On 14/10/2016 12:05, Wanpeng Li wrote: > 2016-10-13 19:34 GMT+08:00 Wanpeng Li <kernel...@gmail.com>: > [...] >> >> The patchset reduces ~300+ clock cycles for each APIC timer oneshot mode >> operation virtualization. However, the performance of periodic mode is >> still bad, so this version is still a RFC. Your comments to improve the >> patchset is a great appreciated. > > I observed that the clock cycles between the start of > kvm_lapic_expired_hv_timer() (periodic mode emulated through restart a > preemption timer after the last preemption timer vmexit) to > kvm_inject_apic_timer_irqs() is almost half of the clock cycles > between the start of apic_timer_fn() (periodic mode emulated through > hrtimer) to kvm_inject_apic_timers_irqs(), so the overhead of > preemption timer is lower in this path. Maybe something is still not > correct in other places which results in performance of periodic mode > emulated by VMX preemption timer is still not good. Any help is a > great appreciated. :) > > Radim, Paolo, ping :)
Just post the patches and we'll see. :) Another useful thing to do is to prepare testcases similar to tscdeadline_latency, but using the periodic and oneshot modes. Paolo
