Skip to contentSkip to conversation

kvm: Ensure writes to the coalesced MMIO ring are within bounds

PATCH In review Will Deacon 3 participants 5 messages Sep 18, 2019 KVM

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 kvm: Ensure writes to the coalesced MMIO ring are within bounds

When records are written to the coalesced MMIO ring in response to a vCPU MMIO exit, the 'ring->last' field is used to index the ring buffer page. Although we hold the 'kvm->ring_lock' at this point, the ring structure is mapped directly into the host userspace and can therefore be modified to point at arbitrary pages within the kernel.

Since this shouldn't happen in normal operation, simply bound the index by KVM_COALESCED_MMIO_MAX to contain the accesses within the ring buffer page.

Cc Greg Kroah-Hartman <[email protected]>
Cc Paolo Bonzini <[email protected]>
Cc <[email protected]> # 5.2.y
Fixes 5f94c1741bdc ("KVM: Add coalesced MMIO support (common part)")

---

I think there are some other fixes kicking around for this, but they still rely on 'ring->last' being stable, which isn't necessarily the case. I'll send the -stable backport for kernels prior to 5.2 once this hits mainline.

 virt/kvm/coalesced_mmio.c | 13 +++++++------
 1 file changed, 7 insertions(+), 6 deletions(-)
virt/kvm/coalesced_mmio.c +7 -6
index 5294abb3f178..09b3e4421550 100644
@@ -67,6 +67,7 @@ static int coalesced_mmio_write(struct kvm_vcpu *vcpu,
 {
 	struct kvm_coalesced_mmio_dev *dev = to_mmio(this);
 	struct kvm_coalesced_mmio_ring *ring = dev->kvm->coalesced_mmio_ring;
+	u32 last;
 
 	if (!coalesced_mmio_in_range(dev, addr, len))
 		return -EOPNOTSUPP;
@@ -79,13 +80,13 @@ static int coalesced_mmio_write(struct kvm_vcpu *vcpu,
 	}
 
 	/* copy data in first free entry of the ring */
-
-	ring->coalesced_mmio[ring->last].phys_addr = addr;
-	ring->coalesced_mmio[ring->last].len = len;
-	memcpy(ring->coalesced_mmio[ring->last].data, val, len);
-	ring->coalesced_mmio[ring->last].pio = dev->zone.pio;
+	last = ring->last % KVM_COALESCED_MMIO_MAX;
+	ring->coalesced_mmio[last].phys_addr = addr;
+	ring->coalesced_mmio[last].len = len;
+	memcpy(ring->coalesced_mmio[last].data, val, len);
+	ring->coalesced_mmio[last].pio = dev->zone.pio;
 	smp_wmb();
-	ring->last = (ring->last + 1) % KVM_COALESCED_MMIO_MAX;
+	ring->last = (last + 1) % KVM_COALESCED_MMIO_MAX;
 	spin_unlock(&dev->kvm->ring_lock);
 	return 0;
 }
Signature · 1 line
2.23.0.351.gc4317032e6-goog
Reviewed-by on v1 1/1
#1 lore raw
Quoted 48 lines · Will Deacon

When records are written to the coalesced MMIO ring in response to a vCPU MMIO exit, the 'ring->last' field is used to index the ring buffer page. Although we hold the 'kvm->ring_lock' at this point, the ring structure is mapped directly into the host userspace and can therefore be modified to point at arbitrary pages within the kernel.

Since this shouldn't happen in normal operation, simply bound the index by KVM_COALESCED_MMIO_MAX to contain the accesses within the ring buffer page.

Cc Greg Kroah-Hartman <[email protected]>
Cc Paolo Bonzini <[email protected]>
Cc <[email protected]> # 5.2.y
Fixes 5f94c1741bdc ("KVM: Add coalesced MMIO support (common part)")

---

I think there are some other fixes kicking around for this, but they still rely on 'ring->last' being stable, which isn't necessarily the case. I'll send the -stable backport for kernels prior to 5.2 once this hits mainline.

 virt/kvm/coalesced_mmio.c | 13 +++++++------
 1 file changed, 7 insertions(+), 6 deletions(-)
virt/kvm/coalesced_mmio.c +2 -5
index 5294abb3f178..09b3e4421550 100644
@@ -67,6 +67,7 @@ static int coalesced_mmio_write(struct kvm_vcpu *vcpu,
 {
 	struct kvm_coalesced_mmio_dev *dev = to_mmio(this);
 	struct kvm_coalesced_mmio_ring *ring = dev->kvm->coalesced_mmio_ring;
+	u32 last;
 
 	if (!coalesced_mmio_in_range(dev, addr, len))
 		return -EOPNOTSUPP;
@@ -79,13 +80,13 @@ static int coalesced_mmio_write(struct kvm_vcpu *vcpu,
 	}
 
 	/* copy data in first free entry of the ring */
-
-	ring->coalesced_mmio[ring->last].phys_addr = addr;
-	ring->coalesced_mmio[ring->last].len = len;
-	memcpy(ring->coalesced_mmio[ring->last].data, val, len);
-	ring->coalesced_mmio[ring->last].pio = dev->zone.pio;
+	last = ring->last % KVM_COALESCED_MMIO_MAX;

virt/kvm/coalesced_mmio.c static int coalesced_mmio_write(struct kvm_vcpu *vcpu, · line 92

Quoted 10 lines
virt/kvm/coalesced_mmio.c +5 -1
+	ring->coalesced_mmio[last].phys_addr = addr;
+	ring->coalesced_mmio[last].len = len;
+	memcpy(ring->coalesced_mmio[last].data, val, len);
+	ring->coalesced_mmio[last].pio = dev->zone.pio;
 	smp_wmb();
-	ring->last = (ring->last + 1) % KVM_COALESCED_MMIO_MAX;
+	ring->last = (last + 1) % KVM_COALESCED_MMIO_MAX;
 	spin_unlock(&dev->kvm->ring_lock);
 	return 0;
 }
maintainer
#2 lore raw
In reply to Will Deacon
Quoted 16 lines · Will Deacon

When records are written to the coalesced MMIO ring in response to a vCPU MMIO exit, the 'ring->last' field is used to index the ring buffer page. Although we hold the 'kvm->ring_lock' at this point, the ring structure is mapped directly into the host userspace and can therefore be modified to point at arbitrary pages within the kernel.

Since this shouldn't happen in normal operation, simply bound the index by KVM_COALESCED_MMIO_MAX to contain the accesses within the ring buffer page.

Cc Greg Kroah-Hartman <[email protected]>
Cc Paolo Bonzini <[email protected]>
Cc <[email protected]> # 5.2.y
Fixes 5f94c1741bdc ("KVM: Add coalesced MMIO support (common part)")
Quoted 6 lines

---

I think there are some other fixes kicking around for this, but they still rely on 'ring->last' being stable, which isn't necessarily the case. I'll send the -stable backport for kernels prior to 5.2 once this hits mainline.

Google's patch, which checks if ring->last is not in range and fails with -EOPNOTSUPP if not, is slightly better. I'll send it in a second and Cc you (and also send it as a pull request to Linus).

Paolo

Quoted 36 lines
 virt/kvm/coalesced_mmio.c | 13 +++++++------
 1 file changed, 7 insertions(+), 6 deletions(-)
virt/kvm/coalesced_mmio.c +7 -6
index 5294abb3f178..09b3e4421550 100644
@@ -67,6 +67,7 @@ static int coalesced_mmio_write(struct kvm_vcpu *vcpu,
 {
 	struct kvm_coalesced_mmio_dev *dev = to_mmio(this);
 	struct kvm_coalesced_mmio_ring *ring = dev->kvm->coalesced_mmio_ring;
+	u32 last;
 
 	if (!coalesced_mmio_in_range(dev, addr, len))
 		return -EOPNOTSUPP;
@@ -79,13 +80,13 @@ static int coalesced_mmio_write(struct kvm_vcpu *vcpu,
 	}
 
 	/* copy data in first free entry of the ring */
-
-	ring->coalesced_mmio[ring->last].phys_addr = addr;
-	ring->coalesced_mmio[ring->last].len = len;
-	memcpy(ring->coalesced_mmio[ring->last].data, val, len);
-	ring->coalesced_mmio[ring->last].pio = dev->zone.pio;
+	last = ring->last % KVM_COALESCED_MMIO_MAX;
+	ring->coalesced_mmio[last].phys_addr = addr;
+	ring->coalesced_mmio[last].len = len;
+	memcpy(ring->coalesced_mmio[last].data, val, len);
+	ring->coalesced_mmio[last].pio = dev->zone.pio;
 	smp_wmb();
-	ring->last = (ring->last + 1) % KVM_COALESCED_MMIO_MAX;
+	ring->last = (last + 1) % KVM_COALESCED_MMIO_MAX;
 	spin_unlock(&dev->kvm->ring_lock);
 	return 0;
 }
#3 lore raw
Quoted 19 lines · Paolo Bonzini
Quoted 18 lines · Will Deacon

When records are written to the coalesced MMIO ring in response to a vCPU MMIO exit, the 'ring->last' field is used to index the ring buffer page. Although we hold the 'kvm->ring_lock' at this point, the ring structure is mapped directly into the host userspace and can therefore be modified to point at arbitrary pages within the kernel.

Since this shouldn't happen in normal operation, simply bound the index by KVM_COALESCED_MMIO_MAX to contain the accesses within the ring buffer page.

Cc Greg Kroah-Hartman <[email protected]>
Cc Paolo Bonzini <[email protected]>
Cc <[email protected]> # 5.2.y
Fixes 5f94c1741bdc ("KVM: Add coalesced MMIO support (common part)")

---

Quoted 8 lines
Quoted 4 lines

I think there are some other fixes kicking around for this, but they still rely on 'ring->last' being stable, which isn't necessarily the case. I'll send the -stable backport for kernels prior to 5.2 once this hits mainline.

Google's patch, which checks if ring->last is not in range and fails with -EOPNOTSUPP if not, is slightly better. I'll send it in a second and Cc you (and also send it as a pull request to Linus).

Okey doke, as long as it gets fixed! My minor concerns with the error-checking variant are:

* Whether or not you need a READ_ONCE to prevent the compiler potentially

    reloading 'ring->last' after validation

* Whether or not this could be part of a spectre-v1 gadget

so, given that I don't think the malicious host deserves an error code if it starts writing the 'last' index, I went with the "obviously safe" version. But up to you.

Will

maintainer
#4 lore raw
Quoted 5 lines · Will Deacon

Okey doke, as long as it gets fixed! My minor concerns with the error-checking variant are:

* Whether or not you need a READ_ONCE to prevent the compiler potentially

    reloading 'ring->last' after validation

Yes, it certainly needs READ_ONCE. I had already added it locally, indeed.

Quoted 1 line

* Whether or not this could be part of a spectre-v1 gadget

I think not, the spectrev1 gadget require a load that is indexed from the contents of ring->coalesced_mmio[insert], but there's none.

Paolo

Quoted 6 lines

so, given that I don't think the malicious host deserves an error code if it starts writing the 'last' index, I went with the "obviously safe" version. But up to you.

Will