blob: a3984f4ef192a1807ed4714d70babb6702a175d8 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/**
2 * @file buffer_sync.c
3 *
Robert Richterae735e92008-12-25 17:26:07 +01004 * @remark Copyright 2002-2009 OProfile authors
Linus Torvalds1da177e2005-04-16 15:20:36 -07005 * @remark Read the file COPYING
6 *
7 * @author John Levon <levon@movementarian.org>
Barry Kasindorf345c2572008-07-22 21:08:54 +02008 * @author Barry Kasindorf
Robert Richterae735e92008-12-25 17:26:07 +01009 * @author Robert Richter <robert.richter@amd.com>
Linus Torvalds1da177e2005-04-16 15:20:36 -070010 *
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.
15 *
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.
19 *
20 * See fs/dcookies.c for a description of the dentry/offset
21 * objects.
22 */
23
24#include <linux/mm.h>
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>
30#include <linux/fs.h>
Bob Nelson14748552007-07-20 21:39:53 +020031#include <linux/oprofile.h>
Alexey Dobriyane8edc6e2007-05-21 01:22:52 +040032#include <linux/sched.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090033#include <linux/gfp.h>
Bob Nelson14748552007-07-20 21:39:53 +020034
Linus Torvalds1da177e2005-04-16 15:20:36 -070035#include "oprofile_stats.h"
36#include "event_buffer.h"
37#include "cpu_buffer.h"
38#include "buffer_sync.h"
Robert Richter73185e02008-07-22 21:08:51 +020039
Linus Torvalds1da177e2005-04-16 15:20:36 -070040static LIST_HEAD(dying_tasks);
41static LIST_HEAD(dead_tasks);
Rusty Russellf7df8ed2009-01-10 21:58:09 -080042static cpumask_var_t marked_cpus;
Linus Torvalds1da177e2005-04-16 15:20:36 -070043static DEFINE_SPINLOCK(task_mortuary);
44static void process_task_mortuary(void);
45
Linus Torvalds1da177e2005-04-16 15:20:36 -070046/* 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.
Paul E. McKenney4369ef32006-01-08 01:01:35 -080050 * Can be invoked from softirq via RCU callback due to
51 * call_rcu() of the task struct, hence the _irqsave.
Linus Torvalds1da177e2005-04-16 15:20:36 -070052 */
Robert Richter73185e02008-07-22 21:08:51 +020053static int
54task_free_notify(struct notifier_block *self, unsigned long val, void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -070055{
Paul E. McKenney4369ef32006-01-08 01:01:35 -080056 unsigned long flags;
Robert Richter73185e02008-07-22 21:08:51 +020057 struct task_struct *task = data;
Paul E. McKenney4369ef32006-01-08 01:01:35 -080058 spin_lock_irqsave(&task_mortuary, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -070059 list_add(&task->tasks, &dying_tasks);
Paul E. McKenney4369ef32006-01-08 01:01:35 -080060 spin_unlock_irqrestore(&task_mortuary, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -070061 return NOTIFY_OK;
62}
63
64
65/* The task is on its way out. A sync of the buffer means we can catch
66 * any remaining samples for this task.
67 */
Robert Richter73185e02008-07-22 21:08:51 +020068static int
69task_exit_notify(struct notifier_block *self, unsigned long val, void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -070070{
71 /* To avoid latency problems, we only process the current CPU,
72 * hoping that most samples for the task are on this CPU
73 */
Ingo Molnar39c715b2005-06-21 17:14:34 -070074 sync_buffer(raw_smp_processor_id());
Robert Richter73185e02008-07-22 21:08:51 +020075 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -070076}
77
78
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
82 * only.
83 */
Robert Richter73185e02008-07-22 21:08:51 +020084static int
85munmap_notify(struct notifier_block *self, unsigned long val, void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -070086{
87 unsigned long addr = (unsigned long)data;
Robert Richter73185e02008-07-22 21:08:51 +020088 struct mm_struct *mm = current->mm;
89 struct vm_area_struct *mpnt;
Linus Torvalds1da177e2005-04-16 15:20:36 -070090
91 down_read(&mm->mmap_sem);
92
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
98 */
Ingo Molnar39c715b2005-06-21 17:14:34 -070099 sync_buffer(raw_smp_processor_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -0700100 return 0;
101 }
102
103 up_read(&mm->mmap_sem);
104 return 0;
105}
106
Robert Richter73185e02008-07-22 21:08:51 +0200107
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108/* 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.
110 */
Robert Richter73185e02008-07-22 21:08:51 +0200111static int
112module_load_notify(struct notifier_block *self, unsigned long val, void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113{
114#ifdef CONFIG_MODULES
115 if (val != MODULE_STATE_COMING)
116 return 0;
117
118 /* FIXME: should we process all CPU buffers ? */
Markus Armbruster59cc1852006-06-25 05:47:33 -0700119 mutex_lock(&buffer_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120 add_event_entry(ESCAPE_CODE);
121 add_event_entry(MODULE_LOADED_CODE);
Markus Armbruster59cc1852006-06-25 05:47:33 -0700122 mutex_unlock(&buffer_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123#endif
124 return 0;
125}
126
Robert Richter73185e02008-07-22 21:08:51 +0200127
Linus Torvalds1da177e2005-04-16 15:20:36 -0700128static struct notifier_block task_free_nb = {
129 .notifier_call = task_free_notify,
130};
131
132static struct notifier_block task_exit_nb = {
133 .notifier_call = task_exit_notify,
134};
135
136static struct notifier_block munmap_nb = {
137 .notifier_call = munmap_notify,
138};
139
140static struct notifier_block module_load_nb = {
141 .notifier_call = module_load_notify,
142};
143
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144int sync_start(void)
145{
146 int err;
147
Li Zefan79f55992009-06-15 14:58:26 +0800148 if (!zalloc_cpumask_var(&marked_cpus, GFP_KERNEL))
Robert Richter4c50d9e2009-01-22 14:14:14 +0100149 return -ENOMEM;
Robert Richter4c50d9e2009-01-22 14:14:14 +0100150
Robert Richter750d8572010-08-13 16:29:04 +0200151 mutex_lock(&buffer_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152
153 err = task_handoff_register(&task_free_nb);
154 if (err)
155 goto out1;
156 err = profile_event_register(PROFILE_TASK_EXIT, &task_exit_nb);
157 if (err)
158 goto out2;
159 err = profile_event_register(PROFILE_MUNMAP, &munmap_nb);
160 if (err)
161 goto out3;
162 err = register_module_notifier(&module_load_nb);
163 if (err)
164 goto out4;
165
Robert Richter750d8572010-08-13 16:29:04 +0200166 start_cpu_work();
167
Linus Torvalds1da177e2005-04-16 15:20:36 -0700168out:
Robert Richter750d8572010-08-13 16:29:04 +0200169 mutex_unlock(&buffer_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700170 return err;
171out4:
172 profile_event_unregister(PROFILE_MUNMAP, &munmap_nb);
173out3:
174 profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb);
175out2:
176 task_handoff_unregister(&task_free_nb);
177out1:
Robert Richter4c50d9e2009-01-22 14:14:14 +0100178 free_cpumask_var(marked_cpus);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700179 goto out;
180}
181
182
183void sync_stop(void)
184{
Robert Richter750d8572010-08-13 16:29:04 +0200185 /* flush buffers */
186 mutex_lock(&buffer_mutex);
187 end_cpu_work();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188 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);
Robert Richter750d8572010-08-13 16:29:04 +0200192 mutex_unlock(&buffer_mutex);
Tejun Heo3d7851b2010-10-15 09:51:08 -0400193 flush_cpu_work();
Robert Richter750d8572010-08-13 16:29:04 +0200194
195 /* make sure we don't leak task structs */
196 process_task_mortuary();
197 process_task_mortuary();
198
Robert Richter4c50d9e2009-01-22 14:14:14 +0100199 free_cpumask_var(marked_cpus);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700200}
201
Jan Blunck448678a2008-02-14 19:38:36 -0800202
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203/* Optimisation. We can manage without taking the dcookie sem
204 * because we cannot reach this code without at least one
205 * dcookie user still being registered (namely, the reader
206 * of the event buffer). */
Jan Blunck448678a2008-02-14 19:38:36 -0800207static inline unsigned long fast_get_dcookie(struct path *path)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208{
209 unsigned long cookie;
Jan Blunck448678a2008-02-14 19:38:36 -0800210
Nick Pigginc2452f32008-12-01 09:33:43 +0100211 if (path->dentry->d_flags & DCACHE_COOKIE)
Jan Blunck448678a2008-02-14 19:38:36 -0800212 return (unsigned long)path->dentry;
213 get_dcookie(path, &cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214 return cookie;
215}
216
Jan Blunck448678a2008-02-14 19:38:36 -0800217
Linus Torvalds1da177e2005-04-16 15:20:36 -0700218/* Look up the dcookie for the task's first VM_EXECUTABLE mapping,
219 * which corresponds loosely to "application name". This is
220 * not strictly necessary but allows oprofile to associate
221 * shared-library samples with particular applications
222 */
Robert Richter73185e02008-07-22 21:08:51 +0200223static unsigned long get_exec_dcookie(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700224{
John Levon0c0a4002005-06-23 22:02:47 -0700225 unsigned long cookie = NO_COOKIE;
Robert Richter73185e02008-07-22 21:08:51 +0200226 struct vm_area_struct *vma;
227
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228 if (!mm)
229 goto out;
Robert Richter73185e02008-07-22 21:08:51 +0200230
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231 for (vma = mm->mmap; vma; vma = vma->vm_next) {
232 if (!vma->vm_file)
233 continue;
234 if (!(vma->vm_flags & VM_EXECUTABLE))
235 continue;
Jan Blunck448678a2008-02-14 19:38:36 -0800236 cookie = fast_get_dcookie(&vma->vm_file->f_path);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700237 break;
238 }
239
240out:
241 return cookie;
242}
243
244
245/* Convert the EIP value of a sample into a persistent dentry/offset
246 * pair that can then be added to the global event buffer. We make
247 * sure to do this lookup before a mm->mmap modification happens so
248 * we don't lose track.
249 */
Robert Richter73185e02008-07-22 21:08:51 +0200250static unsigned long
251lookup_dcookie(struct mm_struct *mm, unsigned long addr, off_t *offset)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700252{
John Levon0c0a4002005-06-23 22:02:47 -0700253 unsigned long cookie = NO_COOKIE;
Robert Richter73185e02008-07-22 21:08:51 +0200254 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700255
256 for (vma = find_vma(mm, addr); vma; vma = vma->vm_next) {
Robert Richter73185e02008-07-22 21:08:51 +0200257
Linus Torvalds1da177e2005-04-16 15:20:36 -0700258 if (addr < vma->vm_start || addr >= vma->vm_end)
259 continue;
260
John Levon0c0a4002005-06-23 22:02:47 -0700261 if (vma->vm_file) {
Jan Blunck448678a2008-02-14 19:38:36 -0800262 cookie = fast_get_dcookie(&vma->vm_file->f_path);
John Levon0c0a4002005-06-23 22:02:47 -0700263 *offset = (vma->vm_pgoff << PAGE_SHIFT) + addr -
264 vma->vm_start;
265 } else {
266 /* must be an anonymous map */
267 *offset = addr;
268 }
269
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270 break;
271 }
272
John Levon0c0a4002005-06-23 22:02:47 -0700273 if (!vma)
274 cookie = INVALID_COOKIE;
275
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276 return cookie;
277}
278
John Levon0c0a4002005-06-23 22:02:47 -0700279static unsigned long last_cookie = INVALID_COOKIE;
Robert Richter73185e02008-07-22 21:08:51 +0200280
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281static void add_cpu_switch(int i)
282{
283 add_event_entry(ESCAPE_CODE);
284 add_event_entry(CPU_SWITCH_CODE);
285 add_event_entry(i);
John Levon0c0a4002005-06-23 22:02:47 -0700286 last_cookie = INVALID_COOKIE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287}
288
289static void add_kernel_ctx_switch(unsigned int in_kernel)
290{
291 add_event_entry(ESCAPE_CODE);
292 if (in_kernel)
Robert Richter73185e02008-07-22 21:08:51 +0200293 add_event_entry(KERNEL_ENTER_SWITCH_CODE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294 else
Robert Richter73185e02008-07-22 21:08:51 +0200295 add_event_entry(KERNEL_EXIT_SWITCH_CODE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296}
Robert Richter73185e02008-07-22 21:08:51 +0200297
Linus Torvalds1da177e2005-04-16 15:20:36 -0700298static void
Robert Richter73185e02008-07-22 21:08:51 +0200299add_user_ctx_switch(struct task_struct const *task, unsigned long cookie)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300{
301 add_event_entry(ESCAPE_CODE);
Robert Richter73185e02008-07-22 21:08:51 +0200302 add_event_entry(CTX_SWITCH_CODE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303 add_event_entry(task->pid);
304 add_event_entry(cookie);
305 /* Another code for daemon back-compat */
306 add_event_entry(ESCAPE_CODE);
307 add_event_entry(CTX_TGID_CODE);
308 add_event_entry(task->tgid);
309}
310
Robert Richter73185e02008-07-22 21:08:51 +0200311
Linus Torvalds1da177e2005-04-16 15:20:36 -0700312static void add_cookie_switch(unsigned long cookie)
313{
314 add_event_entry(ESCAPE_CODE);
315 add_event_entry(COOKIE_SWITCH_CODE);
316 add_event_entry(cookie);
317}
318
Robert Richter73185e02008-07-22 21:08:51 +0200319
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320static void add_trace_begin(void)
321{
322 add_event_entry(ESCAPE_CODE);
323 add_event_entry(TRACE_BEGIN_CODE);
324}
325
Robert Richter1acda872009-01-05 10:35:31 +0100326static void add_data(struct op_entry *entry, struct mm_struct *mm)
Barry Kasindorf345c2572008-07-22 21:08:54 +0200327{
Robert Richter1acda872009-01-05 10:35:31 +0100328 unsigned long code, pc, val;
329 unsigned long cookie;
Barry Kasindorf345c2572008-07-22 21:08:54 +0200330 off_t offset;
331
Robert Richter1acda872009-01-05 10:35:31 +0100332 if (!op_cpu_buffer_get_data(entry, &code))
Robert Richterdbe6e282008-12-16 11:01:18 +0100333 return;
Robert Richter1acda872009-01-05 10:35:31 +0100334 if (!op_cpu_buffer_get_data(entry, &pc))
335 return;
336 if (!op_cpu_buffer_get_size(entry))
337 return;
Barry Kasindorf345c2572008-07-22 21:08:54 +0200338
339 if (mm) {
Robert Richterd358e752009-01-05 13:14:04 +0100340 cookie = lookup_dcookie(mm, pc, &offset);
Barry Kasindorf345c2572008-07-22 21:08:54 +0200341
Robert Richterd358e752009-01-05 13:14:04 +0100342 if (cookie == NO_COOKIE)
343 offset = pc;
344 if (cookie == INVALID_COOKIE) {
Barry Kasindorf345c2572008-07-22 21:08:54 +0200345 atomic_inc(&oprofile_stats.sample_lost_no_mapping);
Robert Richterd358e752009-01-05 13:14:04 +0100346 offset = pc;
Barry Kasindorf345c2572008-07-22 21:08:54 +0200347 }
Robert Richterd358e752009-01-05 13:14:04 +0100348 if (cookie != last_cookie) {
349 add_cookie_switch(cookie);
350 last_cookie = cookie;
Barry Kasindorf345c2572008-07-22 21:08:54 +0200351 }
352 } else
Robert Richterd358e752009-01-05 13:14:04 +0100353 offset = pc;
Barry Kasindorf345c2572008-07-22 21:08:54 +0200354
355 add_event_entry(ESCAPE_CODE);
356 add_event_entry(code);
357 add_event_entry(offset); /* Offset from Dcookie */
358
Robert Richter1acda872009-01-05 10:35:31 +0100359 while (op_cpu_buffer_get_data(entry, &val))
360 add_event_entry(val);
Barry Kasindorf345c2572008-07-22 21:08:54 +0200361}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700362
Robert Richter6368a1f2008-12-29 18:44:21 +0100363static inline void add_sample_entry(unsigned long offset, unsigned long event)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364{
365 add_event_entry(offset);
366 add_event_entry(event);
367}
368
369
Robert Richter9741b302008-12-18 19:44:20 +0100370/*
371 * Add a sample to the global event buffer. If possible the
372 * sample is converted into a persistent dentry/offset pair
373 * for later lookup from userspace. Return 0 on failure.
374 */
375static int
376add_sample(struct mm_struct *mm, struct op_sample *s, int in_kernel)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377{
378 unsigned long cookie;
379 off_t offset;
Robert Richter73185e02008-07-22 21:08:51 +0200380
Robert Richter9741b302008-12-18 19:44:20 +0100381 if (in_kernel) {
382 add_sample_entry(s->eip, s->event);
383 return 1;
384 }
385
386 /* add userspace sample */
387
388 if (!mm) {
389 atomic_inc(&oprofile_stats.sample_lost_no_mm);
390 return 0;
391 }
392
Robert Richter73185e02008-07-22 21:08:51 +0200393 cookie = lookup_dcookie(mm, s->eip, &offset);
394
John Levon0c0a4002005-06-23 22:02:47 -0700395 if (cookie == INVALID_COOKIE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700396 atomic_inc(&oprofile_stats.sample_lost_no_mapping);
397 return 0;
398 }
399
400 if (cookie != last_cookie) {
401 add_cookie_switch(cookie);
402 last_cookie = cookie;
403 }
404
405 add_sample_entry(offset, s->event);
406
407 return 1;
408}
409
Robert Richter73185e02008-07-22 21:08:51 +0200410
Robert Richter73185e02008-07-22 21:08:51 +0200411static void release_mm(struct mm_struct *mm)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412{
413 if (!mm)
414 return;
415 up_read(&mm->mmap_sem);
416 mmput(mm);
417}
418
419
Robert Richter73185e02008-07-22 21:08:51 +0200420static struct mm_struct *take_tasks_mm(struct task_struct *task)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421{
Robert Richter73185e02008-07-22 21:08:51 +0200422 struct mm_struct *mm = get_task_mm(task);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423 if (mm)
424 down_read(&mm->mmap_sem);
425 return mm;
426}
427
428
429static inline int is_code(unsigned long val)
430{
431 return val == ESCAPE_CODE;
432}
Robert Richter73185e02008-07-22 21:08:51 +0200433
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435/* Move tasks along towards death. Any tasks on dead_tasks
436 * will definitely have no remaining references in any
437 * CPU buffers at this point, because we use two lists,
438 * and to have reached the list, it must have gone through
439 * one full sync already.
440 */
441static void process_task_mortuary(void)
442{
Paul E. McKenney4369ef32006-01-08 01:01:35 -0800443 unsigned long flags;
444 LIST_HEAD(local_dead_tasks);
Robert Richter73185e02008-07-22 21:08:51 +0200445 struct task_struct *task;
446 struct task_struct *ttask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700447
Paul E. McKenney4369ef32006-01-08 01:01:35 -0800448 spin_lock_irqsave(&task_mortuary, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449
Paul E. McKenney4369ef32006-01-08 01:01:35 -0800450 list_splice_init(&dead_tasks, &local_dead_tasks);
451 list_splice_init(&dying_tasks, &dead_tasks);
452
453 spin_unlock_irqrestore(&task_mortuary, flags);
454
455 list_for_each_entry_safe(task, ttask, &local_dead_tasks, tasks) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456 list_del(&task->tasks);
457 free_task(task);
458 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459}
460
461
462static void mark_done(int cpu)
463{
464 int i;
465
Rusty Russellf7df8ed2009-01-10 21:58:09 -0800466 cpumask_set_cpu(cpu, marked_cpus);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467
468 for_each_online_cpu(i) {
Rusty Russellf7df8ed2009-01-10 21:58:09 -0800469 if (!cpumask_test_cpu(i, marked_cpus))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700470 return;
471 }
472
473 /* All CPUs have been processed at least once,
474 * we can process the mortuary once
475 */
476 process_task_mortuary();
477
Rusty Russellf7df8ed2009-01-10 21:58:09 -0800478 cpumask_clear(marked_cpus);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700479}
480
481
482/* FIXME: this is not sufficient if we implement syscall barrier backtrace
483 * traversal, the code switch to sb_sample_start at first kernel enter/exit
484 * switch so we need a fifth state and some special handling in sync_buffer()
485 */
486typedef enum {
487 sb_bt_ignore = -2,
488 sb_buffer_start,
489 sb_bt_start,
490 sb_sample_start,
491} sync_buffer_state;
492
493/* Sync one of the CPU's buffers into the global event buffer.
494 * Here we need to go through each batch of samples punctuated
495 * by context switch notes, taking the task's mmap_sem and doing
496 * lookup in task->mm->mmap to convert EIP into dcookie/offset
497 * value.
498 */
499void sync_buffer(int cpu)
500{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501 struct mm_struct *mm = NULL;
Robert Richterfd7826d2008-09-26 17:50:31 -0400502 struct mm_struct *oldmm;
Robert Richterbd7dc462009-01-06 03:56:50 +0100503 unsigned long val;
Robert Richter73185e02008-07-22 21:08:51 +0200504 struct task_struct *new;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505 unsigned long cookie = 0;
506 int in_kernel = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507 sync_buffer_state state = sb_buffer_start;
Barry Kasindorf9b1f2612008-07-15 00:10:36 +0200508 unsigned int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509 unsigned long available;
Robert Richterae735e92008-12-25 17:26:07 +0100510 unsigned long flags;
Robert Richter2d87b142008-12-30 04:10:46 +0100511 struct op_entry entry;
512 struct op_sample *sample;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513
Markus Armbruster59cc1852006-06-25 05:47:33 -0700514 mutex_lock(&buffer_mutex);
Robert Richter73185e02008-07-22 21:08:51 +0200515
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516 add_cpu_switch(cpu);
517
Robert Richter6d2c53f2008-12-24 16:53:53 +0100518 op_cpu_buffer_reset(cpu);
519 available = op_cpu_buffer_entries(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520
521 for (i = 0; i < available; ++i) {
Robert Richter2d87b142008-12-30 04:10:46 +0100522 sample = op_cpu_buffer_read_entry(&entry, cpu);
523 if (!sample)
Robert Richter6dad8282008-12-09 01:21:32 +0100524 break;
Robert Richter73185e02008-07-22 21:08:51 +0200525
Robert Richter2d87b142008-12-30 04:10:46 +0100526 if (is_code(sample->eip)) {
Robert Richterae735e92008-12-25 17:26:07 +0100527 flags = sample->event;
528 if (flags & TRACE_BEGIN) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529 state = sb_bt_start;
530 add_trace_begin();
Robert Richterae735e92008-12-25 17:26:07 +0100531 }
532 if (flags & KERNEL_CTX_SWITCH) {
533 /* kernel/userspace switch */
534 in_kernel = flags & IS_KERNEL;
535 if (state == sb_buffer_start)
536 state = sb_sample_start;
537 add_kernel_ctx_switch(flags & IS_KERNEL);
538 }
Robert Richterbd7dc462009-01-06 03:56:50 +0100539 if (flags & USER_CTX_SWITCH
540 && op_cpu_buffer_get_data(&entry, &val)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541 /* userspace context switch */
Robert Richterbd7dc462009-01-06 03:56:50 +0100542 new = (struct task_struct *)val;
Robert Richterfd7826d2008-09-26 17:50:31 -0400543 oldmm = mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700544 release_mm(oldmm);
545 mm = take_tasks_mm(new);
546 if (mm != oldmm)
547 cookie = get_exec_dcookie(mm);
548 add_user_ctx_switch(new, cookie);
549 }
Robert Richter1acda872009-01-05 10:35:31 +0100550 if (op_cpu_buffer_get_size(&entry))
551 add_data(&entry, mm);
Robert Richter317f33b2008-12-18 19:44:20 +0100552 continue;
553 }
554
555 if (state < sb_bt_start)
556 /* ignore sample */
557 continue;
558
Robert Richter2d87b142008-12-30 04:10:46 +0100559 if (add_sample(mm, sample, in_kernel))
Robert Richter317f33b2008-12-18 19:44:20 +0100560 continue;
561
562 /* ignore backtraces if failed to add a sample */
563 if (state == sb_bt_start) {
564 state = sb_bt_ignore;
565 atomic_inc(&oprofile_stats.bt_lost_no_mapping);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567 }
568 release_mm(mm);
569
570 mark_done(cpu);
571
Markus Armbruster59cc1852006-06-25 05:47:33 -0700572 mutex_unlock(&buffer_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700573}
Carl Lovea5598ca2008-10-14 23:37:01 +0000574
575/* The function can be used to add a buffer worth of data directly to
576 * the kernel buffer. The buffer is assumed to be a circular buffer.
577 * Take the entries from index start and end at index end, wrapping
578 * at max_entries.
579 */
580void oprofile_put_buff(unsigned long *buf, unsigned int start,
581 unsigned int stop, unsigned int max)
582{
583 int i;
584
585 i = start;
586
587 mutex_lock(&buffer_mutex);
588 while (i != stop) {
589 add_event_entry(buf[i++]);
590
591 if (i >= max)
592 i = 0;
593 }
594
595 mutex_unlock(&buffer_mutex);
596}
597