kernel/time/timer_stats.c

   1 /*
   2  * kernel/time/timer_stats.c
   3  *
   4  * Collect timer usage statistics.
   5  *
   6  * Copyright(C) 2006, Red Hat, Inc., Ingo Molnar
   7  * Copyright(C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
   8  *
   9  * timer_stats is based on timer_top, a similar functionality which was part of
  10  * Con Kolivas dyntick patch set. It was developed by Daniel Petrini at the
  11  * Instituto Nokia de Tecnologia - INdT - Manaus. timer_top's design was based
  12  * on dynamic allocation of the statistics entries and linear search based
  13  * lookup combined with a global lock, rather than the static array, hash
  14  * and per-CPU locking which is used by timer_stats. It was written for the
  15  * pre hrtimer kernel code and therefore did not take hrtimers into account.
  16  * Nevertheless it provided the base for the timer_stats implementation and
  17  * was a helpful source of inspiration. Kudos to Daniel and the Nokia folks
  18  * for this effort.
  19  *
  20  * timer_top.c is
  21  *      Copyright (C) 2005 Instituto Nokia de Tecnologia - INdT - Manaus
  22  *      Written by Daniel Petrini <d.pensator@gmail.com>
  23  *      timer_top.c was released under the GNU General Public License version 2
  24  *
  25  * We export the addresses and counting of timer functions being called,
  26  * the pid and cmdline from the owner process if applicable.
  27  *
  28  * Start/stop data collection:
  29  * # echo 1[0] >/proc/timer_stats
  30  *
  31  * Display the information collected so far:
  32  * # cat /proc/timer_stats
  33  *
  34  * This program is free software; you can redistribute it and/or modify
  35  * it under the terms of the GNU General Public License version 2 as
  36  * published by the Free Software Foundation.
  37  */
  38
  39 #include <linux/proc_fs.h>
  40 #include <linux/module.h>
  41 #include <linux/spinlock.h>
  42 #include <linux/sched.h>
  43 #include <linux/seq_file.h>
  44 #include <linux/kallsyms.h>
  45
  46 #include <asm/uaccess.h>
  47
  48 /*
  49  * This is our basic unit of interest: a timer expiry event identified
  50  * by the timer, its start/expire functions and the PID of the task that
  51  * started the timer. We count the number of times an event happens:
  52  */
  53 struct entry {
  54         /*
  55          * Hash list:
  56          */
  57         struct entry            *next;
  58
  59         /*
  60          * Hash keys:
  61          */
  62         void                    *timer;
  63         void                    *start_func;
  64         void                    *expire_func;
  65         pid_t                   pid;
  66
  67         /*
  68          * Number of timeout events:
  69          */
  70         unsigned long           count;
  71
  72         /*
  73          * We save the command-line string to preserve
  74          * this information past task exit:
  75          */
  76         char                    comm[TASK_COMM_LEN + 1];
  77
  78 } ____cacheline_aligned_in_smp;
  79
  80 /*
  81  * Spinlock protecting the tables - not taken during lookup:
  82  */
  83 static DEFINE_SPINLOCK(table_lock);
  84
  85 /*
  86  * Per-CPU lookup locks for fast hash lookup:
  87  */
  88 static DEFINE_PER_CPU(spinlock_t, lookup_lock);
  89
  90 /*
  91  * Mutex to serialize state changes with show-stats activities:
  92  */
  93 static DEFINE_MUTEX(show_mutex);
  94
  95 /*
  96  * Collection status, active/inactive:
  97  */
  98 static int __read_mostly active;
  99
 100 /*
 101  * Beginning/end timestamps of measurement:
 102  */
 103 static ktime_t time_start, time_stop;
 104
 105 /*
 106  * tstat entry structs only get allocated while collection is
 107  * active and never freed during that time - this simplifies
 108  * things quite a bit.
 109  *
 110  * They get freed when a new collection period is started.
 111  */
 112 #define MAX_ENTRIES_BITS        10
 113 #define MAX_ENTRIES             (1UL << MAX_ENTRIES_BITS)
 114
 115 static unsigned long nr_entries;
 116 static struct entry entries[MAX_ENTRIES];
 117
 118 static atomic_t overflow_count;
 119
 120 static void reset_entries(void)
 121 {
 122         nr_entries = 0;
 123         memset(entries, 0, sizeof(entries));
 124         atomic_set(&overflow_count, 0);
 125 }
 126
 127 static struct entry *alloc_entry(void)
 128 {
 129         if (nr_entries >= MAX_ENTRIES)
 130                 return NULL;
 131
 132         return entries + nr_entries++;
 133 }
 134
 135 /*
 136  * The entries are in a hash-table, for fast lookup:
 137  */
 138 #define TSTAT_HASH_BITS         (MAX_ENTRIES_BITS - 1)
 139 #define TSTAT_HASH_SIZE         (1UL << TSTAT_HASH_BITS)
 140 #define TSTAT_HASH_MASK         (TSTAT_HASH_SIZE - 1)
 141
 142 #define __tstat_hashfn(entry)                                           \
 143         (((unsigned long)(entry)->timer       ^                         \
 144           (unsigned long)(entry)->start_func  ^                         \
 145           (unsigned long)(entry)->expire_func ^                         \
 146           (unsigned long)(entry)->pid           ) & TSTAT_HASH_MASK)
 147
 148 #define tstat_hashentry(entry)  (tstat_hash_table + __tstat_hashfn(entry))
 149
 150 static struct entry *tstat_hash_table[TSTAT_HASH_SIZE] __read_mostly;
 151
 152 static int match_entries(struct entry *entry1, struct entry *entry2)
 153 {
 154         return entry1->timer       == entry2->timer       &&
 155                entry1->start_func  == entry2->start_func  &&
 156                entry1->expire_func == entry2->expire_func &&
 157                entry1->pid         == entry2->pid;
 158 }
 159
 160 /*
 161  * Look up whether an entry matching this item is present
 162  * in the hash already. Must be called with irqs off and the
 163  * lookup lock held:
 164  */
 165 static struct entry *tstat_lookup(struct entry *entry, char *comm)
 166 {
 167         struct entry **head, *curr, *prev;
 168
 169         head = tstat_hashentry(entry);
 170         curr = *head;
 171
 172         /*
 173          * The fastpath is when the entry is already hashed,
 174          * we do this with the lookup lock held, but with the
 175          * table lock not held:
 176          */
 177         while (curr) {
 178                 if (match_entries(curr, entry))
 179                         return curr;
 180
 181                 curr = curr->next;
 182         }
 183         /*
 184          * Slowpath: allocate, set up and link a new hash entry:
 185          */
 186         prev = NULL;
 187         curr = *head;
 188
 189         spin_lock(&table_lock);
 190         /*
 191          * Make sure we have not raced with another CPU:
 192          */
 193         while (curr) {
 194                 if (match_entries(curr, entry))
 195                         goto out_unlock;
 196
 197                 prev = curr;
 198                 curr = curr->next;
 199         }
 200
 201         curr = alloc_entry();
 202         if (curr) {
 203                 *curr = *entry;
 204                 curr->count = 0;
 205                 memcpy(curr->comm, comm, TASK_COMM_LEN);
 206                 if (prev)
 207                         prev->next = curr;
 208                 else
 209                         *head = curr;
 210                 curr->next = NULL;
 211         }
 212  out_unlock:
 213         spin_unlock(&table_lock);
 214
 215         return curr;
 216 }
 217
 218 /**
 219  * timer_stats_update_stats - Update the statistics for a timer.
 220  * @timer:      pointer to either a timer_list or a hrtimer
 221  * @pid:        the pid of the task which set up the timer
 222  * @startf:     pointer to the function which did the timer setup
 223  * @timerf:     pointer to the timer callback function of the timer
 224  * @comm:       name of the process which set up the timer
 225  *
 226  * When the timer is already registered, then the event counter is
 227  * incremented. Otherwise the timer is registered in a free slot.
 228  */
 229 void timer_stats_update_stats(void *timer, pid_t pid, void *startf,
 230                               void *timerf, char * comm)
 231 {
 232         /*
 233          * It doesnt matter which lock we take:
 234          */
 235         spinlock_t *lock = &per_cpu(lookup_lock, raw_smp_processor_id());
 236         struct entry *entry, input;
 237         unsigned long flags;
 238
 239         input.timer = timer;
 240         input.start_func = startf;
 241         input.expire_func = timerf;
 242         input.pid = pid;
 243
 244         spin_lock_irqsave(lock, flags);
 245         if (!active)
 246                 goto out_unlock;
 247
 248         entry = tstat_lookup(&input, comm);
 249         if (likely(entry))
 250                 entry->count++;
 251         else
 252                 atomic_inc(&overflow_count);
 253
 254  out_unlock:
 255         spin_unlock_irqrestore(lock, flags);
 256 }
 257
 258 static void print_name_offset(struct seq_file *m, unsigned long addr)
 259 {
 260         char namebuf[KSYM_NAME_LEN+1];
 261         unsigned long size, offset;
 262         const char *sym_name;
 263         char *modname;
 264
 265         sym_name = kallsyms_lookup(addr, &size, &offset, &modname, namebuf);
 266         if (sym_name)
 267                 seq_printf(m, "%s", sym_name);
 268         else
 269                 seq_printf(m, "<%p>", (void *)addr);
 270 }
 271
 272 static int tstats_show(struct seq_file *m, void *v)
 273 {
 274         struct timespec period;
 275         struct entry *entry;
 276         unsigned long ms;
 277         long events = 0;
 278         ktime_t time;
 279         int i;
 280
 281         mutex_lock(&show_mutex);
 282         /*
 283          * If still active then calculate up to now:
 284          */
 285         if (active)
 286                 time_stop = ktime_get();
 287
 288         time = ktime_sub(time_stop, time_start);
 289
 290         period = ktime_to_timespec(time);
 291         ms = period.tv_nsec / 1000000;
 292
 293         seq_puts(m, "Timer Stats Version: v0.1\n");
 294         seq_printf(m, "Sample period: %ld.%03ld s\n", period.tv_sec, ms);
 295         if (atomic_read(&overflow_count))
 296                 seq_printf(m, "Overflow: %d entries\n",
 297                         atomic_read(&overflow_count));
 298
 299         for (i = 0; i < nr_entries; i++) {
 300                 entry = entries + i;
 301                 seq_printf(m, "%4lu, %5d %-16s ",
 302                                 entry->count, entry->pid, entry->comm);
 303
 304                 print_name_offset(m, (unsigned long)entry->start_func);
 305                 seq_puts(m, " (");
 306                 print_name_offset(m, (unsigned long)entry->expire_func);
 307                 seq_puts(m, ")\n");
 308
 309                 events += entry->count;
 310         }
 311
 312         ms += period.tv_sec * 1000;
 313         if (!ms)
 314                 ms = 1;
 315
 316         if (events && period.tv_sec)
 317                 seq_printf(m, "%ld total events, %ld.%ld events/sec\n", events,
 318                            events / period.tv_sec, events * 1000 / ms);
 319         else
 320                 seq_printf(m, "%ld total events\n", events);
 321
 322         mutex_unlock(&show_mutex);
 323
 324         return 0;
 325 }
 326
 327 /*
 328  * After a state change, make sure all concurrent lookup/update
 329  * activities have stopped:
 330  */
 331 static void sync_access(void)
 332 {
 333         unsigned long flags;
 334         int cpu;
 335
 336         for_each_online_cpu(cpu) {
 337                 spin_lock_irqsave(&per_cpu(lookup_lock, cpu), flags);
 338                 /* nothing */
 339                 spin_unlock_irqrestore(&per_cpu(lookup_lock, cpu), flags);
 340         }
 341 }
 342
 343 static ssize_t tstats_write(struct file *file, const char __user *buf,
 344                             size_t count, loff_t *offs)
 345 {
 346         char ctl[2];
 347
 348         if (count != 2 || *offs)
 349                 return -EINVAL;
 350
 351         if (copy_from_user(ctl, buf, count))
 352                 return -EFAULT;
 353
 354         mutex_lock(&show_mutex);
 355         switch (ctl[0]) {
 356         case '0':
 357                 if (active) {
 358                         active = 0;
 359                         time_stop = ktime_get();
 360                         sync_access();
 361                 }
 362                 break;
 363         case '1':
 364                 if (!active) {
 365                         reset_entries();
 366                         time_start = ktime_get();
 367                         active = 1;
 368                 }
 369                 break;
 370         default:
 371                 count = -EINVAL;
 372         }
 373         mutex_unlock(&show_mutex);
 374
 375         return count;
 376 }
 377
 378 static int tstats_open(struct inode *inode, struct file *filp)
 379 {
 380         return single_open(filp, tstats_show, NULL);
 381 }
 382
 383 static struct file_operations tstats_fops = {
 384         .open           = tstats_open,
 385         .read           = seq_read,
 386         .write          = tstats_write,
 387         .llseek         = seq_lseek,
 388         .release        = seq_release,
 389 };
 390
 391 void __init init_timer_stats(void)
 392 {
 393         int cpu;
 394
 395         for_each_possible_cpu(cpu)
 396                 spin_lock_init(&per_cpu(lookup_lock, cpu));
 397 }
 398
 399 static int __init init_tstats_procfs(void)
 400 {
 401         struct proc_dir_entry *pe;
 402
 403         pe = create_proc_entry("timer_stats", 0644, NULL);
 404         if (!pe)
 405                 return -ENOMEM;
 406
 407         pe->proc_fops = &tstats_fops;
 408
 409         return 0;
 410 }
 411 __initcall(init_tstats_procfs);