On 1/8/2022 11:43 pm, Sean Christopherson wrote:
Not directly related to this patch...
Unless I've missed something, every invocation of start_event() and measure() first
sets evt.count=0. Rather than force every caller to ensure count is zeroed, why not
zero the count during start_event() and then drop all of the manual zeroing?
None object to this idea, after all, there is obvious redundancy here.
diff --git a/x86/pmu.c b/x86/pmu.c
index 01be1e90..ef804272 100644
--- a/x86/pmu.c
+++ b/x86/pmu.c
@@ -141,7 +141,7 @@ static void global_disable(pmu_counter_t *cnt)
static void start_event(pmu_counter_t *evt)
{
- wrmsr(evt->ctr, evt->count);
+ wrmsr(evt->ctr, 0);
Now we have to fix the last call to measure() in check_counter_overflow(), since
it will
also call start_event() after it has been modified and in that case, the
requested high count
has to be passed in from another function parameter.
Also, the naming of start_event() does not imply that the counter will be set to
zero implicitly,
it just lets a counter continue to run, not caring about the current value of
the counter,
which is more flexible.
I may try to do that on the test-cases of AMD vPMU, to help verify the gain of
your idea.
if (is_gp(evt))
wrmsr(MSR_P6_EVNTSEL0 + event_to_global_idx(evt),
evt->config | EVNTSEL_EN);
Accumulating counts can be handled by reading the current count before start_event(),
and doing something like stuffing a high count to test an edge case could be handled
by an inner helper, e.g. by adding __start_event().
