KVM: arm64: Reallocate the nested_mmus array under the mmu_lock

Code that walks kvm->arch.nested_mmus[] holds kvm->mmu_lock. By contrast, kvm_vcpu_init_nested() reallocates the array and frees the old buffer while holding only kvm->arch.config_lock, so a walker can reference the freed array.

Allocate the new array outside the lock, as the allocation can sleep, and do only the copy and the pointer swap under the mmu_lock. After the swap no walker can reach the old buffer, so free it once the lock has been released.

Fixes 4f128f8e1aaac ("KVM: arm64: nv: Support multiple nested Stage-2 mmu structures")

Signed-off-by Hyunwoo Kim <[email protected]>

---
 arch/arm64/kvm/nested.c | 33 ++++++++++++++++++++-------------
 1 file changed, 20 insertions(+), 13 deletions(-)

arch/arm64/kvm/nested.c +20 -13

index 38f672e940878..6f7bc9a9992e0 100644
@@ -89,21 +89,28 @@ int kvm_vcpu_init_nested(struct kvm_vcpu *vcpu)
 	 * again, and there is no reason to affect the whole VM for this.
 	 */
 	num_mmus = atomic_read(&kvm->online_vcpus) * S2_MMU_PER_VCPU;
-	tmp = kvrealloc(kvm->arch.nested_mmus,
-			size_mul(sizeof(*kvm->arch.nested_mmus), num_mmus),
-			GFP_KERNEL_ACCOUNT | __GFP_ZERO);
-	if (!tmp)
-		return -ENOMEM;
 
-	swap(kvm->arch.nested_mmus, tmp);
+	if (num_mmus > kvm->arch.nested_mmus_size) {
+		tmp = kvcalloc(num_mmus, sizeof(*tmp), GFP_KERNEL_ACCOUNT);
+		if (!tmp)
+			return -ENOMEM;
 
-	/*
-	 * If we went through a realocation, adjust the MMU back-pointers in
-	 * the previously initialised kvm_pgtable structures.
-	 */
-	if (kvm->arch.nested_mmus != tmp)
-		for (int i = 0; i < kvm->arch.nested_mmus_size; i++)
-			kvm->arch.nested_mmus[i].pgt->mmu = &kvm->arch.nested_mmus[i];
+		write_lock(&kvm->mmu_lock);
+
+		if (kvm->arch.nested_mmus_size) {
+			memcpy(tmp, kvm->arch.nested_mmus,
+			       size_mul(sizeof(*tmp), kvm->arch.nested_mmus_size));
+
+			for (int i = 0; i < kvm->arch.nested_mmus_size; i++)
+				tmp[i].pgt->mmu = &tmp[i];
+		}
+
+		swap(kvm->arch.nested_mmus, tmp);
+
+		write_unlock(&kvm->mmu_lock);
+
+		kvfree(tmp);
+	}
 
 	for (int i = kvm->arch.nested_mmus_size; !ret && i < num_mmus; i++)
 		ret = init_nested_s2_mmu(kvm, &kvm->arch.nested_mmus[i]);

Signature · 1 line

2.43.0

The shortlog is very confusing, since "allocate behind $LOCK" is usually something alarming. Maybe instead:

KVM: arm64: Reassign nested_mmus array behind mmu_lock

Hi,

The shortlog is very confusing, since "allocate behind $LOCK" is usually something alarming. Maybe instead:

KVM: arm64: Reassign nested_mmus array behind mmu_lock

Quoted 4 lines

Code that walks kvm->arch.nested_mmus[] holds kvm->mmu_lock. By contrast, kvm_vcpu_init_nested() reallocates the array and frees the old buffer while holding only kvm->arch.config_lock, so a walker can reference the freed array.

It wouldn't hurt to share slightly more information here. Are you dealing with a concurrent MMU notifier?

Quoted 6 lines

Allocate the new array outside the lock, as the allocation can sleep, and do only the copy and the pointer swap under the mmu_lock. After the swap no walker can reach the old buffer, so free it once the lock has been released.

Quoted 2 lines

Fixes 4f128f8e1aaac ("KVM: arm64: nv: Support multiple nested Stage-2 mmu structures")

Signed-off-by Hyunwoo Kim <[email protected]>

The diff itself LGTM

Reviewed-by Oliver Upton <[email protected]>

Thanks,
Oliver

Quoted 48 lines

---
 arch/arm64/kvm/nested.c | 33 ++++++++++++++++++++-------------
 1 file changed, 20 insertions(+), 13 deletions(-)

arch/arm64/kvm/nested.c +20 -13

index 38f672e940878..6f7bc9a9992e0 100644
@@ -89,21 +89,28 @@ int kvm_vcpu_init_nested(struct kvm_vcpu *vcpu)
 	 * again, and there is no reason to affect the whole VM for this.
 	 */
 	num_mmus = atomic_read(&kvm->online_vcpus) * S2_MMU_PER_VCPU;
-	tmp = kvrealloc(kvm->arch.nested_mmus,
-			size_mul(sizeof(*kvm->arch.nested_mmus), num_mmus),
-			GFP_KERNEL_ACCOUNT | __GFP_ZERO);
-	if (!tmp)
-		return -ENOMEM;
 
-	swap(kvm->arch.nested_mmus, tmp);
+	if (num_mmus > kvm->arch.nested_mmus_size) {
+		tmp = kvcalloc(num_mmus, sizeof(*tmp), GFP_KERNEL_ACCOUNT);
+		if (!tmp)
+			return -ENOMEM;
 
-	/*
-	 * If we went through a realocation, adjust the MMU back-pointers in
-	 * the previously initialised kvm_pgtable structures.
-	 */
-	if (kvm->arch.nested_mmus != tmp)
-		for (int i = 0; i < kvm->arch.nested_mmus_size; i++)
-			kvm->arch.nested_mmus[i].pgt->mmu = &kvm->arch.nested_mmus[i];
+		write_lock(&kvm->mmu_lock);
+
+		if (kvm->arch.nested_mmus_size) {
+			memcpy(tmp, kvm->arch.nested_mmus,
+			       size_mul(sizeof(*tmp), kvm->arch.nested_mmus_size));
+
+			for (int i = 0; i < kvm->arch.nested_mmus_size; i++)
+				tmp[i].pgt->mmu = &tmp[i];
+		}
+
+		swap(kvm->arch.nested_mmus, tmp);
+
+		write_unlock(&kvm->mmu_lock);
+
+		kvfree(tmp);
+	}
 

Quoted 5 lines

arch/arm64/kvm/nested.c +0 -0

 	for (int i = kvm->arch.nested_mmus_size; !ret && i < num_mmus; i++)
 		ret = init_nested_s2_mmu(kvm, &kvm->arch.nested_mmus[i]);

Signature · 1 line

2.43.0

Quoted 6 lines · Oliver Upton

Hi,

The shortlog is very confusing, since "allocate behind $LOCK" is usually something alarming. Maybe instead:

KVM: arm64: Reassign nested_mmus array behind mmu_lock

heh, that's confusing indeed. I'll change it that way.

Quoted 9 lines · Hyunwoo Kim

Quoted 4 lines

Code that walks kvm->arch.nested_mmus[] holds kvm->mmu_lock. By contrast, kvm_vcpu_init_nested() reallocates the array and frees the old buffer while holding only kvm->arch.config_lock, so a walker can reference the freed array.

It wouldn't hurt to share slightly more information here. Are you dealing with a concurrent MMU notifier?

Yes. The MMU notifier path also walks nested_mmus[] under mmu_lock. kvm_vcpu_init_nested() holds only config_lock, so if a notifier fires during vCPU init, it races with the array realloc and free.

Here's the reworked changelog. Should I send v2?

kvm->arch.nested_mmus[] is walked under kvm->mmu_lock, including from the MMU notifier path (kvm_unmap_gfn_range() -> kvm_nested_s2_unmap()), which can run at any time. kvm_vcpu_init_nested() reallocates the array and frees the old buffer while holding only kvm->arch.config_lock, so such a walker can reference the freed array.

Allocate the new array outside of mmu_lock, as the allocation can sleep. Under the lock, copy the existing entries, fix up the back pointers and reassign the array. Free the old buffer after dropping the lock, as kvfree() can sleep as well.