xref: /openbmc/linux/kernel/latencytop.c (revision 752beb5e)
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/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/sched/stat.h>
60 #include <linux/list.h>
61 #include <linux/stacktrace.h>
62 
63 static DEFINE_RAW_SPINLOCK(latency_lock);
64 
65 #define MAXLR 128
66 static struct latency_record latency_record[MAXLR];
67 
68 int latencytop_enabled;
69 
70 void clear_all_latency_tracing(struct task_struct *p)
71 {
72 	unsigned long flags;
73 
74 	if (!latencytop_enabled)
75 		return;
76 
77 	raw_spin_lock_irqsave(&latency_lock, flags);
78 	memset(&p->latency_record, 0, sizeof(p->latency_record));
79 	p->latency_record_count = 0;
80 	raw_spin_unlock_irqrestore(&latency_lock, flags);
81 }
82 
83 static void clear_global_latency_tracing(void)
84 {
85 	unsigned long flags;
86 
87 	raw_spin_lock_irqsave(&latency_lock, flags);
88 	memset(&latency_record, 0, sizeof(latency_record));
89 	raw_spin_unlock_irqrestore(&latency_lock, flags);
90 }
91 
92 static void __sched
93 account_global_scheduler_latency(struct task_struct *tsk,
94 				 struct latency_record *lat)
95 {
96 	int firstnonnull = MAXLR + 1;
97 	int i;
98 
99 	if (!latencytop_enabled)
100 		return;
101 
102 	/* skip kernel threads for now */
103 	if (!tsk->mm)
104 		return;
105 
106 	for (i = 0; i < MAXLR; i++) {
107 		int q, same = 1;
108 
109 		/* Nothing stored: */
110 		if (!latency_record[i].backtrace[0]) {
111 			if (firstnonnull > i)
112 				firstnonnull = i;
113 			continue;
114 		}
115 		for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
116 			unsigned long record = lat->backtrace[q];
117 
118 			if (latency_record[i].backtrace[q] != record) {
119 				same = 0;
120 				break;
121 			}
122 
123 			/* 0 entry marks end of backtrace: */
124 			if (!record)
125 				break;
126 		}
127 		if (same) {
128 			latency_record[i].count++;
129 			latency_record[i].time += lat->time;
130 			if (lat->time > latency_record[i].max)
131 				latency_record[i].max = lat->time;
132 			return;
133 		}
134 	}
135 
136 	i = firstnonnull;
137 	if (i >= MAXLR - 1)
138 		return;
139 
140 	/* Allocted a new one: */
141 	memcpy(&latency_record[i], lat, sizeof(struct latency_record));
142 }
143 
144 /**
145  * __account_scheduler_latency - record an occurred latency
146  * @tsk - the task struct of the task hitting the latency
147  * @usecs - the duration of the latency in microseconds
148  * @inter - 1 if the sleep was interruptible, 0 if uninterruptible
149  *
150  * This function is the main entry point for recording latency entries
151  * as called by the scheduler.
152  *
153  * This function has a few special cases to deal with normal 'non-latency'
154  * sleeps: specifically, interruptible sleep longer than 5 msec is skipped
155  * since this usually is caused by waiting for events via select() and co.
156  *
157  * Negative latencies (caused by time going backwards) are also explicitly
158  * skipped.
159  */
160 void __sched
161 __account_scheduler_latency(struct task_struct *tsk, int usecs, int inter)
162 {
163 	unsigned long flags;
164 	int i, q;
165 	struct latency_record lat;
166 
167 	/* Long interruptible waits are generally user requested... */
168 	if (inter && usecs > 5000)
169 		return;
170 
171 	/* Negative sleeps are time going backwards */
172 	/* Zero-time sleeps are non-interesting */
173 	if (usecs <= 0)
174 		return;
175 
176 	memset(&lat, 0, sizeof(lat));
177 	lat.count = 1;
178 	lat.time = usecs;
179 	lat.max = usecs;
180 
181 	stack_trace_save_tsk(tsk, lat.backtrace, LT_BACKTRACEDEPTH, 0);
182 
183 	raw_spin_lock_irqsave(&latency_lock, flags);
184 
185 	account_global_scheduler_latency(tsk, &lat);
186 
187 	for (i = 0; i < tsk->latency_record_count; i++) {
188 		struct latency_record *mylat;
189 		int same = 1;
190 
191 		mylat = &tsk->latency_record[i];
192 		for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
193 			unsigned long record = lat.backtrace[q];
194 
195 			if (mylat->backtrace[q] != record) {
196 				same = 0;
197 				break;
198 			}
199 
200 			/* 0 entry is end of backtrace */
201 			if (!record)
202 				break;
203 		}
204 		if (same) {
205 			mylat->count++;
206 			mylat->time += lat.time;
207 			if (lat.time > mylat->max)
208 				mylat->max = lat.time;
209 			goto out_unlock;
210 		}
211 	}
212 
213 	/*
214 	 * short term hack; if we're > 32 we stop; future we recycle:
215 	 */
216 	if (tsk->latency_record_count >= LT_SAVECOUNT)
217 		goto out_unlock;
218 
219 	/* Allocated a new one: */
220 	i = tsk->latency_record_count++;
221 	memcpy(&tsk->latency_record[i], &lat, sizeof(struct latency_record));
222 
223 out_unlock:
224 	raw_spin_unlock_irqrestore(&latency_lock, flags);
225 }
226 
227 static int lstats_show(struct seq_file *m, void *v)
228 {
229 	int i;
230 
231 	seq_puts(m, "Latency Top version : v0.1\n");
232 
233 	for (i = 0; i < MAXLR; i++) {
234 		struct latency_record *lr = &latency_record[i];
235 
236 		if (lr->backtrace[0]) {
237 			int q;
238 			seq_printf(m, "%i %lu %lu",
239 				   lr->count, lr->time, lr->max);
240 			for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
241 				unsigned long bt = lr->backtrace[q];
242 
243 				if (!bt)
244 					break;
245 
246 				seq_printf(m, " %ps", (void *)bt);
247 			}
248 			seq_puts(m, "\n");
249 		}
250 	}
251 	return 0;
252 }
253 
254 static ssize_t
255 lstats_write(struct file *file, const char __user *buf, size_t count,
256 	     loff_t *offs)
257 {
258 	clear_global_latency_tracing();
259 
260 	return count;
261 }
262 
263 static int lstats_open(struct inode *inode, struct file *filp)
264 {
265 	return single_open(filp, lstats_show, NULL);
266 }
267 
268 static const struct file_operations lstats_fops = {
269 	.open		= lstats_open,
270 	.read		= seq_read,
271 	.write		= lstats_write,
272 	.llseek		= seq_lseek,
273 	.release	= single_release,
274 };
275 
276 static int __init init_lstats_procfs(void)
277 {
278 	proc_create("latency_stats", 0644, NULL, &lstats_fops);
279 	return 0;
280 }
281 
282 int sysctl_latencytop(struct ctl_table *table, int write,
283 			void __user *buffer, size_t *lenp, loff_t *ppos)
284 {
285 	int err;
286 
287 	err = proc_dointvec(table, write, buffer, lenp, ppos);
288 	if (latencytop_enabled)
289 		force_schedstat_enabled();
290 
291 	return err;
292 }
293 device_initcall(init_lstats_procfs);
294