blob: a0479976956686d5fd91c4470c140498b6d34ddf [file] [log] [blame]
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001/*
Ingo Molnar57c0c152009-09-21 12:20:38 +02002 * Performance events core code:
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003 *
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
8 *
Ingo Molnar57c0c152009-09-21 12:20:38 +02009 * For licensing details see kernel-base/COPYING
Ingo Molnarcdd6c482009-09-21 12:02:48 +020010 */
11
12#include <linux/fs.h>
13#include <linux/mm.h>
14#include <linux/cpu.h>
15#include <linux/smp.h>
16#include <linux/file.h>
17#include <linux/poll.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090018#include <linux/slab.h>
Frederic Weisbecker76e1d902010-04-05 15:35:57 +020019#include <linux/hash.h>
Ingo Molnarcdd6c482009-09-21 12:02:48 +020020#include <linux/sysfs.h>
21#include <linux/dcache.h>
22#include <linux/percpu.h>
23#include <linux/ptrace.h>
24#include <linux/vmstat.h>
Peter Zijlstra906010b2009-09-21 16:08:49 +020025#include <linux/vmalloc.h>
Ingo Molnarcdd6c482009-09-21 12:02:48 +020026#include <linux/hardirq.h>
27#include <linux/rculist.h>
28#include <linux/uaccess.h>
29#include <linux/syscalls.h>
30#include <linux/anon_inodes.h>
31#include <linux/kernel_stat.h>
32#include <linux/perf_event.h>
Li Zefan6fb29152009-10-15 11:21:42 +080033#include <linux/ftrace_event.h>
Jason Wessel3c502e72010-11-04 17:33:01 -050034#include <linux/hw_breakpoint.h>
Ingo Molnarcdd6c482009-09-21 12:02:48 +020035
36#include <asm/irq_regs.h>
37
Peter Zijlstra82cd6de2010-10-14 17:57:23 +020038atomic_t perf_task_events __read_mostly;
Ingo Molnarcdd6c482009-09-21 12:02:48 +020039static atomic_t nr_mmap_events __read_mostly;
40static atomic_t nr_comm_events __read_mostly;
41static atomic_t nr_task_events __read_mostly;
42
Peter Zijlstra108b02c2010-09-06 14:32:03 +020043static LIST_HEAD(pmus);
44static DEFINE_MUTEX(pmus_lock);
45static struct srcu_struct pmus_srcu;
46
Ingo Molnarcdd6c482009-09-21 12:02:48 +020047/*
48 * perf event paranoia level:
49 * -1 - not paranoid at all
50 * 0 - disallow raw tracepoint access for unpriv
51 * 1 - disallow cpu events for unpriv
52 * 2 - disallow kernel profiling for unpriv
53 */
54int sysctl_perf_event_paranoid __read_mostly = 1;
55
Ingo Molnarcdd6c482009-09-21 12:02:48 +020056int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
57
58/*
59 * max perf event sample rate
60 */
61int sysctl_perf_event_sample_rate __read_mostly = 100000;
62
63static atomic64_t perf_event_id;
64
Ingo Molnarcdd6c482009-09-21 12:02:48 +020065void __weak perf_event_print_debug(void) { }
66
Matt Fleming84c79912010-10-03 21:41:13 +010067extern __weak const char *perf_pmu_name(void)
Ingo Molnarcdd6c482009-09-21 12:02:48 +020068{
Matt Fleming84c79912010-10-03 21:41:13 +010069 return "pmu";
Ingo Molnarcdd6c482009-09-21 12:02:48 +020070}
71
Peter Zijlstra33696fc2010-06-14 08:49:00 +020072void perf_pmu_disable(struct pmu *pmu)
Ingo Molnarcdd6c482009-09-21 12:02:48 +020073{
Peter Zijlstra33696fc2010-06-14 08:49:00 +020074 int *count = this_cpu_ptr(pmu->pmu_disable_count);
75 if (!(*count)++)
76 pmu->pmu_disable(pmu);
Ingo Molnarcdd6c482009-09-21 12:02:48 +020077}
78
Peter Zijlstra33696fc2010-06-14 08:49:00 +020079void perf_pmu_enable(struct pmu *pmu)
Ingo Molnarcdd6c482009-09-21 12:02:48 +020080{
Peter Zijlstra33696fc2010-06-14 08:49:00 +020081 int *count = this_cpu_ptr(pmu->pmu_disable_count);
82 if (!--(*count))
83 pmu->pmu_enable(pmu);
Ingo Molnarcdd6c482009-09-21 12:02:48 +020084}
85
Peter Zijlstrae9d2b062010-09-17 11:28:50 +020086static DEFINE_PER_CPU(struct list_head, rotation_list);
87
88/*
89 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
90 * because they're strictly cpu affine and rotate_start is called with IRQs
91 * disabled, while rotate_context is called from IRQ context.
92 */
Peter Zijlstra108b02c2010-09-06 14:32:03 +020093static void perf_pmu_rotate_start(struct pmu *pmu)
Peter Zijlstrab5ab4cd2010-09-06 16:32:21 +020094{
Peter Zijlstra108b02c2010-09-06 14:32:03 +020095 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
Peter Zijlstrae9d2b062010-09-17 11:28:50 +020096 struct list_head *head = &__get_cpu_var(rotation_list);
Peter Zijlstrab5ab4cd2010-09-06 16:32:21 +020097
Peter Zijlstrae9d2b062010-09-17 11:28:50 +020098 WARN_ON(!irqs_disabled());
Peter Zijlstrab5ab4cd2010-09-06 16:32:21 +020099
Peter Zijlstrae9d2b062010-09-17 11:28:50 +0200100 if (list_empty(&cpuctx->rotation_list))
101 list_add(&cpuctx->rotation_list, head);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200102}
103
104static void get_ctx(struct perf_event_context *ctx)
105{
106 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
107}
108
109static void free_ctx(struct rcu_head *head)
110{
111 struct perf_event_context *ctx;
112
113 ctx = container_of(head, struct perf_event_context, rcu_head);
114 kfree(ctx);
115}
116
117static void put_ctx(struct perf_event_context *ctx)
118{
119 if (atomic_dec_and_test(&ctx->refcount)) {
120 if (ctx->parent_ctx)
121 put_ctx(ctx->parent_ctx);
122 if (ctx->task)
123 put_task_struct(ctx->task);
124 call_rcu(&ctx->rcu_head, free_ctx);
125 }
126}
127
128static void unclone_ctx(struct perf_event_context *ctx)
129{
130 if (ctx->parent_ctx) {
131 put_ctx(ctx->parent_ctx);
132 ctx->parent_ctx = NULL;
133 }
134}
135
Arnaldo Carvalho de Melo6844c092010-12-03 16:36:35 -0200136static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
137{
138 /*
139 * only top level events have the pid namespace they were created in
140 */
141 if (event->parent)
142 event = event->parent;
143
144 return task_tgid_nr_ns(p, event->ns);
145}
146
147static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
148{
149 /*
150 * only top level events have the pid namespace they were created in
151 */
152 if (event->parent)
153 event = event->parent;
154
155 return task_pid_nr_ns(p, event->ns);
156}
157
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200158/*
159 * If we inherit events we want to return the parent event id
160 * to userspace.
161 */
162static u64 primary_event_id(struct perf_event *event)
163{
164 u64 id = event->id;
165
166 if (event->parent)
167 id = event->parent->id;
168
169 return id;
170}
171
172/*
173 * Get the perf_event_context for a task and lock it.
174 * This has to cope with with the fact that until it is locked,
175 * the context could get moved to another task.
176 */
177static struct perf_event_context *
Peter Zijlstra8dc85d52010-09-02 16:50:03 +0200178perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200179{
180 struct perf_event_context *ctx;
181
182 rcu_read_lock();
Peter Zijlstra9ed60602010-06-11 17:36:35 +0200183retry:
Peter Zijlstra8dc85d52010-09-02 16:50:03 +0200184 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200185 if (ctx) {
186 /*
187 * If this context is a clone of another, it might
188 * get swapped for another underneath us by
189 * perf_event_task_sched_out, though the
190 * rcu_read_lock() protects us from any context
191 * getting freed. Lock the context and check if it
192 * got swapped before we could get the lock, and retry
193 * if so. If we locked the right context, then it
194 * can't get swapped on us any more.
195 */
Thomas Gleixnere625cce12009-11-17 18:02:06 +0100196 raw_spin_lock_irqsave(&ctx->lock, *flags);
Peter Zijlstra8dc85d52010-09-02 16:50:03 +0200197 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
Thomas Gleixnere625cce12009-11-17 18:02:06 +0100198 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200199 goto retry;
200 }
201
202 if (!atomic_inc_not_zero(&ctx->refcount)) {
Thomas Gleixnere625cce12009-11-17 18:02:06 +0100203 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200204 ctx = NULL;
205 }
206 }
207 rcu_read_unlock();
208 return ctx;
209}
210
211/*
212 * Get the context for a task and increment its pin_count so it
213 * can't get swapped to another task. This also increments its
214 * reference count so that the context can't get freed.
215 */
Peter Zijlstra8dc85d52010-09-02 16:50:03 +0200216static struct perf_event_context *
217perf_pin_task_context(struct task_struct *task, int ctxn)
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200218{
219 struct perf_event_context *ctx;
220 unsigned long flags;
221
Peter Zijlstra8dc85d52010-09-02 16:50:03 +0200222 ctx = perf_lock_task_context(task, ctxn, &flags);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200223 if (ctx) {
224 ++ctx->pin_count;
Thomas Gleixnere625cce12009-11-17 18:02:06 +0100225 raw_spin_unlock_irqrestore(&ctx->lock, flags);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200226 }
227 return ctx;
228}
229
230static void perf_unpin_context(struct perf_event_context *ctx)
231{
232 unsigned long flags;
233
Thomas Gleixnere625cce12009-11-17 18:02:06 +0100234 raw_spin_lock_irqsave(&ctx->lock, flags);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200235 --ctx->pin_count;
Thomas Gleixnere625cce12009-11-17 18:02:06 +0100236 raw_spin_unlock_irqrestore(&ctx->lock, flags);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200237 put_ctx(ctx);
238}
239
Peter Zijlstraf67218c2009-11-23 11:37:27 +0100240static inline u64 perf_clock(void)
241{
Peter Zijlstrac6763292010-05-25 10:48:51 +0200242 return local_clock();
Peter Zijlstraf67218c2009-11-23 11:37:27 +0100243}
244
245/*
246 * Update the record of the current time in a context.
247 */
248static void update_context_time(struct perf_event_context *ctx)
249{
250 u64 now = perf_clock();
251
252 ctx->time += now - ctx->timestamp;
253 ctx->timestamp = now;
254}
255
256/*
257 * Update the total_time_enabled and total_time_running fields for a event.
258 */
259static void update_event_times(struct perf_event *event)
260{
261 struct perf_event_context *ctx = event->ctx;
262 u64 run_end;
263
264 if (event->state < PERF_EVENT_STATE_INACTIVE ||
265 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
266 return;
267
Peter Zijlstraacd1d7c2009-11-23 15:00:36 +0100268 if (ctx->is_active)
269 run_end = ctx->time;
270 else
271 run_end = event->tstamp_stopped;
272
273 event->total_time_enabled = run_end - event->tstamp_enabled;
Peter Zijlstraf67218c2009-11-23 11:37:27 +0100274
275 if (event->state == PERF_EVENT_STATE_INACTIVE)
276 run_end = event->tstamp_stopped;
277 else
278 run_end = ctx->time;
279
280 event->total_time_running = run_end - event->tstamp_running;
281}
282
Peter Zijlstra96c21a42010-05-11 16:19:10 +0200283/*
284 * Update total_time_enabled and total_time_running for all events in a group.
285 */
286static void update_group_times(struct perf_event *leader)
287{
288 struct perf_event *event;
289
290 update_event_times(leader);
291 list_for_each_entry(event, &leader->sibling_list, group_entry)
292 update_event_times(event);
293}
294
Frederic Weisbecker889ff012010-01-09 20:04:47 +0100295static struct list_head *
296ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
297{
298 if (event->attr.pinned)
299 return &ctx->pinned_groups;
300 else
301 return &ctx->flexible_groups;
302}
303
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200304/*
305 * Add a event from the lists for its context.
306 * Must be called with ctx->mutex and ctx->lock held.
307 */
308static void
309list_add_event(struct perf_event *event, struct perf_event_context *ctx)
310{
Peter Zijlstra8a495422010-05-27 15:47:49 +0200311 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
312 event->attach_state |= PERF_ATTACH_CONTEXT;
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200313
314 /*
Peter Zijlstra8a495422010-05-27 15:47:49 +0200315 * If we're a stand alone event or group leader, we go to the context
316 * list, group events are kept attached to the group so that
317 * perf_group_detach can, at all times, locate all siblings.
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200318 */
Peter Zijlstra8a495422010-05-27 15:47:49 +0200319 if (event->group_leader == event) {
Frederic Weisbecker889ff012010-01-09 20:04:47 +0100320 struct list_head *list;
321
Frederic Weisbeckerd6f962b2010-01-10 01:25:51 +0100322 if (is_software_event(event))
323 event->group_flags |= PERF_GROUP_SOFTWARE;
324
Frederic Weisbecker889ff012010-01-09 20:04:47 +0100325 list = ctx_group_list(event, ctx);
326 list_add_tail(&event->group_entry, list);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200327 }
328
329 list_add_rcu(&event->event_entry, &ctx->event_list);
Peter Zijlstrab5ab4cd2010-09-06 16:32:21 +0200330 if (!ctx->nr_events)
Peter Zijlstra108b02c2010-09-06 14:32:03 +0200331 perf_pmu_rotate_start(ctx->pmu);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200332 ctx->nr_events++;
333 if (event->attr.inherit_stat)
334 ctx->nr_stat++;
335}
336
Arnaldo Carvalho de Meloc320c7b2010-10-20 12:50:11 -0200337/*
338 * Called at perf_event creation and when events are attached/detached from a
339 * group.
340 */
341static void perf_event__read_size(struct perf_event *event)
342{
343 int entry = sizeof(u64); /* value */
344 int size = 0;
345 int nr = 1;
346
347 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
348 size += sizeof(u64);
349
350 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
351 size += sizeof(u64);
352
353 if (event->attr.read_format & PERF_FORMAT_ID)
354 entry += sizeof(u64);
355
356 if (event->attr.read_format & PERF_FORMAT_GROUP) {
357 nr += event->group_leader->nr_siblings;
358 size += sizeof(u64);
359 }
360
361 size += entry * nr;
362 event->read_size = size;
363}
364
365static void perf_event__header_size(struct perf_event *event)
366{
367 struct perf_sample_data *data;
368 u64 sample_type = event->attr.sample_type;
369 u16 size = 0;
370
371 perf_event__read_size(event);
372
373 if (sample_type & PERF_SAMPLE_IP)
374 size += sizeof(data->ip);
375
Arnaldo Carvalho de Melo6844c092010-12-03 16:36:35 -0200376 if (sample_type & PERF_SAMPLE_ADDR)
377 size += sizeof(data->addr);
378
379 if (sample_type & PERF_SAMPLE_PERIOD)
380 size += sizeof(data->period);
381
382 if (sample_type & PERF_SAMPLE_READ)
383 size += event->read_size;
384
385 event->header_size = size;
386}
387
388static void perf_event__id_header_size(struct perf_event *event)
389{
390 struct perf_sample_data *data;
391 u64 sample_type = event->attr.sample_type;
392 u16 size = 0;
393
Arnaldo Carvalho de Meloc320c7b2010-10-20 12:50:11 -0200394 if (sample_type & PERF_SAMPLE_TID)
395 size += sizeof(data->tid_entry);
396
397 if (sample_type & PERF_SAMPLE_TIME)
398 size += sizeof(data->time);
399
Arnaldo Carvalho de Meloc320c7b2010-10-20 12:50:11 -0200400 if (sample_type & PERF_SAMPLE_ID)
401 size += sizeof(data->id);
402
403 if (sample_type & PERF_SAMPLE_STREAM_ID)
404 size += sizeof(data->stream_id);
405
406 if (sample_type & PERF_SAMPLE_CPU)
407 size += sizeof(data->cpu_entry);
408
Arnaldo Carvalho de Melo6844c092010-12-03 16:36:35 -0200409 event->id_header_size = size;
Arnaldo Carvalho de Meloc320c7b2010-10-20 12:50:11 -0200410}
411
Peter Zijlstra8a495422010-05-27 15:47:49 +0200412static void perf_group_attach(struct perf_event *event)
413{
Arnaldo Carvalho de Meloc320c7b2010-10-20 12:50:11 -0200414 struct perf_event *group_leader = event->group_leader, *pos;
Peter Zijlstra8a495422010-05-27 15:47:49 +0200415
Peter Zijlstra74c33372010-10-15 11:40:29 +0200416 /*
417 * We can have double attach due to group movement in perf_event_open.
418 */
419 if (event->attach_state & PERF_ATTACH_GROUP)
420 return;
421
Peter Zijlstra8a495422010-05-27 15:47:49 +0200422 event->attach_state |= PERF_ATTACH_GROUP;
423
424 if (group_leader == event)
425 return;
426
427 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
428 !is_software_event(event))
429 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
430
431 list_add_tail(&event->group_entry, &group_leader->sibling_list);
432 group_leader->nr_siblings++;
Arnaldo Carvalho de Meloc320c7b2010-10-20 12:50:11 -0200433
434 perf_event__header_size(group_leader);
435
436 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
437 perf_event__header_size(pos);
Peter Zijlstra8a495422010-05-27 15:47:49 +0200438}
439
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200440/*
441 * Remove a event from the lists for its context.
442 * Must be called with ctx->mutex and ctx->lock held.
443 */
444static void
445list_del_event(struct perf_event *event, struct perf_event_context *ctx)
446{
Peter Zijlstra8a495422010-05-27 15:47:49 +0200447 /*
448 * We can have double detach due to exit/hot-unplug + close.
449 */
450 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200451 return;
Peter Zijlstra8a495422010-05-27 15:47:49 +0200452
453 event->attach_state &= ~PERF_ATTACH_CONTEXT;
454
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200455 ctx->nr_events--;
456 if (event->attr.inherit_stat)
457 ctx->nr_stat--;
458
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200459 list_del_rcu(&event->event_entry);
460
Peter Zijlstra8a495422010-05-27 15:47:49 +0200461 if (event->group_leader == event)
462 list_del_init(&event->group_entry);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200463
Peter Zijlstra96c21a42010-05-11 16:19:10 +0200464 update_group_times(event);
Stephane Eranianb2e74a22009-11-26 09:24:30 -0800465
466 /*
467 * If event was in error state, then keep it
468 * that way, otherwise bogus counts will be
469 * returned on read(). The only way to get out
470 * of error state is by explicit re-enabling
471 * of the event
472 */
473 if (event->state > PERF_EVENT_STATE_OFF)
474 event->state = PERF_EVENT_STATE_OFF;
Peter Zijlstra050735b2010-05-11 11:51:53 +0200475}
476
Peter Zijlstra8a495422010-05-27 15:47:49 +0200477static void perf_group_detach(struct perf_event *event)
Peter Zijlstra050735b2010-05-11 11:51:53 +0200478{
479 struct perf_event *sibling, *tmp;
Peter Zijlstra8a495422010-05-27 15:47:49 +0200480 struct list_head *list = NULL;
481
482 /*
483 * We can have double detach due to exit/hot-unplug + close.
484 */
485 if (!(event->attach_state & PERF_ATTACH_GROUP))
486 return;
487
488 event->attach_state &= ~PERF_ATTACH_GROUP;
489
490 /*
491 * If this is a sibling, remove it from its group.
492 */
493 if (event->group_leader != event) {
494 list_del_init(&event->group_entry);
495 event->group_leader->nr_siblings--;
Arnaldo Carvalho de Meloc320c7b2010-10-20 12:50:11 -0200496 goto out;
Peter Zijlstra8a495422010-05-27 15:47:49 +0200497 }
498
499 if (!list_empty(&event->group_entry))
500 list = &event->group_entry;
Peter Zijlstra2e2af502009-11-23 11:37:25 +0100501
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200502 /*
503 * If this was a group event with sibling events then
504 * upgrade the siblings to singleton events by adding them
Peter Zijlstra8a495422010-05-27 15:47:49 +0200505 * to whatever list we are on.
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200506 */
507 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
Peter Zijlstra8a495422010-05-27 15:47:49 +0200508 if (list)
509 list_move_tail(&sibling->group_entry, list);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200510 sibling->group_leader = sibling;
Frederic Weisbeckerd6f962b2010-01-10 01:25:51 +0100511
512 /* Inherit group flags from the previous leader */
513 sibling->group_flags = event->group_flags;
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200514 }
Arnaldo Carvalho de Meloc320c7b2010-10-20 12:50:11 -0200515
516out:
517 perf_event__header_size(event->group_leader);
518
519 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
520 perf_event__header_size(tmp);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200521}
522
Stephane Eranianfa66f072010-08-26 16:40:01 +0200523static inline int
524event_filter_match(struct perf_event *event)
525{
526 return event->cpu == -1 || event->cpu == smp_processor_id();
527}
528
Stephane Eranian9ffcfa62010-10-20 15:25:01 +0200529static void
530event_sched_out(struct perf_event *event,
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200531 struct perf_cpu_context *cpuctx,
532 struct perf_event_context *ctx)
533{
Stephane Eranianfa66f072010-08-26 16:40:01 +0200534 u64 delta;
535 /*
536 * An event which could not be activated because of
537 * filter mismatch still needs to have its timings
538 * maintained, otherwise bogus information is return
539 * via read() for time_enabled, time_running:
540 */
541 if (event->state == PERF_EVENT_STATE_INACTIVE
542 && !event_filter_match(event)) {
543 delta = ctx->time - event->tstamp_stopped;
544 event->tstamp_running += delta;
545 event->tstamp_stopped = ctx->time;
546 }
547
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200548 if (event->state != PERF_EVENT_STATE_ACTIVE)
Stephane Eranian9ffcfa62010-10-20 15:25:01 +0200549 return;
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200550
551 event->state = PERF_EVENT_STATE_INACTIVE;
552 if (event->pending_disable) {
553 event->pending_disable = 0;
554 event->state = PERF_EVENT_STATE_OFF;
555 }
Stephane Eranian9ffcfa62010-10-20 15:25:01 +0200556 event->tstamp_stopped = ctx->time;
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +0200557 event->pmu->del(event, 0);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200558 event->oncpu = -1;
559
560 if (!is_software_event(event))
561 cpuctx->active_oncpu--;
562 ctx->nr_active--;
563 if (event->attr.exclusive || !cpuctx->active_oncpu)
564 cpuctx->exclusive = 0;
565}
566
567static void
568group_sched_out(struct perf_event *group_event,
569 struct perf_cpu_context *cpuctx,
570 struct perf_event_context *ctx)
571{
572 struct perf_event *event;
Stephane Eranianfa66f072010-08-26 16:40:01 +0200573 int state = group_event->state;
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200574
575 event_sched_out(group_event, cpuctx, ctx);
576
577 /*
578 * Schedule out siblings (if any):
579 */
580 list_for_each_entry(event, &group_event->sibling_list, group_entry)
581 event_sched_out(event, cpuctx, ctx);
582
Stephane Eranianfa66f072010-08-26 16:40:01 +0200583 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200584 cpuctx->exclusive = 0;
585}
586
Peter Zijlstra108b02c2010-09-06 14:32:03 +0200587static inline struct perf_cpu_context *
588__get_cpu_context(struct perf_event_context *ctx)
589{
590 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
591}
592
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200593/*
594 * Cross CPU call to remove a performance event
595 *
596 * We disable the event on the hardware level first. After that we
597 * remove it from the context list.
598 */
599static void __perf_event_remove_from_context(void *info)
600{
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200601 struct perf_event *event = info;
602 struct perf_event_context *ctx = event->ctx;
Peter Zijlstra108b02c2010-09-06 14:32:03 +0200603 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200604
605 /*
606 * If this is a task context, we need to check whether it is
607 * the current task context of this cpu. If not it has been
608 * scheduled out before the smp call arrived.
609 */
610 if (ctx->task && cpuctx->task_ctx != ctx)
611 return;
612
Thomas Gleixnere625cce12009-11-17 18:02:06 +0100613 raw_spin_lock(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200614
615 event_sched_out(event, cpuctx, ctx);
616
617 list_del_event(event, ctx);
618
Thomas Gleixnere625cce12009-11-17 18:02:06 +0100619 raw_spin_unlock(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200620}
621
622
623/*
624 * Remove the event from a task's (or a CPU's) list of events.
625 *
626 * Must be called with ctx->mutex held.
627 *
628 * CPU events are removed with a smp call. For task events we only
629 * call when the task is on a CPU.
630 *
631 * If event->ctx is a cloned context, callers must make sure that
632 * every task struct that event->ctx->task could possibly point to
633 * remains valid. This is OK when called from perf_release since
634 * that only calls us on the top-level context, which can't be a clone.
635 * When called from perf_event_exit_task, it's OK because the
636 * context has been detached from its task.
637 */
638static void perf_event_remove_from_context(struct perf_event *event)
639{
640 struct perf_event_context *ctx = event->ctx;
641 struct task_struct *task = ctx->task;
642
643 if (!task) {
644 /*
645 * Per cpu events are removed via an smp call and
André Goddard Rosaaf901ca2009-11-14 13:09:05 -0200646 * the removal is always successful.
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200647 */
648 smp_call_function_single(event->cpu,
649 __perf_event_remove_from_context,
650 event, 1);
651 return;
652 }
653
654retry:
655 task_oncpu_function_call(task, __perf_event_remove_from_context,
656 event);
657
Thomas Gleixnere625cce12009-11-17 18:02:06 +0100658 raw_spin_lock_irq(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200659 /*
660 * If the context is active we need to retry the smp call.
661 */
662 if (ctx->nr_active && !list_empty(&event->group_entry)) {
Thomas Gleixnere625cce12009-11-17 18:02:06 +0100663 raw_spin_unlock_irq(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200664 goto retry;
665 }
666
667 /*
668 * The lock prevents that this context is scheduled in so we
669 * can remove the event safely, if the call above did not
670 * succeed.
671 */
Peter Zijlstra6c2bfcb2009-11-23 11:37:24 +0100672 if (!list_empty(&event->group_entry))
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200673 list_del_event(event, ctx);
Thomas Gleixnere625cce12009-11-17 18:02:06 +0100674 raw_spin_unlock_irq(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200675}
676
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200677/*
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200678 * Cross CPU call to disable a performance event
679 */
680static void __perf_event_disable(void *info)
681{
682 struct perf_event *event = info;
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200683 struct perf_event_context *ctx = event->ctx;
Peter Zijlstra108b02c2010-09-06 14:32:03 +0200684 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200685
686 /*
687 * If this is a per-task event, need to check whether this
688 * event's task is the current task on this cpu.
689 */
690 if (ctx->task && cpuctx->task_ctx != ctx)
691 return;
692
Thomas Gleixnere625cce12009-11-17 18:02:06 +0100693 raw_spin_lock(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200694
695 /*
696 * If the event is on, turn it off.
697 * If it is in error state, leave it in error state.
698 */
699 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
700 update_context_time(ctx);
701 update_group_times(event);
702 if (event == event->group_leader)
703 group_sched_out(event, cpuctx, ctx);
704 else
705 event_sched_out(event, cpuctx, ctx);
706 event->state = PERF_EVENT_STATE_OFF;
707 }
708
Thomas Gleixnere625cce12009-11-17 18:02:06 +0100709 raw_spin_unlock(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200710}
711
712/*
713 * Disable a event.
714 *
715 * If event->ctx is a cloned context, callers must make sure that
716 * every task struct that event->ctx->task could possibly point to
717 * remains valid. This condition is satisifed when called through
718 * perf_event_for_each_child or perf_event_for_each because they
719 * hold the top-level event's child_mutex, so any descendant that
720 * goes to exit will block in sync_child_event.
721 * When called from perf_pending_event it's OK because event->ctx
722 * is the current context on this CPU and preemption is disabled,
723 * hence we can't get into perf_event_task_sched_out for this context.
724 */
Frederic Weisbecker44234ad2009-12-09 09:25:48 +0100725void perf_event_disable(struct perf_event *event)
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200726{
727 struct perf_event_context *ctx = event->ctx;
728 struct task_struct *task = ctx->task;
729
730 if (!task) {
731 /*
732 * Disable the event on the cpu that it's on
733 */
734 smp_call_function_single(event->cpu, __perf_event_disable,
735 event, 1);
736 return;
737 }
738
Peter Zijlstra9ed60602010-06-11 17:36:35 +0200739retry:
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200740 task_oncpu_function_call(task, __perf_event_disable, event);
741
Thomas Gleixnere625cce12009-11-17 18:02:06 +0100742 raw_spin_lock_irq(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200743 /*
744 * If the event is still active, we need to retry the cross-call.
745 */
746 if (event->state == PERF_EVENT_STATE_ACTIVE) {
Thomas Gleixnere625cce12009-11-17 18:02:06 +0100747 raw_spin_unlock_irq(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200748 goto retry;
749 }
750
751 /*
752 * Since we have the lock this context can't be scheduled
753 * in, so we can change the state safely.
754 */
755 if (event->state == PERF_EVENT_STATE_INACTIVE) {
756 update_group_times(event);
757 event->state = PERF_EVENT_STATE_OFF;
758 }
759
Thomas Gleixnere625cce12009-11-17 18:02:06 +0100760 raw_spin_unlock_irq(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200761}
762
763static int
Stephane Eranian9ffcfa62010-10-20 15:25:01 +0200764event_sched_in(struct perf_event *event,
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200765 struct perf_cpu_context *cpuctx,
Peter Zijlstra6e377382010-02-11 13:21:58 +0100766 struct perf_event_context *ctx)
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200767{
768 if (event->state <= PERF_EVENT_STATE_OFF)
769 return 0;
770
771 event->state = PERF_EVENT_STATE_ACTIVE;
Peter Zijlstra6e377382010-02-11 13:21:58 +0100772 event->oncpu = smp_processor_id();
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200773 /*
774 * The new state must be visible before we turn it on in the hardware:
775 */
776 smp_wmb();
777
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +0200778 if (event->pmu->add(event, PERF_EF_START)) {
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200779 event->state = PERF_EVENT_STATE_INACTIVE;
780 event->oncpu = -1;
781 return -EAGAIN;
782 }
783
Stephane Eranian9ffcfa62010-10-20 15:25:01 +0200784 event->tstamp_running += ctx->time - event->tstamp_stopped;
785
Stephane Eranianeed01522010-10-26 16:08:01 +0200786 event->shadow_ctx_time = ctx->time - ctx->timestamp;
787
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200788 if (!is_software_event(event))
789 cpuctx->active_oncpu++;
790 ctx->nr_active++;
791
792 if (event->attr.exclusive)
793 cpuctx->exclusive = 1;
794
795 return 0;
796}
797
798static int
799group_sched_in(struct perf_event *group_event,
800 struct perf_cpu_context *cpuctx,
Peter Zijlstra6e377382010-02-11 13:21:58 +0100801 struct perf_event_context *ctx)
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200802{
Lin Ming6bde9b62010-04-23 13:56:00 +0800803 struct perf_event *event, *partial_group = NULL;
Peter Zijlstra51b0fe32010-06-11 13:35:57 +0200804 struct pmu *pmu = group_event->pmu;
Stephane Eraniand7842da2010-10-20 15:25:01 +0200805 u64 now = ctx->time;
806 bool simulate = false;
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200807
808 if (group_event->state == PERF_EVENT_STATE_OFF)
809 return 0;
810
Peter Zijlstraad5133b2010-06-15 12:22:39 +0200811 pmu->start_txn(pmu);
Lin Ming6bde9b62010-04-23 13:56:00 +0800812
Stephane Eranian9ffcfa62010-10-20 15:25:01 +0200813 if (event_sched_in(group_event, cpuctx, ctx)) {
Peter Zijlstraad5133b2010-06-15 12:22:39 +0200814 pmu->cancel_txn(pmu);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200815 return -EAGAIN;
Stephane Eranian90151c352010-05-25 16:23:10 +0200816 }
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200817
818 /*
819 * Schedule in siblings as one group (if any):
820 */
821 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
Stephane Eranian9ffcfa62010-10-20 15:25:01 +0200822 if (event_sched_in(event, cpuctx, ctx)) {
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200823 partial_group = event;
824 goto group_error;
825 }
826 }
827
Stephane Eranian9ffcfa62010-10-20 15:25:01 +0200828 if (!pmu->commit_txn(pmu))
Paul Mackerras6e851582010-05-08 20:58:00 +1000829 return 0;
Stephane Eranian9ffcfa62010-10-20 15:25:01 +0200830
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200831group_error:
832 /*
833 * Groups can be scheduled in as one unit only, so undo any
834 * partial group before returning:
Stephane Eraniand7842da2010-10-20 15:25:01 +0200835 * The events up to the failed event are scheduled out normally,
836 * tstamp_stopped will be updated.
837 *
838 * The failed events and the remaining siblings need to have
839 * their timings updated as if they had gone thru event_sched_in()
840 * and event_sched_out(). This is required to get consistent timings
841 * across the group. This also takes care of the case where the group
842 * could never be scheduled by ensuring tstamp_stopped is set to mark
843 * the time the event was actually stopped, such that time delta
844 * calculation in update_event_times() is correct.
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200845 */
846 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
847 if (event == partial_group)
Stephane Eraniand7842da2010-10-20 15:25:01 +0200848 simulate = true;
849
850 if (simulate) {
851 event->tstamp_running += now - event->tstamp_stopped;
852 event->tstamp_stopped = now;
853 } else {
854 event_sched_out(event, cpuctx, ctx);
855 }
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200856 }
Stephane Eranian9ffcfa62010-10-20 15:25:01 +0200857 event_sched_out(group_event, cpuctx, ctx);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200858
Peter Zijlstraad5133b2010-06-15 12:22:39 +0200859 pmu->cancel_txn(pmu);
Stephane Eranian90151c352010-05-25 16:23:10 +0200860
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200861 return -EAGAIN;
862}
863
864/*
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200865 * Work out whether we can put this event group on the CPU now.
866 */
867static int group_can_go_on(struct perf_event *event,
868 struct perf_cpu_context *cpuctx,
869 int can_add_hw)
870{
871 /*
872 * Groups consisting entirely of software events can always go on.
873 */
Frederic Weisbeckerd6f962b2010-01-10 01:25:51 +0100874 if (event->group_flags & PERF_GROUP_SOFTWARE)
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200875 return 1;
876 /*
877 * If an exclusive group is already on, no other hardware
878 * events can go on.
879 */
880 if (cpuctx->exclusive)
881 return 0;
882 /*
883 * If this group is exclusive and there are already
884 * events on the CPU, it can't go on.
885 */
886 if (event->attr.exclusive && cpuctx->active_oncpu)
887 return 0;
888 /*
889 * Otherwise, try to add it if all previous groups were able
890 * to go on.
891 */
892 return can_add_hw;
893}
894
895static void add_event_to_ctx(struct perf_event *event,
896 struct perf_event_context *ctx)
897{
898 list_add_event(event, ctx);
Peter Zijlstra8a495422010-05-27 15:47:49 +0200899 perf_group_attach(event);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200900 event->tstamp_enabled = ctx->time;
901 event->tstamp_running = ctx->time;
902 event->tstamp_stopped = ctx->time;
903}
904
905/*
906 * Cross CPU call to install and enable a performance event
907 *
908 * Must be called with ctx->mutex held
909 */
910static void __perf_install_in_context(void *info)
911{
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200912 struct perf_event *event = info;
913 struct perf_event_context *ctx = event->ctx;
914 struct perf_event *leader = event->group_leader;
Peter Zijlstra108b02c2010-09-06 14:32:03 +0200915 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200916 int err;
917
918 /*
919 * If this is a task context, we need to check whether it is
920 * the current task context of this cpu. If not it has been
921 * scheduled out before the smp call arrived.
922 * Or possibly this is the right context but it isn't
923 * on this cpu because it had no events.
924 */
925 if (ctx->task && cpuctx->task_ctx != ctx) {
926 if (cpuctx->task_ctx || ctx->task != current)
927 return;
928 cpuctx->task_ctx = ctx;
929 }
930
Thomas Gleixnere625cce12009-11-17 18:02:06 +0100931 raw_spin_lock(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200932 ctx->is_active = 1;
933 update_context_time(ctx);
934
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200935 add_event_to_ctx(event, ctx);
936
Peter Zijlstraf4c41762009-12-16 17:55:54 +0100937 if (event->cpu != -1 && event->cpu != smp_processor_id())
938 goto unlock;
939
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200940 /*
941 * Don't put the event on if it is disabled or if
942 * it is in a group and the group isn't on.
943 */
944 if (event->state != PERF_EVENT_STATE_INACTIVE ||
945 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
946 goto unlock;
947
948 /*
949 * An exclusive event can't go on if there are already active
950 * hardware events, and no hardware event can go on if there
951 * is already an exclusive event on.
952 */
953 if (!group_can_go_on(event, cpuctx, 1))
954 err = -EEXIST;
955 else
Peter Zijlstra6e377382010-02-11 13:21:58 +0100956 err = event_sched_in(event, cpuctx, ctx);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200957
958 if (err) {
959 /*
960 * This event couldn't go on. If it is in a group
961 * then we have to pull the whole group off.
962 * If the event group is pinned then put it in error state.
963 */
964 if (leader != event)
965 group_sched_out(leader, cpuctx, ctx);
966 if (leader->attr.pinned) {
967 update_group_times(leader);
968 leader->state = PERF_EVENT_STATE_ERROR;
969 }
970 }
971
Peter Zijlstra9ed60602010-06-11 17:36:35 +0200972unlock:
Thomas Gleixnere625cce12009-11-17 18:02:06 +0100973 raw_spin_unlock(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200974}
975
976/*
977 * Attach a performance event to a context
978 *
979 * First we add the event to the list with the hardware enable bit
980 * in event->hw_config cleared.
981 *
982 * If the event is attached to a task which is on a CPU we use a smp
983 * call to enable it in the task context. The task might have been
984 * scheduled away, but we check this in the smp call again.
985 *
986 * Must be called with ctx->mutex held.
987 */
988static void
989perf_install_in_context(struct perf_event_context *ctx,
990 struct perf_event *event,
991 int cpu)
992{
993 struct task_struct *task = ctx->task;
994
Peter Zijlstrac3f00c72010-08-18 14:37:15 +0200995 event->ctx = ctx;
996
Ingo Molnarcdd6c482009-09-21 12:02:48 +0200997 if (!task) {
998 /*
999 * Per cpu events are installed via an smp call and
André Goddard Rosaaf901ca2009-11-14 13:09:05 -02001000 * the install is always successful.
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001001 */
1002 smp_call_function_single(cpu, __perf_install_in_context,
1003 event, 1);
1004 return;
1005 }
1006
1007retry:
1008 task_oncpu_function_call(task, __perf_install_in_context,
1009 event);
1010
Thomas Gleixnere625cce12009-11-17 18:02:06 +01001011 raw_spin_lock_irq(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001012 /*
1013 * we need to retry the smp call.
1014 */
1015 if (ctx->is_active && list_empty(&event->group_entry)) {
Thomas Gleixnere625cce12009-11-17 18:02:06 +01001016 raw_spin_unlock_irq(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001017 goto retry;
1018 }
1019
1020 /*
1021 * The lock prevents that this context is scheduled in so we
1022 * can add the event safely, if it the call above did not
1023 * succeed.
1024 */
1025 if (list_empty(&event->group_entry))
1026 add_event_to_ctx(event, ctx);
Thomas Gleixnere625cce12009-11-17 18:02:06 +01001027 raw_spin_unlock_irq(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001028}
1029
1030/*
1031 * Put a event into inactive state and update time fields.
1032 * Enabling the leader of a group effectively enables all
1033 * the group members that aren't explicitly disabled, so we
1034 * have to update their ->tstamp_enabled also.
1035 * Note: this works for group members as well as group leaders
1036 * since the non-leader members' sibling_lists will be empty.
1037 */
1038static void __perf_event_mark_enabled(struct perf_event *event,
1039 struct perf_event_context *ctx)
1040{
1041 struct perf_event *sub;
1042
1043 event->state = PERF_EVENT_STATE_INACTIVE;
1044 event->tstamp_enabled = ctx->time - event->total_time_enabled;
Peter Zijlstra9ed60602010-06-11 17:36:35 +02001045 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1046 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001047 sub->tstamp_enabled =
1048 ctx->time - sub->total_time_enabled;
Peter Zijlstra9ed60602010-06-11 17:36:35 +02001049 }
1050 }
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001051}
1052
1053/*
1054 * Cross CPU call to enable a performance event
1055 */
1056static void __perf_event_enable(void *info)
1057{
1058 struct perf_event *event = info;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001059 struct perf_event_context *ctx = event->ctx;
1060 struct perf_event *leader = event->group_leader;
Peter Zijlstra108b02c2010-09-06 14:32:03 +02001061 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001062 int err;
1063
1064 /*
1065 * If this is a per-task event, need to check whether this
1066 * event's task is the current task on this cpu.
1067 */
1068 if (ctx->task && cpuctx->task_ctx != ctx) {
1069 if (cpuctx->task_ctx || ctx->task != current)
1070 return;
1071 cpuctx->task_ctx = ctx;
1072 }
1073
Thomas Gleixnere625cce12009-11-17 18:02:06 +01001074 raw_spin_lock(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001075 ctx->is_active = 1;
1076 update_context_time(ctx);
1077
1078 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1079 goto unlock;
1080 __perf_event_mark_enabled(event, ctx);
1081
Peter Zijlstraf4c41762009-12-16 17:55:54 +01001082 if (event->cpu != -1 && event->cpu != smp_processor_id())
1083 goto unlock;
1084
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001085 /*
1086 * If the event is in a group and isn't the group leader,
1087 * then don't put it on unless the group is on.
1088 */
1089 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
1090 goto unlock;
1091
1092 if (!group_can_go_on(event, cpuctx, 1)) {
1093 err = -EEXIST;
1094 } else {
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001095 if (event == leader)
Peter Zijlstra6e377382010-02-11 13:21:58 +01001096 err = group_sched_in(event, cpuctx, ctx);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001097 else
Peter Zijlstra6e377382010-02-11 13:21:58 +01001098 err = event_sched_in(event, cpuctx, ctx);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001099 }
1100
1101 if (err) {
1102 /*
1103 * If this event can't go on and it's part of a
1104 * group, then the whole group has to come off.
1105 */
1106 if (leader != event)
1107 group_sched_out(leader, cpuctx, ctx);
1108 if (leader->attr.pinned) {
1109 update_group_times(leader);
1110 leader->state = PERF_EVENT_STATE_ERROR;
1111 }
1112 }
1113
Peter Zijlstra9ed60602010-06-11 17:36:35 +02001114unlock:
Thomas Gleixnere625cce12009-11-17 18:02:06 +01001115 raw_spin_unlock(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001116}
1117
1118/*
1119 * Enable a event.
1120 *
1121 * If event->ctx is a cloned context, callers must make sure that
1122 * every task struct that event->ctx->task could possibly point to
1123 * remains valid. This condition is satisfied when called through
1124 * perf_event_for_each_child or perf_event_for_each as described
1125 * for perf_event_disable.
1126 */
Frederic Weisbecker44234ad2009-12-09 09:25:48 +01001127void perf_event_enable(struct perf_event *event)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001128{
1129 struct perf_event_context *ctx = event->ctx;
1130 struct task_struct *task = ctx->task;
1131
1132 if (!task) {
1133 /*
1134 * Enable the event on the cpu that it's on
1135 */
1136 smp_call_function_single(event->cpu, __perf_event_enable,
1137 event, 1);
1138 return;
1139 }
1140
Thomas Gleixnere625cce12009-11-17 18:02:06 +01001141 raw_spin_lock_irq(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001142 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1143 goto out;
1144
1145 /*
1146 * If the event is in error state, clear that first.
1147 * That way, if we see the event in error state below, we
1148 * know that it has gone back into error state, as distinct
1149 * from the task having been scheduled away before the
1150 * cross-call arrived.
1151 */
1152 if (event->state == PERF_EVENT_STATE_ERROR)
1153 event->state = PERF_EVENT_STATE_OFF;
1154
Peter Zijlstra9ed60602010-06-11 17:36:35 +02001155retry:
Thomas Gleixnere625cce12009-11-17 18:02:06 +01001156 raw_spin_unlock_irq(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001157 task_oncpu_function_call(task, __perf_event_enable, event);
1158
Thomas Gleixnere625cce12009-11-17 18:02:06 +01001159 raw_spin_lock_irq(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001160
1161 /*
1162 * If the context is active and the event is still off,
1163 * we need to retry the cross-call.
1164 */
1165 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1166 goto retry;
1167
1168 /*
1169 * Since we have the lock this context can't be scheduled
1170 * in, so we can change the state safely.
1171 */
1172 if (event->state == PERF_EVENT_STATE_OFF)
1173 __perf_event_mark_enabled(event, ctx);
1174
Peter Zijlstra9ed60602010-06-11 17:36:35 +02001175out:
Thomas Gleixnere625cce12009-11-17 18:02:06 +01001176 raw_spin_unlock_irq(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001177}
1178
1179static int perf_event_refresh(struct perf_event *event, int refresh)
1180{
1181 /*
1182 * not supported on inherited events
1183 */
Franck Bui-Huu2e939d12010-11-23 16:21:44 +01001184 if (event->attr.inherit || !is_sampling_event(event))
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001185 return -EINVAL;
1186
1187 atomic_add(refresh, &event->event_limit);
1188 perf_event_enable(event);
1189
1190 return 0;
1191}
1192
Frederic Weisbecker5b0311e2010-01-17 11:59:13 +01001193enum event_type_t {
1194 EVENT_FLEXIBLE = 0x1,
1195 EVENT_PINNED = 0x2,
1196 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1197};
1198
1199static void ctx_sched_out(struct perf_event_context *ctx,
1200 struct perf_cpu_context *cpuctx,
1201 enum event_type_t event_type)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001202{
1203 struct perf_event *event;
1204
Thomas Gleixnere625cce12009-11-17 18:02:06 +01001205 raw_spin_lock(&ctx->lock);
Peter Zijlstra1b9a6442010-09-07 18:32:22 +02001206 perf_pmu_disable(ctx->pmu);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001207 ctx->is_active = 0;
1208 if (likely(!ctx->nr_events))
1209 goto out;
1210 update_context_time(ctx);
1211
Frederic Weisbecker5b0311e2010-01-17 11:59:13 +01001212 if (!ctx->nr_active)
Peter Zijlstra24cd7f52010-06-11 17:32:03 +02001213 goto out;
Frederic Weisbecker5b0311e2010-01-17 11:59:13 +01001214
Peter Zijlstra9ed60602010-06-11 17:36:35 +02001215 if (event_type & EVENT_PINNED) {
Frederic Weisbecker889ff012010-01-09 20:04:47 +01001216 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1217 group_sched_out(event, cpuctx, ctx);
Peter Zijlstra9ed60602010-06-11 17:36:35 +02001218 }
Frederic Weisbecker889ff012010-01-09 20:04:47 +01001219
Peter Zijlstra9ed60602010-06-11 17:36:35 +02001220 if (event_type & EVENT_FLEXIBLE) {
Frederic Weisbecker889ff012010-01-09 20:04:47 +01001221 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
Xiao Guangrong8c9ed8e2009-09-25 13:51:17 +08001222 group_sched_out(event, cpuctx, ctx);
Peter Zijlstra9ed60602010-06-11 17:36:35 +02001223 }
Peter Zijlstra9ed60602010-06-11 17:36:35 +02001224out:
Peter Zijlstra1b9a6442010-09-07 18:32:22 +02001225 perf_pmu_enable(ctx->pmu);
Thomas Gleixnere625cce12009-11-17 18:02:06 +01001226 raw_spin_unlock(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001227}
1228
1229/*
1230 * Test whether two contexts are equivalent, i.e. whether they
1231 * have both been cloned from the same version of the same context
1232 * and they both have the same number of enabled events.
1233 * If the number of enabled events is the same, then the set
1234 * of enabled events should be the same, because these are both
1235 * inherited contexts, therefore we can't access individual events
1236 * in them directly with an fd; we can only enable/disable all
1237 * events via prctl, or enable/disable all events in a family
1238 * via ioctl, which will have the same effect on both contexts.
1239 */
1240static int context_equiv(struct perf_event_context *ctx1,
1241 struct perf_event_context *ctx2)
1242{
1243 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1244 && ctx1->parent_gen == ctx2->parent_gen
1245 && !ctx1->pin_count && !ctx2->pin_count;
1246}
1247
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001248static void __perf_event_sync_stat(struct perf_event *event,
1249 struct perf_event *next_event)
1250{
1251 u64 value;
1252
1253 if (!event->attr.inherit_stat)
1254 return;
1255
1256 /*
1257 * Update the event value, we cannot use perf_event_read()
1258 * because we're in the middle of a context switch and have IRQs
1259 * disabled, which upsets smp_call_function_single(), however
1260 * we know the event must be on the current CPU, therefore we
1261 * don't need to use it.
1262 */
1263 switch (event->state) {
1264 case PERF_EVENT_STATE_ACTIVE:
Peter Zijlstra3dbebf12009-11-20 22:19:52 +01001265 event->pmu->read(event);
1266 /* fall-through */
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001267
1268 case PERF_EVENT_STATE_INACTIVE:
1269 update_event_times(event);
1270 break;
1271
1272 default:
1273 break;
1274 }
1275
1276 /*
1277 * In order to keep per-task stats reliable we need to flip the event
1278 * values when we flip the contexts.
1279 */
Peter Zijlstrae7850592010-05-21 14:43:08 +02001280 value = local64_read(&next_event->count);
1281 value = local64_xchg(&event->count, value);
1282 local64_set(&next_event->count, value);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001283
1284 swap(event->total_time_enabled, next_event->total_time_enabled);
1285 swap(event->total_time_running, next_event->total_time_running);
1286
1287 /*
1288 * Since we swizzled the values, update the user visible data too.
1289 */
1290 perf_event_update_userpage(event);
1291 perf_event_update_userpage(next_event);
1292}
1293
1294#define list_next_entry(pos, member) \
1295 list_entry(pos->member.next, typeof(*pos), member)
1296
1297static void perf_event_sync_stat(struct perf_event_context *ctx,
1298 struct perf_event_context *next_ctx)
1299{
1300 struct perf_event *event, *next_event;
1301
1302 if (!ctx->nr_stat)
1303 return;
1304
Peter Zijlstra02ffdbc2009-11-20 22:19:50 +01001305 update_context_time(ctx);
1306
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001307 event = list_first_entry(&ctx->event_list,
1308 struct perf_event, event_entry);
1309
1310 next_event = list_first_entry(&next_ctx->event_list,
1311 struct perf_event, event_entry);
1312
1313 while (&event->event_entry != &ctx->event_list &&
1314 &next_event->event_entry != &next_ctx->event_list) {
1315
1316 __perf_event_sync_stat(event, next_event);
1317
1318 event = list_next_entry(event, event_entry);
1319 next_event = list_next_entry(next_event, event_entry);
1320 }
1321}
1322
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02001323void perf_event_context_sched_out(struct task_struct *task, int ctxn,
1324 struct task_struct *next)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001325{
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02001326 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001327 struct perf_event_context *next_ctx;
1328 struct perf_event_context *parent;
Peter Zijlstra108b02c2010-09-06 14:32:03 +02001329 struct perf_cpu_context *cpuctx;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001330 int do_switch = 1;
1331
Peter Zijlstra108b02c2010-09-06 14:32:03 +02001332 if (likely(!ctx))
1333 return;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001334
Peter Zijlstra108b02c2010-09-06 14:32:03 +02001335 cpuctx = __get_cpu_context(ctx);
1336 if (!cpuctx->task_ctx)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001337 return;
1338
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001339 rcu_read_lock();
1340 parent = rcu_dereference(ctx->parent_ctx);
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02001341 next_ctx = next->perf_event_ctxp[ctxn];
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001342 if (parent && next_ctx &&
1343 rcu_dereference(next_ctx->parent_ctx) == parent) {
1344 /*
1345 * Looks like the two contexts are clones, so we might be
1346 * able to optimize the context switch. We lock both
1347 * contexts and check that they are clones under the
1348 * lock (including re-checking that neither has been
1349 * uncloned in the meantime). It doesn't matter which
1350 * order we take the locks because no other cpu could
1351 * be trying to lock both of these tasks.
1352 */
Thomas Gleixnere625cce12009-11-17 18:02:06 +01001353 raw_spin_lock(&ctx->lock);
1354 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001355 if (context_equiv(ctx, next_ctx)) {
1356 /*
1357 * XXX do we need a memory barrier of sorts
1358 * wrt to rcu_dereference() of perf_event_ctxp
1359 */
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02001360 task->perf_event_ctxp[ctxn] = next_ctx;
1361 next->perf_event_ctxp[ctxn] = ctx;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001362 ctx->task = next;
1363 next_ctx->task = task;
1364 do_switch = 0;
1365
1366 perf_event_sync_stat(ctx, next_ctx);
1367 }
Thomas Gleixnere625cce12009-11-17 18:02:06 +01001368 raw_spin_unlock(&next_ctx->lock);
1369 raw_spin_unlock(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001370 }
1371 rcu_read_unlock();
1372
1373 if (do_switch) {
Frederic Weisbecker5b0311e2010-01-17 11:59:13 +01001374 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001375 cpuctx->task_ctx = NULL;
1376 }
1377}
1378
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02001379#define for_each_task_context_nr(ctxn) \
1380 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
1381
1382/*
1383 * Called from scheduler to remove the events of the current task,
1384 * with interrupts disabled.
1385 *
1386 * We stop each event and update the event value in event->count.
1387 *
1388 * This does not protect us against NMI, but disable()
1389 * sets the disabled bit in the control field of event _before_
1390 * accessing the event control register. If a NMI hits, then it will
1391 * not restart the event.
1392 */
Peter Zijlstra82cd6de2010-10-14 17:57:23 +02001393void __perf_event_task_sched_out(struct task_struct *task,
1394 struct task_struct *next)
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02001395{
1396 int ctxn;
1397
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02001398 for_each_task_context_nr(ctxn)
1399 perf_event_context_sched_out(task, ctxn, next);
1400}
1401
Frederic Weisbecker5b0311e2010-01-17 11:59:13 +01001402static void task_ctx_sched_out(struct perf_event_context *ctx,
1403 enum event_type_t event_type)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001404{
Peter Zijlstra108b02c2010-09-06 14:32:03 +02001405 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001406
1407 if (!cpuctx->task_ctx)
1408 return;
1409
1410 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1411 return;
1412
Frederic Weisbecker5b0311e2010-01-17 11:59:13 +01001413 ctx_sched_out(ctx, cpuctx, event_type);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001414 cpuctx->task_ctx = NULL;
1415}
1416
1417/*
1418 * Called with IRQs disabled
1419 */
Frederic Weisbecker5b0311e2010-01-17 11:59:13 +01001420static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1421 enum event_type_t event_type)
1422{
1423 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001424}
1425
1426static void
Frederic Weisbecker5b0311e2010-01-17 11:59:13 +01001427ctx_pinned_sched_in(struct perf_event_context *ctx,
Peter Zijlstra6e377382010-02-11 13:21:58 +01001428 struct perf_cpu_context *cpuctx)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001429{
1430 struct perf_event *event;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001431
Frederic Weisbecker889ff012010-01-09 20:04:47 +01001432 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1433 if (event->state <= PERF_EVENT_STATE_OFF)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001434 continue;
Peter Zijlstra6e377382010-02-11 13:21:58 +01001435 if (event->cpu != -1 && event->cpu != smp_processor_id())
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001436 continue;
1437
Xiao Guangrong8c9ed8e2009-09-25 13:51:17 +08001438 if (group_can_go_on(event, cpuctx, 1))
Peter Zijlstra6e377382010-02-11 13:21:58 +01001439 group_sched_in(event, cpuctx, ctx);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001440
1441 /*
1442 * If this pinned group hasn't been scheduled,
1443 * put it in error state.
1444 */
1445 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1446 update_group_times(event);
1447 event->state = PERF_EVENT_STATE_ERROR;
1448 }
1449 }
Frederic Weisbecker5b0311e2010-01-17 11:59:13 +01001450}
1451
1452static void
1453ctx_flexible_sched_in(struct perf_event_context *ctx,
Peter Zijlstra6e377382010-02-11 13:21:58 +01001454 struct perf_cpu_context *cpuctx)
Frederic Weisbecker5b0311e2010-01-17 11:59:13 +01001455{
1456 struct perf_event *event;
1457 int can_add_hw = 1;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001458
Frederic Weisbecker889ff012010-01-09 20:04:47 +01001459 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1460 /* Ignore events in OFF or ERROR state */
1461 if (event->state <= PERF_EVENT_STATE_OFF)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001462 continue;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001463 /*
1464 * Listen to the 'cpu' scheduling filter constraint
1465 * of events:
1466 */
Peter Zijlstra6e377382010-02-11 13:21:58 +01001467 if (event->cpu != -1 && event->cpu != smp_processor_id())
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001468 continue;
1469
Peter Zijlstra9ed60602010-06-11 17:36:35 +02001470 if (group_can_go_on(event, cpuctx, can_add_hw)) {
Peter Zijlstra6e377382010-02-11 13:21:58 +01001471 if (group_sched_in(event, cpuctx, ctx))
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001472 can_add_hw = 0;
Peter Zijlstra9ed60602010-06-11 17:36:35 +02001473 }
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001474 }
Frederic Weisbecker5b0311e2010-01-17 11:59:13 +01001475}
1476
1477static void
1478ctx_sched_in(struct perf_event_context *ctx,
1479 struct perf_cpu_context *cpuctx,
1480 enum event_type_t event_type)
1481{
Frederic Weisbecker5b0311e2010-01-17 11:59:13 +01001482 raw_spin_lock(&ctx->lock);
1483 ctx->is_active = 1;
1484 if (likely(!ctx->nr_events))
1485 goto out;
1486
1487 ctx->timestamp = perf_clock();
1488
Frederic Weisbecker5b0311e2010-01-17 11:59:13 +01001489 /*
1490 * First go through the list and put on any pinned groups
1491 * in order to give them the best chance of going on.
1492 */
1493 if (event_type & EVENT_PINNED)
Peter Zijlstra6e377382010-02-11 13:21:58 +01001494 ctx_pinned_sched_in(ctx, cpuctx);
Frederic Weisbecker5b0311e2010-01-17 11:59:13 +01001495
1496 /* Then walk through the lower prio flexible groups */
1497 if (event_type & EVENT_FLEXIBLE)
Peter Zijlstra6e377382010-02-11 13:21:58 +01001498 ctx_flexible_sched_in(ctx, cpuctx);
Frederic Weisbecker5b0311e2010-01-17 11:59:13 +01001499
Peter Zijlstra9ed60602010-06-11 17:36:35 +02001500out:
Thomas Gleixnere625cce12009-11-17 18:02:06 +01001501 raw_spin_unlock(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001502}
1503
Frederic Weisbecker329c0e02010-01-17 12:56:05 +01001504static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1505 enum event_type_t event_type)
1506{
1507 struct perf_event_context *ctx = &cpuctx->ctx;
1508
1509 ctx_sched_in(ctx, cpuctx, event_type);
1510}
1511
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02001512static void task_ctx_sched_in(struct perf_event_context *ctx,
Frederic Weisbecker5b0311e2010-01-17 11:59:13 +01001513 enum event_type_t event_type)
1514{
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02001515 struct perf_cpu_context *cpuctx;
Frederic Weisbecker5b0311e2010-01-17 11:59:13 +01001516
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02001517 cpuctx = __get_cpu_context(ctx);
Frederic Weisbecker5b0311e2010-01-17 11:59:13 +01001518 if (cpuctx->task_ctx == ctx)
1519 return;
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02001520
Frederic Weisbecker5b0311e2010-01-17 11:59:13 +01001521 ctx_sched_in(ctx, cpuctx, event_type);
1522 cpuctx->task_ctx = ctx;
1523}
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001524
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02001525void perf_event_context_sched_in(struct perf_event_context *ctx)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001526{
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02001527 struct perf_cpu_context *cpuctx;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001528
Peter Zijlstra108b02c2010-09-06 14:32:03 +02001529 cpuctx = __get_cpu_context(ctx);
Frederic Weisbecker329c0e02010-01-17 12:56:05 +01001530 if (cpuctx->task_ctx == ctx)
1531 return;
1532
Peter Zijlstra1b9a6442010-09-07 18:32:22 +02001533 perf_pmu_disable(ctx->pmu);
Frederic Weisbecker329c0e02010-01-17 12:56:05 +01001534 /*
1535 * We want to keep the following priority order:
1536 * cpu pinned (that don't need to move), task pinned,
1537 * cpu flexible, task flexible.
1538 */
1539 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1540
1541 ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1542 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1543 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1544
1545 cpuctx->task_ctx = ctx;
eranian@google.com9b33fa62010-03-10 22:26:05 -08001546
Peter Zijlstrab5ab4cd2010-09-06 16:32:21 +02001547 /*
1548 * Since these rotations are per-cpu, we need to ensure the
1549 * cpu-context we got scheduled on is actually rotating.
1550 */
Peter Zijlstra108b02c2010-09-06 14:32:03 +02001551 perf_pmu_rotate_start(ctx->pmu);
Peter Zijlstra1b9a6442010-09-07 18:32:22 +02001552 perf_pmu_enable(ctx->pmu);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001553}
1554
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02001555/*
1556 * Called from scheduler to add the events of the current task
1557 * with interrupts disabled.
1558 *
1559 * We restore the event value and then enable it.
1560 *
1561 * This does not protect us against NMI, but enable()
1562 * sets the enabled bit in the control field of event _before_
1563 * accessing the event control register. If a NMI hits, then it will
1564 * keep the event running.
1565 */
Peter Zijlstra82cd6de2010-10-14 17:57:23 +02001566void __perf_event_task_sched_in(struct task_struct *task)
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02001567{
1568 struct perf_event_context *ctx;
1569 int ctxn;
1570
1571 for_each_task_context_nr(ctxn) {
1572 ctx = task->perf_event_ctxp[ctxn];
1573 if (likely(!ctx))
1574 continue;
1575
1576 perf_event_context_sched_in(ctx);
1577 }
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001578}
1579
1580#define MAX_INTERRUPTS (~0ULL)
1581
1582static void perf_log_throttle(struct perf_event *event, int enable);
1583
Peter Zijlstraabd50712010-01-26 18:50:16 +01001584static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1585{
1586 u64 frequency = event->attr.sample_freq;
1587 u64 sec = NSEC_PER_SEC;
1588 u64 divisor, dividend;
1589
1590 int count_fls, nsec_fls, frequency_fls, sec_fls;
1591
1592 count_fls = fls64(count);
1593 nsec_fls = fls64(nsec);
1594 frequency_fls = fls64(frequency);
1595 sec_fls = 30;
1596
1597 /*
1598 * We got @count in @nsec, with a target of sample_freq HZ
1599 * the target period becomes:
1600 *
1601 * @count * 10^9
1602 * period = -------------------
1603 * @nsec * sample_freq
1604 *
1605 */
1606
1607 /*
1608 * Reduce accuracy by one bit such that @a and @b converge
1609 * to a similar magnitude.
1610 */
1611#define REDUCE_FLS(a, b) \
1612do { \
1613 if (a##_fls > b##_fls) { \
1614 a >>= 1; \
1615 a##_fls--; \
1616 } else { \
1617 b >>= 1; \
1618 b##_fls--; \
1619 } \
1620} while (0)
1621
1622 /*
1623 * Reduce accuracy until either term fits in a u64, then proceed with
1624 * the other, so that finally we can do a u64/u64 division.
1625 */
1626 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1627 REDUCE_FLS(nsec, frequency);
1628 REDUCE_FLS(sec, count);
1629 }
1630
1631 if (count_fls + sec_fls > 64) {
1632 divisor = nsec * frequency;
1633
1634 while (count_fls + sec_fls > 64) {
1635 REDUCE_FLS(count, sec);
1636 divisor >>= 1;
1637 }
1638
1639 dividend = count * sec;
1640 } else {
1641 dividend = count * sec;
1642
1643 while (nsec_fls + frequency_fls > 64) {
1644 REDUCE_FLS(nsec, frequency);
1645 dividend >>= 1;
1646 }
1647
1648 divisor = nsec * frequency;
1649 }
1650
Peter Zijlstraf6ab91a2010-06-04 15:18:01 +02001651 if (!divisor)
1652 return dividend;
1653
Peter Zijlstraabd50712010-01-26 18:50:16 +01001654 return div64_u64(dividend, divisor);
1655}
1656
1657static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001658{
1659 struct hw_perf_event *hwc = &event->hw;
Peter Zijlstraf6ab91a2010-06-04 15:18:01 +02001660 s64 period, sample_period;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001661 s64 delta;
1662
Peter Zijlstraabd50712010-01-26 18:50:16 +01001663 period = perf_calculate_period(event, nsec, count);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001664
1665 delta = (s64)(period - hwc->sample_period);
1666 delta = (delta + 7) / 8; /* low pass filter */
1667
1668 sample_period = hwc->sample_period + delta;
1669
1670 if (!sample_period)
1671 sample_period = 1;
1672
1673 hwc->sample_period = sample_period;
Peter Zijlstraabd50712010-01-26 18:50:16 +01001674
Peter Zijlstrae7850592010-05-21 14:43:08 +02001675 if (local64_read(&hwc->period_left) > 8*sample_period) {
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02001676 event->pmu->stop(event, PERF_EF_UPDATE);
Peter Zijlstrae7850592010-05-21 14:43:08 +02001677 local64_set(&hwc->period_left, 0);
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02001678 event->pmu->start(event, PERF_EF_RELOAD);
Peter Zijlstraabd50712010-01-26 18:50:16 +01001679 }
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001680}
1681
Peter Zijlstrab5ab4cd2010-09-06 16:32:21 +02001682static void perf_ctx_adjust_freq(struct perf_event_context *ctx, u64 period)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001683{
1684 struct perf_event *event;
1685 struct hw_perf_event *hwc;
Peter Zijlstraabd50712010-01-26 18:50:16 +01001686 u64 interrupts, now;
1687 s64 delta;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001688
Thomas Gleixnere625cce12009-11-17 18:02:06 +01001689 raw_spin_lock(&ctx->lock);
Paul Mackerras03541f82009-10-14 16:58:03 +11001690 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001691 if (event->state != PERF_EVENT_STATE_ACTIVE)
1692 continue;
1693
Peter Zijlstra5d27c232009-12-17 13:16:32 +01001694 if (event->cpu != -1 && event->cpu != smp_processor_id())
1695 continue;
1696
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001697 hwc = &event->hw;
1698
1699 interrupts = hwc->interrupts;
1700 hwc->interrupts = 0;
1701
1702 /*
1703 * unthrottle events on the tick
1704 */
1705 if (interrupts == MAX_INTERRUPTS) {
1706 perf_log_throttle(event, 1);
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02001707 event->pmu->start(event, 0);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001708 }
1709
1710 if (!event->attr.freq || !event->attr.sample_freq)
1711 continue;
1712
Peter Zijlstraabd50712010-01-26 18:50:16 +01001713 event->pmu->read(event);
Peter Zijlstrae7850592010-05-21 14:43:08 +02001714 now = local64_read(&event->count);
Peter Zijlstraabd50712010-01-26 18:50:16 +01001715 delta = now - hwc->freq_count_stamp;
1716 hwc->freq_count_stamp = now;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001717
Peter Zijlstraabd50712010-01-26 18:50:16 +01001718 if (delta > 0)
Peter Zijlstrab5ab4cd2010-09-06 16:32:21 +02001719 perf_adjust_period(event, period, delta);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001720 }
Thomas Gleixnere625cce12009-11-17 18:02:06 +01001721 raw_spin_unlock(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001722}
1723
1724/*
1725 * Round-robin a context's events:
1726 */
1727static void rotate_ctx(struct perf_event_context *ctx)
1728{
Thomas Gleixnere625cce12009-11-17 18:02:06 +01001729 raw_spin_lock(&ctx->lock);
Frederic Weisbecker889ff012010-01-09 20:04:47 +01001730
Thomas Gleixnerdddd3372010-11-24 10:05:55 +01001731 /*
1732 * Rotate the first entry last of non-pinned groups. Rotation might be
1733 * disabled by the inheritance code.
1734 */
1735 if (!ctx->rotate_disable)
1736 list_rotate_left(&ctx->flexible_groups);
Frederic Weisbeckere2864172010-01-09 21:05:28 +01001737
Thomas Gleixnere625cce12009-11-17 18:02:06 +01001738 raw_spin_unlock(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001739}
1740
Peter Zijlstrab5ab4cd2010-09-06 16:32:21 +02001741/*
Peter Zijlstrae9d2b062010-09-17 11:28:50 +02001742 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
1743 * because they're strictly cpu affine and rotate_start is called with IRQs
1744 * disabled, while rotate_context is called from IRQ context.
Peter Zijlstrab5ab4cd2010-09-06 16:32:21 +02001745 */
Peter Zijlstrae9d2b062010-09-17 11:28:50 +02001746static void perf_rotate_context(struct perf_cpu_context *cpuctx)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001747{
Peter Zijlstrae9d2b062010-09-17 11:28:50 +02001748 u64 interval = (u64)cpuctx->jiffies_interval * TICK_NSEC;
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02001749 struct perf_event_context *ctx = NULL;
Peter Zijlstrae9d2b062010-09-17 11:28:50 +02001750 int rotate = 0, remove = 1;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001751
Peter Zijlstrab5ab4cd2010-09-06 16:32:21 +02001752 if (cpuctx->ctx.nr_events) {
Peter Zijlstrae9d2b062010-09-17 11:28:50 +02001753 remove = 0;
Peter Zijlstrab5ab4cd2010-09-06 16:32:21 +02001754 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1755 rotate = 1;
1756 }
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001757
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02001758 ctx = cpuctx->task_ctx;
Peter Zijlstrab5ab4cd2010-09-06 16:32:21 +02001759 if (ctx && ctx->nr_events) {
Peter Zijlstrae9d2b062010-09-17 11:28:50 +02001760 remove = 0;
Peter Zijlstrab5ab4cd2010-09-06 16:32:21 +02001761 if (ctx->nr_events != ctx->nr_active)
1762 rotate = 1;
1763 }
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001764
Peter Zijlstra1b9a6442010-09-07 18:32:22 +02001765 perf_pmu_disable(cpuctx->ctx.pmu);
Peter Zijlstrae9d2b062010-09-17 11:28:50 +02001766 perf_ctx_adjust_freq(&cpuctx->ctx, interval);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001767 if (ctx)
Peter Zijlstrae9d2b062010-09-17 11:28:50 +02001768 perf_ctx_adjust_freq(ctx, interval);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001769
Peter Zijlstrad4944a02010-03-08 13:51:20 +01001770 if (!rotate)
Peter Zijlstrab5ab4cd2010-09-06 16:32:21 +02001771 goto done;
Peter Zijlstrad4944a02010-03-08 13:51:20 +01001772
Frederic Weisbecker7defb0f2010-01-17 12:15:31 +01001773 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001774 if (ctx)
Frederic Weisbecker7defb0f2010-01-17 12:15:31 +01001775 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001776
1777 rotate_ctx(&cpuctx->ctx);
1778 if (ctx)
1779 rotate_ctx(ctx);
1780
Frederic Weisbecker7defb0f2010-01-17 12:15:31 +01001781 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001782 if (ctx)
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02001783 task_ctx_sched_in(ctx, EVENT_FLEXIBLE);
Peter Zijlstrab5ab4cd2010-09-06 16:32:21 +02001784
1785done:
Peter Zijlstrae9d2b062010-09-17 11:28:50 +02001786 if (remove)
1787 list_del_init(&cpuctx->rotation_list);
Peter Zijlstrab5ab4cd2010-09-06 16:32:21 +02001788
Peter Zijlstrae9d2b062010-09-17 11:28:50 +02001789 perf_pmu_enable(cpuctx->ctx.pmu);
1790}
1791
1792void perf_event_task_tick(void)
1793{
1794 struct list_head *head = &__get_cpu_var(rotation_list);
1795 struct perf_cpu_context *cpuctx, *tmp;
1796
1797 WARN_ON(!irqs_disabled());
1798
1799 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
1800 if (cpuctx->jiffies_interval == 1 ||
1801 !(jiffies % cpuctx->jiffies_interval))
1802 perf_rotate_context(cpuctx);
1803 }
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001804}
1805
Frederic Weisbecker889ff012010-01-09 20:04:47 +01001806static int event_enable_on_exec(struct perf_event *event,
1807 struct perf_event_context *ctx)
1808{
1809 if (!event->attr.enable_on_exec)
1810 return 0;
1811
1812 event->attr.enable_on_exec = 0;
1813 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1814 return 0;
1815
1816 __perf_event_mark_enabled(event, ctx);
1817
1818 return 1;
1819}
1820
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001821/*
1822 * Enable all of a task's events that have been marked enable-on-exec.
1823 * This expects task == current.
1824 */
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02001825static void perf_event_enable_on_exec(struct perf_event_context *ctx)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001826{
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001827 struct perf_event *event;
1828 unsigned long flags;
1829 int enabled = 0;
Frederic Weisbecker889ff012010-01-09 20:04:47 +01001830 int ret;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001831
1832 local_irq_save(flags);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001833 if (!ctx || !ctx->nr_events)
1834 goto out;
1835
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02001836 task_ctx_sched_out(ctx, EVENT_ALL);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001837
Thomas Gleixnere625cce12009-11-17 18:02:06 +01001838 raw_spin_lock(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001839
Frederic Weisbecker889ff012010-01-09 20:04:47 +01001840 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1841 ret = event_enable_on_exec(event, ctx);
1842 if (ret)
1843 enabled = 1;
1844 }
1845
1846 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1847 ret = event_enable_on_exec(event, ctx);
1848 if (ret)
1849 enabled = 1;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001850 }
1851
1852 /*
1853 * Unclone this context if we enabled any event.
1854 */
1855 if (enabled)
1856 unclone_ctx(ctx);
1857
Thomas Gleixnere625cce12009-11-17 18:02:06 +01001858 raw_spin_unlock(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001859
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02001860 perf_event_context_sched_in(ctx);
Peter Zijlstra9ed60602010-06-11 17:36:35 +02001861out:
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001862 local_irq_restore(flags);
1863}
1864
1865/*
1866 * Cross CPU call to read the hardware event
1867 */
1868static void __perf_event_read(void *info)
1869{
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001870 struct perf_event *event = info;
1871 struct perf_event_context *ctx = event->ctx;
Peter Zijlstra108b02c2010-09-06 14:32:03 +02001872 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001873
1874 /*
1875 * If this is a task context, we need to check whether it is
1876 * the current task context of this cpu. If not it has been
1877 * scheduled out before the smp call arrived. In that case
1878 * event->count would have been updated to a recent sample
1879 * when the event was scheduled out.
1880 */
1881 if (ctx->task && cpuctx->task_ctx != ctx)
1882 return;
1883
Thomas Gleixnere625cce12009-11-17 18:02:06 +01001884 raw_spin_lock(&ctx->lock);
Peter Zijlstra58e5ad12009-11-20 22:19:53 +01001885 update_context_time(ctx);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001886 update_event_times(event);
Thomas Gleixnere625cce12009-11-17 18:02:06 +01001887 raw_spin_unlock(&ctx->lock);
Peter Zijlstra2b8988c2009-11-20 22:19:54 +01001888
Peter Zijlstra58e5ad12009-11-20 22:19:53 +01001889 event->pmu->read(event);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001890}
1891
Peter Zijlstrab5e58792010-05-21 14:43:12 +02001892static inline u64 perf_event_count(struct perf_event *event)
1893{
Peter Zijlstrae7850592010-05-21 14:43:08 +02001894 return local64_read(&event->count) + atomic64_read(&event->child_count);
Peter Zijlstrab5e58792010-05-21 14:43:12 +02001895}
1896
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001897static u64 perf_event_read(struct perf_event *event)
1898{
1899 /*
1900 * If event is enabled and currently active on a CPU, update the
1901 * value in the event structure:
1902 */
1903 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1904 smp_call_function_single(event->oncpu,
1905 __perf_event_read, event, 1);
1906 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
Peter Zijlstra2b8988c2009-11-20 22:19:54 +01001907 struct perf_event_context *ctx = event->ctx;
1908 unsigned long flags;
1909
Thomas Gleixnere625cce12009-11-17 18:02:06 +01001910 raw_spin_lock_irqsave(&ctx->lock, flags);
Stephane Eranianc530ccd2010-10-15 15:26:01 +02001911 /*
1912 * may read while context is not active
1913 * (e.g., thread is blocked), in that case
1914 * we cannot update context time
1915 */
1916 if (ctx->is_active)
1917 update_context_time(ctx);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001918 update_event_times(event);
Thomas Gleixnere625cce12009-11-17 18:02:06 +01001919 raw_spin_unlock_irqrestore(&ctx->lock, flags);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001920 }
1921
Peter Zijlstrab5e58792010-05-21 14:43:12 +02001922 return perf_event_count(event);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02001923}
1924
1925/*
Frederic Weisbecker927c7a92010-07-01 16:20:36 +02001926 * Callchain support
1927 */
1928
1929struct callchain_cpus_entries {
1930 struct rcu_head rcu_head;
1931 struct perf_callchain_entry *cpu_entries[0];
1932};
1933
Frederic Weisbecker7ae07ea2010-08-14 20:45:13 +02001934static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
Frederic Weisbecker927c7a92010-07-01 16:20:36 +02001935static atomic_t nr_callchain_events;
1936static DEFINE_MUTEX(callchain_mutex);
1937struct callchain_cpus_entries *callchain_cpus_entries;
1938
1939
1940__weak void perf_callchain_kernel(struct perf_callchain_entry *entry,
1941 struct pt_regs *regs)
1942{
1943}
1944
1945__weak void perf_callchain_user(struct perf_callchain_entry *entry,
1946 struct pt_regs *regs)
1947{
1948}
1949
1950static void release_callchain_buffers_rcu(struct rcu_head *head)
1951{
1952 struct callchain_cpus_entries *entries;
1953 int cpu;
1954
1955 entries = container_of(head, struct callchain_cpus_entries, rcu_head);
1956
1957 for_each_possible_cpu(cpu)
1958 kfree(entries->cpu_entries[cpu]);
1959
1960 kfree(entries);
1961}
1962
1963static void release_callchain_buffers(void)
1964{
1965 struct callchain_cpus_entries *entries;
1966
1967 entries = callchain_cpus_entries;
1968 rcu_assign_pointer(callchain_cpus_entries, NULL);
1969 call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
1970}
1971
1972static int alloc_callchain_buffers(void)
1973{
1974 int cpu;
1975 int size;
1976 struct callchain_cpus_entries *entries;
1977
1978 /*
1979 * We can't use the percpu allocation API for data that can be
1980 * accessed from NMI. Use a temporary manual per cpu allocation
1981 * until that gets sorted out.
1982 */
1983 size = sizeof(*entries) + sizeof(struct perf_callchain_entry *) *
1984 num_possible_cpus();
1985
1986 entries = kzalloc(size, GFP_KERNEL);
1987 if (!entries)
1988 return -ENOMEM;
1989
Frederic Weisbecker7ae07ea2010-08-14 20:45:13 +02001990 size = sizeof(struct perf_callchain_entry) * PERF_NR_CONTEXTS;
Frederic Weisbecker927c7a92010-07-01 16:20:36 +02001991
1992 for_each_possible_cpu(cpu) {
1993 entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
1994 cpu_to_node(cpu));
1995 if (!entries->cpu_entries[cpu])
1996 goto fail;
1997 }
1998
1999 rcu_assign_pointer(callchain_cpus_entries, entries);
2000
2001 return 0;
2002
2003fail:
2004 for_each_possible_cpu(cpu)
2005 kfree(entries->cpu_entries[cpu]);
2006 kfree(entries);
2007
2008 return -ENOMEM;
2009}
2010
2011static int get_callchain_buffers(void)
2012{
2013 int err = 0;
2014 int count;
2015
2016 mutex_lock(&callchain_mutex);
2017
2018 count = atomic_inc_return(&nr_callchain_events);
2019 if (WARN_ON_ONCE(count < 1)) {
2020 err = -EINVAL;
2021 goto exit;
2022 }
2023
2024 if (count > 1) {
2025 /* If the allocation failed, give up */
2026 if (!callchain_cpus_entries)
2027 err = -ENOMEM;
2028 goto exit;
2029 }
2030
2031 err = alloc_callchain_buffers();
2032 if (err)
2033 release_callchain_buffers();
2034exit:
2035 mutex_unlock(&callchain_mutex);
2036
2037 return err;
2038}
2039
2040static void put_callchain_buffers(void)
2041{
2042 if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
2043 release_callchain_buffers();
2044 mutex_unlock(&callchain_mutex);
2045 }
2046}
2047
2048static int get_recursion_context(int *recursion)
2049{
2050 int rctx;
2051
2052 if (in_nmi())
2053 rctx = 3;
2054 else if (in_irq())
2055 rctx = 2;
2056 else if (in_softirq())
2057 rctx = 1;
2058 else
2059 rctx = 0;
2060
2061 if (recursion[rctx])
2062 return -1;
2063
2064 recursion[rctx]++;
2065 barrier();
2066
2067 return rctx;
2068}
2069
2070static inline void put_recursion_context(int *recursion, int rctx)
2071{
2072 barrier();
2073 recursion[rctx]--;
2074}
2075
2076static struct perf_callchain_entry *get_callchain_entry(int *rctx)
2077{
2078 int cpu;
2079 struct callchain_cpus_entries *entries;
2080
2081 *rctx = get_recursion_context(__get_cpu_var(callchain_recursion));
2082 if (*rctx == -1)
2083 return NULL;
2084
2085 entries = rcu_dereference(callchain_cpus_entries);
2086 if (!entries)
2087 return NULL;
2088
2089 cpu = smp_processor_id();
2090
2091 return &entries->cpu_entries[cpu][*rctx];
2092}
2093
2094static void
2095put_callchain_entry(int rctx)
2096{
2097 put_recursion_context(__get_cpu_var(callchain_recursion), rctx);
2098}
2099
2100static struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2101{
2102 int rctx;
2103 struct perf_callchain_entry *entry;
2104
2105
2106 entry = get_callchain_entry(&rctx);
2107 if (rctx == -1)
2108 return NULL;
2109
2110 if (!entry)
2111 goto exit_put;
2112
2113 entry->nr = 0;
2114
2115 if (!user_mode(regs)) {
2116 perf_callchain_store(entry, PERF_CONTEXT_KERNEL);
2117 perf_callchain_kernel(entry, regs);
2118 if (current->mm)
2119 regs = task_pt_regs(current);
2120 else
2121 regs = NULL;
2122 }
2123
2124 if (regs) {
2125 perf_callchain_store(entry, PERF_CONTEXT_USER);
2126 perf_callchain_user(entry, regs);
2127 }
2128
2129exit_put:
2130 put_callchain_entry(rctx);
2131
2132 return entry;
2133}
2134
2135/*
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002136 * Initialize the perf_event context in a task_struct:
2137 */
Peter Zijlstraeb184472010-09-07 15:55:13 +02002138static void __perf_event_init_context(struct perf_event_context *ctx)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002139{
Thomas Gleixnere625cce12009-11-17 18:02:06 +01002140 raw_spin_lock_init(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002141 mutex_init(&ctx->mutex);
Frederic Weisbecker889ff012010-01-09 20:04:47 +01002142 INIT_LIST_HEAD(&ctx->pinned_groups);
2143 INIT_LIST_HEAD(&ctx->flexible_groups);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002144 INIT_LIST_HEAD(&ctx->event_list);
2145 atomic_set(&ctx->refcount, 1);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002146}
2147
Peter Zijlstraeb184472010-09-07 15:55:13 +02002148static struct perf_event_context *
2149alloc_perf_context(struct pmu *pmu, struct task_struct *task)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002150{
2151 struct perf_event_context *ctx;
Peter Zijlstraeb184472010-09-07 15:55:13 +02002152
2153 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2154 if (!ctx)
2155 return NULL;
2156
2157 __perf_event_init_context(ctx);
2158 if (task) {
2159 ctx->task = task;
2160 get_task_struct(task);
2161 }
2162 ctx->pmu = pmu;
2163
2164 return ctx;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002165}
2166
Matt Helsley2ebd4ff2010-09-13 13:01:19 -07002167static struct task_struct *
2168find_lively_task_by_vpid(pid_t vpid)
2169{
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002170 struct task_struct *task;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002171 int err;
2172
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002173 rcu_read_lock();
Matt Helsley2ebd4ff2010-09-13 13:01:19 -07002174 if (!vpid)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002175 task = current;
2176 else
Matt Helsley2ebd4ff2010-09-13 13:01:19 -07002177 task = find_task_by_vpid(vpid);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002178 if (task)
2179 get_task_struct(task);
2180 rcu_read_unlock();
2181
2182 if (!task)
2183 return ERR_PTR(-ESRCH);
2184
2185 /*
2186 * Can't attach events to a dying task.
2187 */
2188 err = -ESRCH;
2189 if (task->flags & PF_EXITING)
2190 goto errout;
2191
2192 /* Reuse ptrace permission checks for now. */
2193 err = -EACCES;
2194 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2195 goto errout;
2196
Matt Helsley2ebd4ff2010-09-13 13:01:19 -07002197 return task;
2198errout:
2199 put_task_struct(task);
2200 return ERR_PTR(err);
2201
2202}
2203
Peter Zijlstra108b02c2010-09-06 14:32:03 +02002204static struct perf_event_context *
Matt Helsley38a81da2010-09-13 13:01:20 -07002205find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002206{
2207 struct perf_event_context *ctx;
2208 struct perf_cpu_context *cpuctx;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002209 unsigned long flags;
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02002210 int ctxn, err;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002211
Matt Helsley38a81da2010-09-13 13:01:20 -07002212 if (!task && cpu != -1) {
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002213 /* Must be root to operate on a CPU event: */
2214 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2215 return ERR_PTR(-EACCES);
2216
2217 if (cpu < 0 || cpu >= nr_cpumask_bits)
2218 return ERR_PTR(-EINVAL);
2219
2220 /*
2221 * We could be clever and allow to attach a event to an
2222 * offline CPU and activate it when the CPU comes up, but
2223 * that's for later.
2224 */
2225 if (!cpu_online(cpu))
2226 return ERR_PTR(-ENODEV);
2227
Peter Zijlstra108b02c2010-09-06 14:32:03 +02002228 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002229 ctx = &cpuctx->ctx;
2230 get_ctx(ctx);
2231
2232 return ctx;
2233 }
2234
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02002235 err = -EINVAL;
2236 ctxn = pmu->task_ctx_nr;
2237 if (ctxn < 0)
2238 goto errout;
2239
Peter Zijlstra9ed60602010-06-11 17:36:35 +02002240retry:
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02002241 ctx = perf_lock_task_context(task, ctxn, &flags);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002242 if (ctx) {
2243 unclone_ctx(ctx);
Thomas Gleixnere625cce12009-11-17 18:02:06 +01002244 raw_spin_unlock_irqrestore(&ctx->lock, flags);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002245 }
2246
2247 if (!ctx) {
Peter Zijlstraeb184472010-09-07 15:55:13 +02002248 ctx = alloc_perf_context(pmu, task);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002249 err = -ENOMEM;
2250 if (!ctx)
2251 goto errout;
Peter Zijlstraeb184472010-09-07 15:55:13 +02002252
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002253 get_ctx(ctx);
Peter Zijlstraeb184472010-09-07 15:55:13 +02002254
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02002255 if (cmpxchg(&task->perf_event_ctxp[ctxn], NULL, ctx)) {
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002256 /*
2257 * We raced with some other task; use
2258 * the context they set.
2259 */
Peter Zijlstraeb184472010-09-07 15:55:13 +02002260 put_task_struct(task);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002261 kfree(ctx);
2262 goto retry;
2263 }
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002264 }
2265
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002266 return ctx;
2267
Peter Zijlstra9ed60602010-06-11 17:36:35 +02002268errout:
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002269 return ERR_PTR(err);
2270}
2271
Li Zefan6fb29152009-10-15 11:21:42 +08002272static void perf_event_free_filter(struct perf_event *event);
2273
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002274static void free_event_rcu(struct rcu_head *head)
2275{
2276 struct perf_event *event;
2277
2278 event = container_of(head, struct perf_event, rcu_head);
2279 if (event->ns)
2280 put_pid_ns(event->ns);
Li Zefan6fb29152009-10-15 11:21:42 +08002281 perf_event_free_filter(event);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002282 kfree(event);
2283}
2284
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002285static void perf_buffer_put(struct perf_buffer *buffer);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002286
2287static void free_event(struct perf_event *event)
2288{
Peter Zijlstrae360adb2010-10-14 14:01:34 +08002289 irq_work_sync(&event->pending);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002290
2291 if (!event->parent) {
Peter Zijlstra82cd6de2010-10-14 17:57:23 +02002292 if (event->attach_state & PERF_ATTACH_TASK)
2293 jump_label_dec(&perf_task_events);
Eric B Munson3af9e852010-05-18 15:30:49 +01002294 if (event->attr.mmap || event->attr.mmap_data)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002295 atomic_dec(&nr_mmap_events);
2296 if (event->attr.comm)
2297 atomic_dec(&nr_comm_events);
2298 if (event->attr.task)
2299 atomic_dec(&nr_task_events);
Frederic Weisbecker927c7a92010-07-01 16:20:36 +02002300 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2301 put_callchain_buffers();
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002302 }
2303
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002304 if (event->buffer) {
2305 perf_buffer_put(event->buffer);
2306 event->buffer = NULL;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002307 }
2308
2309 if (event->destroy)
2310 event->destroy(event);
2311
Peter Zijlstra0c67b402010-09-13 11:15:58 +02002312 if (event->ctx)
2313 put_ctx(event->ctx);
2314
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002315 call_rcu(&event->rcu_head, free_event_rcu);
2316}
2317
Arjan van de Venfb0459d2009-09-25 12:25:56 +02002318int perf_event_release_kernel(struct perf_event *event)
2319{
2320 struct perf_event_context *ctx = event->ctx;
2321
Peter Zijlstra050735b2010-05-11 11:51:53 +02002322 /*
2323 * Remove from the PMU, can't get re-enabled since we got
2324 * here because the last ref went.
2325 */
2326 perf_event_disable(event);
2327
Arjan van de Venfb0459d2009-09-25 12:25:56 +02002328 WARN_ON_ONCE(ctx->parent_ctx);
Peter Zijlstraa0507c82010-05-06 15:42:53 +02002329 /*
2330 * There are two ways this annotation is useful:
2331 *
2332 * 1) there is a lock recursion from perf_event_exit_task
2333 * see the comment there.
2334 *
2335 * 2) there is a lock-inversion with mmap_sem through
2336 * perf_event_read_group(), which takes faults while
2337 * holding ctx->mutex, however this is called after
2338 * the last filedesc died, so there is no possibility
2339 * to trigger the AB-BA case.
2340 */
2341 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
Peter Zijlstra050735b2010-05-11 11:51:53 +02002342 raw_spin_lock_irq(&ctx->lock);
Peter Zijlstra8a495422010-05-27 15:47:49 +02002343 perf_group_detach(event);
Peter Zijlstra050735b2010-05-11 11:51:53 +02002344 list_del_event(event, ctx);
Peter Zijlstra050735b2010-05-11 11:51:53 +02002345 raw_spin_unlock_irq(&ctx->lock);
Arjan van de Venfb0459d2009-09-25 12:25:56 +02002346 mutex_unlock(&ctx->mutex);
2347
Arjan van de Venfb0459d2009-09-25 12:25:56 +02002348 free_event(event);
2349
2350 return 0;
2351}
2352EXPORT_SYMBOL_GPL(perf_event_release_kernel);
2353
Peter Zijlstraa66a3052009-11-23 11:37:23 +01002354/*
2355 * Called when the last reference to the file is gone.
2356 */
2357static int perf_release(struct inode *inode, struct file *file)
2358{
2359 struct perf_event *event = file->private_data;
Peter Zijlstra88821352010-11-09 19:01:43 +01002360 struct task_struct *owner;
Peter Zijlstraa66a3052009-11-23 11:37:23 +01002361
2362 file->private_data = NULL;
2363
Peter Zijlstra88821352010-11-09 19:01:43 +01002364 rcu_read_lock();
2365 owner = ACCESS_ONCE(event->owner);
2366 /*
2367 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2368 * !owner it means the list deletion is complete and we can indeed
2369 * free this event, otherwise we need to serialize on
2370 * owner->perf_event_mutex.
2371 */
2372 smp_read_barrier_depends();
2373 if (owner) {
2374 /*
2375 * Since delayed_put_task_struct() also drops the last
2376 * task reference we can safely take a new reference
2377 * while holding the rcu_read_lock().
2378 */
2379 get_task_struct(owner);
2380 }
2381 rcu_read_unlock();
2382
2383 if (owner) {
2384 mutex_lock(&owner->perf_event_mutex);
2385 /*
2386 * We have to re-check the event->owner field, if it is cleared
2387 * we raced with perf_event_exit_task(), acquiring the mutex
2388 * ensured they're done, and we can proceed with freeing the
2389 * event.
2390 */
2391 if (event->owner)
2392 list_del_init(&event->owner_entry);
2393 mutex_unlock(&owner->perf_event_mutex);
2394 put_task_struct(owner);
2395 }
2396
Peter Zijlstraa66a3052009-11-23 11:37:23 +01002397 return perf_event_release_kernel(event);
2398}
2399
Peter Zijlstra59ed4462009-11-20 22:19:55 +01002400u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002401{
2402 struct perf_event *child;
2403 u64 total = 0;
2404
Peter Zijlstra59ed4462009-11-20 22:19:55 +01002405 *enabled = 0;
2406 *running = 0;
2407
Peter Zijlstra6f105812009-11-20 22:19:56 +01002408 mutex_lock(&event->child_mutex);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002409 total += perf_event_read(event);
Peter Zijlstra59ed4462009-11-20 22:19:55 +01002410 *enabled += event->total_time_enabled +
2411 atomic64_read(&event->child_total_time_enabled);
2412 *running += event->total_time_running +
2413 atomic64_read(&event->child_total_time_running);
2414
2415 list_for_each_entry(child, &event->child_list, child_list) {
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002416 total += perf_event_read(child);
Peter Zijlstra59ed4462009-11-20 22:19:55 +01002417 *enabled += child->total_time_enabled;
2418 *running += child->total_time_running;
2419 }
Peter Zijlstra6f105812009-11-20 22:19:56 +01002420 mutex_unlock(&event->child_mutex);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002421
2422 return total;
2423}
Arjan van de Venfb0459d2009-09-25 12:25:56 +02002424EXPORT_SYMBOL_GPL(perf_event_read_value);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002425
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002426static int perf_event_read_group(struct perf_event *event,
2427 u64 read_format, char __user *buf)
2428{
2429 struct perf_event *leader = event->group_leader, *sub;
Peter Zijlstra6f105812009-11-20 22:19:56 +01002430 int n = 0, size = 0, ret = -EFAULT;
2431 struct perf_event_context *ctx = leader->ctx;
Peter Zijlstraabf48682009-11-20 22:19:49 +01002432 u64 values[5];
Peter Zijlstra59ed4462009-11-20 22:19:55 +01002433 u64 count, enabled, running;
Peter Zijlstraabf48682009-11-20 22:19:49 +01002434
Peter Zijlstra6f105812009-11-20 22:19:56 +01002435 mutex_lock(&ctx->mutex);
Peter Zijlstra59ed4462009-11-20 22:19:55 +01002436 count = perf_event_read_value(leader, &enabled, &running);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002437
2438 values[n++] = 1 + leader->nr_siblings;
Peter Zijlstra59ed4462009-11-20 22:19:55 +01002439 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2440 values[n++] = enabled;
2441 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2442 values[n++] = running;
Peter Zijlstraabf48682009-11-20 22:19:49 +01002443 values[n++] = count;
2444 if (read_format & PERF_FORMAT_ID)
2445 values[n++] = primary_event_id(leader);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002446
2447 size = n * sizeof(u64);
2448
2449 if (copy_to_user(buf, values, size))
Peter Zijlstra6f105812009-11-20 22:19:56 +01002450 goto unlock;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002451
Peter Zijlstra6f105812009-11-20 22:19:56 +01002452 ret = size;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002453
2454 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
Peter Zijlstraabf48682009-11-20 22:19:49 +01002455 n = 0;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002456
Peter Zijlstra59ed4462009-11-20 22:19:55 +01002457 values[n++] = perf_event_read_value(sub, &enabled, &running);
Peter Zijlstraabf48682009-11-20 22:19:49 +01002458 if (read_format & PERF_FORMAT_ID)
2459 values[n++] = primary_event_id(sub);
2460
2461 size = n * sizeof(u64);
2462
Stephane Eranian184d3da2009-11-23 21:40:49 -08002463 if (copy_to_user(buf + ret, values, size)) {
Peter Zijlstra6f105812009-11-20 22:19:56 +01002464 ret = -EFAULT;
2465 goto unlock;
2466 }
Peter Zijlstraabf48682009-11-20 22:19:49 +01002467
2468 ret += size;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002469 }
Peter Zijlstra6f105812009-11-20 22:19:56 +01002470unlock:
2471 mutex_unlock(&ctx->mutex);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002472
Peter Zijlstraabf48682009-11-20 22:19:49 +01002473 return ret;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002474}
2475
2476static int perf_event_read_one(struct perf_event *event,
2477 u64 read_format, char __user *buf)
2478{
Peter Zijlstra59ed4462009-11-20 22:19:55 +01002479 u64 enabled, running;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002480 u64 values[4];
2481 int n = 0;
2482
Peter Zijlstra59ed4462009-11-20 22:19:55 +01002483 values[n++] = perf_event_read_value(event, &enabled, &running);
2484 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2485 values[n++] = enabled;
2486 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2487 values[n++] = running;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002488 if (read_format & PERF_FORMAT_ID)
2489 values[n++] = primary_event_id(event);
2490
2491 if (copy_to_user(buf, values, n * sizeof(u64)))
2492 return -EFAULT;
2493
2494 return n * sizeof(u64);
2495}
2496
2497/*
2498 * Read the performance event - simple non blocking version for now
2499 */
2500static ssize_t
2501perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2502{
2503 u64 read_format = event->attr.read_format;
2504 int ret;
2505
2506 /*
2507 * Return end-of-file for a read on a event that is in
2508 * error state (i.e. because it was pinned but it couldn't be
2509 * scheduled on to the CPU at some point).
2510 */
2511 if (event->state == PERF_EVENT_STATE_ERROR)
2512 return 0;
2513
Arnaldo Carvalho de Meloc320c7b2010-10-20 12:50:11 -02002514 if (count < event->read_size)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002515 return -ENOSPC;
2516
2517 WARN_ON_ONCE(event->ctx->parent_ctx);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002518 if (read_format & PERF_FORMAT_GROUP)
2519 ret = perf_event_read_group(event, read_format, buf);
2520 else
2521 ret = perf_event_read_one(event, read_format, buf);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002522
2523 return ret;
2524}
2525
2526static ssize_t
2527perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2528{
2529 struct perf_event *event = file->private_data;
2530
2531 return perf_read_hw(event, buf, count);
2532}
2533
2534static unsigned int perf_poll(struct file *file, poll_table *wait)
2535{
2536 struct perf_event *event = file->private_data;
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002537 struct perf_buffer *buffer;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002538 unsigned int events = POLL_HUP;
2539
2540 rcu_read_lock();
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002541 buffer = rcu_dereference(event->buffer);
2542 if (buffer)
2543 events = atomic_xchg(&buffer->poll, 0);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002544 rcu_read_unlock();
2545
2546 poll_wait(file, &event->waitq, wait);
2547
2548 return events;
2549}
2550
2551static void perf_event_reset(struct perf_event *event)
2552{
2553 (void)perf_event_read(event);
Peter Zijlstrae7850592010-05-21 14:43:08 +02002554 local64_set(&event->count, 0);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002555 perf_event_update_userpage(event);
2556}
2557
2558/*
2559 * Holding the top-level event's child_mutex means that any
2560 * descendant process that has inherited this event will block
2561 * in sync_child_event if it goes to exit, thus satisfying the
2562 * task existence requirements of perf_event_enable/disable.
2563 */
2564static void perf_event_for_each_child(struct perf_event *event,
2565 void (*func)(struct perf_event *))
2566{
2567 struct perf_event *child;
2568
2569 WARN_ON_ONCE(event->ctx->parent_ctx);
2570 mutex_lock(&event->child_mutex);
2571 func(event);
2572 list_for_each_entry(child, &event->child_list, child_list)
2573 func(child);
2574 mutex_unlock(&event->child_mutex);
2575}
2576
2577static void perf_event_for_each(struct perf_event *event,
2578 void (*func)(struct perf_event *))
2579{
2580 struct perf_event_context *ctx = event->ctx;
2581 struct perf_event *sibling;
2582
2583 WARN_ON_ONCE(ctx->parent_ctx);
2584 mutex_lock(&ctx->mutex);
2585 event = event->group_leader;
2586
2587 perf_event_for_each_child(event, func);
2588 func(event);
2589 list_for_each_entry(sibling, &event->sibling_list, group_entry)
2590 perf_event_for_each_child(event, func);
2591 mutex_unlock(&ctx->mutex);
2592}
2593
2594static int perf_event_period(struct perf_event *event, u64 __user *arg)
2595{
2596 struct perf_event_context *ctx = event->ctx;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002597 int ret = 0;
2598 u64 value;
2599
Franck Bui-Huu6c7e5502010-11-23 16:21:43 +01002600 if (!is_sampling_event(event))
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002601 return -EINVAL;
2602
John Blackwoodad0cf342010-09-28 18:03:11 -04002603 if (copy_from_user(&value, arg, sizeof(value)))
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002604 return -EFAULT;
2605
2606 if (!value)
2607 return -EINVAL;
2608
Thomas Gleixnere625cce12009-11-17 18:02:06 +01002609 raw_spin_lock_irq(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002610 if (event->attr.freq) {
2611 if (value > sysctl_perf_event_sample_rate) {
2612 ret = -EINVAL;
2613 goto unlock;
2614 }
2615
2616 event->attr.sample_freq = value;
2617 } else {
2618 event->attr.sample_period = value;
2619 event->hw.sample_period = value;
2620 }
2621unlock:
Thomas Gleixnere625cce12009-11-17 18:02:06 +01002622 raw_spin_unlock_irq(&ctx->lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002623
2624 return ret;
2625}
2626
Peter Zijlstraac9721f2010-05-27 12:54:41 +02002627static const struct file_operations perf_fops;
2628
2629static struct perf_event *perf_fget_light(int fd, int *fput_needed)
2630{
2631 struct file *file;
2632
2633 file = fget_light(fd, fput_needed);
2634 if (!file)
2635 return ERR_PTR(-EBADF);
2636
2637 if (file->f_op != &perf_fops) {
2638 fput_light(file, *fput_needed);
2639 *fput_needed = 0;
2640 return ERR_PTR(-EBADF);
2641 }
2642
2643 return file->private_data;
2644}
2645
2646static int perf_event_set_output(struct perf_event *event,
2647 struct perf_event *output_event);
Li Zefan6fb29152009-10-15 11:21:42 +08002648static int perf_event_set_filter(struct perf_event *event, void __user *arg);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002649
2650static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2651{
2652 struct perf_event *event = file->private_data;
2653 void (*func)(struct perf_event *);
2654 u32 flags = arg;
2655
2656 switch (cmd) {
2657 case PERF_EVENT_IOC_ENABLE:
2658 func = perf_event_enable;
2659 break;
2660 case PERF_EVENT_IOC_DISABLE:
2661 func = perf_event_disable;
2662 break;
2663 case PERF_EVENT_IOC_RESET:
2664 func = perf_event_reset;
2665 break;
2666
2667 case PERF_EVENT_IOC_REFRESH:
2668 return perf_event_refresh(event, arg);
2669
2670 case PERF_EVENT_IOC_PERIOD:
2671 return perf_event_period(event, (u64 __user *)arg);
2672
2673 case PERF_EVENT_IOC_SET_OUTPUT:
Peter Zijlstraac9721f2010-05-27 12:54:41 +02002674 {
2675 struct perf_event *output_event = NULL;
2676 int fput_needed = 0;
2677 int ret;
2678
2679 if (arg != -1) {
2680 output_event = perf_fget_light(arg, &fput_needed);
2681 if (IS_ERR(output_event))
2682 return PTR_ERR(output_event);
2683 }
2684
2685 ret = perf_event_set_output(event, output_event);
2686 if (output_event)
2687 fput_light(output_event->filp, fput_needed);
2688
2689 return ret;
2690 }
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002691
Li Zefan6fb29152009-10-15 11:21:42 +08002692 case PERF_EVENT_IOC_SET_FILTER:
2693 return perf_event_set_filter(event, (void __user *)arg);
2694
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002695 default:
2696 return -ENOTTY;
2697 }
2698
2699 if (flags & PERF_IOC_FLAG_GROUP)
2700 perf_event_for_each(event, func);
2701 else
2702 perf_event_for_each_child(event, func);
2703
2704 return 0;
2705}
2706
2707int perf_event_task_enable(void)
2708{
2709 struct perf_event *event;
2710
2711 mutex_lock(&current->perf_event_mutex);
2712 list_for_each_entry(event, &current->perf_event_list, owner_entry)
2713 perf_event_for_each_child(event, perf_event_enable);
2714 mutex_unlock(&current->perf_event_mutex);
2715
2716 return 0;
2717}
2718
2719int perf_event_task_disable(void)
2720{
2721 struct perf_event *event;
2722
2723 mutex_lock(&current->perf_event_mutex);
2724 list_for_each_entry(event, &current->perf_event_list, owner_entry)
2725 perf_event_for_each_child(event, perf_event_disable);
2726 mutex_unlock(&current->perf_event_mutex);
2727
2728 return 0;
2729}
2730
2731#ifndef PERF_EVENT_INDEX_OFFSET
2732# define PERF_EVENT_INDEX_OFFSET 0
2733#endif
2734
2735static int perf_event_index(struct perf_event *event)
2736{
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02002737 if (event->hw.state & PERF_HES_STOPPED)
2738 return 0;
2739
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002740 if (event->state != PERF_EVENT_STATE_ACTIVE)
2741 return 0;
2742
2743 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2744}
2745
2746/*
2747 * Callers need to ensure there can be no nesting of this function, otherwise
2748 * the seqlock logic goes bad. We can not serialize this because the arch
2749 * code calls this from NMI context.
2750 */
2751void perf_event_update_userpage(struct perf_event *event)
2752{
2753 struct perf_event_mmap_page *userpg;
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002754 struct perf_buffer *buffer;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002755
2756 rcu_read_lock();
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002757 buffer = rcu_dereference(event->buffer);
2758 if (!buffer)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002759 goto unlock;
2760
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002761 userpg = buffer->user_page;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002762
2763 /*
2764 * Disable preemption so as to not let the corresponding user-space
2765 * spin too long if we get preempted.
2766 */
2767 preempt_disable();
2768 ++userpg->lock;
2769 barrier();
2770 userpg->index = perf_event_index(event);
Peter Zijlstrab5e58792010-05-21 14:43:12 +02002771 userpg->offset = perf_event_count(event);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002772 if (event->state == PERF_EVENT_STATE_ACTIVE)
Peter Zijlstrae7850592010-05-21 14:43:08 +02002773 userpg->offset -= local64_read(&event->hw.prev_count);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002774
2775 userpg->time_enabled = event->total_time_enabled +
2776 atomic64_read(&event->child_total_time_enabled);
2777
2778 userpg->time_running = event->total_time_running +
2779 atomic64_read(&event->child_total_time_running);
2780
2781 barrier();
2782 ++userpg->lock;
2783 preempt_enable();
2784unlock:
2785 rcu_read_unlock();
2786}
2787
Peter Zijlstrad57e34f2010-05-28 19:41:35 +02002788static unsigned long perf_data_size(struct perf_buffer *buffer);
2789
2790static void
2791perf_buffer_init(struct perf_buffer *buffer, long watermark, int flags)
2792{
2793 long max_size = perf_data_size(buffer);
2794
2795 if (watermark)
2796 buffer->watermark = min(max_size, watermark);
2797
2798 if (!buffer->watermark)
2799 buffer->watermark = max_size / 2;
2800
2801 if (flags & PERF_BUFFER_WRITABLE)
2802 buffer->writable = 1;
2803
2804 atomic_set(&buffer->refcount, 1);
2805}
2806
Peter Zijlstra906010b2009-09-21 16:08:49 +02002807#ifndef CONFIG_PERF_USE_VMALLOC
2808
2809/*
2810 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2811 */
2812
2813static struct page *
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002814perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
Peter Zijlstra906010b2009-09-21 16:08:49 +02002815{
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002816 if (pgoff > buffer->nr_pages)
Peter Zijlstra906010b2009-09-21 16:08:49 +02002817 return NULL;
2818
2819 if (pgoff == 0)
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002820 return virt_to_page(buffer->user_page);
Peter Zijlstra906010b2009-09-21 16:08:49 +02002821
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002822 return virt_to_page(buffer->data_pages[pgoff - 1]);
Peter Zijlstra906010b2009-09-21 16:08:49 +02002823}
2824
Peter Zijlstraa19d35c2010-05-17 18:48:00 +02002825static void *perf_mmap_alloc_page(int cpu)
2826{
2827 struct page *page;
2828 int node;
2829
2830 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
2831 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
2832 if (!page)
2833 return NULL;
2834
2835 return page_address(page);
2836}
2837
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002838static struct perf_buffer *
Peter Zijlstrad57e34f2010-05-28 19:41:35 +02002839perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002840{
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002841 struct perf_buffer *buffer;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002842 unsigned long size;
2843 int i;
2844
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002845 size = sizeof(struct perf_buffer);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002846 size += nr_pages * sizeof(void *);
2847
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002848 buffer = kzalloc(size, GFP_KERNEL);
2849 if (!buffer)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002850 goto fail;
2851
Peter Zijlstrad57e34f2010-05-28 19:41:35 +02002852 buffer->user_page = perf_mmap_alloc_page(cpu);
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002853 if (!buffer->user_page)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002854 goto fail_user_page;
2855
2856 for (i = 0; i < nr_pages; i++) {
Peter Zijlstrad57e34f2010-05-28 19:41:35 +02002857 buffer->data_pages[i] = perf_mmap_alloc_page(cpu);
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002858 if (!buffer->data_pages[i])
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002859 goto fail_data_pages;
2860 }
2861
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002862 buffer->nr_pages = nr_pages;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002863
Peter Zijlstrad57e34f2010-05-28 19:41:35 +02002864 perf_buffer_init(buffer, watermark, flags);
2865
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002866 return buffer;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002867
2868fail_data_pages:
2869 for (i--; i >= 0; i--)
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002870 free_page((unsigned long)buffer->data_pages[i]);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002871
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002872 free_page((unsigned long)buffer->user_page);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002873
2874fail_user_page:
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002875 kfree(buffer);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002876
2877fail:
Peter Zijlstra906010b2009-09-21 16:08:49 +02002878 return NULL;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002879}
2880
2881static void perf_mmap_free_page(unsigned long addr)
2882{
2883 struct page *page = virt_to_page((void *)addr);
2884
2885 page->mapping = NULL;
2886 __free_page(page);
2887}
2888
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002889static void perf_buffer_free(struct perf_buffer *buffer)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002890{
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002891 int i;
2892
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002893 perf_mmap_free_page((unsigned long)buffer->user_page);
2894 for (i = 0; i < buffer->nr_pages; i++)
2895 perf_mmap_free_page((unsigned long)buffer->data_pages[i]);
2896 kfree(buffer);
Peter Zijlstra906010b2009-09-21 16:08:49 +02002897}
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002898
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002899static inline int page_order(struct perf_buffer *buffer)
Peter Zijlstra3cafa9f2010-05-20 19:07:56 +02002900{
2901 return 0;
2902}
2903
Peter Zijlstra906010b2009-09-21 16:08:49 +02002904#else
2905
2906/*
2907 * Back perf_mmap() with vmalloc memory.
2908 *
2909 * Required for architectures that have d-cache aliasing issues.
2910 */
2911
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002912static inline int page_order(struct perf_buffer *buffer)
Peter Zijlstra3cafa9f2010-05-20 19:07:56 +02002913{
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002914 return buffer->page_order;
Peter Zijlstra3cafa9f2010-05-20 19:07:56 +02002915}
2916
Peter Zijlstra906010b2009-09-21 16:08:49 +02002917static struct page *
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002918perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
Peter Zijlstra906010b2009-09-21 16:08:49 +02002919{
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002920 if (pgoff > (1UL << page_order(buffer)))
Peter Zijlstra906010b2009-09-21 16:08:49 +02002921 return NULL;
2922
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002923 return vmalloc_to_page((void *)buffer->user_page + pgoff * PAGE_SIZE);
Peter Zijlstra906010b2009-09-21 16:08:49 +02002924}
2925
2926static void perf_mmap_unmark_page(void *addr)
2927{
2928 struct page *page = vmalloc_to_page(addr);
2929
2930 page->mapping = NULL;
2931}
2932
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002933static void perf_buffer_free_work(struct work_struct *work)
Peter Zijlstra906010b2009-09-21 16:08:49 +02002934{
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002935 struct perf_buffer *buffer;
Peter Zijlstra906010b2009-09-21 16:08:49 +02002936 void *base;
2937 int i, nr;
2938
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002939 buffer = container_of(work, struct perf_buffer, work);
2940 nr = 1 << page_order(buffer);
Peter Zijlstra906010b2009-09-21 16:08:49 +02002941
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002942 base = buffer->user_page;
Peter Zijlstra906010b2009-09-21 16:08:49 +02002943 for (i = 0; i < nr + 1; i++)
2944 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2945
2946 vfree(base);
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002947 kfree(buffer);
Peter Zijlstra906010b2009-09-21 16:08:49 +02002948}
2949
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002950static void perf_buffer_free(struct perf_buffer *buffer)
Peter Zijlstra906010b2009-09-21 16:08:49 +02002951{
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002952 schedule_work(&buffer->work);
Peter Zijlstra906010b2009-09-21 16:08:49 +02002953}
2954
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002955static struct perf_buffer *
Peter Zijlstrad57e34f2010-05-28 19:41:35 +02002956perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
Peter Zijlstra906010b2009-09-21 16:08:49 +02002957{
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002958 struct perf_buffer *buffer;
Peter Zijlstra906010b2009-09-21 16:08:49 +02002959 unsigned long size;
2960 void *all_buf;
2961
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002962 size = sizeof(struct perf_buffer);
Peter Zijlstra906010b2009-09-21 16:08:49 +02002963 size += sizeof(void *);
2964
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002965 buffer = kzalloc(size, GFP_KERNEL);
2966 if (!buffer)
Peter Zijlstra906010b2009-09-21 16:08:49 +02002967 goto fail;
2968
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002969 INIT_WORK(&buffer->work, perf_buffer_free_work);
Peter Zijlstra906010b2009-09-21 16:08:49 +02002970
2971 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2972 if (!all_buf)
2973 goto fail_all_buf;
2974
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002975 buffer->user_page = all_buf;
2976 buffer->data_pages[0] = all_buf + PAGE_SIZE;
2977 buffer->page_order = ilog2(nr_pages);
2978 buffer->nr_pages = 1;
Peter Zijlstra906010b2009-09-21 16:08:49 +02002979
Peter Zijlstrad57e34f2010-05-28 19:41:35 +02002980 perf_buffer_init(buffer, watermark, flags);
2981
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002982 return buffer;
Peter Zijlstra906010b2009-09-21 16:08:49 +02002983
2984fail_all_buf:
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002985 kfree(buffer);
Peter Zijlstra906010b2009-09-21 16:08:49 +02002986
2987fail:
2988 return NULL;
2989}
2990
2991#endif
2992
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002993static unsigned long perf_data_size(struct perf_buffer *buffer)
Peter Zijlstra3cafa9f2010-05-20 19:07:56 +02002994{
Peter Zijlstraca5135e2010-05-28 19:33:23 +02002995 return buffer->nr_pages << (PAGE_SHIFT + page_order(buffer));
Peter Zijlstra3cafa9f2010-05-20 19:07:56 +02002996}
2997
Peter Zijlstra906010b2009-09-21 16:08:49 +02002998static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2999{
3000 struct perf_event *event = vma->vm_file->private_data;
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003001 struct perf_buffer *buffer;
Peter Zijlstra906010b2009-09-21 16:08:49 +02003002 int ret = VM_FAULT_SIGBUS;
3003
3004 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3005 if (vmf->pgoff == 0)
3006 ret = 0;
3007 return ret;
3008 }
3009
3010 rcu_read_lock();
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003011 buffer = rcu_dereference(event->buffer);
3012 if (!buffer)
Peter Zijlstra906010b2009-09-21 16:08:49 +02003013 goto unlock;
3014
3015 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3016 goto unlock;
3017
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003018 vmf->page = perf_mmap_to_page(buffer, vmf->pgoff);
Peter Zijlstra906010b2009-09-21 16:08:49 +02003019 if (!vmf->page)
3020 goto unlock;
3021
3022 get_page(vmf->page);
3023 vmf->page->mapping = vma->vm_file->f_mapping;
3024 vmf->page->index = vmf->pgoff;
3025
3026 ret = 0;
3027unlock:
3028 rcu_read_unlock();
3029
3030 return ret;
3031}
3032
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003033static void perf_buffer_free_rcu(struct rcu_head *rcu_head)
Peter Zijlstra906010b2009-09-21 16:08:49 +02003034{
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003035 struct perf_buffer *buffer;
Peter Zijlstra906010b2009-09-21 16:08:49 +02003036
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003037 buffer = container_of(rcu_head, struct perf_buffer, rcu_head);
3038 perf_buffer_free(buffer);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003039}
3040
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003041static struct perf_buffer *perf_buffer_get(struct perf_event *event)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003042{
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003043 struct perf_buffer *buffer;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003044
Peter Zijlstraac9721f2010-05-27 12:54:41 +02003045 rcu_read_lock();
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003046 buffer = rcu_dereference(event->buffer);
3047 if (buffer) {
3048 if (!atomic_inc_not_zero(&buffer->refcount))
3049 buffer = NULL;
Peter Zijlstraac9721f2010-05-27 12:54:41 +02003050 }
3051 rcu_read_unlock();
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003052
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003053 return buffer;
Peter Zijlstraac9721f2010-05-27 12:54:41 +02003054}
3055
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003056static void perf_buffer_put(struct perf_buffer *buffer)
Peter Zijlstraac9721f2010-05-27 12:54:41 +02003057{
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003058 if (!atomic_dec_and_test(&buffer->refcount))
Peter Zijlstraac9721f2010-05-27 12:54:41 +02003059 return;
3060
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003061 call_rcu(&buffer->rcu_head, perf_buffer_free_rcu);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003062}
3063
3064static void perf_mmap_open(struct vm_area_struct *vma)
3065{
3066 struct perf_event *event = vma->vm_file->private_data;
3067
3068 atomic_inc(&event->mmap_count);
3069}
3070
3071static void perf_mmap_close(struct vm_area_struct *vma)
3072{
3073 struct perf_event *event = vma->vm_file->private_data;
3074
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003075 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003076 unsigned long size = perf_data_size(event->buffer);
Peter Zijlstraac9721f2010-05-27 12:54:41 +02003077 struct user_struct *user = event->mmap_user;
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003078 struct perf_buffer *buffer = event->buffer;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003079
Peter Zijlstra906010b2009-09-21 16:08:49 +02003080 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
Peter Zijlstraac9721f2010-05-27 12:54:41 +02003081 vma->vm_mm->locked_vm -= event->mmap_locked;
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003082 rcu_assign_pointer(event->buffer, NULL);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003083 mutex_unlock(&event->mmap_mutex);
Peter Zijlstraac9721f2010-05-27 12:54:41 +02003084
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003085 perf_buffer_put(buffer);
Peter Zijlstraac9721f2010-05-27 12:54:41 +02003086 free_uid(user);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003087 }
3088}
3089
Alexey Dobriyanf0f37e22009-09-27 22:29:37 +04003090static const struct vm_operations_struct perf_mmap_vmops = {
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003091 .open = perf_mmap_open,
3092 .close = perf_mmap_close,
3093 .fault = perf_mmap_fault,
3094 .page_mkwrite = perf_mmap_fault,
3095};
3096
3097static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3098{
3099 struct perf_event *event = file->private_data;
3100 unsigned long user_locked, user_lock_limit;
3101 struct user_struct *user = current_user();
3102 unsigned long locked, lock_limit;
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003103 struct perf_buffer *buffer;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003104 unsigned long vma_size;
3105 unsigned long nr_pages;
3106 long user_extra, extra;
Peter Zijlstrad57e34f2010-05-28 19:41:35 +02003107 int ret = 0, flags = 0;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003108
Peter Zijlstrac7920612010-05-18 10:33:24 +02003109 /*
3110 * Don't allow mmap() of inherited per-task counters. This would
3111 * create a performance issue due to all children writing to the
3112 * same buffer.
3113 */
3114 if (event->cpu == -1 && event->attr.inherit)
3115 return -EINVAL;
3116
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003117 if (!(vma->vm_flags & VM_SHARED))
3118 return -EINVAL;
3119
3120 vma_size = vma->vm_end - vma->vm_start;
3121 nr_pages = (vma_size / PAGE_SIZE) - 1;
3122
3123 /*
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003124 * If we have buffer pages ensure they're a power-of-two number, so we
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003125 * can do bitmasks instead of modulo.
3126 */
3127 if (nr_pages != 0 && !is_power_of_2(nr_pages))
3128 return -EINVAL;
3129
3130 if (vma_size != PAGE_SIZE * (1 + nr_pages))
3131 return -EINVAL;
3132
3133 if (vma->vm_pgoff != 0)
3134 return -EINVAL;
3135
3136 WARN_ON_ONCE(event->ctx->parent_ctx);
3137 mutex_lock(&event->mmap_mutex);
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003138 if (event->buffer) {
3139 if (event->buffer->nr_pages == nr_pages)
3140 atomic_inc(&event->buffer->refcount);
Peter Zijlstraac9721f2010-05-27 12:54:41 +02003141 else
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003142 ret = -EINVAL;
3143 goto unlock;
3144 }
3145
3146 user_extra = nr_pages + 1;
3147 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3148
3149 /*
3150 * Increase the limit linearly with more CPUs:
3151 */
3152 user_lock_limit *= num_online_cpus();
3153
3154 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3155
3156 extra = 0;
3157 if (user_locked > user_lock_limit)
3158 extra = user_locked - user_lock_limit;
3159
Jiri Slaby78d7d402010-03-05 13:42:54 -08003160 lock_limit = rlimit(RLIMIT_MEMLOCK);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003161 lock_limit >>= PAGE_SHIFT;
3162 locked = vma->vm_mm->locked_vm + extra;
3163
3164 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3165 !capable(CAP_IPC_LOCK)) {
3166 ret = -EPERM;
3167 goto unlock;
3168 }
3169
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003170 WARN_ON(event->buffer);
Peter Zijlstra906010b2009-09-21 16:08:49 +02003171
Peter Zijlstrad57e34f2010-05-28 19:41:35 +02003172 if (vma->vm_flags & VM_WRITE)
3173 flags |= PERF_BUFFER_WRITABLE;
3174
3175 buffer = perf_buffer_alloc(nr_pages, event->attr.wakeup_watermark,
3176 event->cpu, flags);
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003177 if (!buffer) {
Peter Zijlstraac9721f2010-05-27 12:54:41 +02003178 ret = -ENOMEM;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003179 goto unlock;
Peter Zijlstraac9721f2010-05-27 12:54:41 +02003180 }
Peter Zijlstrad57e34f2010-05-28 19:41:35 +02003181 rcu_assign_pointer(event->buffer, buffer);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003182
Peter Zijlstraac9721f2010-05-27 12:54:41 +02003183 atomic_long_add(user_extra, &user->locked_vm);
3184 event->mmap_locked = extra;
3185 event->mmap_user = get_current_user();
3186 vma->vm_mm->locked_vm += event->mmap_locked;
3187
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003188unlock:
Peter Zijlstraac9721f2010-05-27 12:54:41 +02003189 if (!ret)
3190 atomic_inc(&event->mmap_count);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003191 mutex_unlock(&event->mmap_mutex);
3192
3193 vma->vm_flags |= VM_RESERVED;
3194 vma->vm_ops = &perf_mmap_vmops;
3195
3196 return ret;
3197}
3198
3199static int perf_fasync(int fd, struct file *filp, int on)
3200{
3201 struct inode *inode = filp->f_path.dentry->d_inode;
3202 struct perf_event *event = filp->private_data;
3203 int retval;
3204
3205 mutex_lock(&inode->i_mutex);
3206 retval = fasync_helper(fd, filp, on, &event->fasync);
3207 mutex_unlock(&inode->i_mutex);
3208
3209 if (retval < 0)
3210 return retval;
3211
3212 return 0;
3213}
3214
3215static const struct file_operations perf_fops = {
Arnd Bergmann3326c1c2010-03-23 19:09:33 +01003216 .llseek = no_llseek,
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003217 .release = perf_release,
3218 .read = perf_read,
3219 .poll = perf_poll,
3220 .unlocked_ioctl = perf_ioctl,
3221 .compat_ioctl = perf_ioctl,
3222 .mmap = perf_mmap,
3223 .fasync = perf_fasync,
3224};
3225
3226/*
3227 * Perf event wakeup
3228 *
3229 * If there's data, ensure we set the poll() state and publish everything
3230 * to user-space before waking everybody up.
3231 */
3232
3233void perf_event_wakeup(struct perf_event *event)
3234{
3235 wake_up_all(&event->waitq);
3236
3237 if (event->pending_kill) {
3238 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3239 event->pending_kill = 0;
3240 }
3241}
3242
Peter Zijlstrae360adb2010-10-14 14:01:34 +08003243static void perf_pending_event(struct irq_work *entry)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003244{
3245 struct perf_event *event = container_of(entry,
3246 struct perf_event, pending);
3247
3248 if (event->pending_disable) {
3249 event->pending_disable = 0;
3250 __perf_event_disable(event);
3251 }
3252
3253 if (event->pending_wakeup) {
3254 event->pending_wakeup = 0;
3255 perf_event_wakeup(event);
3256 }
3257}
3258
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003259/*
Zhang, Yanmin39447b32010-04-19 13:32:41 +08003260 * We assume there is only KVM supporting the callbacks.
3261 * Later on, we might change it to a list if there is
3262 * another virtualization implementation supporting the callbacks.
3263 */
3264struct perf_guest_info_callbacks *perf_guest_cbs;
3265
3266int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3267{
3268 perf_guest_cbs = cbs;
3269 return 0;
3270}
3271EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3272
3273int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3274{
3275 perf_guest_cbs = NULL;
3276 return 0;
3277}
3278EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3279
3280/*
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003281 * Output
3282 */
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003283static bool perf_output_space(struct perf_buffer *buffer, unsigned long tail,
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003284 unsigned long offset, unsigned long head)
3285{
3286 unsigned long mask;
3287
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003288 if (!buffer->writable)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003289 return true;
3290
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003291 mask = perf_data_size(buffer) - 1;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003292
3293 offset = (offset - tail) & mask;
3294 head = (head - tail) & mask;
3295
3296 if ((int)(head - offset) < 0)
3297 return false;
3298
3299 return true;
3300}
3301
3302static void perf_output_wakeup(struct perf_output_handle *handle)
3303{
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003304 atomic_set(&handle->buffer->poll, POLL_IN);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003305
3306 if (handle->nmi) {
3307 handle->event->pending_wakeup = 1;
Peter Zijlstrae360adb2010-10-14 14:01:34 +08003308 irq_work_queue(&handle->event->pending);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003309 } else
3310 perf_event_wakeup(handle->event);
3311}
3312
3313/*
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003314 * We need to ensure a later event_id doesn't publish a head when a former
Peter Zijlstraef607772010-05-18 10:50:41 +02003315 * event isn't done writing. However since we need to deal with NMIs we
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003316 * cannot fully serialize things.
3317 *
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003318 * We only publish the head (and generate a wakeup) when the outer-most
Peter Zijlstraef607772010-05-18 10:50:41 +02003319 * event completes.
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003320 */
Peter Zijlstraef607772010-05-18 10:50:41 +02003321static void perf_output_get_handle(struct perf_output_handle *handle)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003322{
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003323 struct perf_buffer *buffer = handle->buffer;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003324
Peter Zijlstraef607772010-05-18 10:50:41 +02003325 preempt_disable();
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003326 local_inc(&buffer->nest);
3327 handle->wakeup = local_read(&buffer->wakeup);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003328}
3329
Peter Zijlstraef607772010-05-18 10:50:41 +02003330static void perf_output_put_handle(struct perf_output_handle *handle)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003331{
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003332 struct perf_buffer *buffer = handle->buffer;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003333 unsigned long head;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003334
3335again:
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003336 head = local_read(&buffer->head);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003337
3338 /*
Peter Zijlstraef607772010-05-18 10:50:41 +02003339 * IRQ/NMI can happen here, which means we can miss a head update.
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003340 */
3341
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003342 if (!local_dec_and_test(&buffer->nest))
Frederic Weisbeckeracd35a42010-05-20 21:28:34 +02003343 goto out;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003344
3345 /*
Peter Zijlstraef607772010-05-18 10:50:41 +02003346 * Publish the known good head. Rely on the full barrier implied
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003347 * by atomic_dec_and_test() order the buffer->head read and this
Peter Zijlstraef607772010-05-18 10:50:41 +02003348 * write.
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003349 */
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003350 buffer->user_page->data_head = head;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003351
Peter Zijlstraef607772010-05-18 10:50:41 +02003352 /*
3353 * Now check if we missed an update, rely on the (compiler)
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003354 * barrier in atomic_dec_and_test() to re-read buffer->head.
Peter Zijlstraef607772010-05-18 10:50:41 +02003355 */
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003356 if (unlikely(head != local_read(&buffer->head))) {
3357 local_inc(&buffer->nest);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003358 goto again;
3359 }
3360
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003361 if (handle->wakeup != local_read(&buffer->wakeup))
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003362 perf_output_wakeup(handle);
Peter Zijlstraef607772010-05-18 10:50:41 +02003363
Peter Zijlstra9ed60602010-06-11 17:36:35 +02003364out:
Peter Zijlstraef607772010-05-18 10:50:41 +02003365 preempt_enable();
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003366}
3367
Peter Zijlstraa94ffaa2010-05-20 19:50:07 +02003368__always_inline void perf_output_copy(struct perf_output_handle *handle,
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003369 const void *buf, unsigned int len)
3370{
Peter Zijlstra5d967a82010-05-20 16:46:39 +02003371 do {
Peter Zijlstraa94ffaa2010-05-20 19:50:07 +02003372 unsigned long size = min_t(unsigned long, handle->size, len);
Peter Zijlstra5d967a82010-05-20 16:46:39 +02003373
3374 memcpy(handle->addr, buf, size);
3375
3376 len -= size;
3377 handle->addr += size;
Frederic Weisbecker74048f82010-05-27 21:34:58 +02003378 buf += size;
Peter Zijlstra5d967a82010-05-20 16:46:39 +02003379 handle->size -= size;
3380 if (!handle->size) {
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003381 struct perf_buffer *buffer = handle->buffer;
Peter Zijlstra3cafa9f2010-05-20 19:07:56 +02003382
Peter Zijlstra5d967a82010-05-20 16:46:39 +02003383 handle->page++;
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003384 handle->page &= buffer->nr_pages - 1;
3385 handle->addr = buffer->data_pages[handle->page];
3386 handle->size = PAGE_SIZE << page_order(buffer);
Peter Zijlstra5d967a82010-05-20 16:46:39 +02003387 }
3388 } while (len);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003389}
3390
Arnaldo Carvalho de Melo6844c092010-12-03 16:36:35 -02003391static void perf_event_header__init_id(struct perf_event_header *header,
3392 struct perf_sample_data *data,
3393 struct perf_event *event)
3394{
3395 u64 sample_type = event->attr.sample_type;
3396
3397 data->type = sample_type;
3398 header->size += event->id_header_size;
3399
3400 if (sample_type & PERF_SAMPLE_TID) {
3401 /* namespace issues */
3402 data->tid_entry.pid = perf_event_pid(event, current);
3403 data->tid_entry.tid = perf_event_tid(event, current);
3404 }
3405
3406 if (sample_type & PERF_SAMPLE_TIME)
3407 data->time = perf_clock();
3408
3409 if (sample_type & PERF_SAMPLE_ID)
3410 data->id = primary_event_id(event);
3411
3412 if (sample_type & PERF_SAMPLE_STREAM_ID)
3413 data->stream_id = event->id;
3414
3415 if (sample_type & PERF_SAMPLE_CPU) {
3416 data->cpu_entry.cpu = raw_smp_processor_id();
3417 data->cpu_entry.reserved = 0;
3418 }
3419}
3420
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003421int perf_output_begin(struct perf_output_handle *handle,
3422 struct perf_event *event, unsigned int size,
3423 int nmi, int sample)
3424{
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003425 struct perf_buffer *buffer;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003426 unsigned long tail, offset, head;
3427 int have_lost;
3428 struct {
3429 struct perf_event_header header;
3430 u64 id;
3431 u64 lost;
3432 } lost_event;
3433
3434 rcu_read_lock();
3435 /*
3436 * For inherited events we send all the output towards the parent.
3437 */
3438 if (event->parent)
3439 event = event->parent;
3440
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003441 buffer = rcu_dereference(event->buffer);
3442 if (!buffer)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003443 goto out;
3444
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003445 handle->buffer = buffer;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003446 handle->event = event;
3447 handle->nmi = nmi;
3448 handle->sample = sample;
3449
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003450 if (!buffer->nr_pages)
Stephane Eranian00d1d0b2010-05-17 12:46:01 +02003451 goto out;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003452
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003453 have_lost = local_read(&buffer->lost);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003454 if (have_lost)
3455 size += sizeof(lost_event);
3456
Peter Zijlstraef607772010-05-18 10:50:41 +02003457 perf_output_get_handle(handle);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003458
3459 do {
3460 /*
3461 * Userspace could choose to issue a mb() before updating the
3462 * tail pointer. So that all reads will be completed before the
3463 * write is issued.
3464 */
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003465 tail = ACCESS_ONCE(buffer->user_page->data_tail);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003466 smp_rmb();
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003467 offset = head = local_read(&buffer->head);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003468 head += size;
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003469 if (unlikely(!perf_output_space(buffer, tail, offset, head)))
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003470 goto fail;
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003471 } while (local_cmpxchg(&buffer->head, offset, head) != offset);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003472
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003473 if (head - local_read(&buffer->wakeup) > buffer->watermark)
3474 local_add(buffer->watermark, &buffer->wakeup);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003475
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003476 handle->page = offset >> (PAGE_SHIFT + page_order(buffer));
3477 handle->page &= buffer->nr_pages - 1;
3478 handle->size = offset & ((PAGE_SIZE << page_order(buffer)) - 1);
3479 handle->addr = buffer->data_pages[handle->page];
Peter Zijlstra5d967a82010-05-20 16:46:39 +02003480 handle->addr += handle->size;
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003481 handle->size = (PAGE_SIZE << page_order(buffer)) - handle->size;
Peter Zijlstra5d967a82010-05-20 16:46:39 +02003482
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003483 if (have_lost) {
3484 lost_event.header.type = PERF_RECORD_LOST;
3485 lost_event.header.misc = 0;
3486 lost_event.header.size = sizeof(lost_event);
3487 lost_event.id = event->id;
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003488 lost_event.lost = local_xchg(&buffer->lost, 0);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003489
3490 perf_output_put(handle, lost_event);
3491 }
3492
3493 return 0;
3494
3495fail:
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003496 local_inc(&buffer->lost);
Peter Zijlstraef607772010-05-18 10:50:41 +02003497 perf_output_put_handle(handle);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003498out:
3499 rcu_read_unlock();
3500
3501 return -ENOSPC;
3502}
3503
3504void perf_output_end(struct perf_output_handle *handle)
3505{
3506 struct perf_event *event = handle->event;
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003507 struct perf_buffer *buffer = handle->buffer;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003508
3509 int wakeup_events = event->attr.wakeup_events;
3510
3511 if (handle->sample && wakeup_events) {
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003512 int events = local_inc_return(&buffer->events);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003513 if (events >= wakeup_events) {
Peter Zijlstraca5135e2010-05-28 19:33:23 +02003514 local_sub(wakeup_events, &buffer->events);
3515 local_inc(&buffer->wakeup);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003516 }
3517 }
3518
Peter Zijlstraef607772010-05-18 10:50:41 +02003519 perf_output_put_handle(handle);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003520 rcu_read_unlock();
3521}
3522
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003523static void perf_output_read_one(struct perf_output_handle *handle,
Stephane Eranianeed01522010-10-26 16:08:01 +02003524 struct perf_event *event,
3525 u64 enabled, u64 running)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003526{
3527 u64 read_format = event->attr.read_format;
3528 u64 values[4];
3529 int n = 0;
3530
Peter Zijlstrab5e58792010-05-21 14:43:12 +02003531 values[n++] = perf_event_count(event);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003532 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
Stephane Eranianeed01522010-10-26 16:08:01 +02003533 values[n++] = enabled +
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003534 atomic64_read(&event->child_total_time_enabled);
3535 }
3536 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
Stephane Eranianeed01522010-10-26 16:08:01 +02003537 values[n++] = running +
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003538 atomic64_read(&event->child_total_time_running);
3539 }
3540 if (read_format & PERF_FORMAT_ID)
3541 values[n++] = primary_event_id(event);
3542
3543 perf_output_copy(handle, values, n * sizeof(u64));
3544}
3545
3546/*
3547 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3548 */
3549static void perf_output_read_group(struct perf_output_handle *handle,
Stephane Eranianeed01522010-10-26 16:08:01 +02003550 struct perf_event *event,
3551 u64 enabled, u64 running)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003552{
3553 struct perf_event *leader = event->group_leader, *sub;
3554 u64 read_format = event->attr.read_format;
3555 u64 values[5];
3556 int n = 0;
3557
3558 values[n++] = 1 + leader->nr_siblings;
3559
3560 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
Stephane Eranianeed01522010-10-26 16:08:01 +02003561 values[n++] = enabled;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003562
3563 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
Stephane Eranianeed01522010-10-26 16:08:01 +02003564 values[n++] = running;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003565
3566 if (leader != event)
3567 leader->pmu->read(leader);
3568
Peter Zijlstrab5e58792010-05-21 14:43:12 +02003569 values[n++] = perf_event_count(leader);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003570 if (read_format & PERF_FORMAT_ID)
3571 values[n++] = primary_event_id(leader);
3572
3573 perf_output_copy(handle, values, n * sizeof(u64));
3574
3575 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3576 n = 0;
3577
3578 if (sub != event)
3579 sub->pmu->read(sub);
3580
Peter Zijlstrab5e58792010-05-21 14:43:12 +02003581 values[n++] = perf_event_count(sub);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003582 if (read_format & PERF_FORMAT_ID)
3583 values[n++] = primary_event_id(sub);
3584
3585 perf_output_copy(handle, values, n * sizeof(u64));
3586 }
3587}
3588
Stephane Eranianeed01522010-10-26 16:08:01 +02003589#define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
3590 PERF_FORMAT_TOTAL_TIME_RUNNING)
3591
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003592static void perf_output_read(struct perf_output_handle *handle,
3593 struct perf_event *event)
3594{
Stephane Eranianeed01522010-10-26 16:08:01 +02003595 u64 enabled = 0, running = 0, now, ctx_time;
3596 u64 read_format = event->attr.read_format;
3597
3598 /*
3599 * compute total_time_enabled, total_time_running
3600 * based on snapshot values taken when the event
3601 * was last scheduled in.
3602 *
3603 * we cannot simply called update_context_time()
3604 * because of locking issue as we are called in
3605 * NMI context
3606 */
3607 if (read_format & PERF_FORMAT_TOTAL_TIMES) {
3608 now = perf_clock();
3609 ctx_time = event->shadow_ctx_time + now;
3610 enabled = ctx_time - event->tstamp_enabled;
3611 running = ctx_time - event->tstamp_running;
3612 }
3613
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003614 if (event->attr.read_format & PERF_FORMAT_GROUP)
Stephane Eranianeed01522010-10-26 16:08:01 +02003615 perf_output_read_group(handle, event, enabled, running);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003616 else
Stephane Eranianeed01522010-10-26 16:08:01 +02003617 perf_output_read_one(handle, event, enabled, running);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003618}
3619
3620void perf_output_sample(struct perf_output_handle *handle,
3621 struct perf_event_header *header,
3622 struct perf_sample_data *data,
3623 struct perf_event *event)
3624{
3625 u64 sample_type = data->type;
3626
3627 perf_output_put(handle, *header);
3628
3629 if (sample_type & PERF_SAMPLE_IP)
3630 perf_output_put(handle, data->ip);
3631
3632 if (sample_type & PERF_SAMPLE_TID)
3633 perf_output_put(handle, data->tid_entry);
3634
3635 if (sample_type & PERF_SAMPLE_TIME)
3636 perf_output_put(handle, data->time);
3637
3638 if (sample_type & PERF_SAMPLE_ADDR)
3639 perf_output_put(handle, data->addr);
3640
3641 if (sample_type & PERF_SAMPLE_ID)
3642 perf_output_put(handle, data->id);
3643
3644 if (sample_type & PERF_SAMPLE_STREAM_ID)
3645 perf_output_put(handle, data->stream_id);
3646
3647 if (sample_type & PERF_SAMPLE_CPU)
3648 perf_output_put(handle, data->cpu_entry);
3649
3650 if (sample_type & PERF_SAMPLE_PERIOD)
3651 perf_output_put(handle, data->period);
3652
3653 if (sample_type & PERF_SAMPLE_READ)
3654 perf_output_read(handle, event);
3655
3656 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3657 if (data->callchain) {
3658 int size = 1;
3659
3660 if (data->callchain)
3661 size += data->callchain->nr;
3662
3663 size *= sizeof(u64);
3664
3665 perf_output_copy(handle, data->callchain, size);
3666 } else {
3667 u64 nr = 0;
3668 perf_output_put(handle, nr);
3669 }
3670 }
3671
3672 if (sample_type & PERF_SAMPLE_RAW) {
3673 if (data->raw) {
3674 perf_output_put(handle, data->raw->size);
3675 perf_output_copy(handle, data->raw->data,
3676 data->raw->size);
3677 } else {
3678 struct {
3679 u32 size;
3680 u32 data;
3681 } raw = {
3682 .size = sizeof(u32),
3683 .data = 0,
3684 };
3685 perf_output_put(handle, raw);
3686 }
3687 }
3688}
3689
3690void perf_prepare_sample(struct perf_event_header *header,
3691 struct perf_sample_data *data,
3692 struct perf_event *event,
3693 struct pt_regs *regs)
3694{
3695 u64 sample_type = event->attr.sample_type;
3696
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003697 header->type = PERF_RECORD_SAMPLE;
Arnaldo Carvalho de Meloc320c7b2010-10-20 12:50:11 -02003698 header->size = sizeof(*header) + event->header_size;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003699
3700 header->misc = 0;
3701 header->misc |= perf_misc_flags(regs);
3702
Arnaldo Carvalho de Melo6844c092010-12-03 16:36:35 -02003703 perf_event_header__init_id(header, data, event);
3704
Arnaldo Carvalho de Meloc320c7b2010-10-20 12:50:11 -02003705 if (sample_type & PERF_SAMPLE_IP)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003706 data->ip = perf_instruction_pointer(regs);
3707
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003708 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3709 int size = 1;
3710
3711 data->callchain = perf_callchain(regs);
3712
3713 if (data->callchain)
3714 size += data->callchain->nr;
3715
3716 header->size += size * sizeof(u64);
3717 }
3718
3719 if (sample_type & PERF_SAMPLE_RAW) {
3720 int size = sizeof(u32);
3721
3722 if (data->raw)
3723 size += data->raw->size;
3724 else
3725 size += sizeof(u32);
3726
3727 WARN_ON_ONCE(size & (sizeof(u64)-1));
3728 header->size += size;
3729 }
3730}
3731
3732static void perf_event_output(struct perf_event *event, int nmi,
3733 struct perf_sample_data *data,
3734 struct pt_regs *regs)
3735{
3736 struct perf_output_handle handle;
3737 struct perf_event_header header;
3738
Frederic Weisbecker927c7a92010-07-01 16:20:36 +02003739 /* protect the callchain buffers */
3740 rcu_read_lock();
3741
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003742 perf_prepare_sample(&header, data, event, regs);
3743
3744 if (perf_output_begin(&handle, event, header.size, nmi, 1))
Frederic Weisbecker927c7a92010-07-01 16:20:36 +02003745 goto exit;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003746
3747 perf_output_sample(&handle, &header, data, event);
3748
3749 perf_output_end(&handle);
Frederic Weisbecker927c7a92010-07-01 16:20:36 +02003750
3751exit:
3752 rcu_read_unlock();
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003753}
3754
3755/*
3756 * read event_id
3757 */
3758
3759struct perf_read_event {
3760 struct perf_event_header header;
3761
3762 u32 pid;
3763 u32 tid;
3764};
3765
3766static void
3767perf_event_read_event(struct perf_event *event,
3768 struct task_struct *task)
3769{
3770 struct perf_output_handle handle;
3771 struct perf_read_event read_event = {
3772 .header = {
3773 .type = PERF_RECORD_READ,
3774 .misc = 0,
Arnaldo Carvalho de Meloc320c7b2010-10-20 12:50:11 -02003775 .size = sizeof(read_event) + event->read_size,
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003776 },
3777 .pid = perf_event_pid(event, task),
3778 .tid = perf_event_tid(event, task),
3779 };
3780 int ret;
3781
3782 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3783 if (ret)
3784 return;
3785
3786 perf_output_put(&handle, read_event);
3787 perf_output_read(&handle, event);
3788
3789 perf_output_end(&handle);
3790}
3791
3792/*
3793 * task tracking -- fork/exit
3794 *
Eric B Munson3af9e852010-05-18 15:30:49 +01003795 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003796 */
3797
3798struct perf_task_event {
3799 struct task_struct *task;
3800 struct perf_event_context *task_ctx;
3801
3802 struct {
3803 struct perf_event_header header;
3804
3805 u32 pid;
3806 u32 ppid;
3807 u32 tid;
3808 u32 ptid;
3809 u64 time;
3810 } event_id;
3811};
3812
3813static void perf_event_task_output(struct perf_event *event,
3814 struct perf_task_event *task_event)
3815{
3816 struct perf_output_handle handle;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003817 struct task_struct *task = task_event->task;
Mike Galbraith8bb39f92010-03-26 11:11:33 +01003818 int size, ret;
3819
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003820 size = task_event->event_id.header.size;
3821 ret = perf_output_begin(&handle, event, size, 0, 0);
3822
Peter Zijlstraef607772010-05-18 10:50:41 +02003823 if (ret)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003824 return;
3825
3826 task_event->event_id.pid = perf_event_pid(event, task);
3827 task_event->event_id.ppid = perf_event_pid(event, current);
3828
3829 task_event->event_id.tid = perf_event_tid(event, task);
3830 task_event->event_id.ptid = perf_event_tid(event, current);
3831
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003832 perf_output_put(&handle, task_event->event_id);
3833
3834 perf_output_end(&handle);
3835}
3836
3837static int perf_event_task_match(struct perf_event *event)
3838{
Peter Zijlstra6f93d0a2010-02-14 11:12:04 +01003839 if (event->state < PERF_EVENT_STATE_INACTIVE)
Peter Zijlstra22e19082010-01-18 09:12:32 +01003840 return 0;
3841
Peter Zijlstra5d27c232009-12-17 13:16:32 +01003842 if (event->cpu != -1 && event->cpu != smp_processor_id())
3843 return 0;
3844
Eric B Munson3af9e852010-05-18 15:30:49 +01003845 if (event->attr.comm || event->attr.mmap ||
3846 event->attr.mmap_data || event->attr.task)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003847 return 1;
3848
3849 return 0;
3850}
3851
3852static void perf_event_task_ctx(struct perf_event_context *ctx,
3853 struct perf_task_event *task_event)
3854{
3855 struct perf_event *event;
3856
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003857 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3858 if (perf_event_task_match(event))
3859 perf_event_task_output(event, task_event);
3860 }
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003861}
3862
3863static void perf_event_task_event(struct perf_task_event *task_event)
3864{
3865 struct perf_cpu_context *cpuctx;
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02003866 struct perf_event_context *ctx;
Peter Zijlstra108b02c2010-09-06 14:32:03 +02003867 struct pmu *pmu;
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02003868 int ctxn;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003869
Peter Zijlstrad6ff86c2009-11-20 22:19:46 +01003870 rcu_read_lock();
Peter Zijlstra108b02c2010-09-06 14:32:03 +02003871 list_for_each_entry_rcu(pmu, &pmus, entry) {
Peter Zijlstra41945f62010-09-16 19:17:24 +02003872 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
Peter Zijlstra108b02c2010-09-06 14:32:03 +02003873 perf_event_task_ctx(&cpuctx->ctx, task_event);
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02003874
3875 ctx = task_event->task_ctx;
3876 if (!ctx) {
3877 ctxn = pmu->task_ctx_nr;
3878 if (ctxn < 0)
Peter Zijlstra41945f62010-09-16 19:17:24 +02003879 goto next;
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02003880 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
3881 }
3882 if (ctx)
3883 perf_event_task_ctx(ctx, task_event);
Peter Zijlstra41945f62010-09-16 19:17:24 +02003884next:
3885 put_cpu_ptr(pmu->pmu_cpu_context);
Peter Zijlstra108b02c2010-09-06 14:32:03 +02003886 }
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003887 rcu_read_unlock();
3888}
3889
3890static void perf_event_task(struct task_struct *task,
3891 struct perf_event_context *task_ctx,
3892 int new)
3893{
3894 struct perf_task_event task_event;
3895
3896 if (!atomic_read(&nr_comm_events) &&
3897 !atomic_read(&nr_mmap_events) &&
3898 !atomic_read(&nr_task_events))
3899 return;
3900
3901 task_event = (struct perf_task_event){
3902 .task = task,
3903 .task_ctx = task_ctx,
3904 .event_id = {
3905 .header = {
3906 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3907 .misc = 0,
3908 .size = sizeof(task_event.event_id),
3909 },
3910 /* .pid */
3911 /* .ppid */
3912 /* .tid */
3913 /* .ptid */
Peter Zijlstra6f93d0a2010-02-14 11:12:04 +01003914 .time = perf_clock(),
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003915 },
3916 };
3917
3918 perf_event_task_event(&task_event);
3919}
3920
3921void perf_event_fork(struct task_struct *task)
3922{
3923 perf_event_task(task, NULL, 1);
3924}
3925
3926/*
3927 * comm tracking
3928 */
3929
3930struct perf_comm_event {
3931 struct task_struct *task;
3932 char *comm;
3933 int comm_size;
3934
3935 struct {
3936 struct perf_event_header header;
3937
3938 u32 pid;
3939 u32 tid;
3940 } event_id;
3941};
3942
3943static void perf_event_comm_output(struct perf_event *event,
3944 struct perf_comm_event *comm_event)
3945{
3946 struct perf_output_handle handle;
3947 int size = comm_event->event_id.header.size;
3948 int ret = perf_output_begin(&handle, event, size, 0, 0);
3949
3950 if (ret)
3951 return;
3952
3953 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3954 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3955
3956 perf_output_put(&handle, comm_event->event_id);
3957 perf_output_copy(&handle, comm_event->comm,
3958 comm_event->comm_size);
3959 perf_output_end(&handle);
3960}
3961
3962static int perf_event_comm_match(struct perf_event *event)
3963{
Peter Zijlstra6f93d0a2010-02-14 11:12:04 +01003964 if (event->state < PERF_EVENT_STATE_INACTIVE)
Peter Zijlstra22e19082010-01-18 09:12:32 +01003965 return 0;
3966
Peter Zijlstra5d27c232009-12-17 13:16:32 +01003967 if (event->cpu != -1 && event->cpu != smp_processor_id())
3968 return 0;
3969
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003970 if (event->attr.comm)
3971 return 1;
3972
3973 return 0;
3974}
3975
3976static void perf_event_comm_ctx(struct perf_event_context *ctx,
3977 struct perf_comm_event *comm_event)
3978{
3979 struct perf_event *event;
3980
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003981 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3982 if (perf_event_comm_match(event))
3983 perf_event_comm_output(event, comm_event);
3984 }
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003985}
3986
3987static void perf_event_comm_event(struct perf_comm_event *comm_event)
3988{
3989 struct perf_cpu_context *cpuctx;
3990 struct perf_event_context *ctx;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003991 char comm[TASK_COMM_LEN];
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003992 unsigned int size;
Peter Zijlstra108b02c2010-09-06 14:32:03 +02003993 struct pmu *pmu;
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02003994 int ctxn;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003995
3996 memset(comm, 0, sizeof(comm));
Márton Németh96b02d72009-11-21 23:10:15 +01003997 strlcpy(comm, comm_event->task->comm, sizeof(comm));
Ingo Molnarcdd6c482009-09-21 12:02:48 +02003998 size = ALIGN(strlen(comm)+1, sizeof(u64));
3999
4000 comm_event->comm = comm;
4001 comm_event->comm_size = size;
4002
4003 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
4004
Peter Zijlstraf6595f32009-11-20 22:19:47 +01004005 rcu_read_lock();
Peter Zijlstra108b02c2010-09-06 14:32:03 +02004006 list_for_each_entry_rcu(pmu, &pmus, entry) {
Peter Zijlstra41945f62010-09-16 19:17:24 +02004007 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
Peter Zijlstra108b02c2010-09-06 14:32:03 +02004008 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02004009
4010 ctxn = pmu->task_ctx_nr;
4011 if (ctxn < 0)
Peter Zijlstra41945f62010-09-16 19:17:24 +02004012 goto next;
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02004013
4014 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4015 if (ctx)
4016 perf_event_comm_ctx(ctx, comm_event);
Peter Zijlstra41945f62010-09-16 19:17:24 +02004017next:
4018 put_cpu_ptr(pmu->pmu_cpu_context);
Peter Zijlstra108b02c2010-09-06 14:32:03 +02004019 }
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004020 rcu_read_unlock();
4021}
4022
4023void perf_event_comm(struct task_struct *task)
4024{
4025 struct perf_comm_event comm_event;
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02004026 struct perf_event_context *ctx;
4027 int ctxn;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004028
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02004029 for_each_task_context_nr(ctxn) {
4030 ctx = task->perf_event_ctxp[ctxn];
4031 if (!ctx)
4032 continue;
4033
4034 perf_event_enable_on_exec(ctx);
4035 }
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004036
4037 if (!atomic_read(&nr_comm_events))
4038 return;
4039
4040 comm_event = (struct perf_comm_event){
4041 .task = task,
4042 /* .comm */
4043 /* .comm_size */
4044 .event_id = {
4045 .header = {
4046 .type = PERF_RECORD_COMM,
4047 .misc = 0,
4048 /* .size */
4049 },
4050 /* .pid */
4051 /* .tid */
4052 },
4053 };
4054
4055 perf_event_comm_event(&comm_event);
4056}
4057
4058/*
4059 * mmap tracking
4060 */
4061
4062struct perf_mmap_event {
4063 struct vm_area_struct *vma;
4064
4065 const char *file_name;
4066 int file_size;
4067
4068 struct {
4069 struct perf_event_header header;
4070
4071 u32 pid;
4072 u32 tid;
4073 u64 start;
4074 u64 len;
4075 u64 pgoff;
4076 } event_id;
4077};
4078
4079static void perf_event_mmap_output(struct perf_event *event,
4080 struct perf_mmap_event *mmap_event)
4081{
4082 struct perf_output_handle handle;
4083 int size = mmap_event->event_id.header.size;
4084 int ret = perf_output_begin(&handle, event, size, 0, 0);
4085
4086 if (ret)
4087 return;
4088
4089 mmap_event->event_id.pid = perf_event_pid(event, current);
4090 mmap_event->event_id.tid = perf_event_tid(event, current);
4091
4092 perf_output_put(&handle, mmap_event->event_id);
4093 perf_output_copy(&handle, mmap_event->file_name,
4094 mmap_event->file_size);
4095 perf_output_end(&handle);
4096}
4097
4098static int perf_event_mmap_match(struct perf_event *event,
Eric B Munson3af9e852010-05-18 15:30:49 +01004099 struct perf_mmap_event *mmap_event,
4100 int executable)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004101{
Peter Zijlstra6f93d0a2010-02-14 11:12:04 +01004102 if (event->state < PERF_EVENT_STATE_INACTIVE)
Peter Zijlstra22e19082010-01-18 09:12:32 +01004103 return 0;
4104
Peter Zijlstra5d27c232009-12-17 13:16:32 +01004105 if (event->cpu != -1 && event->cpu != smp_processor_id())
4106 return 0;
4107
Eric B Munson3af9e852010-05-18 15:30:49 +01004108 if ((!executable && event->attr.mmap_data) ||
4109 (executable && event->attr.mmap))
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004110 return 1;
4111
4112 return 0;
4113}
4114
4115static void perf_event_mmap_ctx(struct perf_event_context *ctx,
Eric B Munson3af9e852010-05-18 15:30:49 +01004116 struct perf_mmap_event *mmap_event,
4117 int executable)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004118{
4119 struct perf_event *event;
4120
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004121 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
Eric B Munson3af9e852010-05-18 15:30:49 +01004122 if (perf_event_mmap_match(event, mmap_event, executable))
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004123 perf_event_mmap_output(event, mmap_event);
4124 }
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004125}
4126
4127static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4128{
4129 struct perf_cpu_context *cpuctx;
4130 struct perf_event_context *ctx;
4131 struct vm_area_struct *vma = mmap_event->vma;
4132 struct file *file = vma->vm_file;
4133 unsigned int size;
4134 char tmp[16];
4135 char *buf = NULL;
4136 const char *name;
Peter Zijlstra108b02c2010-09-06 14:32:03 +02004137 struct pmu *pmu;
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02004138 int ctxn;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004139
4140 memset(tmp, 0, sizeof(tmp));
4141
4142 if (file) {
4143 /*
4144 * d_path works from the end of the buffer backwards, so we
4145 * need to add enough zero bytes after the string to handle
4146 * the 64bit alignment we do later.
4147 */
4148 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4149 if (!buf) {
4150 name = strncpy(tmp, "//enomem", sizeof(tmp));
4151 goto got_name;
4152 }
4153 name = d_path(&file->f_path, buf, PATH_MAX);
4154 if (IS_ERR(name)) {
4155 name = strncpy(tmp, "//toolong", sizeof(tmp));
4156 goto got_name;
4157 }
4158 } else {
4159 if (arch_vma_name(mmap_event->vma)) {
4160 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4161 sizeof(tmp));
4162 goto got_name;
4163 }
4164
4165 if (!vma->vm_mm) {
4166 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4167 goto got_name;
Eric B Munson3af9e852010-05-18 15:30:49 +01004168 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4169 vma->vm_end >= vma->vm_mm->brk) {
4170 name = strncpy(tmp, "[heap]", sizeof(tmp));
4171 goto got_name;
4172 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4173 vma->vm_end >= vma->vm_mm->start_stack) {
4174 name = strncpy(tmp, "[stack]", sizeof(tmp));
4175 goto got_name;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004176 }
4177
4178 name = strncpy(tmp, "//anon", sizeof(tmp));
4179 goto got_name;
4180 }
4181
4182got_name:
4183 size = ALIGN(strlen(name)+1, sizeof(u64));
4184
4185 mmap_event->file_name = name;
4186 mmap_event->file_size = size;
4187
4188 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4189
Peter Zijlstraf6d9dd22009-11-20 22:19:48 +01004190 rcu_read_lock();
Peter Zijlstra108b02c2010-09-06 14:32:03 +02004191 list_for_each_entry_rcu(pmu, &pmus, entry) {
Peter Zijlstra41945f62010-09-16 19:17:24 +02004192 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
Peter Zijlstra108b02c2010-09-06 14:32:03 +02004193 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
4194 vma->vm_flags & VM_EXEC);
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02004195
4196 ctxn = pmu->task_ctx_nr;
4197 if (ctxn < 0)
Peter Zijlstra41945f62010-09-16 19:17:24 +02004198 goto next;
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02004199
4200 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4201 if (ctx) {
4202 perf_event_mmap_ctx(ctx, mmap_event,
4203 vma->vm_flags & VM_EXEC);
4204 }
Peter Zijlstra41945f62010-09-16 19:17:24 +02004205next:
4206 put_cpu_ptr(pmu->pmu_cpu_context);
Peter Zijlstra108b02c2010-09-06 14:32:03 +02004207 }
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004208 rcu_read_unlock();
4209
4210 kfree(buf);
4211}
4212
Eric B Munson3af9e852010-05-18 15:30:49 +01004213void perf_event_mmap(struct vm_area_struct *vma)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004214{
4215 struct perf_mmap_event mmap_event;
4216
4217 if (!atomic_read(&nr_mmap_events))
4218 return;
4219
4220 mmap_event = (struct perf_mmap_event){
4221 .vma = vma,
4222 /* .file_name */
4223 /* .file_size */
4224 .event_id = {
4225 .header = {
4226 .type = PERF_RECORD_MMAP,
Zhang, Yanmin39447b32010-04-19 13:32:41 +08004227 .misc = PERF_RECORD_MISC_USER,
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004228 /* .size */
4229 },
4230 /* .pid */
4231 /* .tid */
4232 .start = vma->vm_start,
4233 .len = vma->vm_end - vma->vm_start,
Peter Zijlstra3a0304e2010-02-26 10:33:41 +01004234 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004235 },
4236 };
4237
4238 perf_event_mmap_event(&mmap_event);
4239}
4240
4241/*
4242 * IRQ throttle logging
4243 */
4244
4245static void perf_log_throttle(struct perf_event *event, int enable)
4246{
4247 struct perf_output_handle handle;
4248 int ret;
4249
4250 struct {
4251 struct perf_event_header header;
4252 u64 time;
4253 u64 id;
4254 u64 stream_id;
4255 } throttle_event = {
4256 .header = {
4257 .type = PERF_RECORD_THROTTLE,
4258 .misc = 0,
4259 .size = sizeof(throttle_event),
4260 },
4261 .time = perf_clock(),
4262 .id = primary_event_id(event),
4263 .stream_id = event->id,
4264 };
4265
4266 if (enable)
4267 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4268
4269 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
4270 if (ret)
4271 return;
4272
4273 perf_output_put(&handle, throttle_event);
4274 perf_output_end(&handle);
4275}
4276
4277/*
4278 * Generic event overflow handling, sampling.
4279 */
4280
4281static int __perf_event_overflow(struct perf_event *event, int nmi,
4282 int throttle, struct perf_sample_data *data,
4283 struct pt_regs *regs)
4284{
4285 int events = atomic_read(&event->event_limit);
4286 struct hw_perf_event *hwc = &event->hw;
4287 int ret = 0;
4288
Peter Zijlstra96398822010-11-24 18:55:29 +01004289 /*
4290 * Non-sampling counters might still use the PMI to fold short
4291 * hardware counters, ignore those.
4292 */
4293 if (unlikely(!is_sampling_event(event)))
4294 return 0;
4295
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004296 if (!throttle) {
4297 hwc->interrupts++;
4298 } else {
4299 if (hwc->interrupts != MAX_INTERRUPTS) {
4300 hwc->interrupts++;
4301 if (HZ * hwc->interrupts >
4302 (u64)sysctl_perf_event_sample_rate) {
4303 hwc->interrupts = MAX_INTERRUPTS;
4304 perf_log_throttle(event, 0);
4305 ret = 1;
4306 }
4307 } else {
4308 /*
4309 * Keep re-disabling events even though on the previous
4310 * pass we disabled it - just in case we raced with a
4311 * sched-in and the event got enabled again:
4312 */
4313 ret = 1;
4314 }
4315 }
4316
4317 if (event->attr.freq) {
4318 u64 now = perf_clock();
Peter Zijlstraabd50712010-01-26 18:50:16 +01004319 s64 delta = now - hwc->freq_time_stamp;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004320
Peter Zijlstraabd50712010-01-26 18:50:16 +01004321 hwc->freq_time_stamp = now;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004322
Peter Zijlstraabd50712010-01-26 18:50:16 +01004323 if (delta > 0 && delta < 2*TICK_NSEC)
4324 perf_adjust_period(event, delta, hwc->last_period);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004325 }
4326
4327 /*
4328 * XXX event_limit might not quite work as expected on inherited
4329 * events
4330 */
4331
4332 event->pending_kill = POLL_IN;
4333 if (events && atomic_dec_and_test(&event->event_limit)) {
4334 ret = 1;
4335 event->pending_kill = POLL_HUP;
4336 if (nmi) {
4337 event->pending_disable = 1;
Peter Zijlstrae360adb2010-10-14 14:01:34 +08004338 irq_work_queue(&event->pending);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004339 } else
4340 perf_event_disable(event);
4341 }
4342
Peter Zijlstra453f19e2009-11-20 22:19:43 +01004343 if (event->overflow_handler)
4344 event->overflow_handler(event, nmi, data, regs);
4345 else
4346 perf_event_output(event, nmi, data, regs);
4347
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004348 return ret;
4349}
4350
4351int perf_event_overflow(struct perf_event *event, int nmi,
4352 struct perf_sample_data *data,
4353 struct pt_regs *regs)
4354{
4355 return __perf_event_overflow(event, nmi, 1, data, regs);
4356}
4357
4358/*
4359 * Generic software event infrastructure
4360 */
4361
Peter Zijlstrab28ab832010-09-06 14:48:15 +02004362struct swevent_htable {
4363 struct swevent_hlist *swevent_hlist;
4364 struct mutex hlist_mutex;
4365 int hlist_refcount;
4366
4367 /* Recursion avoidance in each contexts */
4368 int recursion[PERF_NR_CONTEXTS];
4369};
4370
4371static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
4372
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004373/*
4374 * We directly increment event->count and keep a second value in
4375 * event->hw.period_left to count intervals. This period event
4376 * is kept in the range [-sample_period, 0] so that we can use the
4377 * sign as trigger.
4378 */
4379
4380static u64 perf_swevent_set_period(struct perf_event *event)
4381{
4382 struct hw_perf_event *hwc = &event->hw;
4383 u64 period = hwc->last_period;
4384 u64 nr, offset;
4385 s64 old, val;
4386
4387 hwc->last_period = hwc->sample_period;
4388
4389again:
Peter Zijlstrae7850592010-05-21 14:43:08 +02004390 old = val = local64_read(&hwc->period_left);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004391 if (val < 0)
4392 return 0;
4393
4394 nr = div64_u64(period + val, period);
4395 offset = nr * period;
4396 val -= offset;
Peter Zijlstrae7850592010-05-21 14:43:08 +02004397 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004398 goto again;
4399
4400 return nr;
4401}
4402
Peter Zijlstra0cff7842009-11-20 22:19:44 +01004403static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004404 int nmi, struct perf_sample_data *data,
4405 struct pt_regs *regs)
4406{
4407 struct hw_perf_event *hwc = &event->hw;
4408 int throttle = 0;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004409
4410 data->period = event->hw.last_period;
Peter Zijlstra0cff7842009-11-20 22:19:44 +01004411 if (!overflow)
4412 overflow = perf_swevent_set_period(event);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004413
4414 if (hwc->interrupts == MAX_INTERRUPTS)
4415 return;
4416
4417 for (; overflow; overflow--) {
4418 if (__perf_event_overflow(event, nmi, throttle,
4419 data, regs)) {
4420 /*
4421 * We inhibit the overflow from happening when
4422 * hwc->interrupts == MAX_INTERRUPTS.
4423 */
4424 break;
4425 }
4426 throttle = 1;
4427 }
4428}
4429
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02004430static void perf_swevent_event(struct perf_event *event, u64 nr,
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004431 int nmi, struct perf_sample_data *data,
4432 struct pt_regs *regs)
4433{
4434 struct hw_perf_event *hwc = &event->hw;
4435
Peter Zijlstrae7850592010-05-21 14:43:08 +02004436 local64_add(nr, &event->count);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004437
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004438 if (!regs)
4439 return;
4440
Franck Bui-Huu6c7e5502010-11-23 16:21:43 +01004441 if (!is_sampling_event(event))
Peter Zijlstra0cff7842009-11-20 22:19:44 +01004442 return;
4443
4444 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4445 return perf_swevent_overflow(event, 1, nmi, data, regs);
4446
Peter Zijlstrae7850592010-05-21 14:43:08 +02004447 if (local64_add_negative(nr, &hwc->period_left))
Peter Zijlstra0cff7842009-11-20 22:19:44 +01004448 return;
4449
4450 perf_swevent_overflow(event, 0, nmi, data, regs);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004451}
4452
Frederic Weisbeckerf5ffe022009-11-23 15:42:34 +01004453static int perf_exclude_event(struct perf_event *event,
4454 struct pt_regs *regs)
4455{
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02004456 if (event->hw.state & PERF_HES_STOPPED)
4457 return 0;
4458
Frederic Weisbeckerf5ffe022009-11-23 15:42:34 +01004459 if (regs) {
4460 if (event->attr.exclude_user && user_mode(regs))
4461 return 1;
4462
4463 if (event->attr.exclude_kernel && !user_mode(regs))
4464 return 1;
4465 }
4466
4467 return 0;
4468}
4469
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004470static int perf_swevent_match(struct perf_event *event,
4471 enum perf_type_id type,
Li Zefan6fb29152009-10-15 11:21:42 +08004472 u32 event_id,
4473 struct perf_sample_data *data,
4474 struct pt_regs *regs)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004475{
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004476 if (event->attr.type != type)
4477 return 0;
Frederic Weisbeckerf5ffe022009-11-23 15:42:34 +01004478
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004479 if (event->attr.config != event_id)
4480 return 0;
4481
Frederic Weisbeckerf5ffe022009-11-23 15:42:34 +01004482 if (perf_exclude_event(event, regs))
4483 return 0;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004484
4485 return 1;
4486}
4487
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004488static inline u64 swevent_hash(u64 type, u32 event_id)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004489{
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004490 u64 val = event_id | (type << 32);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004491
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004492 return hash_64(val, SWEVENT_HLIST_BITS);
4493}
4494
Frederic Weisbecker49f135e2010-05-20 10:17:46 +02004495static inline struct hlist_head *
4496__find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004497{
Frederic Weisbecker49f135e2010-05-20 10:17:46 +02004498 u64 hash = swevent_hash(type, event_id);
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004499
Frederic Weisbecker49f135e2010-05-20 10:17:46 +02004500 return &hlist->heads[hash];
4501}
4502
4503/* For the read side: events when they trigger */
4504static inline struct hlist_head *
Peter Zijlstrab28ab832010-09-06 14:48:15 +02004505find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
Frederic Weisbecker49f135e2010-05-20 10:17:46 +02004506{
4507 struct swevent_hlist *hlist;
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004508
Peter Zijlstrab28ab832010-09-06 14:48:15 +02004509 hlist = rcu_dereference(swhash->swevent_hlist);
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004510 if (!hlist)
4511 return NULL;
4512
Frederic Weisbecker49f135e2010-05-20 10:17:46 +02004513 return __find_swevent_head(hlist, type, event_id);
4514}
4515
4516/* For the event head insertion and removal in the hlist */
4517static inline struct hlist_head *
Peter Zijlstrab28ab832010-09-06 14:48:15 +02004518find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
Frederic Weisbecker49f135e2010-05-20 10:17:46 +02004519{
4520 struct swevent_hlist *hlist;
4521 u32 event_id = event->attr.config;
4522 u64 type = event->attr.type;
4523
4524 /*
4525 * Event scheduling is always serialized against hlist allocation
4526 * and release. Which makes the protected version suitable here.
4527 * The context lock guarantees that.
4528 */
Peter Zijlstrab28ab832010-09-06 14:48:15 +02004529 hlist = rcu_dereference_protected(swhash->swevent_hlist,
Frederic Weisbecker49f135e2010-05-20 10:17:46 +02004530 lockdep_is_held(&event->ctx->lock));
4531 if (!hlist)
4532 return NULL;
4533
4534 return __find_swevent_head(hlist, type, event_id);
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004535}
4536
4537static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4538 u64 nr, int nmi,
4539 struct perf_sample_data *data,
4540 struct pt_regs *regs)
4541{
Peter Zijlstrab28ab832010-09-06 14:48:15 +02004542 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004543 struct perf_event *event;
4544 struct hlist_node *node;
4545 struct hlist_head *head;
4546
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004547 rcu_read_lock();
Peter Zijlstrab28ab832010-09-06 14:48:15 +02004548 head = find_swevent_head_rcu(swhash, type, event_id);
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004549 if (!head)
4550 goto end;
4551
4552 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
Li Zefan6fb29152009-10-15 11:21:42 +08004553 if (perf_swevent_match(event, type, event_id, data, regs))
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02004554 perf_swevent_event(event, nr, nmi, data, regs);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004555 }
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004556end:
4557 rcu_read_unlock();
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004558}
4559
Peter Zijlstra4ed7c922009-11-23 11:37:29 +01004560int perf_swevent_get_recursion_context(void)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004561{
Peter Zijlstrab28ab832010-09-06 14:48:15 +02004562 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
Frederic Weisbeckerce71b9d2009-11-22 05:26:55 +01004563
Peter Zijlstrab28ab832010-09-06 14:48:15 +02004564 return get_recursion_context(swhash->recursion);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004565}
Ingo Molnar645e8cc2009-11-22 12:20:19 +01004566EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004567
Peter Zijlstraecc55f82010-05-21 15:11:34 +02004568void inline perf_swevent_put_recursion_context(int rctx)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004569{
Peter Zijlstrab28ab832010-09-06 14:48:15 +02004570 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
Frederic Weisbecker927c7a92010-07-01 16:20:36 +02004571
Peter Zijlstrab28ab832010-09-06 14:48:15 +02004572 put_recursion_context(swhash->recursion, rctx);
Frederic Weisbeckerce71b9d2009-11-22 05:26:55 +01004573}
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004574
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004575void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4576 struct pt_regs *regs, u64 addr)
4577{
Ingo Molnara4234bf2009-11-23 10:57:59 +01004578 struct perf_sample_data data;
Peter Zijlstra4ed7c922009-11-23 11:37:29 +01004579 int rctx;
4580
Peter Zijlstra1c024eca2010-05-19 14:02:22 +02004581 preempt_disable_notrace();
Peter Zijlstra4ed7c922009-11-23 11:37:29 +01004582 rctx = perf_swevent_get_recursion_context();
4583 if (rctx < 0)
4584 return;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004585
Peter Zijlstradc1d6282010-03-03 15:55:04 +01004586 perf_sample_data_init(&data, addr);
Ingo Molnara4234bf2009-11-23 10:57:59 +01004587
4588 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
Peter Zijlstra4ed7c922009-11-23 11:37:29 +01004589
4590 perf_swevent_put_recursion_context(rctx);
Peter Zijlstra1c024eca2010-05-19 14:02:22 +02004591 preempt_enable_notrace();
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004592}
4593
4594static void perf_swevent_read(struct perf_event *event)
4595{
4596}
4597
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02004598static int perf_swevent_add(struct perf_event *event, int flags)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004599{
Peter Zijlstrab28ab832010-09-06 14:48:15 +02004600 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004601 struct hw_perf_event *hwc = &event->hw;
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004602 struct hlist_head *head;
4603
Franck Bui-Huu6c7e5502010-11-23 16:21:43 +01004604 if (is_sampling_event(event)) {
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004605 hwc->last_period = hwc->sample_period;
4606 perf_swevent_set_period(event);
4607 }
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004608
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02004609 hwc->state = !(flags & PERF_EF_START);
4610
Peter Zijlstrab28ab832010-09-06 14:48:15 +02004611 head = find_swevent_head(swhash, event);
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004612 if (WARN_ON_ONCE(!head))
4613 return -EINVAL;
4614
4615 hlist_add_head_rcu(&event->hlist_entry, head);
4616
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004617 return 0;
4618}
4619
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02004620static void perf_swevent_del(struct perf_event *event, int flags)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004621{
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004622 hlist_del_rcu(&event->hlist_entry);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004623}
4624
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02004625static void perf_swevent_start(struct perf_event *event, int flags)
Peter Zijlstrac6df8d52010-06-03 11:21:20 +02004626{
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02004627 event->hw.state = 0;
Peter Zijlstrac6df8d52010-06-03 11:21:20 +02004628}
4629
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02004630static void perf_swevent_stop(struct perf_event *event, int flags)
Peter Zijlstrac6df8d52010-06-03 11:21:20 +02004631{
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02004632 event->hw.state = PERF_HES_STOPPED;
Peter Zijlstrac6df8d52010-06-03 11:21:20 +02004633}
4634
Frederic Weisbecker49f135e2010-05-20 10:17:46 +02004635/* Deref the hlist from the update side */
4636static inline struct swevent_hlist *
Peter Zijlstrab28ab832010-09-06 14:48:15 +02004637swevent_hlist_deref(struct swevent_htable *swhash)
Frederic Weisbecker49f135e2010-05-20 10:17:46 +02004638{
Peter Zijlstrab28ab832010-09-06 14:48:15 +02004639 return rcu_dereference_protected(swhash->swevent_hlist,
4640 lockdep_is_held(&swhash->hlist_mutex));
Frederic Weisbecker49f135e2010-05-20 10:17:46 +02004641}
4642
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004643static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4644{
4645 struct swevent_hlist *hlist;
4646
4647 hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4648 kfree(hlist);
4649}
4650
Peter Zijlstrab28ab832010-09-06 14:48:15 +02004651static void swevent_hlist_release(struct swevent_htable *swhash)
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004652{
Peter Zijlstrab28ab832010-09-06 14:48:15 +02004653 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004654
Frederic Weisbecker49f135e2010-05-20 10:17:46 +02004655 if (!hlist)
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004656 return;
4657
Peter Zijlstrab28ab832010-09-06 14:48:15 +02004658 rcu_assign_pointer(swhash->swevent_hlist, NULL);
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004659 call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4660}
4661
4662static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4663{
Peter Zijlstrab28ab832010-09-06 14:48:15 +02004664 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004665
Peter Zijlstrab28ab832010-09-06 14:48:15 +02004666 mutex_lock(&swhash->hlist_mutex);
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004667
Peter Zijlstrab28ab832010-09-06 14:48:15 +02004668 if (!--swhash->hlist_refcount)
4669 swevent_hlist_release(swhash);
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004670
Peter Zijlstrab28ab832010-09-06 14:48:15 +02004671 mutex_unlock(&swhash->hlist_mutex);
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004672}
4673
4674static void swevent_hlist_put(struct perf_event *event)
4675{
4676 int cpu;
4677
4678 if (event->cpu != -1) {
4679 swevent_hlist_put_cpu(event, event->cpu);
4680 return;
4681 }
4682
4683 for_each_possible_cpu(cpu)
4684 swevent_hlist_put_cpu(event, cpu);
4685}
4686
4687static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4688{
Peter Zijlstrab28ab832010-09-06 14:48:15 +02004689 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004690 int err = 0;
4691
Peter Zijlstrab28ab832010-09-06 14:48:15 +02004692 mutex_lock(&swhash->hlist_mutex);
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004693
Peter Zijlstrab28ab832010-09-06 14:48:15 +02004694 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004695 struct swevent_hlist *hlist;
4696
4697 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4698 if (!hlist) {
4699 err = -ENOMEM;
4700 goto exit;
4701 }
Peter Zijlstrab28ab832010-09-06 14:48:15 +02004702 rcu_assign_pointer(swhash->swevent_hlist, hlist);
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004703 }
Peter Zijlstrab28ab832010-09-06 14:48:15 +02004704 swhash->hlist_refcount++;
Peter Zijlstra9ed60602010-06-11 17:36:35 +02004705exit:
Peter Zijlstrab28ab832010-09-06 14:48:15 +02004706 mutex_unlock(&swhash->hlist_mutex);
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004707
4708 return err;
4709}
4710
4711static int swevent_hlist_get(struct perf_event *event)
4712{
4713 int err;
4714 int cpu, failed_cpu;
4715
4716 if (event->cpu != -1)
4717 return swevent_hlist_get_cpu(event, event->cpu);
4718
4719 get_online_cpus();
4720 for_each_possible_cpu(cpu) {
4721 err = swevent_hlist_get_cpu(event, cpu);
4722 if (err) {
4723 failed_cpu = cpu;
4724 goto fail;
4725 }
4726 }
4727 put_online_cpus();
4728
4729 return 0;
Peter Zijlstra9ed60602010-06-11 17:36:35 +02004730fail:
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004731 for_each_possible_cpu(cpu) {
4732 if (cpu == failed_cpu)
4733 break;
4734 swevent_hlist_put_cpu(event, cpu);
4735 }
4736
4737 put_online_cpus();
4738 return err;
4739}
4740
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02004741atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
Frederic Weisbecker95476b62010-04-14 23:42:18 +02004742
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02004743static void sw_perf_event_destroy(struct perf_event *event)
4744{
4745 u64 event_id = event->attr.config;
4746
4747 WARN_ON(event->parent);
4748
Peter Zijlstra7e54a5a2010-10-14 22:32:45 +02004749 jump_label_dec(&perf_swevent_enabled[event_id]);
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02004750 swevent_hlist_put(event);
4751}
4752
4753static int perf_swevent_init(struct perf_event *event)
4754{
4755 int event_id = event->attr.config;
4756
4757 if (event->attr.type != PERF_TYPE_SOFTWARE)
4758 return -ENOENT;
4759
4760 switch (event_id) {
4761 case PERF_COUNT_SW_CPU_CLOCK:
4762 case PERF_COUNT_SW_TASK_CLOCK:
4763 return -ENOENT;
4764
4765 default:
4766 break;
4767 }
4768
4769 if (event_id > PERF_COUNT_SW_MAX)
4770 return -ENOENT;
4771
4772 if (!event->parent) {
4773 int err;
4774
4775 err = swevent_hlist_get(event);
4776 if (err)
4777 return err;
4778
Peter Zijlstra7e54a5a2010-10-14 22:32:45 +02004779 jump_label_inc(&perf_swevent_enabled[event_id]);
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02004780 event->destroy = sw_perf_event_destroy;
4781 }
4782
4783 return 0;
4784}
4785
4786static struct pmu perf_swevent = {
Peter Zijlstra89a1e182010-09-07 17:34:50 +02004787 .task_ctx_nr = perf_sw_context,
4788
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02004789 .event_init = perf_swevent_init,
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02004790 .add = perf_swevent_add,
4791 .del = perf_swevent_del,
4792 .start = perf_swevent_start,
4793 .stop = perf_swevent_stop,
Peter Zijlstra1c024eca2010-05-19 14:02:22 +02004794 .read = perf_swevent_read,
Peter Zijlstra1c024eca2010-05-19 14:02:22 +02004795};
Frederic Weisbecker95476b62010-04-14 23:42:18 +02004796
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02004797#ifdef CONFIG_EVENT_TRACING
4798
Peter Zijlstra1c024eca2010-05-19 14:02:22 +02004799static int perf_tp_filter_match(struct perf_event *event,
Frederic Weisbecker95476b62010-04-14 23:42:18 +02004800 struct perf_sample_data *data)
4801{
4802 void *record = data->raw->data;
4803
4804 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4805 return 1;
4806 return 0;
4807}
4808
Peter Zijlstra1c024eca2010-05-19 14:02:22 +02004809static int perf_tp_event_match(struct perf_event *event,
4810 struct perf_sample_data *data,
4811 struct pt_regs *regs)
4812{
Peter Zijlstra580d6072010-05-20 20:54:31 +02004813 /*
4814 * All tracepoints are from kernel-space.
4815 */
4816 if (event->attr.exclude_kernel)
Peter Zijlstra1c024eca2010-05-19 14:02:22 +02004817 return 0;
4818
4819 if (!perf_tp_filter_match(event, data))
4820 return 0;
4821
4822 return 1;
4823}
4824
4825void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
Peter Zijlstraecc55f82010-05-21 15:11:34 +02004826 struct pt_regs *regs, struct hlist_head *head, int rctx)
Peter Zijlstra1c024eca2010-05-19 14:02:22 +02004827{
4828 struct perf_sample_data data;
4829 struct perf_event *event;
4830 struct hlist_node *node;
4831
4832 struct perf_raw_record raw = {
4833 .size = entry_size,
4834 .data = record,
4835 };
4836
4837 perf_sample_data_init(&data, addr);
4838 data.raw = &raw;
4839
Peter Zijlstra1c024eca2010-05-19 14:02:22 +02004840 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4841 if (perf_tp_event_match(event, &data, regs))
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02004842 perf_swevent_event(event, count, 1, &data, regs);
Peter Zijlstra1c024eca2010-05-19 14:02:22 +02004843 }
Peter Zijlstraecc55f82010-05-21 15:11:34 +02004844
4845 perf_swevent_put_recursion_context(rctx);
Peter Zijlstra1c024eca2010-05-19 14:02:22 +02004846}
4847EXPORT_SYMBOL_GPL(perf_tp_event);
4848
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004849static void tp_perf_event_destroy(struct perf_event *event)
4850{
Peter Zijlstra1c024eca2010-05-19 14:02:22 +02004851 perf_trace_destroy(event);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004852}
4853
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02004854static int perf_tp_event_init(struct perf_event *event)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004855{
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02004856 int err;
4857
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02004858 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4859 return -ENOENT;
4860
Peter Zijlstra1c024eca2010-05-19 14:02:22 +02004861 err = perf_trace_init(event);
4862 if (err)
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02004863 return err;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004864
4865 event->destroy = tp_perf_event_destroy;
4866
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02004867 return 0;
4868}
4869
4870static struct pmu perf_tracepoint = {
Peter Zijlstra89a1e182010-09-07 17:34:50 +02004871 .task_ctx_nr = perf_sw_context,
4872
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02004873 .event_init = perf_tp_event_init,
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02004874 .add = perf_trace_add,
4875 .del = perf_trace_del,
4876 .start = perf_swevent_start,
4877 .stop = perf_swevent_stop,
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02004878 .read = perf_swevent_read,
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02004879};
4880
4881static inline void perf_tp_register(void)
4882{
4883 perf_pmu_register(&perf_tracepoint);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004884}
Li Zefan6fb29152009-10-15 11:21:42 +08004885
4886static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4887{
4888 char *filter_str;
4889 int ret;
4890
4891 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4892 return -EINVAL;
4893
4894 filter_str = strndup_user(arg, PAGE_SIZE);
4895 if (IS_ERR(filter_str))
4896 return PTR_ERR(filter_str);
4897
4898 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4899
4900 kfree(filter_str);
4901 return ret;
4902}
4903
4904static void perf_event_free_filter(struct perf_event *event)
4905{
4906 ftrace_profile_free_filter(event);
4907}
4908
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004909#else
Li Zefan6fb29152009-10-15 11:21:42 +08004910
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02004911static inline void perf_tp_register(void)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004912{
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004913}
Li Zefan6fb29152009-10-15 11:21:42 +08004914
4915static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4916{
4917 return -ENOENT;
4918}
4919
4920static void perf_event_free_filter(struct perf_event *event)
4921{
4922}
4923
Li Zefan07b139c2009-12-21 14:27:35 +08004924#endif /* CONFIG_EVENT_TRACING */
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004925
Frederic Weisbecker24f1e32c2009-09-09 19:22:48 +02004926#ifdef CONFIG_HAVE_HW_BREAKPOINT
Frederic Weisbeckerf5ffe022009-11-23 15:42:34 +01004927void perf_bp_event(struct perf_event *bp, void *data)
Frederic Weisbecker24f1e32c2009-09-09 19:22:48 +02004928{
Frederic Weisbeckerf5ffe022009-11-23 15:42:34 +01004929 struct perf_sample_data sample;
4930 struct pt_regs *regs = data;
4931
Peter Zijlstradc1d6282010-03-03 15:55:04 +01004932 perf_sample_data_init(&sample, bp->attr.bp_addr);
Frederic Weisbeckerf5ffe022009-11-23 15:42:34 +01004933
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02004934 if (!bp->hw.state && !perf_exclude_event(bp, regs))
4935 perf_swevent_event(bp, 1, 1, &sample, regs);
Frederic Weisbecker24f1e32c2009-09-09 19:22:48 +02004936}
4937#endif
4938
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02004939/*
4940 * hrtimer based swevent callback
4941 */
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004942
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02004943static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004944{
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02004945 enum hrtimer_restart ret = HRTIMER_RESTART;
4946 struct perf_sample_data data;
4947 struct pt_regs *regs;
4948 struct perf_event *event;
4949 u64 period;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004950
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02004951 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4952 event->pmu->read(event);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004953
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02004954 perf_sample_data_init(&data, 0);
4955 data.period = event->hw.last_period;
4956 regs = get_irq_regs();
4957
4958 if (regs && !perf_exclude_event(event, regs)) {
4959 if (!(event->attr.exclude_idle && current->pid == 0))
4960 if (perf_event_overflow(event, 0, &data, regs))
4961 ret = HRTIMER_NORESTART;
4962 }
4963
4964 period = max_t(u64, 10000, event->hw.sample_period);
4965 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4966
4967 return ret;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004968}
4969
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02004970static void perf_swevent_start_hrtimer(struct perf_event *event)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02004971{
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02004972 struct hw_perf_event *hwc = &event->hw;
Franck Bui-Huu5d508e82010-11-23 16:21:45 +01004973 s64 period;
4974
4975 if (!is_sampling_event(event))
4976 return;
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02004977
4978 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4979 hwc->hrtimer.function = perf_swevent_hrtimer;
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02004980
Franck Bui-Huu5d508e82010-11-23 16:21:45 +01004981 period = local64_read(&hwc->period_left);
4982 if (period) {
4983 if (period < 0)
4984 period = 10000;
Peter Zijlstrafa407f32010-06-24 12:35:12 +02004985
Franck Bui-Huu5d508e82010-11-23 16:21:45 +01004986 local64_set(&hwc->period_left, 0);
4987 } else {
4988 period = max_t(u64, 10000, hwc->sample_period);
4989 }
4990 __hrtimer_start_range_ns(&hwc->hrtimer,
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02004991 ns_to_ktime(period), 0,
Peter Zijlstrab5ab4cd2010-09-06 16:32:21 +02004992 HRTIMER_MODE_REL_PINNED, 0);
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02004993}
4994
4995static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4996{
4997 struct hw_perf_event *hwc = &event->hw;
4998
Franck Bui-Huu6c7e5502010-11-23 16:21:43 +01004999 if (is_sampling_event(event)) {
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005000 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
Peter Zijlstrafa407f32010-06-24 12:35:12 +02005001 local64_set(&hwc->period_left, ktime_to_ns(remaining));
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005002
5003 hrtimer_cancel(&hwc->hrtimer);
5004 }
5005}
5006
5007/*
5008 * Software event: cpu wall time clock
5009 */
5010
5011static void cpu_clock_event_update(struct perf_event *event)
5012{
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005013 s64 prev;
5014 u64 now;
5015
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02005016 now = local_clock();
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005017 prev = local64_xchg(&event->hw.prev_count, now);
5018 local64_add(now - prev, &event->count);
5019}
5020
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02005021static void cpu_clock_event_start(struct perf_event *event, int flags)
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005022{
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02005023 local64_set(&event->hw.prev_count, local_clock());
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005024 perf_swevent_start_hrtimer(event);
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02005025}
5026
5027static void cpu_clock_event_stop(struct perf_event *event, int flags)
5028{
5029 perf_swevent_cancel_hrtimer(event);
5030 cpu_clock_event_update(event);
5031}
5032
5033static int cpu_clock_event_add(struct perf_event *event, int flags)
5034{
5035 if (flags & PERF_EF_START)
5036 cpu_clock_event_start(event, flags);
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005037
5038 return 0;
5039}
5040
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02005041static void cpu_clock_event_del(struct perf_event *event, int flags)
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005042{
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02005043 cpu_clock_event_stop(event, flags);
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005044}
5045
5046static void cpu_clock_event_read(struct perf_event *event)
5047{
5048 cpu_clock_event_update(event);
5049}
5050
5051static int cpu_clock_event_init(struct perf_event *event)
5052{
5053 if (event->attr.type != PERF_TYPE_SOFTWARE)
5054 return -ENOENT;
5055
5056 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
5057 return -ENOENT;
5058
5059 return 0;
5060}
5061
5062static struct pmu perf_cpu_clock = {
Peter Zijlstra89a1e182010-09-07 17:34:50 +02005063 .task_ctx_nr = perf_sw_context,
5064
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005065 .event_init = cpu_clock_event_init,
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02005066 .add = cpu_clock_event_add,
5067 .del = cpu_clock_event_del,
5068 .start = cpu_clock_event_start,
5069 .stop = cpu_clock_event_stop,
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005070 .read = cpu_clock_event_read,
5071};
5072
5073/*
5074 * Software event: task time clock
5075 */
5076
5077static void task_clock_event_update(struct perf_event *event, u64 now)
5078{
5079 u64 prev;
5080 s64 delta;
5081
5082 prev = local64_xchg(&event->hw.prev_count, now);
5083 delta = now - prev;
5084 local64_add(delta, &event->count);
5085}
5086
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02005087static void task_clock_event_start(struct perf_event *event, int flags)
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005088{
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02005089 local64_set(&event->hw.prev_count, event->ctx->time);
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005090 perf_swevent_start_hrtimer(event);
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02005091}
5092
5093static void task_clock_event_stop(struct perf_event *event, int flags)
5094{
5095 perf_swevent_cancel_hrtimer(event);
5096 task_clock_event_update(event, event->ctx->time);
5097}
5098
5099static int task_clock_event_add(struct perf_event *event, int flags)
5100{
5101 if (flags & PERF_EF_START)
5102 task_clock_event_start(event, flags);
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005103
5104 return 0;
5105}
5106
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02005107static void task_clock_event_del(struct perf_event *event, int flags)
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005108{
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02005109 task_clock_event_stop(event, PERF_EF_UPDATE);
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005110}
5111
5112static void task_clock_event_read(struct perf_event *event)
5113{
5114 u64 time;
5115
5116 if (!in_nmi()) {
5117 update_context_time(event->ctx);
5118 time = event->ctx->time;
5119 } else {
5120 u64 now = perf_clock();
5121 u64 delta = now - event->ctx->timestamp;
5122 time = event->ctx->time + delta;
5123 }
5124
5125 task_clock_event_update(event, time);
5126}
5127
5128static int task_clock_event_init(struct perf_event *event)
5129{
5130 if (event->attr.type != PERF_TYPE_SOFTWARE)
5131 return -ENOENT;
5132
5133 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5134 return -ENOENT;
5135
5136 return 0;
5137}
5138
5139static struct pmu perf_task_clock = {
Peter Zijlstra89a1e182010-09-07 17:34:50 +02005140 .task_ctx_nr = perf_sw_context,
5141
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005142 .event_init = task_clock_event_init,
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +02005143 .add = task_clock_event_add,
5144 .del = task_clock_event_del,
5145 .start = task_clock_event_start,
5146 .stop = task_clock_event_stop,
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005147 .read = task_clock_event_read,
5148};
5149
Peter Zijlstraad5133b2010-06-15 12:22:39 +02005150static void perf_pmu_nop_void(struct pmu *pmu)
5151{
5152}
5153
5154static int perf_pmu_nop_int(struct pmu *pmu)
5155{
5156 return 0;
5157}
5158
5159static void perf_pmu_start_txn(struct pmu *pmu)
5160{
5161 perf_pmu_disable(pmu);
5162}
5163
5164static int perf_pmu_commit_txn(struct pmu *pmu)
5165{
5166 perf_pmu_enable(pmu);
5167 return 0;
5168}
5169
5170static void perf_pmu_cancel_txn(struct pmu *pmu)
5171{
5172 perf_pmu_enable(pmu);
5173}
5174
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02005175/*
5176 * Ensures all contexts with the same task_ctx_nr have the same
5177 * pmu_cpu_context too.
5178 */
5179static void *find_pmu_context(int ctxn)
5180{
5181 struct pmu *pmu;
5182
5183 if (ctxn < 0)
5184 return NULL;
5185
5186 list_for_each_entry(pmu, &pmus, entry) {
5187 if (pmu->task_ctx_nr == ctxn)
5188 return pmu->pmu_cpu_context;
5189 }
5190
5191 return NULL;
5192}
5193
5194static void free_pmu_context(void * __percpu cpu_context)
5195{
5196 struct pmu *pmu;
5197
5198 mutex_lock(&pmus_lock);
5199 /*
5200 * Like a real lame refcount.
5201 */
5202 list_for_each_entry(pmu, &pmus, entry) {
5203 if (pmu->pmu_cpu_context == cpu_context)
5204 goto out;
5205 }
5206
5207 free_percpu(cpu_context);
5208out:
5209 mutex_unlock(&pmus_lock);
5210}
5211
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005212int perf_pmu_register(struct pmu *pmu)
5213{
Peter Zijlstra108b02c2010-09-06 14:32:03 +02005214 int cpu, ret;
Peter Zijlstra33696fc2010-06-14 08:49:00 +02005215
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005216 mutex_lock(&pmus_lock);
Peter Zijlstra33696fc2010-06-14 08:49:00 +02005217 ret = -ENOMEM;
5218 pmu->pmu_disable_count = alloc_percpu(int);
5219 if (!pmu->pmu_disable_count)
5220 goto unlock;
Peter Zijlstraad5133b2010-06-15 12:22:39 +02005221
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02005222 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
5223 if (pmu->pmu_cpu_context)
5224 goto got_cpu_context;
5225
Peter Zijlstra108b02c2010-09-06 14:32:03 +02005226 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
5227 if (!pmu->pmu_cpu_context)
5228 goto free_pdc;
5229
5230 for_each_possible_cpu(cpu) {
5231 struct perf_cpu_context *cpuctx;
5232
5233 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
Peter Zijlstraeb184472010-09-07 15:55:13 +02005234 __perf_event_init_context(&cpuctx->ctx);
Peter Zijlstrab04243e2010-09-17 11:28:48 +02005235 cpuctx->ctx.type = cpu_context;
Peter Zijlstra108b02c2010-09-06 14:32:03 +02005236 cpuctx->ctx.pmu = pmu;
Peter Zijlstrae9d2b062010-09-17 11:28:50 +02005237 cpuctx->jiffies_interval = 1;
5238 INIT_LIST_HEAD(&cpuctx->rotation_list);
Peter Zijlstra108b02c2010-09-06 14:32:03 +02005239 }
5240
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02005241got_cpu_context:
Peter Zijlstraad5133b2010-06-15 12:22:39 +02005242 if (!pmu->start_txn) {
5243 if (pmu->pmu_enable) {
5244 /*
5245 * If we have pmu_enable/pmu_disable calls, install
5246 * transaction stubs that use that to try and batch
5247 * hardware accesses.
5248 */
5249 pmu->start_txn = perf_pmu_start_txn;
5250 pmu->commit_txn = perf_pmu_commit_txn;
5251 pmu->cancel_txn = perf_pmu_cancel_txn;
5252 } else {
5253 pmu->start_txn = perf_pmu_nop_void;
5254 pmu->commit_txn = perf_pmu_nop_int;
5255 pmu->cancel_txn = perf_pmu_nop_void;
5256 }
5257 }
5258
5259 if (!pmu->pmu_enable) {
5260 pmu->pmu_enable = perf_pmu_nop_void;
5261 pmu->pmu_disable = perf_pmu_nop_void;
5262 }
5263
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005264 list_add_rcu(&pmu->entry, &pmus);
Peter Zijlstra33696fc2010-06-14 08:49:00 +02005265 ret = 0;
5266unlock:
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005267 mutex_unlock(&pmus_lock);
5268
Peter Zijlstra33696fc2010-06-14 08:49:00 +02005269 return ret;
Peter Zijlstra108b02c2010-09-06 14:32:03 +02005270
5271free_pdc:
5272 free_percpu(pmu->pmu_disable_count);
5273 goto unlock;
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005274}
5275
5276void perf_pmu_unregister(struct pmu *pmu)
5277{
5278 mutex_lock(&pmus_lock);
5279 list_del_rcu(&pmu->entry);
5280 mutex_unlock(&pmus_lock);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005281
5282 /*
Peter Zijlstracde8e882010-09-13 11:06:55 +02005283 * We dereference the pmu list under both SRCU and regular RCU, so
5284 * synchronize against both of those.
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005285 */
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005286 synchronize_srcu(&pmus_srcu);
Peter Zijlstracde8e882010-09-13 11:06:55 +02005287 synchronize_rcu();
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005288
Peter Zijlstra33696fc2010-06-14 08:49:00 +02005289 free_percpu(pmu->pmu_disable_count);
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02005290 free_pmu_context(pmu->pmu_cpu_context);
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005291}
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005292
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005293struct pmu *perf_init_event(struct perf_event *event)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005294{
Peter Zijlstra51b0fe32010-06-11 13:35:57 +02005295 struct pmu *pmu = NULL;
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005296 int idx;
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02005297
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005298 idx = srcu_read_lock(&pmus_srcu);
5299 list_for_each_entry_rcu(pmu, &pmus, entry) {
5300 int ret = pmu->event_init(event);
5301 if (!ret)
Peter Zijlstrae5f4d332010-09-10 17:38:06 +02005302 goto unlock;
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02005303
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005304 if (ret != -ENOENT) {
5305 pmu = ERR_PTR(ret);
Peter Zijlstrae5f4d332010-09-10 17:38:06 +02005306 goto unlock;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005307 }
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005308 }
Peter Zijlstrae5f4d332010-09-10 17:38:06 +02005309 pmu = ERR_PTR(-ENOENT);
5310unlock:
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005311 srcu_read_unlock(&pmus_srcu, idx);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005312
5313 return pmu;
5314}
5315
5316/*
5317 * Allocate and initialize a event structure
5318 */
5319static struct perf_event *
Peter Zijlstrac3f00c72010-08-18 14:37:15 +02005320perf_event_alloc(struct perf_event_attr *attr, int cpu,
Peter Zijlstrad580ff82010-10-14 17:43:23 +02005321 struct task_struct *task,
5322 struct perf_event *group_leader,
5323 struct perf_event *parent_event,
5324 perf_overflow_handler_t overflow_handler)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005325{
Peter Zijlstra51b0fe32010-06-11 13:35:57 +02005326 struct pmu *pmu;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005327 struct perf_event *event;
5328 struct hw_perf_event *hwc;
5329 long err;
5330
Peter Zijlstrac3f00c72010-08-18 14:37:15 +02005331 event = kzalloc(sizeof(*event), GFP_KERNEL);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005332 if (!event)
5333 return ERR_PTR(-ENOMEM);
5334
5335 /*
5336 * Single events are their own group leaders, with an
5337 * empty sibling list:
5338 */
5339 if (!group_leader)
5340 group_leader = event;
5341
5342 mutex_init(&event->child_mutex);
5343 INIT_LIST_HEAD(&event->child_list);
5344
5345 INIT_LIST_HEAD(&event->group_entry);
5346 INIT_LIST_HEAD(&event->event_entry);
5347 INIT_LIST_HEAD(&event->sibling_list);
5348 init_waitqueue_head(&event->waitq);
Peter Zijlstrae360adb2010-10-14 14:01:34 +08005349 init_irq_work(&event->pending, perf_pending_event);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005350
5351 mutex_init(&event->mmap_mutex);
5352
5353 event->cpu = cpu;
5354 event->attr = *attr;
5355 event->group_leader = group_leader;
5356 event->pmu = NULL;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005357 event->oncpu = -1;
5358
5359 event->parent = parent_event;
5360
5361 event->ns = get_pid_ns(current->nsproxy->pid_ns);
5362 event->id = atomic64_inc_return(&perf_event_id);
5363
5364 event->state = PERF_EVENT_STATE_INACTIVE;
5365
Peter Zijlstrad580ff82010-10-14 17:43:23 +02005366 if (task) {
5367 event->attach_state = PERF_ATTACH_TASK;
5368#ifdef CONFIG_HAVE_HW_BREAKPOINT
5369 /*
5370 * hw_breakpoint is a bit difficult here..
5371 */
5372 if (attr->type == PERF_TYPE_BREAKPOINT)
5373 event->hw.bp_target = task;
5374#endif
5375 }
5376
Frederic Weisbeckerb326e952009-12-05 09:44:31 +01005377 if (!overflow_handler && parent_event)
5378 overflow_handler = parent_event->overflow_handler;
Frederic Weisbecker97eaf532009-10-18 15:33:50 +02005379
Frederic Weisbeckerb326e952009-12-05 09:44:31 +01005380 event->overflow_handler = overflow_handler;
Frederic Weisbecker97eaf532009-10-18 15:33:50 +02005381
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005382 if (attr->disabled)
5383 event->state = PERF_EVENT_STATE_OFF;
5384
5385 pmu = NULL;
5386
5387 hwc = &event->hw;
5388 hwc->sample_period = attr->sample_period;
5389 if (attr->freq && attr->sample_freq)
5390 hwc->sample_period = 1;
5391 hwc->last_period = hwc->sample_period;
5392
Peter Zijlstrae7850592010-05-21 14:43:08 +02005393 local64_set(&hwc->period_left, hwc->sample_period);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005394
5395 /*
5396 * we currently do not support PERF_FORMAT_GROUP on inherited events
5397 */
5398 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
5399 goto done;
5400
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02005401 pmu = perf_init_event(event);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005402
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005403done:
5404 err = 0;
5405 if (!pmu)
5406 err = -EINVAL;
5407 else if (IS_ERR(pmu))
5408 err = PTR_ERR(pmu);
5409
5410 if (err) {
5411 if (event->ns)
5412 put_pid_ns(event->ns);
5413 kfree(event);
5414 return ERR_PTR(err);
5415 }
5416
5417 event->pmu = pmu;
5418
5419 if (!event->parent) {
Peter Zijlstra82cd6de2010-10-14 17:57:23 +02005420 if (event->attach_state & PERF_ATTACH_TASK)
5421 jump_label_inc(&perf_task_events);
Eric B Munson3af9e852010-05-18 15:30:49 +01005422 if (event->attr.mmap || event->attr.mmap_data)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005423 atomic_inc(&nr_mmap_events);
5424 if (event->attr.comm)
5425 atomic_inc(&nr_comm_events);
5426 if (event->attr.task)
5427 atomic_inc(&nr_task_events);
Frederic Weisbecker927c7a92010-07-01 16:20:36 +02005428 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
5429 err = get_callchain_buffers();
5430 if (err) {
5431 free_event(event);
5432 return ERR_PTR(err);
5433 }
5434 }
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005435 }
5436
5437 return event;
5438}
5439
5440static int perf_copy_attr(struct perf_event_attr __user *uattr,
5441 struct perf_event_attr *attr)
5442{
5443 u32 size;
5444 int ret;
5445
5446 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
5447 return -EFAULT;
5448
5449 /*
5450 * zero the full structure, so that a short copy will be nice.
5451 */
5452 memset(attr, 0, sizeof(*attr));
5453
5454 ret = get_user(size, &uattr->size);
5455 if (ret)
5456 return ret;
5457
5458 if (size > PAGE_SIZE) /* silly large */
5459 goto err_size;
5460
5461 if (!size) /* abi compat */
5462 size = PERF_ATTR_SIZE_VER0;
5463
5464 if (size < PERF_ATTR_SIZE_VER0)
5465 goto err_size;
5466
5467 /*
5468 * If we're handed a bigger struct than we know of,
5469 * ensure all the unknown bits are 0 - i.e. new
5470 * user-space does not rely on any kernel feature
5471 * extensions we dont know about yet.
5472 */
5473 if (size > sizeof(*attr)) {
5474 unsigned char __user *addr;
5475 unsigned char __user *end;
5476 unsigned char val;
5477
5478 addr = (void __user *)uattr + sizeof(*attr);
5479 end = (void __user *)uattr + size;
5480
5481 for (; addr < end; addr++) {
5482 ret = get_user(val, addr);
5483 if (ret)
5484 return ret;
5485 if (val)
5486 goto err_size;
5487 }
5488 size = sizeof(*attr);
5489 }
5490
5491 ret = copy_from_user(attr, uattr, size);
5492 if (ret)
5493 return -EFAULT;
5494
5495 /*
5496 * If the type exists, the corresponding creation will verify
5497 * the attr->config.
5498 */
5499 if (attr->type >= PERF_TYPE_MAX)
5500 return -EINVAL;
5501
Mahesh Salgaonkarcd757642010-01-30 10:25:18 +05305502 if (attr->__reserved_1)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005503 return -EINVAL;
5504
5505 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
5506 return -EINVAL;
5507
5508 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
5509 return -EINVAL;
5510
5511out:
5512 return ret;
5513
5514err_size:
5515 put_user(sizeof(*attr), &uattr->size);
5516 ret = -E2BIG;
5517 goto out;
5518}
5519
Peter Zijlstraac9721f2010-05-27 12:54:41 +02005520static int
5521perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005522{
Peter Zijlstraca5135e2010-05-28 19:33:23 +02005523 struct perf_buffer *buffer = NULL, *old_buffer = NULL;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005524 int ret = -EINVAL;
5525
Peter Zijlstraac9721f2010-05-27 12:54:41 +02005526 if (!output_event)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005527 goto set;
5528
Peter Zijlstraac9721f2010-05-27 12:54:41 +02005529 /* don't allow circular references */
5530 if (event == output_event)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005531 goto out;
5532
Peter Zijlstra0f139302010-05-20 14:35:15 +02005533 /*
5534 * Don't allow cross-cpu buffers
5535 */
5536 if (output_event->cpu != event->cpu)
5537 goto out;
5538
5539 /*
5540 * If its not a per-cpu buffer, it must be the same task.
5541 */
5542 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
5543 goto out;
5544
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005545set:
5546 mutex_lock(&event->mmap_mutex);
Peter Zijlstraac9721f2010-05-27 12:54:41 +02005547 /* Can't redirect output if we've got an active mmap() */
5548 if (atomic_read(&event->mmap_count))
5549 goto unlock;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005550
Peter Zijlstraac9721f2010-05-27 12:54:41 +02005551 if (output_event) {
5552 /* get the buffer we want to redirect to */
Peter Zijlstraca5135e2010-05-28 19:33:23 +02005553 buffer = perf_buffer_get(output_event);
5554 if (!buffer)
Peter Zijlstraac9721f2010-05-27 12:54:41 +02005555 goto unlock;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005556 }
5557
Peter Zijlstraca5135e2010-05-28 19:33:23 +02005558 old_buffer = event->buffer;
5559 rcu_assign_pointer(event->buffer, buffer);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005560 ret = 0;
Peter Zijlstraac9721f2010-05-27 12:54:41 +02005561unlock:
5562 mutex_unlock(&event->mmap_mutex);
5563
Peter Zijlstraca5135e2010-05-28 19:33:23 +02005564 if (old_buffer)
5565 perf_buffer_put(old_buffer);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005566out:
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005567 return ret;
5568}
5569
5570/**
5571 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5572 *
5573 * @attr_uptr: event_id type attributes for monitoring/sampling
5574 * @pid: target pid
5575 * @cpu: target cpu
5576 * @group_fd: group leader event fd
5577 */
5578SYSCALL_DEFINE5(perf_event_open,
5579 struct perf_event_attr __user *, attr_uptr,
5580 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
5581{
Peter Zijlstrab04243e2010-09-17 11:28:48 +02005582 struct perf_event *group_leader = NULL, *output_event = NULL;
5583 struct perf_event *event, *sibling;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005584 struct perf_event_attr attr;
5585 struct perf_event_context *ctx;
5586 struct file *event_file = NULL;
5587 struct file *group_file = NULL;
Matt Helsley38a81da2010-09-13 13:01:20 -07005588 struct task_struct *task = NULL;
Peter Zijlstra89a1e182010-09-07 17:34:50 +02005589 struct pmu *pmu;
Al Viroea635c62010-05-26 17:40:29 -04005590 int event_fd;
Peter Zijlstrab04243e2010-09-17 11:28:48 +02005591 int move_group = 0;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005592 int fput_needed = 0;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005593 int err;
5594
5595 /* for future expandability... */
5596 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
5597 return -EINVAL;
5598
5599 err = perf_copy_attr(attr_uptr, &attr);
5600 if (err)
5601 return err;
5602
5603 if (!attr.exclude_kernel) {
5604 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
5605 return -EACCES;
5606 }
5607
5608 if (attr.freq) {
5609 if (attr.sample_freq > sysctl_perf_event_sample_rate)
5610 return -EINVAL;
5611 }
5612
Al Viroea635c62010-05-26 17:40:29 -04005613 event_fd = get_unused_fd_flags(O_RDWR);
5614 if (event_fd < 0)
5615 return event_fd;
5616
Peter Zijlstraac9721f2010-05-27 12:54:41 +02005617 if (group_fd != -1) {
5618 group_leader = perf_fget_light(group_fd, &fput_needed);
5619 if (IS_ERR(group_leader)) {
5620 err = PTR_ERR(group_leader);
Stephane Eraniand14b12d2010-09-17 11:28:47 +02005621 goto err_fd;
Peter Zijlstraac9721f2010-05-27 12:54:41 +02005622 }
5623 group_file = group_leader->filp;
5624 if (flags & PERF_FLAG_FD_OUTPUT)
5625 output_event = group_leader;
5626 if (flags & PERF_FLAG_FD_NO_GROUP)
5627 group_leader = NULL;
5628 }
5629
Peter Zijlstrac6be5a52010-10-14 16:59:46 +02005630 if (pid != -1) {
5631 task = find_lively_task_by_vpid(pid);
5632 if (IS_ERR(task)) {
5633 err = PTR_ERR(task);
5634 goto err_group_fd;
5635 }
5636 }
5637
Peter Zijlstrad580ff82010-10-14 17:43:23 +02005638 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, NULL);
Stephane Eraniand14b12d2010-09-17 11:28:47 +02005639 if (IS_ERR(event)) {
5640 err = PTR_ERR(event);
Peter Zijlstrac6be5a52010-10-14 16:59:46 +02005641 goto err_task;
Stephane Eraniand14b12d2010-09-17 11:28:47 +02005642 }
5643
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005644 /*
Peter Zijlstra89a1e182010-09-07 17:34:50 +02005645 * Special case software events and allow them to be part of
5646 * any hardware group.
5647 */
5648 pmu = event->pmu;
Peter Zijlstrab04243e2010-09-17 11:28:48 +02005649
5650 if (group_leader &&
5651 (is_software_event(event) != is_software_event(group_leader))) {
5652 if (is_software_event(event)) {
5653 /*
5654 * If event and group_leader are not both a software
5655 * event, and event is, then group leader is not.
5656 *
5657 * Allow the addition of software events to !software
5658 * groups, this is safe because software events never
5659 * fail to schedule.
5660 */
5661 pmu = group_leader->pmu;
5662 } else if (is_software_event(group_leader) &&
5663 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
5664 /*
5665 * In case the group is a pure software group, and we
5666 * try to add a hardware event, move the whole group to
5667 * the hardware context.
5668 */
5669 move_group = 1;
5670 }
5671 }
Peter Zijlstra89a1e182010-09-07 17:34:50 +02005672
5673 /*
5674 * Get the target context (task or percpu):
5675 */
Matt Helsley38a81da2010-09-13 13:01:20 -07005676 ctx = find_get_context(pmu, task, cpu);
Peter Zijlstra89a1e182010-09-07 17:34:50 +02005677 if (IS_ERR(ctx)) {
5678 err = PTR_ERR(ctx);
Peter Zijlstrac6be5a52010-10-14 16:59:46 +02005679 goto err_alloc;
Peter Zijlstra89a1e182010-09-07 17:34:50 +02005680 }
5681
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005682 /*
5683 * Look up the group leader (we will attach this event to it):
5684 */
Peter Zijlstraac9721f2010-05-27 12:54:41 +02005685 if (group_leader) {
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005686 err = -EINVAL;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005687
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005688 /*
5689 * Do not allow a recursive hierarchy (this new sibling
5690 * becoming part of another group-sibling):
5691 */
5692 if (group_leader->group_leader != group_leader)
Peter Zijlstrac3f00c72010-08-18 14:37:15 +02005693 goto err_context;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005694 /*
5695 * Do not allow to attach to a group in a different
5696 * task or CPU context:
5697 */
Peter Zijlstrab04243e2010-09-17 11:28:48 +02005698 if (move_group) {
5699 if (group_leader->ctx->type != ctx->type)
5700 goto err_context;
5701 } else {
5702 if (group_leader->ctx != ctx)
5703 goto err_context;
5704 }
5705
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005706 /*
5707 * Only a group leader can be exclusive or pinned
5708 */
5709 if (attr.exclusive || attr.pinned)
Peter Zijlstrac3f00c72010-08-18 14:37:15 +02005710 goto err_context;
Peter Zijlstraac9721f2010-05-27 12:54:41 +02005711 }
5712
5713 if (output_event) {
5714 err = perf_event_set_output(event, output_event);
5715 if (err)
Peter Zijlstrac3f00c72010-08-18 14:37:15 +02005716 goto err_context;
Peter Zijlstraac9721f2010-05-27 12:54:41 +02005717 }
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005718
Al Viroea635c62010-05-26 17:40:29 -04005719 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
5720 if (IS_ERR(event_file)) {
5721 err = PTR_ERR(event_file);
Peter Zijlstrac3f00c72010-08-18 14:37:15 +02005722 goto err_context;
Al Viroea635c62010-05-26 17:40:29 -04005723 }
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005724
Peter Zijlstrab04243e2010-09-17 11:28:48 +02005725 if (move_group) {
5726 struct perf_event_context *gctx = group_leader->ctx;
5727
5728 mutex_lock(&gctx->mutex);
5729 perf_event_remove_from_context(group_leader);
5730 list_for_each_entry(sibling, &group_leader->sibling_list,
5731 group_entry) {
5732 perf_event_remove_from_context(sibling);
5733 put_ctx(gctx);
5734 }
5735 mutex_unlock(&gctx->mutex);
5736 put_ctx(gctx);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005737 }
5738
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005739 event->filp = event_file;
5740 WARN_ON_ONCE(ctx->parent_ctx);
5741 mutex_lock(&ctx->mutex);
Peter Zijlstrab04243e2010-09-17 11:28:48 +02005742
5743 if (move_group) {
5744 perf_install_in_context(ctx, group_leader, cpu);
5745 get_ctx(ctx);
5746 list_for_each_entry(sibling, &group_leader->sibling_list,
5747 group_entry) {
5748 perf_install_in_context(ctx, sibling, cpu);
5749 get_ctx(ctx);
5750 }
5751 }
5752
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005753 perf_install_in_context(ctx, event, cpu);
5754 ++ctx->generation;
5755 mutex_unlock(&ctx->mutex);
5756
5757 event->owner = current;
Peter Zijlstra88821352010-11-09 19:01:43 +01005758
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005759 mutex_lock(&current->perf_event_mutex);
5760 list_add_tail(&event->owner_entry, &current->perf_event_list);
5761 mutex_unlock(&current->perf_event_mutex);
5762
Peter Zijlstra8a495422010-05-27 15:47:49 +02005763 /*
Arnaldo Carvalho de Meloc320c7b2010-10-20 12:50:11 -02005764 * Precalculate sample_data sizes
5765 */
5766 perf_event__header_size(event);
Arnaldo Carvalho de Melo6844c092010-12-03 16:36:35 -02005767 perf_event__id_header_size(event);
Arnaldo Carvalho de Meloc320c7b2010-10-20 12:50:11 -02005768
5769 /*
Peter Zijlstra8a495422010-05-27 15:47:49 +02005770 * Drop the reference on the group_event after placing the
5771 * new event on the sibling_list. This ensures destruction
5772 * of the group leader will find the pointer to itself in
5773 * perf_group_detach().
5774 */
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005775 fput_light(group_file, fput_needed);
Al Viroea635c62010-05-26 17:40:29 -04005776 fd_install(event_fd, event_file);
5777 return event_fd;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005778
Peter Zijlstrac3f00c72010-08-18 14:37:15 +02005779err_context:
Al Viroea635c62010-05-26 17:40:29 -04005780 put_ctx(ctx);
Peter Zijlstrac6be5a52010-10-14 16:59:46 +02005781err_alloc:
5782 free_event(event);
Peter Zijlstrae7d0bc02010-10-14 16:54:51 +02005783err_task:
5784 if (task)
5785 put_task_struct(task);
Peter Zijlstra89a1e182010-09-07 17:34:50 +02005786err_group_fd:
5787 fput_light(group_file, fput_needed);
Al Viroea635c62010-05-26 17:40:29 -04005788err_fd:
5789 put_unused_fd(event_fd);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005790 return err;
5791}
5792
Arjan van de Venfb0459d2009-09-25 12:25:56 +02005793/**
5794 * perf_event_create_kernel_counter
5795 *
5796 * @attr: attributes of the counter to create
5797 * @cpu: cpu in which the counter is bound
Matt Helsley38a81da2010-09-13 13:01:20 -07005798 * @task: task to profile (NULL for percpu)
Arjan van de Venfb0459d2009-09-25 12:25:56 +02005799 */
5800struct perf_event *
5801perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
Matt Helsley38a81da2010-09-13 13:01:20 -07005802 struct task_struct *task,
Frederic Weisbeckerb326e952009-12-05 09:44:31 +01005803 perf_overflow_handler_t overflow_handler)
Arjan van de Venfb0459d2009-09-25 12:25:56 +02005804{
Arjan van de Venfb0459d2009-09-25 12:25:56 +02005805 struct perf_event_context *ctx;
Peter Zijlstrac3f00c72010-08-18 14:37:15 +02005806 struct perf_event *event;
Arjan van de Venfb0459d2009-09-25 12:25:56 +02005807 int err;
5808
5809 /*
5810 * Get the target context (task or percpu):
5811 */
5812
Peter Zijlstrad580ff82010-10-14 17:43:23 +02005813 event = perf_event_alloc(attr, cpu, task, NULL, NULL, overflow_handler);
Frederic Weisbeckerc6567f62009-11-26 05:35:41 +01005814 if (IS_ERR(event)) {
5815 err = PTR_ERR(event);
Peter Zijlstrac3f00c72010-08-18 14:37:15 +02005816 goto err;
5817 }
5818
Matt Helsley38a81da2010-09-13 13:01:20 -07005819 ctx = find_get_context(event->pmu, task, cpu);
Arjan van de Venfb0459d2009-09-25 12:25:56 +02005820 if (IS_ERR(ctx)) {
5821 err = PTR_ERR(ctx);
Peter Zijlstrac3f00c72010-08-18 14:37:15 +02005822 goto err_free;
Frederic Weisbeckerc6567f62009-11-26 05:35:41 +01005823 }
Arjan van de Venfb0459d2009-09-25 12:25:56 +02005824
5825 event->filp = NULL;
5826 WARN_ON_ONCE(ctx->parent_ctx);
5827 mutex_lock(&ctx->mutex);
5828 perf_install_in_context(ctx, event, cpu);
5829 ++ctx->generation;
5830 mutex_unlock(&ctx->mutex);
5831
Arjan van de Venfb0459d2009-09-25 12:25:56 +02005832 return event;
5833
Peter Zijlstrac3f00c72010-08-18 14:37:15 +02005834err_free:
5835 free_event(event);
5836err:
Frederic Weisbeckerc6567f62009-11-26 05:35:41 +01005837 return ERR_PTR(err);
Arjan van de Venfb0459d2009-09-25 12:25:56 +02005838}
5839EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
5840
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005841static void sync_child_event(struct perf_event *child_event,
5842 struct task_struct *child)
5843{
5844 struct perf_event *parent_event = child_event->parent;
5845 u64 child_val;
5846
5847 if (child_event->attr.inherit_stat)
5848 perf_event_read_event(child_event, child);
5849
Peter Zijlstrab5e58792010-05-21 14:43:12 +02005850 child_val = perf_event_count(child_event);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005851
5852 /*
5853 * Add back the child's count to the parent's count:
5854 */
Peter Zijlstraa6e6dea2010-05-21 14:27:58 +02005855 atomic64_add(child_val, &parent_event->child_count);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005856 atomic64_add(child_event->total_time_enabled,
5857 &parent_event->child_total_time_enabled);
5858 atomic64_add(child_event->total_time_running,
5859 &parent_event->child_total_time_running);
5860
5861 /*
5862 * Remove this event from the parent's list
5863 */
5864 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5865 mutex_lock(&parent_event->child_mutex);
5866 list_del_init(&child_event->child_list);
5867 mutex_unlock(&parent_event->child_mutex);
5868
5869 /*
5870 * Release the parent event, if this was the last
5871 * reference to it.
5872 */
5873 fput(parent_event->filp);
5874}
5875
5876static void
5877__perf_event_exit_task(struct perf_event *child_event,
5878 struct perf_event_context *child_ctx,
5879 struct task_struct *child)
5880{
5881 struct perf_event *parent_event;
5882
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005883 perf_event_remove_from_context(child_event);
5884
5885 parent_event = child_event->parent;
5886 /*
5887 * It can happen that parent exits first, and has events
5888 * that are still around due to the child reference. These
5889 * events need to be zapped - but otherwise linger.
5890 */
5891 if (parent_event) {
5892 sync_child_event(child_event, child);
5893 free_event(child_event);
5894 }
5895}
5896
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02005897static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005898{
5899 struct perf_event *child_event, *tmp;
5900 struct perf_event_context *child_ctx;
5901 unsigned long flags;
5902
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02005903 if (likely(!child->perf_event_ctxp[ctxn])) {
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005904 perf_event_task(child, NULL, 0);
5905 return;
5906 }
5907
5908 local_irq_save(flags);
5909 /*
5910 * We can't reschedule here because interrupts are disabled,
5911 * and either child is current or it is a task that can't be
5912 * scheduled, so we are now safe from rescheduling changing
5913 * our context.
5914 */
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02005915 child_ctx = child->perf_event_ctxp[ctxn];
Peter Zijlstra82cd6de2010-10-14 17:57:23 +02005916 task_ctx_sched_out(child_ctx, EVENT_ALL);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005917
5918 /*
5919 * Take the context lock here so that if find_get_context is
5920 * reading child->perf_event_ctxp, we wait until it has
5921 * incremented the context's refcount before we do put_ctx below.
5922 */
Thomas Gleixnere625cce12009-11-17 18:02:06 +01005923 raw_spin_lock(&child_ctx->lock);
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02005924 child->perf_event_ctxp[ctxn] = NULL;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005925 /*
5926 * If this context is a clone; unclone it so it can't get
5927 * swapped to another process while we're removing all
5928 * the events from it.
5929 */
5930 unclone_ctx(child_ctx);
Peter Zijlstra5e942bb2009-11-23 11:37:26 +01005931 update_context_time(child_ctx);
Thomas Gleixnere625cce12009-11-17 18:02:06 +01005932 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005933
5934 /*
5935 * Report the task dead after unscheduling the events so that we
5936 * won't get any samples after PERF_RECORD_EXIT. We can however still
5937 * get a few PERF_RECORD_READ events.
5938 */
5939 perf_event_task(child, child_ctx, 0);
5940
5941 /*
5942 * We can recurse on the same lock type through:
5943 *
5944 * __perf_event_exit_task()
5945 * sync_child_event()
5946 * fput(parent_event->filp)
5947 * perf_release()
5948 * mutex_lock(&ctx->mutex)
5949 *
5950 * But since its the parent context it won't be the same instance.
5951 */
Peter Zijlstraa0507c82010-05-06 15:42:53 +02005952 mutex_lock(&child_ctx->mutex);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005953
5954again:
Frederic Weisbecker889ff012010-01-09 20:04:47 +01005955 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5956 group_entry)
5957 __perf_event_exit_task(child_event, child_ctx, child);
5958
5959 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005960 group_entry)
5961 __perf_event_exit_task(child_event, child_ctx, child);
5962
5963 /*
5964 * If the last event was a group event, it will have appended all
5965 * its siblings to the list, but we obtained 'tmp' before that which
5966 * will still point to the list head terminating the iteration.
5967 */
Frederic Weisbecker889ff012010-01-09 20:04:47 +01005968 if (!list_empty(&child_ctx->pinned_groups) ||
5969 !list_empty(&child_ctx->flexible_groups))
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005970 goto again;
5971
5972 mutex_unlock(&child_ctx->mutex);
5973
5974 put_ctx(child_ctx);
5975}
5976
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02005977/*
5978 * When a child task exits, feed back event values to parent events.
5979 */
5980void perf_event_exit_task(struct task_struct *child)
5981{
Peter Zijlstra88821352010-11-09 19:01:43 +01005982 struct perf_event *event, *tmp;
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02005983 int ctxn;
5984
Peter Zijlstra88821352010-11-09 19:01:43 +01005985 mutex_lock(&child->perf_event_mutex);
5986 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
5987 owner_entry) {
5988 list_del_init(&event->owner_entry);
5989
5990 /*
5991 * Ensure the list deletion is visible before we clear
5992 * the owner, closes a race against perf_release() where
5993 * we need to serialize on the owner->perf_event_mutex.
5994 */
5995 smp_wmb();
5996 event->owner = NULL;
5997 }
5998 mutex_unlock(&child->perf_event_mutex);
5999
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02006000 for_each_task_context_nr(ctxn)
6001 perf_event_exit_task_context(child, ctxn);
6002}
6003
Frederic Weisbecker889ff012010-01-09 20:04:47 +01006004static void perf_free_event(struct perf_event *event,
6005 struct perf_event_context *ctx)
6006{
6007 struct perf_event *parent = event->parent;
6008
6009 if (WARN_ON_ONCE(!parent))
6010 return;
6011
6012 mutex_lock(&parent->child_mutex);
6013 list_del_init(&event->child_list);
6014 mutex_unlock(&parent->child_mutex);
6015
6016 fput(parent->filp);
6017
Peter Zijlstra8a495422010-05-27 15:47:49 +02006018 perf_group_detach(event);
Frederic Weisbecker889ff012010-01-09 20:04:47 +01006019 list_del_event(event, ctx);
6020 free_event(event);
6021}
6022
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006023/*
6024 * free an unexposed, unused context as created by inheritance by
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02006025 * perf_event_init_task below, used by fork() in case of fail.
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006026 */
6027void perf_event_free_task(struct task_struct *task)
6028{
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02006029 struct perf_event_context *ctx;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006030 struct perf_event *event, *tmp;
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02006031 int ctxn;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006032
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02006033 for_each_task_context_nr(ctxn) {
6034 ctx = task->perf_event_ctxp[ctxn];
6035 if (!ctx)
6036 continue;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006037
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02006038 mutex_lock(&ctx->mutex);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006039again:
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02006040 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
6041 group_entry)
6042 perf_free_event(event, ctx);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006043
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02006044 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
6045 group_entry)
6046 perf_free_event(event, ctx);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006047
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02006048 if (!list_empty(&ctx->pinned_groups) ||
6049 !list_empty(&ctx->flexible_groups))
6050 goto again;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006051
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02006052 mutex_unlock(&ctx->mutex);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006053
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02006054 put_ctx(ctx);
6055 }
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006056}
6057
Peter Zijlstra4e231c72010-09-09 21:01:59 +02006058void perf_event_delayed_put(struct task_struct *task)
6059{
6060 int ctxn;
6061
6062 for_each_task_context_nr(ctxn)
6063 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
6064}
6065
Peter Zijlstra97dee4f2010-09-07 15:35:33 +02006066/*
6067 * inherit a event from parent task to child task:
6068 */
6069static struct perf_event *
6070inherit_event(struct perf_event *parent_event,
6071 struct task_struct *parent,
6072 struct perf_event_context *parent_ctx,
6073 struct task_struct *child,
6074 struct perf_event *group_leader,
6075 struct perf_event_context *child_ctx)
6076{
6077 struct perf_event *child_event;
Peter Zijlstracee010e2010-09-10 12:51:54 +02006078 unsigned long flags;
Peter Zijlstra97dee4f2010-09-07 15:35:33 +02006079
6080 /*
6081 * Instead of creating recursive hierarchies of events,
6082 * we link inherited events back to the original parent,
6083 * which has a filp for sure, which we use as the reference
6084 * count:
6085 */
6086 if (parent_event->parent)
6087 parent_event = parent_event->parent;
6088
6089 child_event = perf_event_alloc(&parent_event->attr,
6090 parent_event->cpu,
Peter Zijlstrad580ff82010-10-14 17:43:23 +02006091 child,
Peter Zijlstra97dee4f2010-09-07 15:35:33 +02006092 group_leader, parent_event,
6093 NULL);
6094 if (IS_ERR(child_event))
6095 return child_event;
6096 get_ctx(child_ctx);
6097
6098 /*
6099 * Make the child state follow the state of the parent event,
6100 * not its attr.disabled bit. We hold the parent's mutex,
6101 * so we won't race with perf_event_{en, dis}able_family.
6102 */
6103 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
6104 child_event->state = PERF_EVENT_STATE_INACTIVE;
6105 else
6106 child_event->state = PERF_EVENT_STATE_OFF;
6107
6108 if (parent_event->attr.freq) {
6109 u64 sample_period = parent_event->hw.sample_period;
6110 struct hw_perf_event *hwc = &child_event->hw;
6111
6112 hwc->sample_period = sample_period;
6113 hwc->last_period = sample_period;
6114
6115 local64_set(&hwc->period_left, sample_period);
6116 }
6117
6118 child_event->ctx = child_ctx;
6119 child_event->overflow_handler = parent_event->overflow_handler;
6120
6121 /*
Thomas Gleixner614b6782010-12-03 16:24:32 -02006122 * Precalculate sample_data sizes
6123 */
6124 perf_event__header_size(child_event);
Arnaldo Carvalho de Melo6844c092010-12-03 16:36:35 -02006125 perf_event__id_header_size(child_event);
Thomas Gleixner614b6782010-12-03 16:24:32 -02006126
6127 /*
Peter Zijlstra97dee4f2010-09-07 15:35:33 +02006128 * Link it up in the child's context:
6129 */
Peter Zijlstracee010e2010-09-10 12:51:54 +02006130 raw_spin_lock_irqsave(&child_ctx->lock, flags);
Peter Zijlstra97dee4f2010-09-07 15:35:33 +02006131 add_event_to_ctx(child_event, child_ctx);
Peter Zijlstracee010e2010-09-10 12:51:54 +02006132 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
Peter Zijlstra97dee4f2010-09-07 15:35:33 +02006133
6134 /*
6135 * Get a reference to the parent filp - we will fput it
6136 * when the child event exits. This is safe to do because
6137 * we are in the parent and we know that the filp still
6138 * exists and has a nonzero count:
6139 */
6140 atomic_long_inc(&parent_event->filp->f_count);
6141
6142 /*
6143 * Link this into the parent event's child list
6144 */
6145 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6146 mutex_lock(&parent_event->child_mutex);
6147 list_add_tail(&child_event->child_list, &parent_event->child_list);
6148 mutex_unlock(&parent_event->child_mutex);
6149
6150 return child_event;
6151}
6152
6153static int inherit_group(struct perf_event *parent_event,
6154 struct task_struct *parent,
6155 struct perf_event_context *parent_ctx,
6156 struct task_struct *child,
6157 struct perf_event_context *child_ctx)
6158{
6159 struct perf_event *leader;
6160 struct perf_event *sub;
6161 struct perf_event *child_ctr;
6162
6163 leader = inherit_event(parent_event, parent, parent_ctx,
6164 child, NULL, child_ctx);
6165 if (IS_ERR(leader))
6166 return PTR_ERR(leader);
6167 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
6168 child_ctr = inherit_event(sub, parent, parent_ctx,
6169 child, leader, child_ctx);
6170 if (IS_ERR(child_ctr))
6171 return PTR_ERR(child_ctr);
6172 }
6173 return 0;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006174}
6175
Frederic Weisbecker889ff012010-01-09 20:04:47 +01006176static int
6177inherit_task_group(struct perf_event *event, struct task_struct *parent,
6178 struct perf_event_context *parent_ctx,
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02006179 struct task_struct *child, int ctxn,
Frederic Weisbecker889ff012010-01-09 20:04:47 +01006180 int *inherited_all)
6181{
6182 int ret;
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02006183 struct perf_event_context *child_ctx;
Frederic Weisbecker889ff012010-01-09 20:04:47 +01006184
6185 if (!event->attr.inherit) {
6186 *inherited_all = 0;
6187 return 0;
6188 }
6189
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02006190 child_ctx = child->perf_event_ctxp[ctxn];
Frederic Weisbecker889ff012010-01-09 20:04:47 +01006191 if (!child_ctx) {
6192 /*
6193 * This is executed from the parent task context, so
6194 * inherit events that have been marked for cloning.
6195 * First allocate and initialize a context for the
6196 * child.
6197 */
6198
Peter Zijlstraeb184472010-09-07 15:55:13 +02006199 child_ctx = alloc_perf_context(event->pmu, child);
Frederic Weisbecker889ff012010-01-09 20:04:47 +01006200 if (!child_ctx)
6201 return -ENOMEM;
6202
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02006203 child->perf_event_ctxp[ctxn] = child_ctx;
Frederic Weisbecker889ff012010-01-09 20:04:47 +01006204 }
6205
6206 ret = inherit_group(event, parent, parent_ctx,
6207 child, child_ctx);
6208
6209 if (ret)
6210 *inherited_all = 0;
6211
6212 return ret;
6213}
6214
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006215/*
6216 * Initialize the perf_event context in task_struct
6217 */
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02006218int perf_event_init_context(struct task_struct *child, int ctxn)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006219{
Frederic Weisbecker889ff012010-01-09 20:04:47 +01006220 struct perf_event_context *child_ctx, *parent_ctx;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006221 struct perf_event_context *cloned_ctx;
6222 struct perf_event *event;
6223 struct task_struct *parent = current;
6224 int inherited_all = 1;
Thomas Gleixnerdddd3372010-11-24 10:05:55 +01006225 unsigned long flags;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006226 int ret = 0;
6227
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02006228 child->perf_event_ctxp[ctxn] = NULL;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006229
6230 mutex_init(&child->perf_event_mutex);
6231 INIT_LIST_HEAD(&child->perf_event_list);
6232
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02006233 if (likely(!parent->perf_event_ctxp[ctxn]))
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006234 return 0;
6235
6236 /*
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006237 * If the parent's context is a clone, pin it so it won't get
6238 * swapped under us.
6239 */
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02006240 parent_ctx = perf_pin_task_context(parent, ctxn);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006241
6242 /*
6243 * No need to check if parent_ctx != NULL here; since we saw
6244 * it non-NULL earlier, the only reason for it to become NULL
6245 * is if we exit, and since we're currently in the middle of
6246 * a fork we can't be exiting at the same time.
6247 */
6248
6249 /*
6250 * Lock the parent list. No need to lock the child - not PID
6251 * hashed yet and not running, so nobody can access it.
6252 */
6253 mutex_lock(&parent_ctx->mutex);
6254
6255 /*
6256 * We dont have to disable NMIs - we are only looking at
6257 * the list, not manipulating it:
6258 */
Frederic Weisbecker889ff012010-01-09 20:04:47 +01006259 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02006260 ret = inherit_task_group(event, parent, parent_ctx,
6261 child, ctxn, &inherited_all);
Frederic Weisbecker889ff012010-01-09 20:04:47 +01006262 if (ret)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006263 break;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006264 }
6265
Thomas Gleixnerdddd3372010-11-24 10:05:55 +01006266 /*
6267 * We can't hold ctx->lock when iterating the ->flexible_group list due
6268 * to allocations, but we need to prevent rotation because
6269 * rotate_ctx() will change the list from interrupt context.
6270 */
6271 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
6272 parent_ctx->rotate_disable = 1;
6273 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
6274
Frederic Weisbecker889ff012010-01-09 20:04:47 +01006275 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02006276 ret = inherit_task_group(event, parent, parent_ctx,
6277 child, ctxn, &inherited_all);
Frederic Weisbecker889ff012010-01-09 20:04:47 +01006278 if (ret)
6279 break;
6280 }
6281
Thomas Gleixnerdddd3372010-11-24 10:05:55 +01006282 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
6283 parent_ctx->rotate_disable = 0;
6284 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
6285
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02006286 child_ctx = child->perf_event_ctxp[ctxn];
Frederic Weisbecker889ff012010-01-09 20:04:47 +01006287
Peter Zijlstra05cbaa22009-12-30 16:00:35 +01006288 if (child_ctx && inherited_all) {
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006289 /*
6290 * Mark the child context as a clone of the parent
6291 * context, or of whatever the parent is a clone of.
6292 * Note that if the parent is a clone, it could get
6293 * uncloned at any point, but that doesn't matter
6294 * because the list of events and the generation
6295 * count can't have changed since we took the mutex.
6296 */
6297 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
6298 if (cloned_ctx) {
6299 child_ctx->parent_ctx = cloned_ctx;
6300 child_ctx->parent_gen = parent_ctx->parent_gen;
6301 } else {
6302 child_ctx->parent_ctx = parent_ctx;
6303 child_ctx->parent_gen = parent_ctx->generation;
6304 }
6305 get_ctx(child_ctx->parent_ctx);
6306 }
6307
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006308 mutex_unlock(&parent_ctx->mutex);
6309
6310 perf_unpin_context(parent_ctx);
6311
6312 return ret;
6313}
6314
Peter Zijlstra8dc85d52010-09-02 16:50:03 +02006315/*
6316 * Initialize the perf_event context in task_struct
6317 */
6318int perf_event_init_task(struct task_struct *child)
6319{
6320 int ctxn, ret;
6321
6322 for_each_task_context_nr(ctxn) {
6323 ret = perf_event_init_context(child, ctxn);
6324 if (ret)
6325 return ret;
6326 }
6327
6328 return 0;
6329}
6330
Paul Mackerras220b1402010-03-10 20:45:52 +11006331static void __init perf_event_init_all_cpus(void)
6332{
Peter Zijlstrab28ab832010-09-06 14:48:15 +02006333 struct swevent_htable *swhash;
Paul Mackerras220b1402010-03-10 20:45:52 +11006334 int cpu;
Paul Mackerras220b1402010-03-10 20:45:52 +11006335
6336 for_each_possible_cpu(cpu) {
Peter Zijlstrab28ab832010-09-06 14:48:15 +02006337 swhash = &per_cpu(swevent_htable, cpu);
6338 mutex_init(&swhash->hlist_mutex);
Peter Zijlstrae9d2b062010-09-17 11:28:50 +02006339 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
Paul Mackerras220b1402010-03-10 20:45:52 +11006340 }
6341}
6342
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006343static void __cpuinit perf_event_init_cpu(int cpu)
6344{
Peter Zijlstra108b02c2010-09-06 14:32:03 +02006345 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006346
Peter Zijlstrab28ab832010-09-06 14:48:15 +02006347 mutex_lock(&swhash->hlist_mutex);
6348 if (swhash->hlist_refcount > 0) {
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02006349 struct swevent_hlist *hlist;
6350
Peter Zijlstrab28ab832010-09-06 14:48:15 +02006351 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
6352 WARN_ON(!hlist);
6353 rcu_assign_pointer(swhash->swevent_hlist, hlist);
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02006354 }
Peter Zijlstrab28ab832010-09-06 14:48:15 +02006355 mutex_unlock(&swhash->hlist_mutex);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006356}
6357
6358#ifdef CONFIG_HOTPLUG_CPU
Peter Zijlstrae9d2b062010-09-17 11:28:50 +02006359static void perf_pmu_rotate_stop(struct pmu *pmu)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006360{
Peter Zijlstrae9d2b062010-09-17 11:28:50 +02006361 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
6362
6363 WARN_ON(!irqs_disabled());
6364
6365 list_del_init(&cpuctx->rotation_list);
6366}
6367
Peter Zijlstra108b02c2010-09-06 14:32:03 +02006368static void __perf_event_exit_context(void *__info)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006369{
Peter Zijlstra108b02c2010-09-06 14:32:03 +02006370 struct perf_event_context *ctx = __info;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006371 struct perf_event *event, *tmp;
6372
Peter Zijlstra108b02c2010-09-06 14:32:03 +02006373 perf_pmu_rotate_stop(ctx->pmu);
Peter Zijlstrab5ab4cd2010-09-06 16:32:21 +02006374
Frederic Weisbecker889ff012010-01-09 20:04:47 +01006375 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
6376 __perf_event_remove_from_context(event);
6377 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006378 __perf_event_remove_from_context(event);
6379}
Peter Zijlstra108b02c2010-09-06 14:32:03 +02006380
6381static void perf_event_exit_cpu_context(int cpu)
6382{
6383 struct perf_event_context *ctx;
6384 struct pmu *pmu;
6385 int idx;
6386
6387 idx = srcu_read_lock(&pmus_srcu);
6388 list_for_each_entry_rcu(pmu, &pmus, entry) {
Peter Zijlstra917bdd12010-09-17 11:28:49 +02006389 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
Peter Zijlstra108b02c2010-09-06 14:32:03 +02006390
6391 mutex_lock(&ctx->mutex);
6392 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
6393 mutex_unlock(&ctx->mutex);
6394 }
6395 srcu_read_unlock(&pmus_srcu, idx);
Peter Zijlstra108b02c2010-09-06 14:32:03 +02006396}
6397
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006398static void perf_event_exit_cpu(int cpu)
6399{
Peter Zijlstrab28ab832010-09-06 14:48:15 +02006400 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006401
Peter Zijlstrab28ab832010-09-06 14:48:15 +02006402 mutex_lock(&swhash->hlist_mutex);
6403 swevent_hlist_release(swhash);
6404 mutex_unlock(&swhash->hlist_mutex);
Frederic Weisbecker76e1d902010-04-05 15:35:57 +02006405
Peter Zijlstra108b02c2010-09-06 14:32:03 +02006406 perf_event_exit_cpu_context(cpu);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006407}
6408#else
6409static inline void perf_event_exit_cpu(int cpu) { }
6410#endif
6411
6412static int __cpuinit
6413perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
6414{
6415 unsigned int cpu = (long)hcpu;
6416
Peter Zijlstra5e116372010-06-11 13:35:08 +02006417 switch (action & ~CPU_TASKS_FROZEN) {
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006418
6419 case CPU_UP_PREPARE:
Peter Zijlstra5e116372010-06-11 13:35:08 +02006420 case CPU_DOWN_FAILED:
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006421 perf_event_init_cpu(cpu);
6422 break;
6423
Peter Zijlstra5e116372010-06-11 13:35:08 +02006424 case CPU_UP_CANCELED:
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006425 case CPU_DOWN_PREPARE:
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006426 perf_event_exit_cpu(cpu);
6427 break;
6428
6429 default:
6430 break;
6431 }
6432
6433 return NOTIFY_OK;
6434}
6435
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006436void __init perf_event_init(void)
6437{
Jason Wessel3c502e72010-11-04 17:33:01 -05006438 int ret;
6439
Paul Mackerras220b1402010-03-10 20:45:52 +11006440 perf_event_init_all_cpus();
Peter Zijlstrab0a873e2010-06-11 13:35:08 +02006441 init_srcu_struct(&pmus_srcu);
6442 perf_pmu_register(&perf_swevent);
6443 perf_pmu_register(&perf_cpu_clock);
6444 perf_pmu_register(&perf_task_clock);
6445 perf_tp_register();
6446 perf_cpu_notifier(perf_cpu_notify);
Jason Wessel3c502e72010-11-04 17:33:01 -05006447
6448 ret = init_hw_breakpoint();
6449 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02006450}