1 // SPDX-License-Identifier: GPL-2.0-only <<
2 /* 1 /*
3 * latencytop.c: Latency display infrastructur 2 * latencytop.c: Latency display infrastructure
4 * 3 *
5 * (C) Copyright 2008 Intel Corporation 4 * (C) Copyright 2008 Intel Corporation
6 * Author: Arjan van de Ven <arjan@linux.intel 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.
7 */ 11 */
8 12
9 /* 13 /*
10 * CONFIG_LATENCYTOP enables a kernel latency 14 * CONFIG_LATENCYTOP enables a kernel latency tracking infrastructure that is
11 * used by the "latencytop" userspace tool. Th 15 * used by the "latencytop" userspace tool. The latency that is tracked is not
12 * the 'traditional' interrupt latency (which 16 * the 'traditional' interrupt latency (which is primarily caused by something
13 * else consuming CPU), but instead, it is the 17 * else consuming CPU), but instead, it is the latency an application encounters
14 * because the kernel sleeps on its behalf for 18 * because the kernel sleeps on its behalf for various reasons.
15 * 19 *
16 * This code tracks 2 levels of statistics: 20 * This code tracks 2 levels of statistics:
17 * 1) System level latency 21 * 1) System level latency
18 * 2) Per process latency 22 * 2) Per process latency
19 * 23 *
20 * The latency is stored in fixed sized data s 24 * The latency is stored in fixed sized data structures in an accumulated form;
21 * if the "same" latency cause is hit twice, t 25 * if the "same" latency cause is hit twice, this will be tracked as one entry
22 * in the data structure. Both the count, tota 26 * in the data structure. Both the count, total accumulated latency and maximum
23 * latency are tracked in this data structure. 27 * latency are tracked in this data structure. When the fixed size structure is
24 * full, no new causes are tracked until the b 28 * full, no new causes are tracked until the buffer is flushed by writing to
25 * the /proc file; the userspace tool does thi 29 * the /proc file; the userspace tool does this on a regular basis.
26 * 30 *
27 * A latency cause is identified by a stringif 31 * A latency cause is identified by a stringified backtrace at the point that
28 * the scheduler gets invoked. The userland to 32 * the scheduler gets invoked. The userland tool will use this string to
29 * identify the cause of the latency in human 33 * identify the cause of the latency in human readable form.
30 * 34 *
31 * The information is exported via /proc/laten 35 * The information is exported via /proc/latency_stats and /proc/<pid>/latency.
32 * These files look like this: 36 * These files look like this:
33 * 37 *
34 * Latency Top version : v0.1 38 * Latency Top version : v0.1
35 * 70 59433 4897 i915_irq_wait drm_ioctl vfs_i 39 * 70 59433 4897 i915_irq_wait drm_ioctl vfs_ioctl do_vfs_ioctl sys_ioctl
36 * | | | | 40 * | | | |
37 * | | | +----> the stringified backt 41 * | | | +----> the stringified backtrace
38 * | | +---------> The maximum latency f 42 * | | +---------> The maximum latency for this entry in microseconds
39 * | +--------------> The accumulated laten 43 * | +--------------> The accumulated latency for this entry (microseconds)
40 * +-------------------> The number of times t 44 * +-------------------> The number of times this entry is hit
41 * 45 *
42 * (note: the average latency is the accumulat 46 * (note: the average latency is the accumulated latency divided by the number
43 * of times) 47 * of times)
44 */ 48 */
45 49
46 #include <linux/kallsyms.h> 50 #include <linux/kallsyms.h>
47 #include <linux/seq_file.h> 51 #include <linux/seq_file.h>
48 #include <linux/notifier.h> 52 #include <linux/notifier.h>
49 #include <linux/spinlock.h> 53 #include <linux/spinlock.h>
50 #include <linux/proc_fs.h> 54 #include <linux/proc_fs.h>
51 #include <linux/latencytop.h> 55 #include <linux/latencytop.h>
52 #include <linux/export.h> 56 #include <linux/export.h>
53 #include <linux/sched.h> 57 #include <linux/sched.h>
54 #include <linux/sched/debug.h> 58 #include <linux/sched/debug.h>
55 #include <linux/sched/stat.h> 59 #include <linux/sched/stat.h>
56 #include <linux/list.h> 60 #include <linux/list.h>
57 #include <linux/stacktrace.h> 61 #include <linux/stacktrace.h>
58 #include <linux/sysctl.h> <<
59 62
60 static DEFINE_RAW_SPINLOCK(latency_lock); 63 static DEFINE_RAW_SPINLOCK(latency_lock);
61 64
62 #define MAXLR 128 65 #define MAXLR 128
63 static struct latency_record latency_record[MA 66 static struct latency_record latency_record[MAXLR];
64 67
65 int latencytop_enabled; 68 int latencytop_enabled;
66 69
67 #ifdef CONFIG_SYSCTL !! 70 void clear_all_latency_tracing(struct task_struct *p)
68 static int sysctl_latencytop(const struct ctl_ <<
69 size_t *lenp, loff_t *ppos) <<
70 { <<
71 int err; <<
72 <<
73 err = proc_dointvec(table, write, buff <<
74 if (latencytop_enabled) <<
75 force_schedstat_enabled(); <<
76 <<
77 return err; <<
78 } <<
79 <<
80 static struct ctl_table latencytop_sysctl[] = <<
81 { <<
82 .procname = "latencytop", <<
83 .data = &latencytop_enab <<
84 .maxlen = sizeof(int), <<
85 .mode = 0644, <<
86 .proc_handler = sysctl_laten <<
87 }, <<
88 }; <<
89 #endif <<
90 <<
91 void clear_tsk_latency_tracing(struct task_str <<
92 { 71 {
93 unsigned long flags; 72 unsigned long flags;
94 73
>> 74 if (!latencytop_enabled)
>> 75 return;
>> 76
95 raw_spin_lock_irqsave(&latency_lock, f 77 raw_spin_lock_irqsave(&latency_lock, flags);
96 memset(&p->latency_record, 0, sizeof(p 78 memset(&p->latency_record, 0, sizeof(p->latency_record));
97 p->latency_record_count = 0; 79 p->latency_record_count = 0;
98 raw_spin_unlock_irqrestore(&latency_lo 80 raw_spin_unlock_irqrestore(&latency_lock, flags);
99 } 81 }
100 82
101 static void clear_global_latency_tracing(void) 83 static void clear_global_latency_tracing(void)
102 { 84 {
103 unsigned long flags; 85 unsigned long flags;
104 86
105 raw_spin_lock_irqsave(&latency_lock, f 87 raw_spin_lock_irqsave(&latency_lock, flags);
106 memset(&latency_record, 0, sizeof(late 88 memset(&latency_record, 0, sizeof(latency_record));
107 raw_spin_unlock_irqrestore(&latency_lo 89 raw_spin_unlock_irqrestore(&latency_lock, flags);
108 } 90 }
109 91
110 static void __sched 92 static void __sched
111 account_global_scheduler_latency(struct task_s 93 account_global_scheduler_latency(struct task_struct *tsk,
112 struct latenc 94 struct latency_record *lat)
113 { 95 {
114 int firstnonnull = MAXLR; !! 96 int firstnonnull = MAXLR + 1;
115 int i; 97 int i;
116 98
>> 99 if (!latencytop_enabled)
>> 100 return;
>> 101
117 /* skip kernel threads for now */ 102 /* skip kernel threads for now */
118 if (!tsk->mm) 103 if (!tsk->mm)
119 return; 104 return;
120 105
121 for (i = 0; i < MAXLR; i++) { 106 for (i = 0; i < MAXLR; i++) {
122 int q, same = 1; 107 int q, same = 1;
123 108
124 /* Nothing stored: */ 109 /* Nothing stored: */
125 if (!latency_record[i].backtra 110 if (!latency_record[i].backtrace[0]) {
126 if (firstnonnull > i) 111 if (firstnonnull > i)
127 firstnonnull = 112 firstnonnull = i;
128 continue; 113 continue;
129 } 114 }
130 for (q = 0; q < LT_BACKTRACEDE 115 for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
131 unsigned long record = 116 unsigned long record = lat->backtrace[q];
132 117
133 if (latency_record[i]. 118 if (latency_record[i].backtrace[q] != record) {
134 same = 0; 119 same = 0;
135 break; 120 break;
136 } 121 }
137 122
138 /* 0 entry marks end o !! 123 /* 0 and ULONG_MAX entries mean end of backtrace: */
139 if (!record) !! 124 if (record == 0 || record == ULONG_MAX)
140 break; 125 break;
141 } 126 }
142 if (same) { 127 if (same) {
143 latency_record[i].coun 128 latency_record[i].count++;
144 latency_record[i].time 129 latency_record[i].time += lat->time;
145 if (lat->time > latenc 130 if (lat->time > latency_record[i].max)
146 latency_record 131 latency_record[i].max = lat->time;
147 return; 132 return;
148 } 133 }
149 } 134 }
150 135
151 i = firstnonnull; 136 i = firstnonnull;
152 if (i >= MAXLR) !! 137 if (i >= MAXLR - 1)
153 return; 138 return;
154 139
155 /* Allocted a new one: */ 140 /* Allocted a new one: */
156 memcpy(&latency_record[i], lat, sizeof 141 memcpy(&latency_record[i], lat, sizeof(struct latency_record));
157 } 142 }
158 143
>> 144 /*
>> 145 * Iterator to store a backtrace into a latency record entry
>> 146 */
>> 147 static inline void store_stacktrace(struct task_struct *tsk,
>> 148 struct latency_record *lat)
>> 149 {
>> 150 struct stack_trace trace;
>> 151
>> 152 memset(&trace, 0, sizeof(trace));
>> 153 trace.max_entries = LT_BACKTRACEDEPTH;
>> 154 trace.entries = &lat->backtrace[0];
>> 155 save_stack_trace_tsk(tsk, &trace);
>> 156 }
>> 157
159 /** 158 /**
160 * __account_scheduler_latency - record an occ 159 * __account_scheduler_latency - record an occurred latency
161 * @tsk - the task struct of the task hitting 160 * @tsk - the task struct of the task hitting the latency
162 * @usecs - the duration of the latency in mic 161 * @usecs - the duration of the latency in microseconds
163 * @inter - 1 if the sleep was interruptible, 162 * @inter - 1 if the sleep was interruptible, 0 if uninterruptible
164 * 163 *
165 * This function is the main entry point for r 164 * This function is the main entry point for recording latency entries
166 * as called by the scheduler. 165 * as called by the scheduler.
167 * 166 *
168 * This function has a few special cases to de 167 * This function has a few special cases to deal with normal 'non-latency'
169 * sleeps: specifically, interruptible sleep l 168 * sleeps: specifically, interruptible sleep longer than 5 msec is skipped
170 * since this usually is caused by waiting for 169 * since this usually is caused by waiting for events via select() and co.
171 * 170 *
172 * Negative latencies (caused by time going ba 171 * Negative latencies (caused by time going backwards) are also explicitly
173 * skipped. 172 * skipped.
174 */ 173 */
175 void __sched 174 void __sched
176 __account_scheduler_latency(struct task_struct 175 __account_scheduler_latency(struct task_struct *tsk, int usecs, int inter)
177 { 176 {
178 unsigned long flags; 177 unsigned long flags;
179 int i, q; 178 int i, q;
180 struct latency_record lat; 179 struct latency_record lat;
181 180
182 /* Long interruptible waits are genera 181 /* Long interruptible waits are generally user requested... */
183 if (inter && usecs > 5000) 182 if (inter && usecs > 5000)
184 return; 183 return;
185 184
186 /* Negative sleeps are time going back 185 /* Negative sleeps are time going backwards */
187 /* Zero-time sleeps are non-interestin 186 /* Zero-time sleeps are non-interesting */
188 if (usecs <= 0) 187 if (usecs <= 0)
189 return; 188 return;
190 189
191 memset(&lat, 0, sizeof(lat)); 190 memset(&lat, 0, sizeof(lat));
192 lat.count = 1; 191 lat.count = 1;
193 lat.time = usecs; 192 lat.time = usecs;
194 lat.max = usecs; 193 lat.max = usecs;
195 !! 194 store_stacktrace(tsk, &lat);
196 stack_trace_save_tsk(tsk, lat.backtrac <<
197 195
198 raw_spin_lock_irqsave(&latency_lock, f 196 raw_spin_lock_irqsave(&latency_lock, flags);
199 197
200 account_global_scheduler_latency(tsk, 198 account_global_scheduler_latency(tsk, &lat);
201 199
202 for (i = 0; i < tsk->latency_record_co 200 for (i = 0; i < tsk->latency_record_count; i++) {
203 struct latency_record *mylat; 201 struct latency_record *mylat;
204 int same = 1; 202 int same = 1;
205 203
206 mylat = &tsk->latency_record[i 204 mylat = &tsk->latency_record[i];
207 for (q = 0; q < LT_BACKTRACEDE 205 for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
208 unsigned long record = 206 unsigned long record = lat.backtrace[q];
209 207
210 if (mylat->backtrace[q 208 if (mylat->backtrace[q] != record) {
211 same = 0; 209 same = 0;
212 break; 210 break;
213 } 211 }
214 212
215 /* 0 entry is end of b !! 213 /* 0 and ULONG_MAX entries mean end of backtrace: */
216 if (!record) !! 214 if (record == 0 || record == ULONG_MAX)
217 break; 215 break;
218 } 216 }
219 if (same) { 217 if (same) {
220 mylat->count++; 218 mylat->count++;
221 mylat->time += lat.tim 219 mylat->time += lat.time;
222 if (lat.time > mylat-> 220 if (lat.time > mylat->max)
223 mylat->max = l 221 mylat->max = lat.time;
224 goto out_unlock; 222 goto out_unlock;
225 } 223 }
226 } 224 }
227 225
228 /* 226 /*
229 * short term hack; if we're > 32 we s 227 * short term hack; if we're > 32 we stop; future we recycle:
230 */ 228 */
231 if (tsk->latency_record_count >= LT_SA 229 if (tsk->latency_record_count >= LT_SAVECOUNT)
232 goto out_unlock; 230 goto out_unlock;
233 231
234 /* Allocated a new one: */ 232 /* Allocated a new one: */
235 i = tsk->latency_record_count++; 233 i = tsk->latency_record_count++;
236 memcpy(&tsk->latency_record[i], &lat, 234 memcpy(&tsk->latency_record[i], &lat, sizeof(struct latency_record));
237 235
238 out_unlock: 236 out_unlock:
239 raw_spin_unlock_irqrestore(&latency_lo 237 raw_spin_unlock_irqrestore(&latency_lock, flags);
240 } 238 }
241 239
242 static int lstats_show(struct seq_file *m, voi 240 static int lstats_show(struct seq_file *m, void *v)
243 { 241 {
244 int i; 242 int i;
245 243
246 seq_puts(m, "Latency Top version : v0. 244 seq_puts(m, "Latency Top version : v0.1\n");
247 245
248 for (i = 0; i < MAXLR; i++) { 246 for (i = 0; i < MAXLR; i++) {
249 struct latency_record *lr = &l 247 struct latency_record *lr = &latency_record[i];
250 248
251 if (lr->backtrace[0]) { 249 if (lr->backtrace[0]) {
252 int q; 250 int q;
253 seq_printf(m, "%i %lu 251 seq_printf(m, "%i %lu %lu",
254 lr->count, 252 lr->count, lr->time, lr->max);
255 for (q = 0; q < LT_BAC 253 for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
256 unsigned long 254 unsigned long bt = lr->backtrace[q];
257 <<
258 if (!bt) 255 if (!bt)
259 break; 256 break;
260 !! 257 if (bt == ULONG_MAX)
>> 258 break;
261 seq_printf(m, 259 seq_printf(m, " %ps", (void *)bt);
262 } 260 }
263 seq_puts(m, "\n"); 261 seq_puts(m, "\n");
264 } 262 }
265 } 263 }
266 return 0; 264 return 0;
267 } 265 }
268 266
269 static ssize_t 267 static ssize_t
270 lstats_write(struct file *file, const char __u 268 lstats_write(struct file *file, const char __user *buf, size_t count,
271 loff_t *offs) 269 loff_t *offs)
272 { 270 {
273 clear_global_latency_tracing(); 271 clear_global_latency_tracing();
274 272
275 return count; 273 return count;
276 } 274 }
277 275
278 static int lstats_open(struct inode *inode, st 276 static int lstats_open(struct inode *inode, struct file *filp)
279 { 277 {
280 return single_open(filp, lstats_show, 278 return single_open(filp, lstats_show, NULL);
281 } 279 }
282 280
283 static const struct proc_ops lstats_proc_ops = !! 281 static const struct file_operations lstats_fops = {
284 .proc_open = lstats_open, !! 282 .open = lstats_open,
285 .proc_read = seq_read, !! 283 .read = seq_read,
286 .proc_write = lstats_write, !! 284 .write = lstats_write,
287 .proc_lseek = seq_lseek, !! 285 .llseek = seq_lseek,
288 .proc_release = single_release, !! 286 .release = single_release,
289 }; 287 };
290 288
291 static int __init init_lstats_procfs(void) 289 static int __init init_lstats_procfs(void)
292 { 290 {
293 proc_create("latency_stats", 0644, NUL !! 291 proc_create("latency_stats", 0644, NULL, &lstats_fops);
294 #ifdef CONFIG_SYSCTL <<
295 register_sysctl_init("kernel", latency <<
296 #endif <<
297 return 0; 292 return 0;
>> 293 }
>> 294
>> 295 int sysctl_latencytop(struct ctl_table *table, int write,
>> 296 void __user *buffer, size_t *lenp, loff_t *ppos)
>> 297 {
>> 298 int err;
>> 299
>> 300 err = proc_dointvec(table, write, buffer, lenp, ppos);
>> 301 if (latencytop_enabled)
>> 302 force_schedstat_enabled();
>> 303
>> 304 return err;
298 } 305 }
299 device_initcall(init_lstats_procfs); 306 device_initcall(init_lstats_procfs);
300 307
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.