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
>> 50 #include <linux/latencytop.h>
46 #include <linux/kallsyms.h> 51 #include <linux/kallsyms.h>
47 #include <linux/seq_file.h> 52 #include <linux/seq_file.h>
48 #include <linux/notifier.h> 53 #include <linux/notifier.h>
49 #include <linux/spinlock.h> 54 #include <linux/spinlock.h>
50 #include <linux/proc_fs.h> 55 #include <linux/proc_fs.h>
51 #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> <<
55 #include <linux/sched/stat.h> <<
56 #include <linux/list.h> 58 #include <linux/list.h>
57 #include <linux/stacktrace.h> 59 #include <linux/stacktrace.h>
58 #include <linux/sysctl.h> <<
59 60
60 static DEFINE_RAW_SPINLOCK(latency_lock); 61 static DEFINE_RAW_SPINLOCK(latency_lock);
61 62
62 #define MAXLR 128 63 #define MAXLR 128
63 static struct latency_record latency_record[MA 64 static struct latency_record latency_record[MAXLR];
64 65
65 int latencytop_enabled; 66 int latencytop_enabled;
66 67
67 #ifdef CONFIG_SYSCTL !! 68 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 { 69 {
93 unsigned long flags; 70 unsigned long flags;
94 71
>> 72 if (!latencytop_enabled)
>> 73 return;
>> 74
95 raw_spin_lock_irqsave(&latency_lock, f 75 raw_spin_lock_irqsave(&latency_lock, flags);
96 memset(&p->latency_record, 0, sizeof(p 76 memset(&p->latency_record, 0, sizeof(p->latency_record));
97 p->latency_record_count = 0; 77 p->latency_record_count = 0;
98 raw_spin_unlock_irqrestore(&latency_lo 78 raw_spin_unlock_irqrestore(&latency_lock, flags);
99 } 79 }
100 80
101 static void clear_global_latency_tracing(void) 81 static void clear_global_latency_tracing(void)
102 { 82 {
103 unsigned long flags; 83 unsigned long flags;
104 84
105 raw_spin_lock_irqsave(&latency_lock, f 85 raw_spin_lock_irqsave(&latency_lock, flags);
106 memset(&latency_record, 0, sizeof(late 86 memset(&latency_record, 0, sizeof(latency_record));
107 raw_spin_unlock_irqrestore(&latency_lo 87 raw_spin_unlock_irqrestore(&latency_lock, flags);
108 } 88 }
109 89
110 static void __sched 90 static void __sched
111 account_global_scheduler_latency(struct task_s !! 91 account_global_scheduler_latency(struct task_struct *tsk, struct latency_record *lat)
112 struct latenc <<
113 { 92 {
114 int firstnonnull = MAXLR; !! 93 int firstnonnull = MAXLR + 1;
115 int i; 94 int i;
116 95
>> 96 if (!latencytop_enabled)
>> 97 return;
>> 98
117 /* skip kernel threads for now */ 99 /* skip kernel threads for now */
118 if (!tsk->mm) 100 if (!tsk->mm)
119 return; 101 return;
120 102
121 for (i = 0; i < MAXLR; i++) { 103 for (i = 0; i < MAXLR; i++) {
122 int q, same = 1; 104 int q, same = 1;
123 105
124 /* Nothing stored: */ 106 /* Nothing stored: */
125 if (!latency_record[i].backtra 107 if (!latency_record[i].backtrace[0]) {
126 if (firstnonnull > i) 108 if (firstnonnull > i)
127 firstnonnull = 109 firstnonnull = i;
128 continue; 110 continue;
129 } 111 }
130 for (q = 0; q < LT_BACKTRACEDE 112 for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
131 unsigned long record = 113 unsigned long record = lat->backtrace[q];
132 114
133 if (latency_record[i]. 115 if (latency_record[i].backtrace[q] != record) {
134 same = 0; 116 same = 0;
135 break; 117 break;
136 } 118 }
137 119
138 /* 0 entry marks end o !! 120 /* 0 and ULONG_MAX entries mean end of backtrace: */
139 if (!record) !! 121 if (record == 0 || record == ULONG_MAX)
140 break; 122 break;
141 } 123 }
142 if (same) { 124 if (same) {
143 latency_record[i].coun 125 latency_record[i].count++;
144 latency_record[i].time 126 latency_record[i].time += lat->time;
145 if (lat->time > latenc 127 if (lat->time > latency_record[i].max)
146 latency_record 128 latency_record[i].max = lat->time;
147 return; 129 return;
148 } 130 }
149 } 131 }
150 132
151 i = firstnonnull; 133 i = firstnonnull;
152 if (i >= MAXLR) !! 134 if (i >= MAXLR - 1)
153 return; 135 return;
154 136
155 /* Allocted a new one: */ 137 /* Allocted a new one: */
156 memcpy(&latency_record[i], lat, sizeof 138 memcpy(&latency_record[i], lat, sizeof(struct latency_record));
157 } 139 }
158 140
>> 141 /*
>> 142 * Iterator to store a backtrace into a latency record entry
>> 143 */
>> 144 static 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
159 /** 155 /**
160 * __account_scheduler_latency - record an occ 156 * __account_scheduler_latency - record an occurred latency
161 * @tsk - the task struct of the task hitting 157 * @tsk - the task struct of the task hitting the latency
162 * @usecs - the duration of the latency in mic 158 * @usecs - the duration of the latency in microseconds
163 * @inter - 1 if the sleep was interruptible, 159 * @inter - 1 if the sleep was interruptible, 0 if uninterruptible
164 * 160 *
165 * This function is the main entry point for r 161 * This function is the main entry point for recording latency entries
166 * as called by the scheduler. 162 * as called by the scheduler.
167 * 163 *
168 * This function has a few special cases to de 164 * This function has a few special cases to deal with normal 'non-latency'
169 * sleeps: specifically, interruptible sleep l 165 * sleeps: specifically, interruptible sleep longer than 5 msec is skipped
170 * since this usually is caused by waiting for 166 * since this usually is caused by waiting for events via select() and co.
171 * 167 *
172 * Negative latencies (caused by time going ba 168 * Negative latencies (caused by time going backwards) are also explicitly
173 * skipped. 169 * skipped.
174 */ 170 */
175 void __sched 171 void __sched
176 __account_scheduler_latency(struct task_struct 172 __account_scheduler_latency(struct task_struct *tsk, int usecs, int inter)
177 { 173 {
178 unsigned long flags; 174 unsigned long flags;
179 int i, q; 175 int i, q;
180 struct latency_record lat; 176 struct latency_record lat;
181 177
182 /* Long interruptible waits are genera 178 /* Long interruptible waits are generally user requested... */
183 if (inter && usecs > 5000) 179 if (inter && usecs > 5000)
184 return; 180 return;
185 181
186 /* Negative sleeps are time going back 182 /* Negative sleeps are time going backwards */
187 /* Zero-time sleeps are non-interestin 183 /* Zero-time sleeps are non-interesting */
188 if (usecs <= 0) 184 if (usecs <= 0)
189 return; 185 return;
190 186
191 memset(&lat, 0, sizeof(lat)); 187 memset(&lat, 0, sizeof(lat));
192 lat.count = 1; 188 lat.count = 1;
193 lat.time = usecs; 189 lat.time = usecs;
194 lat.max = usecs; 190 lat.max = usecs;
195 !! 191 store_stacktrace(tsk, &lat);
196 stack_trace_save_tsk(tsk, lat.backtrac <<
197 192
198 raw_spin_lock_irqsave(&latency_lock, f 193 raw_spin_lock_irqsave(&latency_lock, flags);
199 194
200 account_global_scheduler_latency(tsk, 195 account_global_scheduler_latency(tsk, &lat);
201 196
202 for (i = 0; i < tsk->latency_record_co 197 for (i = 0; i < tsk->latency_record_count; i++) {
203 struct latency_record *mylat; 198 struct latency_record *mylat;
204 int same = 1; 199 int same = 1;
205 200
206 mylat = &tsk->latency_record[i 201 mylat = &tsk->latency_record[i];
207 for (q = 0; q < LT_BACKTRACEDE 202 for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
208 unsigned long record = 203 unsigned long record = lat.backtrace[q];
209 204
210 if (mylat->backtrace[q 205 if (mylat->backtrace[q] != record) {
211 same = 0; 206 same = 0;
212 break; 207 break;
213 } 208 }
214 209
215 /* 0 entry is end of b !! 210 /* 0 and ULONG_MAX entries mean end of backtrace: */
216 if (!record) !! 211 if (record == 0 || record == ULONG_MAX)
217 break; 212 break;
218 } 213 }
219 if (same) { 214 if (same) {
220 mylat->count++; 215 mylat->count++;
221 mylat->time += lat.tim 216 mylat->time += lat.time;
222 if (lat.time > mylat-> 217 if (lat.time > mylat->max)
223 mylat->max = l 218 mylat->max = lat.time;
224 goto out_unlock; 219 goto out_unlock;
225 } 220 }
226 } 221 }
227 222
228 /* 223 /*
229 * short term hack; if we're > 32 we s 224 * short term hack; if we're > 32 we stop; future we recycle:
230 */ 225 */
231 if (tsk->latency_record_count >= LT_SA 226 if (tsk->latency_record_count >= LT_SAVECOUNT)
232 goto out_unlock; 227 goto out_unlock;
233 228
234 /* Allocated a new one: */ 229 /* Allocated a new one: */
235 i = tsk->latency_record_count++; 230 i = tsk->latency_record_count++;
236 memcpy(&tsk->latency_record[i], &lat, 231 memcpy(&tsk->latency_record[i], &lat, sizeof(struct latency_record));
237 232
238 out_unlock: 233 out_unlock:
239 raw_spin_unlock_irqrestore(&latency_lo 234 raw_spin_unlock_irqrestore(&latency_lock, flags);
240 } 235 }
241 236
242 static int lstats_show(struct seq_file *m, voi 237 static int lstats_show(struct seq_file *m, void *v)
243 { 238 {
244 int i; 239 int i;
245 240
246 seq_puts(m, "Latency Top version : v0. 241 seq_puts(m, "Latency Top version : v0.1\n");
247 242
248 for (i = 0; i < MAXLR; i++) { 243 for (i = 0; i < MAXLR; i++) {
249 struct latency_record *lr = &l 244 struct latency_record *lr = &latency_record[i];
250 245
251 if (lr->backtrace[0]) { 246 if (lr->backtrace[0]) {
252 int q; 247 int q;
253 seq_printf(m, "%i %lu 248 seq_printf(m, "%i %lu %lu",
254 lr->count, 249 lr->count, lr->time, lr->max);
255 for (q = 0; q < LT_BAC 250 for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
256 unsigned long 251 unsigned long bt = lr->backtrace[q];
257 <<
258 if (!bt) 252 if (!bt)
259 break; 253 break;
260 !! 254 if (bt == ULONG_MAX)
>> 255 break;
261 seq_printf(m, 256 seq_printf(m, " %ps", (void *)bt);
262 } 257 }
263 seq_puts(m, "\n"); !! 258 seq_printf(m, "\n");
264 } 259 }
265 } 260 }
266 return 0; 261 return 0;
267 } 262 }
268 263
269 static ssize_t 264 static ssize_t
270 lstats_write(struct file *file, const char __u 265 lstats_write(struct file *file, const char __user *buf, size_t count,
271 loff_t *offs) 266 loff_t *offs)
272 { 267 {
273 clear_global_latency_tracing(); 268 clear_global_latency_tracing();
274 269
275 return count; 270 return count;
276 } 271 }
277 272
278 static int lstats_open(struct inode *inode, st 273 static int lstats_open(struct inode *inode, struct file *filp)
279 { 274 {
280 return single_open(filp, lstats_show, 275 return single_open(filp, lstats_show, NULL);
281 } 276 }
282 277
283 static const struct proc_ops lstats_proc_ops = !! 278 static const struct file_operations lstats_fops = {
284 .proc_open = lstats_open, !! 279 .open = lstats_open,
285 .proc_read = seq_read, !! 280 .read = seq_read,
286 .proc_write = lstats_write, !! 281 .write = lstats_write,
287 .proc_lseek = seq_lseek, !! 282 .llseek = seq_lseek,
288 .proc_release = single_release, !! 283 .release = single_release,
289 }; 284 };
290 285
291 static int __init init_lstats_procfs(void) 286 static int __init init_lstats_procfs(void)
292 { 287 {
293 proc_create("latency_stats", 0644, NUL !! 288 proc_create("latency_stats", 0644, NULL, &lstats_fops);
294 #ifdef CONFIG_SYSCTL <<
295 register_sysctl_init("kernel", latency <<
296 #endif <<
297 return 0; 289 return 0;
298 } 290 }
299 device_initcall(init_lstats_procfs); 291 device_initcall(init_lstats_procfs);
300 292
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.