/* * Pressure stall information for CPU, memory and IO * * Copyright (c) 2018 Facebook, Inc. * Author: Johannes Weiner * * 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 #include #include #include #include #include #include #include #include "sched.h" static int psi_bug __read_mostly; bool psi_disabled __read_mostly; core_param(psi_disabled, psi_disabled, bool, 0644); /* 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_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_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 bool update_stats(struct psi_group *group) { u64 deltas[NR_PSI_STATES - 1] = { 0, }; unsigned long missed_periods = 0; unsigned long nonidle_total = 0; u64 now, expires, period; int cpu; int s; mutex_lock(&group->stat_lock); /* * 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; get_recent_times(group, cpu, times); 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[s] += div_u64(deltas[s], max(nonidle_total, 1UL)); /* avgX= */ now = sched_clock(); expires = group->next_update; if (now < expires) goto out; 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. */ group->next_update = expires + ((1 + missed_periods) * psi_period); period = now - (group->last_update + (missed_periods * psi_period)); group->last_update = now; for (s = 0; s < NR_PSI_STATES - 1; s++) { u32 sample; sample = group->total[s] - group->total_prev[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->total_prev[s] += sample; calc_avgs(group->avg[s], missed_periods, sample, period); } out: mutex_unlock(&group->stat_lock); return nonidle_total; } static void psi_update_work(struct work_struct *work) { struct delayed_work *dwork; struct psi_group *group; bool nonidle; dwork = to_delayed_work(work); group = container_of(dwork, struct psi_group, clock_work); /* * 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. */ nonidle = update_stats(group); if (nonidle) { unsigned long delay = 0; u64 now; now = sched_clock(); if (group->next_update > now) delay = nsecs_to_jiffies(group->next_update - now) + 1; schedule_delayed_work(dwork, delay); } } static void record_times(struct psi_group_cpu *groupc, int cpu, bool memstall_tick) { u32 delta; u64 now; now = cpu_clock(cpu); delta = now - groupc->state_start; groupc->state_start = now; if (test_state(groupc->tasks, PSI_IO_SOME)) { groupc->times[PSI_IO_SOME] += delta; if (test_state(groupc->tasks, PSI_IO_FULL)) groupc->times[PSI_IO_FULL] += delta; } if (test_state(groupc->tasks, PSI_MEM_SOME)) { groupc->times[PSI_MEM_SOME] += delta; if (test_state(groupc->tasks, PSI_MEM_FULL)) groupc->times[PSI_MEM_FULL] += delta; else if (memstall_tick) { u32 sample; /* * Since we care about lost potential, a * memstall is FULL when there are no other * working tasks, but also when the CPU is * actively reclaiming and nothing productive * could run even if it were runnable. * * When the timer tick sees a reclaiming CPU, * regardless of runnable tasks, sample a FULL * tick (or less if it hasn't been a full tick * since the last state change). */ sample = min(delta, (u32)jiffies_to_nsecs(1)); groupc->times[PSI_MEM_FULL] += sample; } } if (test_state(groupc->tasks, PSI_CPU_SOME)) groupc->times[PSI_CPU_SOME] += delta; if (test_state(groupc->tasks, PSI_NONIDLE)) groupc->times[PSI_NONIDLE] += delta; } static void psi_group_change(struct psi_group *group, int cpu, unsigned int clear, unsigned int set) { struct psi_group_cpu *groupc; unsigned int t, m; groupc = per_cpu_ptr(group->pcpu, cpu); /* * First we assess the aggregate resource states this CPU's * tasks have been in since the last change, and account any * SOME and FULL time these may have resulted in. * * Then we update the task counts according to the state * change requested through the @clear and @set bits. */ write_seqcount_begin(&groupc->seq); record_times(groupc, cpu, false); for (t = 0, m = clear; m; m &= ~(1 << t), t++) { if (!(m & (1 << t))) continue; if (groupc->tasks[t] == 0 && !psi_bug) { printk_deferred(KERN_ERR "psi: task underflow! cpu=%d t=%d tasks=[%u %u %u] clear=%x set=%x\n", cpu, t, groupc->tasks[0], groupc->tasks[1], groupc->tasks[2], clear, set); psi_bug = 1; } groupc->tasks[t]--; } for (t = 0; set; set &= ~(1 << t), t++) if (set & (1 << t)) groupc->tasks[t]++; write_seqcount_end(&groupc->seq); if (!delayed_work_pending(&group->clock_work)) schedule_delayed_work(&group->clock_work, PSI_FREQ); } static struct psi_group *iterate_groups(struct task_struct *task, void **iter) { #ifdef CONFIG_CGROUPS struct cgroup *cgroup = NULL; if (!*iter) cgroup = task->cgroups->dfl_cgrp; else if (*iter == &psi_system) return NULL; else cgroup = cgroup_parent(*iter); if (cgroup && cgroup_parent(cgroup)) { *iter = cgroup; return cgroup_psi(cgroup); } #else if (*iter) return NULL; #endif *iter = &psi_system; return &psi_system; } void psi_task_change(struct task_struct *task, int clear, int set) { int cpu = task_cpu(task); struct psi_group *group; void *iter = NULL; if (!task->pid) return; if (((task->psi_flags & set) || (task->psi_flags & clear) != clear) && !psi_bug) { printk_deferred(KERN_ERR "psi: inconsistent task state! task=%d:%s cpu=%d psi_flags=%x clear=%x set=%x\n", task->pid, task->comm, cpu, task->psi_flags, clear, set); psi_bug = 1; } task->psi_flags &= ~clear; task->psi_flags |= set; while ((group = iterate_groups(task, &iter))) psi_group_change(group, cpu, clear, set); } void psi_memstall_tick(struct task_struct *task, int cpu) { struct psi_group *group; void *iter = NULL; while ((group = iterate_groups(task, &iter))) { struct psi_group_cpu *groupc; groupc = per_cpu_ptr(group->pcpu, cpu); write_seqcount_begin(&groupc->seq); record_times(groupc, cpu, true); write_seqcount_end(&groupc->seq); } } /** * psi_memstall_enter - mark the beginning of a memory stall section * @flags: flags to handle nested sections * * Marks the calling task as being stalled due to a lack of memory, * such as waiting for a refault or performing reclaim. */ void psi_memstall_enter(unsigned long *flags) { struct rq_flags rf; struct rq *rq; if (psi_disabled) return; *flags = current->flags & PF_MEMSTALL; if (*flags) return; /* * PF_MEMSTALL setting & accounting needs to be atomic wrt * changes to the task's scheduling state, otherwise we can * race with CPU migration. */ rq = this_rq_lock_irq(&rf); current->flags |= PF_MEMSTALL; psi_task_change(current, 0, TSK_MEMSTALL); rq_unlock_irq(rq, &rf); } /** * psi_memstall_leave - mark the end of an memory stall section * @flags: flags to handle nested memdelay sections * * Marks the calling task as no longer stalled due to lack of memory. */ void psi_memstall_leave(unsigned long *flags) { struct rq_flags rf; struct rq *rq; if (psi_disabled) return; if (*flags) return; /* * PF_MEMSTALL clearing & accounting needs to be atomic wrt * changes to the task's scheduling state, otherwise we could * race with CPU migration. */ rq = this_rq_lock_irq(&rf); current->flags &= ~PF_MEMSTALL; psi_task_change(current, TSK_MEMSTALL, 0); rq_unlock_irq(rq, &rf); } #ifdef CONFIG_CGROUPS int psi_cgroup_alloc(struct cgroup *cgroup) { if (psi_disabled) return 0; cgroup->psi.pcpu = alloc_percpu(struct psi_group_cpu); if (!cgroup->psi.pcpu) return -ENOMEM; group_init(&cgroup->psi); return 0; } void psi_cgroup_free(struct cgroup *cgroup) { if (psi_disabled) return; cancel_delayed_work_sync(&cgroup->psi.clock_work); free_percpu(cgroup->psi.pcpu); } /** * cgroup_move_task - move task to a different cgroup * @task: the task * @to: the target css_set * * Move task to a new cgroup and safely migrate its associated stall * state between the different groups. * * This function acquires the task's rq lock to lock out concurrent * changes to the task's scheduling state and - in case the task is * running - concurrent changes to its stall state. */ void cgroup_move_task(struct task_struct *task, struct css_set *to) { bool move_psi = !psi_disabled; unsigned int task_flags = 0; struct rq_flags rf; struct rq *rq; if (move_psi) { rq = task_rq_lock(task, &rf); if (task_on_rq_queued(task)) task_flags = TSK_RUNNING; else if (task->in_iowait) task_flags = TSK_IOWAIT; if (task->flags & PF_MEMSTALL) task_flags |= TSK_MEMSTALL; if (task_flags) psi_task_change(task, task_flags, 0); } /* * Lame to do this here, but the scheduler cannot be locked * from the outside, so we move cgroups from inside sched/. */ rcu_assign_pointer(task->cgroups, to); if (move_psi) { if (task_flags) psi_task_change(task, 0, task_flags); task_rq_unlock(rq, task, &rf); } } #endif /* CONFIG_CGROUPS */ int psi_show(struct seq_file *m, struct psi_group *group, enum psi_res res) { int full; if (psi_disabled) return -EOPNOTSUPP; update_stats(group); for (full = 0; full < 2 - (res == PSI_CPU); full++) { unsigned long avg[3]; u64 total; int w; for (w = 0; w < 3; w++) avg[w] = group->avg[res * 2 + full][w]; total = div_u64(group->total[res * 2 + full], NSEC_PER_USEC); seq_printf(m, "%s avg10=%lu.%02lu avg60=%lu.%02lu avg300=%lu.%02lu total=%llu\n", full ? "full" : "some", LOAD_INT(avg[0]), LOAD_FRAC(avg[0]), LOAD_INT(avg[1]), LOAD_FRAC(avg[1]), LOAD_INT(avg[2]), LOAD_FRAC(avg[2]), total); } return 0; } static int psi_io_show(struct seq_file *m, void *v) { return psi_show(m, &psi_system, PSI_IO); } static int psi_memory_show(struct seq_file *m, void *v) { return psi_show(m, &psi_system, PSI_MEM); } static int psi_cpu_show(struct seq_file *m, void *v) { return psi_show(m, &psi_system, PSI_CPU); } static int psi_io_open(struct inode *inode, struct file *file) { return single_open(file, psi_io_show, NULL); } static int psi_memory_open(struct inode *inode, struct file *file) { return single_open(file, psi_memory_show, NULL); } static int psi_cpu_open(struct inode *inode, struct file *file) { return single_open(file, psi_cpu_show, NULL); } static const struct file_operations psi_io_fops = { .open = psi_io_open, .read = seq_read, .llseek = seq_lseek, .release = single_release, }; static const struct file_operations psi_memory_fops = { .open = psi_memory_open, .read = seq_read, .llseek = seq_lseek, .release = single_release, }; static const struct file_operations psi_cpu_fops = { .open = psi_cpu_open, .read = seq_read, .llseek = seq_lseek, .release = single_release, }; static int __init psi_proc_init(void) { proc_mkdir("pressure", NULL); proc_create("pressure/io", 0, NULL, &psi_io_fops); proc_create("pressure/memory", 0, NULL, &psi_memory_fops); proc_create("pressure/cpu", 0, NULL, &psi_cpu_fops); return 0; } module_init(psi_proc_init);