From owner-svn-src-stable-9@FreeBSD.ORG Thu Apr 12 06:29:03 2012 Return-Path: Delivered-To: svn-src-stable-9@freebsd.org Received: from mx1.freebsd.org (mx1.freebsd.org [IPv6:2001:4f8:fff6::34]) by hub.freebsd.org (Postfix) with ESMTP id 4BB9B106564A; Thu, 12 Apr 2012 06:29:03 +0000 (UTC) (envelope-from mav@FreeBSD.org) Received: from svn.freebsd.org (svn.freebsd.org [IPv6:2001:4f8:fff6::2c]) by mx1.freebsd.org (Postfix) with ESMTP id 34D5C8FC15; Thu, 12 Apr 2012 06:29:03 +0000 (UTC) Received: from svn.freebsd.org (localhost [127.0.0.1]) by svn.freebsd.org (8.14.4/8.14.4) with ESMTP id q3C6T3jt014098; Thu, 12 Apr 2012 06:29:03 GMT (envelope-from mav@svn.freebsd.org) Received: (from mav@localhost) by svn.freebsd.org (8.14.4/8.14.4/Submit) id q3C6T3rX014095; Thu, 12 Apr 2012 06:29:03 GMT (envelope-from mav@svn.freebsd.org) Message-Id: <201204120629.q3C6T3rX014095@svn.freebsd.org> From: Alexander Motin Date: Thu, 12 Apr 2012 06:29:02 +0000 (UTC) To: src-committers@freebsd.org, svn-src-all@freebsd.org, svn-src-stable@freebsd.org, svn-src-stable-9@freebsd.org X-SVN-Group: stable-9 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Cc: Subject: svn commit: r234167 - in stable/9: share/man/man4 sys/kern X-BeenThere: svn-src-stable-9@freebsd.org X-Mailman-Version: 2.1.5 Precedence: list List-Id: SVN commit messages for only the 9-stable src tree List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Thu, 12 Apr 2012 06:29:03 -0000 Author: mav Date: Thu Apr 12 06:29:02 2012 New Revision: 234167 URL: http://svn.freebsd.org/changeset/base/234167 Log: MFC 232919: Add kern.eventtimer.activetick tunable/sysctl, specifying whether each hardclock() tick should be run on every active CPU, or on only one. On my tests, avoiding extra interrupts because of this on 8-CPU Core i7 system with HZ=10000 saves about 2% of performance. At this moment option implemented only for global timers, as reprogramming per-CPU timers is too expensive now to be compensated by this benefit, especially since we still have to regularly run hardclock() on at least one active CPU to update system uptime. For global timer it is quite trivial: timer runs always, but we just skip IPIs to other CPUs when possible. Option is enabled by default now, keeping previous behavior, as periodic hardclock() calls are still used at least to implement setitimer(2) with ITIMER_VIRTUAL and ITIMER_PROF arguments. But since default schedulers don't depend on it since r232917, we are much more free to experiment with it. Modified: stable/9/share/man/man4/eventtimers.4 stable/9/sys/kern/kern_clocksource.c Directory Properties: stable/9/share/man/man4/ (props changed) stable/9/sys/ (props changed) Modified: stable/9/share/man/man4/eventtimers.4 ============================================================================== --- stable/9/share/man/man4/eventtimers.4 Thu Apr 12 06:24:47 2012 (r234166) +++ stable/9/share/man/man4/eventtimers.4 Thu Apr 12 06:29:02 2012 (r234167) @@ -24,7 +24,7 @@ .\" .\" $FreeBSD$ .\" -.Dd September 15, 2010 +.Dd March 13, 2012 .Dt EVENTTIMERS 4 .Os .Sh NAME @@ -143,6 +143,12 @@ By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are always generating. +.It Va kern.eventtimer.activetick +makes each CPU to receive all kinds of timer interrupts when they are busy. +Disabling it allows to skip some hardclock() calls in some cases. +By default this options is enabled. +If chosen timer is per-CPU, this option has no effect - all interrupts are +always generating, as timer reprogramming is too expensive for that case. .El .Sh SEE ALSO .Xr apic 4 , Modified: stable/9/sys/kern/kern_clocksource.c ============================================================================== --- stable/9/sys/kern/kern_clocksource.c Thu Apr 12 06:24:47 2012 (r234166) +++ stable/9/sys/kern/kern_clocksource.c Thu Apr 12 06:29:02 2012 (r234167) @@ -1,5 +1,5 @@ /*- - * Copyright (c) 2010 Alexander Motin + * Copyright (c) 2010-2012 Alexander Motin * All rights reserved. * * Redistribution and use in source and binary forms, with or without @@ -99,6 +99,7 @@ static struct bintime hardperiod; /* har static struct bintime statperiod; /* statclock() events period. */ static struct bintime profperiod; /* profclock() events period. */ static struct bintime nexttick; /* Next global timer tick time. */ +static struct bintime nexthard; /* Next global hardlock() event. */ static u_int busy = 0; /* Reconfiguration is in progress. */ static int profiling = 0; /* Profiling events enabled. */ @@ -110,11 +111,16 @@ TUNABLE_INT("kern.eventtimer.singlemul", SYSCTL_INT(_kern_eventtimer, OID_AUTO, singlemul, CTLFLAG_RW, &singlemul, 0, "Multiplier for periodic mode"); -static u_int idletick = 0; /* Idle mode allowed. */ +static u_int idletick = 0; /* Run periodic events when idle. */ TUNABLE_INT("kern.eventtimer.idletick", &idletick); SYSCTL_UINT(_kern_eventtimer, OID_AUTO, idletick, CTLFLAG_RW, &idletick, 0, "Run periodic events when idle"); +static u_int activetick = 1; /* Run all periodic events when active. */ +TUNABLE_INT("kern.eventtimer.activetick", &activetick); +SYSCTL_UINT(_kern_eventtimer, OID_AUTO, activetick, CTLFLAG_RW, &activetick, + 0, "Run all periodic events when active"); + static int periodic = 0; /* Periodic or one-shot mode. */ static int want_periodic = 0; /* What mode to prefer. */ TUNABLE_INT("kern.eventtimer.periodic", &want_periodic); @@ -202,6 +208,9 @@ handleevents(struct bintime *now, int fa bintime_add(&state->nexthard, &hardperiod); runs++; } + if ((timer->et_flags & ET_FLAGS_PERCPU) == 0 && + bintime_cmp(&state->nexthard, &nexthard, >)) + nexthard = state->nexthard; if (runs && fake < 2) { hardclock_cnt(runs, usermode); done = 1; @@ -263,9 +272,11 @@ getnextcpuevent(struct bintime *event, i int skip; state = DPCPU_PTR(timerstate); + /* Handle hardclock() events. */ *event = state->nexthard; - if (idle) { /* If CPU is idle - ask callouts for how long. */ - skip = 4; + if (idle || (!activetick && !profiling && + (timer->et_flags & ET_FLAGS_PERCPU) == 0)) { + skip = idle ? 4 : (stathz / 2); if (curcpu == CPU_FIRST() && tc_min_ticktock_freq > skip) skip = tc_min_ticktock_freq; skip = callout_tickstofirst(hz / skip) - 1; @@ -273,7 +284,8 @@ getnextcpuevent(struct bintime *event, i tmp = hardperiod; bintime_mul(&tmp, skip); bintime_add(event, &tmp); - } else { /* If CPU is active - handle all types of events. */ + } + if (!idle) { /* If CPU is active - handle other types of events. */ if (bintime_cmp(event, &state->nextstat, >)) *event = state->nextstat; if (profiling && bintime_cmp(event, &state->nextprof, >)) @@ -295,24 +307,28 @@ getnextevent(struct bintime *event) #ifdef SMP int cpu; #endif - int c; + int c, nonidle; state = DPCPU_PTR(timerstate); *event = state->nextevent; c = curcpu; -#ifdef SMP + nonidle = !state->idle; if ((timer->et_flags & ET_FLAGS_PERCPU) == 0) { +#ifdef SMP CPU_FOREACH(cpu) { if (curcpu == cpu) continue; state = DPCPU_ID_PTR(cpu, timerstate); + nonidle += !state->idle; if (bintime_cmp(event, &state->nextevent, >)) { *event = state->nextevent; c = cpu; } } - } #endif + if (nonidle != 0 && bintime_cmp(event, &nexthard, >)) + *event = nexthard; + } CTR5(KTR_SPARE2, "next at %d: next %d.%08x%08x by %d", curcpu, event->sec, (unsigned int)(event->frac >> 32), (unsigned int)(event->frac & 0xffffffff), c);