1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * latencytop.c: Latency display infrastructure
4 *
5 * (C) Copyright 2008 Intel Corporation
6 * Author: Arjan van de Ven <arjan@linux.intel.com>
7 */
8
9/*
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.
15 *
16 * This code tracks 2 levels of statistics:
17 * 1) System level latency
18 * 2) Per process latency
19 *
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.
26 *
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.
30 *
31 * The information is exported via /proc/latency_stats and /proc/<pid>/latency.
32 * These files look like this:
33 *
34 * Latency Top version : v0.1
35 * 70 59433 4897 i915_irq_wait drm_ioctl vfs_ioctl do_vfs_ioctl sys_ioctl
36 * |    |    |    |
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
41 *
42 * (note: the average latency is the accumulated latency divided by the number
43 * of times)
44 */
45
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>
59
60static DEFINE_RAW_SPINLOCK(latency_lock);
61
62#define MAXLR 128
63static struct latency_record latency_record[MAXLR];
64
65int latencytop_enabled;
66
67#ifdef CONFIG_SYSCTL
68static int sysctl_latencytop(struct ctl_table *table, int write, void *buffer,
69		size_t *lenp, loff_t *ppos)
70{
71	int err;
72
73	err = proc_dointvec(table, write, buffer, lenp, ppos);
74	if (latencytop_enabled)
75		force_schedstat_enabled();
76
77	return err;
78}
79
80static struct ctl_table latencytop_sysctl[] = {
81	{
82		.procname   = "latencytop",
83		.data       = &latencytop_enabled,
84		.maxlen     = sizeof(int),
85		.mode       = 0644,
86		.proc_handler   = sysctl_latencytop,
87	},
88};
89#endif
90
91void clear_tsk_latency_tracing(struct task_struct *p)
92{
93	unsigned long flags;
94
95	raw_spin_lock_irqsave(&latency_lock, flags);
96	memset(&p->latency_record, 0, sizeof(p->latency_record));
97	p->latency_record_count = 0;
98	raw_spin_unlock_irqrestore(&latency_lock, flags);
99}
100
101static void clear_global_latency_tracing(void)
102{
103	unsigned long flags;
104
105	raw_spin_lock_irqsave(&latency_lock, flags);
106	memset(&latency_record, 0, sizeof(latency_record));
107	raw_spin_unlock_irqrestore(&latency_lock, flags);
108}
109
110static void __sched
111account_global_scheduler_latency(struct task_struct *tsk,
112				 struct latency_record *lat)
113{
114	int firstnonnull = MAXLR;
115	int i;
116
117	/* skip kernel threads for now */
118	if (!tsk->mm)
119		return;
120
121	for (i = 0; i < MAXLR; i++) {
122		int q, same = 1;
123
124		/* Nothing stored: */
125		if (!latency_record[i].backtrace[0]) {
126			if (firstnonnull > i)
127				firstnonnull = i;
128			continue;
129		}
130		for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
131			unsigned long record = lat->backtrace[q];
132
133			if (latency_record[i].backtrace[q] != record) {
134				same = 0;
135				break;
136			}
137
138			/* 0 entry marks end of backtrace: */
139			if (!record)
140				break;
141		}
142		if (same) {
143			latency_record[i].count++;
144			latency_record[i].time += lat->time;
145			if (lat->time > latency_record[i].max)
146				latency_record[i].max = lat->time;
147			return;
148		}
149	}
150
151	i = firstnonnull;
152	if (i >= MAXLR)
153		return;
154
155	/* Allocted a new one: */
156	memcpy(&latency_record[i], lat, sizeof(struct latency_record));
157}
158
159/**
160 * __account_scheduler_latency - record an occurred latency
161 * @tsk - the task struct of the task hitting the latency
162 * @usecs - the duration of the latency in microseconds
163 * @inter - 1 if the sleep was interruptible, 0 if uninterruptible
164 *
165 * This function is the main entry point for recording latency entries
166 * as called by the scheduler.
167 *
168 * This function has a few special cases to deal with normal 'non-latency'
169 * sleeps: specifically, interruptible sleep longer than 5 msec is skipped
170 * since this usually is caused by waiting for events via select() and co.
171 *
172 * Negative latencies (caused by time going backwards) are also explicitly
173 * skipped.
174 */
175void __sched
176__account_scheduler_latency(struct task_struct *tsk, int usecs, int inter)
177{
178	unsigned long flags;
179	int i, q;
180	struct latency_record lat;
181
182	/* Long interruptible waits are generally user requested... */
183	if (inter && usecs > 5000)
184		return;
185
186	/* Negative sleeps are time going backwards */
187	/* Zero-time sleeps are non-interesting */
188	if (usecs <= 0)
189		return;
190
191	memset(&lat, 0, sizeof(lat));
192	lat.count = 1;
193	lat.time = usecs;
194	lat.max = usecs;
195
196	stack_trace_save_tsk(tsk, lat.backtrace, LT_BACKTRACEDEPTH, 0);
197
198	raw_spin_lock_irqsave(&latency_lock, flags);
199
200	account_global_scheduler_latency(tsk, &lat);
201
202	for (i = 0; i < tsk->latency_record_count; i++) {
203		struct latency_record *mylat;
204		int same = 1;
205
206		mylat = &tsk->latency_record[i];
207		for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
208			unsigned long record = lat.backtrace[q];
209
210			if (mylat->backtrace[q] != record) {
211				same = 0;
212				break;
213			}
214
215			/* 0 entry is end of backtrace */
216			if (!record)
217				break;
218		}
219		if (same) {
220			mylat->count++;
221			mylat->time += lat.time;
222			if (lat.time > mylat->max)
223				mylat->max = lat.time;
224			goto out_unlock;
225		}
226	}
227
228	/*
229	 * short term hack; if we're > 32 we stop; future we recycle:
230	 */
231	if (tsk->latency_record_count >= LT_SAVECOUNT)
232		goto out_unlock;
233
234	/* Allocated a new one: */
235	i = tsk->latency_record_count++;
236	memcpy(&tsk->latency_record[i], &lat, sizeof(struct latency_record));
237
238out_unlock:
239	raw_spin_unlock_irqrestore(&latency_lock, flags);
240}
241
242static int lstats_show(struct seq_file *m, void *v)
243{
244	int i;
245
246	seq_puts(m, "Latency Top version : v0.1\n");
247
248	for (i = 0; i < MAXLR; i++) {
249		struct latency_record *lr = &latency_record[i];
250
251		if (lr->backtrace[0]) {
252			int q;
253			seq_printf(m, "%i %lu %lu",
254				   lr->count, lr->time, lr->max);
255			for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
256				unsigned long bt = lr->backtrace[q];
257
258				if (!bt)
259					break;
260
261				seq_printf(m, " %ps", (void *)bt);
262			}
263			seq_puts(m, "\n");
264		}
265	}
266	return 0;
267}
268
269static ssize_t
270lstats_write(struct file *file, const char __user *buf, size_t count,
271	     loff_t *offs)
272{
273	clear_global_latency_tracing();
274
275	return count;
276}
277
278static int lstats_open(struct inode *inode, struct file *filp)
279{
280	return single_open(filp, lstats_show, NULL);
281}
282
283static const struct proc_ops lstats_proc_ops = {
284	.proc_open	= lstats_open,
285	.proc_read	= seq_read,
286	.proc_write	= lstats_write,
287	.proc_lseek	= seq_lseek,
288	.proc_release	= single_release,
289};
290
291static int __init init_lstats_procfs(void)
292{
293	proc_create("latency_stats", 0644, NULL, &lstats_proc_ops);
294#ifdef CONFIG_SYSCTL
295	register_sysctl_init("kernel", latencytop_sysctl);
296#endif
297	return 0;
298}
299device_initcall(init_lstats_procfs);
300