linux_dsm_epyc7002/drivers/cpufreq/cpufreq_governor.h

185 lines
5.6 KiB
C
Raw Normal View History

/*
* drivers/cpufreq/cpufreq_governor.h
*
* Header file for CPUFreq governors common code
*
* Copyright (C) 2001 Russell King
* (C) 2003 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>.
* (C) 2003 Jun Nakajima <jun.nakajima@intel.com>
* (C) 2009 Alexander Clouter <alex@digriz.org.uk>
* (c) 2012 Viresh Kumar <viresh.kumar@linaro.org>
*
* 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.
*/
#ifndef _CPUFREQ_GOVERNOR_H
#define _CPUFREQ_GOVERNOR_H
#include <linux/atomic.h>
#include <linux/irq_work.h>
#include <linux/cpufreq.h>
#include <linux/kernel_stat.h>
#include <linux/module.h>
#include <linux/mutex.h>
/* Ondemand Sampling types */
enum {OD_NORMAL_SAMPLE, OD_SUB_SAMPLE};
/*
* Abbreviations:
* dbs: used as a shortform for demand based switching It helps to keep variable
* names smaller, simpler
* cdbs: common dbs
* od_*: On-demand governor
* cs_*: Conservative governor
*/
/* Governor demand based switching data (per-policy or global). */
struct dbs_data {
struct gov_attr_set attr_set;
void *tuners;
unsigned int min_sampling_rate;
unsigned int ignore_nice_load;
unsigned int sampling_rate;
unsigned int sampling_down_factor;
unsigned int up_threshold;
unsigned int io_is_busy;
cpufreq: governor: New sysfs show/store callbacks for governor tunables The ondemand and conservative governors use the global-attr or freq-attr structures to represent sysfs attributes corresponding to their tunables (which of them is actually used depends on whether or not different policy objects can use the same governor with different tunables at the same time and, consequently, on where those attributes are located in sysfs). Unfortunately, in the freq-attr case, the standard cpufreq show/store sysfs attribute callbacks are applied to the governor tunable attributes and they always acquire the policy->rwsem lock before carrying out the operation. That may lead to an ABBA deadlock if governor tunable attributes are removed under policy->rwsem while one of them is being accessed concurrently (if sysfs attributes removal wins the race, it will wait for the access to complete with policy->rwsem held while the attribute callback will block on policy->rwsem indefinitely). We attempted to address this issue by dropping policy->rwsem around governor tunable attributes removal (that is, around invocations of the ->governor callback with the event arg equal to CPUFREQ_GOV_POLICY_EXIT) in cpufreq_set_policy(), but that opened up race conditions that had not been possible with policy->rwsem held all the time. Therefore policy->rwsem cannot be dropped in cpufreq_set_policy() at any point, but the deadlock situation described above must be avoided too. To that end, use the observation that in principle governor tunables may be represented by the same data type regardless of whether the governor is system-wide or per-policy and introduce a new structure, struct governor_attr, for representing them and new corresponding macros for creating show/store sysfs callbacks for them. Also make their parent kobject use a new kobject type whose default show/store callbacks are not related to the standard core cpufreq ones in any way (and they don't acquire policy->rwsem in particular). Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Tested-by: Juri Lelli <juri.lelli@arm.com> Tested-by: Shilpasri G Bhat <shilpa.bhat@linux.vnet.ibm.com> [ rjw: Subject & changelog + rebase ] Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2016-02-09 10:31:33 +07:00
};
static inline struct dbs_data *to_dbs_data(struct gov_attr_set *attr_set)
{
return container_of(attr_set, struct dbs_data, attr_set);
}
cpufreq: governor: New sysfs show/store callbacks for governor tunables The ondemand and conservative governors use the global-attr or freq-attr structures to represent sysfs attributes corresponding to their tunables (which of them is actually used depends on whether or not different policy objects can use the same governor with different tunables at the same time and, consequently, on where those attributes are located in sysfs). Unfortunately, in the freq-attr case, the standard cpufreq show/store sysfs attribute callbacks are applied to the governor tunable attributes and they always acquire the policy->rwsem lock before carrying out the operation. That may lead to an ABBA deadlock if governor tunable attributes are removed under policy->rwsem while one of them is being accessed concurrently (if sysfs attributes removal wins the race, it will wait for the access to complete with policy->rwsem held while the attribute callback will block on policy->rwsem indefinitely). We attempted to address this issue by dropping policy->rwsem around governor tunable attributes removal (that is, around invocations of the ->governor callback with the event arg equal to CPUFREQ_GOV_POLICY_EXIT) in cpufreq_set_policy(), but that opened up race conditions that had not been possible with policy->rwsem held all the time. Therefore policy->rwsem cannot be dropped in cpufreq_set_policy() at any point, but the deadlock situation described above must be avoided too. To that end, use the observation that in principle governor tunables may be represented by the same data type regardless of whether the governor is system-wide or per-policy and introduce a new structure, struct governor_attr, for representing them and new corresponding macros for creating show/store sysfs callbacks for them. Also make their parent kobject use a new kobject type whose default show/store callbacks are not related to the standard core cpufreq ones in any way (and they don't acquire policy->rwsem in particular). Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Tested-by: Juri Lelli <juri.lelli@arm.com> Tested-by: Shilpasri G Bhat <shilpa.bhat@linux.vnet.ibm.com> [ rjw: Subject & changelog + rebase ] Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2016-02-09 10:31:33 +07:00
#define gov_show_one(_gov, file_name) \
static ssize_t show_##file_name \
(struct gov_attr_set *attr_set, char *buf) \
cpufreq: governor: New sysfs show/store callbacks for governor tunables The ondemand and conservative governors use the global-attr or freq-attr structures to represent sysfs attributes corresponding to their tunables (which of them is actually used depends on whether or not different policy objects can use the same governor with different tunables at the same time and, consequently, on where those attributes are located in sysfs). Unfortunately, in the freq-attr case, the standard cpufreq show/store sysfs attribute callbacks are applied to the governor tunable attributes and they always acquire the policy->rwsem lock before carrying out the operation. That may lead to an ABBA deadlock if governor tunable attributes are removed under policy->rwsem while one of them is being accessed concurrently (if sysfs attributes removal wins the race, it will wait for the access to complete with policy->rwsem held while the attribute callback will block on policy->rwsem indefinitely). We attempted to address this issue by dropping policy->rwsem around governor tunable attributes removal (that is, around invocations of the ->governor callback with the event arg equal to CPUFREQ_GOV_POLICY_EXIT) in cpufreq_set_policy(), but that opened up race conditions that had not been possible with policy->rwsem held all the time. Therefore policy->rwsem cannot be dropped in cpufreq_set_policy() at any point, but the deadlock situation described above must be avoided too. To that end, use the observation that in principle governor tunables may be represented by the same data type regardless of whether the governor is system-wide or per-policy and introduce a new structure, struct governor_attr, for representing them and new corresponding macros for creating show/store sysfs callbacks for them. Also make their parent kobject use a new kobject type whose default show/store callbacks are not related to the standard core cpufreq ones in any way (and they don't acquire policy->rwsem in particular). Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Tested-by: Juri Lelli <juri.lelli@arm.com> Tested-by: Shilpasri G Bhat <shilpa.bhat@linux.vnet.ibm.com> [ rjw: Subject & changelog + rebase ] Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2016-02-09 10:31:33 +07:00
{ \
struct dbs_data *dbs_data = to_dbs_data(attr_set); \
cpufreq: governor: New sysfs show/store callbacks for governor tunables The ondemand and conservative governors use the global-attr or freq-attr structures to represent sysfs attributes corresponding to their tunables (which of them is actually used depends on whether or not different policy objects can use the same governor with different tunables at the same time and, consequently, on where those attributes are located in sysfs). Unfortunately, in the freq-attr case, the standard cpufreq show/store sysfs attribute callbacks are applied to the governor tunable attributes and they always acquire the policy->rwsem lock before carrying out the operation. That may lead to an ABBA deadlock if governor tunable attributes are removed under policy->rwsem while one of them is being accessed concurrently (if sysfs attributes removal wins the race, it will wait for the access to complete with policy->rwsem held while the attribute callback will block on policy->rwsem indefinitely). We attempted to address this issue by dropping policy->rwsem around governor tunable attributes removal (that is, around invocations of the ->governor callback with the event arg equal to CPUFREQ_GOV_POLICY_EXIT) in cpufreq_set_policy(), but that opened up race conditions that had not been possible with policy->rwsem held all the time. Therefore policy->rwsem cannot be dropped in cpufreq_set_policy() at any point, but the deadlock situation described above must be avoided too. To that end, use the observation that in principle governor tunables may be represented by the same data type regardless of whether the governor is system-wide or per-policy and introduce a new structure, struct governor_attr, for representing them and new corresponding macros for creating show/store sysfs callbacks for them. Also make their parent kobject use a new kobject type whose default show/store callbacks are not related to the standard core cpufreq ones in any way (and they don't acquire policy->rwsem in particular). Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Tested-by: Juri Lelli <juri.lelli@arm.com> Tested-by: Shilpasri G Bhat <shilpa.bhat@linux.vnet.ibm.com> [ rjw: Subject & changelog + rebase ] Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2016-02-09 10:31:33 +07:00
struct _gov##_dbs_tuners *tuners = dbs_data->tuners; \
return sprintf(buf, "%u\n", tuners->file_name); \
}
#define gov_show_one_common(file_name) \
static ssize_t show_##file_name \
(struct gov_attr_set *attr_set, char *buf) \
cpufreq: governor: New sysfs show/store callbacks for governor tunables The ondemand and conservative governors use the global-attr or freq-attr structures to represent sysfs attributes corresponding to their tunables (which of them is actually used depends on whether or not different policy objects can use the same governor with different tunables at the same time and, consequently, on where those attributes are located in sysfs). Unfortunately, in the freq-attr case, the standard cpufreq show/store sysfs attribute callbacks are applied to the governor tunable attributes and they always acquire the policy->rwsem lock before carrying out the operation. That may lead to an ABBA deadlock if governor tunable attributes are removed under policy->rwsem while one of them is being accessed concurrently (if sysfs attributes removal wins the race, it will wait for the access to complete with policy->rwsem held while the attribute callback will block on policy->rwsem indefinitely). We attempted to address this issue by dropping policy->rwsem around governor tunable attributes removal (that is, around invocations of the ->governor callback with the event arg equal to CPUFREQ_GOV_POLICY_EXIT) in cpufreq_set_policy(), but that opened up race conditions that had not been possible with policy->rwsem held all the time. Therefore policy->rwsem cannot be dropped in cpufreq_set_policy() at any point, but the deadlock situation described above must be avoided too. To that end, use the observation that in principle governor tunables may be represented by the same data type regardless of whether the governor is system-wide or per-policy and introduce a new structure, struct governor_attr, for representing them and new corresponding macros for creating show/store sysfs callbacks for them. Also make their parent kobject use a new kobject type whose default show/store callbacks are not related to the standard core cpufreq ones in any way (and they don't acquire policy->rwsem in particular). Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Tested-by: Juri Lelli <juri.lelli@arm.com> Tested-by: Shilpasri G Bhat <shilpa.bhat@linux.vnet.ibm.com> [ rjw: Subject & changelog + rebase ] Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2016-02-09 10:31:33 +07:00
{ \
struct dbs_data *dbs_data = to_dbs_data(attr_set); \
cpufreq: governor: New sysfs show/store callbacks for governor tunables The ondemand and conservative governors use the global-attr or freq-attr structures to represent sysfs attributes corresponding to their tunables (which of them is actually used depends on whether or not different policy objects can use the same governor with different tunables at the same time and, consequently, on where those attributes are located in sysfs). Unfortunately, in the freq-attr case, the standard cpufreq show/store sysfs attribute callbacks are applied to the governor tunable attributes and they always acquire the policy->rwsem lock before carrying out the operation. That may lead to an ABBA deadlock if governor tunable attributes are removed under policy->rwsem while one of them is being accessed concurrently (if sysfs attributes removal wins the race, it will wait for the access to complete with policy->rwsem held while the attribute callback will block on policy->rwsem indefinitely). We attempted to address this issue by dropping policy->rwsem around governor tunable attributes removal (that is, around invocations of the ->governor callback with the event arg equal to CPUFREQ_GOV_POLICY_EXIT) in cpufreq_set_policy(), but that opened up race conditions that had not been possible with policy->rwsem held all the time. Therefore policy->rwsem cannot be dropped in cpufreq_set_policy() at any point, but the deadlock situation described above must be avoided too. To that end, use the observation that in principle governor tunables may be represented by the same data type regardless of whether the governor is system-wide or per-policy and introduce a new structure, struct governor_attr, for representing them and new corresponding macros for creating show/store sysfs callbacks for them. Also make their parent kobject use a new kobject type whose default show/store callbacks are not related to the standard core cpufreq ones in any way (and they don't acquire policy->rwsem in particular). Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Tested-by: Juri Lelli <juri.lelli@arm.com> Tested-by: Shilpasri G Bhat <shilpa.bhat@linux.vnet.ibm.com> [ rjw: Subject & changelog + rebase ] Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2016-02-09 10:31:33 +07:00
return sprintf(buf, "%u\n", dbs_data->file_name); \
}
#define gov_attr_ro(_name) \
static struct governor_attr _name = \
__ATTR(_name, 0444, show_##_name, NULL)
#define gov_attr_rw(_name) \
static struct governor_attr _name = \
__ATTR(_name, 0644, show_##_name, store_##_name)
/* Common to all CPUs of a policy */
struct policy_dbs_info {
struct cpufreq_policy *policy;
/*
cpufreq: governor: replace per-CPU delayed work with timers cpufreq governors evaluate load at sampling rate and based on that they update frequency for a group of CPUs belonging to the same cpufreq policy. This is required to be done in a single thread for all policy->cpus, but because we don't want to wakeup idle CPUs to do just that, we use deferrable work for this. If we would have used a single delayed deferrable work for the entire policy, there were chances that the CPU required to run the handler can be in idle and we might end up not changing the frequency for the entire group with load variations. And so we were forced to keep per-cpu works, and only the one that expires first need to do the real work and others are rescheduled for next sampling time. We have been using the more complex solution until now, where we used a delayed deferrable work for this, which is a combination of a timer and a work. This could be made lightweight by keeping per-cpu deferred timers with a single work item, which is scheduled by the first timer that expires. This patch does just that and here are important changes: - The timer handler will run in irq context and so we need to use a spin_lock instead of the timer_mutex. And so a separate timer_lock is created. This also makes the use of the mutex and lock quite clear, as we know what exactly they are protecting. - A new field 'skip_work' is added to track when the timer handlers can queue a work. More comments present in code. Suggested-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Reviewed-by: Ashwin Chaugule <ashwin.chaugule@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2015-12-09 09:04:42 +07:00
* Per policy mutex that serializes load evaluation from limit-change
* and work-handler.
*/
struct mutex timer_mutex;
cpufreq: governor: replace per-CPU delayed work with timers cpufreq governors evaluate load at sampling rate and based on that they update frequency for a group of CPUs belonging to the same cpufreq policy. This is required to be done in a single thread for all policy->cpus, but because we don't want to wakeup idle CPUs to do just that, we use deferrable work for this. If we would have used a single delayed deferrable work for the entire policy, there were chances that the CPU required to run the handler can be in idle and we might end up not changing the frequency for the entire group with load variations. And so we were forced to keep per-cpu works, and only the one that expires first need to do the real work and others are rescheduled for next sampling time. We have been using the more complex solution until now, where we used a delayed deferrable work for this, which is a combination of a timer and a work. This could be made lightweight by keeping per-cpu deferred timers with a single work item, which is scheduled by the first timer that expires. This patch does just that and here are important changes: - The timer handler will run in irq context and so we need to use a spin_lock instead of the timer_mutex. And so a separate timer_lock is created. This also makes the use of the mutex and lock quite clear, as we know what exactly they are protecting. - A new field 'skip_work' is added to track when the timer handlers can queue a work. More comments present in code. Suggested-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Reviewed-by: Ashwin Chaugule <ashwin.chaugule@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2015-12-09 09:04:42 +07:00
u64 last_sample_time;
s64 sample_delay_ns;
atomic_t work_count;
struct irq_work irq_work;
cpufreq: governor: replace per-CPU delayed work with timers cpufreq governors evaluate load at sampling rate and based on that they update frequency for a group of CPUs belonging to the same cpufreq policy. This is required to be done in a single thread for all policy->cpus, but because we don't want to wakeup idle CPUs to do just that, we use deferrable work for this. If we would have used a single delayed deferrable work for the entire policy, there were chances that the CPU required to run the handler can be in idle and we might end up not changing the frequency for the entire group with load variations. And so we were forced to keep per-cpu works, and only the one that expires first need to do the real work and others are rescheduled for next sampling time. We have been using the more complex solution until now, where we used a delayed deferrable work for this, which is a combination of a timer and a work. This could be made lightweight by keeping per-cpu deferred timers with a single work item, which is scheduled by the first timer that expires. This patch does just that and here are important changes: - The timer handler will run in irq context and so we need to use a spin_lock instead of the timer_mutex. And so a separate timer_lock is created. This also makes the use of the mutex and lock quite clear, as we know what exactly they are protecting. - A new field 'skip_work' is added to track when the timer handlers can queue a work. More comments present in code. Suggested-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Reviewed-by: Ashwin Chaugule <ashwin.chaugule@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2015-12-09 09:04:42 +07:00
struct work_struct work;
/* dbs_data may be shared between multiple policy objects */
struct dbs_data *dbs_data;
struct list_head list;
/* Multiplier for increasing sample delay temporarily. */
unsigned int rate_mult;
/* Status indicators */
bool is_shared; /* This object is used by multiple CPUs */
bool work_in_progress; /* Work is being queued up or in progress */
};
static inline void gov_update_sample_delay(struct policy_dbs_info *policy_dbs,
unsigned int delay_us)
{
policy_dbs->sample_delay_ns = delay_us * NSEC_PER_USEC;
}
/* Per cpu structures */
struct cpu_dbs_info {
u64 prev_cpu_idle;
u64 prev_update_time;
u64 prev_cpu_nice;
cpufreq: governor: Be friendly towards latency-sensitive bursty workloads Cpufreq governors like the ondemand governor calculate the load on the CPU periodically by employing deferrable timers. A deferrable timer won't fire if the CPU is completely idle (and there are no other timers to be run), in order to avoid unnecessary wakeups and thus save CPU power. However, the load calculation logic is agnostic to all this, and this can lead to the problem described below. Time (ms) CPU 1 100 Task-A running 110 Governor's timer fires, finds load as 100% in the last 10ms interval and increases the CPU frequency. 110.5 Task-A running 120 Governor's timer fires, finds load as 100% in the last 10ms interval and increases the CPU frequency. 125 Task-A went to sleep. With nothing else to do, CPU 1 went completely idle. 200 Task-A woke up and started running again. 200.5 Governor's deferred timer (which was originally programmed to fire at time 130) fires now. It calculates load for the time period 120 to 200.5, and finds the load is almost zero. Hence it decreases the CPU frequency to the minimum. 210 Governor's timer fires, finds load as 100% in the last 10ms interval and increases the CPU frequency. So, after the workload woke up and started running, the frequency was suddenly dropped to absolute minimum, and after that, there was an unnecessary delay of 10ms (sampling period) to increase the CPU frequency back to a reasonable value. And this pattern repeats for every wake-up-from-cpu-idle for that workload. This can be quite undesirable for latency- or response-time sensitive bursty workloads. So we need to fix the governor's logic to detect such wake-up-from- cpu-idle scenarios and start the workload at a reasonably high CPU frequency. One extreme solution would be to fake a load of 100% in such scenarios. But that might lead to undesirable side-effects such as frequency spikes (which might also need voltage changes) especially if the previous frequency happened to be very low. We just want to avoid the stupidity of dropping down the frequency to a minimum and then enduring a needless (and long) delay before ramping it up back again. So, let us simply carry forward the previous load - that is, let us just pretend that the 'load' for the current time-window is the same as the load for the previous window. That way, the frequency and voltage will continue to be set to whatever values they were set at previously. This means that bursty workloads will get a chance to influence the CPU frequency at which they wake up from cpu-idle, based on their past execution history. Thus, they might be able to avoid suffering from slow wakeups and long response-times. However, we should take care not to over-do this. For example, such a "copy previous load" logic will benefit cases like this: (where # represents busy and . represents idle) ##########.........#########.........###########...........##########........ but it will be detrimental in cases like the one shown below, because it will retain the high frequency (copied from the previous interval) even in a mostly idle system: ##########.........#.................#.....................#............... (i.e., the workload finished and the remaining tasks are such that their busy periods are smaller than the sampling interval, which causes the timer to always get deferred. So, this will make the copy-previous-load logic copy the initial high load to subsequent idle periods over and over again, thus keeping the frequency high unnecessarily). So, we modify this copy-previous-load logic such that it is used only once upon every wakeup-from-idle. Thus if we have 2 consecutive idle periods, the previous load won't get blindly copied over; cpufreq will freshly evaluate the load in the second idle interval, thus ensuring that the system comes back to its normal state. [ The right way to solve this whole problem is to teach the CPU frequency governors to also track load on a per-task basis, not just a per-CPU basis, and then use both the data sources intelligently to set the appropriate frequency on the CPUs. But that involves redesigning the cpufreq subsystem, so this patch should make the situation bearable until then. ] Experimental results: +-------------------+ I ran a modified version of ebizzy (called 'sleeping-ebizzy') that sleeps in between its execution such that its total utilization can be a user-defined value, say 10% or 20% (higher the utilization specified, lesser the amount of sleeps injected). This ebizzy was run with a single-thread, tied to CPU 8. Behavior observed with tracing (sample taken from 40% utilization runs): ------------------------------------------------------------------------ Without patch: ~~~~~~~~~~~~~~ kworker/8:2-12137 416.335742: cpu_frequency: state=2061000 cpu_id=8 kworker/8:2-12137 416.335744: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy <...>-40753 416.345741: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2 kworker/8:2-12137 416.345744: cpu_frequency: state=4123000 cpu_id=8 kworker/8:2-12137 416.345746: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy <...>-40753 416.355738: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2 <snip> --------------------------------------------------------------------- <snip> <...>-40753 416.402202: sched_switch: prev_comm=ebizzy ==> next_comm=swapper/8 <idle>-0 416.502130: sched_switch: prev_comm=swapper/8 ==> next_comm=ebizzy <...>-40753 416.505738: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2 kworker/8:2-12137 416.505739: cpu_frequency: state=2061000 cpu_id=8 kworker/8:2-12137 416.505741: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy <...>-40753 416.515739: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2 kworker/8:2-12137 416.515742: cpu_frequency: state=4123000 cpu_id=8 kworker/8:2-12137 416.515744: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy Observation: Ebizzy went idle at 416.402202, and started running again at 416.502130. But cpufreq noticed the long idle period, and dropped the frequency at 416.505739, only to increase it back again at 416.515742, realizing that the workload is in-fact CPU bound. Thus ebizzy needlessly ran at the lowest frequency for almost 13 milliseconds (almost 1 full sample period), and this pattern repeats on every sleep-wakeup. This could hurt latency-sensitive workloads quite a lot. With patch: ~~~~~~~~~~~ kworker/8:2-29802 464.832535: cpu_frequency: state=2061000 cpu_id=8 <snip> --------------------------------------------------------------------- <snip> kworker/8:2-29802 464.962538: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy <...>-40738 464.972533: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2 kworker/8:2-29802 464.972536: cpu_frequency: state=4123000 cpu_id=8 kworker/8:2-29802 464.972538: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy <...>-40738 464.982531: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2 <snip> --------------------------------------------------------------------- <snip> kworker/8:2-29802 465.022533: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy <...>-40738 465.032531: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2 kworker/8:2-29802 465.032532: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy <...>-40738 465.035797: sched_switch: prev_comm=ebizzy ==> next_comm=swapper/8 <idle>-0 465.240178: sched_switch: prev_comm=swapper/8 ==> next_comm=ebizzy <...>-40738 465.242533: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2 kworker/8:2-29802 465.242535: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy <...>-40738 465.252531: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2 Observation: Ebizzy went idle at 465.035797, and started running again at 465.240178. Since ebizzy was the only real workload running on this CPU, cpufreq retained the frequency at 4.1Ghz throughout the run of ebizzy, no matter how many times ebizzy slept and woke-up in-between. Thus, ebizzy got the 10ms worth of 4.1 Ghz benefit during every sleep-wakeup (as compared to the run without the patch) and this boost gave a modest improvement in total throughput, as shown below. Sleeping-ebizzy records-per-second: ----------------------------------- Utilization Without patch With patch Difference (Absolute and % values) 10% 274767 277046 + 2279 (+0.829%) 20% 543429 553484 + 10055 (+1.850%) 40% 1090744 1107959 + 17215 (+1.578%) 60% 1634908 1662018 + 27110 (+1.658%) A rudimentary and somewhat approximately latency-sensitive workload such as sleeping-ebizzy itself showed a consistent, noticeable performance improvement with this patch. Hence, workloads that are truly latency-sensitive will benefit quite a bit from this change. Moreover, this is an overall win-win since this patch does not hurt power-savings at all (because, this patch does not reduce the idle time or idle residency; and the high frequency of the CPU when it goes to cpu-idle does not affect/hurt the power-savings of deep idle states). Signed-off-by: Srivatsa S. Bhat <srivatsa.bhat@linux.vnet.ibm.com> Reviewed-by: Gautham R. Shenoy <ego@linux.vnet.ibm.com> Acked-by: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2014-06-08 03:41:43 +07:00
/*
2014-06-09 15:51:24 +07:00
* Used to keep track of load in the previous interval. However, when
* explicitly set to zero, it is used as a flag to ensure that we copy
* the previous load to the current interval only once, upon the first
* wake-up from idle.
cpufreq: governor: Be friendly towards latency-sensitive bursty workloads Cpufreq governors like the ondemand governor calculate the load on the CPU periodically by employing deferrable timers. A deferrable timer won't fire if the CPU is completely idle (and there are no other timers to be run), in order to avoid unnecessary wakeups and thus save CPU power. However, the load calculation logic is agnostic to all this, and this can lead to the problem described below. Time (ms) CPU 1 100 Task-A running 110 Governor's timer fires, finds load as 100% in the last 10ms interval and increases the CPU frequency. 110.5 Task-A running 120 Governor's timer fires, finds load as 100% in the last 10ms interval and increases the CPU frequency. 125 Task-A went to sleep. With nothing else to do, CPU 1 went completely idle. 200 Task-A woke up and started running again. 200.5 Governor's deferred timer (which was originally programmed to fire at time 130) fires now. It calculates load for the time period 120 to 200.5, and finds the load is almost zero. Hence it decreases the CPU frequency to the minimum. 210 Governor's timer fires, finds load as 100% in the last 10ms interval and increases the CPU frequency. So, after the workload woke up and started running, the frequency was suddenly dropped to absolute minimum, and after that, there was an unnecessary delay of 10ms (sampling period) to increase the CPU frequency back to a reasonable value. And this pattern repeats for every wake-up-from-cpu-idle for that workload. This can be quite undesirable for latency- or response-time sensitive bursty workloads. So we need to fix the governor's logic to detect such wake-up-from- cpu-idle scenarios and start the workload at a reasonably high CPU frequency. One extreme solution would be to fake a load of 100% in such scenarios. But that might lead to undesirable side-effects such as frequency spikes (which might also need voltage changes) especially if the previous frequency happened to be very low. We just want to avoid the stupidity of dropping down the frequency to a minimum and then enduring a needless (and long) delay before ramping it up back again. So, let us simply carry forward the previous load - that is, let us just pretend that the 'load' for the current time-window is the same as the load for the previous window. That way, the frequency and voltage will continue to be set to whatever values they were set at previously. This means that bursty workloads will get a chance to influence the CPU frequency at which they wake up from cpu-idle, based on their past execution history. Thus, they might be able to avoid suffering from slow wakeups and long response-times. However, we should take care not to over-do this. For example, such a "copy previous load" logic will benefit cases like this: (where # represents busy and . represents idle) ##########.........#########.........###########...........##########........ but it will be detrimental in cases like the one shown below, because it will retain the high frequency (copied from the previous interval) even in a mostly idle system: ##########.........#.................#.....................#............... (i.e., the workload finished and the remaining tasks are such that their busy periods are smaller than the sampling interval, which causes the timer to always get deferred. So, this will make the copy-previous-load logic copy the initial high load to subsequent idle periods over and over again, thus keeping the frequency high unnecessarily). So, we modify this copy-previous-load logic such that it is used only once upon every wakeup-from-idle. Thus if we have 2 consecutive idle periods, the previous load won't get blindly copied over; cpufreq will freshly evaluate the load in the second idle interval, thus ensuring that the system comes back to its normal state. [ The right way to solve this whole problem is to teach the CPU frequency governors to also track load on a per-task basis, not just a per-CPU basis, and then use both the data sources intelligently to set the appropriate frequency on the CPUs. But that involves redesigning the cpufreq subsystem, so this patch should make the situation bearable until then. ] Experimental results: +-------------------+ I ran a modified version of ebizzy (called 'sleeping-ebizzy') that sleeps in between its execution such that its total utilization can be a user-defined value, say 10% or 20% (higher the utilization specified, lesser the amount of sleeps injected). This ebizzy was run with a single-thread, tied to CPU 8. Behavior observed with tracing (sample taken from 40% utilization runs): ------------------------------------------------------------------------ Without patch: ~~~~~~~~~~~~~~ kworker/8:2-12137 416.335742: cpu_frequency: state=2061000 cpu_id=8 kworker/8:2-12137 416.335744: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy <...>-40753 416.345741: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2 kworker/8:2-12137 416.345744: cpu_frequency: state=4123000 cpu_id=8 kworker/8:2-12137 416.345746: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy <...>-40753 416.355738: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2 <snip> --------------------------------------------------------------------- <snip> <...>-40753 416.402202: sched_switch: prev_comm=ebizzy ==> next_comm=swapper/8 <idle>-0 416.502130: sched_switch: prev_comm=swapper/8 ==> next_comm=ebizzy <...>-40753 416.505738: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2 kworker/8:2-12137 416.505739: cpu_frequency: state=2061000 cpu_id=8 kworker/8:2-12137 416.505741: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy <...>-40753 416.515739: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2 kworker/8:2-12137 416.515742: cpu_frequency: state=4123000 cpu_id=8 kworker/8:2-12137 416.515744: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy Observation: Ebizzy went idle at 416.402202, and started running again at 416.502130. But cpufreq noticed the long idle period, and dropped the frequency at 416.505739, only to increase it back again at 416.515742, realizing that the workload is in-fact CPU bound. Thus ebizzy needlessly ran at the lowest frequency for almost 13 milliseconds (almost 1 full sample period), and this pattern repeats on every sleep-wakeup. This could hurt latency-sensitive workloads quite a lot. With patch: ~~~~~~~~~~~ kworker/8:2-29802 464.832535: cpu_frequency: state=2061000 cpu_id=8 <snip> --------------------------------------------------------------------- <snip> kworker/8:2-29802 464.962538: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy <...>-40738 464.972533: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2 kworker/8:2-29802 464.972536: cpu_frequency: state=4123000 cpu_id=8 kworker/8:2-29802 464.972538: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy <...>-40738 464.982531: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2 <snip> --------------------------------------------------------------------- <snip> kworker/8:2-29802 465.022533: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy <...>-40738 465.032531: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2 kworker/8:2-29802 465.032532: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy <...>-40738 465.035797: sched_switch: prev_comm=ebizzy ==> next_comm=swapper/8 <idle>-0 465.240178: sched_switch: prev_comm=swapper/8 ==> next_comm=ebizzy <...>-40738 465.242533: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2 kworker/8:2-29802 465.242535: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy <...>-40738 465.252531: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2 Observation: Ebizzy went idle at 465.035797, and started running again at 465.240178. Since ebizzy was the only real workload running on this CPU, cpufreq retained the frequency at 4.1Ghz throughout the run of ebizzy, no matter how many times ebizzy slept and woke-up in-between. Thus, ebizzy got the 10ms worth of 4.1 Ghz benefit during every sleep-wakeup (as compared to the run without the patch) and this boost gave a modest improvement in total throughput, as shown below. Sleeping-ebizzy records-per-second: ----------------------------------- Utilization Without patch With patch Difference (Absolute and % values) 10% 274767 277046 + 2279 (+0.829%) 20% 543429 553484 + 10055 (+1.850%) 40% 1090744 1107959 + 17215 (+1.578%) 60% 1634908 1662018 + 27110 (+1.658%) A rudimentary and somewhat approximately latency-sensitive workload such as sleeping-ebizzy itself showed a consistent, noticeable performance improvement with this patch. Hence, workloads that are truly latency-sensitive will benefit quite a bit from this change. Moreover, this is an overall win-win since this patch does not hurt power-savings at all (because, this patch does not reduce the idle time or idle residency; and the high frequency of the CPU when it goes to cpu-idle does not affect/hurt the power-savings of deep idle states). Signed-off-by: Srivatsa S. Bhat <srivatsa.bhat@linux.vnet.ibm.com> Reviewed-by: Gautham R. Shenoy <ego@linux.vnet.ibm.com> Acked-by: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2014-06-08 03:41:43 +07:00
*/
2014-06-09 15:51:24 +07:00
unsigned int prev_load;
struct update_util_data update_util;
struct policy_dbs_info *policy_dbs;
};
/* Common Governor data across policies */
struct dbs_governor {
struct cpufreq_governor gov;
cpufreq: governor: New sysfs show/store callbacks for governor tunables The ondemand and conservative governors use the global-attr or freq-attr structures to represent sysfs attributes corresponding to their tunables (which of them is actually used depends on whether or not different policy objects can use the same governor with different tunables at the same time and, consequently, on where those attributes are located in sysfs). Unfortunately, in the freq-attr case, the standard cpufreq show/store sysfs attribute callbacks are applied to the governor tunable attributes and they always acquire the policy->rwsem lock before carrying out the operation. That may lead to an ABBA deadlock if governor tunable attributes are removed under policy->rwsem while one of them is being accessed concurrently (if sysfs attributes removal wins the race, it will wait for the access to complete with policy->rwsem held while the attribute callback will block on policy->rwsem indefinitely). We attempted to address this issue by dropping policy->rwsem around governor tunable attributes removal (that is, around invocations of the ->governor callback with the event arg equal to CPUFREQ_GOV_POLICY_EXIT) in cpufreq_set_policy(), but that opened up race conditions that had not been possible with policy->rwsem held all the time. Therefore policy->rwsem cannot be dropped in cpufreq_set_policy() at any point, but the deadlock situation described above must be avoided too. To that end, use the observation that in principle governor tunables may be represented by the same data type regardless of whether the governor is system-wide or per-policy and introduce a new structure, struct governor_attr, for representing them and new corresponding macros for creating show/store sysfs callbacks for them. Also make their parent kobject use a new kobject type whose default show/store callbacks are not related to the standard core cpufreq ones in any way (and they don't acquire policy->rwsem in particular). Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Tested-by: Juri Lelli <juri.lelli@arm.com> Tested-by: Shilpasri G Bhat <shilpa.bhat@linux.vnet.ibm.com> [ rjw: Subject & changelog + rebase ] Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2016-02-09 10:31:33 +07:00
struct kobj_type kobj_type;
/*
* Common data for platforms that don't set
* CPUFREQ_HAVE_GOVERNOR_PER_POLICY
*/
struct dbs_data *gdbs_data;
unsigned int (*gov_dbs_timer)(struct cpufreq_policy *policy);
struct policy_dbs_info *(*alloc)(void);
void (*free)(struct policy_dbs_info *policy_dbs);
cpufreq: Drop the 'initialized' field from struct cpufreq_governor The 'initialized' field in struct cpufreq_governor is only used by the conservative governor (as a usage counter) and the way that happens is far from straightforward and arguably incorrect. Namely, the value of 'initialized' is checked by cpufreq_dbs_governor_init() and cpufreq_dbs_governor_exit() and the results of those checks are passed (as the second argument) to the ->init() and ->exit() callbacks in struct dbs_governor. Those callbacks are only implemented by the ondemand and conservative governors and ondemand doesn't use their second argument at all. In turn, the conservative governor uses it to decide whether or not to either register or unregister a transition notifier. That whole mechanism is not only unnecessarily convoluted, but also racy, because the 'initialized' field of struct cpufreq_governor is updated in cpufreq_init_governor() and cpufreq_exit_governor() under policy->rwsem which doesn't help if one of these functions is run twice in parallel for different policies (which isn't impossible in principle), for example. Instead of it, add a proper usage counter to the conservative governor and update it from cs_init() and cs_exit() which is guaranteed to be non-racy, as those functions are only called under gov_dbs_data_mutex which is global. With that in place, drop the 'initialized' field from struct cpufreq_governor as it is not used any more. Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
2016-05-19 03:59:49 +07:00
int (*init)(struct dbs_data *dbs_data);
void (*exit)(struct dbs_data *dbs_data);
void (*start)(struct cpufreq_policy *policy);
};
static inline struct dbs_governor *dbs_governor_of(struct cpufreq_policy *policy)
{
return container_of(policy->governor, struct dbs_governor, gov);
}
/* Governor callback routines */
int cpufreq_dbs_governor_init(struct cpufreq_policy *policy);
void cpufreq_dbs_governor_exit(struct cpufreq_policy *policy);
int cpufreq_dbs_governor_start(struct cpufreq_policy *policy);
void cpufreq_dbs_governor_stop(struct cpufreq_policy *policy);
void cpufreq_dbs_governor_limits(struct cpufreq_policy *policy);
#define CPUFREQ_DBS_GOVERNOR_INITIALIZER(_name_) \
{ \
.name = _name_, \
.max_transition_latency = TRANSITION_LATENCY_LIMIT, \
.owner = THIS_MODULE, \
.init = cpufreq_dbs_governor_init, \
.exit = cpufreq_dbs_governor_exit, \
.start = cpufreq_dbs_governor_start, \
.stop = cpufreq_dbs_governor_stop, \
.limits = cpufreq_dbs_governor_limits, \
}
/* Governor specific operations */
struct od_ops {
unsigned int (*powersave_bias_target)(struct cpufreq_policy *policy,
unsigned int freq_next, unsigned int relation);
};
unsigned int dbs_update(struct cpufreq_policy *policy);
void od_register_powersave_bias_handler(unsigned int (*f)
(struct cpufreq_policy *, unsigned int, unsigned int),
unsigned int powersave_bias);
void od_unregister_powersave_bias_handler(void);
ssize_t store_sampling_rate(struct gov_attr_set *attr_set, const char *buf,
size_t count);
void gov_update_cpu_data(struct dbs_data *dbs_data);
#endif /* _CPUFREQ_GOVERNOR_H */