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