Kai, On Wed, Dec 16 2020 at 22:18, shenkai wrote: > After some tests, the conclusion that time cost is from deep C-state > turns out to be wrong > > Sorry for that.
No problem. > In kexec case, first let APs spinwait like what I didĀ in that patch, > but wake APs up by sending apic INIT and SIPIĀ interrupts as normal > procedure instead of writing to some address and there is no > acceleration (time cost is still 210ms). Ok. > So can we say that the main time cost is from apic INIT and SIPI > interrupts and the handling of them instead of deep C-state? That's a fair conclusion. > I didn't test with play_dead() because in kexec case, one new kernel > will be started and APs can't be waken up by normal interrupts like in > hibernate case for the irq vectors are gone with the old kernel. > > Or maybe I didn't get the point correctly? Not exactly, but your experiment answered the question already. My point was that the regular kexec unplugs the APs which then end up in play_dead() and trying to use the deepest C-state via mwait(). So if the overhead would be related to getting them out of a deep C-state then forcing that play_dead() to use the HLT instruction or the most shallow C-state with mwait() would have brought an improvement, right? But obviously the C-state in which the APs are waiting is not really relevant, as you demonstrated that the cost is due to INIT/SIPI even with spinwait, which is what I suspected. OTOH, the advantage of INIT/SIPI is that the AP comes up in a well known state. Thanks, tglx