cpufreq: interactive: New 'interactive' governor
Mike Chan [Tue, 22 Jun 2010 18:26:45 +0000 (11:26 -0700)]
This governor is designed for latency-sensitive workloads, such as
interactive user interfaces.  The interactive governor aims to be
significantly more responsive to ramp CPU quickly up when CPU-intensive
activity begins.

Existing governors sample CPU load at a particular rate, typically
every X ms.  This can lead to under-powering UI threads for the period of
time during which the user begins interacting with a previously-idle system
until the next sample period happens.

The 'interactive' governor uses a different approach. Instead of sampling
the CPU at a specified rate, the governor will check whether to scale the
CPU frequency up soon after coming out of idle.  When the CPU comes out of
idle, a timer is configured to fire within 1-2 ticks.  If the CPU is very
busy from exiting idle to when the timer fires then we assume the CPU is
underpowered and ramp to MAX speed.

If the CPU was not sufficiently busy to immediately ramp to MAX speed, then
the governor evaluates the CPU load since the last speed adjustment,
choosing the highest value between that longer-term load or the short-term
load since idle exit to determine the CPU speed to ramp to.

A realtime thread is used for scaling up, giving the remaining tasks the
CPU performance benefit, unlike existing governors which are more likely to
schedule rampup work to occur after your performance starved tasks have
completed.

The tuneables for this governor are:
/sys/devices/system/cpu/cpufreq/interactive/min_sample_time:
The minimum amount of time to spend at the current frequency before
ramping down. This is to ensure that the governor has seen enough
historic CPU load data to determine the appropriate workload.
Default is 80000 uS.
/sys/devices/system/cpu/cpufreq/interactive/go_maxspeed_load
The CPU load at which to ramp to max speed.  Default is 85.

Change-Id: Ib2b362607c62f7c56d35f44a9ef3280f98c17585
Signed-off-by: Mike Chan <mike@android.com>
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Bug: 3152864

Documentation/cpu-freq/governors.txt
drivers/cpufreq/Kconfig
drivers/cpufreq/Makefile
drivers/cpufreq/cpufreq_interactive.c [new file with mode: 0644]
include/linux/cpufreq.h

index e74d0a2..16799ce 100644 (file)
@@ -28,6 +28,7 @@ Contents:
 2.3  Userspace
 2.4  Ondemand
 2.5  Conservative
+2.6  Interactive
 
 3.   The Governor Interface in the CPUfreq Core
 
@@ -193,6 +194,41 @@ 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 Interactive
+---------------
+
+The CPUfreq governor "interactive" is designed for latency-sensitive,
+interactive workloads. This governor sets the CPU speed depending on
+usage, similar to "ondemand" and "conservative" governors.  However,
+the governor is more aggressive about scaling the CPU speed up in
+response to CPU-intensive activity.
+
+Sampling the CPU load every X ms can lead to under-powering the CPU
+for X ms, leading to dropped frames, stuttering UI, etc.  Instead of
+sampling the cpu at a specified rate, the interactive governor will
+check whether to scale the cpu frequency up soon after coming out of
+idle.  When the cpu comes out of idle, a timer is configured to fire
+within 1-2 ticks.  If the cpu is very busy between exiting idle and
+when the timer fires then we assume the cpu is underpowered and ramp
+to MAX speed.
+    
+If the cpu was not sufficiently busy to immediately ramp to MAX speed,
+then governor evaluates the cpu load since the last speed adjustment,
+choosing th highest value between that longer-term load or the
+short-term load since idle exit to determine the cpu speed to ramp to.
+
+The tuneable value for this governor are:
+
+min_sample_time: The minimum amount of time to spend at the current
+frequency before ramping down. This is to ensure that the governor has
+seen enough historic cpu load data to determine the appropriate
+workload.  Default is 80000 uS.
+
+go_maxspeed_load: The CPU load at which to ramp to max speed.  Default
+is 85.
+
+
 3. The Governor Interface in the CPUfreq Core
 =============================================
 
index e24a2a1..2cdc38b 100644 (file)
@@ -99,6 +99,16 @@ 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_INTERACTIVE
+       bool "interactive"
+       select CPU_FREQ_GOV_INTERACTIVE
+       help
+         Use the CPUFreq governor 'interactive' as default. This allows
+         you to get a full dynamic cpu frequency capable system by simply
+         loading your cpufreq low-level hardware driver, using the
+         'interactive' governor for latency-sensitive workloads.
+
 endchoice
 
 config CPU_FREQ_GOV_PERFORMANCE
@@ -156,6 +166,12 @@ config CPU_FREQ_GOV_ONDEMAND
 
          If in doubt, say N.
 
+config CPU_FREQ_GOV_INTERACTIVE
+       tristate "'interactive' cpufreq policy governor"
+       help
+         'interactive' - This driver adds a dynamic cpufreq policy governor
+         designed for latency-sensitive workloads.
+
 config CPU_FREQ_GOV_CONSERVATIVE
        tristate "'conservative' cpufreq governor"
        depends on CPU_FREQ
index a48bc02..d43b391 100644 (file)
@@ -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_INTERACTIVE) += cpufreq_interactive.o
 
 # CPUfreq cross-arch helpers
 obj-$(CONFIG_CPU_FREQ_TABLE)           += freq_table.o
diff --git a/drivers/cpufreq/cpufreq_interactive.c b/drivers/cpufreq/cpufreq_interactive.c
new file mode 100644 (file)
index 0000000..6069ca2
--- /dev/null
@@ -0,0 +1,681 @@
+/*
+ * drivers/cpufreq/cpufreq_interactive.c
+ *
+ * Copyright (C) 2010 Google, Inc.
+ *
+ * This software is licensed under the terms of the GNU General Public
+ * License version 2, as published by the Free Software Foundation, and
+ * may be copied, distributed, and modified under those terms.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * Author: Mike Chan (mike@android.com)
+ *
+ */
+
+#include <linux/cpu.h>
+#include <linux/cpumask.h>
+#include <linux/cpufreq.h>
+#include <linux/mutex.h>
+#include <linux/sched.h>
+#include <linux/tick.h>
+#include <linux/timer.h>
+#include <linux/workqueue.h>
+#include <linux/kthread.h>
+
+#include <asm/cputime.h>
+
+static void (*pm_idle_old)(void);
+static atomic_t active_count = ATOMIC_INIT(0);
+
+struct cpufreq_interactive_cpuinfo {
+       struct timer_list cpu_timer;
+       int timer_idlecancel;
+       u64 time_in_idle;
+       u64 idle_exit_time;
+       u64 timer_run_time;
+       int idling;
+       u64 freq_change_time;
+       u64 freq_change_time_in_idle;
+       struct cpufreq_policy *policy;
+       struct cpufreq_frequency_table *freq_table;
+       unsigned int target_freq;
+       int governor_enabled;
+};
+
+static DEFINE_PER_CPU(struct cpufreq_interactive_cpuinfo, cpuinfo);
+
+/* Workqueues handle frequency scaling */
+static struct task_struct *up_task;
+static struct workqueue_struct *down_wq;
+static struct work_struct freq_scale_down_work;
+static cpumask_t up_cpumask;
+static spinlock_t up_cpumask_lock;
+static cpumask_t down_cpumask;
+static spinlock_t down_cpumask_lock;
+
+/* Go to max speed when CPU load at or above this value. */
+#define DEFAULT_GO_MAXSPEED_LOAD 85
+static unsigned long go_maxspeed_load;
+
+/*
+ * The minimum amount of time to spend at a frequency before we can ramp down.
+ */
+#define DEFAULT_MIN_SAMPLE_TIME 80000;
+static unsigned long min_sample_time;
+
+#define DEBUG 0
+#define BUFSZ 128
+
+#if DEBUG
+#include <linux/proc_fs.h>
+
+struct dbgln {
+       int cpu;
+       unsigned long jiffy;
+       unsigned long run;
+       char buf[BUFSZ];
+};
+
+#define NDBGLNS 256
+
+static struct dbgln dbgbuf[NDBGLNS];
+static int dbgbufs;
+static int dbgbufe;
+static struct proc_dir_entry   *dbg_proc;
+static spinlock_t dbgpr_lock;
+
+static u64 up_request_time;
+static unsigned int up_max_latency;
+
+static void dbgpr(char *fmt, ...)
+{
+       va_list args;
+       int n;
+       unsigned long flags;
+
+       spin_lock_irqsave(&dbgpr_lock, flags);
+       n = dbgbufe;
+        va_start(args, fmt);
+        vsnprintf(dbgbuf[n].buf, BUFSZ, fmt, args);
+        va_end(args);
+       dbgbuf[n].cpu = smp_processor_id();
+       dbgbuf[n].run = nr_running();
+       dbgbuf[n].jiffy = jiffies;
+
+       if (++dbgbufe >= NDBGLNS)
+               dbgbufe = 0;
+
+       if (dbgbufe == dbgbufs)
+               if (++dbgbufs >= NDBGLNS)
+                       dbgbufs = 0;
+
+       spin_unlock_irqrestore(&dbgpr_lock, flags);
+}
+
+static void dbgdump(void)
+{
+       int i, j;
+       unsigned long flags;
+       static struct dbgln prbuf[NDBGLNS];
+
+       spin_lock_irqsave(&dbgpr_lock, flags);
+       i = dbgbufs;
+       j = dbgbufe;
+       memcpy(prbuf, dbgbuf, sizeof(dbgbuf));
+       dbgbufs = 0;
+       dbgbufe = 0;
+       spin_unlock_irqrestore(&dbgpr_lock, flags);
+
+       while (i != j)
+       {
+               printk("%lu %d %lu %s",
+                      prbuf[i].jiffy, prbuf[i].cpu, prbuf[i].run,
+                      prbuf[i].buf);
+               if (++i == NDBGLNS)
+                       i = 0;
+       }
+}
+
+static int dbg_proc_read(char *buffer, char **start, off_t offset,
+                              int count, int *peof, void *dat)
+{
+       printk("max up_task latency=%uus\n", up_max_latency);
+       dbgdump();
+       *peof = 1;
+       return 0;
+}
+
+
+#else
+#define dbgpr(...) do {} while (0)
+#endif
+
+static int cpufreq_governor_interactive(struct cpufreq_policy *policy,
+               unsigned int event);
+
+#ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_INTERACTIVE
+static
+#endif
+struct cpufreq_governor cpufreq_gov_interactive = {
+       .name = "interactive",
+       .governor = cpufreq_governor_interactive,
+       .max_transition_latency = 10000000,
+       .owner = THIS_MODULE,
+};
+
+static void cpufreq_interactive_timer(unsigned long data)
+{
+       unsigned int delta_idle;
+       unsigned int delta_time;
+       int cpu_load;
+       int load_since_change;
+       u64 time_in_idle;
+       u64 idle_exit_time;
+       struct cpufreq_interactive_cpuinfo *pcpu =
+               &per_cpu(cpuinfo, data);
+       u64 now_idle;
+       unsigned int new_freq;
+       unsigned int index;
+
+       /*
+        * Once pcpu->timer_run_time is updated to >= pcpu->idle_exit_time,
+        * this lets idle exit know the current idle time sample has
+        * been processed, and idle exit can generate a new sample and
+        * re-arm the timer.  This prevents a concurrent idle
+        * exit on that CPU from writing a new set of info at the same time
+        * the timer function runs (the timer function can't use that info
+        * until more time passes).
+        */
+       time_in_idle = pcpu->time_in_idle;
+       idle_exit_time = pcpu->idle_exit_time;
+       now_idle = get_cpu_idle_time_us(data, &pcpu->timer_run_time);
+       smp_wmb();
+
+       /* If we raced with cancelling a timer, skip. */
+       if (!idle_exit_time) {
+               dbgpr("timer %d: no valid idle exit sample\n", (int) data);
+               goto exit;
+       }
+
+#if DEBUG
+       if ((int) jiffies - (int) pcpu->cpu_timer.expires >= 10)
+               dbgpr("timer %d: late by %d ticks\n",
+                     (int) data, jiffies - pcpu->cpu_timer.expires);
+#endif
+
+       delta_idle = (unsigned int) cputime64_sub(now_idle, time_in_idle);
+       delta_time = (unsigned int) cputime64_sub(pcpu->timer_run_time,
+                                                 idle_exit_time);
+
+       /*
+        * If timer ran less than 1ms after short-term sample started, retry.
+        */
+       if (delta_time < 1000) {
+               dbgpr("timer %d: time delta %u too short exit=%llu now=%llu\n", (int) data,
+                     delta_time, idle_exit_time, pcpu->timer_run_time);
+               goto rearm;
+       }
+
+       if (delta_idle > delta_time)
+               cpu_load = 0;
+       else
+               cpu_load = 100 * (delta_time - delta_idle) / delta_time;
+
+       delta_idle = (unsigned int) cputime64_sub(now_idle,
+                                                pcpu->freq_change_time_in_idle);
+       delta_time = (unsigned int) cputime64_sub(pcpu->timer_run_time,
+                                                 pcpu->freq_change_time);
+
+       if (delta_idle > delta_time)
+               load_since_change = 0;
+       else
+               load_since_change =
+                       100 * (delta_time - delta_idle) / delta_time;
+
+       /*
+        * Choose greater of short-term load (since last idle timer
+        * started or timer function re-armed itself) or long-term load
+        * (since last frequency change).
+        */
+       if (load_since_change > cpu_load)
+               cpu_load = load_since_change;
+
+       if (cpu_load >= go_maxspeed_load)
+               new_freq = pcpu->policy->max;
+       else
+               new_freq = pcpu->policy->max * cpu_load / 100;
+
+       if (cpufreq_frequency_table_target(pcpu->policy, pcpu->freq_table,
+                                          new_freq, CPUFREQ_RELATION_H,
+                                          &index)) {
+               dbgpr("timer %d: cpufreq_frequency_table_target error\n", (int) data);
+               goto rearm;
+       }
+
+       new_freq = pcpu->freq_table[index].frequency;
+
+       if (pcpu->target_freq == new_freq)
+       {
+               dbgpr("timer %d: load=%d, already at %d\n", (int) data, cpu_load, new_freq);
+               goto rearm_if_notmax;
+       }
+
+       /*
+        * Do not scale down unless we have been at this frequency for the
+        * minimum sample time.
+        */
+       if (new_freq < pcpu->target_freq) {
+               if (cputime64_sub(pcpu->timer_run_time, pcpu->freq_change_time) <
+                   min_sample_time) {
+                       dbgpr("timer %d: load=%d cur=%d tgt=%d not yet\n", (int) data, cpu_load, pcpu->target_freq, new_freq);
+                       goto rearm;
+               }
+       }
+
+       dbgpr("timer %d: load=%d cur=%d tgt=%d queue\n", (int) data, cpu_load, pcpu->target_freq, new_freq);
+
+       if (new_freq < pcpu->target_freq) {
+               pcpu->target_freq = new_freq;
+               spin_lock(&down_cpumask_lock);
+               cpumask_set_cpu(data, &down_cpumask);
+               spin_unlock(&down_cpumask_lock);
+               queue_work(down_wq, &freq_scale_down_work);
+       } else {
+               pcpu->target_freq = new_freq;
+#if DEBUG
+               up_request_time = ktime_to_us(ktime_get());
+#endif
+               spin_lock(&up_cpumask_lock);
+               cpumask_set_cpu(data, &up_cpumask);
+               spin_unlock(&up_cpumask_lock);
+               wake_up_process(up_task);
+       }
+
+rearm_if_notmax:
+       /*
+        * Already set max speed and don't see a need to change that,
+        * wait until next idle to re-evaluate, don't need timer.
+        */
+       if (pcpu->target_freq == pcpu->policy->max)
+               goto exit;
+
+rearm:
+       if (!timer_pending(&pcpu->cpu_timer)) {
+               /*
+                * If already at min: if that CPU is idle, don't set timer.
+                * Else cancel the timer if that CPU goes idle.  We don't
+                * need to re-evaluate speed until the next idle exit.
+                */
+               if (pcpu->target_freq == pcpu->policy->min) {
+                       smp_rmb();
+
+                       if (pcpu->idling) {
+                               dbgpr("timer %d: cpu idle, don't re-arm\n", (int) data);
+                               goto exit;
+                       }
+
+                       pcpu->timer_idlecancel = 1;
+               }
+
+               pcpu->time_in_idle = get_cpu_idle_time_us(
+                       data, &pcpu->idle_exit_time);
+               mod_timer(&pcpu->cpu_timer, jiffies + 2);
+               dbgpr("timer %d: set timer for %lu exit=%llu\n", (int) data, pcpu->cpu_timer.expires, pcpu->idle_exit_time);
+       }
+
+exit:
+       return;
+}
+
+static void cpufreq_interactive_idle(void)
+{
+       struct cpufreq_interactive_cpuinfo *pcpu =
+               &per_cpu(cpuinfo, smp_processor_id());
+       int pending;
+
+       if (!pcpu->governor_enabled) {
+               pm_idle_old();
+               return;
+       }
+
+       pcpu->idling = 1;
+       smp_wmb();
+       pending = timer_pending(&pcpu->cpu_timer);
+
+       if (pcpu->target_freq != pcpu->policy->min) {
+#ifdef CONFIG_SMP
+               /*
+                * Entering idle while not at lowest speed.  On some
+                * platforms this can hold the other CPU(s) at that speed
+                * even though the CPU is idle. Set a timer to re-evaluate
+                * speed so this idle CPU doesn't hold the other CPUs above
+                * min indefinitely.  This should probably be a quirk of
+                * the CPUFreq driver.
+                */
+               if (!pending) {
+                       pcpu->time_in_idle = get_cpu_idle_time_us(
+                               smp_processor_id(), &pcpu->idle_exit_time);
+                       pcpu->timer_idlecancel = 0;
+                       mod_timer(&pcpu->cpu_timer, jiffies + 2);
+                       dbgpr("idle: enter at %d, set timer for %lu exit=%llu\n",
+                             pcpu->target_freq, pcpu->cpu_timer.expires,
+                             pcpu->idle_exit_time);
+               }
+#endif
+       } else {
+               /*
+                * If at min speed and entering idle after load has
+                * already been evaluated, and a timer has been set just in
+                * case the CPU suddenly goes busy, cancel that timer.  The
+                * CPU didn't go busy; we'll recheck things upon idle exit.
+                */
+               if (pending && pcpu->timer_idlecancel) {
+                       dbgpr("idle: cancel timer for %lu\n", pcpu->cpu_timer.expires);
+                       del_timer(&pcpu->cpu_timer);
+                       /*
+                        * Ensure last timer run time is after current idle
+                        * sample start time, so next idle exit will always
+                        * start a new idle sampling period.
+                        */
+                       pcpu->idle_exit_time = 0;
+                       pcpu->timer_idlecancel = 0;
+               }
+       }
+
+       pm_idle_old();
+       pcpu->idling = 0;
+       smp_wmb();
+
+       /*
+        * Arm the timer for 1-2 ticks later if not already, and if the timer
+        * function has already processed the previous load sampling
+        * interval.  (If the timer is not pending but has not processed
+        * the previous interval, it is probably racing with us on another
+        * CPU.  Let it compute load based on the previous sample and then
+        * re-arm the timer for another interval when it's done, rather
+        * than updating the interval start time to be "now", which doesn't
+        * give the timer function enough time to make a decision on this
+        * run.)
+        */
+       if (timer_pending(&pcpu->cpu_timer) == 0 &&
+           pcpu->timer_run_time >= pcpu->idle_exit_time) {
+               pcpu->time_in_idle =
+                       get_cpu_idle_time_us(smp_processor_id(),
+                                            &pcpu->idle_exit_time);
+               pcpu->timer_idlecancel = 0;
+               mod_timer(&pcpu->cpu_timer, jiffies + 2);
+               dbgpr("idle: exit, set timer for %lu exit=%llu\n", pcpu->cpu_timer.expires, pcpu->idle_exit_time);
+#if DEBUG
+       } else if (timer_pending(&pcpu->cpu_timer) == 0 &&
+                  pcpu->timer_run_time < pcpu->idle_exit_time) {
+               dbgpr("idle: timer not run yet: exit=%llu tmrrun=%llu\n",
+                     pcpu->idle_exit_time, pcpu->timer_run_time);
+#endif
+       }
+
+}
+
+static int cpufreq_interactive_up_task(void *data)
+{
+       unsigned int cpu;
+       cpumask_t tmp_mask;
+       struct cpufreq_interactive_cpuinfo *pcpu;
+
+#if DEBUG
+       u64 now;
+       u64 then;
+       unsigned int lat;
+#endif
+
+       while (1) {
+               set_current_state(TASK_INTERRUPTIBLE);
+               spin_lock(&up_cpumask_lock);
+
+               if (cpumask_empty(&up_cpumask)) {
+                       spin_unlock(&up_cpumask_lock);
+                       schedule();
+
+                       if (kthread_should_stop())
+                               break;
+
+                       spin_lock(&up_cpumask_lock);
+               }
+
+               set_current_state(TASK_RUNNING);
+
+#if DEBUG
+               then = up_request_time;
+               now = ktime_to_us(ktime_get());
+
+               if (now > then) {
+                       lat = ktime_to_us(ktime_get()) - then;
+
+                       if (lat > up_max_latency)
+                               up_max_latency = lat;
+               }
+#endif
+
+               tmp_mask = up_cpumask;
+               cpumask_clear(&up_cpumask);
+               spin_unlock(&up_cpumask_lock);
+
+               for_each_cpu(cpu, &tmp_mask) {
+                       pcpu = &per_cpu(cpuinfo, cpu);
+
+                       if (nr_running() == 1) {
+                               dbgpr("up %d: tgt=%d nothing else running\n", cpu,
+                                     pcpu->target_freq);
+                       }
+
+                       __cpufreq_driver_target(pcpu->policy,
+                                               pcpu->target_freq,
+                                               CPUFREQ_RELATION_H);
+                       pcpu->freq_change_time_in_idle =
+                               get_cpu_idle_time_us(cpu,
+                                                    &pcpu->freq_change_time);
+                       dbgpr("up %d: set tgt=%d (actual=%d)\n", cpu, pcpu->target_freq, pcpu->policy->cur);
+               }
+       }
+
+       return 0;
+}
+
+static void cpufreq_interactive_freq_down(struct work_struct *work)
+{
+       unsigned int cpu;
+       cpumask_t tmp_mask;
+       struct cpufreq_interactive_cpuinfo *pcpu;
+
+       spin_lock(&down_cpumask_lock);
+       tmp_mask = down_cpumask;
+       cpumask_clear(&down_cpumask);
+       spin_unlock(&down_cpumask_lock);
+
+       for_each_cpu(cpu, &tmp_mask) {
+               pcpu = &per_cpu(cpuinfo, cpu);
+               __cpufreq_driver_target(pcpu->policy,
+                                       pcpu->target_freq,
+                                       CPUFREQ_RELATION_H);
+               pcpu->freq_change_time_in_idle =
+                       get_cpu_idle_time_us(cpu,
+                                            &pcpu->freq_change_time);
+               dbgpr("down %d: set tgt=%d (actual=%d)\n", cpu, pcpu->target_freq, pcpu->policy->cur);
+       }
+}
+
+static ssize_t show_go_maxspeed_load(struct kobject *kobj,
+                                    struct attribute *attr, char *buf)
+{
+       return sprintf(buf, "%lu\n", go_maxspeed_load);
+}
+
+static ssize_t store_go_maxspeed_load(struct kobject *kobj,
+                       struct attribute *attr, const char *buf, size_t count)
+{
+       return strict_strtoul(buf, 0, &go_maxspeed_load);
+}
+
+static struct global_attr go_maxspeed_load_attr = __ATTR(go_maxspeed_load, 0644,
+               show_go_maxspeed_load, store_go_maxspeed_load);
+
+static ssize_t show_min_sample_time(struct kobject *kobj,
+                               struct attribute *attr, char *buf)
+{
+       return sprintf(buf, "%lu\n", min_sample_time);
+}
+
+static ssize_t store_min_sample_time(struct kobject *kobj,
+                       struct attribute *attr, const char *buf, size_t count)
+{
+       return strict_strtoul(buf, 0, &min_sample_time);
+}
+
+static struct global_attr min_sample_time_attr = __ATTR(min_sample_time, 0644,
+               show_min_sample_time, store_min_sample_time);
+
+static struct attribute *interactive_attributes[] = {
+       &go_maxspeed_load_attr.attr,
+       &min_sample_time_attr.attr,
+       NULL,
+};
+
+static struct attribute_group interactive_attr_group = {
+       .attrs = interactive_attributes,
+       .name = "interactive",
+};
+
+static int cpufreq_governor_interactive(struct cpufreq_policy *new_policy,
+               unsigned int event)
+{
+       int rc;
+       struct cpufreq_interactive_cpuinfo *pcpu =
+               &per_cpu(cpuinfo, new_policy->cpu);
+
+       switch (event) {
+       case CPUFREQ_GOV_START:
+               if (!cpu_online(new_policy->cpu))
+                       return -EINVAL;
+
+               pcpu->policy = new_policy;
+               pcpu->freq_table = cpufreq_frequency_get_table(new_policy->cpu);
+               pcpu->target_freq = new_policy->cur;
+               pcpu->freq_change_time_in_idle =
+                       get_cpu_idle_time_us(new_policy->cpu,
+                                            &pcpu->freq_change_time);
+               pcpu->governor_enabled = 1;
+               /*
+                * Do not register the idle hook and create sysfs
+                * entries if we have already done so.
+                */
+               if (atomic_inc_return(&active_count) > 1)
+                       return 0;
+
+               rc = sysfs_create_group(cpufreq_global_kobject,
+                               &interactive_attr_group);
+               if (rc)
+                       return rc;
+
+               pm_idle_old = pm_idle;
+               pm_idle = cpufreq_interactive_idle;
+               break;
+
+       case CPUFREQ_GOV_STOP:
+               pcpu->governor_enabled = 0;
+
+               if (atomic_dec_return(&active_count) > 0)
+                       return 0;
+
+               sysfs_remove_group(cpufreq_global_kobject,
+                               &interactive_attr_group);
+
+               pm_idle = pm_idle_old;
+               del_timer(&pcpu->cpu_timer);
+               break;
+
+       case CPUFREQ_GOV_LIMITS:
+               if (new_policy->max < new_policy->cur)
+                       __cpufreq_driver_target(new_policy,
+                                       new_policy->max, CPUFREQ_RELATION_H);
+               else if (new_policy->min > new_policy->cur)
+                       __cpufreq_driver_target(new_policy,
+                                       new_policy->min, CPUFREQ_RELATION_L);
+               break;
+       }
+       return 0;
+}
+
+static int __init cpufreq_interactive_init(void)
+{
+       unsigned int i;
+       struct cpufreq_interactive_cpuinfo *pcpu;
+       struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
+
+       go_maxspeed_load = DEFAULT_GO_MAXSPEED_LOAD;
+       min_sample_time = DEFAULT_MIN_SAMPLE_TIME;
+
+       /* Initalize per-cpu timers */
+       for_each_possible_cpu(i) {
+               pcpu = &per_cpu(cpuinfo, i);
+               init_timer(&pcpu->cpu_timer);
+               pcpu->cpu_timer.function = cpufreq_interactive_timer;
+               pcpu->cpu_timer.data = i;
+       }
+
+       up_task = kthread_create(cpufreq_interactive_up_task, NULL,
+                                "kinteractiveup");
+       if (IS_ERR(up_task))
+               return PTR_ERR(up_task);
+
+       sched_setscheduler_nocheck(up_task, SCHED_FIFO, &param);
+       get_task_struct(up_task);
+
+       /* No rescuer thread, bind to CPU queuing the work for possibly
+          warm cache (probably doesn't matter much). */
+       down_wq = alloc_workqueue("knteractive_down", 0, 1);
+
+       if (! down_wq)
+               goto err_freeuptask;
+
+       INIT_WORK(&freq_scale_down_work,
+                 cpufreq_interactive_freq_down);
+
+       spin_lock_init(&up_cpumask_lock);
+       spin_lock_init(&down_cpumask_lock);
+
+#if DEBUG
+       spin_lock_init(&dbgpr_lock);
+       dbg_proc = create_proc_entry("igov", S_IWUSR | S_IRUGO, NULL);
+       dbg_proc->read_proc = dbg_proc_read;
+#endif
+
+       return cpufreq_register_governor(&cpufreq_gov_interactive);
+
+err_freeuptask:
+       put_task_struct(up_task);
+       return -ENOMEM;
+}
+
+#ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_INTERACTIVE
+fs_initcall(cpufreq_interactive_init);
+#else
+module_init(cpufreq_interactive_init);
+#endif
+
+static void __exit cpufreq_interactive_exit(void)
+{
+       cpufreq_unregister_governor(&cpufreq_gov_interactive);
+       kthread_stop(up_task);
+       put_task_struct(up_task);
+       destroy_workqueue(down_wq);
+}
+
+module_exit(cpufreq_interactive_exit);
+
+MODULE_AUTHOR("Mike Chan <mike@android.com>");
+MODULE_DESCRIPTION("'cpufreq_interactive' - A cpufreq governor for "
+       "Latency sensitive workloads");
+MODULE_LICENSE("GPL");
index 6216115..c6126b9 100644 (file)
@@ -363,6 +363,9 @@ 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_INTERACTIVE)
+extern struct cpufreq_governor cpufreq_gov_interactive;
+#define CPUFREQ_DEFAULT_GOVERNOR       (&cpufreq_gov_interactive)
 #endif