From: Christian Ehrhardt <ehrhardt@xxxxxxxxxxxxxxxxxx>
The current kvmtrace code uses get_cycles() while the interpretation would be
easier using using nanoseconds. Getnstimeofday should give the same accuracy
as get_cycles on all architectures but at a better unit (e.g. comparable
between two hosts with different frequencies.
The implementation of kvmtrace uses a 64 bit cycle variable while get_cycles
only provided a unsigned long which is != 64 bit on 32 bit architectures.
Since this patch introduced getnstimeofday we can addtionally merge the
tv_sec portion of the struct timespec returned to use the full 64 bit
of the trace interface and therefor avoid some wraparounds. This merge
is done always if the two values of a struct timespec fit into the u64.
An alternative might be ktime_t instead of u64 and using as Sheng Yang
suggested and ktime_get(), I had no time to test&compare that yet.
Any comments or requirements from others what to prefer here ?
Signed-off-by: Christian Ehrhardt <ehrhardt@xxxxxxxxxxxxxxxxxx>
---
[diffstat]
kvm_trace.c | 9 ++++++++-
1 file changed, 8 insertions(+), 1 deletion(-)
[diff]
diff --git a/virt/kvm/kvm_trace.c b/virt/kvm/kvm_trace.c
--- a/virt/kvm/kvm_trace.c
+++ b/virt/kvm/kvm_trace.c
@@ -56,6 +56,7 @@
struct kvm_vcpu *vcpu;
int i, size;
u32 extra;
+ struct timespec tv;
if (unlikely(kt->trace_state != KVM_TRACE_STATE_RUNNING))
return;
@@ -73,7 +74,13 @@
| TRACE_REC_NUM_DATA_ARGS(extra);
if (p->cycle_in) {
- rec.u.cycle.cycle_u64 = get_cycles();
+ getnstimeofday(&tv);
+ /* compress both struct tv values into u64 on 32 bit archs */
+ if (sizeof(tv) == sizeof(u64))
+ rec.u.cycle.cycle_u64 =
+ (((u64)tv.tv_sec) << 32) | tv.tv_nsec;
+ else
+ rec.u.cycle.cycle_u64 = tv.tv_nsec;
for (i = 0; i < extra; i++)
rec.u.cycle.extra_u32[i] = va_arg(*args, u32);
--
To unsubscribe from this list: send the line "unsubscribe kvm-ppc" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
