNDEFINED
> exception.

Urgh, I'm starting to get the sense that we should pursue a generalized
solution. This relies on adding a subtle detail to KVM's trap behavior
when what we're trying to do already has a direct description in the
architecture...

Thanks,
Oliver

[PATCH v2 7/8] KVM: arm64: nv: Enforce sysreg visibility before evaulating guest traps

[Oliver Upton]

Not all sysregs / system instructions can be described by FGUs. Of
these, there are some which may be not implemented for the VM (e.g. MTE
cache maintenance instructions) yet intersect with valid trap bits for
the VM (HCR_EL2.TPCP).

This is annoying to deal with since most of our UNDEF infrastructure is
built around traps being RESx for unimplemented features. Test the
visibility of the sys_reg_desc to avoid injecting traps for non-existent
system registers into the guest hypervisor.

Cc: stable@vger.kernel.org
Fixes: a7484c80e5ca ("KVM: arm64: Allow userspace to request KVM_ARM_VCPU_EL2*")
Signed-off-by: Oliver Upton <oupton@kernel.org>
---
 arch/arm64/kvm/emulate-nested.c | 24 +++++++++++++++++++++---
 arch/arm64/kvm/sys_regs.c       | 18 ++++++++++--------
 arch/arm64/kvm/sys_regs.h       |  4 +++-
 3 files changed, 34 insertions(+), 12 deletions(-)

diff --git a/arch/arm64/kvm/emulate-nested.c b/arch/arm64/kvm/emulate-nested.c
index d04ac5203454..46dc658b041b 100644
--- a/arch/arm64/kvm/emulate-nested.c
+++ b/arch/arm64/kvm/emulate-nested.c
@@ -2546,12 +2546,13 @@ static bool check_fgt_bit(struct kvm_vcpu *vcpu, enum vcpu_sysreg sr,
 	return !(kvm_get_sysreg_res0(kvm, sr) & BIT(tc.bit));
 }
 
-bool triage_sysreg_trap(struct kvm_vcpu *vcpu, int *sr_index)
+bool triage_sysreg_trap(struct kvm_vcpu *vcpu, const struct sys_reg_desc **r)
 {
 	enum vcpu_sysreg fgtreg;
 	union trap_config tc;
 	enum trap_behaviour b;
 	bool is_read;
+	int sr_index;
 	u32 sysreg;
 	u64 esr;
 
@@ -2559,6 +2560,7 @@ bool triage_sysreg_trap(struct kvm_vcpu *vcpu, int *sr_index)
 	sysreg = esr_sys64_to_sysreg(esr);
 	is_read = (esr & ESR_ELx_SYS64_ISS_DIR_MASK) == ESR_ELx_SYS64_ISS_DIR_READ;
 
+	*r = NULL;
 	tc = get_trap_config(sysreg);
 
 	/*
@@ -2569,6 +2571,11 @@ bool triage_sysreg_trap(struct kvm_vcpu *vcpu, int *sr_index)
 	if (!tc.val)
 		goto local;
 
+	if (tc.sri) {
+		sr_index = tc.sri - 1;
+		*r = kvm_sys_reg_idx_to_desc(esr_sys64_to_sysreg(esr), sr_index);
+	}
+
 	/*
 	 * If a sysreg can be trapped using a FGT, first check whether we
 	 * trap for the purpose of forbidding the feature. In that case,
@@ -2595,6 +2602,18 @@ bool triage_sysreg_trap(struct kvm_vcpu *vcpu, int *sr_index)
 	if (is_hyp_ctxt(vcpu) && !vcpu_is_host_el0(vcpu))
 		goto local;
 
+	/*
+	 * Certain sysregs and system instructions could be UNDEF for the VM
+	 * but lack a corresponding FGU description and intersect with a valid
+	 * trap bit for the VM. Test the visibility the sysreg (if we have a
+	 * sys_reg_desc) before determining the effects of the guest
+	 * hypervisor's traps.
+	 */
+	if (*r && sysreg_guest_hidden(vcpu, *r)) {
+		kvm_inject_undefined(vcpu);
+		return true;
+	}
+
 	switch ((enum fgt_group_id)tc.fgt) {
 	case __NO_FGT_GROUP__:
 		break;
@@ -2666,7 +2685,7 @@ bool triage_sysreg_trap(struct kvm_vcpu *vcpu, int *sr_index)
 		goto inject;
 
 local:
-	if (!tc.sri) {
+	if (!(*r)) {
 		struct sys_reg_params params;
 
 		params = esr_sys64_to_params(esr);
@@ -2697,7 +2716,6 @@ bool triage_sysreg_trap(struct kvm_vcpu *vcpu, int *sr_index)
 		return true;
 	}
 
-	*sr_index = tc.sri - 1;
 	return false;
 
 inject:
diff --git a/arch/arm64/kvm/sys_regs.c b/arch/arm64/kvm/sys_regs.c
index 7f7a197a9ad6..7c8277de0494 100644
--- a/arch/arm64/kvm/sys_regs.c
+++ b/arch/arm64/kvm/sys_regs.c
@@ -5407,6 +5407,15 @@ void kvm_reset_sys_regs(struct kvm_vcpu *vcpu)
 		kvm_make_request(KVM_REQ_RELOAD_PMU, vcpu);
 }
 
+const struct sys_reg_desc *kvm_sys_reg_idx_to_desc(u32 encoding, int sr_idx)
+{
+	/* System registers have Op0=={2,3}, as per DDI487 J.a C5.1.2 */
+	if (sys_reg_Op0(encoding) == 2 || sys_reg_Op0(encoding) == 3)
+		return &sys_reg_descs[sr_idx];
+	else
+		return &sys_insn_descs[sr_idx];
+}
+
 /**
  * kvm_handle_sys_reg -- handles a system instruction or mrs/msr instruction
  *			 trap on a guest execution
@@ -5418,22 +5427,15 @@ int kvm_handle_sys_reg(struct kvm_vcpu *vcpu)
 	struct sys_reg_params params;
 	unsigned long esr = kvm_vcpu_get_esr(vcpu);
 	int Rt = kvm_vcpu_sys_get_rt(vcpu);
-	int sr_idx;
 
 	trace_kvm_handle_sys_reg(esr);
 
-	if (triage_sysreg_trap(vcpu, &sr_idx))
+	if (triage_sysreg_trap(vcpu, &desc))
 		return 1;
 
 	params = esr_sys64_to_params(esr);
 	params.regval = vcpu_get_reg(vcpu, Rt);
 
-	/* System registers have Op0=={2,3}, as per DDI487 J.a C5.1.2 */
-	if (params.Op0 == 2 || params.Op0 == 3)
-		desc = &sys_reg_descs[sr_idx];
-	else
-		desc = &sys_insn_descs[sr_idx];
-
 	perform_access(vcpu, &params, desc);
 
 	/* Read from system register? */
diff --git a/arch/arm64/kvm/sys_regs.h b/arch/arm64/kvm/sys_regs.h
index 50214271b24a..d8ae21578665 100644
--- a/arch/arm64/kvm/sys_regs.h
+++ b/arch/arm64/kvm/sys_regs.h
@@ -263,10 +263,12 @@ int kvm_sys_reg_get_user(struct kvm_vcpu *vcpu, const struct kvm_one_reg *reg,
 int kvm_sys_reg_set_user(struct kvm_vcpu *vcpu, const struct kvm_one_reg *reg,
 			 const struct sys_reg_desc table[], unsigned int num);
 
-bool triage_sysreg_trap(struct kvm_vcpu *vcpu, int *sr_index);
+bool triage_sysreg_trap(struct kvm_vcpu *vcpu, const struct sys_reg_desc **r);
 
 int kvm_vcpu_finalize_sys_regs(struct kvm_vcpu *vcpu);
 
+const struct sys_reg_desc *kvm_sys_reg_idx_to_desc(u32 encoding, int sr_idx);
+
 #define AA32(_x)	.aarch32_map = AA32_##_x
 #define Op0(_x) 	.Op0 = _x
 #define Op1(_x) 	.Op1 = _x
-- 
2.47.3

[PATCH v2 8/8] KVM: arm64: Stop leaning on __kvm_inject_sea() for abort injection

[Oliver Upton]

KVM's handling of local abort injection is, politely stated, a giant
fscking mess. At surface level, things look somewhat innocuous: inject
an SEA and directly update ESR_EL1 with the correct FSC afterwards. But
this is broken for a couple reasons.

For one, on implementations with FEAT_DoubleFault2 KVM may inject the
SEA as an SError depending on the falue of SCTLR2_EL1.NMEA, meaning that
the eventual non-SEA abort is incorrectly taken to the SError vector.

AArch32 is also quite bitrotted. No surprise. The FSR value is written
to IFSR not DFSR, when a prefetch abort is injected. So while DFSR is
aliased to ESR_EL1 architecturally, fixing up ESR_EL1.FSC after
injection has no effect on IFSR.

Finally, injecting aborts into non-LPAE relics may use an invalid value
for xFSC.FS, as KVM unconditionally uses the LPAE/VMSAv8-64 format.

Fix the first two by tearing the whole mess, adding __kvm_inject_abort()
for injecting non-SEA aborts which takes the desired FSC from the
caller. Fix the third one by selecting a fault status value based on the
LPAE-ness of the guest.

Cc: stable@vger.kernel.org
Fixes: fff97df2a0bd ("KVM: arm64: Route SEAs to the SError vector when EASE is set")
Fixes: 2937aeec9dc5 ("KVM: arm64: Handle DABT caused by LS64* instructions on unsupported memory")
Signed-off-by: Oliver Upton <oupton@kernel.org>
---
 arch/arm64/kvm/inject_fault.c | 133 +++++++++++++++++++---------------
 1 file changed, 76 insertions(+), 57 deletions(-)

diff --git a/arch/arm64/kvm/inject_fault.c b/arch/arm64/kvm/inject_fault.c
index aa2011cc0d5a..89c7890b8743 100644
--- a/arch/arm64/kvm/inject_fault.c
+++ b/arch/arm64/kvm/inject_fault.c
@@ -102,12 +102,32 @@ static bool effective_sctlr2_nmea(struct kvm_vcpu *vcpu)
 	return __effective_sctlr2_bit(vcpu, SCTLR2_EL1_NMEA_SHIFT);
 }
 
-static void inject_abt64(struct kvm_vcpu *vcpu, bool is_iabt, unsigned long addr)
+static void __inject_abt64(struct kvm_vcpu *vcpu, bool iabt, u32 iss, unsigned long addr)
 {
-	unsigned long cpsr = *vcpu_cpsr(vcpu);
-	bool is_aarch32 = vcpu_mode_is_32bit(vcpu);
-	u64 esr = 0, fsc;
-	int level;
+	u64 esr;
+	u8 ec;
+
+	if (vcpu_mode_priv(vcpu))
+		ec = iabt ? ESR_ELx_EC_IABT_CUR : ESR_ELx_EC_DABT_CUR;
+	else
+		ec = iabt ? ESR_ELx_EC_IABT_LOW : ESR_ELx_EC_DABT_LOW;
+
+	esr = FIELD_PREP(ESR_ELx_EC_MASK, ec) | FIELD_PREP(ESR_ELx_ISS_MASK, iss) |
+	      ESR_ELx_IL;
+
+	vcpu_write_sys_reg(vcpu, addr, exception_far_elx(vcpu));
+	vcpu_write_sys_reg(vcpu, esr, exception_esr_elx(vcpu));
+}
+
+static void inject_abt64(struct kvm_vcpu *vcpu, bool iabt, u32 iss, u64 addr)
+{
+	pend_sync_exception(vcpu);
+	__inject_abt64(vcpu, iabt, iss, addr);
+}
+
+static void inject_sea64(struct kvm_vcpu *vcpu, bool iabt, u64 addr)
+{
+	u8 fsc;
 
 	/*
 	 * If injecting an abort from a failed S1PTW, rewalk the S1 PTs to
@@ -116,7 +136,7 @@ static void inject_abt64(struct kvm_vcpu *vcpu, bool is_iabt, unsigned long addr
 	 */
 	if (kvm_vcpu_abt_iss1tw(vcpu)) {
 		u64 hpfar = kvm_vcpu_get_fault_ipa(vcpu);
-		int ret;
+		int ret, level;
 
 		if (hpfar == INVALID_GPA)
 			return;
@@ -137,28 +157,7 @@ static void inject_abt64(struct kvm_vcpu *vcpu, bool is_iabt, unsigned long addr
 	else
 		pend_sync_exception(vcpu);
 
-	/*
-	 * Build an {i,d}abort, depending on the level.
-	 * Report an external synchronous abort.
-	 */
-	esr |= ESR_ELx_IL;
-
-	/*
-	 * Here, the guest runs in AArch64 mode when in EL1. If we get
-	 * an AArch32 fault, it means we managed to trap an EL0 fault.
-	 */
-	if (is_aarch32 || (cpsr & PSR_MODE_MASK) == PSR_MODE_EL0t)
-		esr |= (ESR_ELx_EC_IABT_LOW << ESR_ELx_EC_SHIFT);
-	else
-		esr |= (ESR_ELx_EC_IABT_CUR << ESR_ELx_EC_SHIFT);
-
-	if (!is_iabt)
-		esr |= ESR_ELx_EC_DABT_LOW << ESR_ELx_EC_SHIFT;
-
-	esr |= fsc;
-
-	vcpu_write_sys_reg(vcpu, addr, exception_far_elx(vcpu));
-	vcpu_write_sys_reg(vcpu, esr, exception_esr_elx(vcpu));
+	__inject_abt64(vcpu, iabt, fsc, addr);
 }
 
 void kvm_inject_sync(struct kvm_vcpu *vcpu, u64 esr)
@@ -174,31 +173,41 @@ static void inject_undef64(struct kvm_vcpu *vcpu)
 	kvm_inject_sync(vcpu, esr);
 }
 
-#define DFSR_FSC_EXTABT_LPAE	0x10
-#define DFSR_FSC_EXTABT_nLPAE	0x08
-#define DFSR_LPAE		BIT(9)
-#define TTBCR_EAE		BIT(31)
+#define DFSR_FS_4			BIT(10)
+#define DFSR_FS_3_0			GENMASK(3, 0)
+#define DFSR_FS_EXTABT_nLPAE		0x08
+#define DFSR_FS_EXCL_ATOMIC_nLPAE	0x15
+#define DFSR_FSC			GENMASK(5, 0)
+#define DFSR_FSC_EXTABT_LPAE		0x10
+#define DFSR_LPAE			BIT(9)
+#define TTBCR_EAE			BIT(31)
 
 static void inject_undef32(struct kvm_vcpu *vcpu)
 {
 	kvm_pend_exception(vcpu, EXCEPT_AA32_UND);
 }
 
+static bool vcpu_trans_32bit_non_lpae(struct kvm_vcpu *vcpu)
+{
+	return vcpu_el1_is_32bit(vcpu) && !(vcpu_read_sys_reg(vcpu, TCR_EL1) & TTBCR_EAE);
+}
+
 /*
  * Modelled after TakeDataAbortException() and TakePrefetchAbortException
  * pseudocode.
  */
-static void inject_abt32(struct kvm_vcpu *vcpu, bool is_pabt, u32 addr)
+static void inject_abt32(struct kvm_vcpu *vcpu, bool is_pabt, u8 status, u32 addr)
 {
 	u64 far;
 	u32 fsr;
 
-	/* Give the guest an IMPLEMENTATION DEFINED exception */
-	if (vcpu_read_sys_reg(vcpu, TCR_EL1) & TTBCR_EAE) {
-		fsr = DFSR_LPAE | DFSR_FSC_EXTABT_LPAE;
+	if (vcpu_trans_32bit_non_lpae(vcpu)) {
+		/* Ensure status actually fits in DFSR.FS */
+		KVM_BUG_ON(status > 0x1F, vcpu->kvm);
+		fsr = FIELD_PREP(DFSR_FS_4, FIELD_GET(BIT(4), status)) |
+		      FIELD_PREP(DFSR_FS_3_0, FIELD_GET(GENMASK(3, 0), status));
 	} else {
-		/* no need to shuffle FS[4] into DFSR[10] as it's 0 */
-		fsr = DFSR_FSC_EXTABT_nLPAE;
+		fsr = FIELD_PREP(DFSR_FSC, status) | DFSR_LPAE;
 	}
 
 	far = vcpu_read_sys_reg(vcpu, FAR_EL1);
@@ -218,12 +227,28 @@ static void inject_abt32(struct kvm_vcpu *vcpu, bool is_pabt, u32 addr)
 	vcpu_write_sys_reg(vcpu, far, FAR_EL1);
 }
 
+static void inject_sea32(struct kvm_vcpu *vcpu, bool is_pabt, u32 addr)
+{
+	if (vcpu_trans_32bit_non_lpae(vcpu))
+		inject_abt32(vcpu, is_pabt, DFSR_FS_EXTABT_nLPAE, addr);
+	else
+		inject_abt32(vcpu, is_pabt, DFSR_FSC_EXTABT_LPAE, addr);
+}
+
 static void __kvm_inject_sea(struct kvm_vcpu *vcpu, bool iabt, u64 addr)
 {
 	if (vcpu_el1_is_32bit(vcpu))
-		inject_abt32(vcpu, iabt, addr);
+		inject_sea32(vcpu, iabt, addr);
 	else
-		inject_abt64(vcpu, iabt, addr);
+		inject_sea64(vcpu, iabt, addr);
+}
+
+static void __kvm_inject_abort(struct kvm_vcpu *vcpu, bool iabt, u8 fsc, u64 addr)
+{
+	if (vcpu_el1_is_32bit(vcpu))
+		inject_abt32(vcpu, iabt, fsc, addr);
+	else
+		inject_abt64(vcpu, iabt, fsc, addr);
 }
 
 static bool kvm_sea_target_is_el2(struct kvm_vcpu *vcpu)
@@ -270,28 +295,25 @@ static int kvm_inject_nested_excl_atomic(struct kvm_vcpu *vcpu, u64 addr)
  */
 int kvm_inject_dabt_excl_atomic(struct kvm_vcpu *vcpu, u64 addr)
 {
-	u64 esr;
-
 	if (is_nested_ctxt(vcpu) && (vcpu_read_sys_reg(vcpu, HCR_EL2) & HCR_EL2_VM))
 		return kvm_inject_nested_excl_atomic(vcpu, addr);
 
-	__kvm_inject_sea(vcpu, false, addr);
-	esr = vcpu_read_sys_reg(vcpu, exception_esr_elx(vcpu));
-	esr &= ~ESR_ELx_FSC;
-	esr |= ESR_ELx_FSC_EXCL_ATOMIC;
-	vcpu_write_sys_reg(vcpu, esr, exception_esr_elx(vcpu));
+	if (vcpu_trans_32bit_non_lpae(vcpu))
+		__kvm_inject_abort(vcpu, false, DFSR_FS_EXCL_ATOMIC_nLPAE, addr);
+	else
+		__kvm_inject_abort(vcpu, false, ESR_ELx_FSC_EXCL_ATOMIC, addr);
+
 	return 1;
 }
 
 void kvm_inject_size_fault(struct kvm_vcpu *vcpu)
 {
-	unsigned long addr, esr;
+	bool iabt = kvm_vcpu_trap_is_iabt(vcpu);
+	unsigned long addr;
 
 	addr  = kvm_vcpu_get_fault_ipa(vcpu);
 	addr |= FAR_TO_FIPA_OFFSET(kvm_vcpu_get_hfar(vcpu));
 
-	__kvm_inject_sea(vcpu, kvm_vcpu_trap_is_iabt(vcpu), addr);
-
 	/*
 	 * If AArch64 or LPAE, set FSC to 0 to indicate an Address
 	 * Size Fault at level 0, as if exceeding PARange.
@@ -299,13 +321,10 @@ void kvm_inject_size_fault(struct kvm_vcpu *vcpu)
 	 * Non-LPAE guests will only get the external abort, as there
 	 * is no way to describe the ASF.
 	 */
-	if (vcpu_el1_is_32bit(vcpu) &&
-	    !(vcpu_read_sys_reg(vcpu, TCR_EL1) & TTBCR_EAE))
-		return;
-
-	esr = vcpu_read_sys_reg(vcpu, exception_esr_elx(vcpu));
-	esr &= ~GENMASK_ULL(5, 0);
-	vcpu_write_sys_reg(vcpu, esr, exception_esr_elx(vcpu));
+	if (vcpu_trans_32bit_non_lpae(vcpu))
+		__kvm_inject_sea(vcpu, iabt, addr);
+	else
+		__kvm_inject_abort(vcpu, iabt, ESR_ELx_FSC_ADDRSZ, addr);
 }
 
 /**
-- 
2.47.3

Re: [PATCH v2 8/8] KVM: arm64: Stop leaning on __kvm_inject_sea() for abort injection

[sashiko-bot]

Thank you for your contribution! Sashiko AI review found 1 potential
issue to consider:
- [Low] KVM: arm64: missing Fixes trailer for kvm_inject_size_fault()
  AArch32 PABT bug
--

commit f2d6dc968339775720bc3ec74dcb33374b8473b1
Author: Oliver Upton <oupton@kernel.org>

> KVM: arm64: Stop leaning on __kvm_inject_sea() for abort injection
>
> KVM's handling of local abort injection is, politely stated, a giant
> fscking mess. At surface level, things look somewhat innocuous: inject
> an SEA and directly update ESR_EL1 with the correct FSC afterwards. But
> this is broken for a couple reasons.
>
> For one, on implementations with FEAT_DoubleFault2 KVM may inject the
> SEA as an SError depending on the falue of SCTLR2_EL1.NMEA, meaning that
> the eventual non-SEA abort is incorrectly taken to the SError vector.
>
> AArch32 is also quite bitrotted. No surprise. The FSR value is written
> to IFSR not DFSR, when a prefetch abort is injected. So while DFSR is
> aliased to ESR_EL1 architecturally, fixing up ESR_EL1.FSC after
> injection has no effect on IFSR.
>
> Finally, injecting aborts into non-LPAE relics may use an invalid value
> for xFSC.FS, as KVM unconditionally uses the LPAE/VMSAv8-64 format.
>
> Fix the first two by tearing the whole mess, adding __kvm_inject_abort()
> for injecting non-SEA aborts which takes the desired FSC from the
> caller. Fix the third one by selecting a fault status value based on the
> LPAE-ness of the guest.
>
> Cc: stable@vger.kernel.org
> Fixes: fff97df2a0bd ("KVM: arm64: Route SEAs to the SError vector when EASE is set")
> Fixes: 2937aeec9dc5 ("KVM: arm64: Handle DABT caused by LS64* instructions on unsupported memory")
> Signed-off-by: Oliver Upton <oupton@kernel.org>

[Severity: Low]
Of the two functions changed here, only kvm_inject_size_fault() can inject
an AArch32 prefetch abort (where FSR is written to IFSR32_EL2 while the
post-injection fixup only cleared ESR_EL1).  Should this also include:

  Fixes: 85ea6b1ec915 ("KVM: arm64: Inject exception on out-of-IPA-range translation fault")

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20261106183154.3590825-1-oupton@kernel.org?part=8