Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1756350Ab2BYLaN (ORCPT ); Sat, 25 Feb 2012 06:30:13 -0500 Received: from atrey.karlin.mff.cuni.cz ([195.113.26.193]:43429 "EHLO atrey.karlin.mff.cuni.cz" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752280Ab2BYLaK (ORCPT ); Sat, 25 Feb 2012 06:30:10 -0500 Date: Sat, 25 Feb 2012 12:30:06 +0100 From: Pavel Machek To: MyungJoo Ham Cc: cpufreq@vger.kernel.org, linux-kernel@vger.kernel.org, linux-pm@vger.kernel.org, Dave Jones , Len Brown , "Rafael J. Wysocki" , Kevin Hilman , Jean Pihet , markgross , kyungmin.park@samsung.com, myungjoo.ham@gmail.com Subject: Re: [RFC PATCH 2/2] CPUfreq ondemand: handle QoS request on DVFS response latency Message-ID: <20120225113005.GC16932@elf.ucw.cz> References: <1329897815-15871-1-git-send-email-myungjoo.ham@samsung.com> <1329897815-15871-3-git-send-email-myungjoo.ham@samsung.com> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <1329897815-15871-3-git-send-email-myungjoo.ham@samsung.com> X-Warning: Reading this can be dangerous to your mental health. User-Agent: Mutt/1.5.20 (2009-06-14) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 6483 Lines: 180 On Wed 2012-02-22 17:03:35, MyungJoo Ham wrote: > With QoS class, DVFS_RESPONSE_LATENCY, users (device drivers and > userspace processes) may express the desired maximum response latency > from DVFS mechanisms such as CPUfreq's ondemand governors. Based on such > QoS requests, the ondemand governor may flexibly adjust sampling rate > accordingly unless it goes below the min_sampling_rate. > > The benefit of having DVFS_RESPONSE_LATENCY is to have faster response > from user inputs (mouse clicks, keyboard inputs, touchscreen touches, > and others) without increasing frequency unconditionally. Because some > input events may not require any performance increases, increasing the > frequency unconditionally for inputs may simply consume too much energy. > Adjusting sampling rate based on user inputs enabled to increase > frequency with less latency if it requires and not to increase frequency > if it does not require. > > Signed-off-by: MyungJoo Ham > Signed-off-by: Kyungmin Park > > -- > This patch depends on the patch > "PM / QoS: Introduce new classes: DMA-Throughput and DVFS-Latency". > and the patch > "CPUfreq ondemand: update sampling rate without waiting for next > sampling" > --- > drivers/cpufreq/cpufreq_ondemand.c | 108 ++++++++++++++++++++++++++++++++---- > 1 files changed, 96 insertions(+), 12 deletions(-) > > diff --git a/drivers/cpufreq/cpufreq_ondemand.c b/drivers/cpufreq/cpufreq_ondemand.c > index 2d66649..b9188f1 100644 > --- a/drivers/cpufreq/cpufreq_ondemand.c > +++ b/drivers/cpufreq/cpufreq_ondemand.c > @@ -22,6 +22,7 @@ > #include > #include > #include > +#include > > /* > * dbs is used in this file as a shortform for demandbased switching > @@ -93,6 +94,7 @@ struct cpu_dbs_info_s { > * when user is changing the governor or limits. > */ > struct mutex timer_mutex; > + bool activated; /* dbs_timer_init is in effect */ > }; > static DEFINE_PER_CPU(struct cpu_dbs_info_s, od_cpu_dbs_info); > > @@ -111,6 +113,8 @@ static struct dbs_tuners { > unsigned int sampling_down_factor; > unsigned int powersave_bias; > unsigned int io_is_busy; > + struct notifier_block dvfs_lat_qos_db; > + unsigned int dvfs_lat_qos_wants; > } dbs_tuners_ins = { > .up_threshold = DEF_FREQUENCY_UP_THRESHOLD, > .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR, > @@ -164,6 +168,23 @@ static inline cputime64_t get_cpu_iowait_time(unsigned int cpu, cputime64_t *wal > } > > /* > + * Find right sampling rate based on sampling_rate and > + * QoS requests on dvfs latency. > + */ > +static unsigned int effective_sampling_rate(void) > +{ > + unsigned int effective; > + > + if (dbs_tuners_ins.dvfs_lat_qos_wants) > + effective = min(dbs_tuners_ins.dvfs_lat_qos_wants, > + dbs_tuners_ins.sampling_rate); > + else > + effective = dbs_tuners_ins.sampling_rate; > + > + return max(effective, min_sampling_rate); > +} > + > +/* > * Find right freq to be set now with powersave_bias on. > * Returns the freq_hi to be used right now and will set freq_hi_jiffies, > * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs. > @@ -207,7 +228,7 @@ static unsigned int powersave_bias_target(struct cpufreq_policy *policy, > dbs_info->freq_lo_jiffies = 0; > return freq_lo; > } > - jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate); > + jiffies_total = usecs_to_jiffies(effective_sampling_rate()); > jiffies_hi = (freq_avg - freq_lo) * jiffies_total; > jiffies_hi += ((freq_hi - freq_lo) / 2); > jiffies_hi /= (freq_hi - freq_lo); > @@ -259,7 +280,8 @@ show_one(powersave_bias, powersave_bias); > > /** > * update_sampling_rate - update sampling rate effective immediately if needed. > - * @new_rate: new sampling rate > + * @new_rate: new sampling rate. if it is 0, regard sampling rate is not > + * changed and assume that qos request value is changed. > * > * If new rate is smaller than the old, simply updaing > * dbs_tuners_int.sampling_rate might not be appropriate. For example, > @@ -273,9 +295,13 @@ show_one(powersave_bias, powersave_bias); > static void update_sampling_rate(unsigned int new_rate) > { > int cpu; > + unsigned int effective; > + > + > + if (new_rate) > + dbs_tuners_ins.sampling_rate = max(new_rate, min_sampling_rate); > > - dbs_tuners_ins.sampling_rate = new_rate > - = max(new_rate, min_sampling_rate); > + effective = effective_sampling_rate(); > > for_each_online_cpu(cpu) { > struct cpufreq_policy *policy; > @@ -283,21 +309,31 @@ static void update_sampling_rate(unsigned int new_rate) > struct timer_list *timer; > unsigned long appointed_at; > > + /* > + * mutex_destory(&dbs_info->timer_mutex) should not happen > + * in this context. > + */ > + mutex_lock(&dbs_mutex); > + > policy = cpufreq_cpu_get(cpu); > if (!policy) > - continue; > + goto next; > dbs_info = &per_cpu(od_cpu_dbs_info, policy->cpu); > cpufreq_cpu_put(policy); > > + /* timer_mutex destroyed or will be destoyed soon */ > + if (!dbs_info->activated) > + goto next; > + > mutex_lock(&dbs_info->timer_mutex); > > if (!delayed_work_pending(&dbs_info->work)) > - goto next; > + goto next_timer_mutex; > > timer = &dbs_info->work.timer; > appointed_at = timer->expires; > > - if (time_before(jiffies + usecs_to_jiffies(new_rate), > + if (time_before(jiffies + usecs_to_jiffies(effective), > appointed_at)) { > > mutex_unlock(&dbs_info->timer_mutex); > @@ -305,12 +341,15 @@ static void update_sampling_rate(unsigned int new_rate) > mutex_lock(&dbs_info->timer_mutex); > > schedule_delayed_work_on(dbs_info->cpu, &dbs_info->work, > - usecs_to_jiffies(new_rate)); > + usecs_to_jiffies(effective)); > > } > -next: > +next_timer_mutex: > mutex_unlock(&dbs_info->timer_mutex); > +next: > + mutex_unlock(&dbs_mutex); > } > + > } I don't think gotos are helpful here. Can you use normal program structure or move it to subroutine...? Pavel -- (english) http://www.livejournal.com/~pavelmachek (cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/