linux/drivers/cpufreq/cpufreq_governor.c

553 lines
15 KiB
C
Raw Normal View History

/*
* drivers/cpufreq/cpufreq_governor.c
*
* 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.
*/
#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
#include <linux/export.h>
#include <linux/kernel_stat.h>
#include <linux/slab.h>
#include "cpufreq_governor.h"
static struct attribute_group *get_sysfs_attr(struct dbs_data *dbs_data)
{
if (have_governor_per_policy())
return dbs_data->cdata->attr_group_gov_pol;
else
return dbs_data->cdata->attr_group_gov_sys;
}
void dbs_check_cpu(struct dbs_data *dbs_data, int cpu)
{
struct cpu_dbs_info *cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
struct od_dbs_tuners *od_tuners = dbs_data->tuners;
struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
struct cpufreq_policy *policy = cdbs->shared->policy;
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>
2014-06-07 20:41:43 +00:00
unsigned int sampling_rate;
unsigned int max_load = 0;
unsigned int ignore_nice;
unsigned int j;
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>
2014-06-07 20:41:43 +00:00
if (dbs_data->cdata->governor == GOV_ONDEMAND) {
struct od_cpu_dbs_info_s *od_dbs_info =
dbs_data->cdata->get_cpu_dbs_info_s(cpu);
/*
* Sometimes, the ondemand governor uses an additional
* multiplier to give long delays. So apply this multiplier to
* the 'sampling_rate', so as to keep the wake-up-from-idle
* detection logic a bit conservative.
*/
sampling_rate = od_tuners->sampling_rate;
sampling_rate *= od_dbs_info->rate_mult;
ignore_nice = od_tuners->ignore_nice_load;
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>
2014-06-07 20:41:43 +00:00
} else {
sampling_rate = cs_tuners->sampling_rate;
ignore_nice = cs_tuners->ignore_nice_load;
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>
2014-06-07 20:41:43 +00:00
}
cpufreq: ondemand: Change the calculation of target frequency The ondemand governor calculates load in terms of frequency and increases it only if load_freq is greater than up_threshold multiplied by the current or average frequency. This appears to produce oscillations of frequency between min and max because, for example, a relatively small load can easily saturate minimum frequency and lead the CPU to the max. Then, it will decrease back to the min due to small load_freq. Change the calculation method of load and target frequency on the basis of the following two observations: - Load computation should not depend on the current or average measured frequency. For example, absolute load of 80% at 100MHz is not necessarily equivalent to 8% at 1000MHz in the next sampling interval. - It should be possible to increase the target frequency to any value present in the frequency table proportional to the absolute load, rather than to the max only, so that: Target frequency = C * load where we take C = policy->cpuinfo.max_freq / 100. Tested on Intel i7-3770 CPU @ 3.40GHz and on Quad core 1500MHz Krait. Phoronix benchmark of Linux Kernel Compilation 3.1 test shows an increase ~1.5% in performance. cpufreq_stats (time_in_state) shows that middle frequencies are used more, with this patch. Highest and lowest frequencies were used less by ~9%. [rjw: We have run multiple other tests on kernels with this change applied and in the vast majority of cases it turns out that the resulting performance improvement also leads to reduced consumption of energy. The change is additionally justified by the overall simplification of the code in question.] Signed-off-by: Stratos Karafotis <stratosk@semaphore.gr> Acked-by: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2013-06-05 16:01:25 +00:00
/* Get Absolute Load */
for_each_cpu(j, policy->cpus) {
struct cpu_dbs_info *j_cdbs;
u64 cur_wall_time, cur_idle_time;
unsigned int idle_time, wall_time;
unsigned int load;
int io_busy = 0;
j_cdbs = dbs_data->cdata->get_cpu_cdbs(j);
/*
* For the purpose of ondemand, waiting for disk IO is
* an indication that you're performance critical, and
* not that the system is actually idle. So do not add
* the iowait time to the cpu idle time.
*/
if (dbs_data->cdata->governor == GOV_ONDEMAND)
io_busy = od_tuners->io_is_busy;
cur_idle_time = get_cpu_idle_time(j, &cur_wall_time, io_busy);
wall_time = (unsigned int)
(cur_wall_time - j_cdbs->prev_cpu_wall);
j_cdbs->prev_cpu_wall = cur_wall_time;
idle_time = (unsigned int)
(cur_idle_time - j_cdbs->prev_cpu_idle);
j_cdbs->prev_cpu_idle = cur_idle_time;
if (ignore_nice) {
u64 cur_nice;
unsigned long cur_nice_jiffies;
cur_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE] -
cdbs->prev_cpu_nice;
/*
* Assumption: nice time between sampling periods will
* be less than 2^32 jiffies for 32 bit sys
*/
cur_nice_jiffies = (unsigned long)
cputime64_to_jiffies64(cur_nice);
cdbs->prev_cpu_nice =
kcpustat_cpu(j).cpustat[CPUTIME_NICE];
idle_time += jiffies_to_usecs(cur_nice_jiffies);
}
if (unlikely(!wall_time || wall_time < idle_time))
continue;
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>
2014-06-07 20:41:43 +00:00
/*
* If the CPU had gone completely idle, and a task just woke up
* on this CPU now, it would be unfair to calculate 'load' the
* usual way for this elapsed time-window, because it will show
* near-zero load, irrespective of how CPU intensive that task
* actually is. This is undesirable for latency-sensitive bursty
* workloads.
*
* To avoid this, we reuse the 'load' from the previous
* time-window and give this task a chance to start with a
* reasonably high CPU frequency. (However, we shouldn't over-do
* this copy, lest we get stuck at a high load (high frequency)
* for too long, even when the current system load has actually
* dropped down. So we perform the copy only once, upon the
* first wake-up from idle.)
*
* Detecting this situation is easy: the governor's deferrable
* timer would not have fired during CPU-idle periods. Hence
* an unusually large 'wall_time' (as compared to the sampling
* rate) indicates this scenario.
2014-06-09 08:51:24 +00:00
*
* prev_load can be zero in two cases and we must recalculate it
* for both cases:
* - during long idle intervals
* - explicitly set to zero
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>
2014-06-07 20:41:43 +00:00
*/
2014-06-09 08:51:24 +00:00
if (unlikely(wall_time > (2 * sampling_rate) &&
j_cdbs->prev_load)) {
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>
2014-06-07 20:41:43 +00:00
load = j_cdbs->prev_load;
2014-06-09 08:51:24 +00:00
/*
* Perform a destructive copy, to ensure that we copy
* the previous load only once, upon the first wake-up
* from idle.
*/
j_cdbs->prev_load = 0;
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>
2014-06-07 20:41:43 +00:00
} else {
load = 100 * (wall_time - idle_time) / wall_time;
j_cdbs->prev_load = load;
}
if (load > max_load)
max_load = load;
}
dbs_data->cdata->gov_check_cpu(cpu, max_load);
}
EXPORT_SYMBOL_GPL(dbs_check_cpu);
static inline void __gov_queue_work(int cpu, struct dbs_data *dbs_data,
unsigned int delay)
{
struct cpu_dbs_info *cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
mod_delayed_work_on(cpu, system_wq, &cdbs->dwork, delay);
}
void gov_queue_work(struct dbs_data *dbs_data, struct cpufreq_policy *policy,
unsigned int delay, bool all_cpus)
{
int i;
cpufreq: Fix timer/workqueue corruption by protecting reading governor_enabled When a CPU is hot removed we'll cancel all the delayed work items via gov_cancel_work(). Sometimes the delayed work function determines that it should adjust the delay for all other CPUs that the policy is managing. If this scenario occurs, the canceling CPU will cancel its own work but queue up the other CPUs works to run. Commit 3617f2 (cpufreq: Fix timer/workqueue corruption due to double queueing) has tried to fix this, but reading governor_enabled is not protected by cpufreq_governor_lock. Even though od_dbs_timer() checks governor_enabled before gov_queue_work(), this scenario may occur. For example: CPU0 CPU1 ---- ---- cpu_down() ... <work runs> __cpufreq_remove_dev() od_dbs_timer() __cpufreq_governor() policy->governor_enabled policy->governor_enabled = false; cpufreq_governor_dbs() case CPUFREQ_GOV_STOP: gov_cancel_work(dbs_data, policy); cpu0 work is canceled timer is canceled cpu1 work is canceled <waits for cpu1> gov_queue_work(*, *, true); cpu0 work queued cpu1 work queued cpu2 work queued ... cpu1 work is canceled cpu2 work is canceled ... At the end of the GOV_STOP case cpu0 still has a work queued to run although the code is expecting all of the works to be canceled. __cpufreq_remove_dev() will then proceed to re-initialize all the other CPUs works except for the CPU that is going down. The CPUFREQ_GOV_START case in cpufreq_governor_dbs() will trample over the queued work and debugobjects will spit out a warning: WARNING: at lib/debugobjects.c:260 debug_print_object+0x94/0xbc() ODEBUG: init active (active state 0) object type: timer_list hint: delayed_work_timer_fn+0x0/0x14 Modules linked in: CPU: 1 PID: 1205 Comm: sh Tainted: G W 3.10.0 #200 [<c01144f0>] (unwind_backtrace+0x0/0xf8) from [<c0111d98>] (show_stack+0x10/0x14) [<c0111d98>] (show_stack+0x10/0x14) from [<c01272cc>] (warn_slowpath_common+0x4c/0x68) [<c01272cc>] (warn_slowpath_common+0x4c/0x68) from [<c012737c>] (warn_slowpath_fmt+0x30/0x40) [<c012737c>] (warn_slowpath_fmt+0x30/0x40) from [<c034c640>] (debug_print_object+0x94/0xbc) [<c034c640>] (debug_print_object+0x94/0xbc) from [<c034c7f8>] (__debug_object_init+0xc8/0x3c0) [<c034c7f8>] (__debug_object_init+0xc8/0x3c0) from [<c01360e0>] (init_timer_key+0x20/0x104) [<c01360e0>] (init_timer_key+0x20/0x104) from [<c04872ac>] (cpufreq_governor_dbs+0x1dc/0x68c) [<c04872ac>] (cpufreq_governor_dbs+0x1dc/0x68c) from [<c04833a8>] (__cpufreq_governor+0x80/0x1b0) [<c04833a8>] (__cpufreq_governor+0x80/0x1b0) from [<c0483704>] (__cpufreq_remove_dev.isra.12+0x22c/0x380) [<c0483704>] (__cpufreq_remove_dev.isra.12+0x22c/0x380) from [<c0692f38>] (cpufreq_cpu_callback+0x48/0x5c) [<c0692f38>] (cpufreq_cpu_callback+0x48/0x5c) from [<c014fb40>] (notifier_call_chain+0x44/0x84) [<c014fb40>] (notifier_call_chain+0x44/0x84) from [<c012ae44>] (__cpu_notify+0x2c/0x48) [<c012ae44>] (__cpu_notify+0x2c/0x48) from [<c068dd40>] (_cpu_down+0x80/0x258) [<c068dd40>] (_cpu_down+0x80/0x258) from [<c068df40>] (cpu_down+0x28/0x3c) [<c068df40>] (cpu_down+0x28/0x3c) from [<c068e4c0>] (store_online+0x30/0x74) [<c068e4c0>] (store_online+0x30/0x74) from [<c03a7308>] (dev_attr_store+0x18/0x24) [<c03a7308>] (dev_attr_store+0x18/0x24) from [<c0256fe0>] (sysfs_write_file+0x100/0x180) [<c0256fe0>] (sysfs_write_file+0x100/0x180) from [<c01fec9c>] (vfs_write+0xbc/0x184) [<c01fec9c>] (vfs_write+0xbc/0x184) from [<c01ff034>] (SyS_write+0x40/0x68) [<c01ff034>] (SyS_write+0x40/0x68) from [<c010e200>] (ret_fast_syscall+0x0/0x48) In gov_queue_work(), lock cpufreq_governor_lock before gov_queue_work, and unlock it after __gov_queue_work(). In this way, governor_enabled is guaranteed not changed in gov_queue_work(). Signed-off-by: Jane Li <jiel@marvell.com> Acked-by: Viresh Kumar <viresh.kumar@linaro.org> Reviewed-by: Dmitry Torokhov <dmitry.torokhov@gmail.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2014-01-03 09:17:41 +00:00
mutex_lock(&cpufreq_governor_lock);
cpufreq: Fix timer/workqueue corruption due to double queueing When a CPU is hot removed we'll cancel all the delayed work items via gov_cancel_work(). Normally this will just cancels a delayed timer on each CPU that the policy is managing and the work won't run, but if the work is already running the workqueue code will wait for the work to finish before continuing to prevent the work items from re-queuing themselves like they normally do. This scheme will work most of the time, except for the case where the work function determines that it should adjust the delay for all other CPUs that the policy is managing. If this scenario occurs, the canceling CPU will cancel its own work but queue up the other CPUs works to run. For example: CPU0 CPU1 ---- ---- cpu_down() ... __cpufreq_remove_dev() cpufreq_governor_dbs() case CPUFREQ_GOV_STOP: gov_cancel_work(dbs_data, policy); cpu0 work is canceled timer is canceled cpu1 work is canceled <work runs> <waits for cpu1> od_dbs_timer() gov_queue_work(*, *, true); cpu0 work queued cpu1 work queued cpu2 work queued ... cpu1 work is canceled cpu2 work is canceled ... At the end of the GOV_STOP case cpu0 still has a work queued to run although the code is expecting all of the works to be canceled. __cpufreq_remove_dev() will then proceed to re-initialize all the other CPUs works except for the CPU that is going down. The CPUFREQ_GOV_START case in cpufreq_governor_dbs() will trample over the queued work and debugobjects will spit out a warning: WARNING: at lib/debugobjects.c:260 debug_print_object+0x94/0xbc() ODEBUG: init active (active state 0) object type: timer_list hint: delayed_work_timer_fn+0x0/0x10 Modules linked in: CPU: 0 PID: 1491 Comm: sh Tainted: G W 3.10.0 #19 [<c010c178>] (unwind_backtrace+0x0/0x11c) from [<c0109dec>] (show_stack+0x10/0x14) [<c0109dec>] (show_stack+0x10/0x14) from [<c01904cc>] (warn_slowpath_common+0x4c/0x6c) [<c01904cc>] (warn_slowpath_common+0x4c/0x6c) from [<c019056c>] (warn_slowpath_fmt+0x2c/0x3c) [<c019056c>] (warn_slowpath_fmt+0x2c/0x3c) from [<c0388a7c>] (debug_print_object+0x94/0xbc) [<c0388a7c>] (debug_print_object+0x94/0xbc) from [<c0388e34>] (__debug_object_init+0x2d0/0x340) [<c0388e34>] (__debug_object_init+0x2d0/0x340) from [<c019e3b0>] (init_timer_key+0x14/0xb0) [<c019e3b0>] (init_timer_key+0x14/0xb0) from [<c0635f78>] (cpufreq_governor_dbs+0x3e8/0x5f8) [<c0635f78>] (cpufreq_governor_dbs+0x3e8/0x5f8) from [<c06325a0>] (__cpufreq_governor+0xdc/0x1a4) [<c06325a0>] (__cpufreq_governor+0xdc/0x1a4) from [<c0633704>] (__cpufreq_remove_dev.isra.10+0x3b4/0x434) [<c0633704>] (__cpufreq_remove_dev.isra.10+0x3b4/0x434) from [<c08989f4>] (cpufreq_cpu_callback+0x60/0x80) [<c08989f4>] (cpufreq_cpu_callback+0x60/0x80) from [<c08a43c0>] (notifier_call_chain+0x38/0x68) [<c08a43c0>] (notifier_call_chain+0x38/0x68) from [<c01938e0>] (__cpu_notify+0x28/0x40) [<c01938e0>] (__cpu_notify+0x28/0x40) from [<c0892ad4>] (_cpu_down+0x7c/0x2c0) [<c0892ad4>] (_cpu_down+0x7c/0x2c0) from [<c0892d3c>] (cpu_down+0x24/0x40) [<c0892d3c>] (cpu_down+0x24/0x40) from [<c0893ea8>] (store_online+0x2c/0x74) [<c0893ea8>] (store_online+0x2c/0x74) from [<c04519d8>] (dev_attr_store+0x18/0x24) [<c04519d8>] (dev_attr_store+0x18/0x24) from [<c02a69d4>] (sysfs_write_file+0x100/0x148) [<c02a69d4>] (sysfs_write_file+0x100/0x148) from [<c0255c18>] (vfs_write+0xcc/0x174) [<c0255c18>] (vfs_write+0xcc/0x174) from [<c0255f70>] (SyS_write+0x38/0x64) [<c0255f70>] (SyS_write+0x38/0x64) from [<c0106120>] (ret_fast_syscall+0x0/0x30) Signed-off-by: Stephen Boyd <sboyd@codeaurora.org> Acked-by: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2013-08-27 18:47:29 +00:00
if (!policy->governor_enabled)
cpufreq: Fix timer/workqueue corruption by protecting reading governor_enabled When a CPU is hot removed we'll cancel all the delayed work items via gov_cancel_work(). Sometimes the delayed work function determines that it should adjust the delay for all other CPUs that the policy is managing. If this scenario occurs, the canceling CPU will cancel its own work but queue up the other CPUs works to run. Commit 3617f2 (cpufreq: Fix timer/workqueue corruption due to double queueing) has tried to fix this, but reading governor_enabled is not protected by cpufreq_governor_lock. Even though od_dbs_timer() checks governor_enabled before gov_queue_work(), this scenario may occur. For example: CPU0 CPU1 ---- ---- cpu_down() ... <work runs> __cpufreq_remove_dev() od_dbs_timer() __cpufreq_governor() policy->governor_enabled policy->governor_enabled = false; cpufreq_governor_dbs() case CPUFREQ_GOV_STOP: gov_cancel_work(dbs_data, policy); cpu0 work is canceled timer is canceled cpu1 work is canceled <waits for cpu1> gov_queue_work(*, *, true); cpu0 work queued cpu1 work queued cpu2 work queued ... cpu1 work is canceled cpu2 work is canceled ... At the end of the GOV_STOP case cpu0 still has a work queued to run although the code is expecting all of the works to be canceled. __cpufreq_remove_dev() will then proceed to re-initialize all the other CPUs works except for the CPU that is going down. The CPUFREQ_GOV_START case in cpufreq_governor_dbs() will trample over the queued work and debugobjects will spit out a warning: WARNING: at lib/debugobjects.c:260 debug_print_object+0x94/0xbc() ODEBUG: init active (active state 0) object type: timer_list hint: delayed_work_timer_fn+0x0/0x14 Modules linked in: CPU: 1 PID: 1205 Comm: sh Tainted: G W 3.10.0 #200 [<c01144f0>] (unwind_backtrace+0x0/0xf8) from [<c0111d98>] (show_stack+0x10/0x14) [<c0111d98>] (show_stack+0x10/0x14) from [<c01272cc>] (warn_slowpath_common+0x4c/0x68) [<c01272cc>] (warn_slowpath_common+0x4c/0x68) from [<c012737c>] (warn_slowpath_fmt+0x30/0x40) [<c012737c>] (warn_slowpath_fmt+0x30/0x40) from [<c034c640>] (debug_print_object+0x94/0xbc) [<c034c640>] (debug_print_object+0x94/0xbc) from [<c034c7f8>] (__debug_object_init+0xc8/0x3c0) [<c034c7f8>] (__debug_object_init+0xc8/0x3c0) from [<c01360e0>] (init_timer_key+0x20/0x104) [<c01360e0>] (init_timer_key+0x20/0x104) from [<c04872ac>] (cpufreq_governor_dbs+0x1dc/0x68c) [<c04872ac>] (cpufreq_governor_dbs+0x1dc/0x68c) from [<c04833a8>] (__cpufreq_governor+0x80/0x1b0) [<c04833a8>] (__cpufreq_governor+0x80/0x1b0) from [<c0483704>] (__cpufreq_remove_dev.isra.12+0x22c/0x380) [<c0483704>] (__cpufreq_remove_dev.isra.12+0x22c/0x380) from [<c0692f38>] (cpufreq_cpu_callback+0x48/0x5c) [<c0692f38>] (cpufreq_cpu_callback+0x48/0x5c) from [<c014fb40>] (notifier_call_chain+0x44/0x84) [<c014fb40>] (notifier_call_chain+0x44/0x84) from [<c012ae44>] (__cpu_notify+0x2c/0x48) [<c012ae44>] (__cpu_notify+0x2c/0x48) from [<c068dd40>] (_cpu_down+0x80/0x258) [<c068dd40>] (_cpu_down+0x80/0x258) from [<c068df40>] (cpu_down+0x28/0x3c) [<c068df40>] (cpu_down+0x28/0x3c) from [<c068e4c0>] (store_online+0x30/0x74) [<c068e4c0>] (store_online+0x30/0x74) from [<c03a7308>] (dev_attr_store+0x18/0x24) [<c03a7308>] (dev_attr_store+0x18/0x24) from [<c0256fe0>] (sysfs_write_file+0x100/0x180) [<c0256fe0>] (sysfs_write_file+0x100/0x180) from [<c01fec9c>] (vfs_write+0xbc/0x184) [<c01fec9c>] (vfs_write+0xbc/0x184) from [<c01ff034>] (SyS_write+0x40/0x68) [<c01ff034>] (SyS_write+0x40/0x68) from [<c010e200>] (ret_fast_syscall+0x0/0x48) In gov_queue_work(), lock cpufreq_governor_lock before gov_queue_work, and unlock it after __gov_queue_work(). In this way, governor_enabled is guaranteed not changed in gov_queue_work(). Signed-off-by: Jane Li <jiel@marvell.com> Acked-by: Viresh Kumar <viresh.kumar@linaro.org> Reviewed-by: Dmitry Torokhov <dmitry.torokhov@gmail.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2014-01-03 09:17:41 +00:00
goto out_unlock;
cpufreq: Fix timer/workqueue corruption due to double queueing When a CPU is hot removed we'll cancel all the delayed work items via gov_cancel_work(). Normally this will just cancels a delayed timer on each CPU that the policy is managing and the work won't run, but if the work is already running the workqueue code will wait for the work to finish before continuing to prevent the work items from re-queuing themselves like they normally do. This scheme will work most of the time, except for the case where the work function determines that it should adjust the delay for all other CPUs that the policy is managing. If this scenario occurs, the canceling CPU will cancel its own work but queue up the other CPUs works to run. For example: CPU0 CPU1 ---- ---- cpu_down() ... __cpufreq_remove_dev() cpufreq_governor_dbs() case CPUFREQ_GOV_STOP: gov_cancel_work(dbs_data, policy); cpu0 work is canceled timer is canceled cpu1 work is canceled <work runs> <waits for cpu1> od_dbs_timer() gov_queue_work(*, *, true); cpu0 work queued cpu1 work queued cpu2 work queued ... cpu1 work is canceled cpu2 work is canceled ... At the end of the GOV_STOP case cpu0 still has a work queued to run although the code is expecting all of the works to be canceled. __cpufreq_remove_dev() will then proceed to re-initialize all the other CPUs works except for the CPU that is going down. The CPUFREQ_GOV_START case in cpufreq_governor_dbs() will trample over the queued work and debugobjects will spit out a warning: WARNING: at lib/debugobjects.c:260 debug_print_object+0x94/0xbc() ODEBUG: init active (active state 0) object type: timer_list hint: delayed_work_timer_fn+0x0/0x10 Modules linked in: CPU: 0 PID: 1491 Comm: sh Tainted: G W 3.10.0 #19 [<c010c178>] (unwind_backtrace+0x0/0x11c) from [<c0109dec>] (show_stack+0x10/0x14) [<c0109dec>] (show_stack+0x10/0x14) from [<c01904cc>] (warn_slowpath_common+0x4c/0x6c) [<c01904cc>] (warn_slowpath_common+0x4c/0x6c) from [<c019056c>] (warn_slowpath_fmt+0x2c/0x3c) [<c019056c>] (warn_slowpath_fmt+0x2c/0x3c) from [<c0388a7c>] (debug_print_object+0x94/0xbc) [<c0388a7c>] (debug_print_object+0x94/0xbc) from [<c0388e34>] (__debug_object_init+0x2d0/0x340) [<c0388e34>] (__debug_object_init+0x2d0/0x340) from [<c019e3b0>] (init_timer_key+0x14/0xb0) [<c019e3b0>] (init_timer_key+0x14/0xb0) from [<c0635f78>] (cpufreq_governor_dbs+0x3e8/0x5f8) [<c0635f78>] (cpufreq_governor_dbs+0x3e8/0x5f8) from [<c06325a0>] (__cpufreq_governor+0xdc/0x1a4) [<c06325a0>] (__cpufreq_governor+0xdc/0x1a4) from [<c0633704>] (__cpufreq_remove_dev.isra.10+0x3b4/0x434) [<c0633704>] (__cpufreq_remove_dev.isra.10+0x3b4/0x434) from [<c08989f4>] (cpufreq_cpu_callback+0x60/0x80) [<c08989f4>] (cpufreq_cpu_callback+0x60/0x80) from [<c08a43c0>] (notifier_call_chain+0x38/0x68) [<c08a43c0>] (notifier_call_chain+0x38/0x68) from [<c01938e0>] (__cpu_notify+0x28/0x40) [<c01938e0>] (__cpu_notify+0x28/0x40) from [<c0892ad4>] (_cpu_down+0x7c/0x2c0) [<c0892ad4>] (_cpu_down+0x7c/0x2c0) from [<c0892d3c>] (cpu_down+0x24/0x40) [<c0892d3c>] (cpu_down+0x24/0x40) from [<c0893ea8>] (store_online+0x2c/0x74) [<c0893ea8>] (store_online+0x2c/0x74) from [<c04519d8>] (dev_attr_store+0x18/0x24) [<c04519d8>] (dev_attr_store+0x18/0x24) from [<c02a69d4>] (sysfs_write_file+0x100/0x148) [<c02a69d4>] (sysfs_write_file+0x100/0x148) from [<c0255c18>] (vfs_write+0xcc/0x174) [<c0255c18>] (vfs_write+0xcc/0x174) from [<c0255f70>] (SyS_write+0x38/0x64) [<c0255f70>] (SyS_write+0x38/0x64) from [<c0106120>] (ret_fast_syscall+0x0/0x30) Signed-off-by: Stephen Boyd <sboyd@codeaurora.org> Acked-by: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2013-08-27 18:47:29 +00:00
if (!all_cpus) {
/*
* Use raw_smp_processor_id() to avoid preemptible warnings.
* We know that this is only called with all_cpus == false from
* works that have been queued with *_work_on() functions and
* those works are canceled during CPU_DOWN_PREPARE so they
* can't possibly run on any other CPU.
*/
__gov_queue_work(raw_smp_processor_id(), dbs_data, delay);
} else {
for_each_cpu(i, policy->cpus)
__gov_queue_work(i, dbs_data, delay);
}
cpufreq: Fix timer/workqueue corruption by protecting reading governor_enabled When a CPU is hot removed we'll cancel all the delayed work items via gov_cancel_work(). Sometimes the delayed work function determines that it should adjust the delay for all other CPUs that the policy is managing. If this scenario occurs, the canceling CPU will cancel its own work but queue up the other CPUs works to run. Commit 3617f2 (cpufreq: Fix timer/workqueue corruption due to double queueing) has tried to fix this, but reading governor_enabled is not protected by cpufreq_governor_lock. Even though od_dbs_timer() checks governor_enabled before gov_queue_work(), this scenario may occur. For example: CPU0 CPU1 ---- ---- cpu_down() ... <work runs> __cpufreq_remove_dev() od_dbs_timer() __cpufreq_governor() policy->governor_enabled policy->governor_enabled = false; cpufreq_governor_dbs() case CPUFREQ_GOV_STOP: gov_cancel_work(dbs_data, policy); cpu0 work is canceled timer is canceled cpu1 work is canceled <waits for cpu1> gov_queue_work(*, *, true); cpu0 work queued cpu1 work queued cpu2 work queued ... cpu1 work is canceled cpu2 work is canceled ... At the end of the GOV_STOP case cpu0 still has a work queued to run although the code is expecting all of the works to be canceled. __cpufreq_remove_dev() will then proceed to re-initialize all the other CPUs works except for the CPU that is going down. The CPUFREQ_GOV_START case in cpufreq_governor_dbs() will trample over the queued work and debugobjects will spit out a warning: WARNING: at lib/debugobjects.c:260 debug_print_object+0x94/0xbc() ODEBUG: init active (active state 0) object type: timer_list hint: delayed_work_timer_fn+0x0/0x14 Modules linked in: CPU: 1 PID: 1205 Comm: sh Tainted: G W 3.10.0 #200 [<c01144f0>] (unwind_backtrace+0x0/0xf8) from [<c0111d98>] (show_stack+0x10/0x14) [<c0111d98>] (show_stack+0x10/0x14) from [<c01272cc>] (warn_slowpath_common+0x4c/0x68) [<c01272cc>] (warn_slowpath_common+0x4c/0x68) from [<c012737c>] (warn_slowpath_fmt+0x30/0x40) [<c012737c>] (warn_slowpath_fmt+0x30/0x40) from [<c034c640>] (debug_print_object+0x94/0xbc) [<c034c640>] (debug_print_object+0x94/0xbc) from [<c034c7f8>] (__debug_object_init+0xc8/0x3c0) [<c034c7f8>] (__debug_object_init+0xc8/0x3c0) from [<c01360e0>] (init_timer_key+0x20/0x104) [<c01360e0>] (init_timer_key+0x20/0x104) from [<c04872ac>] (cpufreq_governor_dbs+0x1dc/0x68c) [<c04872ac>] (cpufreq_governor_dbs+0x1dc/0x68c) from [<c04833a8>] (__cpufreq_governor+0x80/0x1b0) [<c04833a8>] (__cpufreq_governor+0x80/0x1b0) from [<c0483704>] (__cpufreq_remove_dev.isra.12+0x22c/0x380) [<c0483704>] (__cpufreq_remove_dev.isra.12+0x22c/0x380) from [<c0692f38>] (cpufreq_cpu_callback+0x48/0x5c) [<c0692f38>] (cpufreq_cpu_callback+0x48/0x5c) from [<c014fb40>] (notifier_call_chain+0x44/0x84) [<c014fb40>] (notifier_call_chain+0x44/0x84) from [<c012ae44>] (__cpu_notify+0x2c/0x48) [<c012ae44>] (__cpu_notify+0x2c/0x48) from [<c068dd40>] (_cpu_down+0x80/0x258) [<c068dd40>] (_cpu_down+0x80/0x258) from [<c068df40>] (cpu_down+0x28/0x3c) [<c068df40>] (cpu_down+0x28/0x3c) from [<c068e4c0>] (store_online+0x30/0x74) [<c068e4c0>] (store_online+0x30/0x74) from [<c03a7308>] (dev_attr_store+0x18/0x24) [<c03a7308>] (dev_attr_store+0x18/0x24) from [<c0256fe0>] (sysfs_write_file+0x100/0x180) [<c0256fe0>] (sysfs_write_file+0x100/0x180) from [<c01fec9c>] (vfs_write+0xbc/0x184) [<c01fec9c>] (vfs_write+0xbc/0x184) from [<c01ff034>] (SyS_write+0x40/0x68) [<c01ff034>] (SyS_write+0x40/0x68) from [<c010e200>] (ret_fast_syscall+0x0/0x48) In gov_queue_work(), lock cpufreq_governor_lock before gov_queue_work, and unlock it after __gov_queue_work(). In this way, governor_enabled is guaranteed not changed in gov_queue_work(). Signed-off-by: Jane Li <jiel@marvell.com> Acked-by: Viresh Kumar <viresh.kumar@linaro.org> Reviewed-by: Dmitry Torokhov <dmitry.torokhov@gmail.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2014-01-03 09:17:41 +00:00
out_unlock:
mutex_unlock(&cpufreq_governor_lock);
}
EXPORT_SYMBOL_GPL(gov_queue_work);
static inline void gov_cancel_work(struct dbs_data *dbs_data,
struct cpufreq_policy *policy)
{
struct cpu_dbs_info *cdbs;
int i;
for_each_cpu(i, policy->cpus) {
cdbs = dbs_data->cdata->get_cpu_cdbs(i);
cancel_delayed_work_sync(&cdbs->dwork);
}
}
/* Will return if we need to evaluate cpu load again or not */
static bool need_load_eval(struct cpu_common_dbs_info *shared,
unsigned int sampling_rate)
{
if (policy_is_shared(shared->policy)) {
ktime_t time_now = ktime_get();
s64 delta_us = ktime_us_delta(time_now, shared->time_stamp);
/* Do nothing if we recently have sampled */
if (delta_us < (s64)(sampling_rate / 2))
return false;
else
shared->time_stamp = time_now;
}
return true;
}
static void dbs_timer(struct work_struct *work)
{
struct cpu_dbs_info *cdbs = container_of(work, struct cpu_dbs_info,
dwork.work);
struct cpu_common_dbs_info *shared = cdbs->shared;
struct cpufreq_policy *policy = shared->policy;
struct dbs_data *dbs_data = policy->governor_data;
unsigned int sampling_rate, delay;
bool modify_all = true;
mutex_lock(&shared->timer_mutex);
if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
sampling_rate = cs_tuners->sampling_rate;
} else {
struct od_dbs_tuners *od_tuners = dbs_data->tuners;
sampling_rate = od_tuners->sampling_rate;
}
if (!need_load_eval(cdbs->shared, sampling_rate))
modify_all = false;
delay = dbs_data->cdata->gov_dbs_timer(cdbs, dbs_data, modify_all);
gov_queue_work(dbs_data, policy, delay, modify_all);
mutex_unlock(&shared->timer_mutex);
}
static void set_sampling_rate(struct dbs_data *dbs_data,
unsigned int sampling_rate)
{
if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
cs_tuners->sampling_rate = sampling_rate;
} else {
struct od_dbs_tuners *od_tuners = dbs_data->tuners;
od_tuners->sampling_rate = sampling_rate;
}
}
static int alloc_common_dbs_info(struct cpufreq_policy *policy,
struct common_dbs_data *cdata)
{
struct cpu_common_dbs_info *shared;
int j;
/* Allocate memory for the common information for policy->cpus */
shared = kzalloc(sizeof(*shared), GFP_KERNEL);
if (!shared)
return -ENOMEM;
/* Set shared for all CPUs, online+offline */
for_each_cpu(j, policy->related_cpus)
cdata->get_cpu_cdbs(j)->shared = shared;
return 0;
}
static void free_common_dbs_info(struct cpufreq_policy *policy,
struct common_dbs_data *cdata)
{
struct cpu_dbs_info *cdbs = cdata->get_cpu_cdbs(policy->cpu);
struct cpu_common_dbs_info *shared = cdbs->shared;
int j;
for_each_cpu(j, policy->cpus)
cdata->get_cpu_cdbs(j)->shared = NULL;
kfree(shared);
}
static int cpufreq_governor_init(struct cpufreq_policy *policy,
struct dbs_data *dbs_data,
struct common_dbs_data *cdata)
{
unsigned int latency;
int ret;
if (dbs_data) {
if (WARN_ON(have_governor_per_policy()))
return -EINVAL;
ret = alloc_common_dbs_info(policy, cdata);
if (ret)
return ret;
dbs_data->usage_count++;
policy->governor_data = dbs_data;
return 0;
}
dbs_data = kzalloc(sizeof(*dbs_data), GFP_KERNEL);
if (!dbs_data)
return -ENOMEM;
ret = alloc_common_dbs_info(policy, cdata);
if (ret)
goto free_dbs_data;
dbs_data->cdata = cdata;
dbs_data->usage_count = 1;
ret = cdata->init(dbs_data, !policy->governor->initialized);
if (ret)
goto free_common_dbs_info;
/* policy latency is in ns. Convert it to us first */
latency = policy->cpuinfo.transition_latency / 1000;
if (latency == 0)
latency = 1;
/* Bring kernel and HW constraints together */
dbs_data->min_sampling_rate = max(dbs_data->min_sampling_rate,
MIN_LATENCY_MULTIPLIER * latency);
set_sampling_rate(dbs_data, max(dbs_data->min_sampling_rate,
latency * LATENCY_MULTIPLIER));
if (!have_governor_per_policy()) {
if (WARN_ON(cpufreq_get_global_kobject())) {
ret = -EINVAL;
goto cdata_exit;
}
cdata->gdbs_data = dbs_data;
}
ret = sysfs_create_group(get_governor_parent_kobj(policy),
get_sysfs_attr(dbs_data));
if (ret)
goto put_kobj;
policy->governor_data = dbs_data;
return 0;
put_kobj:
if (!have_governor_per_policy()) {
cdata->gdbs_data = NULL;
cpufreq_put_global_kobject();
}
cdata_exit:
cdata->exit(dbs_data, !policy->governor->initialized);
free_common_dbs_info:
free_common_dbs_info(policy, cdata);
free_dbs_data:
kfree(dbs_data);
return ret;
}
static void cpufreq_governor_exit(struct cpufreq_policy *policy,
struct dbs_data *dbs_data)
{
struct common_dbs_data *cdata = dbs_data->cdata;
policy->governor_data = NULL;
if (!--dbs_data->usage_count) {
sysfs_remove_group(get_governor_parent_kobj(policy),
get_sysfs_attr(dbs_data));
if (!have_governor_per_policy()) {
cdata->gdbs_data = NULL;
cpufreq_put_global_kobject();
}
cdata->exit(dbs_data, policy->governor->initialized == 1);
kfree(dbs_data);
}
free_common_dbs_info(policy, cdata);
}
static int cpufreq_governor_start(struct cpufreq_policy *policy,
struct dbs_data *dbs_data)
{
struct common_dbs_data *cdata = dbs_data->cdata;
unsigned int sampling_rate, ignore_nice, j, cpu = policy->cpu;
struct cpu_dbs_info *cdbs = cdata->get_cpu_cdbs(cpu);
struct cpu_common_dbs_info *shared = cdbs->shared;
int io_busy = 0;
if (!policy->cur)
return -EINVAL;
if (cdata->governor == GOV_CONSERVATIVE) {
struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
sampling_rate = cs_tuners->sampling_rate;
ignore_nice = cs_tuners->ignore_nice_load;
} else {
struct od_dbs_tuners *od_tuners = dbs_data->tuners;
sampling_rate = od_tuners->sampling_rate;
ignore_nice = od_tuners->ignore_nice_load;
io_busy = od_tuners->io_is_busy;
}
shared->policy = policy;
shared->time_stamp = ktime_get();
mutex_init(&shared->timer_mutex);
for_each_cpu(j, policy->cpus) {
struct cpu_dbs_info *j_cdbs = cdata->get_cpu_cdbs(j);
unsigned int prev_load;
j_cdbs->prev_cpu_idle =
get_cpu_idle_time(j, &j_cdbs->prev_cpu_wall, io_busy);
prev_load = (unsigned int)(j_cdbs->prev_cpu_wall -
j_cdbs->prev_cpu_idle);
j_cdbs->prev_load = 100 * prev_load /
(unsigned int)j_cdbs->prev_cpu_wall;
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>
2014-06-07 20:41:43 +00:00
if (ignore_nice)
j_cdbs->prev_cpu_nice = kcpustat_cpu(j).cpustat[CPUTIME_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>
2014-06-07 20:41:43 +00:00
INIT_DEFERRABLE_WORK(&j_cdbs->dwork, dbs_timer);
}
if (cdata->governor == GOV_CONSERVATIVE) {
struct cs_cpu_dbs_info_s *cs_dbs_info =
cdata->get_cpu_dbs_info_s(cpu);
cs_dbs_info->down_skip = 0;
cs_dbs_info->enable = 1;
cs_dbs_info->requested_freq = policy->cur;
} else {
struct od_ops *od_ops = cdata->gov_ops;
struct od_cpu_dbs_info_s *od_dbs_info = cdata->get_cpu_dbs_info_s(cpu);
od_dbs_info->rate_mult = 1;
od_dbs_info->sample_type = OD_NORMAL_SAMPLE;
od_ops->powersave_bias_init_cpu(cpu);
}
gov_queue_work(dbs_data, policy, delay_for_sampling_rate(sampling_rate),
true);
return 0;
}
static void cpufreq_governor_stop(struct cpufreq_policy *policy,
struct dbs_data *dbs_data)
{
struct common_dbs_data *cdata = dbs_data->cdata;
unsigned int cpu = policy->cpu;
struct cpu_dbs_info *cdbs = cdata->get_cpu_cdbs(cpu);
struct cpu_common_dbs_info *shared = cdbs->shared;
gov_cancel_work(dbs_data, policy);
if (cdata->governor == GOV_CONSERVATIVE) {
struct cs_cpu_dbs_info_s *cs_dbs_info =
cdata->get_cpu_dbs_info_s(cpu);
cs_dbs_info->enable = 0;
}
shared->policy = NULL;
mutex_destroy(&shared->timer_mutex);
}
static void cpufreq_governor_limits(struct cpufreq_policy *policy,
struct dbs_data *dbs_data)
{
struct common_dbs_data *cdata = dbs_data->cdata;
unsigned int cpu = policy->cpu;
struct cpu_dbs_info *cdbs = cdata->get_cpu_cdbs(cpu);
if (!cdbs->shared || !cdbs->shared->policy)
return;
mutex_lock(&cdbs->shared->timer_mutex);
if (policy->max < cdbs->shared->policy->cur)
__cpufreq_driver_target(cdbs->shared->policy, policy->max,
CPUFREQ_RELATION_H);
else if (policy->min > cdbs->shared->policy->cur)
__cpufreq_driver_target(cdbs->shared->policy, policy->min,
CPUFREQ_RELATION_L);
dbs_check_cpu(dbs_data, cpu);
mutex_unlock(&cdbs->shared->timer_mutex);
}
int cpufreq_governor_dbs(struct cpufreq_policy *policy,
struct common_dbs_data *cdata, unsigned int event)
{
struct dbs_data *dbs_data;
int ret = 0;
cpufreq: governor: Serialize governor callbacks There are several races reported in cpufreq core around governors (only ondemand and conservative) by different people. There are at least two race scenarios present in governor code: (a) Concurrent access/updates of governor internal structures. It is possible that fields such as 'dbs_data->usage_count', etc. are accessed simultaneously for different policies using same governor structure (i.e. CPUFREQ_HAVE_GOVERNOR_PER_POLICY flag unset). And because of this we can dereference bad pointers. For example consider a system with two CPUs with separate 'struct cpufreq_policy' instances. CPU0 governor: ondemand and CPU1: powersave. CPU0 switching to powersave and CPU1 to ondemand: CPU0 CPU1 store* store* cpufreq_governor_exit() cpufreq_governor_init() dbs_data = cdata->gdbs_data; if (!--dbs_data->usage_count) kfree(dbs_data); dbs_data->usage_count++; *Bad pointer dereference* There are other races possible between EXIT and START/STOP/LIMIT as well. Its really complicated. (b) Switching governor state in bad sequence: For example trying to switch a governor to START state, when the governor is in EXIT state. There are some checks present in __cpufreq_governor() but they aren't sufficient as they compare events against 'policy->governor_enabled', where as we need to take governor's state into account, which can be used by multiple policies. These two issues need to be solved separately and the responsibility should be properly divided between cpufreq and governor core. The first problem is more about the governor core, as it needs to protect its structures properly. And the second problem should be fixed in cpufreq core instead of governor, as its all about sequence of events. This patch is trying to solve only the first problem. There are two types of data we need to protect, - 'struct common_dbs_data': No matter what, there is going to be a single copy of this per governor. - 'struct dbs_data': With CPUFREQ_HAVE_GOVERNOR_PER_POLICY flag set, we will have per-policy copy of this data, otherwise a single copy. Because of such complexities, the mutex present in 'struct dbs_data' is insufficient to solve our problem. For example we need to protect fetching of 'dbs_data' from different structures at the beginning of cpufreq_governor_dbs(), to make sure it isn't currently being updated. This can be fixed if we can guarantee serialization of event parsing code for an individual governor. This is best solved with a mutex per governor, and the placeholder for that is 'struct common_dbs_data'. And so this patch moves the mutex from 'struct dbs_data' to 'struct common_dbs_data' and takes it at the beginning and drops it at the end of cpufreq_governor_dbs(). Tested with and without following configuration options: CONFIG_LOCKDEP_SUPPORT=y CONFIG_DEBUG_RT_MUTEXES=y CONFIG_DEBUG_PI_LIST=y CONFIG_DEBUG_SPINLOCK=y CONFIG_DEBUG_MUTEXES=y CONFIG_DEBUG_LOCK_ALLOC=y CONFIG_PROVE_LOCKING=y CONFIG_LOCKDEP=y CONFIG_DEBUG_ATOMIC_SLEEP=y Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Reviewed-by: Preeti U Murthy <preeti@linux.vnet.ibm.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2015-06-03 10:27:13 +00:00
/* Lock governor to block concurrent initialization of governor */
mutex_lock(&cdata->mutex);
if (have_governor_per_policy())
dbs_data = policy->governor_data;
else
dbs_data = cdata->gdbs_data;
cpufreq: governor: Serialize governor callbacks There are several races reported in cpufreq core around governors (only ondemand and conservative) by different people. There are at least two race scenarios present in governor code: (a) Concurrent access/updates of governor internal structures. It is possible that fields such as 'dbs_data->usage_count', etc. are accessed simultaneously for different policies using same governor structure (i.e. CPUFREQ_HAVE_GOVERNOR_PER_POLICY flag unset). And because of this we can dereference bad pointers. For example consider a system with two CPUs with separate 'struct cpufreq_policy' instances. CPU0 governor: ondemand and CPU1: powersave. CPU0 switching to powersave and CPU1 to ondemand: CPU0 CPU1 store* store* cpufreq_governor_exit() cpufreq_governor_init() dbs_data = cdata->gdbs_data; if (!--dbs_data->usage_count) kfree(dbs_data); dbs_data->usage_count++; *Bad pointer dereference* There are other races possible between EXIT and START/STOP/LIMIT as well. Its really complicated. (b) Switching governor state in bad sequence: For example trying to switch a governor to START state, when the governor is in EXIT state. There are some checks present in __cpufreq_governor() but they aren't sufficient as they compare events against 'policy->governor_enabled', where as we need to take governor's state into account, which can be used by multiple policies. These two issues need to be solved separately and the responsibility should be properly divided between cpufreq and governor core. The first problem is more about the governor core, as it needs to protect its structures properly. And the second problem should be fixed in cpufreq core instead of governor, as its all about sequence of events. This patch is trying to solve only the first problem. There are two types of data we need to protect, - 'struct common_dbs_data': No matter what, there is going to be a single copy of this per governor. - 'struct dbs_data': With CPUFREQ_HAVE_GOVERNOR_PER_POLICY flag set, we will have per-policy copy of this data, otherwise a single copy. Because of such complexities, the mutex present in 'struct dbs_data' is insufficient to solve our problem. For example we need to protect fetching of 'dbs_data' from different structures at the beginning of cpufreq_governor_dbs(), to make sure it isn't currently being updated. This can be fixed if we can guarantee serialization of event parsing code for an individual governor. This is best solved with a mutex per governor, and the placeholder for that is 'struct common_dbs_data'. And so this patch moves the mutex from 'struct dbs_data' to 'struct common_dbs_data' and takes it at the beginning and drops it at the end of cpufreq_governor_dbs(). Tested with and without following configuration options: CONFIG_LOCKDEP_SUPPORT=y CONFIG_DEBUG_RT_MUTEXES=y CONFIG_DEBUG_PI_LIST=y CONFIG_DEBUG_SPINLOCK=y CONFIG_DEBUG_MUTEXES=y CONFIG_DEBUG_LOCK_ALLOC=y CONFIG_PROVE_LOCKING=y CONFIG_LOCKDEP=y CONFIG_DEBUG_ATOMIC_SLEEP=y Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Reviewed-by: Preeti U Murthy <preeti@linux.vnet.ibm.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2015-06-03 10:27:13 +00:00
if (WARN_ON(!dbs_data && (event != CPUFREQ_GOV_POLICY_INIT))) {
ret = -EINVAL;
goto unlock;
}
switch (event) {
case CPUFREQ_GOV_POLICY_INIT:
ret = cpufreq_governor_init(policy, dbs_data, cdata);
break;
case CPUFREQ_GOV_POLICY_EXIT:
cpufreq_governor_exit(policy, dbs_data);
break;
case CPUFREQ_GOV_START:
ret = cpufreq_governor_start(policy, dbs_data);
break;
case CPUFREQ_GOV_STOP:
cpufreq_governor_stop(policy, dbs_data);
break;
case CPUFREQ_GOV_LIMITS:
cpufreq_governor_limits(policy, dbs_data);
break;
}
cpufreq: governor: Serialize governor callbacks There are several races reported in cpufreq core around governors (only ondemand and conservative) by different people. There are at least two race scenarios present in governor code: (a) Concurrent access/updates of governor internal structures. It is possible that fields such as 'dbs_data->usage_count', etc. are accessed simultaneously for different policies using same governor structure (i.e. CPUFREQ_HAVE_GOVERNOR_PER_POLICY flag unset). And because of this we can dereference bad pointers. For example consider a system with two CPUs with separate 'struct cpufreq_policy' instances. CPU0 governor: ondemand and CPU1: powersave. CPU0 switching to powersave and CPU1 to ondemand: CPU0 CPU1 store* store* cpufreq_governor_exit() cpufreq_governor_init() dbs_data = cdata->gdbs_data; if (!--dbs_data->usage_count) kfree(dbs_data); dbs_data->usage_count++; *Bad pointer dereference* There are other races possible between EXIT and START/STOP/LIMIT as well. Its really complicated. (b) Switching governor state in bad sequence: For example trying to switch a governor to START state, when the governor is in EXIT state. There are some checks present in __cpufreq_governor() but they aren't sufficient as they compare events against 'policy->governor_enabled', where as we need to take governor's state into account, which can be used by multiple policies. These two issues need to be solved separately and the responsibility should be properly divided between cpufreq and governor core. The first problem is more about the governor core, as it needs to protect its structures properly. And the second problem should be fixed in cpufreq core instead of governor, as its all about sequence of events. This patch is trying to solve only the first problem. There are two types of data we need to protect, - 'struct common_dbs_data': No matter what, there is going to be a single copy of this per governor. - 'struct dbs_data': With CPUFREQ_HAVE_GOVERNOR_PER_POLICY flag set, we will have per-policy copy of this data, otherwise a single copy. Because of such complexities, the mutex present in 'struct dbs_data' is insufficient to solve our problem. For example we need to protect fetching of 'dbs_data' from different structures at the beginning of cpufreq_governor_dbs(), to make sure it isn't currently being updated. This can be fixed if we can guarantee serialization of event parsing code for an individual governor. This is best solved with a mutex per governor, and the placeholder for that is 'struct common_dbs_data'. And so this patch moves the mutex from 'struct dbs_data' to 'struct common_dbs_data' and takes it at the beginning and drops it at the end of cpufreq_governor_dbs(). Tested with and without following configuration options: CONFIG_LOCKDEP_SUPPORT=y CONFIG_DEBUG_RT_MUTEXES=y CONFIG_DEBUG_PI_LIST=y CONFIG_DEBUG_SPINLOCK=y CONFIG_DEBUG_MUTEXES=y CONFIG_DEBUG_LOCK_ALLOC=y CONFIG_PROVE_LOCKING=y CONFIG_LOCKDEP=y CONFIG_DEBUG_ATOMIC_SLEEP=y Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Reviewed-by: Preeti U Murthy <preeti@linux.vnet.ibm.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2015-06-03 10:27:13 +00:00
unlock:
mutex_unlock(&cdata->mutex);
return ret;
}
EXPORT_SYMBOL_GPL(cpufreq_governor_dbs);