Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /** |
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> | |
29 | ||
30 | #include "oprofile_stats.h" | |
31 | #include "event_buffer.h" | |
32 | #include "cpu_buffer.h" | |
33 | #include "buffer_sync.h" | |
34 | ||
35 | static LIST_HEAD(dying_tasks); | |
36 | static LIST_HEAD(dead_tasks); | |
37 | static cpumask_t marked_cpus = CPU_MASK_NONE; | |
38 | static DEFINE_SPINLOCK(task_mortuary); | |
39 | static void process_task_mortuary(void); | |
40 | ||
41 | ||
42 | /* Take ownership of the task struct and place it on the | |
43 | * list for processing. Only after two full buffer syncs | |
44 | * does the task eventually get freed, because by then | |
45 | * we are sure we will not reference it again. | |
46 | */ | |
47 | static int task_free_notify(struct notifier_block * self, unsigned long val, void * data) | |
48 | { | |
49 | struct task_struct * task = data; | |
50 | spin_lock(&task_mortuary); | |
51 | list_add(&task->tasks, &dying_tasks); | |
52 | spin_unlock(&task_mortuary); | |
53 | return NOTIFY_OK; | |
54 | } | |
55 | ||
56 | ||
57 | /* The task is on its way out. A sync of the buffer means we can catch | |
58 | * any remaining samples for this task. | |
59 | */ | |
60 | static int task_exit_notify(struct notifier_block * self, unsigned long val, void * data) | |
61 | { | |
62 | /* To avoid latency problems, we only process the current CPU, | |
63 | * hoping that most samples for the task are on this CPU | |
64 | */ | |
39c715b7 | 65 | sync_buffer(raw_smp_processor_id()); |
1da177e4 LT |
66 | return 0; |
67 | } | |
68 | ||
69 | ||
70 | /* The task is about to try a do_munmap(). We peek at what it's going to | |
71 | * do, and if it's an executable region, process the samples first, so | |
72 | * we don't lose any. This does not have to be exact, it's a QoI issue | |
73 | * only. | |
74 | */ | |
75 | static int munmap_notify(struct notifier_block * self, unsigned long val, void * data) | |
76 | { | |
77 | unsigned long addr = (unsigned long)data; | |
78 | struct mm_struct * mm = current->mm; | |
79 | struct vm_area_struct * mpnt; | |
80 | ||
81 | down_read(&mm->mmap_sem); | |
82 | ||
83 | mpnt = find_vma(mm, addr); | |
84 | if (mpnt && mpnt->vm_file && (mpnt->vm_flags & VM_EXEC)) { | |
85 | up_read(&mm->mmap_sem); | |
86 | /* To avoid latency problems, we only process the current CPU, | |
87 | * hoping that most samples for the task are on this CPU | |
88 | */ | |
39c715b7 | 89 | sync_buffer(raw_smp_processor_id()); |
1da177e4 LT |
90 | return 0; |
91 | } | |
92 | ||
93 | up_read(&mm->mmap_sem); | |
94 | return 0; | |
95 | } | |
96 | ||
97 | ||
98 | /* We need to be told about new modules so we don't attribute to a previously | |
99 | * loaded module, or drop the samples on the floor. | |
100 | */ | |
101 | static int module_load_notify(struct notifier_block * self, unsigned long val, void * data) | |
102 | { | |
103 | #ifdef CONFIG_MODULES | |
104 | if (val != MODULE_STATE_COMING) | |
105 | return 0; | |
106 | ||
107 | /* FIXME: should we process all CPU buffers ? */ | |
108 | down(&buffer_sem); | |
109 | add_event_entry(ESCAPE_CODE); | |
110 | add_event_entry(MODULE_LOADED_CODE); | |
111 | up(&buffer_sem); | |
112 | #endif | |
113 | return 0; | |
114 | } | |
115 | ||
116 | ||
117 | static struct notifier_block task_free_nb = { | |
118 | .notifier_call = task_free_notify, | |
119 | }; | |
120 | ||
121 | static struct notifier_block task_exit_nb = { | |
122 | .notifier_call = task_exit_notify, | |
123 | }; | |
124 | ||
125 | static struct notifier_block munmap_nb = { | |
126 | .notifier_call = munmap_notify, | |
127 | }; | |
128 | ||
129 | static struct notifier_block module_load_nb = { | |
130 | .notifier_call = module_load_notify, | |
131 | }; | |
132 | ||
133 | ||
134 | static void end_sync(void) | |
135 | { | |
136 | end_cpu_work(); | |
137 | /* make sure we don't leak task structs */ | |
138 | process_task_mortuary(); | |
139 | process_task_mortuary(); | |
140 | } | |
141 | ||
142 | ||
143 | int sync_start(void) | |
144 | { | |
145 | int err; | |
146 | ||
147 | start_cpu_work(); | |
148 | ||
149 | err = task_handoff_register(&task_free_nb); | |
150 | if (err) | |
151 | goto out1; | |
152 | err = profile_event_register(PROFILE_TASK_EXIT, &task_exit_nb); | |
153 | if (err) | |
154 | goto out2; | |
155 | err = profile_event_register(PROFILE_MUNMAP, &munmap_nb); | |
156 | if (err) | |
157 | goto out3; | |
158 | err = register_module_notifier(&module_load_nb); | |
159 | if (err) | |
160 | goto out4; | |
161 | ||
162 | out: | |
163 | return err; | |
164 | out4: | |
165 | profile_event_unregister(PROFILE_MUNMAP, &munmap_nb); | |
166 | out3: | |
167 | profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb); | |
168 | out2: | |
169 | task_handoff_unregister(&task_free_nb); | |
170 | out1: | |
171 | end_sync(); | |
172 | goto out; | |
173 | } | |
174 | ||
175 | ||
176 | void sync_stop(void) | |
177 | { | |
178 | unregister_module_notifier(&module_load_nb); | |
179 | profile_event_unregister(PROFILE_MUNMAP, &munmap_nb); | |
180 | profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb); | |
181 | task_handoff_unregister(&task_free_nb); | |
182 | end_sync(); | |
183 | } | |
184 | ||
185 | ||
186 | /* Optimisation. We can manage without taking the dcookie sem | |
187 | * because we cannot reach this code without at least one | |
188 | * dcookie user still being registered (namely, the reader | |
189 | * of the event buffer). */ | |
190 | static inline unsigned long fast_get_dcookie(struct dentry * dentry, | |
191 | struct vfsmount * vfsmnt) | |
192 | { | |
193 | unsigned long cookie; | |
194 | ||
195 | if (dentry->d_cookie) | |
196 | return (unsigned long)dentry; | |
197 | get_dcookie(dentry, vfsmnt, &cookie); | |
198 | return cookie; | |
199 | } | |
200 | ||
201 | ||
202 | /* Look up the dcookie for the task's first VM_EXECUTABLE mapping, | |
203 | * which corresponds loosely to "application name". This is | |
204 | * not strictly necessary but allows oprofile to associate | |
205 | * shared-library samples with particular applications | |
206 | */ | |
207 | static unsigned long get_exec_dcookie(struct mm_struct * mm) | |
208 | { | |
0c0a400d | 209 | unsigned long cookie = NO_COOKIE; |
1da177e4 LT |
210 | struct vm_area_struct * vma; |
211 | ||
212 | if (!mm) | |
213 | goto out; | |
214 | ||
215 | for (vma = mm->mmap; vma; vma = vma->vm_next) { | |
216 | if (!vma->vm_file) | |
217 | continue; | |
218 | if (!(vma->vm_flags & VM_EXECUTABLE)) | |
219 | continue; | |
220 | cookie = fast_get_dcookie(vma->vm_file->f_dentry, | |
221 | vma->vm_file->f_vfsmnt); | |
222 | break; | |
223 | } | |
224 | ||
225 | out: | |
226 | return cookie; | |
227 | } | |
228 | ||
229 | ||
230 | /* Convert the EIP value of a sample into a persistent dentry/offset | |
231 | * pair that can then be added to the global event buffer. We make | |
232 | * sure to do this lookup before a mm->mmap modification happens so | |
233 | * we don't lose track. | |
234 | */ | |
235 | static unsigned long lookup_dcookie(struct mm_struct * mm, unsigned long addr, off_t * offset) | |
236 | { | |
0c0a400d | 237 | unsigned long cookie = NO_COOKIE; |
1da177e4 LT |
238 | struct vm_area_struct * vma; |
239 | ||
240 | for (vma = find_vma(mm, addr); vma; vma = vma->vm_next) { | |
241 | ||
1da177e4 LT |
242 | if (addr < vma->vm_start || addr >= vma->vm_end) |
243 | continue; | |
244 | ||
0c0a400d JL |
245 | if (vma->vm_file) { |
246 | cookie = fast_get_dcookie(vma->vm_file->f_dentry, | |
247 | vma->vm_file->f_vfsmnt); | |
248 | *offset = (vma->vm_pgoff << PAGE_SHIFT) + addr - | |
249 | vma->vm_start; | |
250 | } else { | |
251 | /* must be an anonymous map */ | |
252 | *offset = addr; | |
253 | } | |
254 | ||
1da177e4 LT |
255 | break; |
256 | } | |
257 | ||
0c0a400d JL |
258 | if (!vma) |
259 | cookie = INVALID_COOKIE; | |
260 | ||
1da177e4 LT |
261 | return cookie; |
262 | } | |
263 | ||
264 | ||
0c0a400d | 265 | static unsigned long last_cookie = INVALID_COOKIE; |
1da177e4 LT |
266 | |
267 | static void add_cpu_switch(int i) | |
268 | { | |
269 | add_event_entry(ESCAPE_CODE); | |
270 | add_event_entry(CPU_SWITCH_CODE); | |
271 | add_event_entry(i); | |
0c0a400d | 272 | last_cookie = INVALID_COOKIE; |
1da177e4 LT |
273 | } |
274 | ||
275 | static void add_kernel_ctx_switch(unsigned int in_kernel) | |
276 | { | |
277 | add_event_entry(ESCAPE_CODE); | |
278 | if (in_kernel) | |
279 | add_event_entry(KERNEL_ENTER_SWITCH_CODE); | |
280 | else | |
281 | add_event_entry(KERNEL_EXIT_SWITCH_CODE); | |
282 | } | |
283 | ||
284 | static void | |
285 | add_user_ctx_switch(struct task_struct const * task, unsigned long cookie) | |
286 | { | |
287 | add_event_entry(ESCAPE_CODE); | |
288 | add_event_entry(CTX_SWITCH_CODE); | |
289 | add_event_entry(task->pid); | |
290 | add_event_entry(cookie); | |
291 | /* Another code for daemon back-compat */ | |
292 | add_event_entry(ESCAPE_CODE); | |
293 | add_event_entry(CTX_TGID_CODE); | |
294 | add_event_entry(task->tgid); | |
295 | } | |
296 | ||
297 | ||
298 | static void add_cookie_switch(unsigned long cookie) | |
299 | { | |
300 | add_event_entry(ESCAPE_CODE); | |
301 | add_event_entry(COOKIE_SWITCH_CODE); | |
302 | add_event_entry(cookie); | |
303 | } | |
304 | ||
305 | ||
306 | static void add_trace_begin(void) | |
307 | { | |
308 | add_event_entry(ESCAPE_CODE); | |
309 | add_event_entry(TRACE_BEGIN_CODE); | |
310 | } | |
311 | ||
312 | ||
313 | static void add_sample_entry(unsigned long offset, unsigned long event) | |
314 | { | |
315 | add_event_entry(offset); | |
316 | add_event_entry(event); | |
317 | } | |
318 | ||
319 | ||
320 | static int add_us_sample(struct mm_struct * mm, struct op_sample * s) | |
321 | { | |
322 | unsigned long cookie; | |
323 | off_t offset; | |
324 | ||
325 | cookie = lookup_dcookie(mm, s->eip, &offset); | |
326 | ||
0c0a400d | 327 | if (cookie == INVALID_COOKIE) { |
1da177e4 LT |
328 | atomic_inc(&oprofile_stats.sample_lost_no_mapping); |
329 | return 0; | |
330 | } | |
331 | ||
332 | if (cookie != last_cookie) { | |
333 | add_cookie_switch(cookie); | |
334 | last_cookie = cookie; | |
335 | } | |
336 | ||
337 | add_sample_entry(offset, s->event); | |
338 | ||
339 | return 1; | |
340 | } | |
341 | ||
342 | ||
343 | /* Add a sample to the global event buffer. If possible the | |
344 | * sample is converted into a persistent dentry/offset pair | |
345 | * for later lookup from userspace. | |
346 | */ | |
347 | static int | |
348 | add_sample(struct mm_struct * mm, struct op_sample * s, int in_kernel) | |
349 | { | |
350 | if (in_kernel) { | |
351 | add_sample_entry(s->eip, s->event); | |
352 | return 1; | |
353 | } else if (mm) { | |
354 | return add_us_sample(mm, s); | |
355 | } else { | |
356 | atomic_inc(&oprofile_stats.sample_lost_no_mm); | |
357 | } | |
358 | return 0; | |
359 | } | |
360 | ||
361 | ||
362 | static void release_mm(struct mm_struct * mm) | |
363 | { | |
364 | if (!mm) | |
365 | return; | |
366 | up_read(&mm->mmap_sem); | |
367 | mmput(mm); | |
368 | } | |
369 | ||
370 | ||
371 | static struct mm_struct * take_tasks_mm(struct task_struct * task) | |
372 | { | |
373 | struct mm_struct * mm = get_task_mm(task); | |
374 | if (mm) | |
375 | down_read(&mm->mmap_sem); | |
376 | return mm; | |
377 | } | |
378 | ||
379 | ||
380 | static inline int is_code(unsigned long val) | |
381 | { | |
382 | return val == ESCAPE_CODE; | |
383 | } | |
384 | ||
385 | ||
386 | /* "acquire" as many cpu buffer slots as we can */ | |
387 | static unsigned long get_slots(struct oprofile_cpu_buffer * b) | |
388 | { | |
389 | unsigned long head = b->head_pos; | |
390 | unsigned long tail = b->tail_pos; | |
391 | ||
392 | /* | |
393 | * Subtle. This resets the persistent last_task | |
394 | * and in_kernel values used for switching notes. | |
395 | * BUT, there is a small window between reading | |
396 | * head_pos, and this call, that means samples | |
397 | * can appear at the new head position, but not | |
398 | * be prefixed with the notes for switching | |
399 | * kernel mode or a task switch. This small hole | |
400 | * can lead to mis-attribution or samples where | |
401 | * we don't know if it's in the kernel or not, | |
402 | * at the start of an event buffer. | |
403 | */ | |
404 | cpu_buffer_reset(b); | |
405 | ||
406 | if (head >= tail) | |
407 | return head - tail; | |
408 | ||
409 | return head + (b->buffer_size - tail); | |
410 | } | |
411 | ||
412 | ||
413 | static void increment_tail(struct oprofile_cpu_buffer * b) | |
414 | { | |
415 | unsigned long new_tail = b->tail_pos + 1; | |
416 | ||
417 | rmb(); | |
418 | ||
419 | if (new_tail < b->buffer_size) | |
420 | b->tail_pos = new_tail; | |
421 | else | |
422 | b->tail_pos = 0; | |
423 | } | |
424 | ||
425 | ||
426 | /* Move tasks along towards death. Any tasks on dead_tasks | |
427 | * will definitely have no remaining references in any | |
428 | * CPU buffers at this point, because we use two lists, | |
429 | * and to have reached the list, it must have gone through | |
430 | * one full sync already. | |
431 | */ | |
432 | static void process_task_mortuary(void) | |
433 | { | |
434 | struct list_head * pos; | |
435 | struct list_head * pos2; | |
436 | struct task_struct * task; | |
437 | ||
438 | spin_lock(&task_mortuary); | |
439 | ||
440 | list_for_each_safe(pos, pos2, &dead_tasks) { | |
441 | task = list_entry(pos, struct task_struct, tasks); | |
442 | list_del(&task->tasks); | |
443 | free_task(task); | |
444 | } | |
445 | ||
446 | list_for_each_safe(pos, pos2, &dying_tasks) { | |
447 | task = list_entry(pos, struct task_struct, tasks); | |
448 | list_del(&task->tasks); | |
449 | list_add_tail(&task->tasks, &dead_tasks); | |
450 | } | |
451 | ||
452 | spin_unlock(&task_mortuary); | |
453 | } | |
454 | ||
455 | ||
456 | static void mark_done(int cpu) | |
457 | { | |
458 | int i; | |
459 | ||
460 | cpu_set(cpu, marked_cpus); | |
461 | ||
462 | for_each_online_cpu(i) { | |
463 | if (!cpu_isset(i, marked_cpus)) | |
464 | return; | |
465 | } | |
466 | ||
467 | /* All CPUs have been processed at least once, | |
468 | * we can process the mortuary once | |
469 | */ | |
470 | process_task_mortuary(); | |
471 | ||
472 | cpus_clear(marked_cpus); | |
473 | } | |
474 | ||
475 | ||
476 | /* FIXME: this is not sufficient if we implement syscall barrier backtrace | |
477 | * traversal, the code switch to sb_sample_start at first kernel enter/exit | |
478 | * switch so we need a fifth state and some special handling in sync_buffer() | |
479 | */ | |
480 | typedef enum { | |
481 | sb_bt_ignore = -2, | |
482 | sb_buffer_start, | |
483 | sb_bt_start, | |
484 | sb_sample_start, | |
485 | } sync_buffer_state; | |
486 | ||
487 | /* Sync one of the CPU's buffers into the global event buffer. | |
488 | * Here we need to go through each batch of samples punctuated | |
489 | * by context switch notes, taking the task's mmap_sem and doing | |
490 | * lookup in task->mm->mmap to convert EIP into dcookie/offset | |
491 | * value. | |
492 | */ | |
493 | void sync_buffer(int cpu) | |
494 | { | |
495 | struct oprofile_cpu_buffer * cpu_buf = &cpu_buffer[cpu]; | |
496 | struct mm_struct *mm = NULL; | |
497 | struct task_struct * new; | |
498 | unsigned long cookie = 0; | |
499 | int in_kernel = 1; | |
500 | unsigned int i; | |
501 | sync_buffer_state state = sb_buffer_start; | |
502 | unsigned long available; | |
503 | ||
504 | down(&buffer_sem); | |
505 | ||
506 | add_cpu_switch(cpu); | |
507 | ||
508 | /* Remember, only we can modify tail_pos */ | |
509 | ||
510 | available = get_slots(cpu_buf); | |
511 | ||
512 | for (i = 0; i < available; ++i) { | |
513 | struct op_sample * s = &cpu_buf->buffer[cpu_buf->tail_pos]; | |
514 | ||
515 | if (is_code(s->eip)) { | |
516 | if (s->event <= CPU_IS_KERNEL) { | |
517 | /* kernel/userspace switch */ | |
518 | in_kernel = s->event; | |
519 | if (state == sb_buffer_start) | |
520 | state = sb_sample_start; | |
521 | add_kernel_ctx_switch(s->event); | |
522 | } else if (s->event == CPU_TRACE_BEGIN) { | |
523 | state = sb_bt_start; | |
524 | add_trace_begin(); | |
525 | } else { | |
526 | struct mm_struct * oldmm = mm; | |
527 | ||
528 | /* userspace context switch */ | |
529 | new = (struct task_struct *)s->event; | |
530 | ||
531 | release_mm(oldmm); | |
532 | mm = take_tasks_mm(new); | |
533 | if (mm != oldmm) | |
534 | cookie = get_exec_dcookie(mm); | |
535 | add_user_ctx_switch(new, cookie); | |
536 | } | |
537 | } else { | |
538 | if (state >= sb_bt_start && | |
539 | !add_sample(mm, s, in_kernel)) { | |
540 | if (state == sb_bt_start) { | |
541 | state = sb_bt_ignore; | |
542 | atomic_inc(&oprofile_stats.bt_lost_no_mapping); | |
543 | } | |
544 | } | |
545 | } | |
546 | ||
547 | increment_tail(cpu_buf); | |
548 | } | |
549 | release_mm(mm); | |
550 | ||
551 | mark_done(cpu); | |
552 | ||
553 | up(&buffer_sem); | |
554 | } |