On 2020/2/25 8:08, Eric Hankland wrote:
Hi Like - Thanks for the feedback - is your recommendation to do the read and period change at the same time and only take the lock once or is there another way around this while still handling writes correctly?
For non-running counters(the most common situation), we have too many chances to reflect their new periods. In this case, calling perf_event_period() in the trap of counter msr is redundant and burdensome. A better way is to check if this counter is running via pmc_speculative_in_use (), and if so, just trigger kvm_make_request(KVM_REQ_PMU, pmc->vcpu). Thanks, Like Xu
Eric
