2 * kernel/time/timer_stats.c
4 * Collect timer usage statistics.
6 * Copyright(C) 2006, Red Hat, Inc., Ingo Molnar
7 * Copyright(C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
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
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
25 * We export the addresses and counting of timer functions being called,
26 * the pid and cmdline from the owner process if applicable.
28 * Start/stop data collection:
29 * # echo 1[0] >/proc/timer_stats
31 * Display the information collected so far:
32 * # cat /proc/timer_stats
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.
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>
46 #include <asm/uaccess.h>
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:
68 * Number of timeout events:
73 * We save the command-line string to preserve
74 * this information past task exit:
76 char comm[TASK_COMM_LEN + 1];
78 } ____cacheline_aligned_in_smp;
81 * Spinlock protecting the tables - not taken during lookup:
83 static DEFINE_SPINLOCK(table_lock);
86 * Per-CPU lookup locks for fast hash lookup:
88 static DEFINE_PER_CPU(spinlock_t, lookup_lock);
91 * Mutex to serialize state changes with show-stats activities:
93 static DEFINE_MUTEX(show_mutex);
96 * Collection status, active/inactive:
98 static int __read_mostly active;
101 * Beginning/end timestamps of measurement:
103 static ktime_t time_start, time_stop;
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.
110 * They get freed when a new collection period is started.
112 #define MAX_ENTRIES_BITS 10
113 #define MAX_ENTRIES (1UL << MAX_ENTRIES_BITS)
115 static unsigned long nr_entries;
116 static struct entry entries[MAX_ENTRIES];
118 static atomic_t overflow_count;
120 static void reset_entries(void)
123 memset(entries, 0, sizeof(entries));
124 atomic_set(&overflow_count, 0);
127 static struct entry *alloc_entry(void)
129 if (nr_entries >= MAX_ENTRIES)
132 return entries + nr_entries++;
136 * The entries are in a hash-table, for fast lookup:
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)
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)
148 #define tstat_hashentry(entry) (tstat_hash_table + __tstat_hashfn(entry))
150 static struct entry *tstat_hash_table[TSTAT_HASH_SIZE] __read_mostly;
152 static int match_entries(struct entry *entry1, struct entry *entry2)
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;
161 * Look up whether an entry matching this item is present
162 * in the hash already. Must be called with irqs off and the
165 static struct entry *tstat_lookup(struct entry *entry, char *comm)
167 struct entry **head, *curr, *prev;
169 head = tstat_hashentry(entry);
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:
178 if (match_entries(curr, entry))
184 * Slowpath: allocate, set up and link a new hash entry:
189 spin_lock(&table_lock);
191 * Make sure we have not raced with another CPU:
194 if (match_entries(curr, entry))
201 curr = alloc_entry();
205 memcpy(curr->comm, comm, TASK_COMM_LEN);
213 spin_unlock(&table_lock);
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
226 * When the timer is already registered, then the event counter is
227 * incremented. Otherwise the timer is registered in a free slot.
229 void timer_stats_update_stats(void *timer, pid_t pid, void *startf,
230 void *timerf, char * comm)
233 * It doesnt matter which lock we take:
235 spinlock_t *lock = &per_cpu(lookup_lock, raw_smp_processor_id());
236 struct entry *entry, input;
240 input.start_func = startf;
241 input.expire_func = timerf;
244 spin_lock_irqsave(lock, flags);
248 entry = tstat_lookup(&input, comm);
252 atomic_inc(&overflow_count);
255 spin_unlock_irqrestore(lock, flags);
258 static void print_name_offset(struct seq_file *m, unsigned long addr)
260 char symname[KSYM_NAME_LEN+1];
262 if (lookup_symbol_name(addr, symname) < 0)
263 seq_printf(m, "<%p>", (void *)addr);
265 seq_printf(m, "%s", symname);
268 static int tstats_show(struct seq_file *m, void *v)
270 struct timespec period;
277 mutex_lock(&show_mutex);
279 * If still active then calculate up to now:
282 time_stop = ktime_get();
284 time = ktime_sub(time_stop, time_start);
286 period = ktime_to_timespec(time);
287 ms = period.tv_nsec / 1000000;
289 seq_puts(m, "Timer Stats Version: v0.1\n");
290 seq_printf(m, "Sample period: %ld.%03ld s\n", period.tv_sec, ms);
291 if (atomic_read(&overflow_count))
292 seq_printf(m, "Overflow: %d entries\n",
293 atomic_read(&overflow_count));
295 for (i = 0; i < nr_entries; i++) {
297 seq_printf(m, "%4lu, %5d %-16s ",
298 entry->count, entry->pid, entry->comm);
300 print_name_offset(m, (unsigned long)entry->start_func);
302 print_name_offset(m, (unsigned long)entry->expire_func);
305 events += entry->count;
308 ms += period.tv_sec * 1000;
312 if (events && period.tv_sec)
313 seq_printf(m, "%ld total events, %ld.%ld events/sec\n", events,
314 events / period.tv_sec, events * 1000 / ms);
316 seq_printf(m, "%ld total events\n", events);
318 mutex_unlock(&show_mutex);
324 * After a state change, make sure all concurrent lookup/update
325 * activities have stopped:
327 static void sync_access(void)
332 for_each_online_cpu(cpu) {
333 spin_lock_irqsave(&per_cpu(lookup_lock, cpu), flags);
335 spin_unlock_irqrestore(&per_cpu(lookup_lock, cpu), flags);
339 static ssize_t tstats_write(struct file *file, const char __user *buf,
340 size_t count, loff_t *offs)
344 if (count != 2 || *offs)
347 if (copy_from_user(ctl, buf, count))
350 mutex_lock(&show_mutex);
355 time_stop = ktime_get();
362 time_start = ktime_get();
369 mutex_unlock(&show_mutex);
374 static int tstats_open(struct inode *inode, struct file *filp)
376 return single_open(filp, tstats_show, NULL);
379 static struct file_operations tstats_fops = {
382 .write = tstats_write,
384 .release = seq_release,
387 void __init init_timer_stats(void)
391 for_each_possible_cpu(cpu)
392 spin_lock_init(&per_cpu(lookup_lock, cpu));
395 static int __init init_tstats_procfs(void)
397 struct proc_dir_entry *pe;
399 pe = create_proc_entry("timer_stats", 0644, NULL);
403 pe->proc_fops = &tstats_fops;
407 __initcall(init_tstats_procfs);