1/*
2 * latencytop.c: Latency display infrastructure
3 *
4 * (C) Copyright 2008 Intel Corporation
5 * Author: Arjan van de Ven <arjan@linux.intel.com>
6 *
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
10 * of the License.
11 */
12
13/*
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.
19 *
20 * This code tracks 2 levels of statistics:
21 * 1) System level latency
22 * 2) Per process latency
23 *
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.
30 *
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.
34 *
35 * The information is exported via /proc/latency_stats and /proc/<pid>/latency.
36 * These files look like this:
37 *
38 * Latency Top version : v0.1
39 * 70 59433 4897 i915_irq_wait drm_ioctl vfs_ioctl do_vfs_ioctl sys_ioctl
40 * |    |    |    |
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
45 *
46 * (note: the average latency is the accumulated latency divided by the number
47 * of times)
48 */
49
50#include <linux/latencytop.h>
51#include <linux/kallsyms.h>
52#include <linux/seq_file.h>
53#include <linux/notifier.h>
54#include <linux/spinlock.h>
55#include <linux/proc_fs.h>
56#include <linux/export.h>
57#include <linux/sched.h>
58#include <linux/list.h>
59#include <linux/stacktrace.h>
60
61static DEFINE_RAW_SPINLOCK(latency_lock);
62
63#define MAXLR 128
64static struct latency_record latency_record[MAXLR];
65
66int latencytop_enabled;
67
68void clear_all_latency_tracing(struct task_struct *p)
69{
70	unsigned long flags;
71
72	if (!latencytop_enabled)
73		return;
74
75	raw_spin_lock_irqsave(&latency_lock, flags);
76	memset(&p->latency_record, 0, sizeof(p->latency_record));
77	p->latency_record_count = 0;
78	raw_spin_unlock_irqrestore(&latency_lock, flags);
79}
80
81static void clear_global_latency_tracing(void)
82{
83	unsigned long flags;
84
85	raw_spin_lock_irqsave(&latency_lock, flags);
86	memset(&latency_record, 0, sizeof(latency_record));
87	raw_spin_unlock_irqrestore(&latency_lock, flags);
88}
89
90static void __sched
91account_global_scheduler_latency(struct task_struct *tsk,
92				 struct latency_record *lat)
93{
94	int firstnonnull = MAXLR + 1;
95	int i;
96
97	if (!latencytop_enabled)
98		return;
99
100	/* skip kernel threads for now */
101	if (!tsk->mm)
102		return;
103
104	for (i = 0; i < MAXLR; i++) {
105		int q, same = 1;
106
107		/* Nothing stored: */
108		if (!latency_record[i].backtrace[0]) {
109			if (firstnonnull > i)
110				firstnonnull = i;
111			continue;
112		}
113		for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
114			unsigned long record = lat->backtrace[q];
115
116			if (latency_record[i].backtrace[q] != record) {
117				same = 0;
118				break;
119			}
120
121			/* 0 and ULONG_MAX entries mean end of backtrace: */
122			if (record == 0 || record == ULONG_MAX)
123				break;
124		}
125		if (same) {
126			latency_record[i].count++;
127			latency_record[i].time += lat->time;
128			if (lat->time > latency_record[i].max)
129				latency_record[i].max = lat->time;
130			return;
131		}
132	}
133
134	i = firstnonnull;
135	if (i >= MAXLR - 1)
136		return;
137
138	/* Allocted a new one: */
139	memcpy(&latency_record[i], lat, sizeof(struct latency_record));
140}
141
142/*
143 * Iterator to store a backtrace into a latency record entry
144 */
145static inline void store_stacktrace(struct task_struct *tsk,
146					struct latency_record *lat)
147{
148	struct stack_trace trace;
149
150	memset(&trace, 0, sizeof(trace));
151	trace.max_entries = LT_BACKTRACEDEPTH;
152	trace.entries = &lat->backtrace[0];
153	save_stack_trace_tsk(tsk, &trace);
154}
155
156/**
157 * __account_scheduler_latency - record an occurred latency
158 * @tsk - the task struct of the task hitting the latency
159 * @usecs - the duration of the latency in microseconds
160 * @inter - 1 if the sleep was interruptible, 0 if uninterruptible
161 *
162 * This function is the main entry point for recording latency entries
163 * as called by the scheduler.
164 *
165 * This function has a few special cases to deal with normal 'non-latency'
166 * sleeps: specifically, interruptible sleep longer than 5 msec is skipped
167 * since this usually is caused by waiting for events via select() and co.
168 *
169 * Negative latencies (caused by time going backwards) are also explicitly
170 * skipped.
171 */
172void __sched
173__account_scheduler_latency(struct task_struct *tsk, int usecs, int inter)
174{
175	unsigned long flags;
176	int i, q;
177	struct latency_record lat;
178
179	/* Long interruptible waits are generally user requested... */
180	if (inter && usecs > 5000)
181		return;
182
183	/* Negative sleeps are time going backwards */
184	/* Zero-time sleeps are non-interesting */
185	if (usecs <= 0)
186		return;
187
188	memset(&lat, 0, sizeof(lat));
189	lat.count = 1;
190	lat.time = usecs;
191	lat.max = usecs;
192	store_stacktrace(tsk, &lat);
193
194	raw_spin_lock_irqsave(&latency_lock, flags);
195
196	account_global_scheduler_latency(tsk, &lat);
197
198	for (i = 0; i < tsk->latency_record_count; i++) {
199		struct latency_record *mylat;
200		int same = 1;
201
202		mylat = &tsk->latency_record[i];
203		for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
204			unsigned long record = lat.backtrace[q];
205
206			if (mylat->backtrace[q] != record) {
207				same = 0;
208				break;
209			}
210
211			/* 0 and ULONG_MAX entries mean end of backtrace: */
212			if (record == 0 || record == ULONG_MAX)
213				break;
214		}
215		if (same) {
216			mylat->count++;
217			mylat->time += lat.time;
218			if (lat.time > mylat->max)
219				mylat->max = lat.time;
220			goto out_unlock;
221		}
222	}
223
224	/*
225	 * short term hack; if we're > 32 we stop; future we recycle:
226	 */
227	if (tsk->latency_record_count >= LT_SAVECOUNT)
228		goto out_unlock;
229
230	/* Allocated a new one: */
231	i = tsk->latency_record_count++;
232	memcpy(&tsk->latency_record[i], &lat, sizeof(struct latency_record));
233
234out_unlock:
235	raw_spin_unlock_irqrestore(&latency_lock, flags);
236}
237
238static int lstats_show(struct seq_file *m, void *v)
239{
240	int i;
241
242	seq_puts(m, "Latency Top version : v0.1\n");
243
244	for (i = 0; i < MAXLR; i++) {
245		struct latency_record *lr = &latency_record[i];
246
247		if (lr->backtrace[0]) {
248			int q;
249			seq_printf(m, "%i %lu %lu",
250				   lr->count, lr->time, lr->max);
251			for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
252				unsigned long bt = lr->backtrace[q];
253				if (!bt)
254					break;
255				if (bt == ULONG_MAX)
256					break;
257				seq_printf(m, " %ps", (void *)bt);
258			}
259			seq_puts(m, "\n");
260		}
261	}
262	return 0;
263}
264
265static ssize_t
266lstats_write(struct file *file, const char __user *buf, size_t count,
267	     loff_t *offs)
268{
269	clear_global_latency_tracing();
270
271	return count;
272}
273
274static int lstats_open(struct inode *inode, struct file *filp)
275{
276	return single_open(filp, lstats_show, NULL);
277}
278
279static const struct file_operations lstats_fops = {
280	.open		= lstats_open,
281	.read		= seq_read,
282	.write		= lstats_write,
283	.llseek		= seq_lseek,
284	.release	= single_release,
285};
286
287static int __init init_lstats_procfs(void)
288{
289	proc_create("latency_stats", 0644, NULL, &lstats_fops);
290	return 0;
291}
292device_initcall(init_lstats_procfs);
293