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, same = 1;
  68
  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                         unsigned long record = lat->backtrace[q];
  77
  78                         if (latency_record[i].backtrace[q] != record) {
  79                                 same = 0;
  80                                 break;
  81                         }
  82
  83                         /* 0 and ULONG_MAX entries mean end of backtrace: */
  84                         if (record == 0 || record == ULONG_MAX)
  85                                 break;
  86                 }
  87                 if (same) {
  88                         latency_record[i].count++;
  89                         latency_record[i].time += lat->time;
  90                         if (lat->time > latency_record[i].max)
  91                                 latency_record[i].max = lat->time;
  92                         return;
  93                 }
  94         }
  95
  96         i = firstnonnull;
  97         if (i >= MAXLR - 1)
  98                 return;
  99
 100         /* Allocted a new one: */
 101         memcpy(&latency_record[i], lat, sizeof(struct latency_record));
 102 }
 103
 104 static inline void store_stacktrace(struct task_struct *tsk, struct latency_record *lat)
 105 {
 106         struct stack_trace trace;
 107
 108         memset(&trace, 0, sizeof(trace));
 109         trace.max_entries = LT_BACKTRACEDEPTH;
 110         trace.entries = &lat->backtrace[0];
 111         trace.skip = 0;
 112         save_stack_trace_tsk(tsk, &trace);
 113 }
 114
 115 void __sched
 116 account_scheduler_latency(struct task_struct *tsk, int usecs, int inter)
 117 {
 118         unsigned long flags;
 119         int i, q;
 120         struct latency_record lat;
 121
 122         if (!latencytop_enabled)
 123                 return;
 124
 125         /* Long interruptible waits are generally user requested... */
 126         if (inter && usecs > 5000)
 127                 return;
 128
 129         memset(&lat, 0, sizeof(lat));
 130         lat.count = 1;
 131         lat.time = usecs;
 132         lat.max = usecs;
 133         store_stacktrace(tsk, &lat);
 134
 135         spin_lock_irqsave(&latency_lock, flags);
 136
 137         account_global_scheduler_latency(tsk, &lat);
 138
 139         /*
 140          * short term hack; if we're > 32 we stop; future we recycle:
 141          */
 142         tsk->latency_record_count++;
 143         if (tsk->latency_record_count >= LT_SAVECOUNT)
 144                 goto out_unlock;
 145
 146         for (i = 0; i < LT_SAVECOUNT ; i++) {
 147                 struct latency_record *mylat;
 148                 int same = 1;
 149
 150                 mylat = &tsk->latency_record[i];
 151                 for (q = 0 ; q < LT_BACKTRACEDEPTH ; q++) {
 152                         unsigned long record = lat.backtrace[q];
 153
 154                         if (mylat->backtrace[q] != record) {
 155                                 same = 0;
 156                                 break;
 157                         }
 158
 159                         /* 0 and ULONG_MAX entries mean end of backtrace: */
 160                         if (record == 0 || record == ULONG_MAX)
 161                                 break;
 162                 }
 163                 if (same) {
 164                         mylat->count++;
 165                         mylat->time += lat.time;
 166                         if (lat.time > mylat->max)
 167                                 mylat->max = lat.time;
 168                         goto out_unlock;
 169                 }
 170         }
 171
 172         /* Allocated a new one: */
 173         i = tsk->latency_record_count;
 174         memcpy(&tsk->latency_record[i], &lat, sizeof(struct latency_record));
 175
 176 out_unlock:
 177         spin_unlock_irqrestore(&latency_lock, flags);
 178 }
 179
 180 static int lstats_show(struct seq_file *m, void *v)
 181 {
 182         int i;
 183
 184         seq_puts(m, "Latency Top version : v0.1\n");
 185
 186         for (i = 0; i < MAXLR; i++) {
 187                 if (latency_record[i].backtrace[0]) {
 188                         int q;
 189                         seq_printf(m, "%i %li %li ",
 190                                 latency_record[i].count,
 191                                 latency_record[i].time,
 192                                 latency_record[i].max);
 193                         for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
 194                                 char sym[KSYM_SYMBOL_LEN];
 195                                 char *c;
 196                                 if (!latency_record[i].backtrace[q])
 197                                         break;
 198                                 if (latency_record[i].backtrace[q] == ULONG_MAX)
 199                                         break;
 200                                 sprint_symbol(sym, latency_record[i].backtrace[q]);
 201                                 c = strchr(sym, '+');
 202                                 if (c)
 203                                         *c = 0;
 204                                 seq_printf(m, "%s ", sym);
 205                         }
 206                         seq_printf(m, "\n");
 207                 }
 208         }
 209         return 0;
 210 }
 211
 212 static ssize_t
 213 lstats_write(struct file *file, const char __user *buf, size_t count,
 214              loff_t *offs)
 215 {
 216         clear_global_latency_tracing();
 217
 218         return count;
 219 }
 220
 221 static int lstats_open(struct inode *inode, struct file *filp)
 222 {
 223         return single_open(filp, lstats_show, NULL);
 224 }
 225
 226 static struct file_operations lstats_fops = {
 227         .open           = lstats_open,
 228         .read           = seq_read,
 229         .write          = lstats_write,
 230         .llseek         = seq_lseek,
 231         .release        = single_release,
 232 };
 233
 234 static int __init init_lstats_procfs(void)
 235 {
 236         proc_create("latency_stats", 0644, NULL, &lstats_fops);
 237         return 0;
 238 }
 239 __initcall(init_lstats_procfs);