/* * Pressure stall information for CPU, memory and IO * * Copyright (c) 2018 Facebook, Inc. * Author: Johannes Weiner <hannes@cmpxchg.org> * * Polling support by Suren Baghdasaryan <surenb@google.com> * Copyright (c) 2018 Google, Inc. * * When CPU, memory and IO are contended, tasks experience delays that * reduce throughput and introduce latencies into the workload. Memory * and IO contention, in addition, can cause a full loss of forward * progress in which the CPU goes idle. * * This code aggregates individual task delays into resource pressure * metrics that indicate problems with both workload health and * resource utilization. * *			Model * * The time in which a task can execute on a CPU is our baseline for * productivity. Pressure expresses the amount of time in which this * potential cannot be realized due to resource contention. * * This concept of productivity has two components: the workload and * the CPU. To measure the impact of pressure on both, we define two * contention states for a resource: SOME and FULL. * * In the SOME state of a given resource, one or more tasks are * delayed on that resource. This affects the workload's ability to * perform work, but the CPU may still be executing other tasks. * * In the FULL state of a given resource, all non-idle tasks are * delayed on that resource such that nobody is advancing and the CPU * goes idle. This leaves both workload and CPU unproductive. * * (Naturally, the FULL state doesn't exist for the CPU resource.) * *	SOME = nr_delayed_tasks != 0 *	FULL = nr_delayed_tasks != 0 && nr_running_tasks == 0 * * The percentage of wallclock time spent in those compound stall * states gives pressure numbers between 0 and 100 for each resource, * where the SOME percentage indicates workload slowdowns and the FULL * percentage indicates reduced CPU utilization: * *	%SOME = time(SOME) / period *	%FULL = time(FULL) / period * *			Multiple CPUs * * The more tasks and available CPUs there are, the more work can be * performed concurrently. This means that the potential that can go * unrealized due to resource contention *also* scales with non-idle * tasks and CPUs. * * Consider a scenario where 257 number crunching tasks are trying to * run concurrently on 256 CPUs. If we simply aggregated the task * states, we would have to conclude a CPU SOME pressure number of * 100%, since *somebody* is waiting on a runqueue at all * times. However, that is clearly not the amount of contention the * workload is experiencing: only one out of 256 possible exceution * threads will be contended at any given time, or about 0.4%. * * Conversely, consider a scenario of 4 tasks and 4 CPUs where at any * given time *one* of the tasks is delayed due to a lack of memory. * Again, looking purely at the task state would yield a memory FULL * pressure number of 0%, since *somebody* is always making forward * progress. But again this wouldn't capture the amount of execution * potential lost, which is 1 out of 4 CPUs, or 25%. * * To calculate wasted potential (pressure) with multiple processors, * we have to base our calculation on the number of non-idle tasks in * conjunction with the number of available CPUs, which is the number * of potential execution threads. SOME becomes then the proportion of * delayed tasks to possibe threads, and FULL is the share of possible * threads that are unproductive due to delays: * *	threads = min(nr_nonidle_tasks, nr_cpus) *	   SOME = min(nr_delayed_tasks / threads, 1) *	   FULL = (threads - min(nr_running_tasks, threads)) / threads * * For the 257 number crunchers on 256 CPUs, this yields: * *	threads = min(257, 256) *	   SOME = min(1 / 256, 1)             = 0.4% *	   FULL = (256 - min(257, 256)) / 256 = 0% * * For the 1 out of 4 memory-delayed tasks, this yields: * *	threads = min(4, 4) *	   SOME = min(1 / 4, 1)               = 25% *	   FULL = (4 - min(3, 4)) / 4         = 25% * * [ Substitute nr_cpus with 1, and you can see that it's a natural *   extension of the single-CPU model. ] * *			Implementation * * To assess the precise time spent in each such state, we would have * to freeze the system on task changes and start/stop the state * clocks accordingly. Obviously that doesn't scale in practice. * * Because the scheduler aims to distribute the compute load evenly * among the available CPUs, we can track task state locally to each * CPU and, at much lower frequency, extrapolate the global state for * the cumulative stall times and the running averages. * * For each runqueue, we track: * *	   tSOME[cpu] = time(nr_delayed_tasks[cpu] != 0) *	   tFULL[cpu] = time(nr_delayed_tasks[cpu] && !nr_running_tasks[cpu]) *	tNONIDLE[cpu] = time(nr_nonidle_tasks[cpu] != 0) * * and then periodically aggregate: * *	tNONIDLE = sum(tNONIDLE[i]) * *	   tSOME = sum(tSOME[i] * tNONIDLE[i]) / tNONIDLE *	   tFULL = sum(tFULL[i] * tNONIDLE[i]) / tNONIDLE * *	   %SOME = tSOME / period *	   %FULL = tFULL / period * * This gives us an approximation of pressure that is practical * cost-wise, yet way more sensitive and accurate than periodic * sampling of the aggregate task states would be. */ #include "../workqueue_internal.h"#include <linux/sched/loadavg.h>#include <linux/seq_file.h>#include <linux/proc_fs.h>#include <linux/seqlock.h>#include <linux/uaccess.h>#include <linux/cgroup.h>#include <linux/module.h>#include <linux/sched.h>#include <linux/ctype.h>#include <linux/file.h>#include <linux/poll.h>#include <linux/psi.h>#include "sched.h" static int psi_bug __read_mostly; DEFINE_STATIC_KEY_FALSE(psi_disabled); #ifdef CONFIG_PSI_DEFAULT_DISABLEDstatic bool psi_enable;#elsestatic bool psi_enable = true;#endifstatic int __init setup_psi(char *str){	return kstrtobool(str, &psi_enable) == 0;}__setup("psi=", setup_psi); /* Running averages - we need to be higher-res than loadavg */#define PSI_FREQ	(2*HZ+1)	/* 2 sec intervals */#define EXP_10s		1677		/* 1/exp(2s/10s) as fixed-point */#define EXP_60s		1981		/* 1/exp(2s/60s) */#define EXP_300s	2034		/* 1/exp(2s/300s) */ /* PSI trigger definitions */#define WINDOW_MIN_US 500000	/* Min window size is 500ms */#define WINDOW_MAX_US 10000000	/* Max window size is 10s */#define UPDATES_PER_WINDOW 10	/* 10 updates per window */ /* Sampling frequency in nanoseconds */static u64 psi_period __read_mostly; /* System-level pressure and stall tracking */static DEFINE_PER_CPU(struct psi_group_cpu, system_group_pcpu);struct psi_group psi_system = {	.pcpu = &system_group_pcpu,}; static void psi_avgs_work(struct work_struct *work); static void group_init(struct psi_group *group){	int cpu; 	for_each_possible_cpu(cpu)		seqcount_init(&per_cpu_ptr(group->pcpu, cpu)->seq);	group->avg_last_update = sched_clock();	group->avg_next_update = group->avg_last_update + psi_period;	INIT_DELAYED_WORK(&group->avgs_work, psi_avgs_work);	mutex_init(&group->avgs_lock);	/* Init trigger-related members */	mutex_init(&group->trigger_lock);	INIT_LIST_HEAD(&group->triggers);	memset(group->nr_triggers, 0, sizeof(group->nr_triggers));	group->poll_states = 0;	group->poll_min_period = U32_MAX;	memset(group->polling_total, 0, sizeof(group->polling_total));	group->polling_next_update = ULLONG_MAX;	group->polling_until = 0;	rcu_assign_pointer(group->poll_task, NULL);} void __init psi_init(void){	if (!psi_enable) {		static_branch_enable(&psi_disabled);		return;	} 	psi_period = jiffies_to_nsecs(PSI_FREQ);	group_init(&psi_system);} static bool test_state(unsigned int *tasks, enum psi_states state){	switch (state) {	case PSI_IO_SOME:		return tasks[NR_IOWAIT];	case PSI_IO_FULL:		return tasks[NR_IOWAIT] && !tasks[NR_RUNNING];	case PSI_MEM_SOME:		return tasks[NR_MEMSTALL];	case PSI_MEM_FULL:		return tasks[NR_MEMSTALL] && !tasks[NR_RUNNING];	case PSI_CPU_SOME:		return tasks[NR_RUNNING] > tasks[NR_ONCPU];	case PSI_NONIDLE:		return tasks[NR_IOWAIT] || tasks[NR_MEMSTALL] ||			tasks[NR_RUNNING];	default:		return false;	}} static void get_recent_times(struct psi_group *group, int cpu,			     enum psi_aggregators aggregator, u32 *times,			     u32 *pchanged_states){	struct psi_group_cpu *groupc = per_cpu_ptr(group->pcpu, cpu);	u64 now, state_start;	enum psi_states s;	unsigned int seq;	u32 state_mask; 	*pchanged_states = 0; 	/* Snapshot a coherent view of the CPU state */	do {		seq = read_seqcount_begin(&groupc->seq);		now = cpu_clock(cpu);		memcpy(times, groupc->times, sizeof(groupc->times));		state_mask = groupc->state_mask;		state_start = groupc->state_start;	} while (read_seqcount_retry(&groupc->seq, seq)); 	/* Calculate state time deltas against the previous snapshot */	for (s = 0; s < NR_PSI_STATES; s++) {		u32 delta;		/*		 * In addition to already concluded states, we also		 * incorporate currently active states on the CPU,		 * since states may last for many sampling periods.		 *		 * This way we keep our delta sampling buckets small		 * (u32) and our reported pressure close to what's		 * actually happening.		 */		if (state_mask & (1 << s))			times[s] += now - state_start; 		delta = times[s] - groupc->times_prev[aggregator][s];		groupc->times_prev[aggregator][s] = times[s]; 		times[s] = delta;		if (delta)			*pchanged_states |= (1 << s);	}} static void calc_avgs(unsigned long avg[3], int missed_periods,		      u64 time, u64 period){	unsigned long pct; 	/* Fill in zeroes for periods of no activity */	if (missed_periods) {		avg[0] = calc_load_n(avg[0], EXP_10s, 0, missed_periods);		avg[1] = calc_load_n(avg[1], EXP_60s, 0, missed_periods);		avg[2] = calc_load_n(avg[2], EXP_300s, 0, missed_periods);	} 	/* Sample the most recent active period */	pct = div_u64(time * 100, period);	pct *= FIXED_1;	avg[0] = calc_load(avg[0], EXP_10s, pct);	avg[1] = calc_load(avg[1], EXP_60s, pct);	avg[2] = calc_load(avg[2], EXP_300s, pct);} static void collect_percpu_times(struct psi_group *group,				 enum psi_aggregators aggregator,				 u32 *pchanged_states){	u64 deltas[NR_PSI_STATES - 1] = { 0, };	unsigned long nonidle_total = 0;	u32 changed_states = 0;	int cpu;	int s; 	/*	 * Collect the per-cpu time buckets and average them into a	 * single time sample that is normalized to wallclock time.	 *	 * For averaging, each CPU is weighted by its non-idle time in	 * the sampling period. This eliminates artifacts from uneven	 * loading, or even entirely idle CPUs.	 */	for_each_possible_cpu(cpu) {		u32 times[NR_PSI_STATES];		u32 nonidle;		u32 cpu_changed_states; 		get_recent_times(group, cpu, aggregator, times,				&cpu_changed_states);		changed_states |= cpu_changed_states; 		nonidle = nsecs_to_jiffies(times[PSI_NONIDLE]);		nonidle_total += nonidle; 		for (s = 0; s < PSI_NONIDLE; s++)			deltas[s] += (u64)times[s] * nonidle;	} 	/*	 * Integrate the sample into the running statistics that are	 * reported to userspace: the cumulative stall times and the	 * decaying averages.	 *	 * Pressure percentages are sampled at PSI_FREQ. We might be	 * called more often when the user polls more frequently than	 * that; we might be called less often when there is no task	 * activity, thus no data, and clock ticks are sporadic. The	 * below handles both.	 */ 	/* total= */	for (s = 0; s < NR_PSI_STATES - 1; s++)		group->total[aggregator][s] +=				div_u64(deltas[s], max(nonidle_total, 1UL)); 	if (pchanged_states)		*pchanged_states = changed_states;} static u64 update_averages(struct psi_group *group, u64 now){	unsigned long missed_periods = 0;	u64 expires, period;	u64 avg_next_update;	int s; 	/* avgX= */	expires = group->avg_next_update;	if (now - expires >= psi_period)		missed_periods = div_u64(now - expires, psi_period); 	/*	 * The periodic clock tick can get delayed for various	 * reasons, especially on loaded systems. To avoid clock	 * drift, we schedule the clock in fixed psi_period intervals.	 * But the deltas we sample out of the per-cpu buckets above	 * are based on the actual time elapsing between clock ticks.	 */	avg_next_update = expires + ((1 + missed_periods) * psi_period);	period = now - (group->avg_last_update + (missed_periods * psi_period));	group->avg_last_update = now; 	for (s = 0; s < NR_PSI_STATES - 1; s++) {		u32 sample; 		sample = group->total[PSI_AVGS][s] - group->avg_total[s];		/*		 * Due to the lockless sampling of the time buckets,		 * recorded time deltas can slip into the next period,		 * which under full pressure can result in samples in		 * excess of the period length.		 *		 * We don't want to report non-sensical pressures in		 * excess of 100%, nor do we want to drop such events		 * on the floor. Instead we punt any overage into the		 * future until pressure subsides. By doing this we		 * don't underreport the occurring pressure curve, we		 * just report it delayed by one period length.		 *		 * The error isn't cumulative. As soon as another		 * delta slips from a period P to P+1, by definition		 * it frees up its time T in P.		 */		if (sample > period)			sample = period;		group->avg_total[s] += sample;		calc_avgs(group->avg[s], missed_periods, sample, period);	} 	return avg_next_update;} static void psi_avgs_work(struct work_struct *work){	struct delayed_work *dwork;	struct psi_group *group;	u32 changed_states;	bool nonidle;	u64 now; 	dwork = to_delayed_work(work);	group = container_of(dwork, struct psi_group, avgs_work); 	mutex_lock(&group->avgs_lock); 	now = sched_clock(); 	collect_percpu_times(group, PSI_AVGS, &changed_states);	nonidle = changed_states & (1 << PSI_NONIDLE);	/*	 * If there is task activity, periodically fold the per-cpu	 * times and feed samples into the running averages. If things	 * are idle and there is no data to process, stop the clock.	 * Once restarted, we'll catch up the running averages in one	 * go - see calc_avgs() and missed_periods.	 */	if (now >= group->avg_next_update)		group->avg_next_update = update_averages(group, now); 	if (nonidle) {		schedule_delayed_work(dwork, nsecs_to_jiffies(				group->avg_next_update - now) + 1);	} 	mutex_unlock(&group->avgs_lock);} /* Trigger tracking window manupulations */static void window_reset(struct psi_window *win, u64 now, u64 value,			 u64 prev_growth){	win->start_time = now;	win->start_value = value;	win->prev_growth = prev_growth;} /* * PSI growth tracking window update and growth calculation routine. * * This approximates a sliding tracking window by interpolating * partially elapsed windows using historical growth data from the * previous intervals. This minimizes memory requirements (by not storing * all the intermediate values in the previous window) and simplifies * the calculations. It works well because PSI signal changes only in * positive direction and over relatively small window sizes the growth * is close to linear. */static u64 window_update(struct psi_window *win, u64 now, u64 value){	u64 elapsed;	u64 growth; 	elapsed = now - win->start_time;	growth = value - win->start_value;	/*	 * After each tracking window passes win->start_value and	 * win->start_time get reset and win->prev_growth stores	 * the average per-window growth of the previous window.	 * win->prev_growth is then used to interpolate additional	 * growth from the previous window assuming it was linear.	 */	if (elapsed > win->size)		window_reset(win, now, value, growth);	else {		u32 remaining; 		remaining = win->size - elapsed;		growth += div64_u64(win->prev_growth * remaining, win->size);	} 	return growth;} static void init_triggers(struct psi_group *group, u64 now){	struct psi_trigger *t; 	list_for_each_entry(t, &group->triggers, node)		window_reset(&t->win, now,				group->total[PSI_POLL][t->state], 0);	memcpy(group->polling_total, group->total[PSI_POLL],		   sizeof(group->polling_total));	group->polling_next_update = now + group->poll_min_period;}

Lines 1–500 of 1355. Load more