]>
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> | |
1474855d | 29 | #include <linux/oprofile.h> |
e8edc6e0 | 30 | #include <linux/sched.h> |
1474855d | 31 | |
1da177e4 LT |
32 | #include "oprofile_stats.h" |
33 | #include "event_buffer.h" | |
34 | #include "cpu_buffer.h" | |
35 | #include "buffer_sync.h" | |
36 | ||
37 | static LIST_HEAD(dying_tasks); | |
38 | static LIST_HEAD(dead_tasks); | |
39 | static cpumask_t marked_cpus = CPU_MASK_NONE; | |
40 | static DEFINE_SPINLOCK(task_mortuary); | |
41 | static 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. | |
4369ef3c PM |
48 | * Can be invoked from softirq via RCU callback due to |
49 | * call_rcu() of the task struct, hence the _irqsave. | |
1da177e4 LT |
50 | */ |
51 | static int task_free_notify(struct notifier_block * self, unsigned long val, void * data) | |
52 | { | |
4369ef3c | 53 | unsigned long flags; |
1da177e4 | 54 | struct task_struct * task = data; |
4369ef3c | 55 | spin_lock_irqsave(&task_mortuary, flags); |
1da177e4 | 56 | list_add(&task->tasks, &dying_tasks); |
4369ef3c | 57 | spin_unlock_irqrestore(&task_mortuary, flags); |
1da177e4 LT |
58 | 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 | */ | |
65 | static 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 | */ | |
39c715b7 | 70 | sync_buffer(raw_smp_processor_id()); |
1da177e4 LT |
71 | 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 | */ | |
80 | static 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 | */ | |
39c715b7 | 94 | sync_buffer(raw_smp_processor_id()); |
1da177e4 LT |
95 | 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 | */ | |
106 | static 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 ? */ | |
59cc185a | 113 | mutex_lock(&buffer_mutex); |
1da177e4 LT |
114 | add_event_entry(ESCAPE_CODE); |
115 | add_event_entry(MODULE_LOADED_CODE); | |
59cc185a | 116 | mutex_unlock(&buffer_mutex); |
1da177e4 LT |
117 | #endif |
118 | return 0; | |
119 | } | |
120 | ||
121 | ||
122 | static struct notifier_block task_free_nb = { | |
123 | .notifier_call = task_free_notify, | |
124 | }; | |
125 | ||
126 | static struct notifier_block task_exit_nb = { | |
127 | .notifier_call = task_exit_notify, | |
128 | }; | |
129 | ||
130 | static struct notifier_block munmap_nb = { | |
131 | .notifier_call = munmap_notify, | |
132 | }; | |
133 | ||
134 | static struct notifier_block module_load_nb = { | |
135 | .notifier_call = module_load_notify, | |
136 | }; | |
137 | ||
138 | ||
139 | static 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 | ||
148 | int 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 | ||
167 | out: | |
168 | return err; | |
169 | out4: | |
170 | profile_event_unregister(PROFILE_MUNMAP, &munmap_nb); | |
171 | out3: | |
172 | profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb); | |
173 | out2: | |
174 | task_handoff_unregister(&task_free_nb); | |
175 | out1: | |
176 | end_sync(); | |
177 | goto out; | |
178 | } | |
179 | ||
180 | ||
181 | void 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 | ||
190 | ||
191 | /* 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). */ | |
195 | static inline unsigned long fast_get_dcookie(struct dentry * dentry, | |
196 | struct vfsmount * vfsmnt) | |
197 | { | |
198 | unsigned long cookie; | |
199 | ||
200 | if (dentry->d_cookie) | |
201 | return (unsigned long)dentry; | |
202 | get_dcookie(dentry, vfsmnt, &cookie); | |
203 | return cookie; | |
204 | } | |
205 | ||
206 | ||
207 | /* Look up the dcookie for the task's first VM_EXECUTABLE mapping, | |
208 | * which corresponds loosely to "application name". This is | |
209 | * not strictly necessary but allows oprofile to associate | |
210 | * shared-library samples with particular applications | |
211 | */ | |
212 | static unsigned long get_exec_dcookie(struct mm_struct * mm) | |
213 | { | |
0c0a400d | 214 | unsigned long cookie = NO_COOKIE; |
1da177e4 LT |
215 | struct vm_area_struct * vma; |
216 | ||
217 | if (!mm) | |
218 | goto out; | |
219 | ||
220 | for (vma = mm->mmap; vma; vma = vma->vm_next) { | |
221 | if (!vma->vm_file) | |
222 | continue; | |
223 | if (!(vma->vm_flags & VM_EXECUTABLE)) | |
224 | continue; | |
1fb1430b JS |
225 | cookie = fast_get_dcookie(vma->vm_file->f_path.dentry, |
226 | vma->vm_file->f_path.mnt); | |
1da177e4 LT |
227 | break; |
228 | } | |
229 | ||
230 | out: | |
231 | return cookie; | |
232 | } | |
233 | ||
234 | ||
235 | /* Convert the EIP value of a sample into a persistent dentry/offset | |
236 | * pair that can then be added to the global event buffer. We make | |
237 | * sure to do this lookup before a mm->mmap modification happens so | |
238 | * we don't lose track. | |
239 | */ | |
240 | static unsigned long lookup_dcookie(struct mm_struct * mm, unsigned long addr, off_t * offset) | |
241 | { | |
0c0a400d | 242 | unsigned long cookie = NO_COOKIE; |
1da177e4 LT |
243 | struct vm_area_struct * vma; |
244 | ||
245 | for (vma = find_vma(mm, addr); vma; vma = vma->vm_next) { | |
246 | ||
1da177e4 LT |
247 | if (addr < vma->vm_start || addr >= vma->vm_end) |
248 | continue; | |
249 | ||
0c0a400d | 250 | if (vma->vm_file) { |
1fb1430b JS |
251 | cookie = fast_get_dcookie(vma->vm_file->f_path.dentry, |
252 | vma->vm_file->f_path.mnt); | |
0c0a400d JL |
253 | *offset = (vma->vm_pgoff << PAGE_SHIFT) + addr - |
254 | vma->vm_start; | |
255 | } else { | |
256 | /* must be an anonymous map */ | |
257 | *offset = addr; | |
258 | } | |
259 | ||
1da177e4 LT |
260 | break; |
261 | } | |
262 | ||
0c0a400d JL |
263 | if (!vma) |
264 | cookie = INVALID_COOKIE; | |
265 | ||
1da177e4 LT |
266 | return cookie; |
267 | } | |
268 | ||
269 | ||
0c0a400d | 270 | static unsigned long last_cookie = INVALID_COOKIE; |
1da177e4 LT |
271 | |
272 | static void add_cpu_switch(int i) | |
273 | { | |
274 | add_event_entry(ESCAPE_CODE); | |
275 | add_event_entry(CPU_SWITCH_CODE); | |
276 | add_event_entry(i); | |
0c0a400d | 277 | last_cookie = INVALID_COOKIE; |
1da177e4 LT |
278 | } |
279 | ||
280 | static void add_kernel_ctx_switch(unsigned int in_kernel) | |
281 | { | |
282 | add_event_entry(ESCAPE_CODE); | |
283 | if (in_kernel) | |
284 | add_event_entry(KERNEL_ENTER_SWITCH_CODE); | |
285 | else | |
286 | add_event_entry(KERNEL_EXIT_SWITCH_CODE); | |
287 | } | |
288 | ||
289 | static void | |
290 | add_user_ctx_switch(struct task_struct const * task, unsigned long cookie) | |
291 | { | |
292 | add_event_entry(ESCAPE_CODE); | |
293 | add_event_entry(CTX_SWITCH_CODE); | |
294 | add_event_entry(task->pid); | |
295 | add_event_entry(cookie); | |
296 | /* Another code for daemon back-compat */ | |
297 | add_event_entry(ESCAPE_CODE); | |
298 | add_event_entry(CTX_TGID_CODE); | |
299 | add_event_entry(task->tgid); | |
300 | } | |
301 | ||
302 | ||
303 | static void add_cookie_switch(unsigned long cookie) | |
304 | { | |
305 | add_event_entry(ESCAPE_CODE); | |
306 | add_event_entry(COOKIE_SWITCH_CODE); | |
307 | add_event_entry(cookie); | |
308 | } | |
309 | ||
310 | ||
311 | static void add_trace_begin(void) | |
312 | { | |
313 | add_event_entry(ESCAPE_CODE); | |
314 | add_event_entry(TRACE_BEGIN_CODE); | |
315 | } | |
316 | ||
317 | ||
318 | static void add_sample_entry(unsigned long offset, unsigned long event) | |
319 | { | |
320 | add_event_entry(offset); | |
321 | add_event_entry(event); | |
322 | } | |
323 | ||
324 | ||
325 | static int add_us_sample(struct mm_struct * mm, struct op_sample * s) | |
326 | { | |
327 | unsigned long cookie; | |
328 | off_t offset; | |
329 | ||
330 | cookie = lookup_dcookie(mm, s->eip, &offset); | |
331 | ||
0c0a400d | 332 | if (cookie == INVALID_COOKIE) { |
1da177e4 LT |
333 | atomic_inc(&oprofile_stats.sample_lost_no_mapping); |
334 | return 0; | |
335 | } | |
336 | ||
337 | if (cookie != last_cookie) { | |
338 | add_cookie_switch(cookie); | |
339 | last_cookie = cookie; | |
340 | } | |
341 | ||
342 | add_sample_entry(offset, s->event); | |
343 | ||
344 | return 1; | |
345 | } | |
346 | ||
347 | ||
348 | /* Add a sample to the global event buffer. If possible the | |
349 | * sample is converted into a persistent dentry/offset pair | |
350 | * for later lookup from userspace. | |
351 | */ | |
352 | static int | |
353 | add_sample(struct mm_struct * mm, struct op_sample * s, int in_kernel) | |
354 | { | |
355 | if (in_kernel) { | |
356 | add_sample_entry(s->eip, s->event); | |
357 | return 1; | |
358 | } else if (mm) { | |
359 | return add_us_sample(mm, s); | |
360 | } else { | |
361 | atomic_inc(&oprofile_stats.sample_lost_no_mm); | |
362 | } | |
363 | return 0; | |
364 | } | |
365 | ||
366 | ||
367 | static void release_mm(struct mm_struct * mm) | |
368 | { | |
369 | if (!mm) | |
370 | return; | |
371 | up_read(&mm->mmap_sem); | |
372 | mmput(mm); | |
373 | } | |
374 | ||
375 | ||
376 | static struct mm_struct * take_tasks_mm(struct task_struct * task) | |
377 | { | |
378 | struct mm_struct * mm = get_task_mm(task); | |
379 | if (mm) | |
380 | down_read(&mm->mmap_sem); | |
381 | return mm; | |
382 | } | |
383 | ||
384 | ||
385 | static inline int is_code(unsigned long val) | |
386 | { | |
387 | return val == ESCAPE_CODE; | |
388 | } | |
389 | ||
390 | ||
391 | /* "acquire" as many cpu buffer slots as we can */ | |
392 | static unsigned long get_slots(struct oprofile_cpu_buffer * b) | |
393 | { | |
394 | unsigned long head = b->head_pos; | |
395 | unsigned long tail = b->tail_pos; | |
396 | ||
397 | /* | |
398 | * Subtle. This resets the persistent last_task | |
399 | * and in_kernel values used for switching notes. | |
400 | * BUT, there is a small window between reading | |
401 | * head_pos, and this call, that means samples | |
402 | * can appear at the new head position, but not | |
403 | * be prefixed with the notes for switching | |
404 | * kernel mode or a task switch. This small hole | |
405 | * can lead to mis-attribution or samples where | |
406 | * we don't know if it's in the kernel or not, | |
407 | * at the start of an event buffer. | |
408 | */ | |
409 | cpu_buffer_reset(b); | |
410 | ||
411 | if (head >= tail) | |
412 | return head - tail; | |
413 | ||
414 | return head + (b->buffer_size - tail); | |
415 | } | |
416 | ||
417 | ||
418 | static void increment_tail(struct oprofile_cpu_buffer * b) | |
419 | { | |
420 | unsigned long new_tail = b->tail_pos + 1; | |
421 | ||
422 | rmb(); | |
423 | ||
424 | if (new_tail < b->buffer_size) | |
425 | b->tail_pos = new_tail; | |
426 | else | |
427 | b->tail_pos = 0; | |
428 | } | |
429 | ||
430 | ||
431 | /* Move tasks along towards death. Any tasks on dead_tasks | |
432 | * will definitely have no remaining references in any | |
433 | * CPU buffers at this point, because we use two lists, | |
434 | * and to have reached the list, it must have gone through | |
435 | * one full sync already. | |
436 | */ | |
437 | static void process_task_mortuary(void) | |
438 | { | |
4369ef3c PM |
439 | unsigned long flags; |
440 | LIST_HEAD(local_dead_tasks); | |
1da177e4 | 441 | struct task_struct * task; |
4369ef3c | 442 | struct task_struct * ttask; |
1da177e4 | 443 | |
4369ef3c | 444 | spin_lock_irqsave(&task_mortuary, flags); |
1da177e4 | 445 | |
4369ef3c PM |
446 | list_splice_init(&dead_tasks, &local_dead_tasks); |
447 | list_splice_init(&dying_tasks, &dead_tasks); | |
1da177e4 | 448 | |
4369ef3c PM |
449 | spin_unlock_irqrestore(&task_mortuary, flags); |
450 | ||
451 | list_for_each_entry_safe(task, ttask, &local_dead_tasks, tasks) { | |
1da177e4 | 452 | list_del(&task->tasks); |
4369ef3c | 453 | free_task(task); |
1da177e4 | 454 | } |
1da177e4 LT |
455 | } |
456 | ||
457 | ||
458 | static void mark_done(int cpu) | |
459 | { | |
460 | int i; | |
461 | ||
462 | cpu_set(cpu, marked_cpus); | |
463 | ||
464 | for_each_online_cpu(i) { | |
465 | if (!cpu_isset(i, marked_cpus)) | |
466 | return; | |
467 | } | |
468 | ||
469 | /* All CPUs have been processed at least once, | |
470 | * we can process the mortuary once | |
471 | */ | |
472 | process_task_mortuary(); | |
473 | ||
474 | cpus_clear(marked_cpus); | |
475 | } | |
476 | ||
477 | ||
478 | /* FIXME: this is not sufficient if we implement syscall barrier backtrace | |
479 | * traversal, the code switch to sb_sample_start at first kernel enter/exit | |
480 | * switch so we need a fifth state and some special handling in sync_buffer() | |
481 | */ | |
482 | typedef enum { | |
483 | sb_bt_ignore = -2, | |
484 | sb_buffer_start, | |
485 | sb_bt_start, | |
486 | sb_sample_start, | |
487 | } sync_buffer_state; | |
488 | ||
489 | /* Sync one of the CPU's buffers into the global event buffer. | |
490 | * Here we need to go through each batch of samples punctuated | |
491 | * by context switch notes, taking the task's mmap_sem and doing | |
492 | * lookup in task->mm->mmap to convert EIP into dcookie/offset | |
493 | * value. | |
494 | */ | |
495 | void sync_buffer(int cpu) | |
496 | { | |
497 | struct oprofile_cpu_buffer * cpu_buf = &cpu_buffer[cpu]; | |
498 | struct mm_struct *mm = NULL; | |
499 | struct task_struct * new; | |
500 | unsigned long cookie = 0; | |
501 | int in_kernel = 1; | |
502 | unsigned int i; | |
503 | sync_buffer_state state = sb_buffer_start; | |
504 | unsigned long available; | |
505 | ||
59cc185a | 506 | mutex_lock(&buffer_mutex); |
1da177e4 LT |
507 | |
508 | add_cpu_switch(cpu); | |
509 | ||
510 | /* Remember, only we can modify tail_pos */ | |
511 | ||
512 | available = get_slots(cpu_buf); | |
513 | ||
514 | for (i = 0; i < available; ++i) { | |
515 | struct op_sample * s = &cpu_buf->buffer[cpu_buf->tail_pos]; | |
516 | ||
517 | if (is_code(s->eip)) { | |
518 | if (s->event <= CPU_IS_KERNEL) { | |
519 | /* kernel/userspace switch */ | |
520 | in_kernel = s->event; | |
521 | if (state == sb_buffer_start) | |
522 | state = sb_sample_start; | |
523 | add_kernel_ctx_switch(s->event); | |
524 | } else if (s->event == CPU_TRACE_BEGIN) { | |
525 | state = sb_bt_start; | |
526 | add_trace_begin(); | |
527 | } else { | |
528 | struct mm_struct * oldmm = mm; | |
529 | ||
530 | /* userspace context switch */ | |
531 | new = (struct task_struct *)s->event; | |
532 | ||
533 | release_mm(oldmm); | |
534 | mm = take_tasks_mm(new); | |
535 | if (mm != oldmm) | |
536 | cookie = get_exec_dcookie(mm); | |
537 | add_user_ctx_switch(new, cookie); | |
538 | } | |
539 | } else { | |
540 | if (state >= sb_bt_start && | |
541 | !add_sample(mm, s, in_kernel)) { | |
542 | if (state == sb_bt_start) { | |
543 | state = sb_bt_ignore; | |
544 | atomic_inc(&oprofile_stats.bt_lost_no_mapping); | |
545 | } | |
546 | } | |
547 | } | |
548 | ||
549 | increment_tail(cpu_buf); | |
550 | } | |
551 | release_mm(mm); | |
552 | ||
553 | mark_done(cpu); | |
554 | ||
59cc185a | 555 | mutex_unlock(&buffer_mutex); |
1da177e4 | 556 | } |