blob: b07ba2a1411949cf7751d0830617a1ac34013abe [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/**
2 * @file buffer_sync.c
3 *
4 * @remark Copyright 2002 OProfile authors
5 * @remark Read the file COPYING
6 *
7 * @author John Levon <levon@movementarian.org>
8 *
9 * This is the core of the buffer management. Each
10 * CPU buffer is processed and entered into the
11 * global event buffer. Such processing is necessary
12 * in several circumstances, mentioned below.
13 *
14 * The processing does the job of converting the
15 * transitory EIP value into a persistent dentry/offset
16 * value that the profiler can record at its leisure.
17 *
18 * See fs/dcookies.c for a description of the dentry/offset
19 * objects.
20 */
21
22#include <linux/mm.h>
23#include <linux/workqueue.h>
24#include <linux/notifier.h>
25#include <linux/dcookies.h>
26#include <linux/profile.h>
27#include <linux/module.h>
28#include <linux/fs.h>
Bob Nelson14748552007-07-20 21:39:53 +020029#include <linux/oprofile.h>
Alexey Dobriyane8edc6e2007-05-21 01:22:52 +040030#include <linux/sched.h>
Bob Nelson14748552007-07-20 21:39:53 +020031
Linus Torvalds1da177e2005-04-16 15:20:36 -070032#include "oprofile_stats.h"
33#include "event_buffer.h"
34#include "cpu_buffer.h"
35#include "buffer_sync.h"
36
37static LIST_HEAD(dying_tasks);
38static LIST_HEAD(dead_tasks);
39static cpumask_t marked_cpus = CPU_MASK_NONE;
40static DEFINE_SPINLOCK(task_mortuary);
41static void process_task_mortuary(void);
42
43
44/* Take ownership of the task struct and place it on the
45 * list for processing. Only after two full buffer syncs
46 * does the task eventually get freed, because by then
47 * we are sure we will not reference it again.
Paul E. McKenney4369ef32006-01-08 01:01:35 -080048 * Can be invoked from softirq via RCU callback due to
49 * call_rcu() of the task struct, hence the _irqsave.
Linus Torvalds1da177e2005-04-16 15:20:36 -070050 */
51static int task_free_notify(struct notifier_block * self, unsigned long val, void * data)
52{
Paul E. McKenney4369ef32006-01-08 01:01:35 -080053 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -070054 struct task_struct * task = data;
Paul E. McKenney4369ef32006-01-08 01:01:35 -080055 spin_lock_irqsave(&task_mortuary, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -070056 list_add(&task->tasks, &dying_tasks);
Paul E. McKenney4369ef32006-01-08 01:01:35 -080057 spin_unlock_irqrestore(&task_mortuary, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -070058 return NOTIFY_OK;
59}
60
61
62/* The task is on its way out. A sync of the buffer means we can catch
63 * any remaining samples for this task.
64 */
65static int task_exit_notify(struct notifier_block * self, unsigned long val, void * data)
66{
67 /* To avoid latency problems, we only process the current CPU,
68 * hoping that most samples for the task are on this CPU
69 */
Ingo Molnar39c715b2005-06-21 17:14:34 -070070 sync_buffer(raw_smp_processor_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -070071 return 0;
72}
73
74
75/* The task is about to try a do_munmap(). We peek at what it's going to
76 * do, and if it's an executable region, process the samples first, so
77 * we don't lose any. This does not have to be exact, it's a QoI issue
78 * only.
79 */
80static int munmap_notify(struct notifier_block * self, unsigned long val, void * data)
81{
82 unsigned long addr = (unsigned long)data;
83 struct mm_struct * mm = current->mm;
84 struct vm_area_struct * mpnt;
85
86 down_read(&mm->mmap_sem);
87
88 mpnt = find_vma(mm, addr);
89 if (mpnt && mpnt->vm_file && (mpnt->vm_flags & VM_EXEC)) {
90 up_read(&mm->mmap_sem);
91 /* To avoid latency problems, we only process the current CPU,
92 * hoping that most samples for the task are on this CPU
93 */
Ingo Molnar39c715b2005-06-21 17:14:34 -070094 sync_buffer(raw_smp_processor_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -070095 return 0;
96 }
97
98 up_read(&mm->mmap_sem);
99 return 0;
100}
101
102
103/* We need to be told about new modules so we don't attribute to a previously
104 * loaded module, or drop the samples on the floor.
105 */
106static int module_load_notify(struct notifier_block * self, unsigned long val, void * data)
107{
108#ifdef CONFIG_MODULES
109 if (val != MODULE_STATE_COMING)
110 return 0;
111
112 /* FIXME: should we process all CPU buffers ? */
Markus Armbruster59cc1852006-06-25 05:47:33 -0700113 mutex_lock(&buffer_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114 add_event_entry(ESCAPE_CODE);
115 add_event_entry(MODULE_LOADED_CODE);
Markus Armbruster59cc1852006-06-25 05:47:33 -0700116 mutex_unlock(&buffer_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700117#endif
118 return 0;
119}
120
121
122static struct notifier_block task_free_nb = {
123 .notifier_call = task_free_notify,
124};
125
126static struct notifier_block task_exit_nb = {
127 .notifier_call = task_exit_notify,
128};
129
130static struct notifier_block munmap_nb = {
131 .notifier_call = munmap_notify,
132};
133
134static struct notifier_block module_load_nb = {
135 .notifier_call = module_load_notify,
136};
137
138
139static void end_sync(void)
140{
141 end_cpu_work();
142 /* make sure we don't leak task structs */
143 process_task_mortuary();
144 process_task_mortuary();
145}
146
147
148int sync_start(void)
149{
150 int err;
151
152 start_cpu_work();
153
154 err = task_handoff_register(&task_free_nb);
155 if (err)
156 goto out1;
157 err = profile_event_register(PROFILE_TASK_EXIT, &task_exit_nb);
158 if (err)
159 goto out2;
160 err = profile_event_register(PROFILE_MUNMAP, &munmap_nb);
161 if (err)
162 goto out3;
163 err = register_module_notifier(&module_load_nb);
164 if (err)
165 goto out4;
166
167out:
168 return err;
169out4:
170 profile_event_unregister(PROFILE_MUNMAP, &munmap_nb);
171out3:
172 profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb);
173out2:
174 task_handoff_unregister(&task_free_nb);
175out1:
176 end_sync();
177 goto out;
178}
179
180
181void sync_stop(void)
182{
183 unregister_module_notifier(&module_load_nb);
184 profile_event_unregister(PROFILE_MUNMAP, &munmap_nb);
185 profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb);
186 task_handoff_unregister(&task_free_nb);
187 end_sync();
188}
189
Jan Blunck448678a2008-02-14 19:38:36 -0800190
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191/* Optimisation. We can manage without taking the dcookie sem
192 * because we cannot reach this code without at least one
193 * dcookie user still being registered (namely, the reader
194 * of the event buffer). */
Jan Blunck448678a2008-02-14 19:38:36 -0800195static inline unsigned long fast_get_dcookie(struct path *path)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700196{
197 unsigned long cookie;
Jan Blunck448678a2008-02-14 19:38:36 -0800198
199 if (path->dentry->d_cookie)
200 return (unsigned long)path->dentry;
201 get_dcookie(path, &cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202 return cookie;
203}
204
Jan Blunck448678a2008-02-14 19:38:36 -0800205
Linus Torvalds1da177e2005-04-16 15:20:36 -0700206/* Look up the dcookie for the task's first VM_EXECUTABLE mapping,
207 * which corresponds loosely to "application name". This is
208 * not strictly necessary but allows oprofile to associate
209 * shared-library samples with particular applications
210 */
211static unsigned long get_exec_dcookie(struct mm_struct * mm)
212{
John Levon0c0a4002005-06-23 22:02:47 -0700213 unsigned long cookie = NO_COOKIE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214 struct vm_area_struct * vma;
215
216 if (!mm)
217 goto out;
218
219 for (vma = mm->mmap; vma; vma = vma->vm_next) {
220 if (!vma->vm_file)
221 continue;
222 if (!(vma->vm_flags & VM_EXECUTABLE))
223 continue;
Jan Blunck448678a2008-02-14 19:38:36 -0800224 cookie = fast_get_dcookie(&vma->vm_file->f_path);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700225 break;
226 }
227
228out:
229 return cookie;
230}
231
232
233/* Convert the EIP value of a sample into a persistent dentry/offset
234 * pair that can then be added to the global event buffer. We make
235 * sure to do this lookup before a mm->mmap modification happens so
236 * we don't lose track.
237 */
238static unsigned long lookup_dcookie(struct mm_struct * mm, unsigned long addr, off_t * offset)
239{
John Levon0c0a4002005-06-23 22:02:47 -0700240 unsigned long cookie = NO_COOKIE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700241 struct vm_area_struct * vma;
242
243 for (vma = find_vma(mm, addr); vma; vma = vma->vm_next) {
244
Linus Torvalds1da177e2005-04-16 15:20:36 -0700245 if (addr < vma->vm_start || addr >= vma->vm_end)
246 continue;
247
John Levon0c0a4002005-06-23 22:02:47 -0700248 if (vma->vm_file) {
Jan Blunck448678a2008-02-14 19:38:36 -0800249 cookie = fast_get_dcookie(&vma->vm_file->f_path);
John Levon0c0a4002005-06-23 22:02:47 -0700250 *offset = (vma->vm_pgoff << PAGE_SHIFT) + addr -
251 vma->vm_start;
252 } else {
253 /* must be an anonymous map */
254 *offset = addr;
255 }
256
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257 break;
258 }
259
John Levon0c0a4002005-06-23 22:02:47 -0700260 if (!vma)
261 cookie = INVALID_COOKIE;
262
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263 return cookie;
264}
265
266
John Levon0c0a4002005-06-23 22:02:47 -0700267static unsigned long last_cookie = INVALID_COOKIE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268
269static void add_cpu_switch(int i)
270{
271 add_event_entry(ESCAPE_CODE);
272 add_event_entry(CPU_SWITCH_CODE);
273 add_event_entry(i);
John Levon0c0a4002005-06-23 22:02:47 -0700274 last_cookie = INVALID_COOKIE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275}
276
277static void add_kernel_ctx_switch(unsigned int in_kernel)
278{
279 add_event_entry(ESCAPE_CODE);
280 if (in_kernel)
281 add_event_entry(KERNEL_ENTER_SWITCH_CODE);
282 else
283 add_event_entry(KERNEL_EXIT_SWITCH_CODE);
284}
285
286static void
287add_user_ctx_switch(struct task_struct const * task, unsigned long cookie)
288{
289 add_event_entry(ESCAPE_CODE);
290 add_event_entry(CTX_SWITCH_CODE);
291 add_event_entry(task->pid);
292 add_event_entry(cookie);
293 /* Another code for daemon back-compat */
294 add_event_entry(ESCAPE_CODE);
295 add_event_entry(CTX_TGID_CODE);
296 add_event_entry(task->tgid);
297}
298
299
300static void add_cookie_switch(unsigned long cookie)
301{
302 add_event_entry(ESCAPE_CODE);
303 add_event_entry(COOKIE_SWITCH_CODE);
304 add_event_entry(cookie);
305}
306
307
308static void add_trace_begin(void)
309{
310 add_event_entry(ESCAPE_CODE);
311 add_event_entry(TRACE_BEGIN_CODE);
312}
313
314
315static void add_sample_entry(unsigned long offset, unsigned long event)
316{
317 add_event_entry(offset);
318 add_event_entry(event);
319}
320
321
322static int add_us_sample(struct mm_struct * mm, struct op_sample * s)
323{
324 unsigned long cookie;
325 off_t offset;
326
327 cookie = lookup_dcookie(mm, s->eip, &offset);
328
John Levon0c0a4002005-06-23 22:02:47 -0700329 if (cookie == INVALID_COOKIE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700330 atomic_inc(&oprofile_stats.sample_lost_no_mapping);
331 return 0;
332 }
333
334 if (cookie != last_cookie) {
335 add_cookie_switch(cookie);
336 last_cookie = cookie;
337 }
338
339 add_sample_entry(offset, s->event);
340
341 return 1;
342}
343
344
345/* Add a sample to the global event buffer. If possible the
346 * sample is converted into a persistent dentry/offset pair
347 * for later lookup from userspace.
348 */
349static int
350add_sample(struct mm_struct * mm, struct op_sample * s, int in_kernel)
351{
352 if (in_kernel) {
353 add_sample_entry(s->eip, s->event);
354 return 1;
355 } else if (mm) {
356 return add_us_sample(mm, s);
357 } else {
358 atomic_inc(&oprofile_stats.sample_lost_no_mm);
359 }
360 return 0;
361}
362
363
364static void release_mm(struct mm_struct * mm)
365{
366 if (!mm)
367 return;
368 up_read(&mm->mmap_sem);
369 mmput(mm);
370}
371
372
373static struct mm_struct * take_tasks_mm(struct task_struct * task)
374{
375 struct mm_struct * mm = get_task_mm(task);
376 if (mm)
377 down_read(&mm->mmap_sem);
378 return mm;
379}
380
381
382static inline int is_code(unsigned long val)
383{
384 return val == ESCAPE_CODE;
385}
386
387
388/* "acquire" as many cpu buffer slots as we can */
389static unsigned long get_slots(struct oprofile_cpu_buffer * b)
390{
391 unsigned long head = b->head_pos;
392 unsigned long tail = b->tail_pos;
393
394 /*
395 * Subtle. This resets the persistent last_task
396 * and in_kernel values used for switching notes.
397 * BUT, there is a small window between reading
398 * head_pos, and this call, that means samples
399 * can appear at the new head position, but not
400 * be prefixed with the notes for switching
401 * kernel mode or a task switch. This small hole
402 * can lead to mis-attribution or samples where
403 * we don't know if it's in the kernel or not,
404 * at the start of an event buffer.
405 */
406 cpu_buffer_reset(b);
407
408 if (head >= tail)
409 return head - tail;
410
411 return head + (b->buffer_size - tail);
412}
413
414
415static void increment_tail(struct oprofile_cpu_buffer * b)
416{
417 unsigned long new_tail = b->tail_pos + 1;
418
419 rmb();
420
421 if (new_tail < b->buffer_size)
422 b->tail_pos = new_tail;
423 else
424 b->tail_pos = 0;
425}
426
427
428/* Move tasks along towards death. Any tasks on dead_tasks
429 * will definitely have no remaining references in any
430 * CPU buffers at this point, because we use two lists,
431 * and to have reached the list, it must have gone through
432 * one full sync already.
433 */
434static void process_task_mortuary(void)
435{
Paul E. McKenney4369ef32006-01-08 01:01:35 -0800436 unsigned long flags;
437 LIST_HEAD(local_dead_tasks);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700438 struct task_struct * task;
Paul E. McKenney4369ef32006-01-08 01:01:35 -0800439 struct task_struct * ttask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440
Paul E. McKenney4369ef32006-01-08 01:01:35 -0800441 spin_lock_irqsave(&task_mortuary, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442
Paul E. McKenney4369ef32006-01-08 01:01:35 -0800443 list_splice_init(&dead_tasks, &local_dead_tasks);
444 list_splice_init(&dying_tasks, &dead_tasks);
445
446 spin_unlock_irqrestore(&task_mortuary, flags);
447
448 list_for_each_entry_safe(task, ttask, &local_dead_tasks, tasks) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449 list_del(&task->tasks);
450 free_task(task);
451 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452}
453
454
455static void mark_done(int cpu)
456{
457 int i;
458
459 cpu_set(cpu, marked_cpus);
460
461 for_each_online_cpu(i) {
462 if (!cpu_isset(i, marked_cpus))
463 return;
464 }
465
466 /* All CPUs have been processed at least once,
467 * we can process the mortuary once
468 */
469 process_task_mortuary();
470
471 cpus_clear(marked_cpus);
472}
473
474
475/* FIXME: this is not sufficient if we implement syscall barrier backtrace
476 * traversal, the code switch to sb_sample_start at first kernel enter/exit
477 * switch so we need a fifth state and some special handling in sync_buffer()
478 */
479typedef enum {
480 sb_bt_ignore = -2,
481 sb_buffer_start,
482 sb_bt_start,
483 sb_sample_start,
484} sync_buffer_state;
485
486/* Sync one of the CPU's buffers into the global event buffer.
487 * Here we need to go through each batch of samples punctuated
488 * by context switch notes, taking the task's mmap_sem and doing
489 * lookup in task->mm->mmap to convert EIP into dcookie/offset
490 * value.
491 */
492void sync_buffer(int cpu)
493{
494 struct oprofile_cpu_buffer * cpu_buf = &cpu_buffer[cpu];
495 struct mm_struct *mm = NULL;
496 struct task_struct * new;
497 unsigned long cookie = 0;
498 int in_kernel = 1;
499 unsigned int i;
500 sync_buffer_state state = sb_buffer_start;
501 unsigned long available;
502
Markus Armbruster59cc1852006-06-25 05:47:33 -0700503 mutex_lock(&buffer_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700504
505 add_cpu_switch(cpu);
506
507 /* Remember, only we can modify tail_pos */
508
509 available = get_slots(cpu_buf);
510
511 for (i = 0; i < available; ++i) {
512 struct op_sample * s = &cpu_buf->buffer[cpu_buf->tail_pos];
513
514 if (is_code(s->eip)) {
515 if (s->event <= CPU_IS_KERNEL) {
516 /* kernel/userspace switch */
517 in_kernel = s->event;
518 if (state == sb_buffer_start)
519 state = sb_sample_start;
520 add_kernel_ctx_switch(s->event);
521 } else if (s->event == CPU_TRACE_BEGIN) {
522 state = sb_bt_start;
523 add_trace_begin();
524 } else {
525 struct mm_struct * oldmm = mm;
526
527 /* userspace context switch */
528 new = (struct task_struct *)s->event;
529
530 release_mm(oldmm);
531 mm = take_tasks_mm(new);
532 if (mm != oldmm)
533 cookie = get_exec_dcookie(mm);
534 add_user_ctx_switch(new, cookie);
535 }
536 } else {
537 if (state >= sb_bt_start &&
538 !add_sample(mm, s, in_kernel)) {
539 if (state == sb_bt_start) {
540 state = sb_bt_ignore;
541 atomic_inc(&oprofile_stats.bt_lost_no_mapping);
542 }
543 }
544 }
545
546 increment_tail(cpu_buf);
547 }
548 release_mm(mm);
549
550 mark_done(cpu);
551
Markus Armbruster59cc1852006-06-25 05:47:33 -0700552 mutex_unlock(&buffer_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553}