KVM: arm64: Restore the VM's feature bitmap when kvm_setup_vcpu() fails

__kvm_vcpu_set_target() copies the requested features into the VM-wide bitmap before kvm_setup_vcpu() runs and doesn't undo it when setup fails, so a rejected KVM_ARM_VCPU_INIT leaves the VM recording features that were never set up. With HAS_EL2 | HAS_EL2_E2H0 on a host without FEAT_NV1, kvm_vcpu_init_nested() returns -EINVAL before it allocates any nested stage-2 MMU, and vcpu_has_nv() is then true with nested_mmus_size == 0; its -ENOMEM paths do the same on a VM's first INIT.

Nothing in the tree loads a vCPU whose init failed, so this is latent. The pending KVM_PRE_FAULT_MEMORY series for arm64 does, and the second such load NULL-dereferences in get_s2_mmu_nested() under mmu_lock.

Setup reads the VM-wide bitmap, so the copy can't be deferred; restore the previous value instead when kvm_setup_vcpu() fails.

---
The series' generic kvm_vcpu_pre_fault_memory() calls vcpu_load() before any arm64 hook, so there is no arm64 check it can pass through first. Reproduced on kvmarm/next plus the series under
QEMU (-cpu max with an Apple M2 MIDR, which has_nv1() denies; kvm-arm.mode=nested): KVM_ARM_VCPU_INIT with HAS_EL2 | HAS_EL2_E2H0 returns -EINVAL, then KVM_PRE_FAULT_MEMORY twice. The first call loads with hw_mmu still the canonical MMU and its vcpu_put() clears hw_mmu; the second takes the !hw_mmu path into get_s2_mmu_nested(), whose search over nested_mmus_size == 0 leaves s2_mmu NULL for the BUG_ON(atomic_read(&s2_mmu->refcnt)):

  Unable to handle kernel NULL pointer dereference at virtual address 0000000000000074
  Call trace:
   kvm_vcpu_load_hw_mmu+0x94/0x2c0 (P)
   kvm_arch_vcpu_load+0x2a8/0x5e8
   kvm_vcpu_pre_fault_memory+0xa0/0x1b8
   kvm_vcpu_ioctl+0x40c/0x6d0

The thread dies with mmu_lock held for write and an RCU stall in queued_write_lock_slowpath() follows. With the fix both calls return -ENOENT and the host is unaffected. Applies unchanged to v7.3-rc3.

 arch/arm64/kvm/arm.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

index eaf583b771931..c2eb9b6da80f4 100644
@@ -1685,6 +1685,7 @@ static int kvm_setup_vcpu(struct kvm_vcpu *vcpu)
 static int __kvm_vcpu_set_target(struct kvm_vcpu *vcpu,
 				 const struct kvm_vcpu_init *init)
 {
+	DECLARE_BITMAP(old_features, KVM_VCPU_MAX_FEATURES);
 	unsigned long features = init->features[0];
 	struct kvm *kvm = vcpu->kvm;
 	int ret = -EINVAL;

@@ -1695,11 +1696,17 @@ static int __kvm_vcpu_set_target(struct kvm_vcpu *vcpu,
 	    kvm_vcpu_init_changed(vcpu, init))
 		goto out_unlock;
 
+	/* Setup reads the VM-wide bitmap, so undo the copy if setup fails. */
+	bitmap_copy(old_features, kvm->arch.vcpu_features,
+		    KVM_VCPU_MAX_FEATURES);
 	bitmap_copy(kvm->arch.vcpu_features, &features, KVM_VCPU_MAX_FEATURES);
 
 	ret = kvm_setup_vcpu(vcpu);
-	if (ret)
+	if (ret) {
+		bitmap_copy(kvm->arch.vcpu_features, old_features,
+			    KVM_VCPU_MAX_FEATURES);
 		goto out_unlock;
+	}
 
 	/* Now we know what it is, we can reset it. */
 	kvm_reset_vcpu(vcpu);

base-commit: 089e4f3c4862ba3f29dff2361caa8084879194fd

2.39.5

It seems obviously correct (TM) to undo setting the features on error.

I think this needs clarification :)

The KVM_PRE_FAULT_MEMORY series doesn't do this, or rather it only exposes the ability for the pre-existing generic kvm_vcpu_pre_fault_memory() function to be invoked by setting flags such as to expose pre-faulting as available for arm64.

So something like:

The upcoming series that enables KVM_PRE_FAULT_MEMORY exposes an existing bug in how vCPU initialisation is performed as it allows the generic kvm_vcpu_pre_fault_memory() function to be called which calls vcpu_load() before the vCPU might be initialised.

For that bit.

Also I think this is a bit unclear, esp. the prior reference to 'second', not clear what that refers to. I think you should bring in some of the context you provide under the fold for the patch.

So clarify what you do i.e.:

1. KVM_ARM_VCPU_INIT ioctl with an invalid configuration (HAS_EL2 | HAS_EL2_E2H0) - establishes the invalid kvm->arch.vcpu_features and doesn't clear down.

2. KVM_PRE_FAULT_MEMORY - This call is benign, but it calls vcpu_put() which calls kvm_arch_vcpu_put() and that has:

		if (vcpu_has_nv(vcpu))
			kvm_vcpu_put_hw_mmu(vcpu);

Since the broken config causes vcpu_has_nv() to return true, incorrectly, and causes:

  	vcpu->arch.hw_mmu = NULL;

To be run even though the vCPU in fact has the canonical MMU...

3. KVM_PRE_FAULT_MEMORY again - This time on load there's an issue:

	kvm_vcpu_pre_fault_memory()
	  -> vcpu_load()
	    -> kvm_arch_vcpu_load()

Which has:

if (vcpu_has_nv(vcpu))

		kvm_vcpu_load_hw_mmu(vcpu);

This leads to this path being taken:

		if (!vcpu->arch.hw_mmu) {
			scoped_guard(write_lock, &vcpu->kvm->mmu_lock)
				vcpu->arch.hw_mmu = get_s2_mmu_nested(vcpu);
		}

And in get_s2_mmu_nested() it searches through the non-existent nested mmus leaving s2_mmu NULL, resulting in a NULL pointer deref on this line:

BUG_ON(atomic_read(&s2_mmu->refcnt)); /* We have struct MMUs to spare */

Hmm are 2 fixes really needed? It seems to me that commit 427733579744e ("KVM: arm64: Select default PMU in KVM_ARM_VCPU_INIT handler") is the right one because that added a step _after_ setting the bitmap that can go wrong.

With the comments re: commit message/fixes tag addressed this seems correct and a very good spot thanks! We very much need this for the pre fault functionality.

I also think you should put the stack trace in the commit message, ideally decoded via scripts/decode_stacktrace.sh!

--
Cheers, Lorenzo

Hi Lorenzo,

On Fri, 18 Sep 2026 17:10:43 +0100, Lorenzo Stoakes (ARM) <[email protected]> wrote:
[...]

That's what "does" was standing in for. I'll say that the series enables the generic kvm_vcpu_pre_fault_memory(), which calls vcpu_load() whether or not the vCPU has been initialised.

[...]

I'll bring the two-ioctl walk up from below the fold: the first KVM_PRE_FAULT_MEMORY's vcpu_put() clearing hw_mmu, then the second's vcpu_load() reaching get_s2_mmu_nested() with no nested MMUs to find.

I'd added the first one because it's where the copy into the VM-wide bitmap moved ahead of the reset, but nothing after the copy could fail there. I'll drop it and keep 427733579744e.

[...]

Thanks!

[...]

Sure, will do.

Thanks,
/fuad