Skip to contentSkip to conversation

RISC-V: KVM: Validate AIA MMIO ranges against GPA width

PATCH In review Pengpeng Hou 3 participants 7 messages Sep 20 – Nov 13

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

v2 cover RISC-V: KVM: Validate AIA MMIO ranges against GPA width

This revision keeps the existing mandatory-address checks and narrows the initialization revalidation to APLIC. An IMSIC address is only set after a vCPU exists, which already freezes the selected GPA width. APLIC can be configured earlier, so its range is rechecked at init.

Patch 2 keeps the previously reviewed GPA-width freeze unchanged.

Changes since v1:
https://lore.kernel.org/r/[email protected]/
- Preserve the required-address checks and existing cleanup path.
- Remove the all-vCPU range-validation pass.
- Explain why only APLIC needs a range recheck before initialization.
- Retain Anup's Reviewed-by for unchanged patch 2.

The issues were found by our static-analysis tool. AI assistance was used to revise the patches and descriptions following the review.

Pengpeng Hou (2):
RISC-V: KVM: Validate AIA MMIO address ranges
RISC-V: KVM: Disallow GPA-width changes after AIA init

 arch/riscv/kvm/aia_device.c | 42 +++++++++++++++++++++++++++++++++----
 arch/riscv/kvm/vm.c         |  3 ++-
 2 files changed, 40 insertions(+), 5 deletions(-)

base-commit: 518e5b794c06c0f0eb40df3e202274a66202c137

Signature · 3 lines
kvm-riscv mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/kvm-riscv
#1 lore raw

PATCH v2 1/2 RISC-V: KVM: Validate AIA MMIO address ranges

The AIA address setters check alignment but not the complete fixed-size APLIC and IMSIC MMIO regions. A range can therefore wrap when its end is formed or extend beyond the selected guest physical address space before it is registered on the KVM MMIO bus.

Check the complete range in both setters. Return -EINVAL for alignment or arithmetic overflow and -E2BIG for a range beyond the selected GPA limit, following the arm64 VGIC address-validation convention.

Recheck only the APLIC range during initialization: userspace can set its address and then change KVM_CAP_VM_GPA_BITS while the VM still has no vCPUs or memory slots. An IMSIC address, in contrast, can only be set for an existing vCPU, which already prevents subsequent GPA-width changes. No additional pass over the vCPUs is needed.

Keep the existing required-address checks and IMSIC cleanup path.

The issue was found by our static-analysis tool.

Fixes 89d01306e34d ("RISC-V: KVM: Implement device interface for AIA irqchip")
Assisted-by gpt 5

---
Changes since v1: https://lore.kernel.org/r/38f87205419f328693bc6ab25ae78f021c2dd003.1786512671.git.pengpeng@iscas.ac.cn/
- Keep the original APLIC and per-vCPU IMSIC required-address checks.
- Drop the separate pass over all vCPUs during initialization.
- Recheck only APLIC, whose address can be configured before GPA width
becomes immutable; IMSIC setters already require an existing vCPU.

 arch/riscv/kvm/aia_device.c | 42 +++++++++++++++++++++++++++++++++----
 1 file changed, 38 insertions(+), 4 deletions(-)
arch/riscv/kvm/aia_device.c +38 -4
index efc7c0b..4490015 100644
@@ -10,7 +10,9 @@
 #include <linux/bits.h>
 #include <linux/irqchip/riscv-imsic.h>
 #include <linux/kvm_host.h>
+#include <linux/overflow.h>
 #include <linux/uaccess.h>
+#include <asm/kvm_gstage.h>
 #include <asm/kvm_isa.h>
 
 static int aia_create(struct kvm_device *dev, u32 type)
@@ -142,17 +144,37 @@ static int aia_config(struct kvm *kvm, unsigned long type,
 	return 0;
 }
 
+static int aia_check_addr_range(struct kvm *kvm, u64 addr, u64 alignment,
+				u64 size)
+{
+	u64 end;
+
+	if (!IS_ALIGNED(addr, alignment))
+		return -EINVAL;
+
+	if (check_add_overflow(addr, size, &end))
+		return -EINVAL;
+
+	if (end > kvm_riscv_gstage_gpa_size(kvm->arch.pgd_levels))
+		return -E2BIG;
+
+	return 0;
+}
+
 static int aia_aplic_addr(struct kvm *kvm, u64 *addr, bool write)
 {
 	struct kvm_aia *aia = &kvm->arch.aia;
+	int ret;
 
 	if (write) {
 		/* Writes can only be done before irqchip is initialized */
 		if (kvm_riscv_aia_initialized(kvm))
 			return -EBUSY;
 
-		if (*addr & (KVM_DEV_RISCV_APLIC_ALIGN - 1))
-			return -EINVAL;
+		ret = aia_check_addr_range(kvm, *addr, KVM_DEV_RISCV_APLIC_ALIGN,
+					   KVM_DEV_RISCV_APLIC_SIZE);
+		if (ret)
+			return ret;
 
 		aia->aplic_addr = *addr;
 	} else
@@ -166,6 +188,7 @@ static int aia_imsic_addr(struct kvm *kvm, u64 *addr,
 {
 	struct kvm_vcpu *vcpu;
 	struct kvm_vcpu_aia *vcpu_aia;
+	int ret;
 
 	vcpu = kvm_get_vcpu(kvm, vcpu_idx);
 	if (!vcpu)
@@ -177,8 +200,10 @@ static int aia_imsic_addr(struct kvm *kvm, u64 *addr,
 		if (kvm_riscv_aia_initialized(kvm))
 			return -EBUSY;
 
-		if (*addr & (KVM_DEV_RISCV_IMSIC_ALIGN - 1))
-			return -EINVAL;
+		ret = aia_check_addr_range(kvm, *addr, KVM_DEV_RISCV_IMSIC_ALIGN,
+					   KVM_DEV_RISCV_IMSIC_SIZE);
+		if (ret)
+			return ret;
 	}
 
 	mutex_lock(&vcpu->mutex);
@@ -248,6 +273,15 @@ static int aia_init(struct kvm *kvm)
 	if (aia->nr_sources && aia->aplic_addr == KVM_RISCV_AIA_UNDEF_ADDR)
 		return -EINVAL;
 
+	/* The GPA width may change after setting APLIC's address on an empty VM. */
+	if (aia->nr_sources) {
+		ret = aia_check_addr_range(kvm, aia->aplic_addr,
+					   KVM_DEV_RISCV_APLIC_ALIGN,
+					   KVM_DEV_RISCV_APLIC_SIZE);
+		if (ret)
+			return ret;
+	}
+
 	/* Group index bits must not overlap guest and HART index bits. */
 	if (aia->nr_group_bits &&
 	    aia->nr_group_shift < (IMSIC_MMIO_PAGE_SHIFT +
maintainer Reviewed-by on v2 1/2
#2 lore raw

Re: [PATCH v2 1/2] RISC-V: KVM: Validate AIA MMIO address ranges

Quoted 111 lines · Pengpeng Hou

The AIA address setters check alignment but not the complete fixed-size APLIC and IMSIC MMIO regions. A range can therefore wrap when its end is formed or extend beyond the selected guest physical address space before it is registered on the KVM MMIO bus.

Check the complete range in both setters. Return -EINVAL for alignment or arithmetic overflow and -E2BIG for a range beyond the selected GPA limit, following the arm64 VGIC address-validation convention.

Recheck only the APLIC range during initialization: userspace can set its address and then change KVM_CAP_VM_GPA_BITS while the VM still has no vCPUs or memory slots. An IMSIC address, in contrast, can only be set for an existing vCPU, which already prevents subsequent GPA-width changes. No additional pass over the vCPUs is needed.

Keep the existing required-address checks and IMSIC cleanup path.

The issue was found by our static-analysis tool.

Fixes 89d01306e34d ("RISC-V: KVM: Implement device interface for AIA irqchip")
Assisted-by gpt 5

---
Changes since v1: https://lore.kernel.org/r/38f87205419f328693bc6ab25ae78f021c2dd003.1786512671.git.pengpeng@iscas.ac.cn/
- Keep the original APLIC and per-vCPU IMSIC required-address checks.
- Drop the separate pass over all vCPUs during initialization.
- Recheck only APLIC, whose address can be configured before GPA width
becomes immutable; IMSIC setters already require an existing vCPU.

 arch/riscv/kvm/aia_device.c | 42 +++++++++++++++++++++++++++++++++----
 1 file changed, 38 insertions(+), 4 deletions(-)
arch/riscv/kvm/aia_device.c +29 -4
index efc7c0b..4490015 100644
@@ -10,7 +10,9 @@
 #include <linux/bits.h>
 #include <linux/irqchip/riscv-imsic.h>
 #include <linux/kvm_host.h>
+#include <linux/overflow.h>
 #include <linux/uaccess.h>
+#include <asm/kvm_gstage.h>
 #include <asm/kvm_isa.h>
 
 static int aia_create(struct kvm_device *dev, u32 type)
@@ -142,17 +144,37 @@ static int aia_config(struct kvm *kvm, unsigned long type,
        return 0;
 }
 
+static int aia_check_addr_range(struct kvm *kvm, u64 addr, u64 alignment,
+                               u64 size)
+{
+       u64 end;
+
+       if (!IS_ALIGNED(addr, alignment))
+               return -EINVAL;
+
+       if (check_add_overflow(addr, size, &end))
+               return -EINVAL;
+
+       if (end > kvm_riscv_gstage_gpa_size(kvm->arch.pgd_levels))
+               return -E2BIG;
+
+       return 0;
+}
+
 static int aia_aplic_addr(struct kvm *kvm, u64 *addr, bool write)
 {
        struct kvm_aia *aia = &kvm->arch.aia;
+       int ret;
 
        if (write) {
                /* Writes can only be done before irqchip is initialized */
                if (kvm_riscv_aia_initialized(kvm))
                        return -EBUSY;
 
-               if (*addr & (KVM_DEV_RISCV_APLIC_ALIGN - 1))
-                       return -EINVAL;
+               ret = aia_check_addr_range(kvm, *addr, KVM_DEV_RISCV_APLIC_ALIGN,
+                                          KVM_DEV_RISCV_APLIC_SIZE);
+               if (ret)
+                       return ret;
 
                aia->aplic_addr = *addr;
        } else
@@ -166,6 +188,7 @@ static int aia_imsic_addr(struct kvm *kvm, u64 *addr,
 {
        struct kvm_vcpu *vcpu;
        struct kvm_vcpu_aia *vcpu_aia;
+       int ret;
 
        vcpu = kvm_get_vcpu(kvm, vcpu_idx);
        if (!vcpu)
@@ -177,8 +200,10 @@ static int aia_imsic_addr(struct kvm *kvm, u64 *addr,
                if (kvm_riscv_aia_initialized(kvm))
                        return -EBUSY;
 
-               if (*addr & (KVM_DEV_RISCV_IMSIC_ALIGN - 1))
-                       return -EINVAL;
+               ret = aia_check_addr_range(kvm, *addr, KVM_DEV_RISCV_IMSIC_ALIGN,
+                                          KVM_DEV_RISCV_IMSIC_SIZE);
+               if (ret)
+                       return ret;
        }
 
        mutex_lock(&vcpu->mutex);
@@ -248,6 +273,15 @@ static int aia_init(struct kvm *kvm)
        if (aia->nr_sources && aia->aplic_addr == KVM_RISCV_AIA_UNDEF_ADDR)
                return -EINVAL;

arch/riscv/kvm/aia_device.c static int aia_init(struct kvm *kvm) · line 284

Quoted 10 lines
arch/riscv/kvm/aia_device.c +9 -0
 
+       /* The GPA width may change after setting APLIC's address on an empty VM. */
+       if (aia->nr_sources) {
+               ret = aia_check_addr_range(kvm, aia->aplic_addr,
+                                          KVM_DEV_RISCV_APLIC_ALIGN,
+                                          KVM_DEV_RISCV_APLIC_SIZE);
+               if (ret)
+                       return ret;
+       }
+

We should also check IMSIC addresses one more time in aia_init(). I will take care of it at the time of merging.

arch/riscv/kvm/aia_device.c static int aia_init(struct kvm *kvm) · line 287

Quoted 4 lines
arch/riscv/kvm/aia_device.c +0 -0
        /* Group index bits must not overlap guest and HART index bits. */
        if (aia->nr_group_bits &&
            aia->nr_group_shift < (IMSIC_MMIO_PAGE_SHIFT +

LGTM.

Queued this patch as a fix for Linux-7.4-rcX

Thanks,
Anup

#3 lore raw

PATCH v2 2/2 RISC-V: KVM: Disallow GPA-width changes after AIA init

In reply to Pengpeng Hou

KVM_ENABLE_CAP(KVM_CAP_VM_GPA_BITS) allows userspace to reduce the stage-2 GPA width while a VM has no vCPUs or memory slots.

AIA initialization can complete with an APLIC and no vCPUs or memory slots. The APLIC MMIO device has then been registered, but the existing checks still allow userspace to shrink the GPA width below its address.

Reject GPA-width changes after AIA initialization. Both paths hold kvm->lock, serializing the width change with AIA initialization.

The issue was found by our static-analysis tool.

Fixes 7263b4fdb0b2 ("RISC-V: KVM: Reuse KVM_CAP_VM_GPA_BITS to select HGATP.MODE")
Assisted-by gpt 5

---
Changes since v1: https://lore.kernel.org/r/4852f3b8985c353811fa065e4701d76dc9a7f086.1786512671.git.pengpeng@iscas.ac.cn/
- No change to the code; retain Anup's Reviewed-by.
- Rebase and clarify the locking description.

 arch/riscv/kvm/vm.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
arch/riscv/kvm/vm.c +2 -1
index a9f083f..66edfaa 100644
@@ -250,7 +250,8 @@ int kvm_vm_ioctl_enable_cap(struct kvm *kvm, struct kvm_enable_cap *cap)
 		mutex_lock(&kvm->lock);
 		mutex_lock(&kvm->slots_lock);
 
-		if (kvm->created_vcpus || !kvm_are_all_memslots_empty(kvm))
+		if (kvm->created_vcpus || !kvm_are_all_memslots_empty(kvm) ||
+		    kvm_riscv_aia_initialized(kvm))
 			r = -EBUSY;
 		else
 			kvm->arch.pgd_levels = new_levels;
Reviewed-by on v2 2/2
#4 lore raw

Re: [PATCH v2 2/2] RISC-V: KVM: Disallow GPA-width changes after AIA init

Quoted 10 lines · Pengpeng Hou

KVM_ENABLE_CAP(KVM_CAP_VM_GPA_BITS) allows userspace to reduce the stage-2 GPA width while a VM has no vCPUs or memory slots.

AIA initialization can complete with an APLIC and no vCPUs or memory slots. The APLIC MMIO device has then been registered, but the existing checks still allow userspace to shrink the GPA width below its address.

Reject GPA-width changes after AIA initialization. Both paths hold kvm->lock, serializing the width change with AIA initialization.

Quoted 6 lines

The issue was found by our static-analysis tool.

Fixes 7263b4fdb0b2 ("RISC-V: KVM: Reuse KVM_CAP_VM_GPA_BITS to select HGATP.MODE")
Assisted-by gpt 5

Thanks,
Nutty

Quoted 23 lines

---
Changes since v1: https://lore.kernel.org/r/4852f3b8985c353811fa065e4701d76dc9a7f086.1786512671.git.pengpeng@iscas.ac.cn/
- No change to the code; retain Anup's Reviewed-by.
- Rebase and clarify the locking description.

  arch/riscv/kvm/vm.c | 3 ++-
  1 file changed, 2 insertions(+), 1 deletion(-)
arch/riscv/kvm/vm.c +2 -1
index a9f083f..66edfaa 100644
@@ -250,7 +250,8 @@ int kvm_vm_ioctl_enable_cap(struct kvm *kvm, struct kvm_enable_cap *cap)
  		mutex_lock(&kvm->lock);
  		mutex_lock(&kvm->slots_lock);
 
-		if (kvm->created_vcpus || !kvm_are_all_memslots_empty(kvm))
+		if (kvm->created_vcpus || !kvm_are_all_memslots_empty(kvm) ||
+		    kvm_riscv_aia_initialized(kvm))
  			r = -EBUSY;
  		else
  			kvm->arch.pgd_levels = new_levels;
maintainer
#5 lore raw

Re: [PATCH v2 2/2] RISC-V: KVM: Disallow GPA-width changes after AIA init

In reply to Pengpeng Hou
Quoted 11 lines · Pengpeng Hou

KVM_ENABLE_CAP(KVM_CAP_VM_GPA_BITS) allows userspace to reduce the stage-2 GPA width while a VM has no vCPUs or memory slots.

AIA initialization can complete with an APLIC and no vCPUs or memory slots. The APLIC MMIO device has then been registered, but the existing checks still allow userspace to shrink the GPA width below its address.

Reject GPA-width changes after AIA initialization. Both paths hold kvm->lock, serializing the width change with AIA initialization.

Quoted 6 lines

The issue was found by our static-analysis tool.

Fixes 7263b4fdb0b2 ("RISC-V: KVM: Reuse KVM_CAP_VM_GPA_BITS to select HGATP.MODE")
Assisted-by gpt 5

Queued this patch as a fix for Linux-7.4-rcX

Thanks,
Anup

Quoted 23 lines

---
Changes since v1: https://lore.kernel.org/r/4852f3b8985c353811fa065e4701d76dc9a7f086.1786512671.git.pengpeng@iscas.ac.cn/
- No change to the code; retain Anup's Reviewed-by.
- Rebase and clarify the locking description.

 arch/riscv/kvm/vm.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
arch/riscv/kvm/vm.c +2 -1
index a9f083f..66edfaa 100644
@@ -250,7 +250,8 @@ int kvm_vm_ioctl_enable_cap(struct kvm *kvm, struct kvm_enable_cap *cap)
                mutex_lock(&kvm->lock);
                mutex_lock(&kvm->slots_lock);
 
-               if (kvm->created_vcpus || !kvm_are_all_memslots_empty(kvm))
+               if (kvm->created_vcpus || !kvm_are_all_memslots_empty(kvm) ||
+                   kvm_riscv_aia_initialized(kvm))
                        r = -EBUSY;
                else
                        kvm->arch.pgd_levels = new_levels;
#6 lore raw
In reply to Pengpeng Hou

Hi Anup,

A brief follow-up on this v2 series. Patch 1 preserves the required address checks and limits initialization revalidation to APLIC, as discussed. Patch 2 keeps the reviewed GPA-width freeze unchanged.

Is anything else needed before the series can be picked up?

Thanks,
Pengpeng

Signature · 3 lines
kvm-riscv mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/kvm-riscv