clockevents: optimise tick_nohz_stop_sched_tick() a bit
[linux-2.6] / kernel / latencytop.c
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 #include <linux/latencytop.h>
13 #include <linux/kallsyms.h>
14 #include <linux/seq_file.h>
15 #include <linux/notifier.h>
16 #include <linux/spinlock.h>
17 #include <linux/proc_fs.h>
18 #include <linux/module.h>
19 #include <linux/sched.h>
20 #include <linux/list.h>
21 #include <linux/slab.h>
22 #include <linux/stacktrace.h>
23
24 static DEFINE_SPINLOCK(latency_lock);
25
26 #define MAXLR 128
27 static struct latency_record latency_record[MAXLR];
28
29 int latencytop_enabled;
30
31 void clear_all_latency_tracing(struct task_struct *p)
32 {
33         unsigned long flags;
34
35         if (!latencytop_enabled)
36                 return;
37
38         spin_lock_irqsave(&latency_lock, flags);
39         memset(&p->latency_record, 0, sizeof(p->latency_record));
40         p->latency_record_count = 0;
41         spin_unlock_irqrestore(&latency_lock, flags);
42 }
43
44 static void clear_global_latency_tracing(void)
45 {
46         unsigned long flags;
47
48         spin_lock_irqsave(&latency_lock, flags);
49         memset(&latency_record, 0, sizeof(latency_record));
50         spin_unlock_irqrestore(&latency_lock, flags);
51 }
52
53 static void __sched
54 account_global_scheduler_latency(struct task_struct *tsk, struct latency_record *lat)
55 {
56         int firstnonnull = MAXLR + 1;
57         int i;
58
59         if (!latencytop_enabled)
60                 return;
61
62         /* skip kernel threads for now */
63         if (!tsk->mm)
64                 return;
65
66         for (i = 0; i < MAXLR; i++) {
67                 int q;
68                 int same = 1;
69                 /* Nothing stored: */
70                 if (!latency_record[i].backtrace[0]) {
71                         if (firstnonnull > i)
72                                 firstnonnull = i;
73                         continue;
74                 }
75                 for (q = 0 ; q < LT_BACKTRACEDEPTH ; q++) {
76                         if (latency_record[i].backtrace[q] !=
77                                 lat->backtrace[q])
78                                 same = 0;
79                         if (same && lat->backtrace[q] == 0)
80                                 break;
81                         if (same && lat->backtrace[q] == ULONG_MAX)
82                                 break;
83                 }
84                 if (same) {
85                         latency_record[i].count++;
86                         latency_record[i].time += lat->time;
87                         if (lat->time > latency_record[i].max)
88                                 latency_record[i].max = lat->time;
89                         return;
90                 }
91         }
92
93         i = firstnonnull;
94         if (i >= MAXLR - 1)
95                 return;
96
97         /* Allocted a new one: */
98         memcpy(&latency_record[i], lat, sizeof(struct latency_record));
99 }
100
101 static inline void store_stacktrace(struct task_struct *tsk, struct latency_record *lat)
102 {
103         struct stack_trace trace;
104
105         memset(&trace, 0, sizeof(trace));
106         trace.max_entries = LT_BACKTRACEDEPTH;
107         trace.entries = &lat->backtrace[0];
108         trace.skip = 0;
109         save_stack_trace_tsk(tsk, &trace);
110 }
111
112 void __sched
113 account_scheduler_latency(struct task_struct *tsk, int usecs, int inter)
114 {
115         unsigned long flags;
116         int i, q;
117         struct latency_record lat;
118
119         if (!latencytop_enabled)
120                 return;
121
122         /* Long interruptible waits are generally user requested... */
123         if (inter && usecs > 5000)
124                 return;
125
126         memset(&lat, 0, sizeof(lat));
127         lat.count = 1;
128         lat.time = usecs;
129         lat.max = usecs;
130         store_stacktrace(tsk, &lat);
131
132         spin_lock_irqsave(&latency_lock, flags);
133
134         account_global_scheduler_latency(tsk, &lat);
135
136         /*
137          * short term hack; if we're > 32 we stop; future we recycle:
138          */
139         tsk->latency_record_count++;
140         if (tsk->latency_record_count >= LT_SAVECOUNT)
141                 goto out_unlock;
142
143         for (i = 0; i < LT_SAVECOUNT ; i++) {
144                 struct latency_record *mylat;
145                 int same = 1;
146                 mylat = &tsk->latency_record[i];
147                 for (q = 0 ; q < LT_BACKTRACEDEPTH ; q++) {
148                         if (mylat->backtrace[q] !=
149                                 lat.backtrace[q])
150                                 same = 0;
151                         if (same && lat.backtrace[q] == 0)
152                                 break;
153                         if (same && lat.backtrace[q] == ULONG_MAX)
154                                 break;
155                 }
156                 if (same) {
157                         mylat->count++;
158                         mylat->time += lat.time;
159                         if (lat.time > mylat->max)
160                                 mylat->max = lat.time;
161                         goto out_unlock;
162                 }
163         }
164
165         /* Allocated a new one: */
166         i = tsk->latency_record_count;
167         memcpy(&tsk->latency_record[i], &lat, sizeof(struct latency_record));
168
169 out_unlock:
170         spin_unlock_irqrestore(&latency_lock, flags);
171 }
172
173 static int lstats_show(struct seq_file *m, void *v)
174 {
175         int i;
176
177         seq_puts(m, "Latency Top version : v0.1\n");
178
179         for (i = 0; i < MAXLR; i++) {
180                 if (latency_record[i].backtrace[0]) {
181                         int q;
182                         seq_printf(m, "%i %li %li ",
183                                 latency_record[i].count,
184                                 latency_record[i].time,
185                                 latency_record[i].max);
186                         for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
187                                 char sym[KSYM_NAME_LEN];
188                                 char *c;
189                                 if (!latency_record[i].backtrace[q])
190                                         break;
191                                 if (latency_record[i].backtrace[q] == ULONG_MAX)
192                                         break;
193                                 sprint_symbol(sym, latency_record[i].backtrace[q]);
194                                 c = strchr(sym, '+');
195                                 if (c)
196                                         *c = 0;
197                                 seq_printf(m, "%s ", sym);
198                         }
199                         seq_printf(m, "\n");
200                 }
201         }
202         return 0;
203 }
204
205 static ssize_t
206 lstats_write(struct file *file, const char __user *buf, size_t count,
207              loff_t *offs)
208 {
209         clear_global_latency_tracing();
210
211         return count;
212 }
213
214 static int lstats_open(struct inode *inode, struct file *filp)
215 {
216         return single_open(filp, lstats_show, NULL);
217 }
218
219 static struct file_operations lstats_fops = {
220         .open           = lstats_open,
221         .read           = seq_read,
222         .write          = lstats_write,
223         .llseek         = seq_lseek,
224         .release        = single_release,
225 };
226
227 static int __init init_lstats_procfs(void)
228 {
229         struct proc_dir_entry *pe;
230
231         pe = create_proc_entry("latency_stats", 0644, NULL);
232         if (!pe)
233                 return -ENOMEM;
234
235         pe->proc_fops = &lstats_fops;
236
237         return 0;
238 }
239 __initcall(init_lstats_procfs);