4 * @remark Copyright 2002 OProfile authors
5 * @remark Read the file COPYING
7 * @author John Levon <levon@movementarian.org>
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.
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.
18 * See fs/dcookies.c for a description of the dentry/offset
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>
29 #include <linux/oprofile.h>
30 #include <linux/sched.h>
32 #include "oprofile_stats.h"
33 #include "event_buffer.h"
34 #include "cpu_buffer.h"
35 #include "buffer_sync.h"
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);
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.
48 * Can be invoked from softirq via RCU callback due to
49 * call_rcu() of the task struct, hence the _irqsave.
51 static int task_free_notify(struct notifier_block * self, unsigned long val, void * data)
54 struct task_struct * task = data;
55 spin_lock_irqsave(&task_mortuary, flags);
56 list_add(&task->tasks, &dying_tasks);
57 spin_unlock_irqrestore(&task_mortuary, flags);
62 /* The task is on its way out. A sync of the buffer means we can catch
63 * any remaining samples for this task.
65 static int task_exit_notify(struct notifier_block * self, unsigned long val, void * data)
67 /* To avoid latency problems, we only process the current CPU,
68 * hoping that most samples for the task are on this CPU
70 sync_buffer(raw_smp_processor_id());
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
80 static int munmap_notify(struct notifier_block * self, unsigned long val, void * data)
82 unsigned long addr = (unsigned long)data;
83 struct mm_struct * mm = current->mm;
84 struct vm_area_struct * mpnt;
86 down_read(&mm->mmap_sem);
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
94 sync_buffer(raw_smp_processor_id());
98 up_read(&mm->mmap_sem);
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.
106 static int module_load_notify(struct notifier_block * self, unsigned long val, void * data)
108 #ifdef CONFIG_MODULES
109 if (val != MODULE_STATE_COMING)
112 /* FIXME: should we process all CPU buffers ? */
113 mutex_lock(&buffer_mutex);
114 add_event_entry(ESCAPE_CODE);
115 add_event_entry(MODULE_LOADED_CODE);
116 mutex_unlock(&buffer_mutex);
122 static struct notifier_block task_free_nb = {
123 .notifier_call = task_free_notify,
126 static struct notifier_block task_exit_nb = {
127 .notifier_call = task_exit_notify,
130 static struct notifier_block munmap_nb = {
131 .notifier_call = munmap_notify,
134 static struct notifier_block module_load_nb = {
135 .notifier_call = module_load_notify,
139 static void end_sync(void)
142 /* make sure we don't leak task structs */
143 process_task_mortuary();
144 process_task_mortuary();
154 err = task_handoff_register(&task_free_nb);
157 err = profile_event_register(PROFILE_TASK_EXIT, &task_exit_nb);
160 err = profile_event_register(PROFILE_MUNMAP, &munmap_nb);
163 err = register_module_notifier(&module_load_nb);
170 profile_event_unregister(PROFILE_MUNMAP, &munmap_nb);
172 profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb);
174 task_handoff_unregister(&task_free_nb);
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);
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 path *path)
197 unsigned long cookie;
199 if (path->dentry->d_cookie)
200 return (unsigned long)path->dentry;
201 get_dcookie(path, &cookie);
206 /* 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
211 static unsigned long get_exec_dcookie(struct mm_struct * mm)
213 unsigned long cookie = NO_COOKIE;
214 struct vm_area_struct * vma;
219 for (vma = mm->mmap; vma; vma = vma->vm_next) {
222 if (!(vma->vm_flags & VM_EXECUTABLE))
224 cookie = fast_get_dcookie(&vma->vm_file->f_path);
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.
238 static unsigned long lookup_dcookie(struct mm_struct * mm, unsigned long addr, off_t * offset)
240 unsigned long cookie = NO_COOKIE;
241 struct vm_area_struct * vma;
243 for (vma = find_vma(mm, addr); vma; vma = vma->vm_next) {
245 if (addr < vma->vm_start || addr >= vma->vm_end)
249 cookie = fast_get_dcookie(&vma->vm_file->f_path);
250 *offset = (vma->vm_pgoff << PAGE_SHIFT) + addr -
253 /* must be an anonymous map */
261 cookie = INVALID_COOKIE;
267 static unsigned long last_cookie = INVALID_COOKIE;
269 static void add_cpu_switch(int i)
271 add_event_entry(ESCAPE_CODE);
272 add_event_entry(CPU_SWITCH_CODE);
274 last_cookie = INVALID_COOKIE;
277 static void add_kernel_ctx_switch(unsigned int in_kernel)
279 add_event_entry(ESCAPE_CODE);
281 add_event_entry(KERNEL_ENTER_SWITCH_CODE);
283 add_event_entry(KERNEL_EXIT_SWITCH_CODE);
287 add_user_ctx_switch(struct task_struct const * task, unsigned long cookie)
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);
300 static void add_cookie_switch(unsigned long cookie)
302 add_event_entry(ESCAPE_CODE);
303 add_event_entry(COOKIE_SWITCH_CODE);
304 add_event_entry(cookie);
308 static void add_trace_begin(void)
310 add_event_entry(ESCAPE_CODE);
311 add_event_entry(TRACE_BEGIN_CODE);
315 static void add_sample_entry(unsigned long offset, unsigned long event)
317 add_event_entry(offset);
318 add_event_entry(event);
322 static int add_us_sample(struct mm_struct * mm, struct op_sample * s)
324 unsigned long cookie;
327 cookie = lookup_dcookie(mm, s->eip, &offset);
329 if (cookie == INVALID_COOKIE) {
330 atomic_inc(&oprofile_stats.sample_lost_no_mapping);
334 if (cookie != last_cookie) {
335 add_cookie_switch(cookie);
336 last_cookie = cookie;
339 add_sample_entry(offset, s->event);
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.
350 add_sample(struct mm_struct * mm, struct op_sample * s, int in_kernel)
353 add_sample_entry(s->eip, s->event);
356 return add_us_sample(mm, s);
358 atomic_inc(&oprofile_stats.sample_lost_no_mm);
364 static void release_mm(struct mm_struct * mm)
368 up_read(&mm->mmap_sem);
373 static struct mm_struct * take_tasks_mm(struct task_struct * task)
375 struct mm_struct * mm = get_task_mm(task);
377 down_read(&mm->mmap_sem);
382 static inline int is_code(unsigned long val)
384 return val == ESCAPE_CODE;
388 /* "acquire" as many cpu buffer slots as we can */
389 static unsigned long get_slots(struct oprofile_cpu_buffer * b)
391 unsigned long head = b->head_pos;
392 unsigned long tail = b->tail_pos;
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.
411 return head + (b->buffer_size - tail);
415 static void increment_tail(struct oprofile_cpu_buffer * b)
417 unsigned long new_tail = b->tail_pos + 1;
421 if (new_tail < b->buffer_size)
422 b->tail_pos = new_tail;
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.
434 static void process_task_mortuary(void)
437 LIST_HEAD(local_dead_tasks);
438 struct task_struct * task;
439 struct task_struct * ttask;
441 spin_lock_irqsave(&task_mortuary, flags);
443 list_splice_init(&dead_tasks, &local_dead_tasks);
444 list_splice_init(&dying_tasks, &dead_tasks);
446 spin_unlock_irqrestore(&task_mortuary, flags);
448 list_for_each_entry_safe(task, ttask, &local_dead_tasks, tasks) {
449 list_del(&task->tasks);
455 static void mark_done(int cpu)
459 cpu_set(cpu, marked_cpus);
461 for_each_online_cpu(i) {
462 if (!cpu_isset(i, marked_cpus))
466 /* All CPUs have been processed at least once,
467 * we can process the mortuary once
469 process_task_mortuary();
471 cpus_clear(marked_cpus);
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()
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
492 void sync_buffer(int cpu)
494 struct oprofile_cpu_buffer *cpu_buf = &per_cpu(cpu_buffer, cpu);
495 struct mm_struct *mm = NULL;
496 struct task_struct * new;
497 unsigned long cookie = 0;
500 sync_buffer_state state = sb_buffer_start;
501 unsigned long available;
503 mutex_lock(&buffer_mutex);
507 /* Remember, only we can modify tail_pos */
509 available = get_slots(cpu_buf);
511 for (i = 0; i < available; ++i) {
512 struct op_sample * s = &cpu_buf->buffer[cpu_buf->tail_pos];
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) {
525 struct mm_struct * oldmm = mm;
527 /* userspace context switch */
528 new = (struct task_struct *)s->event;
531 mm = take_tasks_mm(new);
533 cookie = get_exec_dcookie(mm);
534 add_user_ctx_switch(new, cookie);
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);
546 increment_tail(cpu_buf);
552 mutex_unlock(&buffer_mutex);