From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1758274AbZEBTz5 (ORCPT ); Sat, 2 May 2009 15:55:57 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1756016AbZEBTzr (ORCPT ); Sat, 2 May 2009 15:55:47 -0400 Received: from www.tglx.de ([62.245.132.106]:47137 "EHLO www.tglx.de" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1755989AbZEBTzr (ORCPT ); Sat, 2 May 2009 15:55:47 -0400 Date: Sat, 2 May 2009 21:54:49 +0200 (CEST) From: Thomas Gleixner To: Andrew Morton cc: LKML , Dimitri Sivanich , Ingo Molnar , Peter Zijlstra Subject: Re: [patch 2/3] timer: move calc_load to softirq In-Reply-To: <20090502122459.ccc870fc.akpm@linux-foundation.org> Message-ID: References: <20090502190500.513452861@linutronix.de> <20090502190545.489264416@linutronix.de> <20090502122459.ccc870fc.akpm@linux-foundation.org> User-Agent: Alpine 2.00 (LFD 1167 2008-08-23) MIME-Version: 1.0 Content-Type: TEXT/PLAIN; charset=US-ASCII Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Sat, 2 May 2009, Andrew Morton wrote: > > + spin_lock(&avenrun_lock); > > + ticks = atomic_read(&avenrun_ticks); > > + if (ticks >= LOAD_FREQ) { > > + atomic_sub(LOAD_FREQ, &avenrun_ticks); > > + calc_global_load(); > > } > > + spin_unlock(&avenrun_lock); > > + *calc = 0; > > +} > > I wonder if we really really need avenrun_lock. Various bits of code > (eg net/sched/em_meta.c) cheerily read avenrun[] without locking. I don't care about the reader side anyway, the lock is just there to protect the calc_load update from two cpus, but that's probably paranoia. Though, there is a theoretical race between 2 cpus which might want to update avenrun_ticks in the NOHZ case, but thinking more about it we can just prevent this by clever usage of the atomic ops on avenrun_ticks. Thanks, tglx -----------> Subject: timer: move calc_load to softirq From: Thomas Gleixner Date: Sat, 2 May 2009 19:43:41 +0200 xtime_lock is held write locked across calc_load() which iterates over all online CPUs. That can cause long latencies for xtime_lock readers on large SMP systems. The load average calculation is an rough estimate anyway so there is no real need to protect the readers vs. the update. It's not a problem when the avenrun array is updated while a reader copies the values. Move the calculation to the softirq and reduce the xtime_lock write locked section. This also reduces the interrupts off section. Inspired by an inital patch from Dimitri Sivanich. Signed-off-by: Thomas Gleixner --- kernel/time/timekeeping.c | 2 - kernel/timer.c | 57 ++++++++++++++++++++++++++++++++-------------- 2 files changed, 41 insertions(+), 18 deletions(-) Index: linux-2.6/kernel/time/timekeeping.c =================================================================== --- linux-2.6.orig/kernel/time/timekeeping.c +++ linux-2.6/kernel/time/timekeeping.c @@ -22,7 +22,7 @@ /* * This read-write spinlock protects us from races in SMP while - * playing with xtime and avenrun. + * playing with xtime. */ __cacheline_aligned_in_smp DEFINE_SEQLOCK(xtime_lock); Index: linux-2.6/kernel/timer.c =================================================================== --- linux-2.6.orig/kernel/timer.c +++ linux-2.6/kernel/timer.c @@ -1127,12 +1127,13 @@ void update_process_times(int user_tick) * imply that avenrun[] is the standard name for this kind of thing. * Nothing else seems to be standardized: the fractional size etc * all seem to differ on different machines. - * - * Requires xtime_lock to access. */ unsigned long avenrun[3]; EXPORT_SYMBOL(avenrun); +static atomic_t avenrun_ticks; +static DEFINE_PER_CPU(int, avenrun_calculate); + static unsigned long calc_load(unsigned long load, unsigned long exp, unsigned long active) { @@ -1143,23 +1144,44 @@ calc_load(unsigned long load, unsigned l /* * calc_load - given tick count, update the avenrun load estimates. - * This is called while holding a write_lock on xtime_lock. */ -static void calc_global_load(unsigned long ticks) +static void calc_global_load(void) { - unsigned long active_tasks; /* fixed-point */ - static int count = LOAD_FREQ; + unsigned long active_tasks = nr_active() * FIXED_1; - count -= ticks; - if (unlikely(count < 0)) { - active_tasks = nr_active() * FIXED_1; - do { - avenrun[0] = calc_load(avenrun[0], EXP_1, active_tasks); - avenrun[1] = calc_load(avenrun[1], EXP_5, active_tasks); - avenrun[2] = calc_load(avenrun[2], EXP_15, active_tasks); - count += LOAD_FREQ; - } while (count < 0); - } + avenrun[0] = calc_load(avenrun[0], EXP_1, active_tasks); + avenrun[1] = calc_load(avenrun[1], EXP_5, active_tasks); + avenrun[2] = calc_load(avenrun[2], EXP_15, active_tasks); +} + +/* + * Check whether do_timer has set avenrun_calculate. The variable is + * cpu local so we avoid cache line bouncing of avenrun_ticks. + */ +static void check_calc_load(void) +{ + int ticks, *calc = &__get_cpu_var(avenrun_calculate); + + if (!*calc) + return; + + ticks = atomic_sub_return(LOAD_FREQ, &avenrun_ticks); + if (ticks >= 0) + calc_global_load(); + else + atomic_add(LOAD_FREQ, &avenrun_ticks); + *calc = 0; +} + +/* + * Update avenrun_ticks and trigger the load calculation when the + * result is >= LOAD_FREQ. + */ +static void calc_load_update(unsigned long ticks) +{ + ticks = atomic_add_return(ticks, &avenrun_ticks); + if (ticks >= LOAD_FREQ) + __get_cpu_var(avenrun_calculate) = 1; } /* @@ -1169,6 +1191,7 @@ static void run_timer_softirq(struct sof { struct tvec_base *base = __get_cpu_var(tvec_bases); + check_calc_load(); hrtimer_run_pending(); if (time_after_eq(jiffies, base->timer_jiffies)) @@ -1192,7 +1215,7 @@ void run_local_timers(void) static inline void update_times(unsigned long ticks) { update_wall_time(); - calc_global_load(ticks); + calc_load_update(ticks); } /*