Loading...
1/*
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.
7 *
8 * Scheduler profiling support, Arjan van de Ven and Ingo Molnar,
9 * Red Hat, July 2004
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
14 */
15
16#include <linux/export.h>
17#include <linux/profile.h>
18#include <linux/bootmem.h>
19#include <linux/notifier.h>
20#include <linux/mm.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>
30
31struct profile_hit {
32 u32 pc, hits;
33};
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)
38
39/* Oprofile timer tick hook */
40static int (*timer_hook)(struct pt_regs *) __read_mostly;
41
42static atomic_t *prof_buffer;
43static unsigned long prof_len, prof_shift;
44
45int prof_on __read_mostly;
46EXPORT_SYMBOL_GPL(prof_on);
47
48static cpumask_var_t prof_cpu_mask;
49#ifdef CONFIG_SMP
50static DEFINE_PER_CPU(struct profile_hit *[2], cpu_profile_hits);
51static DEFINE_PER_CPU(int, cpu_profile_flip);
52static DEFINE_MUTEX(profile_flip_mutex);
53#endif /* CONFIG_SMP */
54
55int profile_setup(char *str)
56{
57 static char schedstr[] = "schedule";
58 static char sleepstr[] = "sleep";
59 static char kvmstr[] = "kvm";
60 int par;
61
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))
68 prof_shift = par;
69 printk(KERN_INFO
70 "kernel sleep profiling enabled (shift: %ld)\n",
71 prof_shift);
72#else
73 printk(KERN_WARNING
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))
81 prof_shift = par;
82 printk(KERN_INFO
83 "kernel schedule profiling enabled (shift: %ld)\n",
84 prof_shift);
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))
90 prof_shift = par;
91 printk(KERN_INFO
92 "kernel KVM profiling enabled (shift: %ld)\n",
93 prof_shift);
94 } else if (get_option(&str, &par)) {
95 prof_shift = par;
96 prof_on = CPU_PROFILING;
97 printk(KERN_INFO "kernel profiling enabled (shift: %ld)\n",
98 prof_shift);
99 }
100 return 1;
101}
102__setup("profile=", profile_setup);
103
104
105int __ref profile_init(void)
106{
107 int buffer_bytes;
108 if (!prof_on)
109 return 0;
110
111 /* only text is profiled */
112 prof_len = (_etext - _stext) >> prof_shift;
113 buffer_bytes = prof_len*sizeof(atomic_t);
114
115 if (!alloc_cpumask_var(&prof_cpu_mask, GFP_KERNEL))
116 return -ENOMEM;
117
118 cpumask_copy(prof_cpu_mask, cpu_possible_mask);
119
120 prof_buffer = kzalloc(buffer_bytes, GFP_KERNEL|__GFP_NOWARN);
121 if (prof_buffer)
122 return 0;
123
124 prof_buffer = alloc_pages_exact(buffer_bytes,
125 GFP_KERNEL|__GFP_ZERO|__GFP_NOWARN);
126 if (prof_buffer)
127 return 0;
128
129 prof_buffer = vzalloc(buffer_bytes);
130 if (prof_buffer)
131 return 0;
132
133 free_cpumask_var(prof_cpu_mask);
134 return -ENOMEM;
135}
136
137/* Profile event notifications */
138
139static BLOCKING_NOTIFIER_HEAD(task_exit_notifier);
140static ATOMIC_NOTIFIER_HEAD(task_free_notifier);
141static BLOCKING_NOTIFIER_HEAD(munmap_notifier);
142
143void profile_task_exit(struct task_struct *task)
144{
145 blocking_notifier_call_chain(&task_exit_notifier, 0, task);
146}
147
148int profile_handoff_task(struct task_struct *task)
149{
150 int ret;
151 ret = atomic_notifier_call_chain(&task_free_notifier, 0, task);
152 return (ret == NOTIFY_OK) ? 1 : 0;
153}
154
155void profile_munmap(unsigned long addr)
156{
157 blocking_notifier_call_chain(&munmap_notifier, 0, (void *)addr);
158}
159
160int task_handoff_register(struct notifier_block *n)
161{
162 return atomic_notifier_chain_register(&task_free_notifier, n);
163}
164EXPORT_SYMBOL_GPL(task_handoff_register);
165
166int task_handoff_unregister(struct notifier_block *n)
167{
168 return atomic_notifier_chain_unregister(&task_free_notifier, n);
169}
170EXPORT_SYMBOL_GPL(task_handoff_unregister);
171
172int profile_event_register(enum profile_type type, struct notifier_block *n)
173{
174 int err = -EINVAL;
175
176 switch (type) {
177 case PROFILE_TASK_EXIT:
178 err = blocking_notifier_chain_register(
179 &task_exit_notifier, n);
180 break;
181 case PROFILE_MUNMAP:
182 err = blocking_notifier_chain_register(
183 &munmap_notifier, n);
184 break;
185 }
186
187 return err;
188}
189EXPORT_SYMBOL_GPL(profile_event_register);
190
191int profile_event_unregister(enum profile_type type, struct notifier_block *n)
192{
193 int err = -EINVAL;
194
195 switch (type) {
196 case PROFILE_TASK_EXIT:
197 err = blocking_notifier_chain_unregister(
198 &task_exit_notifier, n);
199 break;
200 case PROFILE_MUNMAP:
201 err = blocking_notifier_chain_unregister(
202 &munmap_notifier, n);
203 break;
204 }
205
206 return err;
207}
208EXPORT_SYMBOL_GPL(profile_event_unregister);
209
210int register_timer_hook(int (*hook)(struct pt_regs *))
211{
212 if (timer_hook)
213 return -EBUSY;
214 timer_hook = hook;
215 return 0;
216}
217EXPORT_SYMBOL_GPL(register_timer_hook);
218
219void unregister_timer_hook(int (*hook)(struct pt_regs *))
220{
221 WARN_ON(hook != timer_hook);
222 timer_hook = NULL;
223 /* make sure all CPUs see the NULL hook */
224 synchronize_sched(); /* Allow ongoing interrupts to complete. */
225}
226EXPORT_SYMBOL_GPL(unregister_timer_hook);
227
228
229#ifdef CONFIG_SMP
230/*
231 * Each cpu has a pair of open-addressed hashtables for pending
232 * profile hits. read_profile() IPI's all cpus to request them
233 * to flip buffers and flushes their contents to prof_buffer itself.
234 * Flip requests are serialized by the profile_flip_mutex. The sole
235 * use of having a second hashtable is for avoiding cacheline
236 * contention that would otherwise happen during flushes of pending
237 * profile hits required for the accuracy of reported profile hits
238 * and so resurrect the interrupt livelock issue.
239 *
240 * The open-addressed hashtables are indexed by profile buffer slot
241 * and hold the number of pending hits to that profile buffer slot on
242 * a cpu in an entry. When the hashtable overflows, all pending hits
243 * are accounted to their corresponding profile buffer slots with
244 * atomic_add() and the hashtable emptied. As numerous pending hits
245 * may be accounted to a profile buffer slot in a hashtable entry,
246 * this amortizes a number of atomic profile buffer increments likely
247 * to be far larger than the number of entries in the hashtable,
248 * particularly given that the number of distinct profile buffer
249 * positions to which hits are accounted during short intervals (e.g.
250 * several seconds) is usually very small. Exclusion from buffer
251 * flipping is provided by interrupt disablement (note that for
252 * SCHED_PROFILING or SLEEP_PROFILING profile_hit() may be called from
253 * process context).
254 * The hash function is meant to be lightweight as opposed to strong,
255 * and was vaguely inspired by ppc64 firmware-supported inverted
256 * pagetable hash functions, but uses a full hashtable full of finite
257 * collision chains, not just pairs of them.
258 *
259 * -- wli
260 */
261static void __profile_flip_buffers(void *unused)
262{
263 int cpu = smp_processor_id();
264
265 per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu);
266}
267
268static void profile_flip_buffers(void)
269{
270 int i, j, cpu;
271
272 mutex_lock(&profile_flip_mutex);
273 j = per_cpu(cpu_profile_flip, get_cpu());
274 put_cpu();
275 on_each_cpu(__profile_flip_buffers, NULL, 1);
276 for_each_online_cpu(cpu) {
277 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[j];
278 for (i = 0; i < NR_PROFILE_HIT; ++i) {
279 if (!hits[i].hits) {
280 if (hits[i].pc)
281 hits[i].pc = 0;
282 continue;
283 }
284 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
285 hits[i].hits = hits[i].pc = 0;
286 }
287 }
288 mutex_unlock(&profile_flip_mutex);
289}
290
291static void profile_discard_flip_buffers(void)
292{
293 int i, cpu;
294
295 mutex_lock(&profile_flip_mutex);
296 i = per_cpu(cpu_profile_flip, get_cpu());
297 put_cpu();
298 on_each_cpu(__profile_flip_buffers, NULL, 1);
299 for_each_online_cpu(cpu) {
300 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[i];
301 memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit));
302 }
303 mutex_unlock(&profile_flip_mutex);
304}
305
306static void do_profile_hits(int type, void *__pc, unsigned int nr_hits)
307{
308 unsigned long primary, secondary, flags, pc = (unsigned long)__pc;
309 int i, j, cpu;
310 struct profile_hit *hits;
311
312 pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1);
313 i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
314 secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
315 cpu = get_cpu();
316 hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)];
317 if (!hits) {
318 put_cpu();
319 return;
320 }
321 /*
322 * We buffer the global profiler buffer into a per-CPU
323 * queue and thus reduce the number of global (and possibly
324 * NUMA-alien) accesses. The write-queue is self-coalescing:
325 */
326 local_irq_save(flags);
327 do {
328 for (j = 0; j < PROFILE_GRPSZ; ++j) {
329 if (hits[i + j].pc == pc) {
330 hits[i + j].hits += nr_hits;
331 goto out;
332 } else if (!hits[i + j].hits) {
333 hits[i + j].pc = pc;
334 hits[i + j].hits = nr_hits;
335 goto out;
336 }
337 }
338 i = (i + secondary) & (NR_PROFILE_HIT - 1);
339 } while (i != primary);
340
341 /*
342 * Add the current hit(s) and flush the write-queue out
343 * to the global buffer:
344 */
345 atomic_add(nr_hits, &prof_buffer[pc]);
346 for (i = 0; i < NR_PROFILE_HIT; ++i) {
347 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
348 hits[i].pc = hits[i].hits = 0;
349 }
350out:
351 local_irq_restore(flags);
352 put_cpu();
353}
354
355static int __cpuinit profile_cpu_callback(struct notifier_block *info,
356 unsigned long action, void *__cpu)
357{
358 int node, cpu = (unsigned long)__cpu;
359 struct page *page;
360
361 switch (action) {
362 case CPU_UP_PREPARE:
363 case CPU_UP_PREPARE_FROZEN:
364 node = cpu_to_mem(cpu);
365 per_cpu(cpu_profile_flip, cpu) = 0;
366 if (!per_cpu(cpu_profile_hits, cpu)[1]) {
367 page = alloc_pages_exact_node(node,
368 GFP_KERNEL | __GFP_ZERO,
369 0);
370 if (!page)
371 return notifier_from_errno(-ENOMEM);
372 per_cpu(cpu_profile_hits, cpu)[1] = page_address(page);
373 }
374 if (!per_cpu(cpu_profile_hits, cpu)[0]) {
375 page = alloc_pages_exact_node(node,
376 GFP_KERNEL | __GFP_ZERO,
377 0);
378 if (!page)
379 goto out_free;
380 per_cpu(cpu_profile_hits, cpu)[0] = page_address(page);
381 }
382 break;
383out_free:
384 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
385 per_cpu(cpu_profile_hits, cpu)[1] = NULL;
386 __free_page(page);
387 return notifier_from_errno(-ENOMEM);
388 case CPU_ONLINE:
389 case CPU_ONLINE_FROZEN:
390 if (prof_cpu_mask != NULL)
391 cpumask_set_cpu(cpu, prof_cpu_mask);
392 break;
393 case CPU_UP_CANCELED:
394 case CPU_UP_CANCELED_FROZEN:
395 case CPU_DEAD:
396 case CPU_DEAD_FROZEN:
397 if (prof_cpu_mask != NULL)
398 cpumask_clear_cpu(cpu, prof_cpu_mask);
399 if (per_cpu(cpu_profile_hits, cpu)[0]) {
400 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
401 per_cpu(cpu_profile_hits, cpu)[0] = NULL;
402 __free_page(page);
403 }
404 if (per_cpu(cpu_profile_hits, cpu)[1]) {
405 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
406 per_cpu(cpu_profile_hits, cpu)[1] = NULL;
407 __free_page(page);
408 }
409 break;
410 }
411 return NOTIFY_OK;
412}
413#else /* !CONFIG_SMP */
414#define profile_flip_buffers() do { } while (0)
415#define profile_discard_flip_buffers() do { } while (0)
416#define profile_cpu_callback NULL
417
418static void do_profile_hits(int type, void *__pc, unsigned int nr_hits)
419{
420 unsigned long pc;
421 pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
422 atomic_add(nr_hits, &prof_buffer[min(pc, prof_len - 1)]);
423}
424#endif /* !CONFIG_SMP */
425
426void profile_hits(int type, void *__pc, unsigned int nr_hits)
427{
428 if (prof_on != type || !prof_buffer)
429 return;
430 do_profile_hits(type, __pc, nr_hits);
431}
432EXPORT_SYMBOL_GPL(profile_hits);
433
434void profile_tick(int type)
435{
436 struct pt_regs *regs = get_irq_regs();
437
438 if (type == CPU_PROFILING && timer_hook)
439 timer_hook(regs);
440 if (!user_mode(regs) && prof_cpu_mask != NULL &&
441 cpumask_test_cpu(smp_processor_id(), prof_cpu_mask))
442 profile_hit(type, (void *)profile_pc(regs));
443}
444
445#ifdef CONFIG_PROC_FS
446#include <linux/proc_fs.h>
447#include <linux/seq_file.h>
448#include <asm/uaccess.h>
449
450static int prof_cpu_mask_proc_show(struct seq_file *m, void *v)
451{
452 seq_cpumask(m, prof_cpu_mask);
453 seq_putc(m, '\n');
454 return 0;
455}
456
457static int prof_cpu_mask_proc_open(struct inode *inode, struct file *file)
458{
459 return single_open(file, prof_cpu_mask_proc_show, NULL);
460}
461
462static ssize_t prof_cpu_mask_proc_write(struct file *file,
463 const char __user *buffer, size_t count, loff_t *pos)
464{
465 cpumask_var_t new_value;
466 int err;
467
468 if (!alloc_cpumask_var(&new_value, GFP_KERNEL))
469 return -ENOMEM;
470
471 err = cpumask_parse_user(buffer, count, new_value);
472 if (!err) {
473 cpumask_copy(prof_cpu_mask, new_value);
474 err = count;
475 }
476 free_cpumask_var(new_value);
477 return err;
478}
479
480static const struct file_operations prof_cpu_mask_proc_fops = {
481 .open = prof_cpu_mask_proc_open,
482 .read = seq_read,
483 .llseek = seq_lseek,
484 .release = single_release,
485 .write = prof_cpu_mask_proc_write,
486};
487
488void create_prof_cpu_mask(struct proc_dir_entry *root_irq_dir)
489{
490 /* create /proc/irq/prof_cpu_mask */
491 proc_create("prof_cpu_mask", 0600, root_irq_dir, &prof_cpu_mask_proc_fops);
492}
493
494/*
495 * This function accesses profiling information. The returned data is
496 * binary: the sampling step and the actual contents of the profile
497 * buffer. Use of the program readprofile is recommended in order to
498 * get meaningful info out of these data.
499 */
500static ssize_t
501read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos)
502{
503 unsigned long p = *ppos;
504 ssize_t read;
505 char *pnt;
506 unsigned int sample_step = 1 << prof_shift;
507
508 profile_flip_buffers();
509 if (p >= (prof_len+1)*sizeof(unsigned int))
510 return 0;
511 if (count > (prof_len+1)*sizeof(unsigned int) - p)
512 count = (prof_len+1)*sizeof(unsigned int) - p;
513 read = 0;
514
515 while (p < sizeof(unsigned int) && count > 0) {
516 if (put_user(*((char *)(&sample_step)+p), buf))
517 return -EFAULT;
518 buf++; p++; count--; read++;
519 }
520 pnt = (char *)prof_buffer + p - sizeof(atomic_t);
521 if (copy_to_user(buf, (void *)pnt, count))
522 return -EFAULT;
523 read += count;
524 *ppos += read;
525 return read;
526}
527
528/*
529 * Writing to /proc/profile resets the counters
530 *
531 * Writing a 'profiling multiplier' value into it also re-sets the profiling
532 * interrupt frequency, on architectures that support this.
533 */
534static ssize_t write_profile(struct file *file, const char __user *buf,
535 size_t count, loff_t *ppos)
536{
537#ifdef CONFIG_SMP
538 extern int setup_profiling_timer(unsigned int multiplier);
539
540 if (count == sizeof(int)) {
541 unsigned int multiplier;
542
543 if (copy_from_user(&multiplier, buf, sizeof(int)))
544 return -EFAULT;
545
546 if (setup_profiling_timer(multiplier))
547 return -EINVAL;
548 }
549#endif
550 profile_discard_flip_buffers();
551 memset(prof_buffer, 0, prof_len * sizeof(atomic_t));
552 return count;
553}
554
555static const struct file_operations proc_profile_operations = {
556 .read = read_profile,
557 .write = write_profile,
558 .llseek = default_llseek,
559};
560
561#ifdef CONFIG_SMP
562static void profile_nop(void *unused)
563{
564}
565
566static int create_hash_tables(void)
567{
568 int cpu;
569
570 for_each_online_cpu(cpu) {
571 int node = cpu_to_mem(cpu);
572 struct page *page;
573
574 page = alloc_pages_exact_node(node,
575 GFP_KERNEL | __GFP_ZERO | GFP_THISNODE,
576 0);
577 if (!page)
578 goto out_cleanup;
579 per_cpu(cpu_profile_hits, cpu)[1]
580 = (struct profile_hit *)page_address(page);
581 page = alloc_pages_exact_node(node,
582 GFP_KERNEL | __GFP_ZERO | GFP_THISNODE,
583 0);
584 if (!page)
585 goto out_cleanup;
586 per_cpu(cpu_profile_hits, cpu)[0]
587 = (struct profile_hit *)page_address(page);
588 }
589 return 0;
590out_cleanup:
591 prof_on = 0;
592 smp_mb();
593 on_each_cpu(profile_nop, NULL, 1);
594 for_each_online_cpu(cpu) {
595 struct page *page;
596
597 if (per_cpu(cpu_profile_hits, cpu)[0]) {
598 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
599 per_cpu(cpu_profile_hits, cpu)[0] = NULL;
600 __free_page(page);
601 }
602 if (per_cpu(cpu_profile_hits, cpu)[1]) {
603 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
604 per_cpu(cpu_profile_hits, cpu)[1] = NULL;
605 __free_page(page);
606 }
607 }
608 return -1;
609}
610#else
611#define create_hash_tables() ({ 0; })
612#endif
613
614int __ref create_proc_profile(void) /* false positive from hotcpu_notifier */
615{
616 struct proc_dir_entry *entry;
617
618 if (!prof_on)
619 return 0;
620 if (create_hash_tables())
621 return -ENOMEM;
622 entry = proc_create("profile", S_IWUSR | S_IRUGO,
623 NULL, &proc_profile_operations);
624 if (!entry)
625 return 0;
626 entry->size = (1+prof_len) * sizeof(atomic_t);
627 hotcpu_notifier(profile_cpu_callback, 0);
628 return 0;
629}
630module_init(create_proc_profile);
631#endif /* CONFIG_PROC_FS */