xref: /openbmc/linux/kernel/latencytop.c (revision 04dc82e1)
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 
59 static DEFINE_RAW_SPINLOCK(latency_lock);
60 
61 #define MAXLR 128
62 static struct latency_record latency_record[MAXLR];
63 
64 int latencytop_enabled;
65 
66 void clear_tsk_latency_tracing(struct task_struct *p)
67 {
68 	unsigned long flags;
69 
70 	raw_spin_lock_irqsave(&latency_lock, flags);
71 	memset(&p->latency_record, 0, sizeof(p->latency_record));
72 	p->latency_record_count = 0;
73 	raw_spin_unlock_irqrestore(&latency_lock, flags);
74 }
75 
76 static void clear_global_latency_tracing(void)
77 {
78 	unsigned long flags;
79 
80 	raw_spin_lock_irqsave(&latency_lock, flags);
81 	memset(&latency_record, 0, sizeof(latency_record));
82 	raw_spin_unlock_irqrestore(&latency_lock, flags);
83 }
84 
85 static void __sched
86 account_global_scheduler_latency(struct task_struct *tsk,
87 				 struct latency_record *lat)
88 {
89 	int firstnonnull = MAXLR + 1;
90 	int i;
91 
92 	/* skip kernel threads for now */
93 	if (!tsk->mm)
94 		return;
95 
96 	for (i = 0; i < MAXLR; i++) {
97 		int q, same = 1;
98 
99 		/* Nothing stored: */
100 		if (!latency_record[i].backtrace[0]) {
101 			if (firstnonnull > i)
102 				firstnonnull = i;
103 			continue;
104 		}
105 		for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
106 			unsigned long record = lat->backtrace[q];
107 
108 			if (latency_record[i].backtrace[q] != record) {
109 				same = 0;
110 				break;
111 			}
112 
113 			/* 0 entry marks end of backtrace: */
114 			if (!record)
115 				break;
116 		}
117 		if (same) {
118 			latency_record[i].count++;
119 			latency_record[i].time += lat->time;
120 			if (lat->time > latency_record[i].max)
121 				latency_record[i].max = lat->time;
122 			return;
123 		}
124 	}
125 
126 	i = firstnonnull;
127 	if (i >= MAXLR - 1)
128 		return;
129 
130 	/* Allocted a new one: */
131 	memcpy(&latency_record[i], lat, sizeof(struct latency_record));
132 }
133 
134 /**
135  * __account_scheduler_latency - record an occurred latency
136  * @tsk - the task struct of the task hitting the latency
137  * @usecs - the duration of the latency in microseconds
138  * @inter - 1 if the sleep was interruptible, 0 if uninterruptible
139  *
140  * This function is the main entry point for recording latency entries
141  * as called by the scheduler.
142  *
143  * This function has a few special cases to deal with normal 'non-latency'
144  * sleeps: specifically, interruptible sleep longer than 5 msec is skipped
145  * since this usually is caused by waiting for events via select() and co.
146  *
147  * Negative latencies (caused by time going backwards) are also explicitly
148  * skipped.
149  */
150 void __sched
151 __account_scheduler_latency(struct task_struct *tsk, int usecs, int inter)
152 {
153 	unsigned long flags;
154 	int i, q;
155 	struct latency_record lat;
156 
157 	/* Long interruptible waits are generally user requested... */
158 	if (inter && usecs > 5000)
159 		return;
160 
161 	/* Negative sleeps are time going backwards */
162 	/* Zero-time sleeps are non-interesting */
163 	if (usecs <= 0)
164 		return;
165 
166 	memset(&lat, 0, sizeof(lat));
167 	lat.count = 1;
168 	lat.time = usecs;
169 	lat.max = usecs;
170 
171 	stack_trace_save_tsk(tsk, lat.backtrace, LT_BACKTRACEDEPTH, 0);
172 
173 	raw_spin_lock_irqsave(&latency_lock, flags);
174 
175 	account_global_scheduler_latency(tsk, &lat);
176 
177 	for (i = 0; i < tsk->latency_record_count; i++) {
178 		struct latency_record *mylat;
179 		int same = 1;
180 
181 		mylat = &tsk->latency_record[i];
182 		for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
183 			unsigned long record = lat.backtrace[q];
184 
185 			if (mylat->backtrace[q] != record) {
186 				same = 0;
187 				break;
188 			}
189 
190 			/* 0 entry is end of backtrace */
191 			if (!record)
192 				break;
193 		}
194 		if (same) {
195 			mylat->count++;
196 			mylat->time += lat.time;
197 			if (lat.time > mylat->max)
198 				mylat->max = lat.time;
199 			goto out_unlock;
200 		}
201 	}
202 
203 	/*
204 	 * short term hack; if we're > 32 we stop; future we recycle:
205 	 */
206 	if (tsk->latency_record_count >= LT_SAVECOUNT)
207 		goto out_unlock;
208 
209 	/* Allocated a new one: */
210 	i = tsk->latency_record_count++;
211 	memcpy(&tsk->latency_record[i], &lat, sizeof(struct latency_record));
212 
213 out_unlock:
214 	raw_spin_unlock_irqrestore(&latency_lock, flags);
215 }
216 
217 static int lstats_show(struct seq_file *m, void *v)
218 {
219 	int i;
220 
221 	seq_puts(m, "Latency Top version : v0.1\n");
222 
223 	for (i = 0; i < MAXLR; i++) {
224 		struct latency_record *lr = &latency_record[i];
225 
226 		if (lr->backtrace[0]) {
227 			int q;
228 			seq_printf(m, "%i %lu %lu",
229 				   lr->count, lr->time, lr->max);
230 			for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
231 				unsigned long bt = lr->backtrace[q];
232 
233 				if (!bt)
234 					break;
235 
236 				seq_printf(m, " %ps", (void *)bt);
237 			}
238 			seq_puts(m, "\n");
239 		}
240 	}
241 	return 0;
242 }
243 
244 static ssize_t
245 lstats_write(struct file *file, const char __user *buf, size_t count,
246 	     loff_t *offs)
247 {
248 	clear_global_latency_tracing();
249 
250 	return count;
251 }
252 
253 static int lstats_open(struct inode *inode, struct file *filp)
254 {
255 	return single_open(filp, lstats_show, NULL);
256 }
257 
258 static const struct file_operations lstats_fops = {
259 	.open		= lstats_open,
260 	.read		= seq_read,
261 	.write		= lstats_write,
262 	.llseek		= seq_lseek,
263 	.release	= single_release,
264 };
265 
266 static int __init init_lstats_procfs(void)
267 {
268 	proc_create("latency_stats", 0644, NULL, &lstats_fops);
269 	return 0;
270 }
271 
272 int sysctl_latencytop(struct ctl_table *table, int write,
273 			void __user *buffer, size_t *lenp, loff_t *ppos)
274 {
275 	int err;
276 
277 	err = proc_dointvec(table, write, buffer, lenp, ppos);
278 	if (latencytop_enabled)
279 		force_schedstat_enabled();
280 
281 	return err;
282 }
283 device_initcall(init_lstats_procfs);
284