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>
33 #include <linux/gfp.h>
35 #include "oprofile_stats.h"
36 #include "event_buffer.h"
37 #include "cpu_buffer.h"
38 #include "buffer_sync.h"
40 static LIST_HEAD(dying_tasks);
41 static LIST_HEAD(dead_tasks);
42 static cpumask_var_t marked_cpus;
43 static DEFINE_SPINLOCK(task_mortuary);
44 static void process_task_mortuary(void);
46 /* Take ownership of the task struct and place it on the
47 * list for processing. Only after two full buffer syncs
48 * does the task eventually get freed, because by then
49 * we are sure we will not reference it again.
50 * Can be invoked from softirq via RCU callback due to
51 * call_rcu() of the task struct, hence the _irqsave.
54 task_free_notify(struct notifier_block *self, unsigned long val, void *data)
57 struct task_struct *task = data;
58 spin_lock_irqsave(&task_mortuary, flags);
59 list_add(&task->tasks, &dying_tasks);
60 spin_unlock_irqrestore(&task_mortuary, flags);
65 /* The task is on its way out. A sync of the buffer means we can catch
66 * any remaining samples for this task.
69 task_exit_notify(struct notifier_block *self, unsigned long val, void *data)
71 /* To avoid latency problems, we only process the current CPU,
72 * hoping that most samples for the task are on this CPU
74 sync_buffer(raw_smp_processor_id());
79 /* The task is about to try a do_munmap(). We peek at what it's going to
80 * do, and if it's an executable region, process the samples first, so
81 * we don't lose any. This does not have to be exact, it's a QoI issue
85 munmap_notify(struct notifier_block *self, unsigned long val, void *data)
87 unsigned long addr = (unsigned long)data;
88 struct mm_struct *mm = current->mm;
89 struct vm_area_struct *mpnt;
91 down_read(&mm->mmap_sem);
93 mpnt = find_vma(mm, addr);
94 if (mpnt && mpnt->vm_file && (mpnt->vm_flags & VM_EXEC)) {
95 up_read(&mm->mmap_sem);
96 /* To avoid latency problems, we only process the current CPU,
97 * hoping that most samples for the task are on this CPU
99 sync_buffer(raw_smp_processor_id());
103 up_read(&mm->mmap_sem);
108 /* We need to be told about new modules so we don't attribute to a previously
109 * loaded module, or drop the samples on the floor.
112 module_load_notify(struct notifier_block *self, unsigned long val, void *data)
114 #ifdef CONFIG_MODULES
115 if (val != MODULE_STATE_COMING)
118 /* FIXME: should we process all CPU buffers ? */
119 mutex_lock(&buffer_mutex);
120 add_event_entry(ESCAPE_CODE);
121 add_event_entry(MODULE_LOADED_CODE);
122 mutex_unlock(&buffer_mutex);
128 static struct notifier_block task_free_nb = {
129 .notifier_call = task_free_notify,
132 static struct notifier_block task_exit_nb = {
133 .notifier_call = task_exit_notify,
136 static struct notifier_block munmap_nb = {
137 .notifier_call = munmap_notify,
140 static struct notifier_block module_load_nb = {
141 .notifier_call = module_load_notify,
144 static void free_all_tasks(void)
146 /* make sure we don't leak task structs */
147 process_task_mortuary();
148 process_task_mortuary();
155 if (!zalloc_cpumask_var(&marked_cpus, GFP_KERNEL))
158 err = task_handoff_register(&task_free_nb);
161 err = profile_event_register(PROFILE_TASK_EXIT, &task_exit_nb);
164 err = profile_event_register(PROFILE_MUNMAP, &munmap_nb);
167 err = register_module_notifier(&module_load_nb);
176 profile_event_unregister(PROFILE_MUNMAP, &munmap_nb);
178 profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb);
180 task_handoff_unregister(&task_free_nb);
183 free_cpumask_var(marked_cpus);
191 unregister_module_notifier(&module_load_nb);
192 profile_event_unregister(PROFILE_MUNMAP, &munmap_nb);
193 profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb);
194 task_handoff_unregister(&task_free_nb);
195 barrier(); /* do all of the above first */
200 free_cpumask_var(marked_cpus);
204 /* Optimisation. We can manage without taking the dcookie sem
205 * because we cannot reach this code without at least one
206 * dcookie user still being registered (namely, the reader
207 * of the event buffer). */
208 static inline unsigned long fast_get_dcookie(struct path *path)
210 unsigned long cookie;
212 if (path->dentry->d_flags & DCACHE_COOKIE)
213 return (unsigned long)path->dentry;
214 get_dcookie(path, &cookie);
219 /* Look up the dcookie for the task's mm->exe_file,
220 * which corresponds loosely to "application name". This is
221 * not strictly necessary but allows oprofile to associate
222 * shared-library samples with particular applications
224 static unsigned long get_exec_dcookie(struct mm_struct *mm)
226 unsigned long cookie = NO_COOKIE;
228 if (mm && mm->exe_file)
229 cookie = fast_get_dcookie(&mm->exe_file->f_path);
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.
241 lookup_dcookie(struct mm_struct *mm, unsigned long addr, off_t *offset)
243 unsigned long cookie = NO_COOKIE;
244 struct vm_area_struct *vma;
246 for (vma = find_vma(mm, addr); vma; vma = vma->vm_next) {
248 if (addr < vma->vm_start || addr >= vma->vm_end)
252 cookie = fast_get_dcookie(&vma->vm_file->f_path);
253 *offset = (vma->vm_pgoff << PAGE_SHIFT) + addr -
256 /* must be an anonymous map */
264 cookie = INVALID_COOKIE;
269 static unsigned long last_cookie = INVALID_COOKIE;
271 static void add_cpu_switch(int i)
273 add_event_entry(ESCAPE_CODE);
274 add_event_entry(CPU_SWITCH_CODE);
276 last_cookie = INVALID_COOKIE;
279 static void add_kernel_ctx_switch(unsigned int in_kernel)
281 add_event_entry(ESCAPE_CODE);
283 add_event_entry(KERNEL_ENTER_SWITCH_CODE);
285 add_event_entry(KERNEL_EXIT_SWITCH_CODE);
289 add_user_ctx_switch(struct task_struct const *task, unsigned long cookie)
291 add_event_entry(ESCAPE_CODE);
292 add_event_entry(CTX_SWITCH_CODE);
293 add_event_entry(task->pid);
294 add_event_entry(cookie);
295 /* Another code for daemon back-compat */
296 add_event_entry(ESCAPE_CODE);
297 add_event_entry(CTX_TGID_CODE);
298 add_event_entry(task->tgid);
302 static void add_cookie_switch(unsigned long cookie)
304 add_event_entry(ESCAPE_CODE);
305 add_event_entry(COOKIE_SWITCH_CODE);
306 add_event_entry(cookie);
310 static void add_trace_begin(void)
312 add_event_entry(ESCAPE_CODE);
313 add_event_entry(TRACE_BEGIN_CODE);
316 static void add_data(struct op_entry *entry, struct mm_struct *mm)
318 unsigned long code, pc, val;
319 unsigned long cookie;
322 if (!op_cpu_buffer_get_data(entry, &code))
324 if (!op_cpu_buffer_get_data(entry, &pc))
326 if (!op_cpu_buffer_get_size(entry))
330 cookie = lookup_dcookie(mm, pc, &offset);
332 if (cookie == NO_COOKIE)
334 if (cookie == INVALID_COOKIE) {
335 atomic_inc(&oprofile_stats.sample_lost_no_mapping);
338 if (cookie != last_cookie) {
339 add_cookie_switch(cookie);
340 last_cookie = cookie;
345 add_event_entry(ESCAPE_CODE);
346 add_event_entry(code);
347 add_event_entry(offset); /* Offset from Dcookie */
349 while (op_cpu_buffer_get_data(entry, &val))
350 add_event_entry(val);
353 static inline void add_sample_entry(unsigned long offset, unsigned long event)
355 add_event_entry(offset);
356 add_event_entry(event);
361 * Add a sample to the global event buffer. If possible the
362 * sample is converted into a persistent dentry/offset pair
363 * for later lookup from userspace. Return 0 on failure.
366 add_sample(struct mm_struct *mm, struct op_sample *s, int in_kernel)
368 unsigned long cookie;
372 add_sample_entry(s->eip, s->event);
376 /* add userspace sample */
379 atomic_inc(&oprofile_stats.sample_lost_no_mm);
383 cookie = lookup_dcookie(mm, s->eip, &offset);
385 if (cookie == INVALID_COOKIE) {
386 atomic_inc(&oprofile_stats.sample_lost_no_mapping);
390 if (cookie != last_cookie) {
391 add_cookie_switch(cookie);
392 last_cookie = cookie;
395 add_sample_entry(offset, s->event);
401 static void release_mm(struct mm_struct *mm)
405 up_read(&mm->mmap_sem);
410 static struct mm_struct *take_tasks_mm(struct task_struct *task)
412 struct mm_struct *mm = get_task_mm(task);
414 down_read(&mm->mmap_sem);
419 static inline int is_code(unsigned long val)
421 return val == ESCAPE_CODE;
425 /* Move tasks along towards death. Any tasks on dead_tasks
426 * will definitely have no remaining references in any
427 * CPU buffers at this point, because we use two lists,
428 * and to have reached the list, it must have gone through
429 * one full sync already.
431 static void process_task_mortuary(void)
434 LIST_HEAD(local_dead_tasks);
435 struct task_struct *task;
436 struct task_struct *ttask;
438 spin_lock_irqsave(&task_mortuary, flags);
440 list_splice_init(&dead_tasks, &local_dead_tasks);
441 list_splice_init(&dying_tasks, &dead_tasks);
443 spin_unlock_irqrestore(&task_mortuary, flags);
445 list_for_each_entry_safe(task, ttask, &local_dead_tasks, tasks) {
446 list_del(&task->tasks);
452 static void mark_done(int cpu)
456 cpumask_set_cpu(cpu, marked_cpus);
458 for_each_online_cpu(i) {
459 if (!cpumask_test_cpu(i, marked_cpus))
463 /* All CPUs have been processed at least once,
464 * we can process the mortuary once
466 process_task_mortuary();
468 cpumask_clear(marked_cpus);
472 /* FIXME: this is not sufficient if we implement syscall barrier backtrace
473 * traversal, the code switch to sb_sample_start at first kernel enter/exit
474 * switch so we need a fifth state and some special handling in sync_buffer()
483 /* Sync one of the CPU's buffers into the global event buffer.
484 * Here we need to go through each batch of samples punctuated
485 * by context switch notes, taking the task's mmap_sem and doing
486 * lookup in task->mm->mmap to convert EIP into dcookie/offset
489 void sync_buffer(int cpu)
491 struct mm_struct *mm = NULL;
492 struct mm_struct *oldmm;
494 struct task_struct *new;
495 unsigned long cookie = 0;
497 sync_buffer_state state = sb_buffer_start;
499 unsigned long available;
501 struct op_entry entry;
502 struct op_sample *sample;
504 mutex_lock(&buffer_mutex);
508 op_cpu_buffer_reset(cpu);
509 available = op_cpu_buffer_entries(cpu);
511 for (i = 0; i < available; ++i) {
512 sample = op_cpu_buffer_read_entry(&entry, cpu);
516 if (is_code(sample->eip)) {
517 flags = sample->event;
518 if (flags & TRACE_BEGIN) {
522 if (flags & KERNEL_CTX_SWITCH) {
523 /* kernel/userspace switch */
524 in_kernel = flags & IS_KERNEL;
525 if (state == sb_buffer_start)
526 state = sb_sample_start;
527 add_kernel_ctx_switch(flags & IS_KERNEL);
529 if (flags & USER_CTX_SWITCH
530 && op_cpu_buffer_get_data(&entry, &val)) {
531 /* userspace context switch */
532 new = (struct task_struct *)val;
535 mm = take_tasks_mm(new);
537 cookie = get_exec_dcookie(mm);
538 add_user_ctx_switch(new, cookie);
540 if (op_cpu_buffer_get_size(&entry))
541 add_data(&entry, mm);
545 if (state < sb_bt_start)
549 if (add_sample(mm, sample, in_kernel))
552 /* ignore backtraces if failed to add a sample */
553 if (state == sb_bt_start) {
554 state = sb_bt_ignore;
555 atomic_inc(&oprofile_stats.bt_lost_no_mapping);
562 mutex_unlock(&buffer_mutex);
565 /* The function can be used to add a buffer worth of data directly to
566 * the kernel buffer. The buffer is assumed to be a circular buffer.
567 * Take the entries from index start and end at index end, wrapping
570 void oprofile_put_buff(unsigned long *buf, unsigned int start,
571 unsigned int stop, unsigned int max)
577 mutex_lock(&buffer_mutex);
579 add_event_entry(buf[i++]);
585 mutex_unlock(&buffer_mutex);