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/module.h>
57#include <linux/sched.h>
58#include <linux/list.h>
59#include <linux/stacktrace.h>
60
61static DEFINE_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	spin_lock_irqsave(&latency_lock, flags);
76	memset(&p->latency_record, 0, sizeof(p->latency_record));
77	p->latency_record_count = 0;
78	spin_unlock_irqrestore(&latency_lock, flags);
79}
80
81static void clear_global_latency_tracing(void)
82{
83	unsigned long flags;
84
85	spin_lock_irqsave(&latency_lock, flags);
86	memset(&latency_record, 0, sizeof(latency_record));
87	spin_unlock_irqrestore(&latency_lock, flags);
88}
89
90static void __sched
91account_global_scheduler_latency(struct task_struct *tsk, struct latency_record *lat)
92{
93	int firstnonnull = MAXLR + 1;
94	int i;
95
96	if (!latencytop_enabled)
97		return;
98
99	/* skip kernel threads for now */
100	if (!tsk->mm)
101		return;
102
103	for (i = 0; i < MAXLR; i++) {
104		int q, same = 1;
105
106		/* Nothing stored: */
107		if (!latency_record[i].backtrace[0]) {
108			if (firstnonnull > i)
109				firstnonnull = i;
110			continue;
111		}
112		for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
113			unsigned long record = lat->backtrace[q];
114
115			if (latency_record[i].backtrace[q] != record) {
116				same = 0;
117				break;
118			}
119
120			/* 0 and ULONG_MAX entries mean end of backtrace: */
121			if (record == 0 || record == ULONG_MAX)
122				break;
123		}
124		if (same) {
125			latency_record[i].count++;
126			latency_record[i].time += lat->time;
127			if (lat->time > latency_record[i].max)
128				latency_record[i].max = lat->time;
129			return;
130		}
131	}
132
133	i = firstnonnull;
134	if (i >= MAXLR - 1)
135		return;
136
137	/* Allocted a new one: */
138	memcpy(&latency_record[i], lat, sizeof(struct latency_record));
139}
140
141/*
142 * Iterator to store a backtrace into a latency record entry
143 */
144static inline void store_stacktrace(struct task_struct *tsk,
145					struct latency_record *lat)
146{
147	struct stack_trace trace;
148
149	memset(&trace, 0, sizeof(trace));
150	trace.max_entries = LT_BACKTRACEDEPTH;
151	trace.entries = &lat->backtrace[0];
152	save_stack_trace_tsk(tsk, &trace);
153}
154
155/**
156 * __account_scheduler_latency - record an occured latency
157 * @tsk - the task struct of the task hitting the latency
158 * @usecs - the duration of the latency in microseconds
159 * @inter - 1 if the sleep was interruptible, 0 if uninterruptible
160 *
161 * This function is the main entry point for recording latency entries
162 * as called by the scheduler.
163 *
164 * This function has a few special cases to deal with normal 'non-latency'
165 * sleeps: specifically, interruptible sleep longer than 5 msec is skipped
166 * since this usually is caused by waiting for events via select() and co.
167 *
168 * Negative latencies (caused by time going backwards) are also explicitly
169 * skipped.
170 */
171void __sched
172__account_scheduler_latency(struct task_struct *tsk, int usecs, int inter)
173{
174	unsigned long flags;
175	int i, q;
176	struct latency_record lat;
177
178	/* Long interruptible waits are generally user requested... */
179	if (inter && usecs > 5000)
180		return;
181
182	/* Negative sleeps are time going backwards */
183	/* Zero-time sleeps are non-interesting */
184	if (usecs <= 0)
185		return;
186
187	memset(&lat, 0, sizeof(lat));
188	lat.count = 1;
189	lat.time = usecs;
190	lat.max = usecs;
191	store_stacktrace(tsk, &lat);
192
193	spin_lock_irqsave(&latency_lock, flags);
194
195	account_global_scheduler_latency(tsk, &lat);
196
197	for (i = 0; i < tsk->latency_record_count; i++) {
198		struct latency_record *mylat;
199		int same = 1;
200
201		mylat = &tsk->latency_record[i];
202		for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
203			unsigned long record = lat.backtrace[q];
204
205			if (mylat->backtrace[q] != record) {
206				same = 0;
207				break;
208			}
209
210			/* 0 and ULONG_MAX entries mean end of backtrace: */
211			if (record == 0 || record == ULONG_MAX)
212				break;
213		}
214		if (same) {
215			mylat->count++;
216			mylat->time += lat.time;
217			if (lat.time > mylat->max)
218				mylat->max = lat.time;
219			goto out_unlock;
220		}
221	}
222
223	/*
224	 * short term hack; if we're > 32 we stop; future we recycle:
225	 */
226	if (tsk->latency_record_count >= LT_SAVECOUNT)
227		goto out_unlock;
228
229	/* Allocated a new one: */
230	i = tsk->latency_record_count++;
231	memcpy(&tsk->latency_record[i], &lat, sizeof(struct latency_record));
232
233out_unlock:
234	spin_unlock_irqrestore(&latency_lock, flags);
235}
236
237static int lstats_show(struct seq_file *m, void *v)
238{
239	int i;
240
241	seq_puts(m, "Latency Top version : v0.1\n");
242
243	for (i = 0; i < MAXLR; i++) {
244		if (latency_record[i].backtrace[0]) {
245			int q;
246			seq_printf(m, "%i %lu %lu ",
247				latency_record[i].count,
248				latency_record[i].time,
249				latency_record[i].max);
250			for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
251				char sym[KSYM_SYMBOL_LEN];
252				char *c;
253				if (!latency_record[i].backtrace[q])
254					break;
255				if (latency_record[i].backtrace[q] == ULONG_MAX)
256					break;
257				sprint_symbol(sym, latency_record[i].backtrace[q]);
258				c = strchr(sym, '+');
259				if (c)
260					*c = 0;
261				seq_printf(m, "%s ", sym);
262			}
263			seq_printf(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 file_operations lstats_fops = {
284	.open		= lstats_open,
285	.read		= seq_read,
286	.write		= lstats_write,
287	.llseek		= seq_lseek,
288	.release	= single_release,
289};
290
291static int __init init_lstats_procfs(void)
292{
293	proc_create("latency_stats", 0644, NULL, &lstats_fops);
294	return 0;
295}
296device_initcall(init_lstats_procfs);
297