perf: Annotate cpuctx->ctx.mutex to avoid a lockdep splat
[linux-2.6.git] / kernel / perf_event.c
1 /*
2  * Performance events core code:
3  *
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  *
9  * For licensing details see kernel-base/COPYING
10  */
11
12 #include <linux/fs.h>
13 #include <linux/mm.h>
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/idr.h>
17 #include <linux/file.h>
18 #include <linux/poll.h>
19 #include <linux/slab.h>
20 #include <linux/hash.h>
21 #include <linux/sysfs.h>
22 #include <linux/dcache.h>
23 #include <linux/percpu.h>
24 #include <linux/ptrace.h>
25 #include <linux/reboot.h>
26 #include <linux/vmstat.h>
27 #include <linux/device.h>
28 #include <linux/vmalloc.h>
29 #include <linux/hardirq.h>
30 #include <linux/rculist.h>
31 #include <linux/uaccess.h>
32 #include <linux/syscalls.h>
33 #include <linux/anon_inodes.h>
34 #include <linux/kernel_stat.h>
35 #include <linux/perf_event.h>
36 #include <linux/ftrace_event.h>
37 #include <linux/hw_breakpoint.h>
38
39 #include <asm/irq_regs.h>
40
41 enum event_type_t {
42         EVENT_FLEXIBLE = 0x1,
43         EVENT_PINNED = 0x2,
44         EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
45 };
46
47 atomic_t perf_task_events __read_mostly;
48 static atomic_t nr_mmap_events __read_mostly;
49 static atomic_t nr_comm_events __read_mostly;
50 static atomic_t nr_task_events __read_mostly;
51
52 static LIST_HEAD(pmus);
53 static DEFINE_MUTEX(pmus_lock);
54 static struct srcu_struct pmus_srcu;
55
56 /*
57  * perf event paranoia level:
58  *  -1 - not paranoid at all
59  *   0 - disallow raw tracepoint access for unpriv
60  *   1 - disallow cpu events for unpriv
61  *   2 - disallow kernel profiling for unpriv
62  */
63 int sysctl_perf_event_paranoid __read_mostly = 1;
64
65 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
66
67 /*
68  * max perf event sample rate
69  */
70 int sysctl_perf_event_sample_rate __read_mostly = 100000;
71
72 static atomic64_t perf_event_id;
73
74 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
75                               enum event_type_t event_type);
76
77 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
78                              enum event_type_t event_type);
79
80 void __weak perf_event_print_debug(void)        { }
81
82 extern __weak const char *perf_pmu_name(void)
83 {
84         return "pmu";
85 }
86
87 static inline u64 perf_clock(void)
88 {
89         return local_clock();
90 }
91
92 void perf_pmu_disable(struct pmu *pmu)
93 {
94         int *count = this_cpu_ptr(pmu->pmu_disable_count);
95         if (!(*count)++)
96                 pmu->pmu_disable(pmu);
97 }
98
99 void perf_pmu_enable(struct pmu *pmu)
100 {
101         int *count = this_cpu_ptr(pmu->pmu_disable_count);
102         if (!--(*count))
103                 pmu->pmu_enable(pmu);
104 }
105
106 static DEFINE_PER_CPU(struct list_head, rotation_list);
107
108 /*
109  * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
110  * because they're strictly cpu affine and rotate_start is called with IRQs
111  * disabled, while rotate_context is called from IRQ context.
112  */
113 static void perf_pmu_rotate_start(struct pmu *pmu)
114 {
115         struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
116         struct list_head *head = &__get_cpu_var(rotation_list);
117
118         WARN_ON(!irqs_disabled());
119
120         if (list_empty(&cpuctx->rotation_list))
121                 list_add(&cpuctx->rotation_list, head);
122 }
123
124 static void get_ctx(struct perf_event_context *ctx)
125 {
126         WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
127 }
128
129 static void free_ctx(struct rcu_head *head)
130 {
131         struct perf_event_context *ctx;
132
133         ctx = container_of(head, struct perf_event_context, rcu_head);
134         kfree(ctx);
135 }
136
137 static void put_ctx(struct perf_event_context *ctx)
138 {
139         if (atomic_dec_and_test(&ctx->refcount)) {
140                 if (ctx->parent_ctx)
141                         put_ctx(ctx->parent_ctx);
142                 if (ctx->task)
143                         put_task_struct(ctx->task);
144                 call_rcu(&ctx->rcu_head, free_ctx);
145         }
146 }
147
148 static void unclone_ctx(struct perf_event_context *ctx)
149 {
150         if (ctx->parent_ctx) {
151                 put_ctx(ctx->parent_ctx);
152                 ctx->parent_ctx = NULL;
153         }
154 }
155
156 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
157 {
158         /*
159          * only top level events have the pid namespace they were created in
160          */
161         if (event->parent)
162                 event = event->parent;
163
164         return task_tgid_nr_ns(p, event->ns);
165 }
166
167 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
168 {
169         /*
170          * only top level events have the pid namespace they were created in
171          */
172         if (event->parent)
173                 event = event->parent;
174
175         return task_pid_nr_ns(p, event->ns);
176 }
177
178 /*
179  * If we inherit events we want to return the parent event id
180  * to userspace.
181  */
182 static u64 primary_event_id(struct perf_event *event)
183 {
184         u64 id = event->id;
185
186         if (event->parent)
187                 id = event->parent->id;
188
189         return id;
190 }
191
192 /*
193  * Get the perf_event_context for a task and lock it.
194  * This has to cope with with the fact that until it is locked,
195  * the context could get moved to another task.
196  */
197 static struct perf_event_context *
198 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
199 {
200         struct perf_event_context *ctx;
201
202         rcu_read_lock();
203 retry:
204         ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
205         if (ctx) {
206                 /*
207                  * If this context is a clone of another, it might
208                  * get swapped for another underneath us by
209                  * perf_event_task_sched_out, though the
210                  * rcu_read_lock() protects us from any context
211                  * getting freed.  Lock the context and check if it
212                  * got swapped before we could get the lock, and retry
213                  * if so.  If we locked the right context, then it
214                  * can't get swapped on us any more.
215                  */
216                 raw_spin_lock_irqsave(&ctx->lock, *flags);
217                 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
218                         raw_spin_unlock_irqrestore(&ctx->lock, *flags);
219                         goto retry;
220                 }
221
222                 if (!atomic_inc_not_zero(&ctx->refcount)) {
223                         raw_spin_unlock_irqrestore(&ctx->lock, *flags);
224                         ctx = NULL;
225                 }
226         }
227         rcu_read_unlock();
228         return ctx;
229 }
230
231 /*
232  * Get the context for a task and increment its pin_count so it
233  * can't get swapped to another task.  This also increments its
234  * reference count so that the context can't get freed.
235  */
236 static struct perf_event_context *
237 perf_pin_task_context(struct task_struct *task, int ctxn)
238 {
239         struct perf_event_context *ctx;
240         unsigned long flags;
241
242         ctx = perf_lock_task_context(task, ctxn, &flags);
243         if (ctx) {
244                 ++ctx->pin_count;
245                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
246         }
247         return ctx;
248 }
249
250 static void perf_unpin_context(struct perf_event_context *ctx)
251 {
252         unsigned long flags;
253
254         raw_spin_lock_irqsave(&ctx->lock, flags);
255         --ctx->pin_count;
256         raw_spin_unlock_irqrestore(&ctx->lock, flags);
257         put_ctx(ctx);
258 }
259
260 /*
261  * Update the record of the current time in a context.
262  */
263 static void update_context_time(struct perf_event_context *ctx)
264 {
265         u64 now = perf_clock();
266
267         ctx->time += now - ctx->timestamp;
268         ctx->timestamp = now;
269 }
270
271 static u64 perf_event_time(struct perf_event *event)
272 {
273         struct perf_event_context *ctx = event->ctx;
274         return ctx ? ctx->time : 0;
275 }
276
277 /*
278  * Update the total_time_enabled and total_time_running fields for a event.
279  */
280 static void update_event_times(struct perf_event *event)
281 {
282         struct perf_event_context *ctx = event->ctx;
283         u64 run_end;
284
285         if (event->state < PERF_EVENT_STATE_INACTIVE ||
286             event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
287                 return;
288
289         if (ctx->is_active)
290                 run_end = perf_event_time(event);
291         else
292                 run_end = event->tstamp_stopped;
293
294         event->total_time_enabled = run_end - event->tstamp_enabled;
295
296         if (event->state == PERF_EVENT_STATE_INACTIVE)
297                 run_end = event->tstamp_stopped;
298         else
299                 run_end = perf_event_time(event);
300
301         event->total_time_running = run_end - event->tstamp_running;
302 }
303
304 /*
305  * Update total_time_enabled and total_time_running for all events in a group.
306  */
307 static void update_group_times(struct perf_event *leader)
308 {
309         struct perf_event *event;
310
311         update_event_times(leader);
312         list_for_each_entry(event, &leader->sibling_list, group_entry)
313                 update_event_times(event);
314 }
315
316 static struct list_head *
317 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
318 {
319         if (event->attr.pinned)
320                 return &ctx->pinned_groups;
321         else
322                 return &ctx->flexible_groups;
323 }
324
325 /*
326  * Add a event from the lists for its context.
327  * Must be called with ctx->mutex and ctx->lock held.
328  */
329 static void
330 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
331 {
332         WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
333         event->attach_state |= PERF_ATTACH_CONTEXT;
334
335         /*
336          * If we're a stand alone event or group leader, we go to the context
337          * list, group events are kept attached to the group so that
338          * perf_group_detach can, at all times, locate all siblings.
339          */
340         if (event->group_leader == event) {
341                 struct list_head *list;
342
343                 if (is_software_event(event))
344                         event->group_flags |= PERF_GROUP_SOFTWARE;
345
346                 list = ctx_group_list(event, ctx);
347                 list_add_tail(&event->group_entry, list);
348         }
349
350         list_add_rcu(&event->event_entry, &ctx->event_list);
351         if (!ctx->nr_events)
352                 perf_pmu_rotate_start(ctx->pmu);
353         ctx->nr_events++;
354         if (event->attr.inherit_stat)
355                 ctx->nr_stat++;
356 }
357
358 /*
359  * Called at perf_event creation and when events are attached/detached from a
360  * group.
361  */
362 static void perf_event__read_size(struct perf_event *event)
363 {
364         int entry = sizeof(u64); /* value */
365         int size = 0;
366         int nr = 1;
367
368         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
369                 size += sizeof(u64);
370
371         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
372                 size += sizeof(u64);
373
374         if (event->attr.read_format & PERF_FORMAT_ID)
375                 entry += sizeof(u64);
376
377         if (event->attr.read_format & PERF_FORMAT_GROUP) {
378                 nr += event->group_leader->nr_siblings;
379                 size += sizeof(u64);
380         }
381
382         size += entry * nr;
383         event->read_size = size;
384 }
385
386 static void perf_event__header_size(struct perf_event *event)
387 {
388         struct perf_sample_data *data;
389         u64 sample_type = event->attr.sample_type;
390         u16 size = 0;
391
392         perf_event__read_size(event);
393
394         if (sample_type & PERF_SAMPLE_IP)
395                 size += sizeof(data->ip);
396
397         if (sample_type & PERF_SAMPLE_ADDR)
398                 size += sizeof(data->addr);
399
400         if (sample_type & PERF_SAMPLE_PERIOD)
401                 size += sizeof(data->period);
402
403         if (sample_type & PERF_SAMPLE_READ)
404                 size += event->read_size;
405
406         event->header_size = size;
407 }
408
409 static void perf_event__id_header_size(struct perf_event *event)
410 {
411         struct perf_sample_data *data;
412         u64 sample_type = event->attr.sample_type;
413         u16 size = 0;
414
415         if (sample_type & PERF_SAMPLE_TID)
416                 size += sizeof(data->tid_entry);
417
418         if (sample_type & PERF_SAMPLE_TIME)
419                 size += sizeof(data->time);
420
421         if (sample_type & PERF_SAMPLE_ID)
422                 size += sizeof(data->id);
423
424         if (sample_type & PERF_SAMPLE_STREAM_ID)
425                 size += sizeof(data->stream_id);
426
427         if (sample_type & PERF_SAMPLE_CPU)
428                 size += sizeof(data->cpu_entry);
429
430         event->id_header_size = size;
431 }
432
433 static void perf_group_attach(struct perf_event *event)
434 {
435         struct perf_event *group_leader = event->group_leader, *pos;
436
437         /*
438          * We can have double attach due to group movement in perf_event_open.
439          */
440         if (event->attach_state & PERF_ATTACH_GROUP)
441                 return;
442
443         event->attach_state |= PERF_ATTACH_GROUP;
444
445         if (group_leader == event)
446                 return;
447
448         if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
449                         !is_software_event(event))
450                 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
451
452         list_add_tail(&event->group_entry, &group_leader->sibling_list);
453         group_leader->nr_siblings++;
454
455         perf_event__header_size(group_leader);
456
457         list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
458                 perf_event__header_size(pos);
459 }
460
461 /*
462  * Remove a event from the lists for its context.
463  * Must be called with ctx->mutex and ctx->lock held.
464  */
465 static void
466 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
467 {
468         /*
469          * We can have double detach due to exit/hot-unplug + close.
470          */
471         if (!(event->attach_state & PERF_ATTACH_CONTEXT))
472                 return;
473
474         event->attach_state &= ~PERF_ATTACH_CONTEXT;
475
476         ctx->nr_events--;
477         if (event->attr.inherit_stat)
478                 ctx->nr_stat--;
479
480         list_del_rcu(&event->event_entry);
481
482         if (event->group_leader == event)
483                 list_del_init(&event->group_entry);
484
485         update_group_times(event);
486
487         /*
488          * If event was in error state, then keep it
489          * that way, otherwise bogus counts will be
490          * returned on read(). The only way to get out
491          * of error state is by explicit re-enabling
492          * of the event
493          */
494         if (event->state > PERF_EVENT_STATE_OFF)
495                 event->state = PERF_EVENT_STATE_OFF;
496 }
497
498 static void perf_group_detach(struct perf_event *event)
499 {
500         struct perf_event *sibling, *tmp;
501         struct list_head *list = NULL;
502
503         /*
504          * We can have double detach due to exit/hot-unplug + close.
505          */
506         if (!(event->attach_state & PERF_ATTACH_GROUP))
507                 return;
508
509         event->attach_state &= ~PERF_ATTACH_GROUP;
510
511         /*
512          * If this is a sibling, remove it from its group.
513          */
514         if (event->group_leader != event) {
515                 list_del_init(&event->group_entry);
516                 event->group_leader->nr_siblings--;
517                 goto out;
518         }
519
520         if (!list_empty(&event->group_entry))
521                 list = &event->group_entry;
522
523         /*
524          * If this was a group event with sibling events then
525          * upgrade the siblings to singleton events by adding them
526          * to whatever list we are on.
527          */
528         list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
529                 if (list)
530                         list_move_tail(&sibling->group_entry, list);
531                 sibling->group_leader = sibling;
532
533                 /* Inherit group flags from the previous leader */
534                 sibling->group_flags = event->group_flags;
535         }
536
537 out:
538         perf_event__header_size(event->group_leader);
539
540         list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
541                 perf_event__header_size(tmp);
542 }
543
544 static inline int
545 event_filter_match(struct perf_event *event)
546 {
547         return event->cpu == -1 || event->cpu == smp_processor_id();
548 }
549
550 static void
551 event_sched_out(struct perf_event *event,
552                   struct perf_cpu_context *cpuctx,
553                   struct perf_event_context *ctx)
554 {
555         u64 tstamp = perf_event_time(event);
556         u64 delta;
557         /*
558          * An event which could not be activated because of
559          * filter mismatch still needs to have its timings
560          * maintained, otherwise bogus information is return
561          * via read() for time_enabled, time_running:
562          */
563         if (event->state == PERF_EVENT_STATE_INACTIVE
564             && !event_filter_match(event)) {
565                 delta = ctx->time - event->tstamp_stopped;
566                 event->tstamp_running += delta;
567                 event->tstamp_stopped = tstamp;
568         }
569
570         if (event->state != PERF_EVENT_STATE_ACTIVE)
571                 return;
572
573         event->state = PERF_EVENT_STATE_INACTIVE;
574         if (event->pending_disable) {
575                 event->pending_disable = 0;
576                 event->state = PERF_EVENT_STATE_OFF;
577         }
578         event->tstamp_stopped = tstamp;
579         event->pmu->del(event, 0);
580         event->oncpu = -1;
581
582         if (!is_software_event(event))
583                 cpuctx->active_oncpu--;
584         ctx->nr_active--;
585         if (event->attr.exclusive || !cpuctx->active_oncpu)
586                 cpuctx->exclusive = 0;
587 }
588
589 static void
590 group_sched_out(struct perf_event *group_event,
591                 struct perf_cpu_context *cpuctx,
592                 struct perf_event_context *ctx)
593 {
594         struct perf_event *event;
595         int state = group_event->state;
596
597         event_sched_out(group_event, cpuctx, ctx);
598
599         /*
600          * Schedule out siblings (if any):
601          */
602         list_for_each_entry(event, &group_event->sibling_list, group_entry)
603                 event_sched_out(event, cpuctx, ctx);
604
605         if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
606                 cpuctx->exclusive = 0;
607 }
608
609 static inline struct perf_cpu_context *
610 __get_cpu_context(struct perf_event_context *ctx)
611 {
612         return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
613 }
614
615 /*
616  * Cross CPU call to remove a performance event
617  *
618  * We disable the event on the hardware level first. After that we
619  * remove it from the context list.
620  */
621 static void __perf_event_remove_from_context(void *info)
622 {
623         struct perf_event *event = info;
624         struct perf_event_context *ctx = event->ctx;
625         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
626
627         /*
628          * If this is a task context, we need to check whether it is
629          * the current task context of this cpu. If not it has been
630          * scheduled out before the smp call arrived.
631          */
632         if (ctx->task && cpuctx->task_ctx != ctx)
633                 return;
634
635         raw_spin_lock(&ctx->lock);
636
637         event_sched_out(event, cpuctx, ctx);
638
639         list_del_event(event, ctx);
640
641         raw_spin_unlock(&ctx->lock);
642 }
643
644
645 /*
646  * Remove the event from a task's (or a CPU's) list of events.
647  *
648  * Must be called with ctx->mutex held.
649  *
650  * CPU events are removed with a smp call. For task events we only
651  * call when the task is on a CPU.
652  *
653  * If event->ctx is a cloned context, callers must make sure that
654  * every task struct that event->ctx->task could possibly point to
655  * remains valid.  This is OK when called from perf_release since
656  * that only calls us on the top-level context, which can't be a clone.
657  * When called from perf_event_exit_task, it's OK because the
658  * context has been detached from its task.
659  */
660 static void perf_event_remove_from_context(struct perf_event *event)
661 {
662         struct perf_event_context *ctx = event->ctx;
663         struct task_struct *task = ctx->task;
664
665         if (!task) {
666                 /*
667                  * Per cpu events are removed via an smp call and
668                  * the removal is always successful.
669                  */
670                 smp_call_function_single(event->cpu,
671                                          __perf_event_remove_from_context,
672                                          event, 1);
673                 return;
674         }
675
676 retry:
677         task_oncpu_function_call(task, __perf_event_remove_from_context,
678                                  event);
679
680         raw_spin_lock_irq(&ctx->lock);
681         /*
682          * If the context is active we need to retry the smp call.
683          */
684         if (ctx->nr_active && !list_empty(&event->group_entry)) {
685                 raw_spin_unlock_irq(&ctx->lock);
686                 goto retry;
687         }
688
689         /*
690          * The lock prevents that this context is scheduled in so we
691          * can remove the event safely, if the call above did not
692          * succeed.
693          */
694         if (!list_empty(&event->group_entry))
695                 list_del_event(event, ctx);
696         raw_spin_unlock_irq(&ctx->lock);
697 }
698
699 /*
700  * Cross CPU call to disable a performance event
701  */
702 static void __perf_event_disable(void *info)
703 {
704         struct perf_event *event = info;
705         struct perf_event_context *ctx = event->ctx;
706         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
707
708         /*
709          * If this is a per-task event, need to check whether this
710          * event's task is the current task on this cpu.
711          */
712         if (ctx->task && cpuctx->task_ctx != ctx)
713                 return;
714
715         raw_spin_lock(&ctx->lock);
716
717         /*
718          * If the event is on, turn it off.
719          * If it is in error state, leave it in error state.
720          */
721         if (event->state >= PERF_EVENT_STATE_INACTIVE) {
722                 update_context_time(ctx);
723                 update_group_times(event);
724                 if (event == event->group_leader)
725                         group_sched_out(event, cpuctx, ctx);
726                 else
727                         event_sched_out(event, cpuctx, ctx);
728                 event->state = PERF_EVENT_STATE_OFF;
729         }
730
731         raw_spin_unlock(&ctx->lock);
732 }
733
734 /*
735  * Disable a event.
736  *
737  * If event->ctx is a cloned context, callers must make sure that
738  * every task struct that event->ctx->task could possibly point to
739  * remains valid.  This condition is satisifed when called through
740  * perf_event_for_each_child or perf_event_for_each because they
741  * hold the top-level event's child_mutex, so any descendant that
742  * goes to exit will block in sync_child_event.
743  * When called from perf_pending_event it's OK because event->ctx
744  * is the current context on this CPU and preemption is disabled,
745  * hence we can't get into perf_event_task_sched_out for this context.
746  */
747 void perf_event_disable(struct perf_event *event)
748 {
749         struct perf_event_context *ctx = event->ctx;
750         struct task_struct *task = ctx->task;
751
752         if (!task) {
753                 /*
754                  * Disable the event on the cpu that it's on
755                  */
756                 smp_call_function_single(event->cpu, __perf_event_disable,
757                                          event, 1);
758                 return;
759         }
760
761 retry:
762         task_oncpu_function_call(task, __perf_event_disable, event);
763
764         raw_spin_lock_irq(&ctx->lock);
765         /*
766          * If the event is still active, we need to retry the cross-call.
767          */
768         if (event->state == PERF_EVENT_STATE_ACTIVE) {
769                 raw_spin_unlock_irq(&ctx->lock);
770                 goto retry;
771         }
772
773         /*
774          * Since we have the lock this context can't be scheduled
775          * in, so we can change the state safely.
776          */
777         if (event->state == PERF_EVENT_STATE_INACTIVE) {
778                 update_group_times(event);
779                 event->state = PERF_EVENT_STATE_OFF;
780         }
781
782         raw_spin_unlock_irq(&ctx->lock);
783 }
784
785 static int
786 event_sched_in(struct perf_event *event,
787                  struct perf_cpu_context *cpuctx,
788                  struct perf_event_context *ctx)
789 {
790         u64 tstamp = perf_event_time(event);
791
792         if (event->state <= PERF_EVENT_STATE_OFF)
793                 return 0;
794
795         event->state = PERF_EVENT_STATE_ACTIVE;
796         event->oncpu = smp_processor_id();
797         /*
798          * The new state must be visible before we turn it on in the hardware:
799          */
800         smp_wmb();
801
802         if (event->pmu->add(event, PERF_EF_START)) {
803                 event->state = PERF_EVENT_STATE_INACTIVE;
804                 event->oncpu = -1;
805                 return -EAGAIN;
806         }
807
808         event->tstamp_running += tstamp - event->tstamp_stopped;
809
810         event->shadow_ctx_time = tstamp - ctx->timestamp;
811
812         if (!is_software_event(event))
813                 cpuctx->active_oncpu++;
814         ctx->nr_active++;
815
816         if (event->attr.exclusive)
817                 cpuctx->exclusive = 1;
818
819         return 0;
820 }
821
822 static int
823 group_sched_in(struct perf_event *group_event,
824                struct perf_cpu_context *cpuctx,
825                struct perf_event_context *ctx)
826 {
827         struct perf_event *event, *partial_group = NULL;
828         struct pmu *pmu = group_event->pmu;
829         u64 now = ctx->time;
830         bool simulate = false;
831
832         if (group_event->state == PERF_EVENT_STATE_OFF)
833                 return 0;
834
835         pmu->start_txn(pmu);
836
837         if (event_sched_in(group_event, cpuctx, ctx)) {
838                 pmu->cancel_txn(pmu);
839                 return -EAGAIN;
840         }
841
842         /*
843          * Schedule in siblings as one group (if any):
844          */
845         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
846                 if (event_sched_in(event, cpuctx, ctx)) {
847                         partial_group = event;
848                         goto group_error;
849                 }
850         }
851
852         if (!pmu->commit_txn(pmu))
853                 return 0;
854
855 group_error:
856         /*
857          * Groups can be scheduled in as one unit only, so undo any
858          * partial group before returning:
859          * The events up to the failed event are scheduled out normally,
860          * tstamp_stopped will be updated.
861          *
862          * The failed events and the remaining siblings need to have
863          * their timings updated as if they had gone thru event_sched_in()
864          * and event_sched_out(). This is required to get consistent timings
865          * across the group. This also takes care of the case where the group
866          * could never be scheduled by ensuring tstamp_stopped is set to mark
867          * the time the event was actually stopped, such that time delta
868          * calculation in update_event_times() is correct.
869          */
870         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
871                 if (event == partial_group)
872                         simulate = true;
873
874                 if (simulate) {
875                         event->tstamp_running += now - event->tstamp_stopped;
876                         event->tstamp_stopped = now;
877                 } else {
878                         event_sched_out(event, cpuctx, ctx);
879                 }
880         }
881         event_sched_out(group_event, cpuctx, ctx);
882
883         pmu->cancel_txn(pmu);
884
885         return -EAGAIN;
886 }
887
888 /*
889  * Work out whether we can put this event group on the CPU now.
890  */
891 static int group_can_go_on(struct perf_event *event,
892                            struct perf_cpu_context *cpuctx,
893                            int can_add_hw)
894 {
895         /*
896          * Groups consisting entirely of software events can always go on.
897          */
898         if (event->group_flags & PERF_GROUP_SOFTWARE)
899                 return 1;
900         /*
901          * If an exclusive group is already on, no other hardware
902          * events can go on.
903          */
904         if (cpuctx->exclusive)
905                 return 0;
906         /*
907          * If this group is exclusive and there are already
908          * events on the CPU, it can't go on.
909          */
910         if (event->attr.exclusive && cpuctx->active_oncpu)
911                 return 0;
912         /*
913          * Otherwise, try to add it if all previous groups were able
914          * to go on.
915          */
916         return can_add_hw;
917 }
918
919 static void add_event_to_ctx(struct perf_event *event,
920                                struct perf_event_context *ctx)
921 {
922         u64 tstamp = perf_event_time(event);
923
924         list_add_event(event, ctx);
925         perf_group_attach(event);
926         event->tstamp_enabled = tstamp;
927         event->tstamp_running = tstamp;
928         event->tstamp_stopped = tstamp;
929 }
930
931 /*
932  * Cross CPU call to install and enable a performance event
933  *
934  * Must be called with ctx->mutex held
935  */
936 static void __perf_install_in_context(void *info)
937 {
938         struct perf_event *event = info;
939         struct perf_event_context *ctx = event->ctx;
940         struct perf_event *leader = event->group_leader;
941         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
942         int err;
943
944         /*
945          * If this is a task context, we need to check whether it is
946          * the current task context of this cpu. If not it has been
947          * scheduled out before the smp call arrived.
948          * Or possibly this is the right context but it isn't
949          * on this cpu because it had no events.
950          */
951         if (ctx->task && cpuctx->task_ctx != ctx) {
952                 if (cpuctx->task_ctx || ctx->task != current)
953                         return;
954                 cpuctx->task_ctx = ctx;
955         }
956
957         raw_spin_lock(&ctx->lock);
958         ctx->is_active = 1;
959         update_context_time(ctx);
960
961         add_event_to_ctx(event, ctx);
962
963         if (!event_filter_match(event))
964                 goto unlock;
965
966         /*
967          * Don't put the event on if it is disabled or if
968          * it is in a group and the group isn't on.
969          */
970         if (event->state != PERF_EVENT_STATE_INACTIVE ||
971             (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
972                 goto unlock;
973
974         /*
975          * An exclusive event can't go on if there are already active
976          * hardware events, and no hardware event can go on if there
977          * is already an exclusive event on.
978          */
979         if (!group_can_go_on(event, cpuctx, 1))
980                 err = -EEXIST;
981         else
982                 err = event_sched_in(event, cpuctx, ctx);
983
984         if (err) {
985                 /*
986                  * This event couldn't go on.  If it is in a group
987                  * then we have to pull the whole group off.
988                  * If the event group is pinned then put it in error state.
989                  */
990                 if (leader != event)
991                         group_sched_out(leader, cpuctx, ctx);
992                 if (leader->attr.pinned) {
993                         update_group_times(leader);
994                         leader->state = PERF_EVENT_STATE_ERROR;
995                 }
996         }
997
998 unlock:
999         raw_spin_unlock(&ctx->lock);
1000 }
1001
1002 /*
1003  * Attach a performance event to a context
1004  *
1005  * First we add the event to the list with the hardware enable bit
1006  * in event->hw_config cleared.
1007  *
1008  * If the event is attached to a task which is on a CPU we use a smp
1009  * call to enable it in the task context. The task might have been
1010  * scheduled away, but we check this in the smp call again.
1011  *
1012  * Must be called with ctx->mutex held.
1013  */
1014 static void
1015 perf_install_in_context(struct perf_event_context *ctx,
1016                         struct perf_event *event,
1017                         int cpu)
1018 {
1019         struct task_struct *task = ctx->task;
1020
1021         event->ctx = ctx;
1022
1023         if (!task) {
1024                 /*
1025                  * Per cpu events are installed via an smp call and
1026                  * the install is always successful.
1027                  */
1028                 smp_call_function_single(cpu, __perf_install_in_context,
1029                                          event, 1);
1030                 return;
1031         }
1032
1033 retry:
1034         task_oncpu_function_call(task, __perf_install_in_context,
1035                                  event);
1036
1037         raw_spin_lock_irq(&ctx->lock);
1038         /*
1039          * we need to retry the smp call.
1040          */
1041         if (ctx->is_active && list_empty(&event->group_entry)) {
1042                 raw_spin_unlock_irq(&ctx->lock);
1043                 goto retry;
1044         }
1045
1046         /*
1047          * The lock prevents that this context is scheduled in so we
1048          * can add the event safely, if it the call above did not
1049          * succeed.
1050          */
1051         if (list_empty(&event->group_entry))
1052                 add_event_to_ctx(event, ctx);
1053         raw_spin_unlock_irq(&ctx->lock);
1054 }
1055
1056 /*
1057  * Put a event into inactive state and update time fields.
1058  * Enabling the leader of a group effectively enables all
1059  * the group members that aren't explicitly disabled, so we
1060  * have to update their ->tstamp_enabled also.
1061  * Note: this works for group members as well as group leaders
1062  * since the non-leader members' sibling_lists will be empty.
1063  */
1064 static void __perf_event_mark_enabled(struct perf_event *event,
1065                                         struct perf_event_context *ctx)
1066 {
1067         struct perf_event *sub;
1068         u64 tstamp = perf_event_time(event);
1069
1070         event->state = PERF_EVENT_STATE_INACTIVE;
1071         event->tstamp_enabled = tstamp - event->total_time_enabled;
1072         list_for_each_entry(sub, &event->sibling_list, group_entry) {
1073                 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1074                         sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1075         }
1076 }
1077
1078 /*
1079  * Cross CPU call to enable a performance event
1080  */
1081 static void __perf_event_enable(void *info)
1082 {
1083         struct perf_event *event = info;
1084         struct perf_event_context *ctx = event->ctx;
1085         struct perf_event *leader = event->group_leader;
1086         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1087         int err;
1088
1089         /*
1090          * If this is a per-task event, need to check whether this
1091          * event's task is the current task on this cpu.
1092          */
1093         if (ctx->task && cpuctx->task_ctx != ctx) {
1094                 if (cpuctx->task_ctx || ctx->task != current)
1095                         return;
1096                 cpuctx->task_ctx = ctx;
1097         }
1098
1099         raw_spin_lock(&ctx->lock);
1100         ctx->is_active = 1;
1101         update_context_time(ctx);
1102
1103         if (event->state >= PERF_EVENT_STATE_INACTIVE)
1104                 goto unlock;
1105         __perf_event_mark_enabled(event, ctx);
1106
1107         if (!event_filter_match(event))
1108                 goto unlock;
1109
1110         /*
1111          * If the event is in a group and isn't the group leader,
1112          * then don't put it on unless the group is on.
1113          */
1114         if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
1115                 goto unlock;
1116
1117         if (!group_can_go_on(event, cpuctx, 1)) {
1118                 err = -EEXIST;
1119         } else {
1120                 if (event == leader)
1121                         err = group_sched_in(event, cpuctx, ctx);
1122                 else
1123                         err = event_sched_in(event, cpuctx, ctx);
1124         }
1125
1126         if (err) {
1127                 /*
1128                  * If this event can't go on and it's part of a
1129                  * group, then the whole group has to come off.
1130                  */
1131                 if (leader != event)
1132                         group_sched_out(leader, cpuctx, ctx);
1133                 if (leader->attr.pinned) {
1134                         update_group_times(leader);
1135                         leader->state = PERF_EVENT_STATE_ERROR;
1136                 }
1137         }
1138
1139 unlock:
1140         raw_spin_unlock(&ctx->lock);
1141 }
1142
1143 /*
1144  * Enable a event.
1145  *
1146  * If event->ctx is a cloned context, callers must make sure that
1147  * every task struct that event->ctx->task could possibly point to
1148  * remains valid.  This condition is satisfied when called through
1149  * perf_event_for_each_child or perf_event_for_each as described
1150  * for perf_event_disable.
1151  */
1152 void perf_event_enable(struct perf_event *event)
1153 {
1154         struct perf_event_context *ctx = event->ctx;
1155         struct task_struct *task = ctx->task;
1156
1157         if (!task) {
1158                 /*
1159                  * Enable the event on the cpu that it's on
1160                  */
1161                 smp_call_function_single(event->cpu, __perf_event_enable,
1162                                          event, 1);
1163                 return;
1164         }
1165
1166         raw_spin_lock_irq(&ctx->lock);
1167         if (event->state >= PERF_EVENT_STATE_INACTIVE)
1168                 goto out;
1169
1170         /*
1171          * If the event is in error state, clear that first.
1172          * That way, if we see the event in error state below, we
1173          * know that it has gone back into error state, as distinct
1174          * from the task having been scheduled away before the
1175          * cross-call arrived.
1176          */
1177         if (event->state == PERF_EVENT_STATE_ERROR)
1178                 event->state = PERF_EVENT_STATE_OFF;
1179
1180 retry:
1181         raw_spin_unlock_irq(&ctx->lock);
1182         task_oncpu_function_call(task, __perf_event_enable, event);
1183
1184         raw_spin_lock_irq(&ctx->lock);
1185
1186         /*
1187          * If the context is active and the event is still off,
1188          * we need to retry the cross-call.
1189          */
1190         if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1191                 goto retry;
1192
1193         /*
1194          * Since we have the lock this context can't be scheduled
1195          * in, so we can change the state safely.
1196          */
1197         if (event->state == PERF_EVENT_STATE_OFF)
1198                 __perf_event_mark_enabled(event, ctx);
1199
1200 out:
1201         raw_spin_unlock_irq(&ctx->lock);
1202 }
1203
1204 static int perf_event_refresh(struct perf_event *event, int refresh)
1205 {
1206         /*
1207          * not supported on inherited events
1208          */
1209         if (event->attr.inherit || !is_sampling_event(event))
1210                 return -EINVAL;
1211
1212         atomic_add(refresh, &event->event_limit);
1213         perf_event_enable(event);
1214
1215         return 0;
1216 }
1217
1218 static void ctx_sched_out(struct perf_event_context *ctx,
1219                           struct perf_cpu_context *cpuctx,
1220                           enum event_type_t event_type)
1221 {
1222         struct perf_event *event;
1223
1224         raw_spin_lock(&ctx->lock);
1225         perf_pmu_disable(ctx->pmu);
1226         ctx->is_active = 0;
1227         if (likely(!ctx->nr_events))
1228                 goto out;
1229         update_context_time(ctx);
1230
1231         if (!ctx->nr_active)
1232                 goto out;
1233
1234         if (event_type & EVENT_PINNED) {
1235                 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1236                         group_sched_out(event, cpuctx, ctx);
1237         }
1238
1239         if (event_type & EVENT_FLEXIBLE) {
1240                 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1241                         group_sched_out(event, cpuctx, ctx);
1242         }
1243 out:
1244         perf_pmu_enable(ctx->pmu);
1245         raw_spin_unlock(&ctx->lock);
1246 }
1247
1248 /*
1249  * Test whether two contexts are equivalent, i.e. whether they
1250  * have both been cloned from the same version of the same context
1251  * and they both have the same number of enabled events.
1252  * If the number of enabled events is the same, then the set
1253  * of enabled events should be the same, because these are both
1254  * inherited contexts, therefore we can't access individual events
1255  * in them directly with an fd; we can only enable/disable all
1256  * events via prctl, or enable/disable all events in a family
1257  * via ioctl, which will have the same effect on both contexts.
1258  */
1259 static int context_equiv(struct perf_event_context *ctx1,
1260                          struct perf_event_context *ctx2)
1261 {
1262         return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1263                 && ctx1->parent_gen == ctx2->parent_gen
1264                 && !ctx1->pin_count && !ctx2->pin_count;
1265 }
1266
1267 static void __perf_event_sync_stat(struct perf_event *event,
1268                                      struct perf_event *next_event)
1269 {
1270         u64 value;
1271
1272         if (!event->attr.inherit_stat)
1273                 return;
1274
1275         /*
1276          * Update the event value, we cannot use perf_event_read()
1277          * because we're in the middle of a context switch and have IRQs
1278          * disabled, which upsets smp_call_function_single(), however
1279          * we know the event must be on the current CPU, therefore we
1280          * don't need to use it.
1281          */
1282         switch (event->state) {
1283         case PERF_EVENT_STATE_ACTIVE:
1284                 event->pmu->read(event);
1285                 /* fall-through */
1286
1287         case PERF_EVENT_STATE_INACTIVE:
1288                 update_event_times(event);
1289                 break;
1290
1291         default:
1292                 break;
1293         }
1294
1295         /*
1296          * In order to keep per-task stats reliable we need to flip the event
1297          * values when we flip the contexts.
1298          */
1299         value = local64_read(&next_event->count);
1300         value = local64_xchg(&event->count, value);
1301         local64_set(&next_event->count, value);
1302
1303         swap(event->total_time_enabled, next_event->total_time_enabled);
1304         swap(event->total_time_running, next_event->total_time_running);
1305
1306         /*
1307          * Since we swizzled the values, update the user visible data too.
1308          */
1309         perf_event_update_userpage(event);
1310         perf_event_update_userpage(next_event);
1311 }
1312
1313 #define list_next_entry(pos, member) \
1314         list_entry(pos->member.next, typeof(*pos), member)
1315
1316 static void perf_event_sync_stat(struct perf_event_context *ctx,
1317                                    struct perf_event_context *next_ctx)
1318 {
1319         struct perf_event *event, *next_event;
1320
1321         if (!ctx->nr_stat)
1322                 return;
1323
1324         update_context_time(ctx);
1325
1326         event = list_first_entry(&ctx->event_list,
1327                                    struct perf_event, event_entry);
1328
1329         next_event = list_first_entry(&next_ctx->event_list,
1330                                         struct perf_event, event_entry);
1331
1332         while (&event->event_entry != &ctx->event_list &&
1333                &next_event->event_entry != &next_ctx->event_list) {
1334
1335                 __perf_event_sync_stat(event, next_event);
1336
1337                 event = list_next_entry(event, event_entry);
1338                 next_event = list_next_entry(next_event, event_entry);
1339         }
1340 }
1341
1342 void perf_event_context_sched_out(struct task_struct *task, int ctxn,
1343                                   struct task_struct *next)
1344 {
1345         struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
1346         struct perf_event_context *next_ctx;
1347         struct perf_event_context *parent;
1348         struct perf_cpu_context *cpuctx;
1349         int do_switch = 1;
1350
1351         if (likely(!ctx))
1352                 return;
1353
1354         cpuctx = __get_cpu_context(ctx);
1355         if (!cpuctx->task_ctx)
1356                 return;
1357
1358         rcu_read_lock();
1359         parent = rcu_dereference(ctx->parent_ctx);
1360         next_ctx = next->perf_event_ctxp[ctxn];
1361         if (parent && next_ctx &&
1362             rcu_dereference(next_ctx->parent_ctx) == parent) {
1363                 /*
1364                  * Looks like the two contexts are clones, so we might be
1365                  * able to optimize the context switch.  We lock both
1366                  * contexts and check that they are clones under the
1367                  * lock (including re-checking that neither has been
1368                  * uncloned in the meantime).  It doesn't matter which
1369                  * order we take the locks because no other cpu could
1370                  * be trying to lock both of these tasks.
1371                  */
1372                 raw_spin_lock(&ctx->lock);
1373                 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1374                 if (context_equiv(ctx, next_ctx)) {
1375                         /*
1376                          * XXX do we need a memory barrier of sorts
1377                          * wrt to rcu_dereference() of perf_event_ctxp
1378                          */
1379                         task->perf_event_ctxp[ctxn] = next_ctx;
1380                         next->perf_event_ctxp[ctxn] = ctx;
1381                         ctx->task = next;
1382                         next_ctx->task = task;
1383                         do_switch = 0;
1384
1385                         perf_event_sync_stat(ctx, next_ctx);
1386                 }
1387                 raw_spin_unlock(&next_ctx->lock);
1388                 raw_spin_unlock(&ctx->lock);
1389         }
1390         rcu_read_unlock();
1391
1392         if (do_switch) {
1393                 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1394                 cpuctx->task_ctx = NULL;
1395         }
1396 }
1397
1398 #define for_each_task_context_nr(ctxn)                                  \
1399         for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
1400
1401 /*
1402  * Called from scheduler to remove the events of the current task,
1403  * with interrupts disabled.
1404  *
1405  * We stop each event and update the event value in event->count.
1406  *
1407  * This does not protect us against NMI, but disable()
1408  * sets the disabled bit in the control field of event _before_
1409  * accessing the event control register. If a NMI hits, then it will
1410  * not restart the event.
1411  */
1412 void __perf_event_task_sched_out(struct task_struct *task,
1413                                  struct task_struct *next)
1414 {
1415         int ctxn;
1416
1417         for_each_task_context_nr(ctxn)
1418                 perf_event_context_sched_out(task, ctxn, next);
1419 }
1420
1421 static void task_ctx_sched_out(struct perf_event_context *ctx,
1422                                enum event_type_t event_type)
1423 {
1424         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1425
1426         if (!cpuctx->task_ctx)
1427                 return;
1428
1429         if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1430                 return;
1431
1432         ctx_sched_out(ctx, cpuctx, event_type);
1433         cpuctx->task_ctx = NULL;
1434 }
1435
1436 /*
1437  * Called with IRQs disabled
1438  */
1439 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1440                               enum event_type_t event_type)
1441 {
1442         ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1443 }
1444
1445 static void
1446 ctx_pinned_sched_in(struct perf_event_context *ctx,
1447                     struct perf_cpu_context *cpuctx)
1448 {
1449         struct perf_event *event;
1450
1451         list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1452                 if (event->state <= PERF_EVENT_STATE_OFF)
1453                         continue;
1454                 if (!event_filter_match(event))
1455                         continue;
1456
1457                 if (group_can_go_on(event, cpuctx, 1))
1458                         group_sched_in(event, cpuctx, ctx);
1459
1460                 /*
1461                  * If this pinned group hasn't been scheduled,
1462                  * put it in error state.
1463                  */
1464                 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1465                         update_group_times(event);
1466                         event->state = PERF_EVENT_STATE_ERROR;
1467                 }
1468         }
1469 }
1470
1471 static void
1472 ctx_flexible_sched_in(struct perf_event_context *ctx,
1473                       struct perf_cpu_context *cpuctx)
1474 {
1475         struct perf_event *event;
1476         int can_add_hw = 1;
1477
1478         list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1479                 /* Ignore events in OFF or ERROR state */
1480                 if (event->state <= PERF_EVENT_STATE_OFF)
1481                         continue;
1482                 /*
1483                  * Listen to the 'cpu' scheduling filter constraint
1484                  * of events:
1485                  */
1486                 if (!event_filter_match(event))
1487                         continue;
1488
1489                 if (group_can_go_on(event, cpuctx, can_add_hw)) {
1490                         if (group_sched_in(event, cpuctx, ctx))
1491                                 can_add_hw = 0;
1492                 }
1493         }
1494 }
1495
1496 static void
1497 ctx_sched_in(struct perf_event_context *ctx,
1498              struct perf_cpu_context *cpuctx,
1499              enum event_type_t event_type)
1500 {
1501         raw_spin_lock(&ctx->lock);
1502         ctx->is_active = 1;
1503         if (likely(!ctx->nr_events))
1504                 goto out;
1505
1506         ctx->timestamp = perf_clock();
1507
1508         /*
1509          * First go through the list and put on any pinned groups
1510          * in order to give them the best chance of going on.
1511          */
1512         if (event_type & EVENT_PINNED)
1513                 ctx_pinned_sched_in(ctx, cpuctx);
1514
1515         /* Then walk through the lower prio flexible groups */
1516         if (event_type & EVENT_FLEXIBLE)
1517                 ctx_flexible_sched_in(ctx, cpuctx);
1518
1519 out:
1520         raw_spin_unlock(&ctx->lock);
1521 }
1522
1523 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1524                              enum event_type_t event_type)
1525 {
1526         struct perf_event_context *ctx = &cpuctx->ctx;
1527
1528         ctx_sched_in(ctx, cpuctx, event_type);
1529 }
1530
1531 static void task_ctx_sched_in(struct perf_event_context *ctx,
1532                               enum event_type_t event_type)
1533 {
1534         struct perf_cpu_context *cpuctx;
1535
1536         cpuctx = __get_cpu_context(ctx);
1537         if (cpuctx->task_ctx == ctx)
1538                 return;
1539
1540         ctx_sched_in(ctx, cpuctx, event_type);
1541         cpuctx->task_ctx = ctx;
1542 }
1543
1544 void perf_event_context_sched_in(struct perf_event_context *ctx)
1545 {
1546         struct perf_cpu_context *cpuctx;
1547
1548         cpuctx = __get_cpu_context(ctx);
1549         if (cpuctx->task_ctx == ctx)
1550                 return;
1551
1552         perf_pmu_disable(ctx->pmu);
1553         /*
1554          * We want to keep the following priority order:
1555          * cpu pinned (that don't need to move), task pinned,
1556          * cpu flexible, task flexible.
1557          */
1558         cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1559
1560         ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1561         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1562         ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1563
1564         cpuctx->task_ctx = ctx;
1565
1566         /*
1567          * Since these rotations are per-cpu, we need to ensure the
1568          * cpu-context we got scheduled on is actually rotating.
1569          */
1570         perf_pmu_rotate_start(ctx->pmu);
1571         perf_pmu_enable(ctx->pmu);
1572 }
1573
1574 /*
1575  * Called from scheduler to add the events of the current task
1576  * with interrupts disabled.
1577  *
1578  * We restore the event value and then enable it.
1579  *
1580  * This does not protect us against NMI, but enable()
1581  * sets the enabled bit in the control field of event _before_
1582  * accessing the event control register. If a NMI hits, then it will
1583  * keep the event running.
1584  */
1585 void __perf_event_task_sched_in(struct task_struct *task)
1586 {
1587         struct perf_event_context *ctx;
1588         int ctxn;
1589
1590         for_each_task_context_nr(ctxn) {
1591                 ctx = task->perf_event_ctxp[ctxn];
1592                 if (likely(!ctx))
1593                         continue;
1594
1595                 perf_event_context_sched_in(ctx);
1596         }
1597 }
1598
1599 #define MAX_INTERRUPTS (~0ULL)
1600
1601 static void perf_log_throttle(struct perf_event *event, int enable);
1602
1603 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1604 {
1605         u64 frequency = event->attr.sample_freq;
1606         u64 sec = NSEC_PER_SEC;
1607         u64 divisor, dividend;
1608
1609         int count_fls, nsec_fls, frequency_fls, sec_fls;
1610
1611         count_fls = fls64(count);
1612         nsec_fls = fls64(nsec);
1613         frequency_fls = fls64(frequency);
1614         sec_fls = 30;
1615
1616         /*
1617          * We got @count in @nsec, with a target of sample_freq HZ
1618          * the target period becomes:
1619          *
1620          *             @count * 10^9
1621          * period = -------------------
1622          *          @nsec * sample_freq
1623          *
1624          */
1625
1626         /*
1627          * Reduce accuracy by one bit such that @a and @b converge
1628          * to a similar magnitude.
1629          */
1630 #define REDUCE_FLS(a, b)                \
1631 do {                                    \
1632         if (a##_fls > b##_fls) {        \
1633                 a >>= 1;                \
1634                 a##_fls--;              \
1635         } else {                        \
1636                 b >>= 1;                \
1637                 b##_fls--;              \
1638         }                               \
1639 } while (0)
1640
1641         /*
1642          * Reduce accuracy until either term fits in a u64, then proceed with
1643          * the other, so that finally we can do a u64/u64 division.
1644          */
1645         while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1646                 REDUCE_FLS(nsec, frequency);
1647                 REDUCE_FLS(sec, count);
1648         }
1649
1650         if (count_fls + sec_fls > 64) {
1651                 divisor = nsec * frequency;
1652
1653                 while (count_fls + sec_fls > 64) {
1654                         REDUCE_FLS(count, sec);
1655                         divisor >>= 1;
1656                 }
1657
1658                 dividend = count * sec;
1659         } else {
1660                 dividend = count * sec;
1661
1662                 while (nsec_fls + frequency_fls > 64) {
1663                         REDUCE_FLS(nsec, frequency);
1664                         dividend >>= 1;
1665                 }
1666
1667                 divisor = nsec * frequency;
1668         }
1669
1670         if (!divisor)
1671                 return dividend;
1672
1673         return div64_u64(dividend, divisor);
1674 }
1675
1676 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1677 {
1678         struct hw_perf_event *hwc = &event->hw;
1679         s64 period, sample_period;
1680         s64 delta;
1681
1682         period = perf_calculate_period(event, nsec, count);
1683
1684         delta = (s64)(period - hwc->sample_period);
1685         delta = (delta + 7) / 8; /* low pass filter */
1686
1687         sample_period = hwc->sample_period + delta;
1688
1689         if (!sample_period)
1690                 sample_period = 1;
1691
1692         hwc->sample_period = sample_period;
1693
1694         if (local64_read(&hwc->period_left) > 8*sample_period) {
1695                 event->pmu->stop(event, PERF_EF_UPDATE);
1696                 local64_set(&hwc->period_left, 0);
1697                 event->pmu->start(event, PERF_EF_RELOAD);
1698         }
1699 }
1700
1701 static void perf_ctx_adjust_freq(struct perf_event_context *ctx, u64 period)
1702 {
1703         struct perf_event *event;
1704         struct hw_perf_event *hwc;
1705         u64 interrupts, now;
1706         s64 delta;
1707
1708         raw_spin_lock(&ctx->lock);
1709         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1710                 if (event->state != PERF_EVENT_STATE_ACTIVE)
1711                         continue;
1712
1713                 if (!event_filter_match(event))
1714                         continue;
1715
1716                 hwc = &event->hw;
1717
1718                 interrupts = hwc->interrupts;
1719                 hwc->interrupts = 0;
1720
1721                 /*
1722                  * unthrottle events on the tick
1723                  */
1724                 if (interrupts == MAX_INTERRUPTS) {
1725                         perf_log_throttle(event, 1);
1726                         event->pmu->start(event, 0);
1727                 }
1728
1729                 if (!event->attr.freq || !event->attr.sample_freq)
1730                         continue;
1731
1732                 event->pmu->read(event);
1733                 now = local64_read(&event->count);
1734                 delta = now - hwc->freq_count_stamp;
1735                 hwc->freq_count_stamp = now;
1736
1737                 if (delta > 0)
1738                         perf_adjust_period(event, period, delta);
1739         }
1740         raw_spin_unlock(&ctx->lock);
1741 }
1742
1743 /*
1744  * Round-robin a context's events:
1745  */
1746 static void rotate_ctx(struct perf_event_context *ctx)
1747 {
1748         raw_spin_lock(&ctx->lock);
1749
1750         /*
1751          * Rotate the first entry last of non-pinned groups. Rotation might be
1752          * disabled by the inheritance code.
1753          */
1754         if (!ctx->rotate_disable)
1755                 list_rotate_left(&ctx->flexible_groups);
1756
1757         raw_spin_unlock(&ctx->lock);
1758 }
1759
1760 /*
1761  * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
1762  * because they're strictly cpu affine and rotate_start is called with IRQs
1763  * disabled, while rotate_context is called from IRQ context.
1764  */
1765 static void perf_rotate_context(struct perf_cpu_context *cpuctx)
1766 {
1767         u64 interval = (u64)cpuctx->jiffies_interval * TICK_NSEC;
1768         struct perf_event_context *ctx = NULL;
1769         int rotate = 0, remove = 1;
1770
1771         if (cpuctx->ctx.nr_events) {
1772                 remove = 0;
1773                 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1774                         rotate = 1;
1775         }
1776
1777         ctx = cpuctx->task_ctx;
1778         if (ctx && ctx->nr_events) {
1779                 remove = 0;
1780                 if (ctx->nr_events != ctx->nr_active)
1781                         rotate = 1;
1782         }
1783
1784         perf_pmu_disable(cpuctx->ctx.pmu);
1785         perf_ctx_adjust_freq(&cpuctx->ctx, interval);
1786         if (ctx)
1787                 perf_ctx_adjust_freq(ctx, interval);
1788
1789         if (!rotate)
1790                 goto done;
1791
1792         cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1793         if (ctx)
1794                 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1795
1796         rotate_ctx(&cpuctx->ctx);
1797         if (ctx)
1798                 rotate_ctx(ctx);
1799
1800         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1801         if (ctx)
1802                 task_ctx_sched_in(ctx, EVENT_FLEXIBLE);
1803
1804 done:
1805         if (remove)
1806                 list_del_init(&cpuctx->rotation_list);
1807
1808         perf_pmu_enable(cpuctx->ctx.pmu);
1809 }
1810
1811 void perf_event_task_tick(void)
1812 {
1813         struct list_head *head = &__get_cpu_var(rotation_list);
1814         struct perf_cpu_context *cpuctx, *tmp;
1815
1816         WARN_ON(!irqs_disabled());
1817
1818         list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
1819                 if (cpuctx->jiffies_interval == 1 ||
1820                                 !(jiffies % cpuctx->jiffies_interval))
1821                         perf_rotate_context(cpuctx);
1822         }
1823 }
1824
1825 static int event_enable_on_exec(struct perf_event *event,
1826                                 struct perf_event_context *ctx)
1827 {
1828         if (!event->attr.enable_on_exec)
1829                 return 0;
1830
1831         event->attr.enable_on_exec = 0;
1832         if (event->state >= PERF_EVENT_STATE_INACTIVE)
1833                 return 0;
1834
1835         __perf_event_mark_enabled(event, ctx);
1836
1837         return 1;
1838 }
1839
1840 /*
1841  * Enable all of a task's events that have been marked enable-on-exec.
1842  * This expects task == current.
1843  */
1844 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
1845 {
1846         struct perf_event *event;
1847         unsigned long flags;
1848         int enabled = 0;
1849         int ret;
1850
1851         local_irq_save(flags);
1852         if (!ctx || !ctx->nr_events)
1853                 goto out;
1854
1855         task_ctx_sched_out(ctx, EVENT_ALL);
1856
1857         raw_spin_lock(&ctx->lock);
1858
1859         list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1860                 ret = event_enable_on_exec(event, ctx);
1861                 if (ret)
1862                         enabled = 1;
1863         }
1864
1865         list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1866                 ret = event_enable_on_exec(event, ctx);
1867                 if (ret)
1868                         enabled = 1;
1869         }
1870
1871         /*
1872          * Unclone this context if we enabled any event.
1873          */
1874         if (enabled)
1875                 unclone_ctx(ctx);
1876
1877         raw_spin_unlock(&ctx->lock);
1878
1879         perf_event_context_sched_in(ctx);
1880 out:
1881         local_irq_restore(flags);
1882 }
1883
1884 /*
1885  * Cross CPU call to read the hardware event
1886  */
1887 static void __perf_event_read(void *info)
1888 {
1889         struct perf_event *event = info;
1890         struct perf_event_context *ctx = event->ctx;
1891         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1892
1893         /*
1894          * If this is a task context, we need to check whether it is
1895          * the current task context of this cpu.  If not it has been
1896          * scheduled out before the smp call arrived.  In that case
1897          * event->count would have been updated to a recent sample
1898          * when the event was scheduled out.
1899          */
1900         if (ctx->task && cpuctx->task_ctx != ctx)
1901                 return;
1902
1903         raw_spin_lock(&ctx->lock);
1904         update_context_time(ctx);
1905         update_event_times(event);
1906         raw_spin_unlock(&ctx->lock);
1907
1908         event->pmu->read(event);
1909 }
1910
1911 static inline u64 perf_event_count(struct perf_event *event)
1912 {
1913         return local64_read(&event->count) + atomic64_read(&event->child_count);
1914 }
1915
1916 static u64 perf_event_read(struct perf_event *event)
1917 {
1918         /*
1919          * If event is enabled and currently active on a CPU, update the
1920          * value in the event structure:
1921          */
1922         if (event->state == PERF_EVENT_STATE_ACTIVE) {
1923                 smp_call_function_single(event->oncpu,
1924                                          __perf_event_read, event, 1);
1925         } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1926                 struct perf_event_context *ctx = event->ctx;
1927                 unsigned long flags;
1928
1929                 raw_spin_lock_irqsave(&ctx->lock, flags);
1930                 /*
1931                  * may read while context is not active
1932                  * (e.g., thread is blocked), in that case
1933                  * we cannot update context time
1934                  */
1935                 if (ctx->is_active)
1936                         update_context_time(ctx);
1937                 update_event_times(event);
1938                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1939         }
1940
1941         return perf_event_count(event);
1942 }
1943
1944 /*
1945  * Callchain support
1946  */
1947
1948 struct callchain_cpus_entries {
1949         struct rcu_head                 rcu_head;
1950         struct perf_callchain_entry     *cpu_entries[0];
1951 };
1952
1953 static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
1954 static atomic_t nr_callchain_events;
1955 static DEFINE_MUTEX(callchain_mutex);
1956 struct callchain_cpus_entries *callchain_cpus_entries;
1957
1958
1959 __weak void perf_callchain_kernel(struct perf_callchain_entry *entry,
1960                                   struct pt_regs *regs)
1961 {
1962 }
1963
1964 __weak void perf_callchain_user(struct perf_callchain_entry *entry,
1965                                 struct pt_regs *regs)
1966 {
1967 }
1968
1969 static void release_callchain_buffers_rcu(struct rcu_head *head)
1970 {
1971         struct callchain_cpus_entries *entries;
1972         int cpu;
1973
1974         entries = container_of(head, struct callchain_cpus_entries, rcu_head);
1975
1976         for_each_possible_cpu(cpu)
1977                 kfree(entries->cpu_entries[cpu]);
1978
1979         kfree(entries);
1980 }
1981
1982 static void release_callchain_buffers(void)
1983 {
1984         struct callchain_cpus_entries *entries;
1985
1986         entries = callchain_cpus_entries;
1987         rcu_assign_pointer(callchain_cpus_entries, NULL);
1988         call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
1989 }
1990
1991 static int alloc_callchain_buffers(void)
1992 {
1993         int cpu;
1994         int size;
1995         struct callchain_cpus_entries *entries;
1996
1997         /*
1998          * We can't use the percpu allocation API for data that can be
1999          * accessed from NMI. Use a temporary manual per cpu allocation
2000          * until that gets sorted out.
2001          */
2002         size = sizeof(*entries) + sizeof(struct perf_callchain_entry *) *
2003                 num_possible_cpus();
2004
2005         entries = kzalloc(size, GFP_KERNEL);
2006         if (!entries)
2007                 return -ENOMEM;
2008
2009         size = sizeof(struct perf_callchain_entry) * PERF_NR_CONTEXTS;
2010
2011         for_each_possible_cpu(cpu) {
2012                 entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
2013                                                          cpu_to_node(cpu));
2014                 if (!entries->cpu_entries[cpu])
2015                         goto fail;
2016         }
2017
2018         rcu_assign_pointer(callchain_cpus_entries, entries);
2019
2020         return 0;
2021
2022 fail:
2023         for_each_possible_cpu(cpu)
2024                 kfree(entries->cpu_entries[cpu]);
2025         kfree(entries);
2026
2027         return -ENOMEM;
2028 }
2029
2030 static int get_callchain_buffers(void)
2031 {
2032         int err = 0;
2033         int count;
2034
2035         mutex_lock(&callchain_mutex);
2036
2037         count = atomic_inc_return(&nr_callchain_events);
2038         if (WARN_ON_ONCE(count < 1)) {
2039                 err = -EINVAL;
2040                 goto exit;
2041         }
2042
2043         if (count > 1) {
2044                 /* If the allocation failed, give up */
2045                 if (!callchain_cpus_entries)
2046                         err = -ENOMEM;
2047                 goto exit;
2048         }
2049
2050         err = alloc_callchain_buffers();
2051         if (err)
2052                 release_callchain_buffers();
2053 exit:
2054         mutex_unlock(&callchain_mutex);
2055
2056         return err;
2057 }
2058
2059 static void put_callchain_buffers(void)
2060 {
2061         if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
2062                 release_callchain_buffers();
2063                 mutex_unlock(&callchain_mutex);
2064         }
2065 }
2066
2067 static int get_recursion_context(int *recursion)
2068 {
2069         int rctx;
2070
2071         if (in_nmi())
2072                 rctx = 3;
2073         else if (in_irq())
2074                 rctx = 2;
2075         else if (in_softirq())
2076                 rctx = 1;
2077         else
2078                 rctx = 0;
2079
2080         if (recursion[rctx])
2081                 return -1;
2082
2083         recursion[rctx]++;
2084         barrier();
2085
2086         return rctx;
2087 }
2088
2089 static inline void put_recursion_context(int *recursion, int rctx)
2090 {
2091         barrier();
2092         recursion[rctx]--;
2093 }
2094
2095 static struct perf_callchain_entry *get_callchain_entry(int *rctx)
2096 {
2097         int cpu;
2098         struct callchain_cpus_entries *entries;
2099
2100         *rctx = get_recursion_context(__get_cpu_var(callchain_recursion));
2101         if (*rctx == -1)
2102                 return NULL;
2103
2104         entries = rcu_dereference(callchain_cpus_entries);
2105         if (!entries)
2106                 return NULL;
2107
2108         cpu = smp_processor_id();
2109
2110         return &entries->cpu_entries[cpu][*rctx];
2111 }
2112
2113 static void
2114 put_callchain_entry(int rctx)
2115 {
2116         put_recursion_context(__get_cpu_var(callchain_recursion), rctx);
2117 }
2118
2119 static struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2120 {
2121         int rctx;
2122         struct perf_callchain_entry *entry;
2123
2124
2125         entry = get_callchain_entry(&rctx);
2126         if (rctx == -1)
2127                 return NULL;
2128
2129         if (!entry)
2130                 goto exit_put;
2131
2132         entry->nr = 0;
2133
2134         if (!user_mode(regs)) {
2135                 perf_callchain_store(entry, PERF_CONTEXT_KERNEL);
2136                 perf_callchain_kernel(entry, regs);
2137                 if (current->mm)
2138                         regs = task_pt_regs(current);
2139                 else
2140                         regs = NULL;
2141         }
2142
2143         if (regs) {
2144                 perf_callchain_store(entry, PERF_CONTEXT_USER);
2145                 perf_callchain_user(entry, regs);
2146         }
2147
2148 exit_put:
2149         put_callchain_entry(rctx);
2150
2151         return entry;
2152 }
2153
2154 /*
2155  * Initialize the perf_event context in a task_struct:
2156  */
2157 static void __perf_event_init_context(struct perf_event_context *ctx)
2158 {
2159         raw_spin_lock_init(&ctx->lock);
2160         mutex_init(&ctx->mutex);
2161         INIT_LIST_HEAD(&ctx->pinned_groups);
2162         INIT_LIST_HEAD(&ctx->flexible_groups);
2163         INIT_LIST_HEAD(&ctx->event_list);
2164         atomic_set(&ctx->refcount, 1);
2165 }
2166
2167 static struct perf_event_context *
2168 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2169 {
2170         struct perf_event_context *ctx;
2171
2172         ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2173         if (!ctx)
2174                 return NULL;
2175
2176         __perf_event_init_context(ctx);
2177         if (task) {
2178                 ctx->task = task;
2179                 get_task_struct(task);
2180         }
2181         ctx->pmu = pmu;
2182
2183         return ctx;
2184 }
2185
2186 static struct task_struct *
2187 find_lively_task_by_vpid(pid_t vpid)
2188 {
2189         struct task_struct *task;
2190         int err;
2191
2192         rcu_read_lock();
2193         if (!vpid)
2194                 task = current;
2195         else
2196                 task = find_task_by_vpid(vpid);
2197         if (task)
2198                 get_task_struct(task);
2199         rcu_read_unlock();
2200
2201         if (!task)
2202                 return ERR_PTR(-ESRCH);
2203
2204         /* Reuse ptrace permission checks for now. */
2205         err = -EACCES;
2206         if (!ptrace_may_access(task, PTRACE_MODE_READ))
2207                 goto errout;
2208
2209         return task;
2210 errout:
2211         put_task_struct(task);
2212         return ERR_PTR(err);
2213
2214 }
2215
2216 static struct perf_event_context *
2217 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
2218 {
2219         struct perf_event_context *ctx;
2220         struct perf_cpu_context *cpuctx;
2221         unsigned long flags;
2222         int ctxn, err;
2223
2224         if (!task) {
2225                 /* Must be root to operate on a CPU event: */
2226                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2227                         return ERR_PTR(-EACCES);
2228
2229                 /*
2230                  * We could be clever and allow to attach a event to an
2231                  * offline CPU and activate it when the CPU comes up, but
2232                  * that's for later.
2233                  */
2234                 if (!cpu_online(cpu))
2235                         return ERR_PTR(-ENODEV);
2236
2237                 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2238                 ctx = &cpuctx->ctx;
2239                 get_ctx(ctx);
2240
2241                 return ctx;
2242         }
2243
2244         err = -EINVAL;
2245         ctxn = pmu->task_ctx_nr;
2246         if (ctxn < 0)
2247                 goto errout;
2248
2249 retry:
2250         ctx = perf_lock_task_context(task, ctxn, &flags);
2251         if (ctx) {
2252                 unclone_ctx(ctx);
2253                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2254         }
2255
2256         if (!ctx) {
2257                 ctx = alloc_perf_context(pmu, task);
2258                 err = -ENOMEM;
2259                 if (!ctx)
2260                         goto errout;
2261
2262                 get_ctx(ctx);
2263
2264                 err = 0;
2265                 mutex_lock(&task->perf_event_mutex);
2266                 /*
2267                  * If it has already passed perf_event_exit_task().
2268                  * we must see PF_EXITING, it takes this mutex too.
2269                  */
2270                 if (task->flags & PF_EXITING)
2271                         err = -ESRCH;
2272                 else if (task->perf_event_ctxp[ctxn])
2273                         err = -EAGAIN;
2274                 else
2275                         rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
2276                 mutex_unlock(&task->perf_event_mutex);
2277
2278                 if (unlikely(err)) {
2279                         put_task_struct(task);
2280                         kfree(ctx);
2281
2282                         if (err == -EAGAIN)
2283                                 goto retry;
2284                         goto errout;
2285                 }
2286         }
2287
2288         return ctx;
2289
2290 errout:
2291         return ERR_PTR(err);
2292 }
2293
2294 static void perf_event_free_filter(struct perf_event *event);
2295
2296 static void free_event_rcu(struct rcu_head *head)
2297 {
2298         struct perf_event *event;
2299
2300         event = container_of(head, struct perf_event, rcu_head);
2301         if (event->ns)
2302                 put_pid_ns(event->ns);
2303         perf_event_free_filter(event);
2304         kfree(event);
2305 }
2306
2307 static void perf_buffer_put(struct perf_buffer *buffer);
2308
2309 static void free_event(struct perf_event *event)
2310 {
2311         irq_work_sync(&event->pending);
2312
2313         if (!event->parent) {
2314                 if (event->attach_state & PERF_ATTACH_TASK)
2315                         jump_label_dec(&perf_task_events);
2316                 if (event->attr.mmap || event->attr.mmap_data)
2317                         atomic_dec(&nr_mmap_events);
2318                 if (event->attr.comm)
2319                         atomic_dec(&nr_comm_events);
2320                 if (event->attr.task)
2321                         atomic_dec(&nr_task_events);
2322                 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2323                         put_callchain_buffers();
2324         }
2325
2326         if (event->buffer) {
2327                 perf_buffer_put(event->buffer);
2328                 event->buffer = NULL;
2329         }
2330
2331         if (event->destroy)
2332                 event->destroy(event);
2333
2334         if (event->ctx)
2335                 put_ctx(event->ctx);
2336
2337         call_rcu(&event->rcu_head, free_event_rcu);
2338 }
2339
2340 int perf_event_release_kernel(struct perf_event *event)
2341 {
2342         struct perf_event_context *ctx = event->ctx;
2343
2344         /*
2345          * Remove from the PMU, can't get re-enabled since we got
2346          * here because the last ref went.
2347          */
2348         perf_event_disable(event);
2349
2350         WARN_ON_ONCE(ctx->parent_ctx);
2351         /*
2352          * There are two ways this annotation is useful:
2353          *
2354          *  1) there is a lock recursion from perf_event_exit_task
2355          *     see the comment there.
2356          *
2357          *  2) there is a lock-inversion with mmap_sem through
2358          *     perf_event_read_group(), which takes faults while
2359          *     holding ctx->mutex, however this is called after
2360          *     the last filedesc died, so there is no possibility
2361          *     to trigger the AB-BA case.
2362          */
2363         mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2364         raw_spin_lock_irq(&ctx->lock);
2365         perf_group_detach(event);
2366         list_del_event(event, ctx);
2367         raw_spin_unlock_irq(&ctx->lock);
2368         mutex_unlock(&ctx->mutex);
2369
2370         free_event(event);
2371
2372         return 0;
2373 }
2374 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
2375
2376 /*
2377  * Called when the last reference to the file is gone.
2378  */
2379 static int perf_release(struct inode *inode, struct file *file)
2380 {
2381         struct perf_event *event = file->private_data;
2382         struct task_struct *owner;
2383
2384         file->private_data = NULL;
2385
2386         rcu_read_lock();
2387         owner = ACCESS_ONCE(event->owner);
2388         /*
2389          * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2390          * !owner it means the list deletion is complete and we can indeed
2391          * free this event, otherwise we need to serialize on
2392          * owner->perf_event_mutex.
2393          */
2394         smp_read_barrier_depends();
2395         if (owner) {
2396                 /*
2397                  * Since delayed_put_task_struct() also drops the last
2398                  * task reference we can safely take a new reference
2399                  * while holding the rcu_read_lock().
2400                  */
2401                 get_task_struct(owner);
2402         }
2403         rcu_read_unlock();
2404
2405         if (owner) {
2406                 mutex_lock(&owner->perf_event_mutex);
2407                 /*
2408                  * We have to re-check the event->owner field, if it is cleared
2409                  * we raced with perf_event_exit_task(), acquiring the mutex
2410                  * ensured they're done, and we can proceed with freeing the
2411                  * event.
2412                  */
2413                 if (event->owner)
2414                         list_del_init(&event->owner_entry);
2415                 mutex_unlock(&owner->perf_event_mutex);
2416                 put_task_struct(owner);
2417         }
2418
2419         return perf_event_release_kernel(event);
2420 }
2421
2422 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
2423 {
2424         struct perf_event *child;
2425         u64 total = 0;
2426
2427         *enabled = 0;
2428         *running = 0;
2429
2430         mutex_lock(&event->child_mutex);
2431         total += perf_event_read(event);
2432         *enabled += event->total_time_enabled +
2433                         atomic64_read(&event->child_total_time_enabled);
2434         *running += event->total_time_running +
2435                         atomic64_read(&event->child_total_time_running);
2436
2437         list_for_each_entry(child, &event->child_list, child_list) {
2438                 total += perf_event_read(child);
2439                 *enabled += child->total_time_enabled;
2440                 *running += child->total_time_running;
2441         }
2442         mutex_unlock(&event->child_mutex);
2443
2444         return total;
2445 }
2446 EXPORT_SYMBOL_GPL(perf_event_read_value);
2447
2448 static int perf_event_read_group(struct perf_event *event,
2449                                    u64 read_format, char __user *buf)
2450 {
2451         struct perf_event *leader = event->group_leader, *sub;
2452         int n = 0, size = 0, ret = -EFAULT;
2453         struct perf_event_context *ctx = leader->ctx;
2454         u64 values[5];
2455         u64 count, enabled, running;
2456
2457         mutex_lock(&ctx->mutex);
2458         count = perf_event_read_value(leader, &enabled, &running);
2459
2460         values[n++] = 1 + leader->nr_siblings;
2461         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2462                 values[n++] = enabled;
2463         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2464                 values[n++] = running;
2465         values[n++] = count;
2466         if (read_format & PERF_FORMAT_ID)
2467                 values[n++] = primary_event_id(leader);
2468
2469         size = n * sizeof(u64);
2470
2471         if (copy_to_user(buf, values, size))
2472                 goto unlock;
2473
2474         ret = size;
2475
2476         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2477                 n = 0;
2478
2479                 values[n++] = perf_event_read_value(sub, &enabled, &running);
2480                 if (read_format & PERF_FORMAT_ID)
2481                         values[n++] = primary_event_id(sub);
2482
2483                 size = n * sizeof(u64);
2484
2485                 if (copy_to_user(buf + ret, values, size)) {
2486                         ret = -EFAULT;
2487                         goto unlock;
2488                 }
2489
2490                 ret += size;
2491         }
2492 unlock:
2493         mutex_unlock(&ctx->mutex);
2494
2495         return ret;
2496 }
2497
2498 static int perf_event_read_one(struct perf_event *event,
2499                                  u64 read_format, char __user *buf)
2500 {
2501         u64 enabled, running;
2502         u64 values[4];
2503         int n = 0;
2504
2505         values[n++] = perf_event_read_value(event, &enabled, &running);
2506         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2507                 values[n++] = enabled;
2508         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2509                 values[n++] = running;
2510         if (read_format & PERF_FORMAT_ID)
2511                 values[n++] = primary_event_id(event);
2512
2513         if (copy_to_user(buf, values, n * sizeof(u64)))
2514                 return -EFAULT;
2515
2516         return n * sizeof(u64);
2517 }
2518
2519 /*
2520  * Read the performance event - simple non blocking version for now
2521  */
2522 static ssize_t
2523 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2524 {
2525         u64 read_format = event->attr.read_format;
2526         int ret;
2527
2528         /*
2529          * Return end-of-file for a read on a event that is in
2530          * error state (i.e. because it was pinned but it couldn't be
2531          * scheduled on to the CPU at some point).
2532          */
2533         if (event->state == PERF_EVENT_STATE_ERROR)
2534                 return 0;
2535
2536         if (count < event->read_size)
2537                 return -ENOSPC;
2538
2539         WARN_ON_ONCE(event->ctx->parent_ctx);
2540         if (read_format & PERF_FORMAT_GROUP)
2541                 ret = perf_event_read_group(event, read_format, buf);
2542         else
2543                 ret = perf_event_read_one(event, read_format, buf);
2544
2545         return ret;
2546 }
2547
2548 static ssize_t
2549 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2550 {
2551         struct perf_event *event = file->private_data;
2552
2553         return perf_read_hw(event, buf, count);
2554 }
2555
2556 static unsigned int perf_poll(struct file *file, poll_table *wait)
2557 {
2558         struct perf_event *event = file->private_data;
2559         struct perf_buffer *buffer;
2560         unsigned int events = POLL_HUP;
2561
2562         rcu_read_lock();
2563         buffer = rcu_dereference(event->buffer);
2564         if (buffer)
2565                 events = atomic_xchg(&buffer->poll, 0);
2566         rcu_read_unlock();
2567
2568         poll_wait(file, &event->waitq, wait);
2569
2570         return events;
2571 }
2572
2573 static void perf_event_reset(struct perf_event *event)
2574 {
2575         (void)perf_event_read(event);
2576         local64_set(&event->count, 0);
2577         perf_event_update_userpage(event);
2578 }
2579
2580 /*
2581  * Holding the top-level event's child_mutex means that any
2582  * descendant process that has inherited this event will block
2583  * in sync_child_event if it goes to exit, thus satisfying the
2584  * task existence requirements of perf_event_enable/disable.
2585  */
2586 static void perf_event_for_each_child(struct perf_event *event,
2587                                         void (*func)(struct perf_event *))
2588 {
2589         struct perf_event *child;
2590
2591         WARN_ON_ONCE(event->ctx->parent_ctx);
2592         mutex_lock(&event->child_mutex);
2593         func(event);
2594         list_for_each_entry(child, &event->child_list, child_list)
2595                 func(child);
2596         mutex_unlock(&event->child_mutex);
2597 }
2598
2599 static void perf_event_for_each(struct perf_event *event,
2600                                   void (*func)(struct perf_event *))
2601 {
2602         struct perf_event_context *ctx = event->ctx;
2603         struct perf_event *sibling;
2604
2605         WARN_ON_ONCE(ctx->parent_ctx);
2606         mutex_lock(&ctx->mutex);
2607         event = event->group_leader;
2608
2609         perf_event_for_each_child(event, func);
2610         func(event);
2611         list_for_each_entry(sibling, &event->sibling_list, group_entry)
2612                 perf_event_for_each_child(event, func);
2613         mutex_unlock(&ctx->mutex);
2614 }
2615
2616 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2617 {
2618         struct perf_event_context *ctx = event->ctx;
2619         int ret = 0;
2620         u64 value;
2621
2622         if (!is_sampling_event(event))
2623                 return -EINVAL;
2624
2625         if (copy_from_user(&value, arg, sizeof(value)))
2626                 return -EFAULT;
2627
2628         if (!value)
2629                 return -EINVAL;
2630
2631         raw_spin_lock_irq(&ctx->lock);
2632         if (event->attr.freq) {
2633                 if (value > sysctl_perf_event_sample_rate) {
2634                         ret = -EINVAL;
2635                         goto unlock;
2636                 }
2637
2638                 event->attr.sample_freq = value;
2639         } else {
2640                 event->attr.sample_period = value;
2641                 event->hw.sample_period = value;
2642         }
2643 unlock:
2644         raw_spin_unlock_irq(&ctx->lock);
2645
2646         return ret;
2647 }
2648
2649 static const struct file_operations perf_fops;
2650
2651 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
2652 {
2653         struct file *file;
2654
2655         file = fget_light(fd, fput_needed);
2656         if (!file)
2657                 return ERR_PTR(-EBADF);
2658
2659         if (file->f_op != &perf_fops) {
2660                 fput_light(file, *fput_needed);
2661                 *fput_needed = 0;
2662                 return ERR_PTR(-EBADF);
2663         }
2664
2665         return file->private_data;
2666 }
2667
2668 static int perf_event_set_output(struct perf_event *event,
2669                                  struct perf_event *output_event);
2670 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2671
2672 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2673 {
2674         struct perf_event *event = file->private_data;
2675         void (*func)(struct perf_event *);
2676         u32 flags = arg;
2677
2678         switch (cmd) {
2679         case PERF_EVENT_IOC_ENABLE:
2680                 func = perf_event_enable;
2681                 break;
2682         case PERF_EVENT_IOC_DISABLE:
2683                 func = perf_event_disable;
2684                 break;
2685         case PERF_EVENT_IOC_RESET:
2686                 func = perf_event_reset;
2687                 break;
2688
2689         case PERF_EVENT_IOC_REFRESH:
2690                 return perf_event_refresh(event, arg);
2691
2692         case PERF_EVENT_IOC_PERIOD:
2693                 return perf_event_period(event, (u64 __user *)arg);
2694
2695         case PERF_EVENT_IOC_SET_OUTPUT:
2696         {
2697                 struct perf_event *output_event = NULL;
2698                 int fput_needed = 0;
2699                 int ret;
2700
2701                 if (arg != -1) {
2702                         output_event = perf_fget_light(arg, &fput_needed);
2703                         if (IS_ERR(output_event))
2704                                 return PTR_ERR(output_event);
2705                 }
2706
2707                 ret = perf_event_set_output(event, output_event);
2708                 if (output_event)
2709                         fput_light(output_event->filp, fput_needed);
2710
2711                 return ret;
2712         }
2713
2714         case PERF_EVENT_IOC_SET_FILTER:
2715                 return perf_event_set_filter(event, (void __user *)arg);
2716
2717         default:
2718                 return -ENOTTY;
2719         }
2720
2721         if (flags & PERF_IOC_FLAG_GROUP)
2722                 perf_event_for_each(event, func);
2723         else
2724                 perf_event_for_each_child(event, func);
2725
2726         return 0;
2727 }
2728
2729 int perf_event_task_enable(void)
2730 {
2731         struct perf_event *event;
2732
2733         mutex_lock(&current->perf_event_mutex);
2734         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2735                 perf_event_for_each_child(event, perf_event_enable);
2736         mutex_unlock(&current->perf_event_mutex);
2737
2738         return 0;
2739 }
2740
2741 int perf_event_task_disable(void)
2742 {
2743         struct perf_event *event;
2744
2745         mutex_lock(&current->perf_event_mutex);
2746         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2747                 perf_event_for_each_child(event, perf_event_disable);
2748         mutex_unlock(&current->perf_event_mutex);
2749
2750         return 0;
2751 }
2752
2753 #ifndef PERF_EVENT_INDEX_OFFSET
2754 # define PERF_EVENT_INDEX_OFFSET 0
2755 #endif
2756
2757 static int perf_event_index(struct perf_event *event)
2758 {
2759         if (event->hw.state & PERF_HES_STOPPED)
2760                 return 0;
2761
2762         if (event->state != PERF_EVENT_STATE_ACTIVE)
2763                 return 0;
2764
2765         return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2766 }
2767
2768 /*
2769  * Callers need to ensure there can be no nesting of this function, otherwise
2770  * the seqlock logic goes bad. We can not serialize this because the arch
2771  * code calls this from NMI context.
2772  */
2773 void perf_event_update_userpage(struct perf_event *event)
2774 {
2775         struct perf_event_mmap_page *userpg;
2776         struct perf_buffer *buffer;
2777
2778         rcu_read_lock();
2779         buffer = rcu_dereference(event->buffer);
2780         if (!buffer)
2781                 goto unlock;
2782
2783         userpg = buffer->user_page;
2784
2785         /*
2786          * Disable preemption so as to not let the corresponding user-space
2787          * spin too long if we get preempted.
2788          */
2789         preempt_disable();
2790         ++userpg->lock;
2791         barrier();
2792         userpg->index = perf_event_index(event);
2793         userpg->offset = perf_event_count(event);
2794         if (event->state == PERF_EVENT_STATE_ACTIVE)
2795                 userpg->offset -= local64_read(&event->hw.prev_count);
2796
2797         userpg->time_enabled = event->total_time_enabled +
2798                         atomic64_read(&event->child_total_time_enabled);
2799
2800         userpg->time_running = event->total_time_running +
2801                         atomic64_read(&event->child_total_time_running);
2802
2803         barrier();
2804         ++userpg->lock;
2805         preempt_enable();
2806 unlock:
2807         rcu_read_unlock();
2808 }
2809
2810 static unsigned long perf_data_size(struct perf_buffer *buffer);
2811
2812 static void
2813 perf_buffer_init(struct perf_buffer *buffer, long watermark, int flags)
2814 {
2815         long max_size = perf_data_size(buffer);
2816
2817         if (watermark)
2818                 buffer->watermark = min(max_size, watermark);
2819
2820         if (!buffer->watermark)
2821                 buffer->watermark = max_size / 2;
2822
2823         if (flags & PERF_BUFFER_WRITABLE)
2824                 buffer->writable = 1;
2825
2826         atomic_set(&buffer->refcount, 1);
2827 }
2828
2829 #ifndef CONFIG_PERF_USE_VMALLOC
2830
2831 /*
2832  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2833  */
2834
2835 static struct page *
2836 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2837 {
2838         if (pgoff > buffer->nr_pages)
2839                 return NULL;
2840
2841         if (pgoff == 0)
2842                 return virt_to_page(buffer->user_page);
2843
2844         return virt_to_page(buffer->data_pages[pgoff - 1]);
2845 }
2846
2847 static void *perf_mmap_alloc_page(int cpu)
2848 {
2849         struct page *page;
2850         int node;
2851
2852         node = (cpu == -1) ? cpu : cpu_to_node(cpu);
2853         page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
2854         if (!page)
2855                 return NULL;
2856
2857         return page_address(page);
2858 }
2859
2860 static struct perf_buffer *
2861 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2862 {
2863         struct perf_buffer *buffer;
2864         unsigned long size;
2865         int i;
2866
2867         size = sizeof(struct perf_buffer);
2868         size += nr_pages * sizeof(void *);
2869
2870         buffer = kzalloc(size, GFP_KERNEL);
2871         if (!buffer)
2872                 goto fail;
2873
2874         buffer->user_page = perf_mmap_alloc_page(cpu);
2875         if (!buffer->user_page)
2876                 goto fail_user_page;
2877
2878         for (i = 0; i < nr_pages; i++) {
2879                 buffer->data_pages[i] = perf_mmap_alloc_page(cpu);
2880                 if (!buffer->data_pages[i])
2881                         goto fail_data_pages;
2882         }
2883
2884         buffer->nr_pages = nr_pages;
2885
2886         perf_buffer_init(buffer, watermark, flags);
2887
2888         return buffer;
2889
2890 fail_data_pages:
2891         for (i--; i >= 0; i--)
2892                 free_page((unsigned long)buffer->data_pages[i]);
2893
2894         free_page((unsigned long)buffer->user_page);
2895
2896 fail_user_page:
2897         kfree(buffer);
2898
2899 fail:
2900         return NULL;
2901 }
2902
2903 static void perf_mmap_free_page(unsigned long addr)
2904 {
2905         struct page *page = virt_to_page((void *)addr);
2906
2907         page->mapping = NULL;
2908         __free_page(page);
2909 }
2910
2911 static void perf_buffer_free(struct perf_buffer *buffer)
2912 {
2913         int i;
2914
2915         perf_mmap_free_page((unsigned long)buffer->user_page);
2916         for (i = 0; i < buffer->nr_pages; i++)
2917                 perf_mmap_free_page((unsigned long)buffer->data_pages[i]);
2918         kfree(buffer);
2919 }
2920
2921 static inline int page_order(struct perf_buffer *buffer)
2922 {
2923         return 0;
2924 }
2925
2926 #else
2927
2928 /*
2929  * Back perf_mmap() with vmalloc memory.
2930  *
2931  * Required for architectures that have d-cache aliasing issues.
2932  */
2933
2934 static inline int page_order(struct perf_buffer *buffer)
2935 {
2936         return buffer->page_order;
2937 }
2938
2939 static struct page *
2940 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2941 {
2942         if (pgoff > (1UL << page_order(buffer)))
2943                 return NULL;
2944
2945         return vmalloc_to_page((void *)buffer->user_page + pgoff * PAGE_SIZE);
2946 }
2947
2948 static void perf_mmap_unmark_page(void *addr)
2949 {
2950         struct page *page = vmalloc_to_page(addr);
2951
2952         page->mapping = NULL;
2953 }
2954
2955 static void perf_buffer_free_work(struct work_struct *work)
2956 {
2957         struct perf_buffer *buffer;
2958         void *base;
2959         int i, nr;
2960
2961         buffer = container_of(work, struct perf_buffer, work);
2962         nr = 1 << page_order(buffer);
2963
2964         base = buffer->user_page;
2965         for (i = 0; i < nr + 1; i++)
2966                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2967
2968         vfree(base);
2969         kfree(buffer);
2970 }
2971
2972 static void perf_buffer_free(struct perf_buffer *buffer)
2973 {
2974         schedule_work(&buffer->work);
2975 }
2976
2977 static struct perf_buffer *
2978 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2979 {
2980         struct perf_buffer *buffer;
2981         unsigned long size;
2982         void *all_buf;
2983
2984         size = sizeof(struct perf_buffer);
2985         size += sizeof(void *);
2986
2987         buffer = kzalloc(size, GFP_KERNEL);
2988         if (!buffer)
2989                 goto fail;
2990
2991         INIT_WORK(&buffer->work, perf_buffer_free_work);
2992
2993         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2994         if (!all_buf)
2995                 goto fail_all_buf;
2996
2997         buffer->user_page = all_buf;
2998         buffer->data_pages[0] = all_buf + PAGE_SIZE;
2999         buffer->page_order = ilog2(nr_pages);
3000         buffer->nr_pages = 1;
3001
3002         perf_buffer_init(buffer, watermark, flags);
3003
3004         return buffer;
3005
3006 fail_all_buf:
3007         kfree(buffer);
3008
3009 fail:
3010         return NULL;
3011 }
3012
3013 #endif
3014
3015 static unsigned long perf_data_size(struct perf_buffer *buffer)
3016 {
3017         return buffer->nr_pages << (PAGE_SHIFT + page_order(buffer));
3018 }
3019
3020 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3021 {
3022         struct perf_event *event = vma->vm_file->private_data;
3023         struct perf_buffer *buffer;
3024         int ret = VM_FAULT_SIGBUS;
3025
3026         if (vmf->flags & FAULT_FLAG_MKWRITE) {
3027                 if (vmf->pgoff == 0)
3028                         ret = 0;
3029                 return ret;
3030         }
3031
3032         rcu_read_lock();
3033         buffer = rcu_dereference(event->buffer);
3034         if (!buffer)
3035                 goto unlock;
3036
3037         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3038                 goto unlock;
3039
3040         vmf->page = perf_mmap_to_page(buffer, vmf->pgoff);
3041         if (!vmf->page)
3042                 goto unlock;
3043
3044         get_page(vmf->page);
3045         vmf->page->mapping = vma->vm_file->f_mapping;
3046         vmf->page->index   = vmf->pgoff;
3047
3048         ret = 0;
3049 unlock:
3050         rcu_read_unlock();
3051
3052         return ret;
3053 }
3054
3055 static void perf_buffer_free_rcu(struct rcu_head *rcu_head)
3056 {
3057         struct perf_buffer *buffer;
3058
3059         buffer = container_of(rcu_head, struct perf_buffer, rcu_head);
3060         perf_buffer_free(buffer);
3061 }
3062
3063 static struct perf_buffer *perf_buffer_get(struct perf_event *event)
3064 {
3065         struct perf_buffer *buffer;
3066
3067         rcu_read_lock();
3068         buffer = rcu_dereference(event->buffer);
3069         if (buffer) {
3070                 if (!atomic_inc_not_zero(&buffer->refcount))
3071                         buffer = NULL;
3072         }
3073         rcu_read_unlock();
3074
3075         return buffer;
3076 }
3077
3078 static void perf_buffer_put(struct perf_buffer *buffer)
3079 {
3080         if (!atomic_dec_and_test(&buffer->refcount))
3081                 return;
3082
3083         call_rcu(&buffer->rcu_head, perf_buffer_free_rcu);
3084 }
3085
3086 static void perf_mmap_open(struct vm_area_struct *vma)
3087 {
3088         struct perf_event *event = vma->vm_file->private_data;
3089
3090         atomic_inc(&event->mmap_count);
3091 }
3092
3093 static void perf_mmap_close(struct vm_area_struct *vma)
3094 {
3095         struct perf_event *event = vma->vm_file->private_data;
3096
3097         if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
3098                 unsigned long size = perf_data_size(event->buffer);
3099                 struct user_struct *user = event->mmap_user;
3100                 struct perf_buffer *buffer = event->buffer;
3101
3102                 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
3103                 vma->vm_mm->locked_vm -= event->mmap_locked;
3104                 rcu_assign_pointer(event->buffer, NULL);
3105                 mutex_unlock(&event->mmap_mutex);
3106
3107                 perf_buffer_put(buffer);
3108                 free_uid(user);
3109         }
3110 }
3111
3112 static const struct vm_operations_struct perf_mmap_vmops = {
3113         .open           = perf_mmap_open,
3114         .close          = perf_mmap_close,
3115         .fault          = perf_mmap_fault,
3116         .page_mkwrite   = perf_mmap_fault,
3117 };
3118
3119 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3120 {
3121         struct perf_event *event = file->private_data;
3122         unsigned long user_locked, user_lock_limit;
3123         struct user_struct *user = current_user();
3124         unsigned long locked, lock_limit;
3125         struct perf_buffer *buffer;
3126         unsigned long vma_size;
3127         unsigned long nr_pages;
3128         long user_extra, extra;
3129         int ret = 0, flags = 0;
3130
3131         /*
3132          * Don't allow mmap() of inherited per-task counters. This would
3133          * create a performance issue due to all children writing to the
3134          * same buffer.
3135          */
3136         if (event->cpu == -1 && event->attr.inherit)
3137                 return -EINVAL;
3138
3139         if (!(vma->vm_flags & VM_SHARED))
3140                 return -EINVAL;
3141
3142         vma_size = vma->vm_end - vma->vm_start;
3143         nr_pages = (vma_size / PAGE_SIZE) - 1;
3144
3145         /*
3146          * If we have buffer pages ensure they're a power-of-two number, so we
3147          * can do bitmasks instead of modulo.
3148          */
3149         if (nr_pages != 0 && !is_power_of_2(nr_pages))
3150                 return -EINVAL;
3151
3152         if (vma_size != PAGE_SIZE * (1 + nr_pages))
3153                 return -EINVAL;
3154
3155         if (vma->vm_pgoff != 0)
3156                 return -EINVAL;
3157
3158         WARN_ON_ONCE(event->ctx->parent_ctx);
3159         mutex_lock(&event->mmap_mutex);
3160         if (event->buffer) {
3161                 if (event->buffer->nr_pages == nr_pages)
3162                         atomic_inc(&event->buffer->refcount);
3163                 else
3164                         ret = -EINVAL;
3165                 goto unlock;
3166         }
3167
3168         user_extra = nr_pages + 1;
3169         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3170
3171         /*
3172          * Increase the limit linearly with more CPUs:
3173          */
3174         user_lock_limit *= num_online_cpus();
3175
3176         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3177
3178         extra = 0;
3179         if (user_locked > user_lock_limit)
3180                 extra = user_locked - user_lock_limit;
3181
3182         lock_limit = rlimit(RLIMIT_MEMLOCK);
3183         lock_limit >>= PAGE_SHIFT;
3184         locked = vma->vm_mm->locked_vm + extra;
3185
3186         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3187                 !capable(CAP_IPC_LOCK)) {
3188                 ret = -EPERM;
3189                 goto unlock;
3190         }
3191
3192         WARN_ON(event->buffer);
3193
3194         if (vma->vm_flags & VM_WRITE)
3195                 flags |= PERF_BUFFER_WRITABLE;
3196
3197         buffer = perf_buffer_alloc(nr_pages, event->attr.wakeup_watermark,
3198                                    event->cpu, flags);
3199         if (!buffer) {
3200                 ret = -ENOMEM;
3201                 goto unlock;
3202         }
3203         rcu_assign_pointer(event->buffer, buffer);
3204
3205         atomic_long_add(user_extra, &user->locked_vm);
3206         event->mmap_locked = extra;
3207         event->mmap_user = get_current_user();
3208         vma->vm_mm->locked_vm += event->mmap_locked;
3209
3210 unlock:
3211         if (!ret)
3212                 atomic_inc(&event->mmap_count);
3213         mutex_unlock(&event->mmap_mutex);
3214
3215         vma->vm_flags |= VM_RESERVED;
3216         vma->vm_ops = &perf_mmap_vmops;
3217
3218         return ret;
3219 }
3220
3221 static int perf_fasync(int fd, struct file *filp, int on)
3222 {
3223         struct inode *inode = filp->f_path.dentry->d_inode;
3224         struct perf_event *event = filp->private_data;
3225         int retval;
3226
3227         mutex_lock(&inode->i_mutex);
3228         retval = fasync_helper(fd, filp, on, &event->fasync);
3229         mutex_unlock(&inode->i_mutex);
3230
3231         if (retval < 0)
3232                 return retval;
3233
3234         return 0;
3235 }
3236
3237 static const struct file_operations perf_fops = {
3238         .llseek                 = no_llseek,
3239         .release                = perf_release,
3240         .read                   = perf_read,
3241         .poll                   = perf_poll,
3242         .unlocked_ioctl         = perf_ioctl,
3243         .compat_ioctl           = perf_ioctl,
3244         .mmap                   = perf_mmap,
3245         .fasync                 = perf_fasync,
3246 };
3247
3248 /*
3249  * Perf event wakeup
3250  *
3251  * If there's data, ensure we set the poll() state and publish everything
3252  * to user-space before waking everybody up.
3253  */
3254
3255 void perf_event_wakeup(struct perf_event *event)
3256 {
3257         wake_up_all(&event->waitq);
3258
3259         if (event->pending_kill) {
3260                 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3261                 event->pending_kill = 0;
3262         }
3263 }
3264
3265 static void perf_pending_event(struct irq_work *entry)
3266 {
3267         struct perf_event *event = container_of(entry,
3268                         struct perf_event, pending);
3269
3270         if (event->pending_disable) {
3271                 event->pending_disable = 0;
3272                 __perf_event_disable(event);
3273         }
3274
3275         if (event->pending_wakeup) {
3276                 event->pending_wakeup = 0;
3277                 perf_event_wakeup(event);
3278         }
3279 }
3280
3281 /*
3282  * We assume there is only KVM supporting the callbacks.
3283  * Later on, we might change it to a list if there is
3284  * another virtualization implementation supporting the callbacks.
3285  */
3286 struct perf_guest_info_callbacks *perf_guest_cbs;
3287
3288 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3289 {
3290         perf_guest_cbs = cbs;
3291         return 0;
3292 }
3293 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3294
3295 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3296 {
3297         perf_guest_cbs = NULL;
3298         return 0;
3299 }
3300 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3301
3302 /*
3303  * Output
3304  */
3305 static bool perf_output_space(struct perf_buffer *buffer, unsigned long tail,
3306                               unsigned long offset, unsigned long head)
3307 {
3308         unsigned long mask;
3309
3310         if (!buffer->writable)
3311                 return true;
3312
3313         mask = perf_data_size(buffer) - 1;
3314
3315         offset = (offset - tail) & mask;
3316         head   = (head   - tail) & mask;
3317
3318         if ((int)(head - offset) < 0)
3319                 return false;
3320
3321         return true;
3322 }
3323
3324 static void perf_output_wakeup(struct perf_output_handle *handle)
3325 {
3326         atomic_set(&handle->buffer->poll, POLL_IN);
3327
3328         if (handle->nmi) {
3329                 handle->event->pending_wakeup = 1;
3330                 irq_work_queue(&handle->event->pending);
3331         } else
3332                 perf_event_wakeup(handle->event);
3333 }
3334
3335 /*
3336  * We need to ensure a later event_id doesn't publish a head when a former
3337  * event isn't done writing. However since we need to deal with NMIs we
3338  * cannot fully serialize things.
3339  *
3340  * We only publish the head (and generate a wakeup) when the outer-most
3341  * event completes.
3342  */
3343 static void perf_output_get_handle(struct perf_output_handle *handle)
3344 {
3345         struct perf_buffer *buffer = handle->buffer;
3346
3347         preempt_disable();
3348         local_inc(&buffer->nest);
3349         handle->wakeup = local_read(&buffer->wakeup);
3350 }
3351
3352 static void perf_output_put_handle(struct perf_output_handle *handle)
3353 {
3354         struct perf_buffer *buffer = handle->buffer;
3355         unsigned long head;
3356
3357 again:
3358         head = local_read(&buffer->head);
3359
3360         /*
3361          * IRQ/NMI can happen here, which means we can miss a head update.
3362          */
3363
3364         if (!local_dec_and_test(&buffer->nest))
3365                 goto out;
3366
3367         /*
3368          * Publish the known good head. Rely on the full barrier implied
3369          * by atomic_dec_and_test() order the buffer->head read and this
3370          * write.
3371          */
3372         buffer->user_page->data_head = head;
3373
3374         /*
3375          * Now check if we missed an update, rely on the (compiler)
3376          * barrier in atomic_dec_and_test() to re-read buffer->head.
3377          */
3378         if (unlikely(head != local_read(&buffer->head))) {
3379                 local_inc(&buffer->nest);
3380                 goto again;
3381         }
3382
3383         if (handle->wakeup != local_read(&buffer->wakeup))
3384                 perf_output_wakeup(handle);
3385
3386 out:
3387         preempt_enable();
3388 }
3389
3390 __always_inline void perf_output_copy(struct perf_output_handle *handle,
3391                       const void *buf, unsigned int len)
3392 {
3393         do {
3394                 unsigned long size = min_t(unsigned long, handle->size, len);
3395
3396                 memcpy(handle->addr, buf, size);
3397
3398                 len -= size;
3399                 handle->addr += size;
3400                 buf += size;
3401                 handle->size -= size;
3402                 if (!handle->size) {
3403                         struct perf_buffer *buffer = handle->buffer;
3404
3405                         handle->page++;
3406                         handle->page &= buffer->nr_pages - 1;
3407                         handle->addr = buffer->data_pages[handle->page];
3408                         handle->size = PAGE_SIZE << page_order(buffer);
3409                 }
3410         } while (len);
3411 }
3412
3413 static void __perf_event_header__init_id(struct perf_event_header *header,
3414                                          struct perf_sample_data *data,
3415                                          struct perf_event *event)
3416 {
3417         u64 sample_type = event->attr.sample_type;
3418
3419         data->type = sample_type;
3420         header->size += event->id_header_size;
3421
3422         if (sample_type & PERF_SAMPLE_TID) {
3423                 /* namespace issues */
3424                 data->tid_entry.pid = perf_event_pid(event, current);
3425                 data->tid_entry.tid = perf_event_tid(event, current);
3426         }
3427
3428         if (sample_type & PERF_SAMPLE_TIME)
3429                 data->time = perf_clock();
3430
3431         if (sample_type & PERF_SAMPLE_ID)
3432                 data->id = primary_event_id(event);
3433
3434         if (sample_type & PERF_SAMPLE_STREAM_ID)
3435                 data->stream_id = event->id;
3436
3437         if (sample_type & PERF_SAMPLE_CPU) {
3438                 data->cpu_entry.cpu      = raw_smp_processor_id();
3439                 data->cpu_entry.reserved = 0;
3440         }
3441 }
3442
3443 static void perf_event_header__init_id(struct perf_event_header *header,
3444                                        struct perf_sample_data *data,
3445                                        struct perf_event *event)
3446 {
3447         if (event->attr.sample_id_all)
3448                 __perf_event_header__init_id(header, data, event);
3449 }
3450
3451 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
3452                                            struct perf_sample_data *data)
3453 {
3454         u64 sample_type = data->type;
3455
3456         if (sample_type & PERF_SAMPLE_TID)
3457                 perf_output_put(handle, data->tid_entry);
3458
3459         if (sample_type & PERF_SAMPLE_TIME)
3460                 perf_output_put(handle, data->time);
3461
3462         if (sample_type & PERF_SAMPLE_ID)
3463                 perf_output_put(handle, data->id);
3464
3465         if (sample_type & PERF_SAMPLE_STREAM_ID)
3466                 perf_output_put(handle, data->stream_id);
3467
3468         if (sample_type & PERF_SAMPLE_CPU)
3469                 perf_output_put(handle, data->cpu_entry);
3470 }
3471
3472 static void perf_event__output_id_sample(struct perf_event *event,
3473                                          struct perf_output_handle *handle,
3474                                          struct perf_sample_data *sample)
3475 {
3476         if (event->attr.sample_id_all)
3477                 __perf_event__output_id_sample(handle, sample);
3478 }
3479
3480 int perf_output_begin(struct perf_output_handle *handle,
3481                       struct perf_event *event, unsigned int size,
3482                       int nmi, int sample)
3483 {
3484         struct perf_buffer *buffer;
3485         unsigned long tail, offset, head;
3486         int have_lost;
3487         struct perf_sample_data sample_data;
3488         struct {
3489                 struct perf_event_header header;
3490                 u64                      id;
3491                 u64                      lost;
3492         } lost_event;
3493
3494         rcu_read_lock();
3495         /*
3496          * For inherited events we send all the output towards the parent.
3497          */
3498         if (event->parent)
3499                 event = event->parent;
3500
3501         buffer = rcu_dereference(event->buffer);
3502         if (!buffer)
3503                 goto out;
3504
3505         handle->buffer  = buffer;
3506         handle->event   = event;
3507         handle->nmi     = nmi;
3508         handle->sample  = sample;
3509
3510         if (!buffer->nr_pages)
3511                 goto out;
3512
3513         have_lost = local_read(&buffer->lost);
3514         if (have_lost) {
3515                 lost_event.header.size = sizeof(lost_event);
3516                 perf_event_header__init_id(&lost_event.header, &sample_data,
3517                                            event);
3518                 size += lost_event.header.size;
3519         }
3520
3521         perf_output_get_handle(handle);
3522
3523         do {
3524                 /*
3525                  * Userspace could choose to issue a mb() before updating the
3526                  * tail pointer. So that all reads will be completed before the
3527                  * write is issued.
3528                  */
3529                 tail = ACCESS_ONCE(buffer->user_page->data_tail);
3530                 smp_rmb();
3531                 offset = head = local_read(&buffer->head);
3532                 head += size;
3533                 if (unlikely(!perf_output_space(buffer, tail, offset, head)))
3534                         goto fail;
3535         } while (local_cmpxchg(&buffer->head, offset, head) != offset);
3536
3537         if (head - local_read(&buffer->wakeup) > buffer->watermark)
3538                 local_add(buffer->watermark, &buffer->wakeup);
3539
3540         handle->page = offset >> (PAGE_SHIFT + page_order(buffer));
3541         handle->page &= buffer->nr_pages - 1;
3542         handle->size = offset & ((PAGE_SIZE << page_order(buffer)) - 1);
3543         handle->addr = buffer->data_pages[handle->page];
3544         handle->addr += handle->size;
3545         handle->size = (PAGE_SIZE << page_order(buffer)) - handle->size;
3546
3547         if (have_lost) {
3548                 lost_event.header.type = PERF_RECORD_LOST;
3549                 lost_event.header.misc = 0;
3550                 lost_event.id          = event->id;
3551                 lost_event.lost        = local_xchg(&buffer->lost, 0);
3552
3553                 perf_output_put(handle, lost_event);
3554                 perf_event__output_id_sample(event, handle, &sample_data);
3555         }
3556
3557         return 0;
3558
3559 fail:
3560         local_inc(&buffer->lost);
3561         perf_output_put_handle(handle);
3562 out:
3563         rcu_read_unlock();
3564
3565         return -ENOSPC;
3566 }
3567
3568 void perf_output_end(struct perf_output_handle *handle)
3569 {
3570         struct perf_event *event = handle->event;
3571         struct perf_buffer *buffer = handle->buffer;
3572
3573         int wakeup_events = event->attr.wakeup_events;
3574
3575         if (handle->sample && wakeup_events) {
3576                 int events = local_inc_return(&buffer->events);
3577                 if (events >= wakeup_events) {
3578                         local_sub(wakeup_events, &buffer->events);
3579                         local_inc(&buffer->wakeup);
3580                 }
3581         }
3582
3583         perf_output_put_handle(handle);
3584         rcu_read_unlock();
3585 }
3586
3587 static void perf_output_read_one(struct perf_output_handle *handle,
3588                                  struct perf_event *event,
3589                                  u64 enabled, u64 running)
3590 {
3591         u64 read_format = event->attr.read_format;
3592         u64 values[4];
3593         int n = 0;
3594
3595         values[n++] = perf_event_count(event);
3596         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3597                 values[n++] = enabled +
3598                         atomic64_read(&event->child_total_time_enabled);
3599         }
3600         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3601                 values[n++] = running +
3602                         atomic64_read(&event->child_total_time_running);
3603         }
3604         if (read_format & PERF_FORMAT_ID)
3605                 values[n++] = primary_event_id(event);
3606
3607         perf_output_copy(handle, values, n * sizeof(u64));
3608 }
3609
3610 /*
3611  * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3612  */
3613 static void perf_output_read_group(struct perf_output_handle *handle,
3614                             struct perf_event *event,
3615                             u64 enabled, u64 running)
3616 {
3617         struct perf_event *leader = event->group_leader, *sub;
3618         u64 read_format = event->attr.read_format;
3619         u64 values[5];
3620         int n = 0;
3621
3622         values[n++] = 1 + leader->nr_siblings;
3623
3624         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3625                 values[n++] = enabled;
3626
3627         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3628                 values[n++] = running;
3629
3630         if (leader != event)
3631                 leader->pmu->read(leader);
3632
3633         values[n++] = perf_event_count(leader);
3634         if (read_format & PERF_FORMAT_ID)
3635                 values[n++] = primary_event_id(leader);
3636
3637         perf_output_copy(handle, values, n * sizeof(u64));
3638
3639         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3640                 n = 0;
3641
3642                 if (sub != event)
3643                         sub->pmu->read(sub);
3644
3645                 values[n++] = perf_event_count(sub);
3646                 if (read_format & PERF_FORMAT_ID)
3647                         values[n++] = primary_event_id(sub);
3648
3649                 perf_output_copy(handle, values, n * sizeof(u64));
3650         }
3651 }
3652
3653 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
3654                                  PERF_FORMAT_TOTAL_TIME_RUNNING)
3655
3656 static void perf_output_read(struct perf_output_handle *handle,
3657                              struct perf_event *event)
3658 {
3659         u64 enabled = 0, running = 0, now, ctx_time;
3660         u64 read_format = event->attr.read_format;
3661
3662         /*
3663          * compute total_time_enabled, total_time_running
3664          * based on snapshot values taken when the event
3665          * was last scheduled in.
3666          *
3667          * we cannot simply called update_context_time()
3668          * because of locking issue as we are called in
3669          * NMI context
3670          */
3671         if (read_format & PERF_FORMAT_TOTAL_TIMES) {
3672                 now = perf_clock();
3673                 ctx_time = event->shadow_ctx_time + now;
3674                 enabled = ctx_time - event->tstamp_enabled;
3675                 running = ctx_time - event->tstamp_running;
3676         }
3677
3678         if (event->attr.read_format & PERF_FORMAT_GROUP)
3679                 perf_output_read_group(handle, event, enabled, running);
3680         else
3681                 perf_output_read_one(handle, event, enabled, running);
3682 }
3683
3684 void perf_output_sample(struct perf_output_handle *handle,
3685                         struct perf_event_header *header,
3686                         struct perf_sample_data *data,
3687                         struct perf_event *event)
3688 {
3689         u64 sample_type = data->type;
3690
3691         perf_output_put(handle, *header);
3692
3693         if (sample_type & PERF_SAMPLE_IP)
3694                 perf_output_put(handle, data->ip);
3695
3696         if (sample_type & PERF_SAMPLE_TID)
3697                 perf_output_put(handle, data->tid_entry);
3698
3699         if (sample_type & PERF_SAMPLE_TIME)
3700                 perf_output_put(handle, data->time);
3701
3702         if (sample_type & PERF_SAMPLE_ADDR)
3703                 perf_output_put(handle, data->addr);
3704
3705         if (sample_type & PERF_SAMPLE_ID)
3706                 perf_output_put(handle, data->id);
3707
3708         if (sample_type & PERF_SAMPLE_STREAM_ID)
3709                 perf_output_put(handle, data->stream_id);
3710
3711         if (sample_type & PERF_SAMPLE_CPU)
3712                 perf_output_put(handle, data->cpu_entry);
3713
3714         if (sample_type & PERF_SAMPLE_PERIOD)
3715                 perf_output_put(handle, data->period);
3716
3717         if (sample_type & PERF_SAMPLE_READ)
3718                 perf_output_read(handle, event);
3719
3720         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3721                 if (data->callchain) {
3722                         int size = 1;
3723
3724                         if (data->callchain)
3725                                 size += data->callchain->nr;
3726
3727                         size *= sizeof(u64);
3728
3729                         perf_output_copy(handle, data->callchain, size);
3730                 } else {
3731                         u64 nr = 0;
3732                         perf_output_put(handle, nr);
3733                 }
3734         }
3735
3736         if (sample_type & PERF_SAMPLE_RAW) {
3737                 if (data->raw) {
3738                         perf_output_put(handle, data->raw->size);
3739                         perf_output_copy(handle, data->raw->data,
3740                                          data->raw->size);
3741                 } else {
3742                         struct {
3743                                 u32     size;
3744                                 u32     data;
3745                         } raw = {
3746                                 .size = sizeof(u32),
3747                                 .data = 0,
3748                         };
3749                         perf_output_put(handle, raw);
3750                 }
3751         }
3752 }
3753
3754 void perf_prepare_sample(struct perf_event_header *header,
3755                          struct perf_sample_data *data,
3756                          struct perf_event *event,
3757                          struct pt_regs *regs)
3758 {
3759         u64 sample_type = event->attr.sample_type;
3760
3761         header->type = PERF_RECORD_SAMPLE;
3762         header->size = sizeof(*header) + event->header_size;
3763
3764         header->misc = 0;
3765         header->misc |= perf_misc_flags(regs);
3766
3767         __perf_event_header__init_id(header, data, event);
3768
3769         if (sample_type & PERF_SAMPLE_IP)
3770                 data->ip = perf_instruction_pointer(regs);
3771
3772         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3773                 int size = 1;
3774
3775                 data->callchain = perf_callchain(regs);
3776
3777                 if (data->callchain)
3778                         size += data->callchain->nr;
3779
3780                 header->size += size * sizeof(u64);
3781         }
3782
3783         if (sample_type & PERF_SAMPLE_RAW) {
3784                 int size = sizeof(u32);
3785
3786                 if (data->raw)
3787                         size += data->raw->size;
3788                 else
3789                         size += sizeof(u32);
3790
3791                 WARN_ON_ONCE(size & (sizeof(u64)-1));
3792                 header->size += size;
3793         }
3794 }
3795
3796 static void perf_event_output(struct perf_event *event, int nmi,
3797                                 struct perf_sample_data *data,
3798                                 struct pt_regs *regs)
3799 {
3800         struct perf_output_handle handle;
3801         struct perf_event_header header;
3802
3803         /* protect the callchain buffers */
3804         rcu_read_lock();
3805
3806         perf_prepare_sample(&header, data, event, regs);
3807
3808         if (perf_output_begin(&handle, event, header.size, nmi, 1))
3809                 goto exit;
3810
3811         perf_output_sample(&handle, &header, data, event);
3812
3813         perf_output_end(&handle);
3814
3815 exit:
3816         rcu_read_unlock();
3817 }
3818
3819 /*
3820  * read event_id
3821  */
3822
3823 struct perf_read_event {
3824         struct perf_event_header        header;
3825
3826         u32                             pid;
3827         u32                             tid;
3828 };
3829
3830 static void
3831 perf_event_read_event(struct perf_event *event,
3832                         struct task_struct *task)
3833 {
3834         struct perf_output_handle handle;
3835         struct perf_sample_data sample;
3836         struct perf_read_event read_event = {
3837                 .header = {
3838                         .type = PERF_RECORD_READ,
3839                         .misc = 0,
3840                         .size = sizeof(read_event) + event->read_size,
3841                 },
3842                 .pid = perf_event_pid(event, task),
3843                 .tid = perf_event_tid(event, task),
3844         };
3845         int ret;
3846
3847         perf_event_header__init_id(&read_event.header, &sample, event);
3848         ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3849         if (ret)
3850                 return;
3851
3852         perf_output_put(&handle, read_event);
3853         perf_output_read(&handle, event);
3854         perf_event__output_id_sample(event, &handle, &sample);
3855
3856         perf_output_end(&handle);
3857 }
3858
3859 /*
3860  * task tracking -- fork/exit
3861  *
3862  * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3863  */
3864
3865 struct perf_task_event {
3866         struct task_struct              *task;
3867         struct perf_event_context       *task_ctx;
3868
3869         struct {
3870                 struct perf_event_header        header;
3871
3872                 u32                             pid;
3873                 u32                             ppid;
3874                 u32                             tid;
3875                 u32                             ptid;
3876                 u64                             time;
3877         } event_id;
3878 };
3879
3880 static void perf_event_task_output(struct perf_event *event,
3881                                      struct perf_task_event *task_event)
3882 {
3883         struct perf_output_handle handle;
3884         struct perf_sample_data sample;
3885         struct task_struct *task = task_event->task;
3886         int ret, size = task_event->event_id.header.size;
3887
3888         perf_event_header__init_id(&task_event->event_id.header, &sample, event);
3889
3890         ret = perf_output_begin(&handle, event,
3891                                 task_event->event_id.header.size, 0, 0);
3892         if (ret)
3893                 goto out;
3894
3895         task_event->event_id.pid = perf_event_pid(event, task);
3896         task_event->event_id.ppid = perf_event_pid(event, current);
3897
3898         task_event->event_id.tid = perf_event_tid(event, task);
3899         task_event->event_id.ptid = perf_event_tid(event, current);
3900
3901         perf_output_put(&handle, task_event->event_id);
3902
3903         perf_event__output_id_sample(event, &handle, &sample);
3904
3905         perf_output_end(&handle);
3906 out:
3907         task_event->event_id.header.size = size;
3908 }
3909
3910 static int perf_event_task_match(struct perf_event *event)
3911 {
3912         if (event->state < PERF_EVENT_STATE_INACTIVE)
3913                 return 0;
3914
3915         if (!event_filter_match(event))
3916                 return 0;
3917
3918         if (event->attr.comm || event->attr.mmap ||
3919             event->attr.mmap_data || event->attr.task)
3920                 return 1;
3921
3922         return 0;
3923 }
3924
3925 static void perf_event_task_ctx(struct perf_event_context *ctx,
3926                                   struct perf_task_event *task_event)
3927 {
3928         struct perf_event *event;
3929
3930         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3931                 if (perf_event_task_match(event))
3932                         perf_event_task_output(event, task_event);
3933         }
3934 }
3935
3936 static void perf_event_task_event(struct perf_task_event *task_event)
3937 {
3938         struct perf_cpu_context *cpuctx;
3939         struct perf_event_context *ctx;
3940         struct pmu *pmu;
3941         int ctxn;
3942
3943         rcu_read_lock();
3944         list_for_each_entry_rcu(pmu, &pmus, entry) {
3945                 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
3946                 if (cpuctx->active_pmu != pmu)
3947                         goto next;
3948                 perf_event_task_ctx(&cpuctx->ctx, task_event);
3949
3950                 ctx = task_event->task_ctx;
3951                 if (!ctx) {
3952                         ctxn = pmu->task_ctx_nr;
3953                         if (ctxn < 0)
3954                                 goto next;
3955                         ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
3956                 }
3957                 if (ctx)
3958                         perf_event_task_ctx(ctx, task_event);
3959 next:
3960                 put_cpu_ptr(pmu->pmu_cpu_context);
3961         }
3962         rcu_read_unlock();
3963 }
3964
3965 static void perf_event_task(struct task_struct *task,
3966                               struct perf_event_context *task_ctx,
3967                               int new)
3968 {
3969         struct perf_task_event task_event;
3970
3971         if (!atomic_read(&nr_comm_events) &&
3972             !atomic_read(&nr_mmap_events) &&
3973             !atomic_read(&nr_task_events))
3974                 return;
3975
3976         task_event = (struct perf_task_event){
3977                 .task     = task,
3978                 .task_ctx = task_ctx,
3979                 .event_id    = {
3980                         .header = {
3981                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3982                                 .misc = 0,
3983                                 .size = sizeof(task_event.event_id),
3984                         },
3985                         /* .pid  */
3986                         /* .ppid */
3987                         /* .tid  */
3988                         /* .ptid */
3989                         .time = perf_clock(),
3990                 },
3991         };
3992
3993         perf_event_task_event(&task_event);
3994 }
3995
3996 void perf_event_fork(struct task_struct *task)
3997 {
3998         perf_event_task(task, NULL, 1);
3999 }
4000
4001 /*
4002  * comm tracking
4003  */
4004
4005 struct perf_comm_event {
4006         struct task_struct      *task;
4007         char                    *comm;
4008         int                     comm_size;
4009
4010         struct {
4011                 struct perf_event_header        header;
4012
4013                 u32                             pid;
4014                 u32                             tid;
4015         } event_id;
4016 };
4017
4018 static void perf_event_comm_output(struct perf_event *event,
4019                                      struct perf_comm_event *comm_event)
4020 {
4021         struct perf_output_handle handle;
4022         struct perf_sample_data sample;
4023         int size = comm_event->event_id.header.size;
4024         int ret;
4025
4026         perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4027         ret = perf_output_begin(&handle, event,
4028                                 comm_event->event_id.header.size, 0, 0);
4029
4030         if (ret)
4031                 goto out;
4032
4033         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4034         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4035
4036         perf_output_put(&handle, comm_event->event_id);
4037         perf_output_copy(&handle, comm_event->comm,
4038                                    comm_event->comm_size);
4039
4040         perf_event__output_id_sample(event, &handle, &sample);
4041
4042         perf_output_end(&handle);
4043 out:
4044         comm_event->event_id.header.size = size;
4045 }
4046
4047 static int perf_event_comm_match(struct perf_event *event)
4048 {
4049         if (event->state < PERF_EVENT_STATE_INACTIVE)
4050                 return 0;
4051
4052         if (!event_filter_match(event))
4053                 return 0;
4054
4055         if (event->attr.comm)
4056                 return 1;
4057
4058         return 0;
4059 }
4060
4061 static void perf_event_comm_ctx(struct perf_event_context *ctx,
4062                                   struct perf_comm_event *comm_event)
4063 {
4064         struct perf_event *event;
4065
4066         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4067                 if (perf_event_comm_match(event))
4068                         perf_event_comm_output(event, comm_event);
4069         }
4070 }
4071
4072 static void perf_event_comm_event(struct perf_comm_event *comm_event)
4073 {
4074         struct perf_cpu_context *cpuctx;
4075         struct perf_event_context *ctx;
4076         char comm[TASK_COMM_LEN];
4077         unsigned int size;
4078         struct pmu *pmu;
4079         int ctxn;
4080
4081         memset(comm, 0, sizeof(comm));
4082         strlcpy(comm, comm_event->task->comm, sizeof(comm));
4083         size = ALIGN(strlen(comm)+1, sizeof(u64));
4084
4085         comm_event->comm = comm;
4086         comm_event->comm_size = size;
4087
4088         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
4089         rcu_read_lock();
4090         list_for_each_entry_rcu(pmu, &pmus, entry) {
4091                 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4092                 if (cpuctx->active_pmu != pmu)
4093                         goto next;
4094                 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
4095
4096                 ctxn = pmu->task_ctx_nr;
4097                 if (ctxn < 0)
4098                         goto next;
4099
4100                 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4101                 if (ctx)
4102                         perf_event_comm_ctx(ctx, comm_event);
4103 next:
4104                 put_cpu_ptr(pmu->pmu_cpu_context);
4105         }
4106         rcu_read_unlock();
4107 }
4108
4109 void perf_event_comm(struct task_struct *task)
4110 {
4111         struct perf_comm_event comm_event;
4112         struct perf_event_context *ctx;
4113         int ctxn;
4114
4115         for_each_task_context_nr(ctxn) {
4116                 ctx = task->perf_event_ctxp[ctxn];
4117                 if (!ctx)
4118                         continue;
4119
4120                 perf_event_enable_on_exec(ctx);
4121         }
4122
4123         if (!atomic_read(&nr_comm_events))
4124                 return;
4125
4126         comm_event = (struct perf_comm_event){
4127                 .task   = task,
4128                 /* .comm      */
4129                 /* .comm_size */
4130                 .event_id  = {
4131                         .header = {
4132                                 .type = PERF_RECORD_COMM,
4133                                 .misc = 0,
4134                                 /* .size */
4135                         },
4136                         /* .pid */
4137                         /* .tid */
4138                 },
4139         };
4140
4141         perf_event_comm_event(&comm_event);
4142 }
4143
4144 /*
4145  * mmap tracking
4146  */
4147
4148 struct perf_mmap_event {
4149         struct vm_area_struct   *vma;
4150
4151         const char              *file_name;
4152         int                     file_size;
4153
4154         struct {
4155                 struct perf_event_header        header;
4156
4157                 u32                             pid;
4158                 u32                             tid;
4159                 u64                             start;
4160                 u64                             len;
4161                 u64                             pgoff;
4162         } event_id;
4163 };
4164
4165 static void perf_event_mmap_output(struct perf_event *event,
4166                                      struct perf_mmap_event *mmap_event)
4167 {
4168         struct perf_output_handle handle;
4169         struct perf_sample_data sample;
4170         int size = mmap_event->event_id.header.size;
4171         int ret;
4172
4173         perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
4174         ret = perf_output_begin(&handle, event,
4175                                 mmap_event->event_id.header.size, 0, 0);
4176         if (ret)
4177                 goto out;
4178
4179         mmap_event->event_id.pid = perf_event_pid(event, current);
4180         mmap_event->event_id.tid = perf_event_tid(event, current);
4181
4182         perf_output_put(&handle, mmap_event->event_id);
4183         perf_output_copy(&handle, mmap_event->file_name,
4184                                    mmap_event->file_size);
4185
4186         perf_event__output_id_sample(event, &handle, &sample);
4187
4188         perf_output_end(&handle);
4189 out:
4190         mmap_event->event_id.header.size = size;
4191 }
4192
4193 static int perf_event_mmap_match(struct perf_event *event,
4194                                    struct perf_mmap_event *mmap_event,
4195                                    int executable)
4196 {
4197         if (event->state < PERF_EVENT_STATE_INACTIVE)
4198                 return 0;
4199
4200         if (!event_filter_match(event))
4201                 return 0;
4202
4203         if ((!executable && event->attr.mmap_data) ||
4204             (executable && event->attr.mmap))
4205                 return 1;
4206
4207         return 0;
4208 }
4209
4210 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
4211                                   struct perf_mmap_event *mmap_event,
4212                                   int executable)
4213 {
4214         struct perf_event *event;
4215
4216         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4217                 if (perf_event_mmap_match(event, mmap_event, executable))
4218                         perf_event_mmap_output(event, mmap_event);
4219         }
4220 }
4221
4222 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4223 {
4224         struct perf_cpu_context *cpuctx;
4225         struct perf_event_context *ctx;
4226         struct vm_area_struct *vma = mmap_event->vma;
4227         struct file *file = vma->vm_file;
4228         unsigned int size;
4229         char tmp[16];
4230         char *buf = NULL;
4231         const char *name;
4232         struct pmu *pmu;
4233         int ctxn;
4234
4235         memset(tmp, 0, sizeof(tmp));
4236
4237         if (file) {
4238                 /*
4239                  * d_path works from the end of the buffer backwards, so we
4240                  * need to add enough zero bytes after the string to handle
4241                  * the 64bit alignment we do later.
4242                  */
4243                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4244                 if (!buf) {
4245                         name = strncpy(tmp, "//enomem", sizeof(tmp));
4246                         goto got_name;
4247                 }
4248                 name = d_path(&file->f_path, buf, PATH_MAX);
4249                 if (IS_ERR(name)) {
4250                         name = strncpy(tmp, "//toolong", sizeof(tmp));
4251                         goto got_name;
4252                 }
4253         } else {
4254                 if (arch_vma_name(mmap_event->vma)) {
4255                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4256                                        sizeof(tmp));
4257                         goto got_name;
4258                 }
4259
4260                 if (!vma->vm_mm) {
4261                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
4262                         goto got_name;
4263                 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4264                                 vma->vm_end >= vma->vm_mm->brk) {
4265                         name = strncpy(tmp, "[heap]", sizeof(tmp));
4266                         goto got_name;
4267                 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4268                                 vma->vm_end >= vma->vm_mm->start_stack) {
4269                         name = strncpy(tmp, "[stack]", sizeof(tmp));
4270                         goto got_name;
4271                 }
4272
4273                 name = strncpy(tmp, "//anon", sizeof(tmp));
4274                 goto got_name;
4275         }
4276
4277 got_name:
4278         size = ALIGN(strlen(name)+1, sizeof(u64));
4279
4280         mmap_event->file_name = name;
4281         mmap_event->file_size = size;
4282
4283         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4284
4285         rcu_read_lock();
4286         list_for_each_entry_rcu(pmu, &pmus, entry) {
4287                 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4288                 if (cpuctx->active_pmu != pmu)
4289                         goto next;
4290                 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
4291                                         vma->vm_flags & VM_EXEC);
4292
4293                 ctxn = pmu->task_ctx_nr;
4294                 if (ctxn < 0)
4295                         goto next;
4296
4297                 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4298                 if (ctx) {
4299                         perf_event_mmap_ctx(ctx, mmap_event,
4300                                         vma->vm_flags & VM_EXEC);
4301                 }
4302 next:
4303                 put_cpu_ptr(pmu->pmu_cpu_context);
4304         }
4305         rcu_read_unlock();
4306
4307         kfree(buf);
4308 }
4309
4310 void perf_event_mmap(struct vm_area_struct *vma)
4311 {
4312         struct perf_mmap_event mmap_event;
4313
4314         if (!atomic_read(&nr_mmap_events))
4315                 return;
4316
4317         mmap_event = (struct perf_mmap_event){
4318                 .vma    = vma,
4319                 /* .file_name */
4320                 /* .file_size */
4321                 .event_id  = {
4322                         .header = {
4323                                 .type = PERF_RECORD_MMAP,
4324                                 .misc = PERF_RECORD_MISC_USER,
4325                                 /* .size */
4326                         },
4327                         /* .pid */
4328                         /* .tid */
4329                         .start  = vma->vm_start,
4330                         .len    = vma->vm_end - vma->vm_start,
4331                         .pgoff  = (u64)vma->vm_pgoff << PAGE_SHIFT,
4332                 },
4333         };
4334
4335         perf_event_mmap_event(&mmap_event);
4336 }
4337
4338 /*
4339  * IRQ throttle logging
4340  */
4341
4342 static void perf_log_throttle(struct perf_event *event, int enable)
4343 {
4344         struct perf_output_handle handle;
4345         struct perf_sample_data sample;
4346         int ret;
4347
4348         struct {
4349                 struct perf_event_header        header;
4350                 u64                             time;
4351                 u64                             id;
4352                 u64                             stream_id;
4353         } throttle_event = {
4354                 .header = {
4355                         .type = PERF_RECORD_THROTTLE,
4356                         .misc = 0,
4357                         .size = sizeof(throttle_event),
4358                 },
4359                 .time           = perf_clock(),
4360                 .id             = primary_event_id(event),
4361                 .stream_id      = event->id,
4362         };
4363
4364         if (enable)
4365                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4366
4367         perf_event_header__init_id(&throttle_event.header, &sample, event);
4368
4369         ret = perf_output_begin(&handle, event,
4370                                 throttle_event.header.size, 1, 0);
4371         if (ret)
4372                 return;
4373
4374         perf_output_put(&handle, throttle_event);
4375         perf_event__output_id_sample(event, &handle, &sample);
4376         perf_output_end(&handle);
4377 }
4378
4379 /*
4380  * Generic event overflow handling, sampling.
4381  */
4382
4383 static int __perf_event_overflow(struct perf_event *event, int nmi,
4384                                    int throttle, struct perf_sample_data *data,
4385                                    struct pt_regs *regs)
4386 {
4387         int events = atomic_read(&event->event_limit);
4388         struct hw_perf_event *hwc = &event->hw;
4389         int ret = 0;
4390
4391         /*
4392          * Non-sampling counters might still use the PMI to fold short
4393          * hardware counters, ignore those.
4394          */
4395         if (unlikely(!is_sampling_event(event)))
4396                 return 0;
4397
4398         if (!throttle) {
4399                 hwc->interrupts++;
4400         } else {
4401                 if (hwc->interrupts != MAX_INTERRUPTS) {
4402                         hwc->interrupts++;
4403                         if (HZ * hwc->interrupts >
4404                                         (u64)sysctl_perf_event_sample_rate) {
4405                                 hwc->interrupts = MAX_INTERRUPTS;
4406                                 perf_log_throttle(event, 0);
4407                                 ret = 1;
4408                         }
4409                 } else {
4410                         /*
4411                          * Keep re-disabling events even though on the previous
4412                          * pass we disabled it - just in case we raced with a
4413                          * sched-in and the event got enabled again:
4414                          */
4415                         ret = 1;
4416                 }
4417         }
4418
4419         if (event->attr.freq) {
4420                 u64 now = perf_clock();
4421                 s64 delta = now - hwc->freq_time_stamp;
4422
4423                 hwc->freq_time_stamp = now;
4424
4425                 if (delta > 0 && delta < 2*TICK_NSEC)
4426                         perf_adjust_period(event, delta, hwc->last_period);
4427         }
4428
4429         /*
4430          * XXX event_limit might not quite work as expected on inherited
4431          * events
4432          */
4433
4434         event->pending_kill = POLL_IN;
4435         if (events && atomic_dec_and_test(&event->event_limit)) {
4436                 ret = 1;
4437                 event->pending_kill = POLL_HUP;
4438                 if (nmi) {
4439                         event->pending_disable = 1;
4440                         irq_work_queue(&event->pending);
4441                 } else
4442                         perf_event_disable(event);
4443         }
4444
4445         if (event->overflow_handler)
4446                 event->overflow_handler(event, nmi, data, regs);
4447         else
4448                 perf_event_output(event, nmi, data, regs);
4449
4450         return ret;
4451 }
4452
4453 int perf_event_overflow(struct perf_event *event, int nmi,
4454                           struct perf_sample_data *data,
4455                           struct pt_regs *regs)
4456 {
4457         return __perf_event_overflow(event, nmi, 1, data, regs);
4458 }
4459
4460 /*
4461  * Generic software event infrastructure
4462  */
4463
4464 struct swevent_htable {
4465         struct swevent_hlist            *swevent_hlist;
4466         struct mutex                    hlist_mutex;
4467         int                             hlist_refcount;
4468
4469         /* Recursion avoidance in each contexts */
4470         int                             recursion[PERF_NR_CONTEXTS];
4471 };
4472
4473 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
4474
4475 /*
4476  * We directly increment event->count and keep a second value in
4477  * event->hw.period_left to count intervals. This period event
4478  * is kept in the range [-sample_period, 0] so that we can use the
4479  * sign as trigger.
4480  */
4481
4482 static u64 perf_swevent_set_period(struct perf_event *event)
4483 {
4484         struct hw_perf_event *hwc = &event->hw;
4485         u64 period = hwc->last_period;
4486         u64 nr, offset;
4487         s64 old, val;
4488
4489         hwc->last_period = hwc->sample_period;
4490
4491 again:
4492         old = val = local64_read(&hwc->period_left);
4493         if (val < 0)
4494                 return 0;
4495
4496         nr = div64_u64(period + val, period);
4497         offset = nr * period;
4498         val -= offset;
4499         if (local64_cmpxchg(&hwc->period_left, old, val) != old)
4500                 goto again;
4501
4502         return nr;
4503 }
4504
4505 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4506                                     int nmi, struct perf_sample_data *data,
4507                                     struct pt_regs *regs)
4508 {
4509         struct hw_perf_event *hwc = &event->hw;
4510         int throttle = 0;
4511
4512         data->period = event->hw.last_period;
4513         if (!overflow)
4514                 overflow = perf_swevent_set_period(event);
4515
4516         if (hwc->interrupts == MAX_INTERRUPTS)
4517                 return;
4518
4519         for (; overflow; overflow--) {
4520                 if (__perf_event_overflow(event, nmi, throttle,
4521                                             data, regs)) {
4522                         /*
4523                          * We inhibit the overflow from happening when
4524                          * hwc->interrupts == MAX_INTERRUPTS.
4525                          */
4526                         break;
4527                 }
4528                 throttle = 1;
4529         }
4530 }
4531
4532 static void perf_swevent_event(struct perf_event *event, u64 nr,
4533                                int nmi, struct perf_sample_data *data,
4534                                struct pt_regs *regs)
4535 {
4536         struct hw_perf_event *hwc = &event->hw;
4537
4538         local64_add(nr, &event->count);
4539
4540         if (!regs)
4541                 return;
4542
4543         if (!is_sampling_event(event))
4544                 return;
4545
4546         if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4547                 return perf_swevent_overflow(event, 1, nmi, data, regs);
4548
4549         if (local64_add_negative(nr, &hwc->period_left))
4550                 return;
4551
4552         perf_swevent_overflow(event, 0, nmi, data, regs);
4553 }
4554
4555 static int perf_exclude_event(struct perf_event *event,
4556                               struct pt_regs *regs)
4557 {
4558         if (event->hw.state & PERF_HES_STOPPED)
4559                 return 0;
4560
4561         if (regs) {
4562                 if (event->attr.exclude_user && user_mode(regs))
4563                         return 1;
4564
4565                 if (event->attr.exclude_kernel && !user_mode(regs))
4566                         return 1;
4567         }
4568
4569         return 0;
4570 }
4571
4572 static int perf_swevent_match(struct perf_event *event,
4573                                 enum perf_type_id type,
4574                                 u32 event_id,
4575                                 struct perf_sample_data *data,
4576                                 struct pt_regs *regs)
4577 {
4578         if (event->attr.type != type)
4579                 return 0;
4580
4581         if (event->attr.config != event_id)
4582                 return 0;
4583
4584         if (perf_exclude_event(event, regs))
4585                 return 0;
4586
4587         return 1;
4588 }
4589
4590 static inline u64 swevent_hash(u64 type, u32 event_id)
4591 {
4592         u64 val = event_id | (type << 32);
4593
4594         return hash_64(val, SWEVENT_HLIST_BITS);
4595 }
4596
4597 static inline struct hlist_head *
4598 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4599 {
4600         u64 hash = swevent_hash(type, event_id);
4601
4602         return &hlist->heads[hash];
4603 }
4604
4605 /* For the read side: events when they trigger */
4606 static inline struct hlist_head *
4607 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
4608 {
4609         struct swevent_hlist *hlist;
4610
4611         hlist = rcu_dereference(swhash->swevent_hlist);
4612         if (!hlist)
4613                 return NULL;
4614
4615         return __find_swevent_head(hlist, type, event_id);
4616 }
4617
4618 /* For the event head insertion and removal in the hlist */
4619 static inline struct hlist_head *
4620 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
4621 {
4622         struct swevent_hlist *hlist;
4623         u32 event_id = event->attr.config;
4624         u64 type = event->attr.type;
4625
4626         /*
4627          * Event scheduling is always serialized against hlist allocation
4628          * and release. Which makes the protected version suitable here.
4629          * The context lock guarantees that.
4630          */
4631         hlist = rcu_dereference_protected(swhash->swevent_hlist,
4632                                           lockdep_is_held(&event->ctx->lock));
4633         if (!hlist)
4634                 return NULL;
4635
4636         return __find_swevent_head(hlist, type, event_id);
4637 }
4638
4639 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4640                                     u64 nr, int nmi,
4641                                     struct perf_sample_data *data,
4642                                     struct pt_regs *regs)
4643 {
4644         struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4645         struct perf_event *event;
4646         struct hlist_node *node;
4647         struct hlist_head *head;
4648
4649         rcu_read_lock();
4650         head = find_swevent_head_rcu(swhash, type, event_id);
4651         if (!head)
4652                 goto end;
4653
4654         hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4655                 if (perf_swevent_match(event, type, event_id, data, regs))
4656                         perf_swevent_event(event, nr, nmi, data, regs);
4657         }
4658 end:
4659         rcu_read_unlock();
4660 }
4661
4662 int perf_swevent_get_recursion_context(void)
4663 {
4664         struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4665
4666         return get_recursion_context(swhash->recursion);
4667 }
4668 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4669
4670 inline void perf_swevent_put_recursion_context(int rctx)
4671 {
4672         struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4673
4674         put_recursion_context(swhash->recursion, rctx);
4675 }
4676
4677 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4678                             struct pt_regs *regs, u64 addr)
4679 {
4680         struct perf_sample_data data;
4681         int rctx;
4682
4683         preempt_disable_notrace();
4684         rctx = perf_swevent_get_recursion_context();
4685         if (rctx < 0)
4686                 return;
4687
4688         perf_sample_data_init(&data, addr);
4689
4690         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4691
4692         perf_swevent_put_recursion_context(rctx);
4693         preempt_enable_notrace();
4694 }
4695
4696 static void perf_swevent_read(struct perf_event *event)
4697 {
4698 }
4699
4700 static int perf_swevent_add(struct perf_event *event, int flags)
4701 {
4702         struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4703         struct hw_perf_event *hwc = &event->hw;
4704         struct hlist_head *head;
4705
4706         if (is_sampling_event(event)) {
4707                 hwc->last_period = hwc->sample_period;
4708                 perf_swevent_set_period(event);
4709         }
4710
4711         hwc->state = !(flags & PERF_EF_START);
4712
4713         head = find_swevent_head(swhash, event);
4714         if (WARN_ON_ONCE(!head))
4715                 return -EINVAL;
4716
4717         hlist_add_head_rcu(&event->hlist_entry, head);
4718
4719         return 0;
4720 }
4721
4722 static void perf_swevent_del(struct perf_event *event, int flags)
4723 {
4724         hlist_del_rcu(&event->hlist_entry);
4725 }
4726
4727 static void perf_swevent_start(struct perf_event *event, int flags)
4728 {
4729         event->hw.state = 0;
4730 }
4731
4732 static void perf_swevent_stop(struct perf_event *event, int flags)
4733 {
4734         event->hw.state = PERF_HES_STOPPED;
4735 }
4736
4737 /* Deref the hlist from the update side */
4738 static inline struct swevent_hlist *
4739 swevent_hlist_deref(struct swevent_htable *swhash)
4740 {
4741         return rcu_dereference_protected(swhash->swevent_hlist,
4742                                          lockdep_is_held(&swhash->hlist_mutex));
4743 }
4744
4745 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4746 {
4747         struct swevent_hlist *hlist;
4748
4749         hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4750         kfree(hlist);
4751 }
4752
4753 static void swevent_hlist_release(struct swevent_htable *swhash)
4754 {
4755         struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
4756
4757         if (!hlist)
4758                 return;
4759
4760         rcu_assign_pointer(swhash->swevent_hlist, NULL);
4761         call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4762 }
4763
4764 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4765 {
4766         struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4767
4768         mutex_lock(&swhash->hlist_mutex);
4769
4770         if (!--swhash->hlist_refcount)
4771                 swevent_hlist_release(swhash);
4772
4773         mutex_unlock(&swhash->hlist_mutex);
4774 }
4775
4776 static void swevent_hlist_put(struct perf_event *event)
4777 {
4778         int cpu;
4779
4780         if (event->cpu != -1) {
4781                 swevent_hlist_put_cpu(event, event->cpu);
4782                 return;
4783         }
4784
4785         for_each_possible_cpu(cpu)
4786                 swevent_hlist_put_cpu(event, cpu);
4787 }
4788
4789 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4790 {
4791         struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4792         int err = 0;
4793
4794         mutex_lock(&swhash->hlist_mutex);
4795
4796         if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
4797                 struct swevent_hlist *hlist;
4798
4799                 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4800                 if (!hlist) {
4801                         err = -ENOMEM;
4802                         goto exit;
4803                 }
4804                 rcu_assign_pointer(swhash->swevent_hlist, hlist);
4805         }
4806         swhash->hlist_refcount++;
4807 exit:
4808         mutex_unlock(&swhash->hlist_mutex);
4809
4810         return err;
4811 }
4812
4813 static int swevent_hlist_get(struct perf_event *event)
4814 {
4815         int err;
4816         int cpu, failed_cpu;
4817
4818         if (event->cpu != -1)
4819                 return swevent_hlist_get_cpu(event, event->cpu);
4820
4821         get_online_cpus();
4822         for_each_possible_cpu(cpu) {
4823                 err = swevent_hlist_get_cpu(event, cpu);
4824                 if (err) {
4825                         failed_cpu = cpu;
4826                         goto fail;
4827                 }
4828         }
4829         put_online_cpus();
4830
4831         return 0;
4832 fail:
4833         for_each_possible_cpu(cpu) {
4834                 if (cpu == failed_cpu)
4835                         break;
4836                 swevent_hlist_put_cpu(event, cpu);
4837         }
4838
4839         put_online_cpus();
4840         return err;
4841 }
4842
4843 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4844
4845 static void sw_perf_event_destroy(struct perf_event *event)
4846 {
4847         u64 event_id = event->attr.config;
4848
4849         WARN_ON(event->parent);
4850
4851         jump_label_dec(&perf_swevent_enabled[event_id]);
4852         swevent_hlist_put(event);
4853 }
4854
4855 static int perf_swevent_init(struct perf_event *event)
4856 {
4857         int event_id = event->attr.config;
4858
4859         if (event->attr.type != PERF_TYPE_SOFTWARE)
4860                 return -ENOENT;
4861
4862         switch (event_id) {
4863         case PERF_COUNT_SW_CPU_CLOCK:
4864         case PERF_COUNT_SW_TASK_CLOCK:
4865                 return -ENOENT;
4866
4867         default:
4868                 break;
4869         }
4870
4871         if (event_id >= PERF_COUNT_SW_MAX)
4872                 return -ENOENT;
4873
4874         if (!event->parent) {
4875                 int err;
4876
4877                 err = swevent_hlist_get(event);
4878                 if (err)
4879                         return err;
4880
4881                 jump_label_inc(&perf_swevent_enabled[event_id]);
4882                 event->destroy = sw_perf_event_destroy;
4883         }
4884
4885         return 0;
4886 }
4887
4888 static struct pmu perf_swevent = {
4889         .task_ctx_nr    = perf_sw_context,
4890
4891         .event_init     = perf_swevent_init,
4892         .add            = perf_swevent_add,
4893         .del            = perf_swevent_del,
4894         .start          = perf_swevent_start,
4895         .stop           = perf_swevent_stop,
4896         .read           = perf_swevent_read,
4897 };
4898
4899 #ifdef CONFIG_EVENT_TRACING
4900
4901 static int perf_tp_filter_match(struct perf_event *event,
4902                                 struct perf_sample_data *data)
4903 {
4904         void *record = data->raw->data;
4905
4906         if (likely(!event->filter) || filter_match_preds(event->filter, record))
4907                 return 1;
4908         return 0;
4909 }
4910
4911 static int perf_tp_event_match(struct perf_event *event,
4912                                 struct perf_sample_data *data,
4913                                 struct pt_regs *regs)
4914 {
4915         /*
4916          * All tracepoints are from kernel-space.
4917          */
4918         if (event->attr.exclude_kernel)
4919                 return 0;
4920
4921         if (!perf_tp_filter_match(event, data))
4922                 return 0;
4923
4924         return 1;
4925 }
4926
4927 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
4928                    struct pt_regs *regs, struct hlist_head *head, int rctx)
4929 {
4930         struct perf_sample_data data;
4931         struct perf_event *event;
4932         struct hlist_node *node;
4933
4934         struct perf_raw_record raw = {
4935                 .size = entry_size,
4936                 .data = record,
4937         };
4938
4939         perf_sample_data_init(&data, addr);
4940         data.raw = &raw;
4941
4942         hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4943                 if (perf_tp_event_match(event, &data, regs))
4944                         perf_swevent_event(event, count, 1, &data, regs);
4945         }
4946
4947         perf_swevent_put_recursion_context(rctx);
4948 }
4949 EXPORT_SYMBOL_GPL(perf_tp_event);
4950
4951 static void tp_perf_event_destroy(struct perf_event *event)
4952 {
4953         perf_trace_destroy(event);
4954 }
4955
4956 static int perf_tp_event_init(struct perf_event *event)
4957 {
4958         int err;
4959
4960         if (event->attr.type != PERF_TYPE_TRACEPOINT)
4961                 return -ENOENT;
4962
4963         err = perf_trace_init(event);
4964         if (err)
4965                 return err;
4966
4967         event->destroy = tp_perf_event_destroy;
4968
4969         return 0;
4970 }
4971
4972 static struct pmu perf_tracepoint = {
4973         .task_ctx_nr    = perf_sw_context,
4974
4975         .event_init     = perf_tp_event_init,
4976         .add            = perf_trace_add,
4977         .del            = perf_trace_del,
4978         .start          = perf_swevent_start,
4979         .stop           = perf_swevent_stop,
4980         .read           = perf_swevent_read,
4981 };
4982
4983 static inline void perf_tp_register(void)
4984 {
4985         perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
4986 }
4987
4988 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4989 {
4990         char *filter_str;
4991         int ret;
4992
4993         if (event->attr.type != PERF_TYPE_TRACEPOINT)
4994                 return -EINVAL;
4995
4996         filter_str = strndup_user(arg, PAGE_SIZE);
4997         if (IS_ERR(filter_str))
4998                 return PTR_ERR(filter_str);
4999
5000         ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5001
5002         kfree(filter_str);
5003         return ret;
5004 }
5005
5006 static void perf_event_free_filter(struct perf_event *event)
5007 {
5008         ftrace_profile_free_filter(event);
5009 }
5010
5011 #else
5012
5013 static inline void perf_tp_register(void)
5014 {
5015 }
5016
5017 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5018 {
5019         return -ENOENT;
5020 }
5021
5022 static void perf_event_free_filter(struct perf_event *event)
5023 {
5024 }
5025
5026 #endif /* CONFIG_EVENT_TRACING */
5027
5028 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5029 void perf_bp_event(struct perf_event *bp, void *data)
5030 {
5031         struct perf_sample_data sample;
5032         struct pt_regs *regs = data;
5033
5034         perf_sample_data_init(&sample, bp->attr.bp_addr);
5035
5036         if (!bp->hw.state && !perf_exclude_event(bp, regs))
5037                 perf_swevent_event(bp, 1, 1, &sample, regs);
5038 }
5039 #endif
5040
5041 /*
5042  * hrtimer based swevent callback
5043  */
5044
5045 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
5046 {
5047         enum hrtimer_restart ret = HRTIMER_RESTART;
5048         struct perf_sample_data data;
5049         struct pt_regs *regs;
5050         struct perf_event *event;
5051         u64 period;
5052
5053         event = container_of(hrtimer, struct perf_event, hw.hrtimer);
5054         event->pmu->read(event);
5055
5056         perf_sample_data_init(&data, 0);
5057         data.period = event->hw.last_period;
5058         regs = get_irq_regs();
5059
5060         if (regs && !perf_exclude_event(event, regs)) {
5061                 if (!(event->attr.exclude_idle && current->pid == 0))
5062                         if (perf_event_overflow(event, 0, &data, regs))
5063                                 ret = HRTIMER_NORESTART;
5064         }
5065
5066         period = max_t(u64, 10000, event->hw.sample_period);
5067         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
5068
5069         return ret;
5070 }
5071
5072 static void perf_swevent_start_hrtimer(struct perf_event *event)
5073 {
5074         struct hw_perf_event *hwc = &event->hw;
5075         s64 period;
5076
5077         if (!is_sampling_event(event))
5078                 return;
5079
5080         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5081         hwc->hrtimer.function = perf_swevent_hrtimer;
5082
5083         period = local64_read(&hwc->period_left);
5084         if (period) {
5085                 if (period < 0)
5086                         period = 10000;
5087
5088                 local64_set(&hwc->period_left, 0);
5089         } else {
5090                 period = max_t(u64, 10000, hwc->sample_period);
5091         }
5092         __hrtimer_start_range_ns(&hwc->hrtimer,
5093                                 ns_to_ktime(period), 0,
5094                                 HRTIMER_MODE_REL_PINNED, 0);
5095 }
5096
5097 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
5098 {
5099         struct hw_perf_event *hwc = &event->hw;
5100
5101         if (is_sampling_event(event)) {
5102                 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
5103                 local64_set(&hwc->period_left, ktime_to_ns(remaining));
5104
5105                 hrtimer_cancel(&hwc->hrtimer);
5106         }
5107 }
5108
5109 /*
5110  * Software event: cpu wall time clock
5111  */
5112
5113 static void cpu_clock_event_update(struct perf_event *event)
5114 {
5115         s64 prev;
5116         u64 now;
5117
5118         now = local_clock();
5119         prev = local64_xchg(&event->hw.prev_count, now);
5120         local64_add(now - prev, &event->count);
5121 }
5122
5123 static void cpu_clock_event_start(struct perf_event *event, int flags)
5124 {
5125         local64_set(&event->hw.prev_count, local_clock());
5126         perf_swevent_start_hrtimer(event);
5127 }
5128
5129 static void cpu_clock_event_stop(struct perf_event *event, int flags)
5130 {
5131         perf_swevent_cancel_hrtimer(event);
5132         cpu_clock_event_update(event);
5133 }
5134
5135 static int cpu_clock_event_add(struct perf_event *event, int flags)
5136 {
5137         if (flags & PERF_EF_START)
5138                 cpu_clock_event_start(event, flags);
5139
5140         return 0;
5141 }
5142
5143 static void cpu_clock_event_del(struct perf_event *event, int flags)
5144 {
5145         cpu_clock_event_stop(event, flags);
5146 }
5147
5148 static void cpu_clock_event_read(struct perf_event *event)
5149 {
5150         cpu_clock_event_update(event);
5151 }
5152
5153 static int cpu_clock_event_init(struct perf_event *event)
5154 {
5155         if (event->attr.type != PERF_TYPE_SOFTWARE)
5156                 return -ENOENT;
5157
5158         if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
5159                 return -ENOENT;
5160
5161         return 0;
5162 }
5163
5164 static struct pmu perf_cpu_clock = {
5165         .task_ctx_nr    = perf_sw_context,
5166
5167         .event_init     = cpu_clock_event_init,
5168         .add            = cpu_clock_event_add,
5169         .del            = cpu_clock_event_del,
5170         .start          = cpu_clock_event_start,
5171         .stop           = cpu_clock_event_stop,
5172         .read           = cpu_clock_event_read,
5173 };
5174
5175 /*
5176  * Software event: task time clock
5177  */
5178
5179 static void task_clock_event_update(struct perf_event *event, u64 now)
5180 {
5181         u64 prev;
5182         s64 delta;
5183
5184         prev = local64_xchg(&event->hw.prev_count, now);
5185         delta = now - prev;
5186         local64_add(delta, &event->count);
5187 }
5188
5189 static void task_clock_event_start(struct perf_event *event, int flags)
5190 {
5191         local64_set(&event->hw.prev_count, event->ctx->time);
5192         perf_swevent_start_hrtimer(event);
5193 }
5194
5195 static void task_clock_event_stop(struct perf_event *event, int flags)
5196 {
5197         perf_swevent_cancel_hrtimer(event);
5198         task_clock_event_update(event, event->ctx->time);
5199 }
5200
5201 static int task_clock_event_add(struct perf_event *event, int flags)
5202 {
5203         if (flags & PERF_EF_START)
5204                 task_clock_event_start(event, flags);
5205
5206         return 0;
5207 }
5208
5209 static void task_clock_event_del(struct perf_event *event, int flags)
5210 {
5211         task_clock_event_stop(event, PERF_EF_UPDATE);
5212 }
5213
5214 static void task_clock_event_read(struct perf_event *event)
5215 {
5216         u64 time;
5217
5218         if (!in_nmi()) {
5219                 update_context_time(event->ctx);
5220                 time = event->ctx->time;
5221         } else {
5222                 u64 now = perf_clock();
5223                 u64 delta = now - event->ctx->timestamp;
5224                 time = event->ctx->time + delta;
5225         }
5226
5227         task_clock_event_update(event, time);
5228 }
5229
5230 static int task_clock_event_init(struct perf_event *event)
5231 {
5232         if (event->attr.type != PERF_TYPE_SOFTWARE)
5233                 return -ENOENT;
5234
5235         if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5236                 return -ENOENT;
5237
5238         return 0;
5239 }
5240
5241 static struct pmu perf_task_clock = {
5242         .task_ctx_nr    = perf_sw_context,
5243
5244         .event_init     = task_clock_event_init,
5245         .add            = task_clock_event_add,
5246         .del            = task_clock_event_del,
5247         .start          = task_clock_event_start,
5248         .stop           = task_clock_event_stop,
5249         .read           = task_clock_event_read,
5250 };
5251
5252 static void perf_pmu_nop_void(struct pmu *pmu)
5253 {
5254 }
5255
5256 static int perf_pmu_nop_int(struct pmu *pmu)
5257 {
5258         return 0;
5259 }
5260
5261 static void perf_pmu_start_txn(struct pmu *pmu)
5262 {
5263         perf_pmu_disable(pmu);
5264 }
5265
5266 static int perf_pmu_commit_txn(struct pmu *pmu)
5267 {
5268         perf_pmu_enable(pmu);
5269         return 0;
5270 }
5271
5272 static void perf_pmu_cancel_txn(struct pmu *pmu)
5273 {
5274         perf_pmu_enable(pmu);
5275 }
5276
5277 /*
5278  * Ensures all contexts with the same task_ctx_nr have the same
5279  * pmu_cpu_context too.
5280  */
5281 static void *find_pmu_context(int ctxn)
5282 {
5283         struct pmu *pmu;
5284
5285         if (ctxn < 0)
5286                 return NULL;
5287
5288         list_for_each_entry(pmu, &pmus, entry) {
5289                 if (pmu->task_ctx_nr == ctxn)
5290                         return pmu->pmu_cpu_context;
5291         }
5292
5293         return NULL;
5294 }
5295
5296 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
5297 {
5298         int cpu;
5299
5300         for_each_possible_cpu(cpu) {
5301                 struct perf_cpu_context *cpuctx;
5302
5303                 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5304
5305                 if (cpuctx->active_pmu == old_pmu)
5306                         cpuctx->active_pmu = pmu;
5307         }
5308 }
5309
5310 static void free_pmu_context(struct pmu *pmu)
5311 {
5312         struct pmu *i;
5313
5314         mutex_lock(&pmus_lock);
5315         /*
5316          * Like a real lame refcount.
5317          */
5318         list_for_each_entry(i, &pmus, entry) {
5319                 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
5320                         update_pmu_context(i, pmu);
5321                         goto out;
5322                 }
5323         }
5324
5325         free_percpu(pmu->pmu_cpu_context);
5326 out:
5327         mutex_unlock(&pmus_lock);
5328 }
5329 static struct idr pmu_idr;
5330
5331 static ssize_t
5332 type_show(struct device *dev, struct device_attribute *attr, char *page)
5333 {
5334         struct pmu *pmu = dev_get_drvdata(dev);
5335
5336         return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
5337 }
5338
5339 static struct device_attribute pmu_dev_attrs[] = {
5340        __ATTR_RO(type),
5341        __ATTR_NULL,
5342 };
5343
5344 static int pmu_bus_running;
5345 static struct bus_type pmu_bus = {
5346         .name           = "event_source",
5347         .dev_attrs      = pmu_dev_attrs,
5348 };
5349
5350 static void pmu_dev_release(struct device *dev)
5351 {
5352         kfree(dev);
5353 }
5354
5355 static int pmu_dev_alloc(struct pmu *pmu)
5356 {
5357         int ret = -ENOMEM;
5358
5359         pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
5360         if (!pmu->dev)
5361                 goto out;
5362
5363         device_initialize(pmu->dev);
5364         ret = dev_set_name(pmu->dev, "%s", pmu->name);
5365         if (ret)
5366                 goto free_dev;
5367
5368         dev_set_drvdata(pmu->dev, pmu);
5369         pmu->dev->bus = &pmu_bus;
5370         pmu->dev->release = pmu_dev_release;
5371         ret = device_add(pmu->dev);
5372         if (ret)
5373                 goto free_dev;
5374
5375 out:
5376         return ret;
5377
5378 free_dev:
5379         put_device(pmu->dev);
5380         goto out;
5381 }
5382
5383 static struct lock_class_key cpuctx_mutex;
5384
5385 int perf_pmu_register(struct pmu *pmu, char *name, int type)
5386 {
5387         int cpu, ret;
5388
5389         mutex_lock(&pmus_lock);
5390         ret = -ENOMEM;
5391         pmu->pmu_disable_count = alloc_percpu(int);
5392         if (!pmu->pmu_disable_count)
5393                 goto unlock;
5394
5395         pmu->type = -1;
5396         if (!name)
5397                 goto skip_type;
5398         pmu->name = name;
5399
5400         if (type < 0) {
5401                 int err = idr_pre_get(&pmu_idr, GFP_KERNEL);
5402                 if (!err)
5403                         goto free_pdc;
5404
5405                 err = idr_get_new_above(&pmu_idr, pmu, PERF_TYPE_MAX, &type);
5406                 if (err) {
5407                         ret = err;
5408                         goto free_pdc;
5409                 }
5410         }
5411         pmu->type = type;
5412
5413         if (pmu_bus_running) {
5414                 ret = pmu_dev_alloc(pmu);
5415                 if (ret)
5416                         goto free_idr;
5417         }
5418
5419 skip_type:
5420         pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
5421         if (pmu->pmu_cpu_context)
5422                 goto got_cpu_context;
5423
5424         pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
5425         if (!pmu->pmu_cpu_context)
5426                 goto free_dev;
5427
5428         for_each_possible_cpu(cpu) {
5429                 struct perf_cpu_context *cpuctx;
5430
5431                 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5432                 __perf_event_init_context(&cpuctx->ctx);
5433                 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
5434                 cpuctx->ctx.type = cpu_context;
5435                 cpuctx->ctx.pmu = pmu;
5436                 cpuctx->jiffies_interval = 1;
5437                 INIT_LIST_HEAD(&cpuctx->rotation_list);
5438                 cpuctx->active_pmu = pmu;
5439         }
5440
5441 got_cpu_context:
5442         if (!pmu->start_txn) {
5443                 if (pmu->pmu_enable) {
5444                         /*
5445                          * If we have pmu_enable/pmu_disable calls, install
5446                          * transaction stubs that use that to try and batch
5447                          * hardware accesses.
5448                          */
5449                         pmu->start_txn  = perf_pmu_start_txn;
5450                         pmu->commit_txn = perf_pmu_commit_txn;
5451                         pmu->cancel_txn = perf_pmu_cancel_txn;
5452                 } else {
5453                         pmu->start_txn  = perf_pmu_nop_void;
5454                         pmu->commit_txn = perf_pmu_nop_int;
5455                         pmu->cancel_txn = perf_pmu_nop_void;
5456                 }
5457         }
5458
5459         if (!pmu->pmu_enable) {
5460                 pmu->pmu_enable  = perf_pmu_nop_void;
5461                 pmu->pmu_disable = perf_pmu_nop_void;
5462         }
5463
5464         list_add_rcu(&pmu->entry, &pmus);
5465         ret = 0;
5466 unlock:
5467         mutex_unlock(&pmus_lock);
5468
5469         return ret;
5470
5471 free_dev:
5472         device_del(pmu->dev);
5473         put_device(pmu->dev);
5474
5475 free_idr:
5476         if (pmu->type >= PERF_TYPE_MAX)
5477                 idr_remove(&pmu_idr, pmu->type);
5478
5479 free_pdc:
5480         free_percpu(pmu->pmu_disable_count);
5481         goto unlock;
5482 }
5483
5484 void perf_pmu_unregister(struct pmu *pmu)
5485 {
5486         mutex_lock(&pmus_lock);
5487         list_del_rcu(&pmu->entry);
5488         mutex_unlock(&pmus_lock);
5489
5490         /*
5491          * We dereference the pmu list under both SRCU and regular RCU, so
5492          * synchronize against both of those.
5493          */
5494         synchronize_srcu(&pmus_srcu);
5495         synchronize_rcu();
5496
5497         free_percpu(pmu->pmu_disable_count);
5498         if (pmu->type >= PERF_TYPE_MAX)
5499                 idr_remove(&pmu_idr, pmu->type);
5500         device_del(pmu->dev);
5501         put_device(pmu->dev);
5502         free_pmu_context(pmu);
5503 }
5504
5505 struct pmu *perf_init_event(struct perf_event *event)
5506 {
5507         struct pmu *pmu = NULL;
5508         int idx;
5509
5510         idx = srcu_read_lock(&pmus_srcu);
5511
5512         rcu_read_lock();
5513         pmu = idr_find(&pmu_idr, event->attr.type);
5514         rcu_read_unlock();
5515         if (pmu)
5516                 goto unlock;
5517
5518         list_for_each_entry_rcu(pmu, &pmus, entry) {
5519                 int ret = pmu->event_init(event);
5520                 if (!ret)
5521                         goto unlock;
5522
5523                 if (ret != -ENOENT) {
5524                         pmu = ERR_PTR(ret);
5525                         goto unlock;
5526                 }
5527         }
5528         pmu = ERR_PTR(-ENOENT);
5529 unlock:
5530         srcu_read_unlock(&pmus_srcu, idx);
5531
5532         return pmu;
5533 }
5534
5535 /*
5536  * Allocate and initialize a event structure
5537  */
5538 static struct perf_event *
5539 perf_event_alloc(struct perf_event_attr *attr, int cpu,
5540                  struct task_struct *task,
5541                  struct perf_event *group_leader,
5542                  struct perf_event *parent_event,
5543                  perf_overflow_handler_t overflow_handler)
5544 {
5545         struct pmu *pmu;
5546         struct perf_event *event;
5547         struct hw_perf_event *hwc;
5548         long err;
5549
5550         if ((unsigned)cpu >= nr_cpu_ids) {
5551                 if (!task || cpu != -1)
5552                         return ERR_PTR(-EINVAL);
5553         }
5554
5555         event = kzalloc(sizeof(*event), GFP_KERNEL);
5556         if (!event)
5557                 return ERR_PTR(-ENOMEM);
5558
5559         /*
5560          * Single events are their own group leaders, with an
5561          * empty sibling list:
5562          */
5563         if (!group_leader)
5564                 group_leader = event;
5565
5566         mutex_init(&event->child_mutex);
5567         INIT_LIST_HEAD(&event->child_list);
5568
5569         INIT_LIST_HEAD(&event->group_entry);
5570         INIT_LIST_HEAD(&event->event_entry);
5571         INIT_LIST_HEAD(&event->sibling_list);
5572         init_waitqueue_head(&event->waitq);
5573         init_irq_work(&event->pending, perf_pending_event);
5574
5575         mutex_init(&event->mmap_mutex);
5576
5577         event->cpu              = cpu;
5578         event->attr             = *attr;
5579         event->group_leader     = group_leader;
5580         event->pmu              = NULL;
5581         event->oncpu            = -1;
5582
5583         event->parent           = parent_event;
5584
5585         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
5586         event->id               = atomic64_inc_return(&perf_event_id);
5587
5588         event->state            = PERF_EVENT_STATE_INACTIVE;
5589
5590         if (task) {
5591                 event->attach_state = PERF_ATTACH_TASK;
5592 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5593                 /*
5594                  * hw_breakpoint is a bit difficult here..
5595                  */
5596                 if (attr->type == PERF_TYPE_BREAKPOINT)
5597                         event->hw.bp_target = task;
5598 #endif
5599         }
5600
5601         if (!overflow_handler && parent_event)
5602                 overflow_handler = parent_event->overflow_handler;
5603
5604         event->overflow_handler = overflow_handler;
5605
5606         if (attr->disabled)
5607                 event->state = PERF_EVENT_STATE_OFF;
5608
5609         pmu = NULL;
5610
5611         hwc = &event->hw;
5612         hwc->sample_period = attr->sample_period;
5613         if (attr->freq && attr->sample_freq)
5614                 hwc->sample_period = 1;
5615         hwc->last_period = hwc->sample_period;
5616
5617         local64_set(&hwc->period_left, hwc->sample_period);
5618
5619         /*
5620          * we currently do not support PERF_FORMAT_GROUP on inherited events
5621          */
5622         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
5623                 goto done;
5624
5625         pmu = perf_init_event(event);
5626
5627 done:
5628         err = 0;
5629         if (!pmu)
5630                 err = -EINVAL;
5631         else if (IS_ERR(pmu))
5632                 err = PTR_ERR(pmu);
5633
5634         if (err) {
5635                 if (event->ns)
5636                         put_pid_ns(event->ns);
5637                 kfree(event);
5638                 return ERR_PTR(err);
5639         }
5640
5641         event->pmu = pmu;
5642
5643         if (!event->parent) {
5644                 if (event->attach_state & PERF_ATTACH_TASK)
5645                         jump_label_inc(&perf_task_events);
5646                 if (event->attr.mmap || event->attr.mmap_data)
5647                         atomic_inc(&nr_mmap_events);
5648                 if (event->attr.comm)
5649                         atomic_inc(&nr_comm_events);
5650                 if (event->attr.task)
5651                         atomic_inc(&nr_task_events);
5652                 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
5653                         err = get_callchain_buffers();
5654                         if (err) {
5655                                 free_event(event);
5656                                 return ERR_PTR(err);
5657                         }
5658                 }
5659         }
5660
5661         return event;
5662 }
5663
5664 static int perf_copy_attr(struct perf_event_attr __user *uattr,
5665                           struct perf_event_attr *attr)
5666 {
5667         u32 size;
5668         int ret;
5669
5670         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
5671                 return -EFAULT;
5672
5673         /*
5674          * zero the full structure, so that a short copy will be nice.
5675          */
5676         memset(attr, 0, sizeof(*attr));
5677
5678         ret = get_user(size, &uattr->size);
5679         if (ret)
5680                 return ret;
5681
5682         if (size > PAGE_SIZE)   /* silly large */
5683                 goto err_size;
5684
5685         if (!size)              /* abi compat */
5686                 size = PERF_ATTR_SIZE_VER0;
5687
5688         if (size < PERF_ATTR_SIZE_VER0)
5689                 goto err_size;
5690
5691         /*
5692          * If we're handed a bigger struct than we know of,
5693          * ensure all the unknown bits are 0 - i.e. new
5694          * user-space does not rely on any kernel feature
5695          * extensions we dont know about yet.
5696          */
5697         if (size > sizeof(*attr)) {
5698                 unsigned char __user *addr;
5699                 unsigned char __user *end;
5700                 unsigned char val;
5701
5702                 addr = (void __user *)uattr + sizeof(*attr);
5703                 end  = (void __user *)uattr + size;
5704
5705                 for (; addr < end; addr++) {
5706                         ret = get_user(val, addr);
5707                         if (ret)
5708                                 return ret;
5709                         if (val)
5710                                 goto err_size;
5711                 }
5712                 size = sizeof(*attr);
5713         }
5714
5715         ret = copy_from_user(attr, uattr, size);
5716         if (ret)
5717                 return -EFAULT;
5718
5719         /*
5720          * If the type exists, the corresponding creation will verify
5721          * the attr->config.
5722          */
5723         if (attr->type >= PERF_TYPE_MAX)
5724                 return -EINVAL;
5725
5726         if (attr->__reserved_1)
5727                 return -EINVAL;
5728
5729         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
5730                 return -EINVAL;
5731
5732         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
5733                 return -EINVAL;
5734
5735 out:
5736         return ret;
5737
5738 err_size:
5739         put_user(sizeof(*attr), &uattr->size);
5740         ret = -E2BIG;
5741         goto out;
5742 }
5743
5744 static int
5745 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
5746 {
5747         struct perf_buffer *buffer = NULL, *old_buffer = NULL;
5748         int ret = -EINVAL;
5749
5750         if (!output_event)
5751                 goto set;
5752
5753         /* don't allow circular references */
5754         if (event == output_event)
5755                 goto out;
5756
5757         /*
5758          * Don't allow cross-cpu buffers
5759          */
5760         if (output_event->cpu != event->cpu)
5761                 goto out;
5762
5763         /*
5764          * If its not a per-cpu buffer, it must be the same task.
5765          */
5766         if (output_event->cpu == -1 && output_event->ctx != event->ctx)
5767                 goto out;
5768
5769 set:
5770         mutex_lock(&event->mmap_mutex);
5771         /* Can't redirect output if we've got an active mmap() */
5772         if (atomic_read(&event->mmap_count))
5773                 goto unlock;
5774
5775         if (output_event) {
5776                 /* get the buffer we want to redirect to */
5777                 buffer = perf_buffer_get(output_event);
5778                 if (!buffer)
5779                         goto unlock;
5780         }
5781
5782         old_buffer = event->buffer;
5783         rcu_assign_pointer(event->buffer, buffer);
5784         ret = 0;
5785 unlock:
5786         mutex_unlock(&event->mmap_mutex);
5787
5788         if (old_buffer)
5789                 perf_buffer_put(old_buffer);
5790 out:
5791         return ret;
5792 }
5793
5794 /**
5795  * sys_perf_event_open - open a performance event, associate it to a task/cpu
5796  *
5797  * @attr_uptr:  event_id type attributes for monitoring/sampling
5798  * @pid:                target pid
5799  * @cpu:                target cpu
5800  * @group_fd:           group leader event fd
5801  */
5802 SYSCALL_DEFINE5(perf_event_open,
5803                 struct perf_event_attr __user *, attr_uptr,
5804                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
5805 {
5806         struct perf_event *group_leader = NULL, *output_event = NULL;
5807         struct perf_event *event, *sibling;
5808         struct perf_event_attr attr;
5809         struct perf_event_context *ctx;
5810         struct file *event_file = NULL;
5811         struct file *group_file = NULL;
5812         struct task_struct *task = NULL;
5813         struct pmu *pmu;
5814         int event_fd;
5815         int move_group = 0;
5816         int fput_needed = 0;
5817         int err;
5818
5819         /* for future expandability... */
5820         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
5821                 return -EINVAL;
5822
5823         err = perf_copy_attr(attr_uptr, &attr);
5824         if (err)
5825                 return err;
5826
5827         if (!attr.exclude_kernel) {
5828                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
5829                         return -EACCES;
5830         }
5831
5832         if (attr.freq) {
5833                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
5834                         return -EINVAL;
5835         }
5836
5837         event_fd = get_unused_fd_flags(O_RDWR);
5838         if (event_fd < 0)
5839                 return event_fd;
5840
5841         if (group_fd != -1) {
5842                 group_leader = perf_fget_light(group_fd, &fput_needed);
5843                 if (IS_ERR(group_leader)) {
5844                         err = PTR_ERR(group_leader);
5845                         goto err_fd;
5846                 }
5847                 group_file = group_leader->filp;
5848                 if (flags & PERF_FLAG_FD_OUTPUT)
5849                         output_event = group_leader;
5850                 if (flags & PERF_FLAG_FD_NO_GROUP)
5851                         group_leader = NULL;
5852         }
5853
5854         if (pid != -1) {
5855                 task = find_lively_task_by_vpid(pid);
5856                 if (IS_ERR(task)) {
5857                         err = PTR_ERR(task);
5858                         goto err_group_fd;
5859                 }
5860         }
5861
5862         event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, NULL);
5863         if (IS_ERR(event)) {
5864                 err = PTR_ERR(event);
5865                 goto err_task;
5866         }
5867
5868         /*
5869          * Special case software events and allow them to be part of
5870          * any hardware group.
5871          */
5872         pmu = event->pmu;
5873
5874         if (group_leader &&
5875             (is_software_event(event) != is_software_event(group_leader))) {
5876                 if (is_software_event(event)) {
5877                         /*
5878                          * If event and group_leader are not both a software
5879                          * event, and event is, then group leader is not.
5880                          *
5881                          * Allow the addition of software events to !software
5882                          * groups, this is safe because software events never
5883                          * fail to schedule.
5884                          */
5885                         pmu = group_leader->pmu;
5886                 } else if (is_software_event(group_leader) &&
5887                            (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
5888                         /*
5889                          * In case the group is a pure software group, and we
5890                          * try to add a hardware event, move the whole group to
5891                          * the hardware context.
5892                          */
5893                         move_group = 1;
5894                 }
5895         }
5896
5897         /*
5898          * Get the target context (task or percpu):
5899          */
5900         ctx = find_get_context(pmu, task, cpu);
5901         if (IS_ERR(ctx)) {
5902                 err = PTR_ERR(ctx);
5903                 goto err_alloc;
5904         }
5905
5906         /*
5907          * Look up the group leader (we will attach this event to it):
5908          */
5909         if (group_leader) {
5910                 err = -EINVAL;
5911
5912                 /*
5913                  * Do not allow a recursive hierarchy (this new sibling
5914                  * becoming part of another group-sibling):
5915                  */
5916                 if (group_leader->group_leader != group_leader)
5917                         goto err_context;
5918                 /*
5919                  * Do not allow to attach to a group in a different
5920                  * task or CPU context:
5921                  */
5922                 if (move_group) {
5923                         if (group_leader->ctx->type != ctx->type)
5924                                 goto err_context;
5925                 } else {
5926                         if (group_leader->ctx != ctx)
5927                                 goto err_context;
5928                 }
5929
5930                 /*
5931                  * Only a group leader can be exclusive or pinned
5932                  */
5933                 if (attr.exclusive || attr.pinned)
5934                         goto err_context;
5935         }
5936
5937         if (output_event) {
5938                 err = perf_event_set_output(event, output_event);
5939                 if (err)
5940                         goto err_context;
5941         }
5942
5943         event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
5944         if (IS_ERR(event_file)) {
5945                 err = PTR_ERR(event_file);
5946                 goto err_context;
5947         }
5948
5949         if (move_group) {
5950                 struct perf_event_context *gctx = group_leader->ctx;
5951
5952                 mutex_lock(&gctx->mutex);
5953                 perf_event_remove_from_context(group_leader);
5954                 list_for_each_entry(sibling, &group_leader->sibling_list,
5955                                     group_entry) {
5956                         perf_event_remove_from_context(sibling);
5957                         put_ctx(gctx);
5958                 }
5959                 mutex_unlock(&gctx->mutex);
5960                 put_ctx(gctx);
5961         }
5962
5963         event->filp = event_file;
5964         WARN_ON_ONCE(ctx->parent_ctx);
5965         mutex_lock(&ctx->mutex);
5966
5967         if (move_group) {
5968                 perf_install_in_context(ctx, group_leader, cpu);
5969                 get_ctx(ctx);
5970                 list_for_each_entry(sibling, &group_leader->sibling_list,
5971                                     group_entry) {
5972                         perf_install_in_context(ctx, sibling, cpu);
5973                         get_ctx(ctx);
5974                 }
5975         }
5976
5977         perf_install_in_context(ctx, event, cpu);
5978         ++ctx->generation;
5979         mutex_unlock(&ctx->mutex);
5980
5981         event->owner = current;
5982
5983         mutex_lock(&current->perf_event_mutex);
5984         list_add_tail(&event->owner_entry, &current->perf_event_list);
5985         mutex_unlock(&current->perf_event_mutex);
5986
5987         /*
5988          * Precalculate sample_data sizes
5989          */
5990         perf_event__header_size(event);
5991         perf_event__id_header_size(event);
5992
5993         /*
5994          * Drop the reference on the group_event after placing the
5995          * new event on the sibling_list. This ensures destruction
5996          * of the group leader will find the pointer to itself in
5997          * perf_group_detach().
5998          */
5999         fput_light(group_file, fput_needed);
6000         fd_install(event_fd, event_file);
6001         return event_fd;
6002
6003 err_context:
6004         put_ctx(ctx);
6005 err_alloc:
6006         free_event(event);
6007 err_task:
6008         if (task)
6009                 put_task_struct(task);
6010 err_group_fd:
6011         fput_light(group_file, fput_needed);
6012 err_fd:
6013         put_unused_fd(event_fd);
6014         return err;
6015 }
6016
6017 /**
6018  * perf_event_create_kernel_counter
6019  *
6020  * @attr: attributes of the counter to create
6021  * @cpu: cpu in which the counter is bound
6022  * @task: task to profile (NULL for percpu)
6023  */
6024 struct perf_event *
6025 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
6026                                  struct task_struct *task,
6027                                  perf_overflow_handler_t overflow_handler)
6028 {
6029         struct perf_event_context *ctx;
6030         struct perf_event *event;
6031         int err;
6032
6033         /*
6034          * Get the target context (task or percpu):
6035          */
6036
6037         event = perf_event_alloc(attr, cpu, task, NULL, NULL, overflow_handler);
6038         if (IS_ERR(event)) {
6039                 err = PTR_ERR(event);
6040                 goto err;
6041         }
6042
6043         ctx = find_get_context(event->pmu, task, cpu);
6044         if (IS_ERR(ctx)) {
6045                 err = PTR_ERR(ctx);
6046                 goto err_free;
6047         }
6048
6049         event->filp = NULL;
6050         WARN_ON_ONCE(ctx->parent_ctx);
6051         mutex_lock(&ctx->mutex);
6052         perf_install_in_context(ctx, event, cpu);
6053         ++ctx->generation;
6054         mutex_unlock(&ctx->mutex);
6055
6056         return event;
6057
6058 err_free:
6059         free_event(event);
6060 err:
6061         return ERR_PTR(err);
6062 }
6063 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
6064
6065 static void sync_child_event(struct perf_event *child_event,
6066                                struct task_struct *child)
6067 {
6068         struct perf_event *parent_event = child_event->parent;
6069         u64 child_val;
6070
6071         if (child_event->attr.inherit_stat)
6072                 perf_event_read_event(child_event, child);
6073
6074         child_val = perf_event_count(child_event);
6075
6076         /*
6077          * Add back the child's count to the parent's count:
6078          */
6079         atomic64_add(child_val, &parent_event->child_count);
6080         atomic64_add(child_event->total_time_enabled,
6081                      &parent_event->child_total_time_enabled);
6082         atomic64_add(child_event->total_time_running,
6083                      &parent_event->child_total_time_running);
6084
6085         /*
6086          * Remove this event from the parent's list
6087          */
6088         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6089         mutex_lock(&parent_event->child_mutex);
6090         list_del_init(&child_event->child_list);
6091         mutex_unlock(&parent_event->child_mutex);
6092
6093         /*
6094          * Release the parent event, if this was the last
6095          * reference to it.
6096          */
6097         fput(parent_event->filp);
6098 }
6099
6100 static void
6101 __perf_event_exit_task(struct perf_event *child_event,
6102                          struct perf_event_context *child_ctx,
6103                          struct task_struct *child)
6104 {
6105         struct perf_event *parent_event;
6106
6107         perf_event_remove_from_context(child_event);
6108
6109         parent_event = child_event->parent;
6110         /*
6111          * It can happen that parent exits first, and has events
6112          * that are still around due to the child reference. These
6113          * events need to be zapped - but otherwise linger.
6114          */
6115         if (parent_event) {
6116                 sync_child_event(child_event, child);
6117                 free_event(child_event);
6118         }
6119 }
6120
6121 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
6122 {
6123         struct perf_event *child_event, *tmp;
6124         struct perf_event_context *child_ctx;
6125         unsigned long flags;
6126
6127         if (likely(!child->perf_event_ctxp[ctxn])) {
6128                 perf_event_task(child, NULL, 0);
6129                 return;
6130         }
6131
6132         local_irq_save(flags);
6133         /*
6134          * We can't reschedule here because interrupts are disabled,
6135          * and either child is current or it is a task that can't be
6136          * scheduled, so we are now safe from rescheduling changing
6137          * our context.
6138          */
6139         child_ctx = rcu_dereference(child->perf_event_ctxp[ctxn]);
6140         task_ctx_sched_out(child_ctx, EVENT_ALL);
6141
6142         /*
6143          * Take the context lock here so that if find_get_context is
6144          * reading child->perf_event_ctxp, we wait until it has
6145          * incremented the context's refcount before we do put_ctx below.
6146          */
6147         raw_spin_lock(&child_ctx->lock);
6148         child->perf_event_ctxp[ctxn] = NULL;
6149         /*
6150          * If this context is a clone; unclone it so it can't get
6151          * swapped to another process while we're removing all
6152          * the events from it.
6153          */
6154         unclone_ctx(child_ctx);
6155         update_context_time(child_ctx);
6156         raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6157
6158         /*
6159          * Report the task dead after unscheduling the events so that we
6160          * won't get any samples after PERF_RECORD_EXIT. We can however still
6161          * get a few PERF_RECORD_READ events.
6162          */
6163         perf_event_task(child, child_ctx, 0);
6164
6165         /*
6166          * We can recurse on the same lock type through:
6167          *
6168          *   __perf_event_exit_task()
6169          *     sync_child_event()
6170          *       fput(parent_event->filp)
6171          *         perf_release()
6172          *           mutex_lock(&ctx->mutex)
6173          *
6174          * But since its the parent context it won't be the same instance.
6175          */
6176         mutex_lock(&child_ctx->mutex);
6177
6178 again:
6179         list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
6180                                  group_entry)
6181                 __perf_event_exit_task(child_event, child_ctx, child);
6182
6183         list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
6184                                  group_entry)
6185                 __perf_event_exit_task(child_event, child_ctx, child);
6186
6187         /*
6188          * If the last event was a group event, it will have appended all
6189          * its siblings to the list, but we obtained 'tmp' before that which
6190          * will still point to the list head terminating the iteration.
6191          */
6192         if (!list_empty(&child_ctx->pinned_groups) ||
6193             !list_empty(&child_ctx->flexible_groups))
6194                 goto again;
6195
6196         mutex_unlock(&child_ctx->mutex);
6197
6198         put_ctx(child_ctx);
6199 }
6200
6201 /*
6202  * When a child task exits, feed back event values to parent events.
6203  */
6204 void perf_event_exit_task(struct task_struct *child)
6205 {
6206         struct perf_event *event, *tmp;
6207         int ctxn;
6208
6209         mutex_lock(&child->perf_event_mutex);
6210         list_for_each_entry_safe(event, tmp, &child->perf_event_list,
6211                                  owner_entry) {
6212                 list_del_init(&event->owner_entry);
6213
6214                 /*
6215                  * Ensure the list deletion is visible before we clear
6216                  * the owner, closes a race against perf_release() where
6217                  * we need to serialize on the owner->perf_event_mutex.
6218                  */
6219                 smp_wmb();
6220                 event->owner = NULL;
6221         }
6222         mutex_unlock(&child->perf_event_mutex);
6223
6224         for_each_task_context_nr(ctxn)
6225                 perf_event_exit_task_context(child, ctxn);
6226 }
6227
6228 static void perf_free_event(struct perf_event *event,
6229                             struct perf_event_context *ctx)
6230 {
6231         struct perf_event *parent = event->parent;
6232
6233         if (WARN_ON_ONCE(!parent))
6234                 return;
6235
6236         mutex_lock(&parent->child_mutex);
6237         list_del_init(&event->child_list);
6238         mutex_unlock(&parent->child_mutex);
6239
6240         fput(parent->filp);
6241
6242         perf_group_detach(event);
6243         list_del_event(event, ctx);
6244         free_event(event);
6245 }
6246
6247 /*
6248  * free an unexposed, unused context as created by inheritance by
6249  * perf_event_init_task below, used by fork() in case of fail.
6250  */
6251 void perf_event_free_task(struct task_struct *task)
6252 {
6253         struct perf_event_context *ctx;
6254         struct perf_event *event, *tmp;
6255         int ctxn;
6256
6257         for_each_task_context_nr(ctxn) {
6258                 ctx = task->perf_event_ctxp[ctxn];
6259                 if (!ctx)
6260                         continue;
6261
6262                 mutex_lock(&ctx->mutex);
6263 again:
6264                 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
6265                                 group_entry)
6266                         perf_free_event(event, ctx);
6267
6268                 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
6269                                 group_entry)
6270                         perf_free_event(event, ctx);
6271
6272                 if (!list_empty(&ctx->pinned_groups) ||
6273                                 !list_empty(&ctx->flexible_groups))
6274                         goto again;
6275
6276                 mutex_unlock(&ctx->mutex);
6277
6278                 put_ctx(ctx);
6279         }
6280 }
6281
6282 void perf_event_delayed_put(struct task_struct *task)
6283 {
6284         int ctxn;
6285
6286         for_each_task_context_nr(ctxn)
6287                 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
6288 }
6289
6290 /*
6291  * inherit a event from parent task to child task:
6292  */
6293 static struct perf_event *
6294 inherit_event(struct perf_event *parent_event,
6295               struct task_struct *parent,
6296               struct perf_event_context *parent_ctx,
6297               struct task_struct *child,
6298               struct perf_event *group_leader,
6299               struct perf_event_context *child_ctx)
6300 {
6301         struct perf_event *child_event;
6302         unsigned long flags;
6303
6304         /*
6305          * Instead of creating recursive hierarchies of events,
6306          * we link inherited events back to the original parent,
6307          * which has a filp for sure, which we use as the reference
6308          * count:
6309          */
6310         if (parent_event->parent)
6311                 parent_event = parent_event->parent;
6312
6313         child_event = perf_event_alloc(&parent_event->attr,
6314                                            parent_event->cpu,
6315                                            child,
6316                                            group_leader, parent_event,
6317                                            NULL);
6318         if (IS_ERR(child_event))
6319                 return child_event;
6320         get_ctx(child_ctx);
6321
6322         /*
6323          * Make the child state follow the state of the parent event,
6324          * not its attr.disabled bit.  We hold the parent's mutex,
6325          * so we won't race with perf_event_{en, dis}able_family.
6326          */
6327         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
6328                 child_event->state = PERF_EVENT_STATE_INACTIVE;
6329         else
6330                 child_event->state = PERF_EVENT_STATE_OFF;
6331
6332         if (parent_event->attr.freq) {
6333                 u64 sample_period = parent_event->hw.sample_period;
6334                 struct hw_perf_event *hwc = &child_event->hw;
6335
6336                 hwc->sample_period = sample_period;
6337                 hwc->last_period   = sample_period;
6338
6339                 local64_set(&hwc->period_left, sample_period);
6340         }
6341
6342         child_event->ctx = child_ctx;
6343         child_event->overflow_handler = parent_event->overflow_handler;
6344
6345         /*
6346          * Precalculate sample_data sizes
6347          */
6348         perf_event__header_size(child_event);
6349         perf_event__id_header_size(child_event);
6350
6351         /*
6352          * Link it up in the child's context:
6353          */
6354         raw_spin_lock_irqsave(&child_ctx->lock, flags);
6355         add_event_to_ctx(child_event, child_ctx);
6356         raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6357
6358         /*
6359          * Get a reference to the parent filp - we will fput it
6360          * when the child event exits. This is safe to do because
6361          * we are in the parent and we know that the filp still
6362          * exists and has a nonzero count:
6363          */
6364         atomic_long_inc(&parent_event->filp->f_count);
6365
6366         /*
6367          * Link this into the parent event's child list
6368          */
6369         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6370         mutex_lock(&parent_event->child_mutex);
6371         list_add_tail(&child_event->child_list, &parent_event->child_list);
6372         mutex_unlock(&parent_event->child_mutex);
6373
6374         return child_event;
6375 }
6376
6377 static int inherit_group(struct perf_event *parent_event,
6378               struct task_struct *parent,
6379               struct perf_event_context *parent_ctx,
6380               struct task_struct *child,
6381               struct perf_event_context *child_ctx)
6382 {
6383         struct perf_event *leader;
6384         struct perf_event *sub;
6385         struct perf_event *child_ctr;
6386
6387         leader = inherit_event(parent_event, parent, parent_ctx,
6388                                  child, NULL, child_ctx);
6389         if (IS_ERR(leader))
6390                 return PTR_ERR(leader);
6391         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
6392                 child_ctr = inherit_event(sub, parent, parent_ctx,
6393                                             child, leader, child_ctx);
6394                 if (IS_ERR(child_ctr))
6395                         return PTR_ERR(child_ctr);
6396         }
6397         return 0;
6398 }
6399
6400 static int
6401 inherit_task_group(struct perf_event *event, struct task_struct *parent,
6402                    struct perf_event_context *parent_ctx,
6403                    struct task_struct *child, int ctxn,
6404                    int *inherited_all)
6405 {
6406         int ret;
6407         struct perf_event_context *child_ctx;
6408
6409         if (!event->attr.inherit) {
6410                 *inherited_all = 0;
6411                 return 0;
6412         }
6413
6414         child_ctx = child->perf_event_ctxp[ctxn];
6415         if (!child_ctx) {
6416                 /*
6417                  * This is executed from the parent task context, so
6418                  * inherit events that have been marked for cloning.
6419                  * First allocate and initialize a context for the
6420                  * child.
6421                  */
6422
6423                 child_ctx = alloc_perf_context(event->pmu, child);
6424                 if (!child_ctx)
6425                         return -ENOMEM;
6426
6427                 child->perf_event_ctxp[ctxn] = child_ctx;
6428         }
6429
6430         ret = inherit_group(event, parent, parent_ctx,
6431                             child, child_ctx);
6432
6433         if (ret)
6434                 *inherited_all = 0;
6435
6436         return ret;
6437 }
6438
6439 /*
6440  * Initialize the perf_event context in task_struct
6441  */
6442 int perf_event_init_context(struct task_struct *child, int ctxn)
6443 {
6444         struct perf_event_context *child_ctx, *parent_ctx;
6445         struct perf_event_context *cloned_ctx;
6446         struct perf_event *event;
6447         struct task_struct *parent = current;
6448         int inherited_all = 1;
6449         unsigned long flags;
6450         int ret = 0;
6451
6452         if (likely(!parent->perf_event_ctxp[ctxn]))
6453                 return 0;
6454
6455         /*
6456          * If the parent's context is a clone, pin it so it won't get
6457          * swapped under us.
6458          */
6459         parent_ctx = perf_pin_task_context(parent, ctxn);
6460
6461         /*
6462          * No need to check if parent_ctx != NULL here; since we saw
6463          * it non-NULL earlier, the only reason for it to become NULL
6464          * is if we exit, and since we're currently in the middle of
6465          * a fork we can't be exiting at the same time.
6466          */
6467
6468         /*
6469          * Lock the parent list. No need to lock the child - not PID
6470          * hashed yet and not running, so nobody can access it.
6471          */
6472         mutex_lock(&parent_ctx->mutex);
6473
6474         /*
6475          * We dont have to disable NMIs - we are only looking at
6476          * the list, not manipulating it:
6477          */
6478         list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
6479                 ret = inherit_task_group(event, parent, parent_ctx,
6480                                          child, ctxn, &inherited_all);
6481                 if (ret)
6482                         break;
6483         }
6484
6485         /*
6486          * We can't hold ctx->lock when iterating the ->flexible_group list due
6487          * to allocations, but we need to prevent rotation because
6488          * rotate_ctx() will change the list from interrupt context.
6489          */
6490         raw_spin_lock_irqsave(&parent_ctx->lock, flags);
6491         parent_ctx->rotate_disable = 1;
6492         raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
6493
6494         list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
6495                 ret = inherit_task_group(event, parent, parent_ctx,
6496                                          child, ctxn, &inherited_all);
6497                 if (ret)
6498                         break;
6499         }
6500
6501         raw_spin_lock_irqsave(&parent_ctx->lock, flags);
6502         parent_ctx->rotate_disable = 0;
6503
6504         child_ctx = child->perf_event_ctxp[ctxn];
6505
6506         if (child_ctx && inherited_all) {
6507                 /*
6508                  * Mark the child context as a clone of the parent
6509                  * context, or of whatever the parent is a clone of.
6510                  *
6511                  * Note that if the parent is a clone, the holding of
6512                  * parent_ctx->lock avoids it from being uncloned.
6513                  */
6514                 cloned_ctx = parent_ctx->parent_ctx;
6515                 if (cloned_ctx) {
6516                         child_ctx->parent_ctx = cloned_ctx;
6517                         child_ctx->parent_gen = parent_ctx->parent_gen;
6518                 } else {
6519                         child_ctx->parent_ctx = parent_ctx;
6520                         child_ctx->parent_gen = parent_ctx->generation;
6521                 }
6522                 get_ctx(child_ctx->parent_ctx);
6523         }
6524
6525         raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
6526         mutex_unlock(&parent_ctx->mutex);
6527
6528         perf_unpin_context(parent_ctx);
6529
6530         return ret;
6531 }
6532
6533 /*
6534  * Initialize the perf_event context in task_struct
6535  */
6536 int perf_event_init_task(struct task_struct *child)
6537 {
6538         int ctxn, ret;
6539
6540         memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
6541         mutex_init(&child->perf_event_mutex);
6542         INIT_LIST_HEAD(&child->perf_event_list);
6543
6544         for_each_task_context_nr(ctxn) {
6545                 ret = perf_event_init_context(child, ctxn);
6546                 if (ret)
6547                         return ret;
6548         }
6549
6550         return 0;
6551 }
6552
6553 static void __init perf_event_init_all_cpus(void)
6554 {
6555         struct swevent_htable *swhash;
6556         int cpu;
6557
6558         for_each_possible_cpu(cpu) {
6559                 swhash = &per_cpu(swevent_htable, cpu);
6560                 mutex_init(&swhash->hlist_mutex);
6561                 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
6562         }
6563 }
6564
6565 static void __cpuinit perf_event_init_cpu(int cpu)
6566 {
6567         struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6568
6569         mutex_lock(&swhash->hlist_mutex);
6570         if (swhash->hlist_refcount > 0) {
6571                 struct swevent_hlist *hlist;
6572
6573                 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
6574                 WARN_ON(!hlist);
6575                 rcu_assign_pointer(swhash->swevent_hlist, hlist);
6576         }
6577         mutex_unlock(&swhash->hlist_mutex);
6578 }
6579
6580 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
6581 static void perf_pmu_rotate_stop(struct pmu *pmu)
6582 {
6583         struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
6584
6585         WARN_ON(!irqs_disabled());
6586
6587         list_del_init(&cpuctx->rotation_list);
6588 }
6589
6590 static void __perf_event_exit_context(void *__info)
6591 {
6592         struct perf_event_context *ctx = __info;
6593         struct perf_event *event, *tmp;
6594
6595         perf_pmu_rotate_stop(ctx->pmu);
6596
6597         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
6598                 __perf_event_remove_from_context(event);
6599         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
6600                 __perf_event_remove_from_context(event);
6601 }
6602
6603 static void perf_event_exit_cpu_context(int cpu)
6604 {
6605         struct perf_event_context *ctx;
6606         struct pmu *pmu;
6607         int idx;
6608
6609         idx = srcu_read_lock(&pmus_srcu);
6610         list_for_each_entry_rcu(pmu, &pmus, entry) {
6611                 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
6612
6613                 mutex_lock(&ctx->mutex);
6614                 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
6615                 mutex_unlock(&ctx->mutex);
6616         }
6617         srcu_read_unlock(&pmus_srcu, idx);
6618 }
6619
6620 static void perf_event_exit_cpu(int cpu)
6621 {
6622         struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6623
6624         mutex_lock(&swhash->hlist_mutex);
6625         swevent_hlist_release(swhash);
6626         mutex_unlock(&swhash->hlist_mutex);
6627
6628         perf_event_exit_cpu_context(cpu);
6629 }
6630 #else
6631 static inline void perf_event_exit_cpu(int cpu) { }
6632 #endif
6633
6634 static int
6635 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
6636 {
6637         int cpu;
6638
6639         for_each_online_cpu(cpu)
6640                 perf_event_exit_cpu(cpu);
6641
6642         return NOTIFY_OK;
6643 }
6644
6645 /*
6646  * Run the perf reboot notifier at the very last possible moment so that
6647  * the generic watchdog code runs as long as possible.
6648  */
6649 static struct notifier_block perf_reboot_notifier = {
6650         .notifier_call = perf_reboot,
6651         .priority = INT_MIN,
6652 };
6653
6654 static int __cpuinit
6655 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
6656 {
6657         unsigned int cpu = (long)hcpu;
6658
6659         switch (action & ~CPU_TASKS_FROZEN) {
6660
6661         case CPU_UP_PREPARE:
6662         case CPU_DOWN_FAILED:
6663                 perf_event_init_cpu(cpu);
6664                 break;
6665
6666         case CPU_UP_CANCELED:
6667         case CPU_DOWN_PREPARE:
6668                 perf_event_exit_cpu(cpu);
6669                 break;
6670
6671         default:
6672                 break;
6673         }
6674
6675         return NOTIFY_OK;
6676 }
6677
6678 void __init perf_event_init(void)
6679 {
6680         int ret;
6681
6682         idr_init(&pmu_idr);
6683
6684         perf_event_init_all_cpus();
6685         init_srcu_struct(&pmus_srcu);
6686         perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
6687         perf_pmu_register(&perf_cpu_clock, NULL, -1);
6688         perf_pmu_register(&perf_task_clock, NULL, -1);
6689         perf_tp_register();
6690         perf_cpu_notifier(perf_cpu_notify);
6691         register_reboot_notifier(&perf_reboot_notifier);
6692
6693         ret = init_hw_breakpoint();
6694         WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
6695 }
6696
6697 static int __init perf_event_sysfs_init(void)
6698 {
6699         struct pmu *pmu;
6700         int ret;
6701
6702         mutex_lock(&pmus_lock);
6703
6704         ret = bus_register(&pmu_bus);
6705         if (ret)
6706                 goto unlock;
6707
6708         list_for_each_entry(pmu, &pmus, entry) {
6709                 if (!pmu->name || pmu->type < 0)
6710                         continue;
6711
6712                 ret = pmu_dev_alloc(pmu);
6713                 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
6714         }
6715         pmu_bus_running = 1;
6716         ret = 0;
6717
6718 unlock:
6719         mutex_unlock(&pmus_lock);
6720
6721         return ret;
6722 }
6723 device_initcall(perf_event_sysfs_init);