CONTENTS 1. Introduction 1.1 Heterogeneous Systems 1.2 CPU Frequency Guidance 2. Window-Based Load Tracking Scheme 2.1 Synchronized Windows 2.2 struct ravg 2.3 Scaling Load Statistics 2.4 sched_window_stats_policy 2.5 Task Events 2.6 update_task_ravg() 2.7 update_history() 2.8 Per-task 'initial task load' 3. CPU Capacity 3.1 Load scale factor 3.2 CPU Power 4. CPU Power 5. HMP Scheduler 5.1 Classification of Tasks and CPUs 5.2 select_best_cpu() 5.2.1 sched_boost 5.2.2 task_will_fit() 5.2.3 Tunables affecting select_best_cpu() 5.2.4 Wakeup Logic 5.3 Scheduler Tick 5.4 Load Balancer 5.5 Real Time Tasks 5.6 Task packing 6. Frequency Guidance 6.1 Per-CPU Window-Based Stats 6.2 Per-task Window-Based Stats 6.3 Effect of various task events 6.4 Tying it all together 7. Tunables 8. HMP Scheduler Trace Points 8.1 sched_enq_deq_task 8.2 sched_task_load 8.3 sched_cpu_load_* 8.4 sched_update_task_ravg 8.5 sched_update_history 8.6 sched_reset_all_windows_stats 8.7 sched_migration_update_sum 8.8 sched_get_busy 8.9 sched_freq_alert 8.10 sched_set_boost 9. Device Tree bindings =============== 1. INTRODUCTION =============== Scheduler extensions described in this document serves two goals: 1) handle heterogeneous multi-processor (HMP) systems 2) guide cpufreq governor on proactive changes to cpu frequency *** 1.1 Heterogeneous systems Heterogeneous systems have cpus that differ with regard to their performance and power characteristics. Some cpus could offer peak performance better than others, although at cost of consuming more power. We shall refer such cpus as "high performance" or "performance efficient" cpus. Other cpus that offer lesser peak performance are referred to as "power efficient". In this situation the scheduler is tasked with the responsibility of assigning tasks to run on the right cpus where their performance requirements can be met at the least expense of power. Achieving that goal is made complicated by the fact that the scheduler has little clue about performance requirements of tasks and how they may change by running on power or performance efficient cpus! One simplifying assumption here could be that a task's desire for more performance is expressed by its cpu utilization. A task demanding high cpu utilization on a power-efficient cpu would likely improve in its performance by running on a performance-efficient cpu. This idea forms the basis for HMP-related scheduler extensions. Key inputs required by the HMP scheduler for its task placement decisions are: a) task load - this reflects cpu utilization or demand of tasks b) CPU capacity - this reflects peak performance offered by cpus c) CPU power - this reflects power or energy cost of cpus Once all 3 pieces of information are available, the HMP scheduler can place tasks on the lowest power cpus where their demand can be satisfied. *** 1.2 CPU Frequency guidance A somewhat separate but related goal of the scheduler extensions described here is to provide guidance to the cpufreq governor on the need to change cpu frequency. Most governors that control cpu frequency work on a reactive basis. CPU utilization is sampled at regular intervals, based on which the need to change frequency is determined. Higher utilization leads to a frequency increase and vice-versa. There are several problems with this approach that scheduler can help resolve. a) latency Reactive nature introduces latency for cpus to ramp up to desired speed which can hurt application performance. This is inevitable as cpufreq governors can only track cpu utilization as a whole and not tasks which are driving that demand. Scheduler can however keep track of individual task demand and can alert the governor on changing task activity. For example, request raise in frequency when tasks activity is increasing on a cpu because of wakeup or migration or request frequency to be lowered when task activity is decreasing because of sleep/exit or migration. b) part-picture Most governors track utilization of each CPU independently. When a task migrates from one cpu to another the task's execution time is split across the two cpus. The governor can fail to see the full picture of task demand in this case and thus the need for increasing frequency, affecting the task's performance. Scheduler can keep track of task migrations, fix up busy time upon migration and report per-cpu busy time to the governor that reflects task demand accurately. The rest of this document explains key enhancements made to the scheduler to accomplish both of the aforementioned goals. ==================================== 2. WINDOW-BASED LOAD TRACKING SCHEME ==================================== As mentioned in the introduction section, knowledge of the CPU demand exerted by a task is a prerequisite to knowing where to best place the task in an HMP system. The per-entity load tracking (PELT) scheme, present in Linux kernel since v3.7, has some perceived shortcomings when used to place tasks on HMP systems or provide recommendations on CPU frequency. Per-entity load tracking does not make a distinction between the ramp up vs ramp down time of task load. It also decays task load without exception when a task sleeps. As an example, a cpu bound task at its peak load (LOAD_AVG_MAX or 47742) can see its load decay to 0 after a sleep of just 213ms! A cpu-bound task running on a performance-efficient cpu could thus get re-classified as not requiring such a cpu after a short sleep. In the case of mobile workloads, tasks could go to sleep due to a lack of user input. When they wakeup it is very likely their cpu utilization pattern repeats. Resetting their load across sleep and incurring latency to reclassify them as requiring a high performance cpu can hurt application performance. The window-based load tracking scheme described in this document avoids these drawbacks. It keeps track of N windows of execution for every task. Windows where a task had no activity are ignored and not recorded. N can be tuned at compile time (RAVG_HIST_SIZE defined in include/linux/sched.h) or at runtime (/proc/sys/kernel/sched_ravg_hist_size). The window size, W, is common for all tasks and currently defaults to 10ms ('sched_ravg_window' defined in kernel/sched/core.c). The window size can be tuned at boot time via the sched_ravg_window=W argument to kernel. Alternately it can be tuned after boot via tunables provided by the interactive governor. More on this later. Based on the N samples available per-task, a per-task "demand" attribute is calculated which represents the cpu demand of that task. The demand attribute is used to classify tasks as to whether or not they need a performance-efficient CPU and also serves to provide inputs on frequency to the cpufreq governor. More on this later. The 'sched_window_stats_policy' tunable (defined in kernel/sched/core.c) controls how the demand field for a task is derived from its N past samples. *** 2.1 Synchronized windows Windows of observation for task activity are synchronized across cpus. This greatly aids in the scheduler's frequency guidance feature. Scheduler currently relies on a synchronized clock (sched_clock()) for this feature to work. It may be possible to extend this feature to work on systems having an unsynchronized sched_clock(). struct rq { .. u64 window_start; .. }; The 'window_start' attribute represents the time when current window began on a cpu. It is updated when key task events such as wakeup or context-switch call update_task_ravg() to record task activity. The window_start value is expected to be the same for all cpus, although it could be behind on some cpus where it has not yet been updated because update_task_ravg() has not been recently called. For example, when a cpu is idle for a long time its window_start could be stale. The window_start value for such cpus is rolled forward upon occurrence of a task event resulting in a call to update_task_ravg(). *** 2.2 struct ravg The ravg struct contains information tracked per-task. struct ravg { u64 mark_start; u32 sum, demand; u32 sum_history[RAVG_HIST_SIZE]; }; struct task_struct { .. struct ravg ravg; .. }; sum_history[] - stores cpu utilization samples from N previous windows where task had activity sum - stores cpu utilization of the task in its most recently tracked window. Once the corresponding window terminates, 'sum' will be pushed into the sum_history[] array and is then reset to 0. It is possible that the window corresponding to sum is not the current window being tracked on a cpu. For example, a task could go to sleep in window X and wakeup in window Y (Y > X). In this case, sum would correspond to the task's activity seen in window X. When update_task_ravg() is called during the task's wakeup event it will be seen that window X has elapsed. The sum value will be pushed to 'sum_history[]' array before being reset to 0. demand - represents task's cpu demand and is derived from the elements in sum_history[]. The section on 'sched_window_stats_policy' provides more details on how 'demand' is derived from elements in sum_history[] array mark_start - records timestamp of the beginning of the most recent task event. See section on 'Task events' for possible events that update 'mark_start' curr_window - this is described in the section on 'Frequency guidance' prev_window - this is described in the section on 'Frequency guidance' *** 2.3 Scaling load statistics Time required for a task to complete its work (and hence its load) depends on, among various other factors, cpu frequency and its efficiency. In a HMP system, some cpus are more performance efficient than others. Performance efficiency of a cpu can be described by its "instructions-per-cycle" (IPC) attribute. History of task execution could involve task having run at different frequencies and on cpus with different IPC attributes. To avoid ambiguity of how task load relates to the frequency and IPC of cpus on which a task has run, task load is captured in a scaled form, with scaling being done in reference to an "ideal" cpu that has best possible IPC and frequency. Such an "ideal" cpu, having the best possible frequency and IPC, may or may not exist in system. As an example, consider a HMP system, with two types of cpus, A53 and A57. A53 has IPC count of 1024 and can run at maximum frequency of 1 GHz, while A57 has IPC count of 2048 and can run at maximum frequency of 2 GHz. Ideal cpu in this case is A57 running at 2 GHz. A unit of work that takes 100ms to finish on A53 running at 100MHz would get done in 10ms on A53 running at 1GHz, in 5 ms running on A57 at 1 GHz and 2.5ms on A57 running at 2 GHz. Thus a load of 100ms can be expressed as 2.5ms in reference to ideal cpu of A57 running at 2 GHz. In order to understand how much load a task will consume on a given cpu, its scaled load needs to be multiplied by a factor (load scale factor). In above example, scaled load of 2.5ms needs to be multiplied by a factor of 4 in order to estimate the load of task on A53 running at 1 GHz. /proc/sched_debug provides IPC attribute and load scale factor for every cpu. In summary, task load information stored in a task's sum_history[] array is scaled for both frequency and efficiency. If a task runs for X ms, then the value stored in its 'sum' field is derived as: X_s = X * (f_cur / max_possible_freq) * (efficiency / max_possible_efficiency) where: X = cpu utilization that needs to be accounted X_s = Scaled derivative of X f_cur = current frequency of the cpu where the task was running max_possible_freq = maximum possible frequency (across all cpus) efficiency = instructions per cycle (IPC) of cpu where task was running max_possible_efficiency = maximum IPC offered by any cpu in system *** 2.4 sched_window_stats_policy sched_window_stats_policy controls how the 'demand' attribute for a task is derived from elements in its 'sum_history[]' array. WINDOW_STATS_RECENT (0) demand = recent WINDOW_STATS_MAX (1) demand = max WINDOW_STATS_MAX_RECENT_AVG (2) demand = maximum(average, recent) WINDOW_STATS_AVG (3) demand = average where: M = history size specified by /proc/sys/kernel/sched_ravg_hist_size average = average of first M samples found in the sum_history[] array max = maximum value of first M samples found in the sum_history[] array recent = most recent sample (sum_history[0]) demand = demand attribute found in 'struct ravg' This policy can be changed at runtime via /proc/sys/kernel/sched_window_stats_policy. For example, the command below would select WINDOW_STATS_USE_MAX policy echo 1 > /proc/sys/kernel/sched_window_stats_policy *** 2.5 Task events A number of events results in the window-based stats of a task being updated. These are: PICK_NEXT_TASK - the task is about to start running on a cpu PUT_PREV_TASK - the task stopped running on a cpu TASK_WAKE - the task is waking from sleep TASK_MIGRATE - the task is migrating from one cpu to another TASK_UPDATE - this event is invoked on a currently running task to update the task's window-stats and also the cpu's window-stats such as 'window_start' IRQ_UPDATE - event to record the busy time spent by an idle cpu processing interrupts *** 2.6 update_task_ravg() update_task_ravg() is called to mark the beginning of an event for a task or a cpu. It serves to accomplish these functions: a. Update a cpu's window_start value b. Update a task's window-stats (sum, sum_history[], demand and mark_start) In addition update_task_ravg() updates the busy time information for the given cpu, which is used for frequency guidance. This is described further in section 6. *** 2.7 update_history() update_history() is called on a task to record its activity in an elapsed window. 'sum', which represents task's cpu demand in its elapsed window is pushed onto sum_history[] array and its 'demand' attribute is updated based on the sched_window_stats_policy in effect. *** 2.8 Initial task load attribute for a task (init_load_pct) In some cases, it may be desirable for children of a task to be assigned a "high" load so that they can start running on best capacity cluster. By default, newly created tasks are assigned a load defined by tunable sched_init_task_load (Sec 7.8). Some specialized tasks may need a higher value than the global default for their child tasks. This will let child tasks run on cpus with best capacity. This is accomplished by setting the 'initial task load' attribute (init_load_pct) for a task. Child tasks starting load (ravg.demand and ravg.sum_history[]) is initialized from their parent's 'initial task load' attribute. Note that child task's 'initial task load' attribute itself will be 0 by default (i.e it is not inherited from parent). A task's 'initial task load' attribute can be set in two ways: **** /proc interface /proc/[pid]/sched_init_task_load can be written to for setting a task's 'initial task load' attribute. A numeric value between 0 - 100 (in percent scale) is accepted for task's 'initial task load' attribute. Reading /proc/[pid]/sched_init_task_load returns the 'initial task load' attribute for the given task. **** kernel API Following kernel APIs are provided to set or retrieve a given task's 'initial task load' attribute: int sched_set_init_task_load(struct task_struct *p, int init_load_pct); int sched_get_init_task_load(struct task_struct *p); =============== 3. CPU CAPACITY =============== CPU capacity reflects peak performance offered by a cpu. It is defined both by maximum frequency at which cpu can run and its efficiency attribute. Capacity of a cpu is defined in reference to "least" performing cpu such that "least" performing cpu has capacity of 1024. capacity = 1024 * (fmax_cur * / min_max_freq) * (efficiency / min_possible_efficiency) where: fmax_cur = maximum frequency at which cpu is currently allowed to run at efficiency = IPC of cpu min_max_freq = max frequency at which "least" performing cpu can run min_possible_efficiency = IPC of "least" performing cpu 'fmax_cur' reflects the fact that a cpu may be constrained at runtime to run at a maximum frequency less than what is supported. This may be a constraint placed by user or drivers such as thermal that intends to reduce temperature of a cpu by restricting its maximum frequency. 'max_possible_capacity' reflects the maximum capacity of a cpu based on the maximum frequency it supports. max_possible_capacity = 1024 * (fmax * / min_max_freq) * (efficiency / min_possible_efficiency) where: fmax = maximum frequency supported by a cpu /proc/sched_debug lists capacity and maximum_capacity information for a cpu. In the example HMP system quoted in Sec 2.3, "least" performing CPU is A53 and thus min_max_freq = 1GHz and min_possible_efficiency = 1024. Capacity of A57 = 1024 * (2GHz / 1GHz) * (2048 / 1024) = 4096 Capacity of A53 = 1024 * (1GHz / 1GHz) * (1024 / 1024) = 1024 Capacity of A57 when constrained to run at maximum frequency of 500MHz can be calculated as: Capacity of A57 = 1024 * (500MHz / 1GHz) * (2048 / 1024) = 1024 *** 3.1 load_scale_factor 'lsf' or load scale factor attribute of a cpu is used to estimate load of a task on that cpu when running at its fmax_cur frequency. 'lsf' is defined in reference to "best" performing cpu such that it's lsf is 1024. 'lsf' for a cpu is defined as: lsf = 1024 * (max_possible_freq / fmax_cur) * (max_possible_efficiency / ipc) where: fmax_cur = maximum frequency at which cpu is currently allowed to run at ipc = IPC of cpu max_possible_freq = max frequency at which "best" performing cpu can run max_possible_efficiency = IPC of "best" performing cpu In the example HMP system quoted in Sec 2.3, "best" performing CPU is A57 and thus max_possible_freq = 2 GHz, max_possible_efficiency = 2048 lsf of A57 = 1024 * (2GHz / 2GHz) * (2048 / 2048) = 1024 lsf of A53 = 1024 * (2GHz / 1 GHz) * (2048 / 1024) = 4096 lsf of A57 constrained to run at maximum frequency of 500MHz can be calculated as: lsf of A57 = 1024 * (2GHz / 500Mhz) * (2048 / 2048) = 4096 To estimate load of a task on a given cpu running at its fmax_cur: load = scaled_load * lsf / 1024 A task with scaled load of 20% would thus be estimated to consume 80% bandwidth of A53 running at 1GHz. The same task with scaled load of 20% would be estimated to consume 160% bandwidth on A53 constrained to run at maximum frequency of 500MHz. load_scale_factor, thus, is very useful to estimate load of a task on a given cpu and thus to decide whether it can fit in a cpu or not. *** 3.2 cpu_power A metric 'cpu_power' related to 'capacity' is also listed in /proc/sched_debug. 'cpu_power' is ideally same for all cpus (1024) when they are idle and running at the same frequency. 'cpu_power' of a cpu can be scaled down from its ideal value to reflect reduced frequency it is operating at and also to reflect the amount of cpu bandwidth consumed by real-time tasks executing on it. 'cpu_power' metric is used by scheduler to decide task load distribution among cpus. CPUs with low 'cpu_power' will be assigned less task load compared to cpus with higher 'cpu_power' ============ 4. CPU POWER ============ The HMP scheduler extensions currently depend on an architecture-specific driver to provide runtime information on cpu power. In the absence of an architecture-specific driver, the scheduler will resort to using the max_possible_capacity metric of a cpu as a measure of its power. ================ 5. HMP SCHEDULER ================ For normal (SCHED_OTHER/fair class) tasks there are three paths in the scheduler which these HMP extensions affect. The task wakeup path, the load balancer, and the scheduler tick are each modified. Real-time and stop-class tasks are served by different code paths. These will be discussed separately. Prior to delving further into the algorithm and implementation however some definitions are required. *** 5.1 Classification of Tasks and CPUs With the extensions described thus far, the following information is available to the HMP scheduler: - per-task CPU demand information from either Per-Entity Load Tracking (PELT) or the window-based algorithm described above - a power value for each frequency supported by each CPU via the API described in section 4 - current CPU frequency, maximum CPU frequency (may be throttled by at runtime due to thermal conditions), maximum possible CPU frequency supported by hardware - data previously maintained within the scheduler such as the number of currently runnable tasks on each CPU Combined with tunable parameters, this information can be used to classify both tasks and CPUs to aid in the placement of tasks. - big task A big task is one that exerts a CPU demand too high for a particular CPU to satisfy. The scheduler will attempt to find a CPU with more capacity for such a task. The definition of "big" is specific to a task *and* a CPU. A task may be considered big on one CPU in the system and not big on another if the first CPU has less capacity than the second. What task demand is "too high" for a particular CPU? One obvious answer would be a task demand which, as measured by PELT or window-based load tracking, matches or exceeds the capacity of that CPU. A task which runs on a CPU for a long time, for example, might meet this criteria as it would report 100% demand of that CPU. It may be desirable however to classify tasks which use less than 100% of a particular CPU as big so that the task has some "headroom" to grow without its CPU bandwidth getting capped and its performance requirements not being met. This task demand is therefore a tunable parameter: /proc/sys/kernel/sched_upmigrate This value is a percentage. If a task consumes more than this much of a particular CPU, that CPU will be considered too small for the task. The task will thus be seen as a "big" task on the cpu and will reflect in nr_big_tasks statistics maintained for that cpu. Note that certain tasks (whose nice value exceeds SCHED_UPMIGRATE_MIN_NICE value or those that belong to a cgroup whose upmigrate_discourage flag is set) will never be classified as big tasks despite their high demand. As the load scale factor is calculated against current fmax, it gets boosted when a lower capacity CPU is restricted to run at lower fmax. The task demand is inflated in this scenario and the task upmigrates early to the maximum capacity CPU. Hence this threshold is auto-adjusted by a factor equal to max_possible_frequency/current_frequency of a lower capacity CPU. This adjustment happens only when the lower capacity CPU frequency is restricted. The same adjustment is applied to the downmigrate threshold as well. When the frequency restriction is relaxed, the previous values are restored. sched_up_down_migrate_auto_update macro defined in kernel/sched/core.c controls this auto-adjustment behavior and it is enabled by default. If the adjusted upmigrate threshold exceeds the window size, it is clipped to the window size. If the adjusted downmigrate threshold decreases the difference between the upmigrate and downmigrate, it is clipped to a value such that the difference between the modified and the original thresholds is same. - spill threshold Tasks will normally be placed on lowest power-cost cluster where they can fit. This could result in power-efficient cluster becoming overcrowded when there are "too" many low-demand tasks. Spill threshold provides a spill over criteria, wherein low-demand task are allowed to be placed on idle or busy cpus in high-performance cluster. Scheduler will avoid placing a task on a cpu if it can result in cpu exceeding its spill threshold, which is defined by two tunables: /proc/sys/kernel/sched_spill_nr_run (default: 10) /proc/sys/kernel/sched_spill_load (default : 100%) A cpu is considered to be above its spill level if it already has 10 tasks or if the sum of task load (scaled in reference to given cpu) and rq->cumulative_runnable_avg exceeds 'sched_spill_load'. - power band The scheduler may be faced with a tradeoff between power and performance when placing a task. If the scheduler sees two CPUs which can accommodate a task: CPU 1, power cost of 20, load of 10 CPU 2, power cost of 10, load of 15 It is not clear what the right choice of CPU is. The HMP scheduler offers the sched_powerband_limit tunable to determine how this situation should be handled. When the power delta between two CPUs is less than sched_powerband_limit_pct, load will be prioritized as the deciding factor as to which CPU is selected. If the power delta between two CPUs exceeds that, the lower power CPU is considered to be in a different "band" and it is selected, despite perhaps having a higher current task load. *** 5.2 select_best_cpu() CPU placement decisions for a task at its wakeup or creation time are the most important decisions made by the HMP scheduler. This section will describe the call flow and algorithm used in detail. The primary entry point for a task wakeup operation is try_to_wake_up(), located in kernel/sched/core.c. This function relies on select_task_rq() to determine the target CPU for the waking task. For fair-class (SCHED_OTHER) tasks, that request will be routed to select_task_rq_fair() in kernel/sched/fair.c. As part of these scheduler extensions a hook has been inserted into the top of that function. If HMP scheduling is enabled the normal scheduling behavior will be replaced by a call to select_best_cpu(). This function, select_best_cpu(), represents the heart of the HMP scheduling algorithm described in this document. Note that select_best_cpu() is also invoked for a task being created. The behavior of select_best_cpu() depends on several factors such as boost setting, choice of several tunables and on task demand. **** 5.2.1 Boost The task placement policy changes signifincantly when scheduler boost is in effect. When boost is in effect the scheduler ignores the power cost of placing tasks on CPUs. Instead it figures out the load on each CPU and then places task on the least loaded CPU. If the load of two or more CPUs is the same (generally when CPUs are idle) the task prefers to go highest capacity CPU in the system. A further enhancement during boost is the scheduler' early detection feature. While boost is in effect the scheduler checks for the precence of tasks that have been runnable for over some period of time within the tick. For such tasks the scheduler informs the governor of imminent need for high frequency. If there exists a task on the runqueue at the tick that has been runnable for greater than SCHED_EARLY_DETECTION_DURATION amount of time, it notifies the governor with a fabricated load of the full window at the highest frequency. The fabricated load is maintained until the task is no longer runnable or until the next tick. Boost can be set via either /proc/sys/kernel/sched_boost or by invoking kernel API sched_set_boost(). int sched_set_boost(int enable); Once turned on, boost will remain in effect until it is explicitly turned off. To allow for boost to be controlled by multiple external entities (application or kernel module) at same time, boost setting is reference counted. This means that two applications can turn on boost and the effect of boost is eliminated only after both applications have turned off boost. boost_refcount variable represents this reference count. **** 5.2.2 task_will_fit() The overall goal of select_best_cpu() is to place a task on the least power cluster where it can "fit" i.e where its cpu usage shall be below the capacity offered by cluster. Criteria for a task to be considered as fitting in a cluster is: i) A low-priority task, whose nice value is greater than SCHED_UPMIGRATE_MIN_NICE or whose cgroup has its upmigrate_discourage flag set, is considered to be fitting in all clusters, irrespective of their capacity and task's cpu demand. ii) All tasks are considered to fit in highest capacity cluster. iii) Task demand scaled in reference to the given cluster should be less than a threshold. See section on load_scale_factor to know more about how task demand is scaled in reference to a given cpu (cluster). The threshold used is normally sched_upmigrate. Its possible for a task's demand to exceed sched_upmigrate threshold in reference to a cluster when its upmigrated to higher capacity cluster. To prevent it from coming back immediately to lower capacity cluster, the task is not considered to "fit" on its earlier cluster until its demand has dropped below sched_downmigrate in reference to that earlier cluster. sched_downmigrate thus provides for some hysteresis control. **** 5.2.3 Factors affecting select_best_cpu() Behavior of select_best_cpu() is further controlled by several tunables and synchronous nature of wakeup. a. /proc/sys/kernel/sched_cpu_high_irqload A cpu whose irq load is greater than this threshold will not be considered eligible for placement. This threshold value in expressed in nanoseconds scale, with default threshold being 10000000 (10ms). See notes on sched_cpu_high_irqload tunable to understand how irq load on a cpu is measured. b. Synchronous nature of wakeup Synchronous wakeup is a hint to scheduler that the task issuing wakeup (i.e task currently running on cpu where wakeup is being processed by scheduler) will "soon" relinquish CPU. A simple example is two tasks communicating with each other using a pipe structure. When reader task blocks waiting for data, its woken by writer task after it has written data to pipe. Writer task usually blocks waiting for reader task to consume data in pipe (which may not have any more room for writes). Synchronous wakeup is accounted for by adjusting load of a cpu to not include load of currently running task. As a result, a cpu that has only one runnable task and which is currently processing synchronous wakeup will be considered idle. c. PF_WAKE_UP_IDLE Any task with this flag set will be woken up to an idle cpu (if one is available) independent of sched_prefer_idle flag setting, its demand and synchronous nature of wakeup. Similarly idle cpu is preferred during wakeup for any task that does not have this flag set but is being woken by a task with PF_WAKE_UP_IDLE flag set. For simplicity, we will use the term "PF_WAKE_UP_IDLE wakeup" to signify wakeups involving a task with PF_WAKE_UP_IDLE flag set. d. /proc/sys/kernel/sched_select_prev_cpu_us This threshold controls whether task placement goes through fast path or not. If task's wakeup time since last sleep is short there are high chances that it's better to place the task on its previous CPU. This reduces task placement latency, cache miss and number of migrations. Default value of sched_select_prev_cpu_us is 2000 (2ms). This can be turned off by setting it to 0. e. /proc/sys/kernel/sched_short_burst_ns This threshold controls whether a task is considered as "short-burst" or not. "short-burst" tasks are eligible for packing to avoid overhead associated with waking up an idle CPU. "non-idle" CPUs which are not loaded with IRQs and can accommodate the waking task without exceeding spill limits are considered. The ties are broken with load followed by previous CPU. This tunable does not affect cluster selection. It only affects CPU selection in a given cluster. This packing is skipped for tasks that are eligible for "wake-up-idle" and "boost". **** 5.2.4 Wakeup Logic for Task "p" Wakeup task placement logic is as follows: 1) Eliminate CPUs with high irq load based on sched_cpu_high_irqload tunable. 2) Eliminate CPUs where either the task does not fit or CPUs where placement will result in exceeding the spill threshold tunables. CPUs elimiated at this stage will be considered as backup choices incase none of the CPUs get past this stage. 3) Find out and return the least power CPU that satisfies all conditions above. 4) If two or more CPUs are projected to have the same power, break ties in the following preference order: a) The CPU is the task's previous CPU. b) The CPU is in the same cluster as the task's previous CPU. c) The CPU has the least load The placement logic described above does not apply when PF_WAKE_UP_IDLE is set for either the waker task or the wakee task. Instead the scheduler chooses the most power efficient idle CPU. 5) If no CPU is found after step 2, resort to backup CPU selection logic whereby the CPU with highest amount of spare capacity is selected. 6) If none of the CPUs have any spare capacity, return the task's previous CPU. *** 5.3 Scheduler Tick Every CPU is interrupted periodically to let kernel update various statistics and possibly preempt the currently running task in favor of a waiting task. This periodicity, determined by CONFIG_HZ value, is set at 10ms. There are various optimizations by which a CPU however can skip taking these interrupts (ticks). A cpu going idle for considerable time in one such case. HMP scheduler extensions brings in a change in processing of tick (scheduler_tick()) that can result in task migration. In case the currently running task on a cpu belongs to fair_sched class, a check is made if it needs to be migrated. Possible reasons for migrating task could be: a) A big task is running on a power-efficient cpu and a high-performance cpu is available (idle) to service it b) A task is starving on a CPU with high irq load. c) A task with upmigration discouraged is running on a performance cluster. See notes on 'cpu.upmigrate_discourage'. In case the test for migration turns out positive (which is expected to be rare event), a candidate cpu is identified for task migration. To avoid multiple task migrations to the same candidate cpu(s), identification of candidate cpu is serialized via global spinlock (migration_lock). *** 5.4 Load Balancer Load balance is a key functionality of scheduler that strives to distribute task across available cpus in a "fair" manner. Most of the complexity associated with this feature involves balancing fair_sched class tasks. Changes made to load balance code serve these goals: 1. Restrict flow of tasks from power-efficient cpus to high-performance cpu. Provide a spill-over threshold, defined in terms of number of tasks (sched_spill_nr_run) and cpu demand (sched_spill_load), beyond which tasks can spill over from power-efficient cpu to high-performance cpus. 2. Allow idle power-efficient cpus to pick up extra load from over-loaded performance-efficient cpu 3. Allow idle high-performance cpu to pick up big tasks from power-efficient cpu *** 5.5 Real Time Tasks Minimal changes introduced in treatment of real-time tasks by HMP scheduler aims at preferring scheduling of real-time tasks on cpus with low load on a power efficient cluster. Prior to HMP scheduler, the fast-path cpu selection for placing a real-time task (at wakeup) is its previous cpu, provided the currently running task on its previous cpu is not a real-time task or a real-time task with lower priority. Failing this, cpu selection in slow-path involves building a list of candidate cpus where the waking real-time task will be of highest priority and thus can be run immediately. The first cpu from this candidate list is chosen for the waking real-time task. Much of the premise for this simple approach is the assumption that real-time tasks often execute for very short intervals and thus the focus is to place them on a cpu where they can be run immediately. HMP scheduler brings in a change which avoids fast-path and always resorts to slow-path. Further cpu with lowest load in a power efficient cluster from candidate list of cpus is chosen as cpu for placing waking real-time task. - PF_WAKE_UP_IDLE Idle cpu is preferred for any waking task that has this flag set in its 'task_struct.flags' field. Further idle cpu is preferred for any task woken by such tasks. PF_WAKE_UP_IDLE flag of a task is inherited by it's children. It can be modified for a task in two ways: > kernel-space interface set_wake_up_idle() needs to be called in the context of a task to set or clear its PF_WAKE_UP_IDLE flag. > user-space interface /proc/[pid]/sched_wake_up_idle file needs to be written to for setting or clearing PF_WAKE_UP_IDLE flag for a given task ===================== 6. FREQUENCY GUIDANCE ===================== As mentioned in the introduction section the scheduler is in a unique position to assist with the determination of CPU frequency. Because the scheduler now maintains an estimate of per-task CPU demand, task activity can be tracked, aggregated and provided to the CPUfreq governor as a replacement for simple CPU busy time. Two of the most popular CPUfreq governors, interactive and ondemand, utilize a window-based approach for measuring CPU busy time. This works well with the window-based load tracking scheme previously described. The following APIs are provided to allow the CPUfreq governor to query busy time from the scheduler instead of using the basic CPU busy time value derived via get_cpu_idle_time_us() and get_cpu_iowait_time_us() APIs. int sched_set_window(u64 window_start, unsigned int window_size) This API is invoked by governor at initialization time or whenever window size is changed. 'window_size' argument (in jiffy units) indicates the size of window to be used. The first window of size 'window_size' is set to begin at jiffy 'window_start' -EINVAL is returned if per-entity load tracking is in use rather than window-based load tracking, otherwise a success value of 0 is returned. int sched_get_busy(int cpu) Returns the busy time for the given CPU in the most recent complete window. The value returned is microseconds of busy time at fmax of given CPU. The values returned by sched_get_busy() take a bit of explanation, both in what they mean and also how they are derived. *** 6.1 Per-CPU Window-Based Stats The scheduler tracks two separate types of quantities on a per CPU basis. The first type has to deal with the aggregate load on a CPU and the second type deals with top-tasks on that same CPU. We will first proceed to explain what is maintained as part of each type of statistics and then provide the connection between these two types of statistics at the end. First lets describe the HMP scheduler extensions to track the aggregate load seen on each CPU. This is done using the same windows that the task demand is tracked with (which is in turn set by the governor when frequency guidance is in use). There are four quantities maintained for each CPU by the HMP scheduler for tracking CPU load: curr_runnable_sum: aggregate demand from all tasks which executed during the current (not yet completed) window prev_runnable_sum: aggregate demand from all tasks which executed during the most recent completed window nt_curr_runnable_sum: aggregate demand from all 'new' tasks which executed during the current (not yet completed) window nt_prev_runnable_sum: aggregate demand from all 'new' tasks which executed during the most recent completed window. When the scheduler is updating a task's window-based stats it also updates these values. Like per-task window-based demand these quantities are normalized against the max possible frequency and max efficiency (instructions per cycle) in the system. If an update occurs and a window rollover is observed, curr_runnable_sum is copied into prev_runnable_sum before being reset to 0. The sched_get_busy() API returns prev_runnable_sum, scaled to the efficiency and fmax of given CPU. The same applies to nt_curr_runnable_sum and nt_prev_runnable_sum. A 'new' task is defined as a task whose number of active windows since fork is less than SCHED_NEW_TASK_WINDOWS. An active window is defined as a window where a task was observed to be runnable. Moving on the second type of statistics; top-tasks, the scheduler tracks a list of top tasks per CPU. A top-task is defined as the task that runs the most in a given window on that CPU. This includes task that ran on that CPU through out the window or were migrated to that CPU prior to window expiration. It does not include tasks that were migrated away from that CPU prior to window expiration. To track top tasks, we first realize that there is no strict need to maintain the task struct itself as long as we know the load exerted by the top task. We also realize that to maintain top tasks on every CPU we have to track the execution of every single task that runs during the window. The load associated with a task needs to be migrated when the task migrates from one CPU to another. When the top task migrates away, we need to locate the second top task and so on. Given the above realizations, we use hashmaps to track top task load both for the current and the previous window. This hashmap is implemented as an array of fixed size. The key of the hashmap is given by task_execution_time_in_a_window / array_size. The size of the array (number of buckets in the hashmap) dictate the load granularity of each bucket. The value stored in each bucket is a refcount of all the tasks that executed long enough to be in that bucket. This approach has a few benefits. Firstly, any top task stats update now take O(1) time. While task migration is also O(1), it does still involve going through up to the size of the array to find the second top task. We optimize this search by using bitmaps. The next set bit in the bitmap gives the position of the second top task in our hashamp. Secondly, and more importantly, not having to store the task struct itself saves a lot of memory usage in that 1) there is no need to retrieve task structs later causing cache misses and 2) we don't have to unnecessarily hold up task memory for up to 2 full windows by calling get_task_struct() after a task exits. Given the motivation above, here are a list of quantities tracked as part of per CPU task top-tasks management top_tasks[NUM_TRACKED_WINDOWS] - Hashmap of top-task load for the current and previous window BITMAP_ARRAY(top_tasks_bitmap) - Two bitmaps for the current and previous windows corresponding to the top-task hashmap. load_subs[NUM_TRACKED_WINDOWS] - An array of load subtractions to be carried out form curr/prev_runnable_sums for each CPU prior to reporting load to the governor. The purpose for this will be explained later in the section pertaining to the TASK_MIGRATE event. The type struct load_subtractions, stores the value of the subtraction along with the window start value for the window for which the subtraction has to take place. curr_table - Indication of which index of the array points to the current window. curr_top - The top task on a CPU at any given moment in the current window prev_top - The top task on a CPU in the previous window *** 6.2 Per-task window-based stats Corresponding to curr_runnable_sum and prev_runnable_sum, two counters are maintained per-task curr_window_cpu - represents task's contribution to cpu busy time on various CPUs in the current window prev_window_cpu - represents task's contribution to cpu busy time on various CPUs in the previous window curr_window - represents the sum of all entries in curr_window_cpu prev_window - represents the sum of all entries in prev_window_cpu "cpu demand" of a task includes its execution time and can also include its wait time. 'SCHED_FREQ_ACCOUNT_WAIT_TIME' controls whether task's wait time is included in its CPU load counters or not. Curr_runnable_sum counter of a cpu is derived from curr_window_cpu[cpu] counter of various tasks that ran on it in its most recent window. *** 6.3 Effect of various task events We now consider various events and how they affect above mentioned counters. PICK_NEXT_TASK This represents beginning of execution for a task. Provided the task refers to a non-idle task, a portion of task's wait time that corresponds to the current window being tracked on a cpu is added to task's curr_window_cpu and curr_window counter, provided SCHED_FREQ_ACCOUNT_WAIT_TIME is set. The same quantum is also added to cpu's curr_runnable_sum counter. The remaining portion, which corresponds to task's wait time in previous window is added to task's prev_window, prev_window_cpu and cpu's prev_runnable_sum counters. CPUs top_tasks hashmap is updated if needed with the new information. Any previous entries in the hashmap are deleted and newer entries are created. The top_tasks_bitmap reflects the updated state of the hashmap. If the top task for the current and/or previous window has changed, curr_top and prev_top are updated accordingly. PUT_PREV_TASK This represents end of execution of a time-slice for a task, where the task could refer to a cpu's idle task also. In case the task is non-idle or (in case of task being idle with cpu having non-zero rq->nr_iowait count and sched_io_is_busy =1), a portion of task's execution time, that corresponds to current window being tracked on a cpu is added to task's curr_window_cpu and curr_window counter and also to cpu's curr_runnable_sum counter. Portion of task's execution that corresponds to the previous window is added to task's prev_window, prev_window_cpu and cpu's prev_runnable_sum counters. CPUs top_tasks hashmap is updated if needed with the new information. Any previous entries in the hashmap are deleted and newer entries are created. The top_tasks_bitmap reflects the updated state of the hashmap. If the top task for the current and/or previous window has changed, curr_top and prev_top are updated accordingly. TASK_UPDATE This event is called on a cpu's currently running task and hence behaves effectively as PUT_PREV_TASK. Task continues executing after this event, until PUT_PREV_TASK event occurs on the task (during context switch). TASK_WAKE This event signifies a task waking from sleep. Since many windows could have elapsed since the task went to sleep, its curr_window_cpu/curr_window and prev_window_cpu/prev_window are updated to reflect task's demand in the most recent and its previous window that is being tracked on a cpu. Updated stats will trigger the same book-keeping for top-tasks as other events. TASK_MIGRATE This event signifies task migration across cpus. It is invoked on the task prior to being moved. Thus at the time of this event, the task can be considered to be in "waiting" state on src_cpu. In that way this event reflects actions taken under PICK_NEXT_TASK (i.e its wait time is added to task's curr/prev_window/_cpu counters as well as src_cpu's curr/prev_runnable_sum counters, provided SCHED_FREQ_ACCOUNT_WAIT_TIME is non-zero). After that update, we make a distinction between intra-cluster and inter-cluster migrations for further book-keeping. For intra-cluster migrations, we simply remove the entry for the task in the top_tasks hashmap from the source CPU and add the entry to the destination CPU. The top_tasks_bitmap, curr_top and prev_top are updated accordingly. We then find the second top-task top in our top_tasks hashmap for both the current and previous window and set curr_top and prev_top to their new values. For inter-cluster migrations we have a much more complicated scheme. Firstly we add to the destination CPU's curr/prev_runnable_sum the tasks curr/prev_window. Note we add the sum and not the contribution any individual CPU. This is because when a tasks migrates across clusters, we need the new cluster to ramp up to the appropriate frequency given the task's total execution summed up across all CPUs in the previous cluster. Secondly the src_cpu's curr/prev_runnable_sum are reduced by task's curr/prev_window_cpu values. Thirdly, we need to walk all the CPUs in the cluster and subtract from each CPU's curr/prev_runnable_sum the task's respective curr/prev_window_cpu values. However, subtracting load from each of the source CPUs is not trivial, as it would require all runqueue locks to be held. To get around this we introduce a deferred load subtraction mechanism whereby subtracting load from each of the source CPUs is deferred until an opportune moment. This opportune moment is when the governor comes asking the scheduler for load. At that time, all necessary runqueue locks are already held. There are a few cases to consider when doing deferred subtraction. Since we are not holding all runqueue locks other CPUs in the source cluster can be in a different window than the source CPU where the task is migrating from. Case 1: Other CPU in the source cluster is in the same window. No special consideration. Case 2: Other CPU in the source cluster is ahead by 1 window. In this case, we will be doing redundant updates to subtraction load for the prev window. There is no way to avoid this redundant update though, without holding the rq lock. Case 3: Other CPU in the source cluster is trailing by 1 window In this case, we might end up overwriting old data for that CPU. But this is not a problem as when the other CPU calls update_task_ravg() it will move to the same window. This relies on maintaining synchronized windows between CPUs, which is true today. To achieve all the above, we simple add the task's curr/prev_window_cpu contributions to the per CPU load_subtractions array. These load subtractions are subtracted from the respective CPU's curr/prev_runnable_sums before the governor queries CPU load. Once this is complete, the scheduler sets all curr/prev_window_cpu contributions of the task to 0 for all CPUs in the source cluster. The destination CPUs's curr/prev_window_cpu is updated with the tasks curr/prev_window sums. Finally, we must deal with frequency aggregation. When frequency aggregation is in effect, there is little point in dealing with per CPU footprint since the load of all related tasks have to be reported on a single CPU. Therefore when a task enters a related group we clear out all per CPU contributions and add it to the task CPU's cpu_time struct. From that point onwards we stop managing per CPU contributions upon inter cluster migrations since that work is redundant. Finally when a task exits a related group we must walk every CPU in reset all CPU contributions. We then set the task CPU contribution to the respective curr/prev sum values and add that sum to the task CPU rq runnable sum. Top-task management is the same as in the case of intra-cluster migrations. IRQ_UPDATE This event signifies end of execution of an interrupt handler. This event results in update of cpu's busy time counters, curr_runnable_sum and prev_runnable_sum, provided cpu was idle. When sched_io_is_busy = 0, only the interrupt handling time is added to cpu's curr_runnable_sum and prev_runnable_sum counters. When sched_io_is_busy = 1, the event mirrors actions taken under TASK_UPDATED event i.e time since last accounting of idle task's cpu usage is added to cpu's curr_runnable_sum and prev_runnable_sum counters. No update is needed for top-tasks in this case. *** 6.4 Tying it all together Now the scheduler maintains two independent quantities for load reporing 1) CPU load as represented by prev_runnable_sum and 2) top-tasks. The reported load is governed by tunable sched_freq_reporting_policy. The default choice is FREQ_REPORT_MAX_CPU_LOAD_TOP_TASK. In other words: max(prev_runnable_sum, top_task load) Let's explain the rationale behind the choice. CPU load tracks the exact amount of execution observed on a CPU. This is close to the quantity that the vanilla governor used to track. It offers the advantages of no load over-reporting that our earlier load fixup mechanisms had deal with. It then also tackles the part picture problem by keeping of track of tasks that might be migrating across CPUs leaving a small footprint on each CPU. Since we maintain one top task per CPU, we can handle as many top tasks as the number of CPUs in a cluster. We might miss a few cases where the combined load of the top and non-top tasks on a CPU are more representative of the true load. However, those cases have been deemed to rare and have little impact on overall load/frequency behavior. =========== 7. TUNABLES =========== *** 7.1 sched_spill_load Appears at: /proc/sys/kernel/sched_spill_load Default value: 100 CPU selection criteria for fair-sched class tasks is the lowest power cpu where they can fit. When the most power-efficient cpu where a task can fit is overloaded (aggregate demand of tasks currently queued on it exceeds sched_spill_load), a task can be placed on a higher-performance cpu, even though the task strictly doesn't need one. *** 7.2 sched_spill_nr_run Appears at: /proc/sys/kernel/sched_spill_nr_run Default value: 10 The intent of this tunable is similar to sched_spill_load, except it applies to nr_running count of a cpu. A task can spill over to a higher-performance cpu when the most power-efficient cpu where it can normally fit has more tasks than sched_spill_nr_run. *** 7.3 sched_upmigrate Appears at: /proc/sys/kernel/sched_upmigrate Default value: 80 This tunable is a percentage. If a task consumes more than this much of a CPU, the CPU is considered too small for the task and the scheduler will try to find a bigger CPU to place the task on. *** 7.4 sched_init_task_load Appears at: /proc/sys/kernel/sched_init_task_load Default value: 15 This tunable is a percentage. When a task is first created it has no history, so the task load tracking mechanism cannot determine a historical load value to assign to it. This tunable specifies the initial load value for newly created tasks. Also see Sec 2.8 on per-task 'initial task load' attribute. *** 7.5 sched_ravg_hist_size Appears at: /proc/sys/kernel/sched_ravg_hist_size Default value: 5 This tunable controls the number of samples used from task's sum_history[] array for determination of its demand. *** 7.6 sched_window_stats_policy Appears at: /proc/sys/kernel/sched_window_stats_policy Default value: 2 This tunable controls the policy in how window-based load tracking calculates an overall demand value based on the windows of CPU utilization it has collected for a task. Possible values for this tunable are: 0: Just use the most recent window sample of task activity when calculating task demand. 1: Use the maximum value of first M samples found in task's cpu demand history (sum_history[] array), where M = sysctl_sched_ravg_hist_size 2: Use the maximum of (the most recent window sample, average of first M samples), where M = sysctl_sched_ravg_hist_size 3. Use average of first M samples, where M = sysctl_sched_ravg_hist_size *** 7.7 sched_ravg_window Appears at: kernel command line argument Default value: 10000000 (10ms, units of tunable are nanoseconds) This specifies the duration of each window in window-based load tracking. By default each window is 10ms long. This quantity must currently be set at boot time on the kernel command line (or the default value of 10ms can be used). *** 7.8 RAVG_HIST_SIZE Appears at: compile time only (see RAVG_HIST_SIZE in include/linux/sched.h) Default value: 5 This macro specifies the number of windows the window-based load tracking mechanism maintains per task. If default values are used for both this and sched_ravg_window then a total of 50ms of task history would be maintained in 5 10ms windows. *** 7.9 sched_freq_inc_notify Appears at: /proc/sys/kernel/sched_freq_inc_notify Default value: 10 * 1024 * 1024 (10 Ghz) When scheduler detects that cur_freq of a cluster is insufficient to meet demand, it sends notification to governor, provided (freq_required - cur_freq) exceeds sched_freq_inc_notify, where freq_required is the frequency calculated by scheduler to meet current task demand. Note that sched_freq_inc_notify is specified in kHz units. *** 7.10 sched_freq_dec_notify Appears at: /proc/sys/kernel/sched_freq_dec_notify Default value: 10 * 1024 * 1024 (10 Ghz) When scheduler detects that cur_freq of a cluster is far greater than what is needed to serve current task demand, it will send notification to governor. More specifically, notification is sent when (cur_freq - freq_required) exceeds sched_freq_dec_notify, where freq_required is the frequency calculated by scheduler to meet current task demand. Note that sched_freq_dec_notify is specified in kHz units. *** 7.11 sched_cpu_high_irqload Appears at: /proc/sys/kernel/sched_cpu_high_irqload Default value: 10000000 (10ms) The scheduler keeps a decaying average of the amount of irq and softirq activity seen on each CPU within a ten millisecond window. Note that this "irqload" (reported in the sched_cpu_load_* tracepoint) will be higher than the typical load in a single window since every time the window rolls over, the value is decayed by some fraction and then added to the irq/softirq time spent in the next window. When the irqload on a CPU exceeds the value of this tunable, the CPU is no longer eligible for placement. This will affect the task placement logic described above, causing the scheduler to try and steer tasks away from the CPU. *** 7.12 cpu.upmigrate_discourage Default value : 0 This is a cgroup attribute supported by the cpu resource controller. It normally appears at [root_cpu]/[name1]/../[name2]/cpu.upmigrate_discourage. Here "root_cpu" is the mount point for cgroup (cpu resource control) filesystem and name1, name2 etc are names of cgroups that form a hierarchy. Setting this flag to 1 discourages upmigration for all tasks of a cgroup. High demand tasks of such a cgroup will never be classified as big tasks and hence not upmigrated. Any task of the cgroup is allowed to upmigrate only under overcommitted scenario. See notes on sched_spill_nr_run and sched_spill_load for how overcommitment threshold is defined. *** 7.13 sched_static_cpu_pwr_cost Default value: 0 Appears at /sys/devices/system/cpu/cpu/sched_static_cpu_pwr_cost This is the power cost associated with bringing an idle CPU out of low power mode. It ignores the actual C-state that a CPU may be in and assumes the worst case power cost of the highest C-state. It is means of biasing task placement away from idle CPUs when necessary. It can be defined per CPU, however, a more appropriate usage to define the same value for every CPU within a cluster and possibly have differing value between clusters as needed. *** 7.14 sched_static_cluster_pwr_cost Default value: 0 Appears at /sys/devices/system/cpu/cpu/sched_static_cluster_pwr_cost This is the power cost associated with bringing an idle cluster out of low power mode. It ignores the actual D-state that a cluster may be in and assumes the worst case power cost of the highest D-state. It is means of biasing task placement away from idle clusters when necessary. *** 7.15 sched_restrict_cluster_spill Default value: 0 Appears at /proc/sys/kernel/sched_restrict_cluster_spill This tunable can be used to restrict tasks spilling to the higher capacity (higher power) cluster. When this tunable is enabled, - Restrict the higher capacity cluster pulling tasks from the lower capacity cluster in the load balance path. The restriction is lifted if all of the CPUS in the lower capacity cluster are above spill. The power cost is used to break the ties if the capacity of clusters are same for applying this restriction. - The current CPU selection algorithm for RT tasks looks for the least loaded CPU across all clusters. When this tunable is enabled, the RT tasks are restricted to the lowest possible power cluster. *** 7.16 sched_downmigrate Appears at: /proc/sys/kernel/sched_downmigrate Default value: 60 This tunable is a percentage. It exists to control hysteresis. Lets say a task migrated to a high-performance cpu when it crossed 80% demand on a power-efficient cpu. We don't let it come back to a power-efficient cpu until its demand *in reference to the power-efficient cpu* drops less than 60% (sched_downmigrate). *** 7.17 sched_small_wakee_task_load Appears at: /proc/sys/kernel/sched_small_wakee_task_load Default value: 10 This tunable is a percentage. Configure the maximum demand of small wakee task. Sync wakee tasks which have demand less than sched_small_wakee_task_load are categorized as small wakee tasks. Scheduler places small wakee tasks on the waker's cluster. *** 7.18 sched_big_waker_task_load Appears at: /proc/sys/kernel/sched_big_waker_task_load Default value: 25 This tunable is a percentage. Configure the minimum demand of big sync waker task. Scheduler places small wakee tasks woken up by big sync waker on the waker's cluster. *** 7.19 sched_prefer_sync_wakee_to_waker Appears at: /proc/sys/kernel/sched_prefer_sync_wakee_to_waker Default value: 0 The default sync wakee policy has a preference to select an idle CPU in the waker cluster compared to the waker CPU running only 1 task. By selecting an idle CPU, it eliminates the chance of waker migrating to a different CPU after the wakee preempts it. This policy is also not susceptible to the incorrect "sync" usage i.e the waker does not goto sleep after waking up the wakee. However LPM exit latency associated with an idle CPU outweigh the above benefits on some targets. When this knob is turned on, the waker CPU is selected if it has only 1 runnable task. *** 7.20 sched_freq_reporting_policy Appears at: /proc/sys/kernel/sched_freq_reporting_policy Default value: 0 This dictates what the load reporting policy to the governor should be. The default value is FREQ_REPORT_MAX_CPU_LOAD_TOP_TASK. Other values include FREQ_REPORT_CPU_LOAD which only reports CPU load to the governor and FREQ_REPORT_TOP_TASK which only reports the load of the top task on a CPU to the governor. ========================= 8. HMP SCHEDULER TRACE POINTS ========================= *** 8.1 sched_enq_deq_task Logged when a task is either enqueued or dequeued on a CPU's run queue. -0 [004] d.h4 12700.711665: sched_enq_deq_task: cpu=4 enqueue comm=powertop pid=13227 prio=120 nr_running=1 cpu_load=0 rt_nr_running=0 affine=ff demand=13364423 - cpu: the CPU that the task is being enqueued on to or dequeued off of - enqueue/dequeue: whether this was an enqueue or dequeue event - comm: name of task - pid: PID of task - prio: priority of task - nr_running: number of runnable tasks on this CPU - cpu_load: current priority-weighted load on the CPU (note, this is *not* the same as CPU utilization or a metric tracked by PELT/window-based tracking) - rt_nr_running: number of real-time processes running on this CPU - affine: CPU affinity mask in hex for this task (so ff is a task eligible to run on CPUs 0-7) - demand: window-based task demand computed based on selected policy (recent, max, or average) (ns) *** 8.2 sched_task_load Logged when selecting the best CPU to run the task (select_best_cpu()). sched_task_load: 4004 (adbd): demand=698425 boost=0 reason=0 sync=0 need_idle=0 best_cpu=0 latency=103177 - demand: window-based task demand computed based on selected policy (recent, max, or average) (ns) - boost: whether boost is in effect - reason: reason we are picking a new CPU: 0: no migration - selecting a CPU for a wakeup or new task wakeup 1: move to big CPU (migration) 2: move to little CPU (migration) 3: move to low irq load CPU (migration) - sync: is the nature synchronous in nature - need_idle: is an idle CPU required for this task based on PF_WAKE_UP_IDLE - best_cpu: The CPU selected by the select_best_cpu() function for placement - latency: The execution time of the function select_best_cpu() *** 8.3 sched_cpu_load_* Logged when selecting the best CPU to run a task (select_best_cpu() for fair class tasks, find_lowest_rq_hmp() for RT tasks) and load balancing (update_sg_lb_stats()). -0 [004] d.h3 12700.711541: sched_cpu_load_*: cpu 0 idle 1 nr_run 0 nr_big 0 lsf 1119 capacity 1024 cr_avg 0 irqload 3301121 fcur 729600 fmax 1459200 power_cost 5 cstate 2 temp 38 - cpu: the CPU being described - idle: boolean indicating whether the CPU is idle - nr_run: number of tasks running on CPU - nr_big: number of BIG tasks running on CPU - lsf: load scale factor - multiply normalized load by this factor to determine how much load task will exert on CPU - capacity: capacity of CPU (based on max possible frequency and efficiency) - cr_avg: cumulative runnable average, instantaneous sum of the demand (either PELT or window-based) of all the runnable task on a CPU (ns) - irqload: decaying average of irq activity on CPU (ns) - fcur: current CPU frequency (Khz) - fmax: max CPU frequency (but not maximum _possible_ frequency) (KHz) - power_cost: cost of running this CPU at the current frequency - cstate: current cstate of CPU - temp: current temperature of the CPU The power_cost value above differs in how it is calculated depending on the callsite of this tracepoint. The select_best_cpu() call to this tracepoint finds the minimum frequency required to satisfy the existing load on the CPU as well as the task being placed, and returns the power cost of that frequency. The load balance and real time task placement paths used a fixed frequency (highest frequency common to all CPUs for load balancing, minimum frequency of the CPU for real time task placement). *** 8.4 sched_update_task_ravg Logged when window-based stats are updated for a task. The update may happen for a variety of reasons, see section 2.5, "Task Events." rcu_preempt-7 [000] d..3 262857.738888: sched_update_task_ravg: wc 262857521127957 ws 262857490000000 delta 31127957 event PICK_NEXT_TASK cpu 0 cur_freq 291055 cur_pid 7 task 9309 (kworker/u16:0) ms 262857520627280 delta 500677 demand 282196 sum 156201 irqtime 0 pred_demand 267103 rq_cs 478718 rq_ps 0 cur_window 78433 (78433 0 0 0 0 0 0 0 ) prev_window 146430 (0 146430 0 0 0 0 0 0 ) nt_cs 0 nt_ps 0 active_wins 149 grp_cs 0 grp_ps 0, grp_nt_cs 0, grp_nt_ps: 0 curr_top 6 prev_top 2 - wc: wallclock, output of sched_clock(), monotonically increasing time since boot (will roll over in 585 years) (ns) - ws: window start, time when the current window started (ns) - delta: time since the window started (wc - ws) (ns) - event: What event caused this trace event to occur (see section 2.5 for more details) - cpu: which CPU the task is running on - cur_freq: CPU's current frequency in KHz - curr_pid: PID of the current running task (current) - task: PID and name of task being updated - ms: mark start - timestamp of the beginning of a segment of task activity, either sleeping or runnable/running (ns) - delta: time since last event within the window (wc - ms) (ns) - demand: task demand computed based on selected policy (recent, max, or average) (ns) - sum: the task's run time during current window scaled by frequency and efficiency (ns) - irqtime: length of interrupt activity (ns). A non-zero irqtime is seen when an idle cpu handles interrupts, the time for which needs to be accounted as cpu busy time - cs: curr_runnable_sum of cpu (ns). See section 6.1 for more details of this counter. - ps: prev_runnable_sum of cpu (ns). See section 6.1 for more details of this counter. - cur_window: cpu demand of task in its most recently tracked window summed up across all CPUs (ns). This is followed by a list of contributions on each individual CPU. - prev_window: cpu demand of task in its previous window summed up across all CPUs (ns). This is followed by a list of contributions on each individual CPU. - nt_cs: curr_runnable_sum of a cpu for new tasks only (ns). - nt_ps: prev_runnable_sum of a cpu for new tasks only (ns). - active_wins: No. of active windows since task statistics were initialized - grp_cs: curr_runnable_sum for colocated tasks. This is independent from cs described above. The addition of these two fields give the total CPU load for the most recent window - grp_ps: prev_runnable_sum for colocated tasks. This is independent from ps described above. The addition of these two fields give the total CPU load for the previous window. - grp_nt_cs: curr_runnable_sum of a cpu for grouped new tasks only (ns). - grp_nt_ps: prev_runnable_sum for a cpu for grouped new tasks only (ns). - curr_top: index of the top task in the top_tasks array in the current window for a CPU. - prev_top: index of the top task in the top_tasks array in the previous window for a CPU *** 8.5 sched_update_history Logged when update_task_ravg() is accounting task activity into one or more windows that have completed. This may occur more than once for a single call into update_task_ravg(). A task that ran for 24ms spanning four 10ms windows (the last 2ms of window 1, all of windows 2 and 3, and the first 2ms of window 4) would result in two calls into update_history() from update_task_ravg(). The first call would record activity in completed window 1 and second call would record activity for windows 2 and 3 together (samples will be 2 in second call). -0 [004] d.h4 12700.711489: sched_update_history: 13227 (powertop): runtime 13364423 samples 1 event TASK_WAKE demand 13364423 (hist: 13364423 9871252 2236009 6162476 10282078) cpu 4 nr_big 0 - runtime: task cpu demand in recently completed window(s). This value is scaled to max_possible_freq and max_possible_efficiency. This value is pushed into task's demand history array. The number of windows to which runtime applies is provided by samples field. - samples: Number of samples (windows), each having value of runtime, that is recorded in task's demand history array. - event: What event caused this trace event to occur (see section 2.5 for more details) - PUT_PREV_TASK, PICK_NEXT_TASK, TASK_WAKE, TASK_MIGRATE, TASK_UPDATE - demand: task demand computed based on selected policy (recent, max, or average) (ns) - hist: last 5 windows of history for the task with the most recent window listed first - cpu: CPU the task is associated with - nr_big: number of big tasks on the CPU *** 8.6 sched_reset_all_windows_stats Logged when key parameters controlling window-based statistics collection are changed. This event signifies that all window-based statistics for tasks and cpus are being reset. Changes to below attributes result in such a reset: * sched_ravg_window (See Sec 2) * sched_window_stats_policy (See Sec 2.4) * sched_ravg_hist_size (See Sec 7.11) -0 [004] d.h4 12700.711489: sched_reset_all_windows_stats: time_taken 1123 window_start 0 window_size 0 reason POLICY_CHANGE old_val 0 new_val 1 - time_taken: time taken for the reset function to complete (ns) - window_start: Beginning of first window following change to window size (ns) - window_size: Size of window. Non-zero if window-size is changing (in ticks) - reason: Reason for reset of statistics. - old_val: Old value of variable, change of which is triggering reset - new_val: New value of variable, change of which is triggering reset *** 8.7 sched_migration_update_sum Logged when a task is migrating to another cpu. -0 [000] d..8 5020.404137: sched_migration_update_sum: cpu 0: cs 471278 ps 902463 nt_cs 0 nt_ps 0 pid 2645 - cpu: cpu, away from which or to which, task is migrating - cs: curr_runnable_sum of cpu (ns). See Sec 6.1 for more details of this counter. - ps: prev_runnable_sum of cpu (ns). See Sec 6.1 for more details of this counter. - nt_cs: nt_curr_runnable_sum of cpu (ns). See Sec 6.1 for more details of this counter. - nt_ps: nt_prev_runnable_sum of cpu (ns). See Sec 6.1 for more details of this counter - pid: PID of migrating task *** 8.8 sched_get_busy Logged when scheduler is returning busy time statistics for a cpu. <...>-4331 [003] d.s3 313.700108: sched_get_busy: cpu 3 load 19076 new_task_load 0 early 0 - cpu: cpu, for which busy time statistic (prev_runnable_sum) is being returned (ns) - load: corresponds to prev_runnable_sum (ns), scaled to fmax of cpu - new_task_load: corresponds to nt_prev_runnable_sum to fmax of cpu - early: A flag indicating whether the scheduler is passing regular load or early detection load 0 - regular load 1 - early detection load *** 8.9 sched_freq_alert Logged when scheduler is alerting cpufreq governor about need to change frequency -0 [004] d.h4 12700.711489: sched_freq_alert: cpu 0 old_load=XXX new_load=YYY - cpu: cpu in cluster that has highest load (prev_runnable_sum) - old_load: cpu busy time last reported to governor. This is load scaled in reference to max_possible_freq and max_possible_efficiency. - new_load: recent cpu busy time. This is load scaled in reference to max_possible_freq and max_possible_efficiency. *** 8.10 sched_set_boost Logged when boost settings are being changed -0 [004] d.h4 12700.711489: sched_set_boost: ref_count=1 - ref_count: A non-zero value indicates boost is in effect ======================== 9. Device Tree bindings ======================== The device tree bindings for the HMP scheduler are defined in Documentation/devicetree/bindings/sched/sched_hmp.txt