2 * linux/kernel/profile.c
3 * Simple profiling. Manages a direct-mapped profile hit count buffer,
4 * with configurable resolution, support for restricting the cpus on
5 * which profiling is done, and switching between cpu time and
6 * schedule() calls via kernel command line parameters passed at boot.
8 * Scheduler profiling support, Arjan van de Ven and Ingo Molnar,
10 * Consolidation of architecture support code for profiling,
11 * William Irwin, Oracle, July 2004
12 * Amortized hit count accounting via per-cpu open-addressed hashtables
13 * to resolve timer interrupt livelocks, William Irwin, Oracle, 2004
16 #include <linux/module.h>
17 #include <linux/profile.h>
18 #include <linux/bootmem.h>
19 #include <linux/notifier.h>
21 #include <linux/cpumask.h>
22 #include <linux/cpu.h>
23 #include <linux/highmem.h>
24 #include <linux/mutex.h>
25 #include <linux/slab.h>
26 #include <linux/vmalloc.h>
27 #include <asm/sections.h>
28 #include <asm/irq_regs.h>
29 #include <asm/ptrace.h>
34 #define PROFILE_GRPSHIFT 3
35 #define PROFILE_GRPSZ (1 << PROFILE_GRPSHIFT)
36 #define NR_PROFILE_HIT (PAGE_SIZE/sizeof(struct profile_hit))
37 #define NR_PROFILE_GRP (NR_PROFILE_HIT/PROFILE_GRPSZ)
39 /* Oprofile timer tick hook */
40 static int (*timer_hook)(struct pt_regs *) __read_mostly;
42 static atomic_t *prof_buffer;
43 static unsigned long prof_len, prof_shift;
45 int prof_on __read_mostly;
46 EXPORT_SYMBOL_GPL(prof_on);
48 static cpumask_var_t prof_cpu_mask;
50 static DEFINE_PER_CPU(struct profile_hit *[2], cpu_profile_hits);
51 static DEFINE_PER_CPU(int, cpu_profile_flip);
52 static DEFINE_MUTEX(profile_flip_mutex);
53 #endif /* CONFIG_SMP */
55 int profile_setup(char *str)
57 static char schedstr[] = "schedule";
58 static char sleepstr[] = "sleep";
59 static char kvmstr[] = "kvm";
62 if (!strncmp(str, sleepstr, strlen(sleepstr))) {
63 #ifdef CONFIG_SCHEDSTATS
64 prof_on = SLEEP_PROFILING;
65 if (str[strlen(sleepstr)] == ',')
66 str += strlen(sleepstr) + 1;
67 if (get_option(&str, &par))
70 "kernel sleep profiling enabled (shift: %ld)\n",
74 "kernel sleep profiling requires CONFIG_SCHEDSTATS\n");
75 #endif /* CONFIG_SCHEDSTATS */
76 } else if (!strncmp(str, schedstr, strlen(schedstr))) {
77 prof_on = SCHED_PROFILING;
78 if (str[strlen(schedstr)] == ',')
79 str += strlen(schedstr) + 1;
80 if (get_option(&str, &par))
83 "kernel schedule profiling enabled (shift: %ld)\n",
85 } else if (!strncmp(str, kvmstr, strlen(kvmstr))) {
86 prof_on = KVM_PROFILING;
87 if (str[strlen(kvmstr)] == ',')
88 str += strlen(kvmstr) + 1;
89 if (get_option(&str, &par))
92 "kernel KVM profiling enabled (shift: %ld)\n",
94 } else if (get_option(&str, &par)) {
96 prof_on = CPU_PROFILING;
97 printk(KERN_INFO "kernel profiling enabled (shift: %ld)\n",
102 __setup("profile=", profile_setup);
105 int __ref profile_init(void)
111 /* only text is profiled */
112 prof_len = (_etext - _stext) >> prof_shift;
113 buffer_bytes = prof_len*sizeof(atomic_t);
115 if (!alloc_cpumask_var(&prof_cpu_mask, GFP_KERNEL))
118 cpumask_copy(prof_cpu_mask, cpu_possible_mask);
120 prof_buffer = kzalloc(buffer_bytes, GFP_KERNEL);
124 prof_buffer = alloc_pages_exact(buffer_bytes, GFP_KERNEL|__GFP_ZERO);
128 prof_buffer = vmalloc(buffer_bytes);
132 free_cpumask_var(prof_cpu_mask);
136 /* Profile event notifications */
138 static BLOCKING_NOTIFIER_HEAD(task_exit_notifier);
139 static ATOMIC_NOTIFIER_HEAD(task_free_notifier);
140 static BLOCKING_NOTIFIER_HEAD(munmap_notifier);
142 void profile_task_exit(struct task_struct *task)
144 blocking_notifier_call_chain(&task_exit_notifier, 0, task);
147 int profile_handoff_task(struct task_struct *task)
150 ret = atomic_notifier_call_chain(&task_free_notifier, 0, task);
151 return (ret == NOTIFY_OK) ? 1 : 0;
154 void profile_munmap(unsigned long addr)
156 blocking_notifier_call_chain(&munmap_notifier, 0, (void *)addr);
159 int task_handoff_register(struct notifier_block *n)
161 return atomic_notifier_chain_register(&task_free_notifier, n);
163 EXPORT_SYMBOL_GPL(task_handoff_register);
165 int task_handoff_unregister(struct notifier_block *n)
167 return atomic_notifier_chain_unregister(&task_free_notifier, n);
169 EXPORT_SYMBOL_GPL(task_handoff_unregister);
171 int profile_event_register(enum profile_type type, struct notifier_block *n)
176 case PROFILE_TASK_EXIT:
177 err = blocking_notifier_chain_register(
178 &task_exit_notifier, n);
181 err = blocking_notifier_chain_register(
182 &munmap_notifier, n);
188 EXPORT_SYMBOL_GPL(profile_event_register);
190 int profile_event_unregister(enum profile_type type, struct notifier_block *n)
195 case PROFILE_TASK_EXIT:
196 err = blocking_notifier_chain_unregister(
197 &task_exit_notifier, n);
200 err = blocking_notifier_chain_unregister(
201 &munmap_notifier, n);
207 EXPORT_SYMBOL_GPL(profile_event_unregister);
209 int register_timer_hook(int (*hook)(struct pt_regs *))
216 EXPORT_SYMBOL_GPL(register_timer_hook);
218 void unregister_timer_hook(int (*hook)(struct pt_regs *))
220 WARN_ON(hook != timer_hook);
222 /* make sure all CPUs see the NULL hook */
223 synchronize_sched(); /* Allow ongoing interrupts to complete. */
225 EXPORT_SYMBOL_GPL(unregister_timer_hook);
230 * Each cpu has a pair of open-addressed hashtables for pending
231 * profile hits. read_profile() IPI's all cpus to request them
232 * to flip buffers and flushes their contents to prof_buffer itself.
233 * Flip requests are serialized by the profile_flip_mutex. The sole
234 * use of having a second hashtable is for avoiding cacheline
235 * contention that would otherwise happen during flushes of pending
236 * profile hits required for the accuracy of reported profile hits
237 * and so resurrect the interrupt livelock issue.
239 * The open-addressed hashtables are indexed by profile buffer slot
240 * and hold the number of pending hits to that profile buffer slot on
241 * a cpu in an entry. When the hashtable overflows, all pending hits
242 * are accounted to their corresponding profile buffer slots with
243 * atomic_add() and the hashtable emptied. As numerous pending hits
244 * may be accounted to a profile buffer slot in a hashtable entry,
245 * this amortizes a number of atomic profile buffer increments likely
246 * to be far larger than the number of entries in the hashtable,
247 * particularly given that the number of distinct profile buffer
248 * positions to which hits are accounted during short intervals (e.g.
249 * several seconds) is usually very small. Exclusion from buffer
250 * flipping is provided by interrupt disablement (note that for
251 * SCHED_PROFILING or SLEEP_PROFILING profile_hit() may be called from
253 * The hash function is meant to be lightweight as opposed to strong,
254 * and was vaguely inspired by ppc64 firmware-supported inverted
255 * pagetable hash functions, but uses a full hashtable full of finite
256 * collision chains, not just pairs of them.
260 static void __profile_flip_buffers(void *unused)
262 int cpu = smp_processor_id();
264 per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu);
267 static void profile_flip_buffers(void)
271 mutex_lock(&profile_flip_mutex);
272 j = per_cpu(cpu_profile_flip, get_cpu());
274 on_each_cpu(__profile_flip_buffers, NULL, 1);
275 for_each_online_cpu(cpu) {
276 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[j];
277 for (i = 0; i < NR_PROFILE_HIT; ++i) {
283 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
284 hits[i].hits = hits[i].pc = 0;
287 mutex_unlock(&profile_flip_mutex);
290 static void profile_discard_flip_buffers(void)
294 mutex_lock(&profile_flip_mutex);
295 i = per_cpu(cpu_profile_flip, get_cpu());
297 on_each_cpu(__profile_flip_buffers, NULL, 1);
298 for_each_online_cpu(cpu) {
299 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[i];
300 memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit));
302 mutex_unlock(&profile_flip_mutex);
305 void profile_hits(int type, void *__pc, unsigned int nr_hits)
307 unsigned long primary, secondary, flags, pc = (unsigned long)__pc;
309 struct profile_hit *hits;
311 if (prof_on != type || !prof_buffer)
313 pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1);
314 i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
315 secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
317 hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)];
323 * We buffer the global profiler buffer into a per-CPU
324 * queue and thus reduce the number of global (and possibly
325 * NUMA-alien) accesses. The write-queue is self-coalescing:
327 local_irq_save(flags);
329 for (j = 0; j < PROFILE_GRPSZ; ++j) {
330 if (hits[i + j].pc == pc) {
331 hits[i + j].hits += nr_hits;
333 } else if (!hits[i + j].hits) {
335 hits[i + j].hits = nr_hits;
339 i = (i + secondary) & (NR_PROFILE_HIT - 1);
340 } while (i != primary);
343 * Add the current hit(s) and flush the write-queue out
344 * to the global buffer:
346 atomic_add(nr_hits, &prof_buffer[pc]);
347 for (i = 0; i < NR_PROFILE_HIT; ++i) {
348 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
349 hits[i].pc = hits[i].hits = 0;
352 local_irq_restore(flags);
356 static int __cpuinit profile_cpu_callback(struct notifier_block *info,
357 unsigned long action, void *__cpu)
359 int node, cpu = (unsigned long)__cpu;
364 case CPU_UP_PREPARE_FROZEN:
365 node = cpu_to_node(cpu);
366 per_cpu(cpu_profile_flip, cpu) = 0;
367 if (!per_cpu(cpu_profile_hits, cpu)[1]) {
368 page = alloc_pages_exact_node(node,
369 GFP_KERNEL | __GFP_ZERO,
373 per_cpu(cpu_profile_hits, cpu)[1] = page_address(page);
375 if (!per_cpu(cpu_profile_hits, cpu)[0]) {
376 page = alloc_pages_exact_node(node,
377 GFP_KERNEL | __GFP_ZERO,
381 per_cpu(cpu_profile_hits, cpu)[0] = page_address(page);
385 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
386 per_cpu(cpu_profile_hits, cpu)[1] = NULL;
390 case CPU_ONLINE_FROZEN:
391 if (prof_cpu_mask != NULL)
392 cpumask_set_cpu(cpu, prof_cpu_mask);
394 case CPU_UP_CANCELED:
395 case CPU_UP_CANCELED_FROZEN:
397 case CPU_DEAD_FROZEN:
398 if (prof_cpu_mask != NULL)
399 cpumask_clear_cpu(cpu, prof_cpu_mask);
400 if (per_cpu(cpu_profile_hits, cpu)[0]) {
401 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
402 per_cpu(cpu_profile_hits, cpu)[0] = NULL;
405 if (per_cpu(cpu_profile_hits, cpu)[1]) {
406 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
407 per_cpu(cpu_profile_hits, cpu)[1] = NULL;
414 #else /* !CONFIG_SMP */
415 #define profile_flip_buffers() do { } while (0)
416 #define profile_discard_flip_buffers() do { } while (0)
417 #define profile_cpu_callback NULL
419 void profile_hits(int type, void *__pc, unsigned int nr_hits)
423 if (prof_on != type || !prof_buffer)
425 pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
426 atomic_add(nr_hits, &prof_buffer[min(pc, prof_len - 1)]);
428 #endif /* !CONFIG_SMP */
429 EXPORT_SYMBOL_GPL(profile_hits);
431 void profile_tick(int type)
433 struct pt_regs *regs = get_irq_regs();
435 if (type == CPU_PROFILING && timer_hook)
437 if (!user_mode(regs) && prof_cpu_mask != NULL &&
438 cpumask_test_cpu(smp_processor_id(), prof_cpu_mask))
439 profile_hit(type, (void *)profile_pc(regs));
442 #ifdef CONFIG_PROC_FS
443 #include <linux/proc_fs.h>
444 #include <asm/uaccess.h>
446 static int prof_cpu_mask_read_proc(char *page, char **start, off_t off,
447 int count, int *eof, void *data)
449 int len = cpumask_scnprintf(page, count, data);
452 len += sprintf(page + len, "\n");
456 static int prof_cpu_mask_write_proc(struct file *file,
457 const char __user *buffer, unsigned long count, void *data)
459 struct cpumask *mask = data;
460 unsigned long full_count = count, err;
461 cpumask_var_t new_value;
463 if (!alloc_cpumask_var(&new_value, GFP_KERNEL))
466 err = cpumask_parse_user(buffer, count, new_value);
468 cpumask_copy(mask, new_value);
471 free_cpumask_var(new_value);
475 void create_prof_cpu_mask(struct proc_dir_entry *root_irq_dir)
477 struct proc_dir_entry *entry;
479 /* create /proc/irq/prof_cpu_mask */
480 entry = create_proc_entry("prof_cpu_mask", 0600, root_irq_dir);
483 entry->data = prof_cpu_mask;
484 entry->read_proc = prof_cpu_mask_read_proc;
485 entry->write_proc = prof_cpu_mask_write_proc;
489 * This function accesses profiling information. The returned data is
490 * binary: the sampling step and the actual contents of the profile
491 * buffer. Use of the program readprofile is recommended in order to
492 * get meaningful info out of these data.
495 read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos)
497 unsigned long p = *ppos;
500 unsigned int sample_step = 1 << prof_shift;
502 profile_flip_buffers();
503 if (p >= (prof_len+1)*sizeof(unsigned int))
505 if (count > (prof_len+1)*sizeof(unsigned int) - p)
506 count = (prof_len+1)*sizeof(unsigned int) - p;
509 while (p < sizeof(unsigned int) && count > 0) {
510 if (put_user(*((char *)(&sample_step)+p), buf))
512 buf++; p++; count--; read++;
514 pnt = (char *)prof_buffer + p - sizeof(atomic_t);
515 if (copy_to_user(buf, (void *)pnt, count))
523 * Writing to /proc/profile resets the counters
525 * Writing a 'profiling multiplier' value into it also re-sets the profiling
526 * interrupt frequency, on architectures that support this.
528 static ssize_t write_profile(struct file *file, const char __user *buf,
529 size_t count, loff_t *ppos)
532 extern int setup_profiling_timer(unsigned int multiplier);
534 if (count == sizeof(int)) {
535 unsigned int multiplier;
537 if (copy_from_user(&multiplier, buf, sizeof(int)))
540 if (setup_profiling_timer(multiplier))
544 profile_discard_flip_buffers();
545 memset(prof_buffer, 0, prof_len * sizeof(atomic_t));
549 static const struct file_operations proc_profile_operations = {
550 .read = read_profile,
551 .write = write_profile,
555 static void profile_nop(void *unused)
559 static int create_hash_tables(void)
563 for_each_online_cpu(cpu) {
564 int node = cpu_to_node(cpu);
567 page = alloc_pages_exact_node(node,
568 GFP_KERNEL | __GFP_ZERO | GFP_THISNODE,
572 per_cpu(cpu_profile_hits, cpu)[1]
573 = (struct profile_hit *)page_address(page);
574 page = alloc_pages_exact_node(node,
575 GFP_KERNEL | __GFP_ZERO | GFP_THISNODE,
579 per_cpu(cpu_profile_hits, cpu)[0]
580 = (struct profile_hit *)page_address(page);
586 on_each_cpu(profile_nop, NULL, 1);
587 for_each_online_cpu(cpu) {
590 if (per_cpu(cpu_profile_hits, cpu)[0]) {
591 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
592 per_cpu(cpu_profile_hits, cpu)[0] = NULL;
595 if (per_cpu(cpu_profile_hits, cpu)[1]) {
596 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
597 per_cpu(cpu_profile_hits, cpu)[1] = NULL;
604 #define create_hash_tables() ({ 0; })
607 int __ref create_proc_profile(void) /* false positive from hotcpu_notifier */
609 struct proc_dir_entry *entry;
613 if (create_hash_tables())
615 entry = proc_create("profile", S_IWUSR | S_IRUGO,
616 NULL, &proc_profile_operations);
619 entry->size = (1+prof_len) * sizeof(atomic_t);
620 hotcpu_notifier(profile_cpu_callback, 0);
623 module_init(create_proc_profile);
624 #endif /* CONFIG_PROC_FS */