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 * Nadia Yvette Chambers, Oracle, July 2004
12 * Amortized hit count accounting via per-cpu open-addressed hashtables
13 * to resolve timer interrupt livelocks, Nadia Yvette Chambers,
17 #include <linux/export.h>
18 #include <linux/profile.h>
19 #include <linux/bootmem.h>
20 #include <linux/notifier.h>
22 #include <linux/cpumask.h>
23 #include <linux/cpu.h>
24 #include <linux/highmem.h>
25 #include <linux/mutex.h>
26 #include <linux/slab.h>
27 #include <linux/vmalloc.h>
28 #include <asm/sections.h>
29 #include <asm/irq_regs.h>
30 #include <asm/ptrace.h>
35 #define PROFILE_GRPSHIFT 3
36 #define PROFILE_GRPSZ (1 << PROFILE_GRPSHIFT)
37 #define NR_PROFILE_HIT (PAGE_SIZE/sizeof(struct profile_hit))
38 #define NR_PROFILE_GRP (NR_PROFILE_HIT/PROFILE_GRPSZ)
40 static atomic_t *prof_buffer;
41 static unsigned long prof_len, prof_shift;
43 int prof_on __read_mostly;
44 EXPORT_SYMBOL_GPL(prof_on);
46 static cpumask_var_t prof_cpu_mask;
47 #if defined(CONFIG_SMP) && defined(CONFIG_PROC_FS)
48 static DEFINE_PER_CPU(struct profile_hit *[2], cpu_profile_hits);
49 static DEFINE_PER_CPU(int, cpu_profile_flip);
50 static DEFINE_MUTEX(profile_flip_mutex);
51 #endif /* CONFIG_SMP */
53 int profile_setup(char *str)
55 static const char schedstr[] = "schedule";
56 static const char sleepstr[] = "sleep";
57 static const char kvmstr[] = "kvm";
60 if (!strncmp(str, sleepstr, strlen(sleepstr))) {
61 #ifdef CONFIG_SCHEDSTATS
62 force_schedstat_enabled();
63 prof_on = SLEEP_PROFILING;
64 if (str[strlen(sleepstr)] == ',')
65 str += strlen(sleepstr) + 1;
66 if (get_option(&str, &par))
68 pr_info("kernel sleep profiling enabled (shift: %ld)\n",
71 pr_warn("kernel sleep profiling requires CONFIG_SCHEDSTATS\n");
72 #endif /* CONFIG_SCHEDSTATS */
73 } else if (!strncmp(str, schedstr, strlen(schedstr))) {
74 prof_on = SCHED_PROFILING;
75 if (str[strlen(schedstr)] == ',')
76 str += strlen(schedstr) + 1;
77 if (get_option(&str, &par))
79 pr_info("kernel schedule profiling enabled (shift: %ld)\n",
81 } else if (!strncmp(str, kvmstr, strlen(kvmstr))) {
82 prof_on = KVM_PROFILING;
83 if (str[strlen(kvmstr)] == ',')
84 str += strlen(kvmstr) + 1;
85 if (get_option(&str, &par))
87 pr_info("kernel KVM profiling enabled (shift: %ld)\n",
89 } else if (get_option(&str, &par)) {
91 prof_on = CPU_PROFILING;
92 pr_info("kernel profiling enabled (shift: %ld)\n",
97 __setup("profile=", profile_setup);
100 int __ref profile_init(void)
106 /* only text is profiled */
107 prof_len = (_etext - _stext) >> prof_shift;
108 buffer_bytes = prof_len*sizeof(atomic_t);
110 if (!alloc_cpumask_var(&prof_cpu_mask, GFP_KERNEL))
113 cpumask_copy(prof_cpu_mask, cpu_possible_mask);
115 prof_buffer = kzalloc(buffer_bytes, GFP_KERNEL|__GFP_NOWARN);
119 prof_buffer = alloc_pages_exact(buffer_bytes,
120 GFP_KERNEL|__GFP_ZERO|__GFP_NOWARN);
124 prof_buffer = vzalloc(buffer_bytes);
128 free_cpumask_var(prof_cpu_mask);
132 /* Profile event notifications */
134 static BLOCKING_NOTIFIER_HEAD(task_exit_notifier);
135 static ATOMIC_NOTIFIER_HEAD(task_free_notifier);
136 static BLOCKING_NOTIFIER_HEAD(munmap_notifier);
138 void profile_task_exit(struct task_struct *task)
140 blocking_notifier_call_chain(&task_exit_notifier, 0, task);
143 int profile_handoff_task(struct task_struct *task)
146 ret = atomic_notifier_call_chain(&task_free_notifier, 0, task);
147 return (ret == NOTIFY_OK) ? 1 : 0;
150 void profile_munmap(unsigned long addr)
152 blocking_notifier_call_chain(&munmap_notifier, 0, (void *)addr);
155 int task_handoff_register(struct notifier_block *n)
157 return atomic_notifier_chain_register(&task_free_notifier, n);
159 EXPORT_SYMBOL_GPL(task_handoff_register);
161 int task_handoff_unregister(struct notifier_block *n)
163 return atomic_notifier_chain_unregister(&task_free_notifier, n);
165 EXPORT_SYMBOL_GPL(task_handoff_unregister);
167 int profile_event_register(enum profile_type type, struct notifier_block *n)
172 case PROFILE_TASK_EXIT:
173 err = blocking_notifier_chain_register(
174 &task_exit_notifier, n);
177 err = blocking_notifier_chain_register(
178 &munmap_notifier, n);
184 EXPORT_SYMBOL_GPL(profile_event_register);
186 int profile_event_unregister(enum profile_type type, struct notifier_block *n)
191 case PROFILE_TASK_EXIT:
192 err = blocking_notifier_chain_unregister(
193 &task_exit_notifier, n);
196 err = blocking_notifier_chain_unregister(
197 &munmap_notifier, n);
203 EXPORT_SYMBOL_GPL(profile_event_unregister);
205 #if defined(CONFIG_SMP) && defined(CONFIG_PROC_FS)
207 * Each cpu has a pair of open-addressed hashtables for pending
208 * profile hits. read_profile() IPI's all cpus to request them
209 * to flip buffers and flushes their contents to prof_buffer itself.
210 * Flip requests are serialized by the profile_flip_mutex. The sole
211 * use of having a second hashtable is for avoiding cacheline
212 * contention that would otherwise happen during flushes of pending
213 * profile hits required for the accuracy of reported profile hits
214 * and so resurrect the interrupt livelock issue.
216 * The open-addressed hashtables are indexed by profile buffer slot
217 * and hold the number of pending hits to that profile buffer slot on
218 * a cpu in an entry. When the hashtable overflows, all pending hits
219 * are accounted to their corresponding profile buffer slots with
220 * atomic_add() and the hashtable emptied. As numerous pending hits
221 * may be accounted to a profile buffer slot in a hashtable entry,
222 * this amortizes a number of atomic profile buffer increments likely
223 * to be far larger than the number of entries in the hashtable,
224 * particularly given that the number of distinct profile buffer
225 * positions to which hits are accounted during short intervals (e.g.
226 * several seconds) is usually very small. Exclusion from buffer
227 * flipping is provided by interrupt disablement (note that for
228 * SCHED_PROFILING or SLEEP_PROFILING profile_hit() may be called from
230 * The hash function is meant to be lightweight as opposed to strong,
231 * and was vaguely inspired by ppc64 firmware-supported inverted
232 * pagetable hash functions, but uses a full hashtable full of finite
233 * collision chains, not just pairs of them.
237 static void __profile_flip_buffers(void *unused)
239 int cpu = smp_processor_id();
241 per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu);
244 static void profile_flip_buffers(void)
248 mutex_lock(&profile_flip_mutex);
249 j = per_cpu(cpu_profile_flip, get_cpu());
251 on_each_cpu(__profile_flip_buffers, NULL, 1);
252 for_each_online_cpu(cpu) {
253 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[j];
254 for (i = 0; i < NR_PROFILE_HIT; ++i) {
260 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
261 hits[i].hits = hits[i].pc = 0;
264 mutex_unlock(&profile_flip_mutex);
267 static void profile_discard_flip_buffers(void)
271 mutex_lock(&profile_flip_mutex);
272 i = 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)[i];
277 memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit));
279 mutex_unlock(&profile_flip_mutex);
282 static void do_profile_hits(int type, void *__pc, unsigned int nr_hits)
284 unsigned long primary, secondary, flags, pc = (unsigned long)__pc;
286 struct profile_hit *hits;
288 pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1);
289 i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
290 secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
292 hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)];
298 * We buffer the global profiler buffer into a per-CPU
299 * queue and thus reduce the number of global (and possibly
300 * NUMA-alien) accesses. The write-queue is self-coalescing:
302 local_irq_save(flags);
304 for (j = 0; j < PROFILE_GRPSZ; ++j) {
305 if (hits[i + j].pc == pc) {
306 hits[i + j].hits += nr_hits;
308 } else if (!hits[i + j].hits) {
310 hits[i + j].hits = nr_hits;
314 i = (i + secondary) & (NR_PROFILE_HIT - 1);
315 } while (i != primary);
318 * Add the current hit(s) and flush the write-queue out
319 * to the global buffer:
321 atomic_add(nr_hits, &prof_buffer[pc]);
322 for (i = 0; i < NR_PROFILE_HIT; ++i) {
323 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
324 hits[i].pc = hits[i].hits = 0;
327 local_irq_restore(flags);
331 static int profile_cpu_callback(struct notifier_block *info,
332 unsigned long action, void *__cpu)
334 int node, cpu = (unsigned long)__cpu;
339 case CPU_UP_PREPARE_FROZEN:
340 node = cpu_to_mem(cpu);
341 per_cpu(cpu_profile_flip, cpu) = 0;
342 if (!per_cpu(cpu_profile_hits, cpu)[1]) {
343 page = __alloc_pages_node(node,
344 GFP_KERNEL | __GFP_ZERO,
347 return notifier_from_errno(-ENOMEM);
348 per_cpu(cpu_profile_hits, cpu)[1] = page_address(page);
350 if (!per_cpu(cpu_profile_hits, cpu)[0]) {
351 page = __alloc_pages_node(node,
352 GFP_KERNEL | __GFP_ZERO,
356 per_cpu(cpu_profile_hits, cpu)[0] = page_address(page);
360 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
361 per_cpu(cpu_profile_hits, cpu)[1] = NULL;
363 return notifier_from_errno(-ENOMEM);
365 case CPU_ONLINE_FROZEN:
366 if (prof_cpu_mask != NULL)
367 cpumask_set_cpu(cpu, prof_cpu_mask);
369 case CPU_UP_CANCELED:
370 case CPU_UP_CANCELED_FROZEN:
372 case CPU_DEAD_FROZEN:
373 if (prof_cpu_mask != NULL)
374 cpumask_clear_cpu(cpu, prof_cpu_mask);
375 if (per_cpu(cpu_profile_hits, cpu)[0]) {
376 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
377 per_cpu(cpu_profile_hits, cpu)[0] = NULL;
380 if (per_cpu(cpu_profile_hits, cpu)[1]) {
381 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
382 per_cpu(cpu_profile_hits, cpu)[1] = NULL;
389 #else /* !CONFIG_SMP */
390 #define profile_flip_buffers() do { } while (0)
391 #define profile_discard_flip_buffers() do { } while (0)
392 #define profile_cpu_callback NULL
394 static void do_profile_hits(int type, void *__pc, unsigned int nr_hits)
397 pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
398 atomic_add(nr_hits, &prof_buffer[min(pc, prof_len - 1)]);
400 #endif /* !CONFIG_SMP */
402 void profile_hits(int type, void *__pc, unsigned int nr_hits)
404 if (prof_on != type || !prof_buffer)
406 do_profile_hits(type, __pc, nr_hits);
408 EXPORT_SYMBOL_GPL(profile_hits);
410 void profile_tick(int type)
412 struct pt_regs *regs = get_irq_regs();
414 if (!user_mode(regs) && prof_cpu_mask != NULL &&
415 cpumask_test_cpu(smp_processor_id(), prof_cpu_mask))
416 profile_hit(type, (void *)profile_pc(regs));
419 #ifdef CONFIG_PROC_FS
420 #include <linux/proc_fs.h>
421 #include <linux/seq_file.h>
422 #include <asm/uaccess.h>
424 static int prof_cpu_mask_proc_show(struct seq_file *m, void *v)
426 seq_printf(m, "%*pb\n", cpumask_pr_args(prof_cpu_mask));
430 static int prof_cpu_mask_proc_open(struct inode *inode, struct file *file)
432 return single_open(file, prof_cpu_mask_proc_show, NULL);
435 static ssize_t prof_cpu_mask_proc_write(struct file *file,
436 const char __user *buffer, size_t count, loff_t *pos)
438 cpumask_var_t new_value;
441 if (!alloc_cpumask_var(&new_value, GFP_KERNEL))
444 err = cpumask_parse_user(buffer, count, new_value);
446 cpumask_copy(prof_cpu_mask, new_value);
449 free_cpumask_var(new_value);
453 static const struct file_operations prof_cpu_mask_proc_fops = {
454 .open = prof_cpu_mask_proc_open,
457 .release = single_release,
458 .write = prof_cpu_mask_proc_write,
461 void create_prof_cpu_mask(void)
463 /* create /proc/irq/prof_cpu_mask */
464 proc_create("irq/prof_cpu_mask", 0600, NULL, &prof_cpu_mask_proc_fops);
468 * This function accesses profiling information. The returned data is
469 * binary: the sampling step and the actual contents of the profile
470 * buffer. Use of the program readprofile is recommended in order to
471 * get meaningful info out of these data.
474 read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos)
476 unsigned long p = *ppos;
479 unsigned int sample_step = 1 << prof_shift;
481 profile_flip_buffers();
482 if (p >= (prof_len+1)*sizeof(unsigned int))
484 if (count > (prof_len+1)*sizeof(unsigned int) - p)
485 count = (prof_len+1)*sizeof(unsigned int) - p;
488 while (p < sizeof(unsigned int) && count > 0) {
489 if (put_user(*((char *)(&sample_step)+p), buf))
491 buf++; p++; count--; read++;
493 pnt = (char *)prof_buffer + p - sizeof(atomic_t);
494 if (copy_to_user(buf, (void *)pnt, count))
502 * Writing to /proc/profile resets the counters
504 * Writing a 'profiling multiplier' value into it also re-sets the profiling
505 * interrupt frequency, on architectures that support this.
507 static ssize_t write_profile(struct file *file, const char __user *buf,
508 size_t count, loff_t *ppos)
511 extern int setup_profiling_timer(unsigned int multiplier);
513 if (count == sizeof(int)) {
514 unsigned int multiplier;
516 if (copy_from_user(&multiplier, buf, sizeof(int)))
519 if (setup_profiling_timer(multiplier))
523 profile_discard_flip_buffers();
524 memset(prof_buffer, 0, prof_len * sizeof(atomic_t));
528 static const struct file_operations proc_profile_operations = {
529 .read = read_profile,
530 .write = write_profile,
531 .llseek = default_llseek,
535 static void profile_nop(void *unused)
539 static int create_hash_tables(void)
543 for_each_online_cpu(cpu) {
544 int node = cpu_to_mem(cpu);
547 page = __alloc_pages_node(node,
548 GFP_KERNEL | __GFP_ZERO | __GFP_THISNODE,
552 per_cpu(cpu_profile_hits, cpu)[1]
553 = (struct profile_hit *)page_address(page);
554 page = __alloc_pages_node(node,
555 GFP_KERNEL | __GFP_ZERO | __GFP_THISNODE,
559 per_cpu(cpu_profile_hits, cpu)[0]
560 = (struct profile_hit *)page_address(page);
566 on_each_cpu(profile_nop, NULL, 1);
567 for_each_online_cpu(cpu) {
570 if (per_cpu(cpu_profile_hits, cpu)[0]) {
571 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
572 per_cpu(cpu_profile_hits, cpu)[0] = NULL;
575 if (per_cpu(cpu_profile_hits, cpu)[1]) {
576 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
577 per_cpu(cpu_profile_hits, cpu)[1] = NULL;
584 #define create_hash_tables() ({ 0; })
587 int __ref create_proc_profile(void) /* false positive from hotcpu_notifier */
589 struct proc_dir_entry *entry;
595 cpu_notifier_register_begin();
597 if (create_hash_tables()) {
602 entry = proc_create("profile", S_IWUSR | S_IRUGO,
603 NULL, &proc_profile_operations);
606 proc_set_size(entry, (1 + prof_len) * sizeof(atomic_t));
607 __hotcpu_notifier(profile_cpu_callback, 0);
610 cpu_notifier_register_done();
613 subsys_initcall(create_proc_profile);
614 #endif /* CONFIG_PROC_FS */