KVM: x86: avoid underflow when scaling TSC frequency
authorPaolo Bonzini <pbonzini@redhat.com>
Wed, 9 Jul 2025 17:45:14 +0000 (13:45 -0400)
committerPaolo Bonzini <pbonzini@redhat.com>
Wed, 9 Jul 2025 17:52:50 +0000 (13:52 -0400)
In function kvm_guest_time_update(), __scale_tsc() is used to calculate
a TSC *frequency* rather than a TSC value.  With low-enough ratios,
a TSC value that is less than 1 would underflow to 0 and to an infinite
while loop in kvm_get_time_scale():

  kvm_guest_time_update(struct kvm_vcpu *v)
    if (kvm_caps.has_tsc_control)
      tgt_tsc_khz = kvm_scale_tsc(tgt_tsc_khz,
                                  v->arch.l1_tsc_scaling_ratio);
        __scale_tsc(u64 ratio, u64 tsc)
          ratio=122380531, tsc=2299998, N=48
          ratio*tsc >> N = 0.999... -> 0

Later in the function:

  Call Trace:
   <TASK>
   kvm_get_time_scale arch/x86/kvm/x86.c:2458 [inline]
   kvm_guest_time_update+0x926/0xb00 arch/x86/kvm/x86.c:3268
   vcpu_enter_guest.constprop.0+0x1e70/0x3cf0 arch/x86/kvm/x86.c:10678
   vcpu_run+0x129/0x8d0 arch/x86/kvm/x86.c:11126
   kvm_arch_vcpu_ioctl_run+0x37a/0x13d0 arch/x86/kvm/x86.c:11352
   kvm_vcpu_ioctl+0x56b/0xe60 virt/kvm/kvm_main.c:4188
   vfs_ioctl fs/ioctl.c:51 [inline]
   __do_sys_ioctl fs/ioctl.c:871 [inline]
   __se_sys_ioctl+0x12d/0x190 fs/ioctl.c:857
   do_syscall_x64 arch/x86/entry/common.c:51 [inline]
   do_syscall_64+0x59/0x110 arch/x86/entry/common.c:81
   entry_SYSCALL_64_after_hwframe+0x78/0xe2

This can really happen only when fuzzing, since the TSC frequency
would have to be nonsensically low.

Fixes: 35181e86df97 ("KVM: x86: Add a common TSC scaling function")
Reported-by: Yuntao Liu <liuyuntao12@huawei.com>
Suggested-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
arch/x86/kvm/x86.c

index b58a74c1722de3f2d180cf8d6a3cf0ce9cf534fd..de51dbd85a58341bdad3a40a36ac04872602d5cc 100644 (file)
@@ -3258,9 +3258,11 @@ int kvm_guest_time_update(struct kvm_vcpu *v)
 
        /* With all the info we got, fill in the values */
 
-       if (kvm_caps.has_tsc_control)
+       if (kvm_caps.has_tsc_control) {
                tgt_tsc_khz = kvm_scale_tsc(tgt_tsc_khz,
                                            v->arch.l1_tsc_scaling_ratio);
+               tgt_tsc_khz = tgt_tsc_khz ? : 1;
+       }
 
        if (unlikely(vcpu->hw_tsc_khz != tgt_tsc_khz)) {
                kvm_get_time_scale(NSEC_PER_SEC, tgt_tsc_khz * 1000LL,