KVM: x86: fix kvm_write_tsc() TSC matching thinko
authorMarcelo Tosatti <mtosatti@redhat.com>
Thu, 8 Mar 2012 21:46:57 +0000 (18:46 -0300)
committerAvi Kivity <avi@redhat.com>
Tue, 20 Mar 2012 10:40:36 +0000 (12:40 +0200)
kvm_write_tsc() converts from guest TSC to microseconds, not nanoseconds
as intended. The result is that the window for matching is 1000 seconds,
not 1 second.

Microsecond precision is enough for checking whether the TSC write delta
is within the heuristic values, so use it instead of nanoseconds.

Noted by Avi Kivity.

Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
arch/x86/kvm/x86.c

index 32096cf6c6c9c6a231c2d3de6c2368842f7f6b36..7287812eeb729745ba0db6f0be2ec4a6f2a025a3 100644 (file)
@@ -1025,7 +1025,7 @@ void kvm_write_tsc(struct kvm_vcpu *vcpu, u64 data)
        struct kvm *kvm = vcpu->kvm;
        u64 offset, ns, elapsed;
        unsigned long flags;
-       s64 nsdiff;
+       s64 usdiff;
 
        raw_spin_lock_irqsave(&kvm->arch.tsc_write_lock, flags);
        offset = kvm_x86_ops->compute_tsc_offset(vcpu, data);
@@ -1033,18 +1033,19 @@ void kvm_write_tsc(struct kvm_vcpu *vcpu, u64 data)
        elapsed = ns - kvm->arch.last_tsc_nsec;
 
        /* n.b - signed multiplication and division required */
-       nsdiff = data - kvm->arch.last_tsc_write;
+       usdiff = data - kvm->arch.last_tsc_write;
 #ifdef CONFIG_X86_64
-       nsdiff = (nsdiff * 1000) / vcpu->arch.virtual_tsc_khz;
+       usdiff = (usdiff * 1000) / vcpu->arch.virtual_tsc_khz;
 #else
        /* do_div() only does unsigned */
        asm("idivl %2; xor %%edx, %%edx"
-           : "=A"(nsdiff)
-           : "A"(nsdiff * 1000), "rm"(vcpu->arch.virtual_tsc_khz));
+           : "=A"(usdiff)
+           : "A"(usdiff * 1000), "rm"(vcpu->arch.virtual_tsc_khz));
 #endif
-       nsdiff -= elapsed;
-       if (nsdiff < 0)
-               nsdiff = -nsdiff;
+       do_div(elapsed, 1000);
+       usdiff -= elapsed;
+       if (usdiff < 0)
+               usdiff = -usdiff;
 
        /*
         * Special case: TSC write with a small delta (1 second) of virtual
@@ -1056,7 +1057,7 @@ void kvm_write_tsc(struct kvm_vcpu *vcpu, u64 data)
         * compensation code attempt to catch up if we fall behind, but
         * it's better to try to match offsets from the beginning.
          */
-       if (nsdiff < NSEC_PER_SEC &&
+       if (usdiff < USEC_PER_SEC &&
            vcpu->arch.virtual_tsc_khz == kvm->arch.last_tsc_khz) {
                if (!check_tsc_unstable()) {
                        offset = kvm->arch.cur_tsc_offset;