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;
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 char schedstr[] = "schedule";
56 static char sleepstr[] = "sleep";
57 static char kvmstr[] = "kvm";
60 if (!strncmp(str, sleepstr, strlen(sleepstr))) {
61 #ifdef CONFIG_SCHEDSTATS
62 prof_on = SLEEP_PROFILING;
63 if (str[strlen(sleepstr)] == ',')
64 str += strlen(sleepstr) + 1;
65 if (get_option(&str, &par))
68 "kernel sleep profiling enabled (shift: %ld)\n",
72 "kernel sleep profiling requires CONFIG_SCHEDSTATS\n");
73 #endif /* CONFIG_SCHEDSTATS */
74 } else if (!strncmp(str, schedstr, strlen(schedstr))) {
75 prof_on = SCHED_PROFILING;
76 if (str[strlen(schedstr)] == ',')
77 str += strlen(schedstr) + 1;
78 if (get_option(&str, &par))
81 "kernel schedule profiling enabled (shift: %ld)\n",
83 } else if (!strncmp(str, kvmstr, strlen(kvmstr))) {
84 prof_on = KVM_PROFILING;
85 if (str[strlen(kvmstr)] == ',')
86 str += strlen(kvmstr) + 1;
87 if (get_option(&str, &par))
90 "kernel KVM profiling enabled (shift: %ld)\n",
92 } else if (get_option(&str, &par)) {
94 prof_on = CPU_PROFILING;
95 printk(KERN_INFO "kernel profiling enabled (shift: %ld)\n",
100 __setup("profile=", profile_setup);
103 int __ref profile_init(void)
109 /* only text is profiled */
110 prof_len = (_etext - _stext) >> prof_shift;
111 buffer_bytes = prof_len*sizeof(atomic_t);
113 if (!alloc_cpumask_var(&prof_cpu_mask, GFP_KERNEL))
116 cpumask_copy(prof_cpu_mask, cpu_possible_mask);
118 prof_buffer = kzalloc(buffer_bytes, GFP_KERNEL|__GFP_NOWARN);
122 prof_buffer = alloc_pages_exact(buffer_bytes,
123 GFP_KERNEL|__GFP_ZERO|__GFP_NOWARN);
127 prof_buffer = vzalloc(buffer_bytes);
131 free_cpumask_var(prof_cpu_mask);
135 /* Profile event notifications */
137 static BLOCKING_NOTIFIER_HEAD(task_exit_notifier);
138 static ATOMIC_NOTIFIER_HEAD(task_free_notifier);
139 static BLOCKING_NOTIFIER_HEAD(munmap_notifier);
141 void profile_task_exit(struct task_struct *task)
143 blocking_notifier_call_chain(&task_exit_notifier, 0, task);
146 int profile_handoff_task(struct task_struct *task)
149 ret = atomic_notifier_call_chain(&task_free_notifier, 0, task);
150 return (ret == NOTIFY_OK) ? 1 : 0;
153 void profile_munmap(unsigned long addr)
155 blocking_notifier_call_chain(&munmap_notifier, 0, (void *)addr);
158 int task_handoff_register(struct notifier_block *n)
160 return atomic_notifier_chain_register(&task_free_notifier, n);
162 EXPORT_SYMBOL_GPL(task_handoff_register);
164 int task_handoff_unregister(struct notifier_block *n)
166 return atomic_notifier_chain_unregister(&task_free_notifier, n);
168 EXPORT_SYMBOL_GPL(task_handoff_unregister);
170 int profile_event_register(enum profile_type type, struct notifier_block *n)
175 case PROFILE_TASK_EXIT:
176 err = blocking_notifier_chain_register(
177 &task_exit_notifier, n);
180 err = blocking_notifier_chain_register(
181 &munmap_notifier, n);
187 EXPORT_SYMBOL_GPL(profile_event_register);
189 int profile_event_unregister(enum profile_type type, struct notifier_block *n)
194 case PROFILE_TASK_EXIT:
195 err = blocking_notifier_chain_unregister(
196 &task_exit_notifier, n);
199 err = blocking_notifier_chain_unregister(
200 &munmap_notifier, n);
206 EXPORT_SYMBOL_GPL(profile_event_unregister);
210 * Each cpu has a pair of open-addressed hashtables for pending
211 * profile hits. read_profile() IPI's all cpus to request them
212 * to flip buffers and flushes their contents to prof_buffer itself.
213 * Flip requests are serialized by the profile_flip_mutex. The sole
214 * use of having a second hashtable is for avoiding cacheline
215 * contention that would otherwise happen during flushes of pending
216 * profile hits required for the accuracy of reported profile hits
217 * and so resurrect the interrupt livelock issue.
219 * The open-addressed hashtables are indexed by profile buffer slot
220 * and hold the number of pending hits to that profile buffer slot on
221 * a cpu in an entry. When the hashtable overflows, all pending hits
222 * are accounted to their corresponding profile buffer slots with
223 * atomic_add() and the hashtable emptied. As numerous pending hits
224 * may be accounted to a profile buffer slot in a hashtable entry,
225 * this amortizes a number of atomic profile buffer increments likely
226 * to be far larger than the number of entries in the hashtable,
227 * particularly given that the number of distinct profile buffer
228 * positions to which hits are accounted during short intervals (e.g.
229 * several seconds) is usually very small. Exclusion from buffer
230 * flipping is provided by interrupt disablement (note that for
231 * SCHED_PROFILING or SLEEP_PROFILING profile_hit() may be called from
233 * The hash function is meant to be lightweight as opposed to strong,
234 * and was vaguely inspired by ppc64 firmware-supported inverted
235 * pagetable hash functions, but uses a full hashtable full of finite
236 * collision chains, not just pairs of them.
240 static void __profile_flip_buffers(void *unused)
242 int cpu = smp_processor_id();
244 per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu);
247 static void profile_flip_buffers(void)
251 mutex_lock(&profile_flip_mutex);
252 j = per_cpu(cpu_profile_flip, get_cpu());
254 on_each_cpu(__profile_flip_buffers, NULL, 1);
255 for_each_online_cpu(cpu) {
256 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[j];
257 for (i = 0; i < NR_PROFILE_HIT; ++i) {
263 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
264 hits[i].hits = hits[i].pc = 0;
267 mutex_unlock(&profile_flip_mutex);
270 static void profile_discard_flip_buffers(void)
274 mutex_lock(&profile_flip_mutex);
275 i = per_cpu(cpu_profile_flip, get_cpu());
277 on_each_cpu(__profile_flip_buffers, NULL, 1);
278 for_each_online_cpu(cpu) {
279 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[i];
280 memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit));
282 mutex_unlock(&profile_flip_mutex);
285 static void do_profile_hits(int type, void *__pc, unsigned int nr_hits)
287 unsigned long primary, secondary, flags, pc = (unsigned long)__pc;
289 struct profile_hit *hits;
291 pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1);
292 i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
293 secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
295 hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)];
301 * We buffer the global profiler buffer into a per-CPU
302 * queue and thus reduce the number of global (and possibly
303 * NUMA-alien) accesses. The write-queue is self-coalescing:
305 local_irq_save(flags);
307 for (j = 0; j < PROFILE_GRPSZ; ++j) {
308 if (hits[i + j].pc == pc) {
309 hits[i + j].hits += nr_hits;
311 } else if (!hits[i + j].hits) {
313 hits[i + j].hits = nr_hits;
317 i = (i + secondary) & (NR_PROFILE_HIT - 1);
318 } while (i != primary);
321 * Add the current hit(s) and flush the write-queue out
322 * to the global buffer:
324 atomic_add(nr_hits, &prof_buffer[pc]);
325 for (i = 0; i < NR_PROFILE_HIT; ++i) {
326 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
327 hits[i].pc = hits[i].hits = 0;
330 local_irq_restore(flags);
334 static int profile_cpu_callback(struct notifier_block *info,
335 unsigned long action, void *__cpu)
337 int node, cpu = (unsigned long)__cpu;
342 case CPU_UP_PREPARE_FROZEN:
343 node = cpu_to_mem(cpu);
344 per_cpu(cpu_profile_flip, cpu) = 0;
345 if (!per_cpu(cpu_profile_hits, cpu)[1]) {
346 page = alloc_pages_exact_node(node,
347 GFP_KERNEL | __GFP_ZERO,
350 return notifier_from_errno(-ENOMEM);
351 per_cpu(cpu_profile_hits, cpu)[1] = page_address(page);
353 if (!per_cpu(cpu_profile_hits, cpu)[0]) {
354 page = alloc_pages_exact_node(node,
355 GFP_KERNEL | __GFP_ZERO,
359 per_cpu(cpu_profile_hits, cpu)[0] = page_address(page);
363 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
364 per_cpu(cpu_profile_hits, cpu)[1] = NULL;
366 return notifier_from_errno(-ENOMEM);
368 case CPU_ONLINE_FROZEN:
369 if (prof_cpu_mask != NULL)
370 cpumask_set_cpu(cpu, prof_cpu_mask);
372 case CPU_UP_CANCELED:
373 case CPU_UP_CANCELED_FROZEN:
375 case CPU_DEAD_FROZEN:
376 if (prof_cpu_mask != NULL)
377 cpumask_clear_cpu(cpu, prof_cpu_mask);
378 if (per_cpu(cpu_profile_hits, cpu)[0]) {
379 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
380 per_cpu(cpu_profile_hits, cpu)[0] = NULL;
383 if (per_cpu(cpu_profile_hits, cpu)[1]) {
384 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
385 per_cpu(cpu_profile_hits, cpu)[1] = NULL;
392 #else /* !CONFIG_SMP */
393 #define profile_flip_buffers() do { } while (0)
394 #define profile_discard_flip_buffers() do { } while (0)
395 #define profile_cpu_callback NULL
397 static void do_profile_hits(int type, void *__pc, unsigned int nr_hits)
400 pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
401 atomic_add(nr_hits, &prof_buffer[min(pc, prof_len - 1)]);
403 #endif /* !CONFIG_SMP */
405 void profile_hits(int type, void *__pc, unsigned int nr_hits)
407 if (prof_on != type || !prof_buffer)
409 do_profile_hits(type, __pc, nr_hits);
411 EXPORT_SYMBOL_GPL(profile_hits);
413 void profile_tick(int type)
415 struct pt_regs *regs = get_irq_regs();
417 if (!user_mode(regs) && prof_cpu_mask != NULL &&
418 cpumask_test_cpu(smp_processor_id(), prof_cpu_mask))
419 profile_hit(type, (void *)profile_pc(regs));
422 #ifdef CONFIG_PROC_FS
423 #include <linux/proc_fs.h>
424 #include <linux/seq_file.h>
425 #include <asm/uaccess.h>
427 static int prof_cpu_mask_proc_show(struct seq_file *m, void *v)
429 seq_cpumask(m, prof_cpu_mask);
434 static int prof_cpu_mask_proc_open(struct inode *inode, struct file *file)
436 return single_open(file, prof_cpu_mask_proc_show, NULL);
439 static ssize_t prof_cpu_mask_proc_write(struct file *file,
440 const char __user *buffer, size_t count, loff_t *pos)
442 cpumask_var_t new_value;
445 if (!alloc_cpumask_var(&new_value, GFP_KERNEL))
448 err = cpumask_parse_user(buffer, count, new_value);
450 cpumask_copy(prof_cpu_mask, new_value);
453 free_cpumask_var(new_value);
457 static const struct file_operations prof_cpu_mask_proc_fops = {
458 .open = prof_cpu_mask_proc_open,
461 .release = single_release,
462 .write = prof_cpu_mask_proc_write,
465 void create_prof_cpu_mask(void)
467 /* create /proc/irq/prof_cpu_mask */
468 proc_create("irq/prof_cpu_mask", 0600, NULL, &prof_cpu_mask_proc_fops);
472 * This function accesses profiling information. The returned data is
473 * binary: the sampling step and the actual contents of the profile
474 * buffer. Use of the program readprofile is recommended in order to
475 * get meaningful info out of these data.
478 read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos)
480 unsigned long p = *ppos;
483 unsigned int sample_step = 1 << prof_shift;
485 profile_flip_buffers();
486 if (p >= (prof_len+1)*sizeof(unsigned int))
488 if (count > (prof_len+1)*sizeof(unsigned int) - p)
489 count = (prof_len+1)*sizeof(unsigned int) - p;
492 while (p < sizeof(unsigned int) && count > 0) {
493 if (put_user(*((char *)(&sample_step)+p), buf))
495 buf++; p++; count--; read++;
497 pnt = (char *)prof_buffer + p - sizeof(atomic_t);
498 if (copy_to_user(buf, (void *)pnt, count))
506 * Writing to /proc/profile resets the counters
508 * Writing a 'profiling multiplier' value into it also re-sets the profiling
509 * interrupt frequency, on architectures that support this.
511 static ssize_t write_profile(struct file *file, const char __user *buf,
512 size_t count, loff_t *ppos)
515 extern int setup_profiling_timer(unsigned int multiplier);
517 if (count == sizeof(int)) {
518 unsigned int multiplier;
520 if (copy_from_user(&multiplier, buf, sizeof(int)))
523 if (setup_profiling_timer(multiplier))
527 profile_discard_flip_buffers();
528 memset(prof_buffer, 0, prof_len * sizeof(atomic_t));
532 static const struct file_operations proc_profile_operations = {
533 .read = read_profile,
534 .write = write_profile,
535 .llseek = default_llseek,
539 static void profile_nop(void *unused)
543 static int create_hash_tables(void)
547 for_each_online_cpu(cpu) {
548 int node = cpu_to_mem(cpu);
551 page = alloc_pages_exact_node(node,
552 GFP_KERNEL | __GFP_ZERO | __GFP_THISNODE,
556 per_cpu(cpu_profile_hits, cpu)[1]
557 = (struct profile_hit *)page_address(page);
558 page = alloc_pages_exact_node(node,
559 GFP_KERNEL | __GFP_ZERO | __GFP_THISNODE,
563 per_cpu(cpu_profile_hits, cpu)[0]
564 = (struct profile_hit *)page_address(page);
570 on_each_cpu(profile_nop, NULL, 1);
571 for_each_online_cpu(cpu) {
574 if (per_cpu(cpu_profile_hits, cpu)[0]) {
575 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
576 per_cpu(cpu_profile_hits, cpu)[0] = NULL;
579 if (per_cpu(cpu_profile_hits, cpu)[1]) {
580 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
581 per_cpu(cpu_profile_hits, cpu)[1] = NULL;
588 #define create_hash_tables() ({ 0; })
591 int __ref create_proc_profile(void) /* false positive from hotcpu_notifier */
593 struct proc_dir_entry *entry;
597 if (create_hash_tables())
599 entry = proc_create("profile", S_IWUSR | S_IRUGO,
600 NULL, &proc_profile_operations);
603 proc_set_size(entry, (1 + prof_len) * sizeof(atomic_t));
604 hotcpu_notifier(profile_cpu_callback, 0);
607 module_init(create_proc_profile);
608 #endif /* CONFIG_PROC_FS */