4 * @remark Copyright 2002 OProfile authors
5 * @remark Read the file COPYING
7 * @author John Levon <levon@movementarian.org>
8 * @author Barry Kasindorf
10 * This is the core of the buffer management. Each
11 * CPU buffer is processed and entered into the
12 * global event buffer. Such processing is necessary
13 * in several circumstances, mentioned below.
15 * The processing does the job of converting the
16 * transitory EIP value into a persistent dentry/offset
17 * value that the profiler can record at its leisure.
19 * See fs/dcookies.c for a description of the dentry/offset
24 #include <linux/workqueue.h>
25 #include <linux/notifier.h>
26 #include <linux/dcookies.h>
27 #include <linux/profile.h>
28 #include <linux/module.h>
30 #include <linux/oprofile.h>
31 #include <linux/sched.h>
33 #include "oprofile_stats.h"
34 #include "event_buffer.h"
35 #include "cpu_buffer.h"
36 #include "buffer_sync.h"
38 static LIST_HEAD(dying_tasks);
39 static LIST_HEAD(dead_tasks);
40 static cpumask_t marked_cpus = CPU_MASK_NONE;
41 static DEFINE_SPINLOCK(task_mortuary);
42 static void process_task_mortuary(void);
45 /* Take ownership of the task struct and place it on the
46 * list for processing. Only after two full buffer syncs
47 * does the task eventually get freed, because by then
48 * we are sure we will not reference it again.
49 * Can be invoked from softirq via RCU callback due to
50 * call_rcu() of the task struct, hence the _irqsave.
53 task_free_notify(struct notifier_block *self, unsigned long val, void *data)
56 struct task_struct *task = data;
57 spin_lock_irqsave(&task_mortuary, flags);
58 list_add(&task->tasks, &dying_tasks);
59 spin_unlock_irqrestore(&task_mortuary, flags);
64 /* The task is on its way out. A sync of the buffer means we can catch
65 * any remaining samples for this task.
68 task_exit_notify(struct notifier_block *self, unsigned long val, void *data)
70 /* To avoid latency problems, we only process the current CPU,
71 * hoping that most samples for the task are on this CPU
73 sync_buffer(raw_smp_processor_id());
78 /* The task is about to try a do_munmap(). We peek at what it's going to
79 * do, and if it's an executable region, process the samples first, so
80 * we don't lose any. This does not have to be exact, it's a QoI issue
84 munmap_notify(struct notifier_block *self, unsigned long val, void *data)
86 unsigned long addr = (unsigned long)data;
87 struct mm_struct *mm = current->mm;
88 struct vm_area_struct *mpnt;
90 down_read(&mm->mmap_sem);
92 mpnt = find_vma(mm, addr);
93 if (mpnt && mpnt->vm_file && (mpnt->vm_flags & VM_EXEC)) {
94 up_read(&mm->mmap_sem);
95 /* To avoid latency problems, we only process the current CPU,
96 * hoping that most samples for the task are on this CPU
98 sync_buffer(raw_smp_processor_id());
102 up_read(&mm->mmap_sem);
107 /* We need to be told about new modules so we don't attribute to a previously
108 * loaded module, or drop the samples on the floor.
111 module_load_notify(struct notifier_block *self, unsigned long val, void *data)
113 #ifdef CONFIG_MODULES
114 if (val != MODULE_STATE_COMING)
117 /* FIXME: should we process all CPU buffers ? */
118 mutex_lock(&buffer_mutex);
119 add_event_entry(ESCAPE_CODE);
120 add_event_entry(MODULE_LOADED_CODE);
121 mutex_unlock(&buffer_mutex);
127 static struct notifier_block task_free_nb = {
128 .notifier_call = task_free_notify,
131 static struct notifier_block task_exit_nb = {
132 .notifier_call = task_exit_notify,
135 static struct notifier_block munmap_nb = {
136 .notifier_call = munmap_notify,
139 static struct notifier_block module_load_nb = {
140 .notifier_call = module_load_notify,
144 static void end_sync(void)
147 /* make sure we don't leak task structs */
148 process_task_mortuary();
149 process_task_mortuary();
159 err = task_handoff_register(&task_free_nb);
162 err = profile_event_register(PROFILE_TASK_EXIT, &task_exit_nb);
165 err = profile_event_register(PROFILE_MUNMAP, &munmap_nb);
168 err = register_module_notifier(&module_load_nb);
175 profile_event_unregister(PROFILE_MUNMAP, &munmap_nb);
177 profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb);
179 task_handoff_unregister(&task_free_nb);
188 unregister_module_notifier(&module_load_nb);
189 profile_event_unregister(PROFILE_MUNMAP, &munmap_nb);
190 profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb);
191 task_handoff_unregister(&task_free_nb);
196 /* Optimisation. We can manage without taking the dcookie sem
197 * because we cannot reach this code without at least one
198 * dcookie user still being registered (namely, the reader
199 * of the event buffer). */
200 static inline unsigned long fast_get_dcookie(struct path *path)
202 unsigned long cookie;
204 if (path->dentry->d_cookie)
205 return (unsigned long)path->dentry;
206 get_dcookie(path, &cookie);
211 /* Look up the dcookie for the task's first VM_EXECUTABLE mapping,
212 * which corresponds loosely to "application name". This is
213 * not strictly necessary but allows oprofile to associate
214 * shared-library samples with particular applications
216 static unsigned long get_exec_dcookie(struct mm_struct *mm)
218 unsigned long cookie = NO_COOKIE;
219 struct vm_area_struct *vma;
224 for (vma = mm->mmap; vma; vma = vma->vm_next) {
227 if (!(vma->vm_flags & VM_EXECUTABLE))
229 cookie = fast_get_dcookie(&vma->vm_file->f_path);
238 /* Convert the EIP value of a sample into a persistent dentry/offset
239 * pair that can then be added to the global event buffer. We make
240 * sure to do this lookup before a mm->mmap modification happens so
241 * we don't lose track.
244 lookup_dcookie(struct mm_struct *mm, unsigned long addr, off_t *offset)
246 unsigned long cookie = NO_COOKIE;
247 struct vm_area_struct *vma;
249 for (vma = find_vma(mm, addr); vma; vma = vma->vm_next) {
251 if (addr < vma->vm_start || addr >= vma->vm_end)
255 cookie = fast_get_dcookie(&vma->vm_file->f_path);
256 *offset = (vma->vm_pgoff << PAGE_SHIFT) + addr -
259 /* must be an anonymous map */
267 cookie = INVALID_COOKIE;
272 static void increment_tail(struct oprofile_cpu_buffer *b)
274 unsigned long new_tail = b->tail_pos + 1;
276 rmb(); /* be sure fifo pointers are synchromized */
278 if (new_tail < b->buffer_size)
279 b->tail_pos = new_tail;
284 static unsigned long last_cookie = INVALID_COOKIE;
286 static void add_cpu_switch(int i)
288 add_event_entry(ESCAPE_CODE);
289 add_event_entry(CPU_SWITCH_CODE);
291 last_cookie = INVALID_COOKIE;
294 static void add_kernel_ctx_switch(unsigned int in_kernel)
296 add_event_entry(ESCAPE_CODE);
298 add_event_entry(KERNEL_ENTER_SWITCH_CODE);
300 add_event_entry(KERNEL_EXIT_SWITCH_CODE);
304 add_user_ctx_switch(struct task_struct const *task, unsigned long cookie)
306 add_event_entry(ESCAPE_CODE);
307 add_event_entry(CTX_SWITCH_CODE);
308 add_event_entry(task->pid);
309 add_event_entry(cookie);
310 /* Another code for daemon back-compat */
311 add_event_entry(ESCAPE_CODE);
312 add_event_entry(CTX_TGID_CODE);
313 add_event_entry(task->tgid);
317 static void add_cookie_switch(unsigned long cookie)
319 add_event_entry(ESCAPE_CODE);
320 add_event_entry(COOKIE_SWITCH_CODE);
321 add_event_entry(cookie);
325 static void add_trace_begin(void)
327 add_event_entry(ESCAPE_CODE);
328 add_event_entry(TRACE_BEGIN_CODE);
331 #ifdef CONFIG_OPROFILE_IBS
333 #define IBS_FETCH_CODE_SIZE 2
334 #define IBS_OP_CODE_SIZE 5
335 #define IBS_EIP(offset) \
336 (((struct op_sample *)&cpu_buf->buffer[(offset)])->eip)
337 #define IBS_EVENT(offset) \
338 (((struct op_sample *)&cpu_buf->buffer[(offset)])->event)
341 * Add IBS fetch and op entries to event buffer
343 static void add_ibs_begin(struct oprofile_cpu_buffer *cpu_buf, int code,
344 int in_kernel, struct mm_struct *mm)
348 unsigned long ibs_cookie = 0;
351 increment_tail(cpu_buf); /* move to RIP entry */
353 rip = IBS_EIP(cpu_buf->tail_pos);
356 rip += IBS_EVENT(cpu_buf->tail_pos) << 32;
360 ibs_cookie = lookup_dcookie(mm, rip, &offset);
362 if (ibs_cookie == NO_COOKIE)
364 if (ibs_cookie == INVALID_COOKIE) {
365 atomic_inc(&oprofile_stats.sample_lost_no_mapping);
368 if (ibs_cookie != last_cookie) {
369 add_cookie_switch(ibs_cookie);
370 last_cookie = ibs_cookie;
375 add_event_entry(ESCAPE_CODE);
376 add_event_entry(code);
377 add_event_entry(offset); /* Offset from Dcookie */
379 /* we send the Dcookie offset, but send the raw Linear Add also*/
380 add_event_entry(IBS_EIP(cpu_buf->tail_pos));
381 add_event_entry(IBS_EVENT(cpu_buf->tail_pos));
383 if (code == IBS_FETCH_CODE)
384 count = IBS_FETCH_CODE_SIZE; /*IBS FETCH is 2 int64s*/
386 count = IBS_OP_CODE_SIZE; /*IBS OP is 5 int64s*/
388 for (i = 0; i < count; i++) {
389 increment_tail(cpu_buf);
390 add_event_entry(IBS_EIP(cpu_buf->tail_pos));
391 add_event_entry(IBS_EVENT(cpu_buf->tail_pos));
397 static void add_sample_entry(unsigned long offset, unsigned long event)
399 add_event_entry(offset);
400 add_event_entry(event);
404 static int add_us_sample(struct mm_struct *mm, struct op_sample *s)
406 unsigned long cookie;
409 cookie = lookup_dcookie(mm, s->eip, &offset);
411 if (cookie == INVALID_COOKIE) {
412 atomic_inc(&oprofile_stats.sample_lost_no_mapping);
416 if (cookie != last_cookie) {
417 add_cookie_switch(cookie);
418 last_cookie = cookie;
421 add_sample_entry(offset, s->event);
427 /* Add a sample to the global event buffer. If possible the
428 * sample is converted into a persistent dentry/offset pair
429 * for later lookup from userspace.
432 add_sample(struct mm_struct *mm, struct op_sample *s, int in_kernel)
435 add_sample_entry(s->eip, s->event);
438 return add_us_sample(mm, s);
440 atomic_inc(&oprofile_stats.sample_lost_no_mm);
446 static void release_mm(struct mm_struct *mm)
450 up_read(&mm->mmap_sem);
455 static struct mm_struct *take_tasks_mm(struct task_struct *task)
457 struct mm_struct *mm = get_task_mm(task);
459 down_read(&mm->mmap_sem);
464 static inline int is_code(unsigned long val)
466 return val == ESCAPE_CODE;
470 /* "acquire" as many cpu buffer slots as we can */
471 static unsigned long get_slots(struct oprofile_cpu_buffer *b)
473 unsigned long head = b->head_pos;
474 unsigned long tail = b->tail_pos;
477 * Subtle. This resets the persistent last_task
478 * and in_kernel values used for switching notes.
479 * BUT, there is a small window between reading
480 * head_pos, and this call, that means samples
481 * can appear at the new head position, but not
482 * be prefixed with the notes for switching
483 * kernel mode or a task switch. This small hole
484 * can lead to mis-attribution or samples where
485 * we don't know if it's in the kernel or not,
486 * at the start of an event buffer.
493 return head + (b->buffer_size - tail);
497 /* Move tasks along towards death. Any tasks on dead_tasks
498 * will definitely have no remaining references in any
499 * CPU buffers at this point, because we use two lists,
500 * and to have reached the list, it must have gone through
501 * one full sync already.
503 static void process_task_mortuary(void)
506 LIST_HEAD(local_dead_tasks);
507 struct task_struct *task;
508 struct task_struct *ttask;
510 spin_lock_irqsave(&task_mortuary, flags);
512 list_splice_init(&dead_tasks, &local_dead_tasks);
513 list_splice_init(&dying_tasks, &dead_tasks);
515 spin_unlock_irqrestore(&task_mortuary, flags);
517 list_for_each_entry_safe(task, ttask, &local_dead_tasks, tasks) {
518 list_del(&task->tasks);
524 static void mark_done(int cpu)
528 cpu_set(cpu, marked_cpus);
530 for_each_online_cpu(i) {
531 if (!cpu_isset(i, marked_cpus))
535 /* All CPUs have been processed at least once,
536 * we can process the mortuary once
538 process_task_mortuary();
540 cpus_clear(marked_cpus);
544 /* FIXME: this is not sufficient if we implement syscall barrier backtrace
545 * traversal, the code switch to sb_sample_start at first kernel enter/exit
546 * switch so we need a fifth state and some special handling in sync_buffer()
555 /* Sync one of the CPU's buffers into the global event buffer.
556 * Here we need to go through each batch of samples punctuated
557 * by context switch notes, taking the task's mmap_sem and doing
558 * lookup in task->mm->mmap to convert EIP into dcookie/offset
561 void sync_buffer(int cpu)
563 struct oprofile_cpu_buffer *cpu_buf = &per_cpu(cpu_buffer, cpu);
564 struct mm_struct *mm = NULL;
565 struct task_struct *new;
566 unsigned long cookie = 0;
569 sync_buffer_state state = sb_buffer_start;
570 unsigned long available;
572 mutex_lock(&buffer_mutex);
576 /* Remember, only we can modify tail_pos */
578 available = get_slots(cpu_buf);
580 for (i = 0; i < available; ++i) {
581 struct op_sample *s = &cpu_buf->buffer[cpu_buf->tail_pos];
583 if (is_code(s->eip)) {
584 if (s->event <= CPU_IS_KERNEL) {
585 /* kernel/userspace switch */
586 in_kernel = s->event;
587 if (state == sb_buffer_start)
588 state = sb_sample_start;
589 add_kernel_ctx_switch(s->event);
590 } else if (s->event == CPU_TRACE_BEGIN) {
593 #ifdef CONFIG_OPROFILE_IBS
594 } else if (s->event == IBS_FETCH_BEGIN) {
596 add_ibs_begin(cpu_buf,
597 IBS_FETCH_CODE, in_kernel, mm);
598 } else if (s->event == IBS_OP_BEGIN) {
600 add_ibs_begin(cpu_buf,
601 IBS_OP_CODE, in_kernel, mm);
604 struct mm_struct *oldmm = mm;
606 /* userspace context switch */
607 new = (struct task_struct *)s->event;
610 mm = take_tasks_mm(new);
612 cookie = get_exec_dcookie(mm);
613 add_user_ctx_switch(new, cookie);
615 } else if (state >= sb_bt_start &&
616 !add_sample(mm, s, in_kernel)) {
617 if (state == sb_bt_start) {
618 state = sb_bt_ignore;
619 atomic_inc(&oprofile_stats.bt_lost_no_mapping);
623 increment_tail(cpu_buf);
629 mutex_unlock(&buffer_mutex);