1 // SPDX-License-Identifier: GPL-2.0-only
3 * latencytop.c: Latency display infrastructure
5 * (C) Copyright 2008 Intel Corporation
6 * Author: Arjan van de Ven <arjan@linux.intel.com>
10 * CONFIG_LATENCYTOP enables a kernel latency tracking infrastructure that is
11 * used by the "latencytop" userspace tool. The latency that is tracked is not
12 * the 'traditional' interrupt latency (which is primarily caused by something
13 * else consuming CPU), but instead, it is the latency an application encounters
14 * because the kernel sleeps on its behalf for various reasons.
16 * This code tracks 2 levels of statistics:
17 * 1) System level latency
18 * 2) Per process latency
20 * The latency is stored in fixed sized data structures in an accumulated form;
21 * if the "same" latency cause is hit twice, this will be tracked as one entry
22 * in the data structure. Both the count, total accumulated latency and maximum
23 * latency are tracked in this data structure. When the fixed size structure is
24 * full, no new causes are tracked until the buffer is flushed by writing to
25 * the /proc file; the userspace tool does this on a regular basis.
27 * A latency cause is identified by a stringified backtrace at the point that
28 * the scheduler gets invoked. The userland tool will use this string to
29 * identify the cause of the latency in human readable form.
31 * The information is exported via /proc/latency_stats and /proc/<pid>/latency.
32 * These files look like this:
34 * Latency Top version : v0.1
35 * 70 59433 4897 i915_irq_wait drm_ioctl vfs_ioctl do_vfs_ioctl sys_ioctl
37 * | | | +----> the stringified backtrace
38 * | | +---------> The maximum latency for this entry in microseconds
39 * | +--------------> The accumulated latency for this entry (microseconds)
40 * +-------------------> The number of times this entry is hit
42 * (note: the average latency is the accumulated latency divided by the number
46 #include <linux/kallsyms.h>
47 #include <linux/seq_file.h>
48 #include <linux/notifier.h>
49 #include <linux/spinlock.h>
50 #include <linux/proc_fs.h>
51 #include <linux/latencytop.h>
52 #include <linux/export.h>
53 #include <linux/sched.h>
54 #include <linux/sched/debug.h>
55 #include <linux/sched/stat.h>
56 #include <linux/list.h>
57 #include <linux/stacktrace.h>
58 #include <linux/sysctl.h>
60 static DEFINE_RAW_SPINLOCK(latency_lock);
63 static struct latency_record latency_record[MAXLR];
65 int latencytop_enabled;
68 static int sysctl_latencytop(struct ctl_table *table, int write, void *buffer,
69 size_t *lenp, loff_t *ppos)
73 err = proc_dointvec(table, write, buffer, lenp, ppos);
74 if (latencytop_enabled)
75 force_schedstat_enabled();
80 static struct ctl_table latencytop_sysctl[] = {
82 .procname = "latencytop",
83 .data = &latencytop_enabled,
84 .maxlen = sizeof(int),
86 .proc_handler = sysctl_latencytop,
92 void clear_tsk_latency_tracing(struct task_struct *p)
96 raw_spin_lock_irqsave(&latency_lock, flags);
97 memset(&p->latency_record, 0, sizeof(p->latency_record));
98 p->latency_record_count = 0;
99 raw_spin_unlock_irqrestore(&latency_lock, flags);
102 static void clear_global_latency_tracing(void)
106 raw_spin_lock_irqsave(&latency_lock, flags);
107 memset(&latency_record, 0, sizeof(latency_record));
108 raw_spin_unlock_irqrestore(&latency_lock, flags);
112 account_global_scheduler_latency(struct task_struct *tsk,
113 struct latency_record *lat)
115 int firstnonnull = MAXLR + 1;
118 /* skip kernel threads for now */
122 for (i = 0; i < MAXLR; i++) {
125 /* Nothing stored: */
126 if (!latency_record[i].backtrace[0]) {
127 if (firstnonnull > i)
131 for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
132 unsigned long record = lat->backtrace[q];
134 if (latency_record[i].backtrace[q] != record) {
139 /* 0 entry marks end of backtrace: */
144 latency_record[i].count++;
145 latency_record[i].time += lat->time;
146 if (lat->time > latency_record[i].max)
147 latency_record[i].max = lat->time;
156 /* Allocted a new one: */
157 memcpy(&latency_record[i], lat, sizeof(struct latency_record));
161 * __account_scheduler_latency - record an occurred latency
162 * @tsk - the task struct of the task hitting the latency
163 * @usecs - the duration of the latency in microseconds
164 * @inter - 1 if the sleep was interruptible, 0 if uninterruptible
166 * This function is the main entry point for recording latency entries
167 * as called by the scheduler.
169 * This function has a few special cases to deal with normal 'non-latency'
170 * sleeps: specifically, interruptible sleep longer than 5 msec is skipped
171 * since this usually is caused by waiting for events via select() and co.
173 * Negative latencies (caused by time going backwards) are also explicitly
177 __account_scheduler_latency(struct task_struct *tsk, int usecs, int inter)
181 struct latency_record lat;
183 /* Long interruptible waits are generally user requested... */
184 if (inter && usecs > 5000)
187 /* Negative sleeps are time going backwards */
188 /* Zero-time sleeps are non-interesting */
192 memset(&lat, 0, sizeof(lat));
197 stack_trace_save_tsk(tsk, lat.backtrace, LT_BACKTRACEDEPTH, 0);
199 raw_spin_lock_irqsave(&latency_lock, flags);
201 account_global_scheduler_latency(tsk, &lat);
203 for (i = 0; i < tsk->latency_record_count; i++) {
204 struct latency_record *mylat;
207 mylat = &tsk->latency_record[i];
208 for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
209 unsigned long record = lat.backtrace[q];
211 if (mylat->backtrace[q] != record) {
216 /* 0 entry is end of backtrace */
222 mylat->time += lat.time;
223 if (lat.time > mylat->max)
224 mylat->max = lat.time;
230 * short term hack; if we're > 32 we stop; future we recycle:
232 if (tsk->latency_record_count >= LT_SAVECOUNT)
235 /* Allocated a new one: */
236 i = tsk->latency_record_count++;
237 memcpy(&tsk->latency_record[i], &lat, sizeof(struct latency_record));
240 raw_spin_unlock_irqrestore(&latency_lock, flags);
243 static int lstats_show(struct seq_file *m, void *v)
247 seq_puts(m, "Latency Top version : v0.1\n");
249 for (i = 0; i < MAXLR; i++) {
250 struct latency_record *lr = &latency_record[i];
252 if (lr->backtrace[0]) {
254 seq_printf(m, "%i %lu %lu",
255 lr->count, lr->time, lr->max);
256 for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
257 unsigned long bt = lr->backtrace[q];
262 seq_printf(m, " %ps", (void *)bt);
271 lstats_write(struct file *file, const char __user *buf, size_t count,
274 clear_global_latency_tracing();
279 static int lstats_open(struct inode *inode, struct file *filp)
281 return single_open(filp, lstats_show, NULL);
284 static const struct proc_ops lstats_proc_ops = {
285 .proc_open = lstats_open,
286 .proc_read = seq_read,
287 .proc_write = lstats_write,
288 .proc_lseek = seq_lseek,
289 .proc_release = single_release,
292 static int __init init_lstats_procfs(void)
294 proc_create("latency_stats", 0644, NULL, &lstats_proc_ops);
296 register_sysctl_init("kernel", latencytop_sysctl);
300 device_initcall(init_lstats_procfs);