path: root/kernel/sched/psi.c

/*
 * Pressure stall information for CPU, memory and IO
 *
 * Copyright (c) 2018 Facebook, Inc.
 * Author: Johannes Weiner <hannes@cmpxchg.org>
 *
 * 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 <linux/sched/loadavg.h>
#include <linux/seq_file.h>
#include <linux/proc_fs.h>
#include <linux/seqlock.h>
#include <linux/cgroup.h>
#include <linux/module.h>
#include <linux/sched.h>
#include <linux/psi.h>
#include "sched.h"

static int psi_bug __read_mostly;

DEFINE_STATIC_KEY_FALSE(psi_disabled);

#ifdef CONFIG_PSI_DEFAULT_DISABLED
bool psi_enable;
#else
bool psi_enable = true;
#endif
static 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) */

/* 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);
static struct psi_group psi_system = {
	.pcpu = &system_group_pcpu,
};

static void psi_update_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->next_update = sched_clock() + psi_period;
	INIT_DELAYED_WORK(&group->clock_work, psi_update_work);
	mutex_init(&group->stat_lock);
}

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] > 1;
	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, u32 *times)
{
	struct psi_group_cpu *groupc = per_cpu_ptr(group->pcpu, cpu);
	unsigned int tasks[NR_PSI_TASK_COUNTS];
	u64 now, state_start;
	unsigned int seq;
	int s;

	/* 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));
		memcpy(tasks, groupc->tasks, sizeof(groupc->tasks));
		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 (test_state(tasks, s))
			times[s] += now - state_start;

		delta = times[s] - groupc->times_prev[s];
		groupc->times_prev[s] = times[s];

		times[s] = delta;
	}
}

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_u6