On Tue, 30 Sep 2025 09:52:37 +0100, Oliver Upton oliver.upton@linux.dev wrote:

Another day, another syzkaller bug. KVM erroneously allows userspace to pend vCPU events for a vCPU that hasn't been initialized yet, leading to KVM interpreting a bunch of uninitialized garbage for routing / injecting the exception.

In one case the injection code and the hyp disagree on whether the vCPU has a 32bit EL1 and put the vCPU into an illegal mode for AArch64, tripping the BUG() in exception_target_el() during the next injection:

kernel BUG at arch/arm64/kvm/inject_fault.c:40! Internal error: Oops - BUG: 00000000f2000800 [#1] SMP CPU: 3 UID: 0 PID: 318 Comm: repro Not tainted 6.17.0-rc4-00104-g10fd0285305d #6 PREEMPT Hardware name: linux,dummy-virt (DT) pstate: 21402009 (nzCv daif +PAN -UAO -TCO +DIT -SSBS BTYPE=--) pc : exception_target_el+0x88/0x8c lr : pend_serror_exception+0x18/0x13c sp : ffff800082f03a10 x29: ffff800082f03a10 x28: ffff0000cb132280 x27: 0000000000000000 x26: 0000000000000000 x25: ffff0000c2a99c20 x24: 0000000000000000 x23: 0000000000008000 x22: 0000000000000002 x21: 0000000000000004 x20: 0000000000008000 x19: ffff0000c2a99c20 x18: 0000000000000000 x17: 0000000000000000 x16: 0000000000000000 x15: 00000000200000c0 x14: 0000000000000000 x13: 0000000000000000 x12: 0000000000000000 x11: 0000000000000000 x10: 0000000000000000 x9 : 0000000000000000 x8 : ffff800082f03af8 x7 : 0000000000000000 x6 : 0000000000000000 x5 : ffff800080f621f0 x4 : 0000000000000000 x3 : 0000000000000000 x2 : 000000000040009b x1 : 0000000000000003 x0 : ffff0000c2a99c20 Call trace: exception_target_el+0x88/0x8c (P) kvm_inject_serror_esr+0x40/0x3b4 __kvm_arm_vcpu_set_events+0xf0/0x100 kvm_arch_vcpu_ioctl+0x180/0x9d4 kvm_vcpu_ioctl+0x60c/0x9f4 __arm64_sys_ioctl+0xac/0x104 invoke_syscall+0x48/0x110 el0_svc_common.constprop.0+0x40/0xe0 do_el0_svc+0x1c/0x28 el0_svc+0x34/0xf0 el0t_64_sync_handler+0xa0/0xe4 el0t_64_sync+0x198/0x19c Code: f946bc01 b4fffe61 9101e020 17fffff2 (d4210000)

Reject the ioctls outright as no sane VMM would call these before KVM_ARM_VCPU_INIT anyway. Even if it did the exception would've been thrown away by the eventual reset of the vCPU's state.

Cc: stable@vger.kernel.org # 6.17 Fixes: b7b27facc7b5 ("arm/arm64: KVM: Add KVM_GET/SET_VCPU_EVENTS") Signed-off-by: Oliver Upton oliver.upton@linux.dev

---

While the blamed commit is indeed broken, only 6.17+ kernels actually hit the BUG() due to commit efa1368ba9f4 ("KVM: arm64: Commit exceptions from KVM_SET_VCPU_EVENTS immediately).

arch/arm64/kvm/arm.c | 6 ++++++ 1 file changed, 6 insertions(+)

diff --git a/arch/arm64/kvm/arm.c b/arch/arm64/kvm/arm.c index a59b4046617c..c44357d26ee8 100644 --- a/arch/arm64/kvm/arm.c +++ b/arch/arm64/kvm/arm.c @@ -1795,6 +1795,9 @@ long kvm_arch_vcpu_ioctl(struct file *filp, case KVM_GET_VCPU_EVENTS: { struct kvm_vcpu_events events;

• if (!kvm_vcpu_initialized(vcpu))
  

• 	return -ENOEXEC;
  

• if (kvm_arm_vcpu_get_events(vcpu, &events)) return -EINVAL;

@@ -1806,6 +1809,9 @@ long kvm_arch_vcpu_ioctl(struct file *filp, case KVM_SET_VCPU_EVENTS: { struct kvm_vcpu_events events;

• if (!kvm_vcpu_initialized(vcpu))
  

• 	return -ENOEXEC;
  

• if (copy_from_user(&events, argp, sizeof(events))) return -EFAULT;

LGTM, but the -ENOEXEC part will eventually need being documented. I'll queue this now though, as it fixes an annoying bug.

Thanks,

M.