4 * @remark Copyright 2002-2009 OProfile authors
5 * @remark Read the file COPYING
7 * @author John Levon <levon@movementarian.org>
8 * @author Barry Kasindorf
9 * @author Robert Richter <robert.richter@amd.com>
11 * This is the core of the buffer management. Each
12 * CPU buffer is processed and entered into the
13 * global event buffer. Such processing is necessary
14 * in several circumstances, mentioned below.
16 * The processing does the job of converting the
17 * transitory EIP value into a persistent dentry/offset
18 * value that the profiler can record at its leisure.
20 * See fs/dcookies.c for a description of the dentry/offset
25 #include <linux/workqueue.h>
26 #include <linux/notifier.h>
27 #include <linux/dcookies.h>
28 #include <linux/profile.h>
29 #include <linux/module.h>
31 #include <linux/oprofile.h>
32 #include <linux/sched.h>
34 #include "oprofile_stats.h"
35 #include "event_buffer.h"
36 #include "cpu_buffer.h"
37 #include "buffer_sync.h"
39 static LIST_HEAD(dying_tasks);
40 static LIST_HEAD(dead_tasks);
41 static cpumask_t marked_cpus = CPU_MASK_NONE;
42 static DEFINE_SPINLOCK(task_mortuary);
43 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 unsigned long last_cookie = INVALID_COOKIE;
274 static void add_cpu_switch(int i)
276 add_event_entry(ESCAPE_CODE);
277 add_event_entry(CPU_SWITCH_CODE);
279 last_cookie = INVALID_COOKIE;
282 static void add_kernel_ctx_switch(unsigned int in_kernel)
284 add_event_entry(ESCAPE_CODE);
286 add_event_entry(KERNEL_ENTER_SWITCH_CODE);
288 add_event_entry(KERNEL_EXIT_SWITCH_CODE);
292 add_user_ctx_switch(struct task_struct const *task, unsigned long cookie)
294 add_event_entry(ESCAPE_CODE);
295 add_event_entry(CTX_SWITCH_CODE);
296 add_event_entry(task->pid);
297 add_event_entry(cookie);
298 /* Another code for daemon back-compat */
299 add_event_entry(ESCAPE_CODE);
300 add_event_entry(CTX_TGID_CODE);
301 add_event_entry(task->tgid);
305 static void add_cookie_switch(unsigned long cookie)
307 add_event_entry(ESCAPE_CODE);
308 add_event_entry(COOKIE_SWITCH_CODE);
309 add_event_entry(cookie);
313 static void add_trace_begin(void)
315 add_event_entry(ESCAPE_CODE);
316 add_event_entry(TRACE_BEGIN_CODE);
319 #ifdef CONFIG_OPROFILE_IBS
321 static void add_data(struct op_entry *entry, struct mm_struct *mm)
323 unsigned long code, pc, val;
324 unsigned long cookie;
327 if (!op_cpu_buffer_get_data(entry, &code))
329 if (!op_cpu_buffer_get_data(entry, &pc))
331 if (!op_cpu_buffer_get_size(entry))
335 cookie = lookup_dcookie(mm, pc, &offset);
337 if (cookie == NO_COOKIE)
339 if (cookie == INVALID_COOKIE) {
340 atomic_inc(&oprofile_stats.sample_lost_no_mapping);
343 if (cookie != last_cookie) {
344 add_cookie_switch(cookie);
345 last_cookie = cookie;
350 add_event_entry(ESCAPE_CODE);
351 add_event_entry(code);
352 add_event_entry(offset); /* Offset from Dcookie */
354 while (op_cpu_buffer_get_data(entry, &val))
355 add_event_entry(val);
360 static inline void add_sample_entry(unsigned long offset, unsigned long event)
362 add_event_entry(offset);
363 add_event_entry(event);
368 * Add a sample to the global event buffer. If possible the
369 * sample is converted into a persistent dentry/offset pair
370 * for later lookup from userspace. Return 0 on failure.
373 add_sample(struct mm_struct *mm, struct op_sample *s, int in_kernel)
375 unsigned long cookie;
379 add_sample_entry(s->eip, s->event);
383 /* add userspace sample */
386 atomic_inc(&oprofile_stats.sample_lost_no_mm);
390 cookie = lookup_dcookie(mm, s->eip, &offset);
392 if (cookie == INVALID_COOKIE) {
393 atomic_inc(&oprofile_stats.sample_lost_no_mapping);
397 if (cookie != last_cookie) {
398 add_cookie_switch(cookie);
399 last_cookie = cookie;
402 add_sample_entry(offset, s->event);
408 static void release_mm(struct mm_struct *mm)
412 up_read(&mm->mmap_sem);
417 static struct mm_struct *take_tasks_mm(struct task_struct *task)
419 struct mm_struct *mm = get_task_mm(task);
421 down_read(&mm->mmap_sem);
426 static inline int is_code(unsigned long val)
428 return val == ESCAPE_CODE;
432 /* Move tasks along towards death. Any tasks on dead_tasks
433 * will definitely have no remaining references in any
434 * CPU buffers at this point, because we use two lists,
435 * and to have reached the list, it must have gone through
436 * one full sync already.
438 static void process_task_mortuary(void)
441 LIST_HEAD(local_dead_tasks);
442 struct task_struct *task;
443 struct task_struct *ttask;
445 spin_lock_irqsave(&task_mortuary, flags);
447 list_splice_init(&dead_tasks, &local_dead_tasks);
448 list_splice_init(&dying_tasks, &dead_tasks);
450 spin_unlock_irqrestore(&task_mortuary, flags);
452 list_for_each_entry_safe(task, ttask, &local_dead_tasks, tasks) {
453 list_del(&task->tasks);
459 static void mark_done(int cpu)
463 cpu_set(cpu, marked_cpus);
465 for_each_online_cpu(i) {
466 if (!cpu_isset(i, marked_cpus))
470 /* All CPUs have been processed at least once,
471 * we can process the mortuary once
473 process_task_mortuary();
475 cpus_clear(marked_cpus);
479 /* FIXME: this is not sufficient if we implement syscall barrier backtrace
480 * traversal, the code switch to sb_sample_start at first kernel enter/exit
481 * switch so we need a fifth state and some special handling in sync_buffer()
490 /* Sync one of the CPU's buffers into the global event buffer.
491 * Here we need to go through each batch of samples punctuated
492 * by context switch notes, taking the task's mmap_sem and doing
493 * lookup in task->mm->mmap to convert EIP into dcookie/offset
496 void sync_buffer(int cpu)
498 struct mm_struct *mm = NULL;
499 struct mm_struct *oldmm;
501 struct task_struct *new;
502 unsigned long cookie = 0;
504 sync_buffer_state state = sb_buffer_start;
506 unsigned long available;
508 struct op_entry entry;
509 struct op_sample *sample;
511 mutex_lock(&buffer_mutex);
515 op_cpu_buffer_reset(cpu);
516 available = op_cpu_buffer_entries(cpu);
518 for (i = 0; i < available; ++i) {
519 sample = op_cpu_buffer_read_entry(&entry, cpu);
523 if (is_code(sample->eip)) {
524 flags = sample->event;
525 if (flags & TRACE_BEGIN) {
529 if (flags & KERNEL_CTX_SWITCH) {
530 /* kernel/userspace switch */
531 in_kernel = flags & IS_KERNEL;
532 if (state == sb_buffer_start)
533 state = sb_sample_start;
534 add_kernel_ctx_switch(flags & IS_KERNEL);
536 if (flags & USER_CTX_SWITCH
537 && op_cpu_buffer_get_data(&entry, &val)) {
538 /* userspace context switch */
539 new = (struct task_struct *)val;
542 mm = take_tasks_mm(new);
544 cookie = get_exec_dcookie(mm);
545 add_user_ctx_switch(new, cookie);
547 #ifdef CONFIG_OPROFILE_IBS
548 if (op_cpu_buffer_get_size(&entry))
549 add_data(&entry, mm);
554 if (state < sb_bt_start)
558 if (add_sample(mm, sample, in_kernel))
561 /* ignore backtraces if failed to add a sample */
562 if (state == sb_bt_start) {
563 state = sb_bt_ignore;
564 atomic_inc(&oprofile_stats.bt_lost_no_mapping);
571 mutex_unlock(&buffer_mutex);
574 /* The function can be used to add a buffer worth of data directly to
575 * the kernel buffer. The buffer is assumed to be a circular buffer.
576 * Take the entries from index start and end at index end, wrapping
579 void oprofile_put_buff(unsigned long *buf, unsigned int start,
580 unsigned int stop, unsigned int max)
586 mutex_lock(&buffer_mutex);
588 add_event_entry(buf[i++]);
594 mutex_unlock(&buffer_mutex);