From e4c777d8314d7925e4895f00b3a7ebd64a4d830b Mon Sep 17 00:00:00 2001 From: Mike Turquette Date: Tue, 17 May 2011 09:43:09 -0500 Subject: [PATCH 2/2] cpufreq: introduce hotplug governor The "hotplug" governor scales CPU frequency based on load, similar to "ondemand". It scales up to the highest frequency when "up_threshold" is crossed and scales down one frequency at a time when "down_threshold" is crossed. Unlike those governors, target frequencies are determined by directly accessing the CPUfreq frequency table, instead of taking some percentage of maximum available frequency. The key difference in the "hotplug" governor is that it will disable auxillary CPUs when the system is very idle, and enable them again once the system becomes busy. This is achieved by averaging load over multiple sampling periods; if CPUs were online or offlined based on a single sampling period then thrashing will occur. Sysfs entries exist for "hotplug_in_sampling_periods" and for "hotplug_out_sampling_periods" which determine how many consecutive periods get averaged to determine if auxillery CPUs should be onlined or offlined. Defaults are 5 periods and 20 periods respectively. Otherwise the standard sysfs entries you might find for "ondemand" and "conservative" governors are there. To use this governor it is assumed that your CPUfreq driver has populated the CPUfreq table, CONFIG_NO_HZ is enabled and CONFIG_HOTPLUG_CPU is enabled. Changes in V2: Corrected default sampling periods Optimized load history array resizing Maintain load history when resizing array Add locking to dbs_check_cpu Switch from enable_nonboot_cpus to cpu_up Switch from disable_nonboot_cpus to down_cpu Fix some printks Coding style around for-loops Signed-off-by: Mike Turquette Signed-off-by: Santosh Shilimkar Signed-off-by: Koen Kooi --- Documentation/cpu-freq/governors.txt | 28 ++ drivers/cpufreq/Kconfig | 33 ++ drivers/cpufreq/Makefile | 1 + drivers/cpufreq/cpufreq_hotplug.c | 705 ++++++++++++++++++++++++++++++++++ include/linux/cpufreq.h | 3 + 5 files changed, 770 insertions(+), 0 deletions(-) create mode 100644 drivers/cpufreq/cpufreq_hotplug.c diff --git a/Documentation/cpu-freq/governors.txt b/Documentation/cpu-freq/governors.txt index e74d0a2..c2e3d3d 100644 --- a/Documentation/cpu-freq/governors.txt +++ b/Documentation/cpu-freq/governors.txt @@ -193,6 +193,34 @@ governor but for the opposite direction. For example when set to its default value of '20' it means that if the CPU usage needs to be below 20% between samples to have the frequency decreased. + +2.6 Hotplug +----------- + +The CPUfreq governor "hotplug" operates similary to "ondemand" and +"conservative". It's decisions are based primarily on CPU load. Like +"ondemand" the "hotplug" governor will ramp up to the highest frequency +once the run-time tunable "up_threshold" parameter is crossed. Like +"conservative", the "hotplug" governor exports a "down_threshold" +parameter that is also tunable at run-time. When the "down_threshold" +is crossed the CPU transitions to the next lowest frequency in the +CPUfreq frequency table instead of decrementing the frequency based on a +percentage of maximum load. + +The main reason "hotplug" governor exists is for architectures requiring +that only the master CPU be online in order to hit low-power states +(C-states). OMAP4 is one such example of this. The "hotplug" governor +is also helpful in reducing thermal output in devices with tight thermal +constraints. + +Auxillary CPUs are onlined/offline based on CPU load, but the decision +to do so is made after averaging several sampling windows. This is to +reduce CPU hotplug "thrashing", which can be caused by normal system +entropy and leads to lots of spurious plug-in and plug-out transitions. +The number of sampling periods averaged together is tunable via the +"hotplug_in_sampling_periods" and "hotplug_out_sampling_periods" +run-time tunable parameters. + 3. The Governor Interface in the CPUfreq Core ============================================= diff --git a/drivers/cpufreq/Kconfig b/drivers/cpufreq/Kconfig index ca8ee80..c716a0e 100644 --- a/drivers/cpufreq/Kconfig +++ b/drivers/cpufreq/Kconfig @@ -110,6 +110,19 @@ config CPU_FREQ_DEFAULT_GOV_CONSERVATIVE Be aware that not all cpufreq drivers support the conservative governor. If unsure have a look at the help section of the driver. Fallback governor will be the performance governor. + +config CPU_FREQ_DEFAULT_GOV_HOTPLUG + bool "hotplug" + select CPU_FREQ_GOV_HOTPLUG + select CPU_FREQ_GOV_PERFORMANCE + help + Use the CPUFreq governor 'hotplug' as default. This allows you + to get a full dynamic frequency capable system with CPU + hotplug support by simply loading your cpufreq low-level + hardware driver. Be aware that not all cpufreq drivers + support the hotplug governor. If unsure have a look at + the help section of the driver. Fallback governor will be the + performance governor. endchoice config CPU_FREQ_GOV_PERFORMANCE @@ -190,4 +203,24 @@ config CPU_FREQ_GOV_CONSERVATIVE If in doubt, say N. +config CPU_FREQ_GOV_HOTPLUG + tristate "'hotplug' cpufreq governor" + depends on CPU_FREQ && NO_HZ && HOTPLUG_CPU + help + 'hotplug' - this driver mimics the frequency scaling behavior + in 'ondemand', but with several key differences. First is + that frequency transitions use the CPUFreq table directly, + instead of incrementing in a percentage of the maximum + available frequency. Second 'hotplug' will offline auxillary + CPUs when the system is idle, and online those CPUs once the + system becomes busy again. This last feature is needed for + architectures which transition to low power states when only + the "master" CPU is online, or for thermally constrained + devices. + + If you don't have one of these architectures or devices, use + 'ondemand' instead. + + If in doubt, say N. + endif # CPU_FREQ diff --git a/drivers/cpufreq/Makefile b/drivers/cpufreq/Makefile index 71fc3b4..05d564c 100644 --- a/drivers/cpufreq/Makefile +++ b/drivers/cpufreq/Makefile @@ -9,6 +9,7 @@ obj-$(CONFIG_CPU_FREQ_GOV_POWERSAVE) += cpufreq_powersave.o obj-$(CONFIG_CPU_FREQ_GOV_USERSPACE) += cpufreq_userspace.o obj-$(CONFIG_CPU_FREQ_GOV_ONDEMAND) += cpufreq_ondemand.o obj-$(CONFIG_CPU_FREQ_GOV_CONSERVATIVE) += cpufreq_conservative.o +obj-$(CONFIG_CPU_FREQ_GOV_HOTPLUG) += cpufreq_hotplug.o # CPUfreq cross-arch helpers obj-$(CONFIG_CPU_FREQ_TABLE) += freq_table.o diff --git a/drivers/cpufreq/cpufreq_hotplug.c b/drivers/cpufreq/cpufreq_hotplug.c new file mode 100644 index 0000000..85aa6d2 --- /dev/null +++ b/drivers/cpufreq/cpufreq_hotplug.c @@ -0,0 +1,705 @@ +/* + * CPUFreq hotplug governor + * + * Copyright (C) 2010 Texas Instruments, Inc. + * Mike Turquette + * Santosh Shilimkar + * + * Based on ondemand governor + * Copyright (C) 2001 Russell King + * (C) 2003 Venkatesh Pallipadi , + * Jun Nakajima + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2 as + * published by the Free Software Foundation. + */ + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +/* greater than 80% avg load across online CPUs increases frequency */ +#define DEFAULT_UP_FREQ_MIN_LOAD (80) + +/* less than 20% avg load across online CPUs decreases frequency */ +#define DEFAULT_DOWN_FREQ_MAX_LOAD (20) + +/* default sampling period (uSec) is bogus; 10x ondemand's default for x86 */ +#define DEFAULT_SAMPLING_PERIOD (100000) + +/* default number of sampling periods to average before hotplug-in decision */ +#define DEFAULT_HOTPLUG_IN_SAMPLING_PERIODS (5) + +/* default number of sampling periods to average before hotplug-out decision */ +#define DEFAULT_HOTPLUG_OUT_SAMPLING_PERIODS (20) + +static void do_dbs_timer(struct work_struct *work); +static int cpufreq_governor_dbs(struct cpufreq_policy *policy, + unsigned int event); + +#ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_HOTPLUG +static +#endif +struct cpufreq_governor cpufreq_gov_hotplug = { + .name = "hotplug", + .governor = cpufreq_governor_dbs, + .owner = THIS_MODULE, +}; + +struct cpu_dbs_info_s { + cputime64_t prev_cpu_idle; + cputime64_t prev_cpu_wall; + cputime64_t prev_cpu_nice; + struct cpufreq_policy *cur_policy; + struct delayed_work work; + struct cpufreq_frequency_table *freq_table; + int cpu; + /* + * percpu mutex that serializes governor limit change with + * do_dbs_timer invocation. We do not want do_dbs_timer to run + * when user is changing the governor or limits. + */ + struct mutex timer_mutex; +}; +static DEFINE_PER_CPU(struct cpu_dbs_info_s, hp_cpu_dbs_info); + +static unsigned int dbs_enable; /* number of CPUs using this policy */ + +/* + * dbs_mutex protects data in dbs_tuners_ins from concurrent changes on + * different CPUs. It protects dbs_enable in governor start/stop. + */ +static DEFINE_MUTEX(dbs_mutex); + +static struct workqueue_struct *khotplug_wq; + +static struct dbs_tuners { + unsigned int sampling_rate; + unsigned int up_threshold; + unsigned int down_threshold; + unsigned int hotplug_in_sampling_periods; + unsigned int hotplug_out_sampling_periods; + unsigned int hotplug_load_index; + unsigned int *hotplug_load_history; + unsigned int ignore_nice; + unsigned int io_is_busy; +} dbs_tuners_ins = { + .sampling_rate = DEFAULT_SAMPLING_PERIOD, + .up_threshold = DEFAULT_UP_FREQ_MIN_LOAD, + .down_threshold = DEFAULT_DOWN_FREQ_MAX_LOAD, + .hotplug_in_sampling_periods = DEFAULT_HOTPLUG_IN_SAMPLING_PERIODS, + .hotplug_out_sampling_periods = DEFAULT_HOTPLUG_OUT_SAMPLING_PERIODS, + .hotplug_load_index = 0, + .ignore_nice = 0, + .io_is_busy = 0, +}; + +/* + * A corner case exists when switching io_is_busy at run-time: comparing idle + * times from a non-io_is_busy period to an io_is_busy period (or vice-versa) + * will misrepresent the actual change in system idleness. We ignore this + * corner case: enabling io_is_busy might cause freq increase and disabling + * might cause freq decrease, which probably matches the original intent. + */ +static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall) +{ + u64 idle_time; + u64 iowait_time; + + /* cpufreq-hotplug always assumes CONFIG_NO_HZ */ + idle_time = get_cpu_idle_time_us(cpu, wall); + + /* add time spent doing I/O to idle time */ + if (dbs_tuners_ins.io_is_busy) { + iowait_time = get_cpu_iowait_time_us(cpu, wall); + /* cpufreq-hotplug always assumes CONFIG_NO_HZ */ + if (iowait_time != -1ULL && idle_time >= iowait_time) + idle_time -= iowait_time; + } + + return idle_time; +} + +/************************** sysfs interface ************************/ + +/* XXX look at global sysfs macros in cpufreq.h, can those be used here? */ + +/* cpufreq_hotplug Governor Tunables */ +#define show_one(file_name, object) \ +static ssize_t show_##file_name \ +(struct kobject *kobj, struct attribute *attr, char *buf) \ +{ \ + return sprintf(buf, "%u\n", dbs_tuners_ins.object); \ +} +show_one(sampling_rate, sampling_rate); +show_one(up_threshold, up_threshold); +show_one(down_threshold, down_threshold); +show_one(hotplug_in_sampling_periods, hotplug_in_sampling_periods); +show_one(hotplug_out_sampling_periods, hotplug_out_sampling_periods); +show_one(ignore_nice_load, ignore_nice); +show_one(io_is_busy, io_is_busy); + +static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b, + const char *buf, size_t count) +{ + unsigned int input; + int ret; + ret = sscanf(buf, "%u", &input); + if (ret != 1) + return -EINVAL; + + mutex_lock(&dbs_mutex); + dbs_tuners_ins.sampling_rate = input; + mutex_unlock(&dbs_mutex); + + return count; +} + +static ssize_t store_up_threshold(struct kobject *a, struct attribute *b, + const char *buf, size_t count) +{ + unsigned int input; + int ret; + ret = sscanf(buf, "%u", &input); + + if (ret != 1 || input <= dbs_tuners_ins.down_threshold) { + return -EINVAL; + } + + mutex_lock(&dbs_mutex); + dbs_tuners_ins.up_threshold = input; + mutex_unlock(&dbs_mutex); + + return count; +} + +static ssize_t store_down_threshold(struct kobject *a, struct attribute *b, + const char *buf, size_t count) +{ + unsigned int input; + int ret; + ret = sscanf(buf, "%u", &input); + + if (ret != 1 || input >= dbs_tuners_ins.up_threshold) { + return -EINVAL; + } + + mutex_lock(&dbs_mutex); + dbs_tuners_ins.down_threshold = input; + mutex_unlock(&dbs_mutex); + + return count; +} + +static ssize_t store_hotplug_in_sampling_periods(struct kobject *a, + struct attribute *b, const char *buf, size_t count) +{ + unsigned int input; + unsigned int *temp; + unsigned int max_windows; + int ret; + ret = sscanf(buf, "%u", &input); + + if (ret != 1) + return -EINVAL; + + /* already using this value, bail out */ + if (input == dbs_tuners_ins.hotplug_in_sampling_periods) + return count; + + mutex_lock(&dbs_mutex); + ret = count; + max_windows = max(dbs_tuners_ins.hotplug_in_sampling_periods, + dbs_tuners_ins.hotplug_out_sampling_periods); + + /* no need to resize array */ + if (input <= max_windows) { + dbs_tuners_ins.hotplug_in_sampling_periods = input; + goto out; + } + + /* resize array */ + temp = kmalloc((sizeof(unsigned int) * input), GFP_KERNEL); + + if (!temp || IS_ERR(temp)) { + ret = -ENOMEM; + goto out; + } + + memcpy(temp, dbs_tuners_ins.hotplug_load_history, + (max_windows * sizeof(unsigned int))); + kfree(dbs_tuners_ins.hotplug_load_history); + + /* replace old buffer, old number of sampling periods & old index */ + dbs_tuners_ins.hotplug_load_history = temp; + dbs_tuners_ins.hotplug_in_sampling_periods = input; + dbs_tuners_ins.hotplug_load_index = max_windows; +out: + mutex_unlock(&dbs_mutex); + + return ret; +} + +static ssize_t store_hotplug_out_sampling_periods(struct kobject *a, + struct attribute *b, const char *buf, size_t count) +{ + unsigned int input; + unsigned int *temp; + unsigned int max_windows; + int ret; + ret = sscanf(buf, "%u", &input); + + if (ret != 1) + return -EINVAL; + + /* already using this value, bail out */ + if (input == dbs_tuners_ins.hotplug_out_sampling_periods) + return count; + + mutex_lock(&dbs_mutex); + ret = count; + max_windows = max(dbs_tuners_ins.hotplug_in_sampling_periods, + dbs_tuners_ins.hotplug_out_sampling_periods); + + /* no need to resize array */ + if (input <= max_windows) { + dbs_tuners_ins.hotplug_out_sampling_periods = input; + goto out; + } + + /* resize array */ + temp = kmalloc((sizeof(unsigned int) * input), GFP_KERNEL); + + if (!temp || IS_ERR(temp)) { + ret = -ENOMEM; + goto out; + } + + memcpy(temp, dbs_tuners_ins.hotplug_load_history, + (max_windows * sizeof(unsigned int))); + kfree(dbs_tuners_ins.hotplug_load_history); + + /* replace old buffer, old number of sampling periods & old index */ + dbs_tuners_ins.hotplug_load_history = temp; + dbs_tuners_ins.hotplug_out_sampling_periods = input; + dbs_tuners_ins.hotplug_load_index = max_windows; +out: + mutex_unlock(&dbs_mutex); + + return ret; +} + +static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b, + const char *buf, size_t count) +{ + unsigned int input; + int ret; + + unsigned int j; + + ret = sscanf(buf, "%u", &input); + if (ret != 1) + return -EINVAL; + + if (input > 1) + input = 1; + + mutex_lock(&dbs_mutex); + if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */ + mutex_unlock(&dbs_mutex); + return count; + } + dbs_tuners_ins.ignore_nice = input; + + /* we need to re-evaluate prev_cpu_idle */ + for_each_online_cpu(j) { + struct cpu_dbs_info_s *dbs_info; + dbs_info = &per_cpu(hp_cpu_dbs_info, j); + dbs_info->prev_cpu_idle = get_cpu_idle_time(j, + &dbs_info->prev_cpu_wall); + if (dbs_tuners_ins.ignore_nice) + dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice; + + } + mutex_unlock(&dbs_mutex); + + return count; +} + +static ssize_t store_io_is_busy(struct kobject *a, struct attribute *b, + const char *buf, size_t count) +{ + unsigned int input; + int ret; + + ret = sscanf(buf, "%u", &input); + if (ret != 1) + return -EINVAL; + + mutex_lock(&dbs_mutex); + dbs_tuners_ins.io_is_busy = !!input; + mutex_unlock(&dbs_mutex); + + return count; +} + +define_one_global_rw(sampling_rate); +define_one_global_rw(up_threshold); +define_one_global_rw(down_threshold); +define_one_global_rw(hotplug_in_sampling_periods); +define_one_global_rw(hotplug_out_sampling_periods); +define_one_global_rw(ignore_nice_load); +define_one_global_rw(io_is_busy); + +static struct attribute *dbs_attributes[] = { + &sampling_rate.attr, + &up_threshold.attr, + &down_threshold.attr, + &hotplug_in_sampling_periods.attr, + &hotplug_out_sampling_periods.attr, + &ignore_nice_load.attr, + &io_is_busy.attr, + NULL +}; + +static struct attribute_group dbs_attr_group = { + .attrs = dbs_attributes, + .name = "hotplug", +}; + +/************************** sysfs end ************************/ + +static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info) +{ + /* combined load of all enabled CPUs */ + unsigned int total_load = 0; + /* single largest CPU load */ + unsigned int max_load = 0; + /* average load across all enabled CPUs */ + unsigned int avg_load = 0; + /* average load across multiple sampling periods for hotplug events */ + unsigned int hotplug_in_avg_load = 0; + unsigned int hotplug_out_avg_load = 0; + /* number of sampling periods averaged for hotplug decisions */ + unsigned int periods; + + struct cpufreq_policy *policy; + unsigned int index = 0; + unsigned int i, j; + + policy = this_dbs_info->cur_policy; + + /* + * cpu load accounting + * get highest load, total load and average load across all CPUs + */ + for_each_cpu(j, policy->cpus) { + unsigned int load; + unsigned int idle_time, wall_time; + cputime64_t cur_wall_time, cur_idle_time; + struct cpu_dbs_info_s *j_dbs_info; + + j_dbs_info = &per_cpu(hp_cpu_dbs_info, j); + + /* update both cur_idle_time and cur_wall_time */ + cur_idle_time = get_cpu_idle_time(j, &cur_wall_time); + + /* how much wall time has passed since last iteration? */ + wall_time = (unsigned int) cputime64_sub(cur_wall_time, + j_dbs_info->prev_cpu_wall); + j_dbs_info->prev_cpu_wall = cur_wall_time; + + /* how much idle time has passed since last iteration? */ + idle_time = (unsigned int) cputime64_sub(cur_idle_time, + j_dbs_info->prev_cpu_idle); + j_dbs_info->prev_cpu_idle = cur_idle_time; + + if (unlikely(!wall_time || wall_time < idle_time)) + continue; + + /* load is the percentage of time not spent in idle */ + load = 100 * (wall_time - idle_time) / wall_time; + + /* keep track of combined load across all CPUs */ + total_load += load; + + /* keep track of highest single load across all CPUs */ + if (load > max_load) + max_load = load; + } + + /* calculate the average load across all related CPUs */ + avg_load = total_load / num_online_cpus(); + + + /* + * hotplug load accounting + * average load over multiple sampling periods + */ + + /* how many sampling periods do we use for hotplug decisions? */ + periods = max(dbs_tuners_ins.hotplug_in_sampling_periods, + dbs_tuners_ins.hotplug_out_sampling_periods); + + /* store avg_load in the circular buffer */ + dbs_tuners_ins.hotplug_load_history[dbs_tuners_ins.hotplug_load_index] + = avg_load; + + /* compute average load across in & out sampling periods */ + for (i = 0, j = dbs_tuners_ins.hotplug_load_index; + i < periods; i++, j--) { + if (i < dbs_tuners_ins.hotplug_in_sampling_periods) + hotplug_in_avg_load += + dbs_tuners_ins.hotplug_load_history[j]; + if (i < dbs_tuners_ins.hotplug_out_sampling_periods) + hotplug_out_avg_load += + dbs_tuners_ins.hotplug_load_history[j]; + + if (j == 0) + j = periods; + } + + hotplug_in_avg_load = hotplug_in_avg_load / + dbs_tuners_ins.hotplug_in_sampling_periods; + + hotplug_out_avg_load = hotplug_out_avg_load / + dbs_tuners_ins.hotplug_out_sampling_periods; + + /* return to first element if we're at the circular buffer's end */ + if (++dbs_tuners_ins.hotplug_load_index == periods) + dbs_tuners_ins.hotplug_load_index = 0; + + /* check for frequency increase */ + if (avg_load > dbs_tuners_ins.up_threshold) { + /* should we enable auxillary CPUs? */ + if (num_online_cpus() < 2 && hotplug_in_avg_load > + dbs_tuners_ins.up_threshold) { + /* hotplug with cpufreq is nasty + * a call to cpufreq_governor_dbs may cause a lockup. + * wq is not running here so its safe. + */ + mutex_unlock(&this_dbs_info->timer_mutex); + cpu_up(1); + mutex_lock(&this_dbs_info->timer_mutex); + goto out; + } + + /* increase to highest frequency supported */ + if (policy->cur < policy->max) + __cpufreq_driver_target(policy, policy->max, + CPUFREQ_RELATION_H); + + goto out; + } + + /* check for frequency decrease */ + if (avg_load < dbs_tuners_ins.down_threshold) { + /* are we at the minimum frequency already? */ + if (policy->cur == policy->min) { + /* should we disable auxillary CPUs? */ + if (num_online_cpus() > 1 && hotplug_out_avg_load < + dbs_tuners_ins.down_threshold) { + mutex_unlock(&this_dbs_info->timer_mutex); + cpu_down(1); + mutex_lock(&this_dbs_info->timer_mutex); + } + goto out; + } + + /* bump down to the next lowest frequency in the table */ + if (cpufreq_frequency_table_next_lowest(policy, + this_dbs_info->freq_table, &index)) { + pr_err("%s: failed to get next lowest frequency\n", + __func__); + goto out; + } + + __cpufreq_driver_target(policy, + this_dbs_info->freq_table[index].frequency, + CPUFREQ_RELATION_L); + } +out: + return; +} + +static void do_dbs_timer(struct work_struct *work) +{ + struct cpu_dbs_info_s *dbs_info = + container_of(work, struct cpu_dbs_info_s, work.work); + unsigned int cpu = dbs_info->cpu; + + /* We want all related CPUs to do sampling nearly on same jiffy */ + int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate); + + mutex_lock(&dbs_info->timer_mutex); + dbs_check_cpu(dbs_info); + queue_delayed_work_on(cpu, khotplug_wq, &dbs_info->work, delay); + mutex_unlock(&dbs_info->timer_mutex); +} + +static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info) +{ + /* We want all related CPUs to do sampling nearly on same jiffy */ + int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate); + delay -= jiffies % delay; + + INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer); + queue_delayed_work_on(dbs_info->cpu, khotplug_wq, &dbs_info->work, + delay); +} + +static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info) +{ + cancel_delayed_work_sync(&dbs_info->work); +} + +static int cpufreq_governor_dbs(struct cpufreq_policy *policy, + unsigned int event) +{ + unsigned int cpu = policy->cpu; + struct cpu_dbs_info_s *this_dbs_info; + unsigned int i, j, max_periods; + int rc; + + this_dbs_info = &per_cpu(hp_cpu_dbs_info, cpu); + + switch (event) { + case CPUFREQ_GOV_START: + if ((!cpu_online(cpu)) || (!policy->cur)) + return -EINVAL; + + mutex_lock(&dbs_mutex); + dbs_enable++; + for_each_cpu(j, policy->cpus) { + struct cpu_dbs_info_s *j_dbs_info; + j_dbs_info = &per_cpu(hp_cpu_dbs_info, j); + j_dbs_info->cur_policy = policy; + + j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j, + &j_dbs_info->prev_cpu_wall); + if (dbs_tuners_ins.ignore_nice) { + j_dbs_info->prev_cpu_nice = + kstat_cpu(j).cpustat.nice; + } + + max_periods = max(DEFAULT_HOTPLUG_IN_SAMPLING_PERIODS, + DEFAULT_HOTPLUG_OUT_SAMPLING_PERIODS); + dbs_tuners_ins.hotplug_load_history = kmalloc( + (sizeof(unsigned int) * max_periods), + GFP_KERNEL); + if (!dbs_tuners_ins.hotplug_load_history) { + WARN_ON(1); + return -ENOMEM; + } + for (i = 0; i < max_periods; i++) + dbs_tuners_ins.hotplug_load_history[i] = 50; + } + this_dbs_info->cpu = cpu; + this_dbs_info->freq_table = cpufreq_frequency_get_table(cpu); + /* + * Start the timerschedule work, when this governor + * is used for first time + */ + if (dbs_enable == 1) { + rc = sysfs_create_group(cpufreq_global_kobject, + &dbs_attr_group); + if (rc) { + mutex_unlock(&dbs_mutex); + return rc; + } + } + mutex_unlock(&dbs_mutex); + + mutex_init(&this_dbs_info->timer_mutex); + dbs_timer_init(this_dbs_info); + break; + + case CPUFREQ_GOV_STOP: + dbs_timer_exit(this_dbs_info); + + mutex_lock(&dbs_mutex); + mutex_destroy(&this_dbs_info->timer_mutex); + dbs_enable--; + mutex_unlock(&dbs_mutex); + if (!dbs_enable) + sysfs_remove_group(cpufreq_global_kobject, + &dbs_attr_group); + kfree(dbs_tuners_ins.hotplug_load_history); + /* + * XXX BIG CAVEAT: Stopping the governor with CPU1 offline + * will result in it remaining offline until the user onlines + * it again. It is up to the user to do this (for now). + */ + break; + + case CPUFREQ_GOV_LIMITS: + mutex_lock(&this_dbs_info->timer_mutex); + if (policy->max < this_dbs_info->cur_policy->cur) + __cpufreq_driver_target(this_dbs_info->cur_policy, + policy->max, CPUFREQ_RELATION_H); + else if (policy->min > this_dbs_info->cur_policy->cur) + __cpufreq_driver_target(this_dbs_info->cur_policy, + policy->min, CPUFREQ_RELATION_L); + mutex_unlock(&this_dbs_info->timer_mutex); + break; + } + return 0; +} + +static int __init cpufreq_gov_dbs_init(void) +{ + int err; + cputime64_t wall; + u64 idle_time; + int cpu = get_cpu(); + + idle_time = get_cpu_idle_time_us(cpu, &wall); + put_cpu(); + if (idle_time != -1ULL) { + dbs_tuners_ins.up_threshold = DEFAULT_UP_FREQ_MIN_LOAD; + } else { + pr_err("cpufreq-hotplug: %s: assumes CONFIG_NO_HZ\n", + __func__); + return -EINVAL; + } + + khotplug_wq = create_workqueue("khotplug"); + if (!khotplug_wq) { + pr_err("Creation of khotplug failed\n"); + return -EFAULT; + } + err = cpufreq_register_governor(&cpufreq_gov_hotplug); + if (err) + destroy_workqueue(khotplug_wq); + + return err; +} + +static void __exit cpufreq_gov_dbs_exit(void) +{ + cpufreq_unregister_governor(&cpufreq_gov_hotplug); + destroy_workqueue(khotplug_wq); +} + +MODULE_AUTHOR("Mike Turquette "); +MODULE_DESCRIPTION("'cpufreq_hotplug' - cpufreq governor for dynamic frequency scaling and CPU hotplugging"); +MODULE_LICENSE("GPL"); + +#ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_HOTPLUG +fs_initcall(cpufreq_gov_dbs_init); +#else +module_init(cpufreq_gov_dbs_init); +#endif +module_exit(cpufreq_gov_dbs_exit); diff --git a/include/linux/cpufreq.h b/include/linux/cpufreq.h index a38fca8..6cbc3df 100644 --- a/include/linux/cpufreq.h +++ b/include/linux/cpufreq.h @@ -355,6 +355,9 @@ extern struct cpufreq_governor cpufreq_gov_ondemand; #elif defined(CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE) extern struct cpufreq_governor cpufreq_gov_conservative; #define CPUFREQ_DEFAULT_GOVERNOR (&cpufreq_gov_conservative) +#elif defined(CONFIG_CPU_FREQ_DEFAULT_GOV_HOTPLUG) +extern struct cpufreq_governor cpufreq_gov_hotplug; +#define CPUFREQ_DEFAULT_GOVERNOR (&cpufreq_gov_hotplug) #endif -- 1.6.6.1