2 * latencytop.c: Latency display infrastructure
4 * (C) Copyright 2008 Intel Corporation
5 * Author: Arjan van de Ven <arjan@linux.intel.com>
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public License
9 * as published by the Free Software Foundation; version 2
14 * CONFIG_LATENCYTOP enables a kernel latency tracking infrastructure that is
15 * used by the "latencytop" userspace tool. The latency that is tracked is not
16 * the 'traditional' interrupt latency (which is primarily caused by something
17 * else consuming CPU), but instead, it is the latency an application encounters
18 * because the kernel sleeps on its behalf for various reasons.
20 * This code tracks 2 levels of statistics:
21 * 1) System level latency
22 * 2) Per process latency
24 * The latency is stored in fixed sized data structures in an accumulated form;
25 * if the "same" latency cause is hit twice, this will be tracked as one entry
26 * in the data structure. Both the count, total accumulated latency and maximum
27 * latency are tracked in this data structure. When the fixed size structure is
28 * full, no new causes are tracked until the buffer is flushed by writing to
29 * the /proc file; the userspace tool does this on a regular basis.
31 * A latency cause is identified by a stringified backtrace at the point that
32 * the scheduler gets invoked. The userland tool will use this string to
33 * identify the cause of the latency in human readable form.
35 * The information is exported via /proc/latency_stats and /proc/<pid>/latency.
36 * These files look like this:
38 * Latency Top version : v0.1
39 * 70 59433 4897 i915_irq_wait drm_ioctl vfs_ioctl do_vfs_ioctl sys_ioctl
41 * | | | +----> the stringified backtrace
42 * | | +---------> The maximum latency for this entry in microseconds
43 * | +--------------> The accumulated latency for this entry (microseconds)
44 * +-------------------> The number of times this entry is hit
46 * (note: the average latency is the accumulated latency divided by the number
50 #include <linux/kallsyms.h>
51 #include <linux/seq_file.h>
52 #include <linux/notifier.h>
53 #include <linux/spinlock.h>
54 #include <linux/proc_fs.h>
55 #include <linux/latencytop.h>
56 #include <linux/export.h>
57 #include <linux/sched.h>
58 #include <linux/sched/debug.h>
59 #include <linux/list.h>
60 #include <linux/stacktrace.h>
62 static DEFINE_RAW_SPINLOCK(latency_lock);
65 static struct latency_record latency_record[MAXLR];
67 int latencytop_enabled;
69 void clear_all_latency_tracing(struct task_struct *p)
73 if (!latencytop_enabled)
76 raw_spin_lock_irqsave(&latency_lock, flags);
77 memset(&p->latency_record, 0, sizeof(p->latency_record));
78 p->latency_record_count = 0;
79 raw_spin_unlock_irqrestore(&latency_lock, flags);
82 static void clear_global_latency_tracing(void)
86 raw_spin_lock_irqsave(&latency_lock, flags);
87 memset(&latency_record, 0, sizeof(latency_record));
88 raw_spin_unlock_irqrestore(&latency_lock, flags);
92 account_global_scheduler_latency(struct task_struct *tsk,
93 struct latency_record *lat)
95 int firstnonnull = MAXLR + 1;
98 if (!latencytop_enabled)
101 /* skip kernel threads for now */
105 for (i = 0; i < MAXLR; i++) {
108 /* Nothing stored: */
109 if (!latency_record[i].backtrace[0]) {
110 if (firstnonnull > i)
114 for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
115 unsigned long record = lat->backtrace[q];
117 if (latency_record[i].backtrace[q] != record) {
122 /* 0 and ULONG_MAX entries mean end of backtrace: */
123 if (record == 0 || record == ULONG_MAX)
127 latency_record[i].count++;
128 latency_record[i].time += lat->time;
129 if (lat->time > latency_record[i].max)
130 latency_record[i].max = lat->time;
139 /* Allocted a new one: */
140 memcpy(&latency_record[i], lat, sizeof(struct latency_record));
144 * Iterator to store a backtrace into a latency record entry
146 static inline void store_stacktrace(struct task_struct *tsk,
147 struct latency_record *lat)
149 struct stack_trace trace;
151 memset(&trace, 0, sizeof(trace));
152 trace.max_entries = LT_BACKTRACEDEPTH;
153 trace.entries = &lat->backtrace[0];
154 save_stack_trace_tsk(tsk, &trace);
158 * __account_scheduler_latency - record an occurred latency
159 * @tsk - the task struct of the task hitting the latency
160 * @usecs - the duration of the latency in microseconds
161 * @inter - 1 if the sleep was interruptible, 0 if uninterruptible
163 * This function is the main entry point for recording latency entries
164 * as called by the scheduler.
166 * This function has a few special cases to deal with normal 'non-latency'
167 * sleeps: specifically, interruptible sleep longer than 5 msec is skipped
168 * since this usually is caused by waiting for events via select() and co.
170 * Negative latencies (caused by time going backwards) are also explicitly
174 __account_scheduler_latency(struct task_struct *tsk, int usecs, int inter)
178 struct latency_record lat;
180 /* Long interruptible waits are generally user requested... */
181 if (inter && usecs > 5000)
184 /* Negative sleeps are time going backwards */
185 /* Zero-time sleeps are non-interesting */
189 memset(&lat, 0, sizeof(lat));
193 store_stacktrace(tsk, &lat);
195 raw_spin_lock_irqsave(&latency_lock, flags);
197 account_global_scheduler_latency(tsk, &lat);
199 for (i = 0; i < tsk->latency_record_count; i++) {
200 struct latency_record *mylat;
203 mylat = &tsk->latency_record[i];
204 for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
205 unsigned long record = lat.backtrace[q];
207 if (mylat->backtrace[q] != record) {
212 /* 0 and ULONG_MAX entries mean end of backtrace: */
213 if (record == 0 || record == ULONG_MAX)
218 mylat->time += lat.time;
219 if (lat.time > mylat->max)
220 mylat->max = lat.time;
226 * short term hack; if we're > 32 we stop; future we recycle:
228 if (tsk->latency_record_count >= LT_SAVECOUNT)
231 /* Allocated a new one: */
232 i = tsk->latency_record_count++;
233 memcpy(&tsk->latency_record[i], &lat, sizeof(struct latency_record));
236 raw_spin_unlock_irqrestore(&latency_lock, flags);
239 static int lstats_show(struct seq_file *m, void *v)
243 seq_puts(m, "Latency Top version : v0.1\n");
245 for (i = 0; i < MAXLR; i++) {
246 struct latency_record *lr = &latency_record[i];
248 if (lr->backtrace[0]) {
250 seq_printf(m, "%i %lu %lu",
251 lr->count, lr->time, lr->max);
252 for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
253 unsigned long bt = lr->backtrace[q];
258 seq_printf(m, " %ps", (void *)bt);
267 lstats_write(struct file *file, const char __user *buf, size_t count,
270 clear_global_latency_tracing();
275 static int lstats_open(struct inode *inode, struct file *filp)
277 return single_open(filp, lstats_show, NULL);
280 static const struct file_operations lstats_fops = {
283 .write = lstats_write,
285 .release = single_release,
288 static int __init init_lstats_procfs(void)
290 proc_create("latency_stats", 0644, NULL, &lstats_fops);
294 int sysctl_latencytop(struct ctl_table *table, int write,
295 void __user *buffer, size_t *lenp, loff_t *ppos)
299 err = proc_dointvec(table, write, buffer, lenp, ppos);
300 if (latencytop_enabled)
301 force_schedstat_enabled();
305 device_initcall(init_lstats_procfs);