Revert "perf: Fix exit() vs PERF_FORMAT_GROUP"
[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/file.h>
17 #include <linux/poll.h>
18 #include <linux/slab.h>
19 #include <linux/hash.h>
20 #include <linux/sysfs.h>
21 #include <linux/dcache.h>
22 #include <linux/percpu.h>
23 #include <linux/ptrace.h>
24 #include <linux/vmstat.h>
25 #include <linux/vmalloc.h>
26 #include <linux/hardirq.h>
27 #include <linux/rculist.h>
28 #include <linux/uaccess.h>
29 #include <linux/syscalls.h>
30 #include <linux/anon_inodes.h>
31 #include <linux/kernel_stat.h>
32 #include <linux/perf_event.h>
33 #include <linux/ftrace_event.h>
34 #include <linux/hw_breakpoint.h>
35
36 #include <asm/irq_regs.h>
37
38 /*
39  * Each CPU has a list of per CPU events:
40  */
41 static DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
42
43 int perf_max_events __read_mostly = 1;
44 static int perf_reserved_percpu __read_mostly;
45 static int perf_overcommit __read_mostly = 1;
46
47 static atomic_t nr_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 /*
53  * perf event paranoia level:
54  *  -1 - not paranoid at all
55  *   0 - disallow raw tracepoint access for unpriv
56  *   1 - disallow cpu events for unpriv
57  *   2 - disallow kernel profiling for unpriv
58  */
59 int sysctl_perf_event_paranoid __read_mostly = 1;
60
61 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
62
63 /*
64  * max perf event sample rate
65  */
66 int sysctl_perf_event_sample_rate __read_mostly = 100000;
67
68 static atomic64_t perf_event_id;
69
70 /*
71  * Lock for (sysadmin-configurable) event reservations:
72  */
73 static DEFINE_SPINLOCK(perf_resource_lock);
74
75 /*
76  * Architecture provided APIs - weak aliases:
77  */
78 extern __weak const struct pmu *hw_perf_event_init(struct perf_event *event)
79 {
80         return NULL;
81 }
82
83 void __weak hw_perf_disable(void)               { barrier(); }
84 void __weak hw_perf_enable(void)                { barrier(); }
85
86 void __weak perf_event_print_debug(void)        { }
87
88 static DEFINE_PER_CPU(int, perf_disable_count);
89
90 void perf_disable(void)
91 {
92         if (!__get_cpu_var(perf_disable_count)++)
93                 hw_perf_disable();
94 }
95
96 void perf_enable(void)
97 {
98         if (!--__get_cpu_var(perf_disable_count))
99                 hw_perf_enable();
100 }
101
102 static void get_ctx(struct perf_event_context *ctx)
103 {
104         WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
105 }
106
107 static void free_ctx(struct rcu_head *head)
108 {
109         struct perf_event_context *ctx;
110
111         ctx = container_of(head, struct perf_event_context, rcu_head);
112         kfree(ctx);
113 }
114
115 static void put_ctx(struct perf_event_context *ctx)
116 {
117         if (atomic_dec_and_test(&ctx->refcount)) {
118                 if (ctx->parent_ctx)
119                         put_ctx(ctx->parent_ctx);
120                 if (ctx->task)
121                         put_task_struct(ctx->task);
122                 call_rcu(&ctx->rcu_head, free_ctx);
123         }
124 }
125
126 static void unclone_ctx(struct perf_event_context *ctx)
127 {
128         if (ctx->parent_ctx) {
129                 put_ctx(ctx->parent_ctx);
130                 ctx->parent_ctx = NULL;
131         }
132 }
133
134 /*
135  * If we inherit events we want to return the parent event id
136  * to userspace.
137  */
138 static u64 primary_event_id(struct perf_event *event)
139 {
140         u64 id = event->id;
141
142         if (event->parent)
143                 id = event->parent->id;
144
145         return id;
146 }
147
148 /*
149  * Get the perf_event_context for a task and lock it.
150  * This has to cope with with the fact that until it is locked,
151  * the context could get moved to another task.
152  */
153 static struct perf_event_context *
154 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
155 {
156         struct perf_event_context *ctx;
157
158         rcu_read_lock();
159  retry:
160         ctx = rcu_dereference(task->perf_event_ctxp);
161         if (ctx) {
162                 /*
163                  * If this context is a clone of another, it might
164                  * get swapped for another underneath us by
165                  * perf_event_task_sched_out, though the
166                  * rcu_read_lock() protects us from any context
167                  * getting freed.  Lock the context and check if it
168                  * got swapped before we could get the lock, and retry
169                  * if so.  If we locked the right context, then it
170                  * can't get swapped on us any more.
171                  */
172                 raw_spin_lock_irqsave(&ctx->lock, *flags);
173                 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
174                         raw_spin_unlock_irqrestore(&ctx->lock, *flags);
175                         goto retry;
176                 }
177
178                 if (!atomic_inc_not_zero(&ctx->refcount)) {
179                         raw_spin_unlock_irqrestore(&ctx->lock, *flags);
180                         ctx = NULL;
181                 }
182         }
183         rcu_read_unlock();
184         return ctx;
185 }
186
187 /*
188  * Get the context for a task and increment its pin_count so it
189  * can't get swapped to another task.  This also increments its
190  * reference count so that the context can't get freed.
191  */
192 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
193 {
194         struct perf_event_context *ctx;
195         unsigned long flags;
196
197         ctx = perf_lock_task_context(task, &flags);
198         if (ctx) {
199                 ++ctx->pin_count;
200                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
201         }
202         return ctx;
203 }
204
205 static void perf_unpin_context(struct perf_event_context *ctx)
206 {
207         unsigned long flags;
208
209         raw_spin_lock_irqsave(&ctx->lock, flags);
210         --ctx->pin_count;
211         raw_spin_unlock_irqrestore(&ctx->lock, flags);
212         put_ctx(ctx);
213 }
214
215 static inline u64 perf_clock(void)
216 {
217         return cpu_clock(raw_smp_processor_id());
218 }
219
220 /*
221  * Update the record of the current time in a context.
222  */
223 static void update_context_time(struct perf_event_context *ctx)
224 {
225         u64 now = perf_clock();
226
227         ctx->time += now - ctx->timestamp;
228         ctx->timestamp = now;
229 }
230
231 /*
232  * Update the total_time_enabled and total_time_running fields for a event.
233  */
234 static void update_event_times(struct perf_event *event)
235 {
236         struct perf_event_context *ctx = event->ctx;
237         u64 run_end;
238
239         if (event->state < PERF_EVENT_STATE_INACTIVE ||
240             event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
241                 return;
242
243         if (ctx->is_active)
244                 run_end = ctx->time;
245         else
246                 run_end = event->tstamp_stopped;
247
248         event->total_time_enabled = run_end - event->tstamp_enabled;
249
250         if (event->state == PERF_EVENT_STATE_INACTIVE)
251                 run_end = event->tstamp_stopped;
252         else
253                 run_end = ctx->time;
254
255         event->total_time_running = run_end - event->tstamp_running;
256 }
257
258 static struct list_head *
259 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
260 {
261         if (event->attr.pinned)
262                 return &ctx->pinned_groups;
263         else
264                 return &ctx->flexible_groups;
265 }
266
267 /*
268  * Add a event from the lists for its context.
269  * Must be called with ctx->mutex and ctx->lock held.
270  */
271 static void
272 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
273 {
274         struct perf_event *group_leader = event->group_leader;
275
276         /*
277          * Depending on whether it is a standalone or sibling event,
278          * add it straight to the context's event list, or to the group
279          * leader's sibling list:
280          */
281         if (group_leader == event) {
282                 struct list_head *list;
283
284                 if (is_software_event(event))
285                         event->group_flags |= PERF_GROUP_SOFTWARE;
286
287                 list = ctx_group_list(event, ctx);
288                 list_add_tail(&event->group_entry, list);
289         } else {
290                 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
291                     !is_software_event(event))
292                         group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
293
294                 list_add_tail(&event->group_entry, &group_leader->sibling_list);
295                 group_leader->nr_siblings++;
296         }
297
298         list_add_rcu(&event->event_entry, &ctx->event_list);
299         ctx->nr_events++;
300         if (event->attr.inherit_stat)
301                 ctx->nr_stat++;
302 }
303
304 /*
305  * Remove a event from the lists for its context.
306  * Must be called with ctx->mutex and ctx->lock held.
307  */
308 static void
309 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
310 {
311         struct perf_event *sibling, *tmp;
312
313         if (list_empty(&event->group_entry))
314                 return;
315         ctx->nr_events--;
316         if (event->attr.inherit_stat)
317                 ctx->nr_stat--;
318
319         list_del_init(&event->group_entry);
320         list_del_rcu(&event->event_entry);
321
322         if (event->group_leader != event)
323                 event->group_leader->nr_siblings--;
324
325         update_event_times(event);
326
327         /*
328          * If event was in error state, then keep it
329          * that way, otherwise bogus counts will be
330          * returned on read(). The only way to get out
331          * of error state is by explicit re-enabling
332          * of the event
333          */
334         if (event->state > PERF_EVENT_STATE_OFF)
335                 event->state = PERF_EVENT_STATE_OFF;
336
337         /*
338          * If this was a group event with sibling events then
339          * upgrade the siblings to singleton events by adding them
340          * to the context list directly:
341          */
342         list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
343                 struct list_head *list;
344
345                 list = ctx_group_list(event, ctx);
346                 list_move_tail(&sibling->group_entry, list);
347                 sibling->group_leader = sibling;
348
349                 /* Inherit group flags from the previous leader */
350                 sibling->group_flags = event->group_flags;
351         }
352 }
353
354 static void
355 event_sched_out(struct perf_event *event,
356                   struct perf_cpu_context *cpuctx,
357                   struct perf_event_context *ctx)
358 {
359         if (event->state != PERF_EVENT_STATE_ACTIVE)
360                 return;
361
362         event->state = PERF_EVENT_STATE_INACTIVE;
363         if (event->pending_disable) {
364                 event->pending_disable = 0;
365                 event->state = PERF_EVENT_STATE_OFF;
366         }
367         event->tstamp_stopped = ctx->time;
368         event->pmu->disable(event);
369         event->oncpu = -1;
370
371         if (!is_software_event(event))
372                 cpuctx->active_oncpu--;
373         ctx->nr_active--;
374         if (event->attr.exclusive || !cpuctx->active_oncpu)
375                 cpuctx->exclusive = 0;
376 }
377
378 static void
379 group_sched_out(struct perf_event *group_event,
380                 struct perf_cpu_context *cpuctx,
381                 struct perf_event_context *ctx)
382 {
383         struct perf_event *event;
384
385         if (group_event->state != PERF_EVENT_STATE_ACTIVE)
386                 return;
387
388         event_sched_out(group_event, cpuctx, ctx);
389
390         /*
391          * Schedule out siblings (if any):
392          */
393         list_for_each_entry(event, &group_event->sibling_list, group_entry)
394                 event_sched_out(event, cpuctx, ctx);
395
396         if (group_event->attr.exclusive)
397                 cpuctx->exclusive = 0;
398 }
399
400 /*
401  * Cross CPU call to remove a performance event
402  *
403  * We disable the event on the hardware level first. After that we
404  * remove it from the context list.
405  */
406 static void __perf_event_remove_from_context(void *info)
407 {
408         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
409         struct perf_event *event = info;
410         struct perf_event_context *ctx = event->ctx;
411
412         /*
413          * If this is a task context, we need to check whether it is
414          * the current task context of this cpu. If not it has been
415          * scheduled out before the smp call arrived.
416          */
417         if (ctx->task && cpuctx->task_ctx != ctx)
418                 return;
419
420         raw_spin_lock(&ctx->lock);
421         /*
422          * Protect the list operation against NMI by disabling the
423          * events on a global level.
424          */
425         perf_disable();
426
427         event_sched_out(event, cpuctx, ctx);
428
429         list_del_event(event, ctx);
430
431         if (!ctx->task) {
432                 /*
433                  * Allow more per task events with respect to the
434                  * reservation:
435                  */
436                 cpuctx->max_pertask =
437                         min(perf_max_events - ctx->nr_events,
438                             perf_max_events - perf_reserved_percpu);
439         }
440
441         perf_enable();
442         raw_spin_unlock(&ctx->lock);
443 }
444
445
446 /*
447  * Remove the event from a task's (or a CPU's) list of events.
448  *
449  * Must be called with ctx->mutex held.
450  *
451  * CPU events are removed with a smp call. For task events we only
452  * call when the task is on a CPU.
453  *
454  * If event->ctx is a cloned context, callers must make sure that
455  * every task struct that event->ctx->task could possibly point to
456  * remains valid.  This is OK when called from perf_release since
457  * that only calls us on the top-level context, which can't be a clone.
458  * When called from perf_event_exit_task, it's OK because the
459  * context has been detached from its task.
460  */
461 static void perf_event_remove_from_context(struct perf_event *event)
462 {
463         struct perf_event_context *ctx = event->ctx;
464         struct task_struct *task = ctx->task;
465
466         if (!task) {
467                 /*
468                  * Per cpu events are removed via an smp call and
469                  * the removal is always successful.
470                  */
471                 smp_call_function_single(event->cpu,
472                                          __perf_event_remove_from_context,
473                                          event, 1);
474                 return;
475         }
476
477 retry:
478         task_oncpu_function_call(task, __perf_event_remove_from_context,
479                                  event);
480
481         raw_spin_lock_irq(&ctx->lock);
482         /*
483          * If the context is active we need to retry the smp call.
484          */
485         if (ctx->nr_active && !list_empty(&event->group_entry)) {
486                 raw_spin_unlock_irq(&ctx->lock);
487                 goto retry;
488         }
489
490         /*
491          * The lock prevents that this context is scheduled in so we
492          * can remove the event safely, if the call above did not
493          * succeed.
494          */
495         if (!list_empty(&event->group_entry))
496                 list_del_event(event, ctx);
497         raw_spin_unlock_irq(&ctx->lock);
498 }
499
500 /*
501  * Update total_time_enabled and total_time_running for all events in a group.
502  */
503 static void update_group_times(struct perf_event *leader)
504 {
505         struct perf_event *event;
506
507         update_event_times(leader);
508         list_for_each_entry(event, &leader->sibling_list, group_entry)
509                 update_event_times(event);
510 }
511
512 /*
513  * Cross CPU call to disable a performance event
514  */
515 static void __perf_event_disable(void *info)
516 {
517         struct perf_event *event = info;
518         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
519         struct perf_event_context *ctx = event->ctx;
520
521         /*
522          * If this is a per-task event, need to check whether this
523          * event's task is the current task on this cpu.
524          */
525         if (ctx->task && cpuctx->task_ctx != ctx)
526                 return;
527
528         raw_spin_lock(&ctx->lock);
529
530         /*
531          * If the event is on, turn it off.
532          * If it is in error state, leave it in error state.
533          */
534         if (event->state >= PERF_EVENT_STATE_INACTIVE) {
535                 update_context_time(ctx);
536                 update_group_times(event);
537                 if (event == event->group_leader)
538                         group_sched_out(event, cpuctx, ctx);
539                 else
540                         event_sched_out(event, cpuctx, ctx);
541                 event->state = PERF_EVENT_STATE_OFF;
542         }
543
544         raw_spin_unlock(&ctx->lock);
545 }
546
547 /*
548  * Disable a event.
549  *
550  * If event->ctx is a cloned context, callers must make sure that
551  * every task struct that event->ctx->task could possibly point to
552  * remains valid.  This condition is satisifed when called through
553  * perf_event_for_each_child or perf_event_for_each because they
554  * hold the top-level event's child_mutex, so any descendant that
555  * goes to exit will block in sync_child_event.
556  * When called from perf_pending_event it's OK because event->ctx
557  * is the current context on this CPU and preemption is disabled,
558  * hence we can't get into perf_event_task_sched_out for this context.
559  */
560 void perf_event_disable(struct perf_event *event)
561 {
562         struct perf_event_context *ctx = event->ctx;
563         struct task_struct *task = ctx->task;
564
565         if (!task) {
566                 /*
567                  * Disable the event on the cpu that it's on
568                  */
569                 smp_call_function_single(event->cpu, __perf_event_disable,
570                                          event, 1);
571                 return;
572         }
573
574  retry:
575         task_oncpu_function_call(task, __perf_event_disable, event);
576
577         raw_spin_lock_irq(&ctx->lock);
578         /*
579          * If the event is still active, we need to retry the cross-call.
580          */
581         if (event->state == PERF_EVENT_STATE_ACTIVE) {
582                 raw_spin_unlock_irq(&ctx->lock);
583                 goto retry;
584         }
585
586         /*
587          * Since we have the lock this context can't be scheduled
588          * in, so we can change the state safely.
589          */
590         if (event->state == PERF_EVENT_STATE_INACTIVE) {
591                 update_group_times(event);
592                 event->state = PERF_EVENT_STATE_OFF;
593         }
594
595         raw_spin_unlock_irq(&ctx->lock);
596 }
597
598 static int
599 event_sched_in(struct perf_event *event,
600                  struct perf_cpu_context *cpuctx,
601                  struct perf_event_context *ctx)
602 {
603         if (event->state <= PERF_EVENT_STATE_OFF)
604                 return 0;
605
606         event->state = PERF_EVENT_STATE_ACTIVE;
607         event->oncpu = smp_processor_id();
608         /*
609          * The new state must be visible before we turn it on in the hardware:
610          */
611         smp_wmb();
612
613         if (event->pmu->enable(event)) {
614                 event->state = PERF_EVENT_STATE_INACTIVE;
615                 event->oncpu = -1;
616                 return -EAGAIN;
617         }
618
619         event->tstamp_running += ctx->time - event->tstamp_stopped;
620
621         if (!is_software_event(event))
622                 cpuctx->active_oncpu++;
623         ctx->nr_active++;
624
625         if (event->attr.exclusive)
626                 cpuctx->exclusive = 1;
627
628         return 0;
629 }
630
631 static int
632 group_sched_in(struct perf_event *group_event,
633                struct perf_cpu_context *cpuctx,
634                struct perf_event_context *ctx)
635 {
636         struct perf_event *event, *partial_group = NULL;
637         const struct pmu *pmu = group_event->pmu;
638         bool txn = false;
639         int ret;
640
641         if (group_event->state == PERF_EVENT_STATE_OFF)
642                 return 0;
643
644         /* Check if group transaction availabe */
645         if (pmu->start_txn)
646                 txn = true;
647
648         if (txn)
649                 pmu->start_txn(pmu);
650
651         if (event_sched_in(group_event, cpuctx, ctx))
652                 return -EAGAIN;
653
654         /*
655          * Schedule in siblings as one group (if any):
656          */
657         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
658                 if (event_sched_in(event, cpuctx, ctx)) {
659                         partial_group = event;
660                         goto group_error;
661                 }
662         }
663
664         if (!txn)
665                 return 0;
666
667         ret = pmu->commit_txn(pmu);
668         if (!ret) {
669                 pmu->cancel_txn(pmu);
670                 return 0;
671         }
672
673 group_error:
674         if (txn)
675                 pmu->cancel_txn(pmu);
676
677         /*
678          * Groups can be scheduled in as one unit only, so undo any
679          * partial group before returning:
680          */
681         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
682                 if (event == partial_group)
683                         break;
684                 event_sched_out(event, cpuctx, ctx);
685         }
686         event_sched_out(group_event, cpuctx, ctx);
687
688         return -EAGAIN;
689 }
690
691 /*
692  * Work out whether we can put this event group on the CPU now.
693  */
694 static int group_can_go_on(struct perf_event *event,
695                            struct perf_cpu_context *cpuctx,
696                            int can_add_hw)
697 {
698         /*
699          * Groups consisting entirely of software events can always go on.
700          */
701         if (event->group_flags & PERF_GROUP_SOFTWARE)
702                 return 1;
703         /*
704          * If an exclusive group is already on, no other hardware
705          * events can go on.
706          */
707         if (cpuctx->exclusive)
708                 return 0;
709         /*
710          * If this group is exclusive and there are already
711          * events on the CPU, it can't go on.
712          */
713         if (event->attr.exclusive && cpuctx->active_oncpu)
714                 return 0;
715         /*
716          * Otherwise, try to add it if all previous groups were able
717          * to go on.
718          */
719         return can_add_hw;
720 }
721
722 static void add_event_to_ctx(struct perf_event *event,
723                                struct perf_event_context *ctx)
724 {
725         list_add_event(event, ctx);
726         event->tstamp_enabled = ctx->time;
727         event->tstamp_running = ctx->time;
728         event->tstamp_stopped = ctx->time;
729 }
730
731 /*
732  * Cross CPU call to install and enable a performance event
733  *
734  * Must be called with ctx->mutex held
735  */
736 static void __perf_install_in_context(void *info)
737 {
738         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
739         struct perf_event *event = info;
740         struct perf_event_context *ctx = event->ctx;
741         struct perf_event *leader = event->group_leader;
742         int err;
743
744         /*
745          * If this is a task context, we need to check whether it is
746          * the current task context of this cpu. If not it has been
747          * scheduled out before the smp call arrived.
748          * Or possibly this is the right context but it isn't
749          * on this cpu because it had no events.
750          */
751         if (ctx->task && cpuctx->task_ctx != ctx) {
752                 if (cpuctx->task_ctx || ctx->task != current)
753                         return;
754                 cpuctx->task_ctx = ctx;
755         }
756
757         raw_spin_lock(&ctx->lock);
758         ctx->is_active = 1;
759         update_context_time(ctx);
760
761         /*
762          * Protect the list operation against NMI by disabling the
763          * events on a global level. NOP for non NMI based events.
764          */
765         perf_disable();
766
767         add_event_to_ctx(event, ctx);
768
769         if (event->cpu != -1 && event->cpu != smp_processor_id())
770                 goto unlock;
771
772         /*
773          * Don't put the event on if it is disabled or if
774          * it is in a group and the group isn't on.
775          */
776         if (event->state != PERF_EVENT_STATE_INACTIVE ||
777             (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
778                 goto unlock;
779
780         /*
781          * An exclusive event can't go on if there are already active
782          * hardware events, and no hardware event can go on if there
783          * is already an exclusive event on.
784          */
785         if (!group_can_go_on(event, cpuctx, 1))
786                 err = -EEXIST;
787         else
788                 err = event_sched_in(event, cpuctx, ctx);
789
790         if (err) {
791                 /*
792                  * This event couldn't go on.  If it is in a group
793                  * then we have to pull the whole group off.
794                  * If the event group is pinned then put it in error state.
795                  */
796                 if (leader != event)
797                         group_sched_out(leader, cpuctx, ctx);
798                 if (leader->attr.pinned) {
799                         update_group_times(leader);
800                         leader->state = PERF_EVENT_STATE_ERROR;
801                 }
802         }
803
804         if (!err && !ctx->task && cpuctx->max_pertask)
805                 cpuctx->max_pertask--;
806
807  unlock:
808         perf_enable();
809
810         raw_spin_unlock(&ctx->lock);
811 }
812
813 /*
814  * Attach a performance event to a context
815  *
816  * First we add the event to the list with the hardware enable bit
817  * in event->hw_config cleared.
818  *
819  * If the event is attached to a task which is on a CPU we use a smp
820  * call to enable it in the task context. The task might have been
821  * scheduled away, but we check this in the smp call again.
822  *
823  * Must be called with ctx->mutex held.
824  */
825 static void
826 perf_install_in_context(struct perf_event_context *ctx,
827                         struct perf_event *event,
828                         int cpu)
829 {
830         struct task_struct *task = ctx->task;
831
832         if (!task) {
833                 /*
834                  * Per cpu events are installed via an smp call and
835                  * the install is always successful.
836                  */
837                 smp_call_function_single(cpu, __perf_install_in_context,
838                                          event, 1);
839                 return;
840         }
841
842 retry:
843         task_oncpu_function_call(task, __perf_install_in_context,
844                                  event);
845
846         raw_spin_lock_irq(&ctx->lock);
847         /*
848          * we need to retry the smp call.
849          */
850         if (ctx->is_active && list_empty(&event->group_entry)) {
851                 raw_spin_unlock_irq(&ctx->lock);
852                 goto retry;
853         }
854
855         /*
856          * The lock prevents that this context is scheduled in so we
857          * can add the event safely, if it the call above did not
858          * succeed.
859          */
860         if (list_empty(&event->group_entry))
861                 add_event_to_ctx(event, ctx);
862         raw_spin_unlock_irq(&ctx->lock);
863 }
864
865 /*
866  * Put a event into inactive state and update time fields.
867  * Enabling the leader of a group effectively enables all
868  * the group members that aren't explicitly disabled, so we
869  * have to update their ->tstamp_enabled also.
870  * Note: this works for group members as well as group leaders
871  * since the non-leader members' sibling_lists will be empty.
872  */
873 static void __perf_event_mark_enabled(struct perf_event *event,
874                                         struct perf_event_context *ctx)
875 {
876         struct perf_event *sub;
877
878         event->state = PERF_EVENT_STATE_INACTIVE;
879         event->tstamp_enabled = ctx->time - event->total_time_enabled;
880         list_for_each_entry(sub, &event->sibling_list, group_entry)
881                 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
882                         sub->tstamp_enabled =
883                                 ctx->time - sub->total_time_enabled;
884 }
885
886 /*
887  * Cross CPU call to enable a performance event
888  */
889 static void __perf_event_enable(void *info)
890 {
891         struct perf_event *event = info;
892         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
893         struct perf_event_context *ctx = event->ctx;
894         struct perf_event *leader = event->group_leader;
895         int err;
896
897         /*
898          * If this is a per-task event, need to check whether this
899          * event's task is the current task on this cpu.
900          */
901         if (ctx->task && cpuctx->task_ctx != ctx) {
902                 if (cpuctx->task_ctx || ctx->task != current)
903                         return;
904                 cpuctx->task_ctx = ctx;
905         }
906
907         raw_spin_lock(&ctx->lock);
908         ctx->is_active = 1;
909         update_context_time(ctx);
910
911         if (event->state >= PERF_EVENT_STATE_INACTIVE)
912                 goto unlock;
913         __perf_event_mark_enabled(event, ctx);
914
915         if (event->cpu != -1 && event->cpu != smp_processor_id())
916                 goto unlock;
917
918         /*
919          * If the event is in a group and isn't the group leader,
920          * then don't put it on unless the group is on.
921          */
922         if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
923                 goto unlock;
924
925         if (!group_can_go_on(event, cpuctx, 1)) {
926                 err = -EEXIST;
927         } else {
928                 perf_disable();
929                 if (event == leader)
930                         err = group_sched_in(event, cpuctx, ctx);
931                 else
932                         err = event_sched_in(event, cpuctx, ctx);
933                 perf_enable();
934         }
935
936         if (err) {
937                 /*
938                  * If this event can't go on and it's part of a
939                  * group, then the whole group has to come off.
940                  */
941                 if (leader != event)
942                         group_sched_out(leader, cpuctx, ctx);
943                 if (leader->attr.pinned) {
944                         update_group_times(leader);
945                         leader->state = PERF_EVENT_STATE_ERROR;
946                 }
947         }
948
949  unlock:
950         raw_spin_unlock(&ctx->lock);
951 }
952
953 /*
954  * Enable a event.
955  *
956  * If event->ctx is a cloned context, callers must make sure that
957  * every task struct that event->ctx->task could possibly point to
958  * remains valid.  This condition is satisfied when called through
959  * perf_event_for_each_child or perf_event_for_each as described
960  * for perf_event_disable.
961  */
962 void perf_event_enable(struct perf_event *event)
963 {
964         struct perf_event_context *ctx = event->ctx;
965         struct task_struct *task = ctx->task;
966
967         if (!task) {
968                 /*
969                  * Enable the event on the cpu that it's on
970                  */
971                 smp_call_function_single(event->cpu, __perf_event_enable,
972                                          event, 1);
973                 return;
974         }
975
976         raw_spin_lock_irq(&ctx->lock);
977         if (event->state >= PERF_EVENT_STATE_INACTIVE)
978                 goto out;
979
980         /*
981          * If the event is in error state, clear that first.
982          * That way, if we see the event in error state below, we
983          * know that it has gone back into error state, as distinct
984          * from the task having been scheduled away before the
985          * cross-call arrived.
986          */
987         if (event->state == PERF_EVENT_STATE_ERROR)
988                 event->state = PERF_EVENT_STATE_OFF;
989
990  retry:
991         raw_spin_unlock_irq(&ctx->lock);
992         task_oncpu_function_call(task, __perf_event_enable, event);
993
994         raw_spin_lock_irq(&ctx->lock);
995
996         /*
997          * If the context is active and the event is still off,
998          * we need to retry the cross-call.
999          */
1000         if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1001                 goto retry;
1002
1003         /*
1004          * Since we have the lock this context can't be scheduled
1005          * in, so we can change the state safely.
1006          */
1007         if (event->state == PERF_EVENT_STATE_OFF)
1008                 __perf_event_mark_enabled(event, ctx);
1009
1010  out:
1011         raw_spin_unlock_irq(&ctx->lock);
1012 }
1013
1014 static int perf_event_refresh(struct perf_event *event, int refresh)
1015 {
1016         /*
1017          * not supported on inherited events
1018          */
1019         if (event->attr.inherit)
1020                 return -EINVAL;
1021
1022         atomic_add(refresh, &event->event_limit);
1023         perf_event_enable(event);
1024
1025         return 0;
1026 }
1027
1028 enum event_type_t {
1029         EVENT_FLEXIBLE = 0x1,
1030         EVENT_PINNED = 0x2,
1031         EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1032 };
1033
1034 static void ctx_sched_out(struct perf_event_context *ctx,
1035                           struct perf_cpu_context *cpuctx,
1036                           enum event_type_t event_type)
1037 {
1038         struct perf_event *event;
1039
1040         raw_spin_lock(&ctx->lock);
1041         ctx->is_active = 0;
1042         if (likely(!ctx->nr_events))
1043                 goto out;
1044         update_context_time(ctx);
1045
1046         perf_disable();
1047         if (!ctx->nr_active)
1048                 goto out_enable;
1049
1050         if (event_type & EVENT_PINNED)
1051                 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1052                         group_sched_out(event, cpuctx, ctx);
1053
1054         if (event_type & EVENT_FLEXIBLE)
1055                 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1056                         group_sched_out(event, cpuctx, ctx);
1057
1058  out_enable:
1059         perf_enable();
1060  out:
1061         raw_spin_unlock(&ctx->lock);
1062 }
1063
1064 /*
1065  * Test whether two contexts are equivalent, i.e. whether they
1066  * have both been cloned from the same version of the same context
1067  * and they both have the same number of enabled events.
1068  * If the number of enabled events is the same, then the set
1069  * of enabled events should be the same, because these are both
1070  * inherited contexts, therefore we can't access individual events
1071  * in them directly with an fd; we can only enable/disable all
1072  * events via prctl, or enable/disable all events in a family
1073  * via ioctl, which will have the same effect on both contexts.
1074  */
1075 static int context_equiv(struct perf_event_context *ctx1,
1076                          struct perf_event_context *ctx2)
1077 {
1078         return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1079                 && ctx1->parent_gen == ctx2->parent_gen
1080                 && !ctx1->pin_count && !ctx2->pin_count;
1081 }
1082
1083 static void __perf_event_sync_stat(struct perf_event *event,
1084                                      struct perf_event *next_event)
1085 {
1086         u64 value;
1087
1088         if (!event->attr.inherit_stat)
1089                 return;
1090
1091         /*
1092          * Update the event value, we cannot use perf_event_read()
1093          * because we're in the middle of a context switch and have IRQs
1094          * disabled, which upsets smp_call_function_single(), however
1095          * we know the event must be on the current CPU, therefore we
1096          * don't need to use it.
1097          */
1098         switch (event->state) {
1099         case PERF_EVENT_STATE_ACTIVE:
1100                 event->pmu->read(event);
1101                 /* fall-through */
1102
1103         case PERF_EVENT_STATE_INACTIVE:
1104                 update_event_times(event);
1105                 break;
1106
1107         default:
1108                 break;
1109         }
1110
1111         /*
1112          * In order to keep per-task stats reliable we need to flip the event
1113          * values when we flip the contexts.
1114          */
1115         value = atomic64_read(&next_event->count);
1116         value = atomic64_xchg(&event->count, value);
1117         atomic64_set(&next_event->count, value);
1118
1119         swap(event->total_time_enabled, next_event->total_time_enabled);
1120         swap(event->total_time_running, next_event->total_time_running);
1121
1122         /*
1123          * Since we swizzled the values, update the user visible data too.
1124          */
1125         perf_event_update_userpage(event);
1126         perf_event_update_userpage(next_event);
1127 }
1128
1129 #define list_next_entry(pos, member) \
1130         list_entry(pos->member.next, typeof(*pos), member)
1131
1132 static void perf_event_sync_stat(struct perf_event_context *ctx,
1133                                    struct perf_event_context *next_ctx)
1134 {
1135         struct perf_event *event, *next_event;
1136
1137         if (!ctx->nr_stat)
1138                 return;
1139
1140         update_context_time(ctx);
1141
1142         event = list_first_entry(&ctx->event_list,
1143                                    struct perf_event, event_entry);
1144
1145         next_event = list_first_entry(&next_ctx->event_list,
1146                                         struct perf_event, event_entry);
1147
1148         while (&event->event_entry != &ctx->event_list &&
1149                &next_event->event_entry != &next_ctx->event_list) {
1150
1151                 __perf_event_sync_stat(event, next_event);
1152
1153                 event = list_next_entry(event, event_entry);
1154                 next_event = list_next_entry(next_event, event_entry);
1155         }
1156 }
1157
1158 /*
1159  * Called from scheduler to remove the events of the current task,
1160  * with interrupts disabled.
1161  *
1162  * We stop each event and update the event value in event->count.
1163  *
1164  * This does not protect us against NMI, but disable()
1165  * sets the disabled bit in the control field of event _before_
1166  * accessing the event control register. If a NMI hits, then it will
1167  * not restart the event.
1168  */
1169 void perf_event_task_sched_out(struct task_struct *task,
1170                                  struct task_struct *next)
1171 {
1172         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1173         struct perf_event_context *ctx = task->perf_event_ctxp;
1174         struct perf_event_context *next_ctx;
1175         struct perf_event_context *parent;
1176         int do_switch = 1;
1177
1178         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0);
1179
1180         if (likely(!ctx || !cpuctx->task_ctx))
1181                 return;
1182
1183         rcu_read_lock();
1184         parent = rcu_dereference(ctx->parent_ctx);
1185         next_ctx = next->perf_event_ctxp;
1186         if (parent && next_ctx &&
1187             rcu_dereference(next_ctx->parent_ctx) == parent) {
1188                 /*
1189                  * Looks like the two contexts are clones, so we might be
1190                  * able to optimize the context switch.  We lock both
1191                  * contexts and check that they are clones under the
1192                  * lock (including re-checking that neither has been
1193                  * uncloned in the meantime).  It doesn't matter which
1194                  * order we take the locks because no other cpu could
1195                  * be trying to lock both of these tasks.
1196                  */
1197                 raw_spin_lock(&ctx->lock);
1198                 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1199                 if (context_equiv(ctx, next_ctx)) {
1200                         /*
1201                          * XXX do we need a memory barrier of sorts
1202                          * wrt to rcu_dereference() of perf_event_ctxp
1203                          */
1204                         task->perf_event_ctxp = next_ctx;
1205                         next->perf_event_ctxp = ctx;
1206                         ctx->task = next;
1207                         next_ctx->task = task;
1208                         do_switch = 0;
1209
1210                         perf_event_sync_stat(ctx, next_ctx);
1211                 }
1212                 raw_spin_unlock(&next_ctx->lock);
1213                 raw_spin_unlock(&ctx->lock);
1214         }
1215         rcu_read_unlock();
1216
1217         if (do_switch) {
1218                 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1219                 cpuctx->task_ctx = NULL;
1220         }
1221 }
1222
1223 static void task_ctx_sched_out(struct perf_event_context *ctx,
1224                                enum event_type_t event_type)
1225 {
1226         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1227
1228         if (!cpuctx->task_ctx)
1229                 return;
1230
1231         if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1232                 return;
1233
1234         ctx_sched_out(ctx, cpuctx, event_type);
1235         cpuctx->task_ctx = NULL;
1236 }
1237
1238 /*
1239  * Called with IRQs disabled
1240  */
1241 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1242 {
1243         task_ctx_sched_out(ctx, EVENT_ALL);
1244 }
1245
1246 /*
1247  * Called with IRQs disabled
1248  */
1249 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1250                               enum event_type_t event_type)
1251 {
1252         ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1253 }
1254
1255 static void
1256 ctx_pinned_sched_in(struct perf_event_context *ctx,
1257                     struct perf_cpu_context *cpuctx)
1258 {
1259         struct perf_event *event;
1260
1261         list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1262                 if (event->state <= PERF_EVENT_STATE_OFF)
1263                         continue;
1264                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1265                         continue;
1266
1267                 if (group_can_go_on(event, cpuctx, 1))
1268                         group_sched_in(event, cpuctx, ctx);
1269
1270                 /*
1271                  * If this pinned group hasn't been scheduled,
1272                  * put it in error state.
1273                  */
1274                 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1275                         update_group_times(event);
1276                         event->state = PERF_EVENT_STATE_ERROR;
1277                 }
1278         }
1279 }
1280
1281 static void
1282 ctx_flexible_sched_in(struct perf_event_context *ctx,
1283                       struct perf_cpu_context *cpuctx)
1284 {
1285         struct perf_event *event;
1286         int can_add_hw = 1;
1287
1288         list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1289                 /* Ignore events in OFF or ERROR state */
1290                 if (event->state <= PERF_EVENT_STATE_OFF)
1291                         continue;
1292                 /*
1293                  * Listen to the 'cpu' scheduling filter constraint
1294                  * of events:
1295                  */
1296                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1297                         continue;
1298
1299                 if (group_can_go_on(event, cpuctx, can_add_hw))
1300                         if (group_sched_in(event, cpuctx, ctx))
1301                                 can_add_hw = 0;
1302         }
1303 }
1304
1305 static void
1306 ctx_sched_in(struct perf_event_context *ctx,
1307              struct perf_cpu_context *cpuctx,
1308              enum event_type_t event_type)
1309 {
1310         raw_spin_lock(&ctx->lock);
1311         ctx->is_active = 1;
1312         if (likely(!ctx->nr_events))
1313                 goto out;
1314
1315         ctx->timestamp = perf_clock();
1316
1317         perf_disable();
1318
1319         /*
1320          * First go through the list and put on any pinned groups
1321          * in order to give them the best chance of going on.
1322          */
1323         if (event_type & EVENT_PINNED)
1324                 ctx_pinned_sched_in(ctx, cpuctx);
1325
1326         /* Then walk through the lower prio flexible groups */
1327         if (event_type & EVENT_FLEXIBLE)
1328                 ctx_flexible_sched_in(ctx, cpuctx);
1329
1330         perf_enable();
1331  out:
1332         raw_spin_unlock(&ctx->lock);
1333 }
1334
1335 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1336                              enum event_type_t event_type)
1337 {
1338         struct perf_event_context *ctx = &cpuctx->ctx;
1339
1340         ctx_sched_in(ctx, cpuctx, event_type);
1341 }
1342
1343 static void task_ctx_sched_in(struct task_struct *task,
1344                               enum event_type_t event_type)
1345 {
1346         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1347         struct perf_event_context *ctx = task->perf_event_ctxp;
1348
1349         if (likely(!ctx))
1350                 return;
1351         if (cpuctx->task_ctx == ctx)
1352                 return;
1353         ctx_sched_in(ctx, cpuctx, event_type);
1354         cpuctx->task_ctx = ctx;
1355 }
1356 /*
1357  * Called from scheduler to add the events of the current task
1358  * with interrupts disabled.
1359  *
1360  * We restore the event value and then enable it.
1361  *
1362  * This does not protect us against NMI, but enable()
1363  * sets the enabled bit in the control field of event _before_
1364  * accessing the event control register. If a NMI hits, then it will
1365  * keep the event running.
1366  */
1367 void perf_event_task_sched_in(struct task_struct *task)
1368 {
1369         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1370         struct perf_event_context *ctx = task->perf_event_ctxp;
1371
1372         if (likely(!ctx))
1373                 return;
1374
1375         if (cpuctx->task_ctx == ctx)
1376                 return;
1377
1378         perf_disable();
1379
1380         /*
1381          * We want to keep the following priority order:
1382          * cpu pinned (that don't need to move), task pinned,
1383          * cpu flexible, task flexible.
1384          */
1385         cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1386
1387         ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1388         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1389         ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1390
1391         cpuctx->task_ctx = ctx;
1392
1393         perf_enable();
1394 }
1395
1396 #define MAX_INTERRUPTS (~0ULL)
1397
1398 static void perf_log_throttle(struct perf_event *event, int enable);
1399
1400 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1401 {
1402         u64 frequency = event->attr.sample_freq;
1403         u64 sec = NSEC_PER_SEC;
1404         u64 divisor, dividend;
1405
1406         int count_fls, nsec_fls, frequency_fls, sec_fls;
1407
1408         count_fls = fls64(count);
1409         nsec_fls = fls64(nsec);
1410         frequency_fls = fls64(frequency);
1411         sec_fls = 30;
1412
1413         /*
1414          * We got @count in @nsec, with a target of sample_freq HZ
1415          * the target period becomes:
1416          *
1417          *             @count * 10^9
1418          * period = -------------------
1419          *          @nsec * sample_freq
1420          *
1421          */
1422
1423         /*
1424          * Reduce accuracy by one bit such that @a and @b converge
1425          * to a similar magnitude.
1426          */
1427 #define REDUCE_FLS(a, b)                \
1428 do {                                    \
1429         if (a##_fls > b##_fls) {        \
1430                 a >>= 1;                \
1431                 a##_fls--;              \
1432         } else {                        \
1433                 b >>= 1;                \
1434                 b##_fls--;              \
1435         }                               \
1436 } while (0)
1437
1438         /*
1439          * Reduce accuracy until either term fits in a u64, then proceed with
1440          * the other, so that finally we can do a u64/u64 division.
1441          */
1442         while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1443                 REDUCE_FLS(nsec, frequency);
1444                 REDUCE_FLS(sec, count);
1445         }
1446
1447         if (count_fls + sec_fls > 64) {
1448                 divisor = nsec * frequency;
1449
1450                 while (count_fls + sec_fls > 64) {
1451                         REDUCE_FLS(count, sec);
1452                         divisor >>= 1;
1453                 }
1454
1455                 dividend = count * sec;
1456         } else {
1457                 dividend = count * sec;
1458
1459                 while (nsec_fls + frequency_fls > 64) {
1460                         REDUCE_FLS(nsec, frequency);
1461                         dividend >>= 1;
1462                 }
1463
1464                 divisor = nsec * frequency;
1465         }
1466
1467         return div64_u64(dividend, divisor);
1468 }
1469
1470 static void perf_event_stop(struct perf_event *event)
1471 {
1472         if (!event->pmu->stop)
1473                 return event->pmu->disable(event);
1474
1475         return event->pmu->stop(event);
1476 }
1477
1478 static int perf_event_start(struct perf_event *event)
1479 {
1480         if (!event->pmu->start)
1481                 return event->pmu->enable(event);
1482
1483         return event->pmu->start(event);
1484 }
1485
1486 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1487 {
1488         struct hw_perf_event *hwc = &event->hw;
1489         u64 period, sample_period;
1490         s64 delta;
1491
1492         period = perf_calculate_period(event, nsec, count);
1493
1494         delta = (s64)(period - hwc->sample_period);
1495         delta = (delta + 7) / 8; /* low pass filter */
1496
1497         sample_period = hwc->sample_period + delta;
1498
1499         if (!sample_period)
1500                 sample_period = 1;
1501
1502         hwc->sample_period = sample_period;
1503
1504         if (atomic64_read(&hwc->period_left) > 8*sample_period) {
1505                 perf_disable();
1506                 perf_event_stop(event);
1507                 atomic64_set(&hwc->period_left, 0);
1508                 perf_event_start(event);
1509                 perf_enable();
1510         }
1511 }
1512
1513 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1514 {
1515         struct perf_event *event;
1516         struct hw_perf_event *hwc;
1517         u64 interrupts, now;
1518         s64 delta;
1519
1520         raw_spin_lock(&ctx->lock);
1521         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1522                 if (event->state != PERF_EVENT_STATE_ACTIVE)
1523                         continue;
1524
1525                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1526                         continue;
1527
1528                 hwc = &event->hw;
1529
1530                 interrupts = hwc->interrupts;
1531                 hwc->interrupts = 0;
1532
1533                 /*
1534                  * unthrottle events on the tick
1535                  */
1536                 if (interrupts == MAX_INTERRUPTS) {
1537                         perf_log_throttle(event, 1);
1538                         perf_disable();
1539                         event->pmu->unthrottle(event);
1540                         perf_enable();
1541                 }
1542
1543                 if (!event->attr.freq || !event->attr.sample_freq)
1544                         continue;
1545
1546                 perf_disable();
1547                 event->pmu->read(event);
1548                 now = atomic64_read(&event->count);
1549                 delta = now - hwc->freq_count_stamp;
1550                 hwc->freq_count_stamp = now;
1551
1552                 if (delta > 0)
1553                         perf_adjust_period(event, TICK_NSEC, delta);
1554                 perf_enable();
1555         }
1556         raw_spin_unlock(&ctx->lock);
1557 }
1558
1559 /*
1560  * Round-robin a context's events:
1561  */
1562 static void rotate_ctx(struct perf_event_context *ctx)
1563 {
1564         raw_spin_lock(&ctx->lock);
1565
1566         /* Rotate the first entry last of non-pinned groups */
1567         list_rotate_left(&ctx->flexible_groups);
1568
1569         raw_spin_unlock(&ctx->lock);
1570 }
1571
1572 void perf_event_task_tick(struct task_struct *curr)
1573 {
1574         struct perf_cpu_context *cpuctx;
1575         struct perf_event_context *ctx;
1576         int rotate = 0;
1577
1578         if (!atomic_read(&nr_events))
1579                 return;
1580
1581         cpuctx = &__get_cpu_var(perf_cpu_context);
1582         if (cpuctx->ctx.nr_events &&
1583             cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1584                 rotate = 1;
1585
1586         ctx = curr->perf_event_ctxp;
1587         if (ctx && ctx->nr_events && ctx->nr_events != ctx->nr_active)
1588                 rotate = 1;
1589
1590         perf_ctx_adjust_freq(&cpuctx->ctx);
1591         if (ctx)
1592                 perf_ctx_adjust_freq(ctx);
1593
1594         if (!rotate)
1595                 return;
1596
1597         perf_disable();
1598         cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1599         if (ctx)
1600                 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1601
1602         rotate_ctx(&cpuctx->ctx);
1603         if (ctx)
1604                 rotate_ctx(ctx);
1605
1606         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1607         if (ctx)
1608                 task_ctx_sched_in(curr, EVENT_FLEXIBLE);
1609         perf_enable();
1610 }
1611
1612 static int event_enable_on_exec(struct perf_event *event,
1613                                 struct perf_event_context *ctx)
1614 {
1615         if (!event->attr.enable_on_exec)
1616                 return 0;
1617
1618         event->attr.enable_on_exec = 0;
1619         if (event->state >= PERF_EVENT_STATE_INACTIVE)
1620                 return 0;
1621
1622         __perf_event_mark_enabled(event, ctx);
1623
1624         return 1;
1625 }
1626
1627 /*
1628  * Enable all of a task's events that have been marked enable-on-exec.
1629  * This expects task == current.
1630  */
1631 static void perf_event_enable_on_exec(struct task_struct *task)
1632 {
1633         struct perf_event_context *ctx;
1634         struct perf_event *event;
1635         unsigned long flags;
1636         int enabled = 0;
1637         int ret;
1638
1639         local_irq_save(flags);
1640         ctx = task->perf_event_ctxp;
1641         if (!ctx || !ctx->nr_events)
1642                 goto out;
1643
1644         __perf_event_task_sched_out(ctx);
1645
1646         raw_spin_lock(&ctx->lock);
1647
1648         list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1649                 ret = event_enable_on_exec(event, ctx);
1650                 if (ret)
1651                         enabled = 1;
1652         }
1653
1654         list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1655                 ret = event_enable_on_exec(event, ctx);
1656                 if (ret)
1657                         enabled = 1;
1658         }
1659
1660         /*
1661          * Unclone this context if we enabled any event.
1662          */
1663         if (enabled)
1664                 unclone_ctx(ctx);
1665
1666         raw_spin_unlock(&ctx->lock);
1667
1668         perf_event_task_sched_in(task);
1669  out:
1670         local_irq_restore(flags);
1671 }
1672
1673 /*
1674  * Cross CPU call to read the hardware event
1675  */
1676 static void __perf_event_read(void *info)
1677 {
1678         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1679         struct perf_event *event = info;
1680         struct perf_event_context *ctx = event->ctx;
1681
1682         /*
1683          * If this is a task context, we need to check whether it is
1684          * the current task context of this cpu.  If not it has been
1685          * scheduled out before the smp call arrived.  In that case
1686          * event->count would have been updated to a recent sample
1687          * when the event was scheduled out.
1688          */
1689         if (ctx->task && cpuctx->task_ctx != ctx)
1690                 return;
1691
1692         raw_spin_lock(&ctx->lock);
1693         update_context_time(ctx);
1694         update_event_times(event);
1695         raw_spin_unlock(&ctx->lock);
1696
1697         event->pmu->read(event);
1698 }
1699
1700 static u64 perf_event_read(struct perf_event *event)
1701 {
1702         /*
1703          * If event is enabled and currently active on a CPU, update the
1704          * value in the event structure:
1705          */
1706         if (event->state == PERF_EVENT_STATE_ACTIVE) {
1707                 smp_call_function_single(event->oncpu,
1708                                          __perf_event_read, event, 1);
1709         } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1710                 struct perf_event_context *ctx = event->ctx;
1711                 unsigned long flags;
1712
1713                 raw_spin_lock_irqsave(&ctx->lock, flags);
1714                 update_context_time(ctx);
1715                 update_event_times(event);
1716                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1717         }
1718
1719         return atomic64_read(&event->count);
1720 }
1721
1722 /*
1723  * Initialize the perf_event context in a task_struct:
1724  */
1725 static void
1726 __perf_event_init_context(struct perf_event_context *ctx,
1727                             struct task_struct *task)
1728 {
1729         raw_spin_lock_init(&ctx->lock);
1730         mutex_init(&ctx->mutex);
1731         INIT_LIST_HEAD(&ctx->pinned_groups);
1732         INIT_LIST_HEAD(&ctx->flexible_groups);
1733         INIT_LIST_HEAD(&ctx->event_list);
1734         atomic_set(&ctx->refcount, 1);
1735         ctx->task = task;
1736 }
1737
1738 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1739 {
1740         struct perf_event_context *ctx;
1741         struct perf_cpu_context *cpuctx;
1742         struct task_struct *task;
1743         unsigned long flags;
1744         int err;
1745
1746         if (pid == -1 && cpu != -1) {
1747                 /* Must be root to operate on a CPU event: */
1748                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1749                         return ERR_PTR(-EACCES);
1750
1751                 if (cpu < 0 || cpu >= nr_cpumask_bits)
1752                         return ERR_PTR(-EINVAL);
1753
1754                 /*
1755                  * We could be clever and allow to attach a event to an
1756                  * offline CPU and activate it when the CPU comes up, but
1757                  * that's for later.
1758                  */
1759                 if (!cpu_online(cpu))
1760                         return ERR_PTR(-ENODEV);
1761
1762                 cpuctx = &per_cpu(perf_cpu_context, cpu);
1763                 ctx = &cpuctx->ctx;
1764                 get_ctx(ctx);
1765
1766                 return ctx;
1767         }
1768
1769         rcu_read_lock();
1770         if (!pid)
1771                 task = current;
1772         else
1773                 task = find_task_by_vpid(pid);
1774         if (task)
1775                 get_task_struct(task);
1776         rcu_read_unlock();
1777
1778         if (!task)
1779                 return ERR_PTR(-ESRCH);
1780
1781         /*
1782          * Can't attach events to a dying task.
1783          */
1784         err = -ESRCH;
1785         if (task->flags & PF_EXITING)
1786                 goto errout;
1787
1788         /* Reuse ptrace permission checks for now. */
1789         err = -EACCES;
1790         if (!ptrace_may_access(task, PTRACE_MODE_READ))
1791                 goto errout;
1792
1793  retry:
1794         ctx = perf_lock_task_context(task, &flags);
1795         if (ctx) {
1796                 unclone_ctx(ctx);
1797                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1798         }
1799
1800         if (!ctx) {
1801                 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1802                 err = -ENOMEM;
1803                 if (!ctx)
1804                         goto errout;
1805                 __perf_event_init_context(ctx, task);
1806                 get_ctx(ctx);
1807                 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1808                         /*
1809                          * We raced with some other task; use
1810                          * the context they set.
1811                          */
1812                         kfree(ctx);
1813                         goto retry;
1814                 }
1815                 get_task_struct(task);
1816         }
1817
1818         put_task_struct(task);
1819         return ctx;
1820
1821  errout:
1822         put_task_struct(task);
1823         return ERR_PTR(err);
1824 }
1825
1826 static void perf_event_free_filter(struct perf_event *event);
1827
1828 static void free_event_rcu(struct rcu_head *head)
1829 {
1830         struct perf_event *event;
1831
1832         event = container_of(head, struct perf_event, rcu_head);
1833         if (event->ns)
1834                 put_pid_ns(event->ns);
1835         perf_event_free_filter(event);
1836         kfree(event);
1837 }
1838
1839 static void perf_pending_sync(struct perf_event *event);
1840
1841 static void free_event(struct perf_event *event)
1842 {
1843         perf_pending_sync(event);
1844
1845         if (!event->parent) {
1846                 atomic_dec(&nr_events);
1847                 if (event->attr.mmap)
1848                         atomic_dec(&nr_mmap_events);
1849                 if (event->attr.comm)
1850                         atomic_dec(&nr_comm_events);
1851                 if (event->attr.task)
1852                         atomic_dec(&nr_task_events);
1853         }
1854
1855         if (event->output) {
1856                 fput(event->output->filp);
1857                 event->output = NULL;
1858         }
1859
1860         if (event->destroy)
1861                 event->destroy(event);
1862
1863         put_ctx(event->ctx);
1864         call_rcu(&event->rcu_head, free_event_rcu);
1865 }
1866
1867 int perf_event_release_kernel(struct perf_event *event)
1868 {
1869         struct perf_event_context *ctx = event->ctx;
1870
1871         WARN_ON_ONCE(ctx->parent_ctx);
1872         /*
1873          * There are two ways this annotation is useful:
1874          *
1875          *  1) there is a lock recursion from perf_event_exit_task
1876          *     see the comment there.
1877          *
1878          *  2) there is a lock-inversion with mmap_sem through
1879          *     perf_event_read_group(), which takes faults while
1880          *     holding ctx->mutex, however this is called after
1881          *     the last filedesc died, so there is no possibility
1882          *     to trigger the AB-BA case.
1883          */
1884         mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
1885         perf_event_remove_from_context(event);
1886         mutex_unlock(&ctx->mutex);
1887
1888         mutex_lock(&event->owner->perf_event_mutex);
1889         list_del_init(&event->owner_entry);
1890         mutex_unlock(&event->owner->perf_event_mutex);
1891         put_task_struct(event->owner);
1892
1893         free_event(event);
1894
1895         return 0;
1896 }
1897 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1898
1899 /*
1900  * Called when the last reference to the file is gone.
1901  */
1902 static int perf_release(struct inode *inode, struct file *file)
1903 {
1904         struct perf_event *event = file->private_data;
1905
1906         file->private_data = NULL;
1907
1908         return perf_event_release_kernel(event);
1909 }
1910
1911 static int perf_event_read_size(struct perf_event *event)
1912 {
1913         int entry = sizeof(u64); /* value */
1914         int size = 0;
1915         int nr = 1;
1916
1917         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1918                 size += sizeof(u64);
1919
1920         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1921                 size += sizeof(u64);
1922
1923         if (event->attr.read_format & PERF_FORMAT_ID)
1924                 entry += sizeof(u64);
1925
1926         if (event->attr.read_format & PERF_FORMAT_GROUP) {
1927                 nr += event->group_leader->nr_siblings;
1928                 size += sizeof(u64);
1929         }
1930
1931         size += entry * nr;
1932
1933         return size;
1934 }
1935
1936 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1937 {
1938         struct perf_event *child;
1939         u64 total = 0;
1940
1941         *enabled = 0;
1942         *running = 0;
1943
1944         mutex_lock(&event->child_mutex);
1945         total += perf_event_read(event);
1946         *enabled += event->total_time_enabled +
1947                         atomic64_read(&event->child_total_time_enabled);
1948         *running += event->total_time_running +
1949                         atomic64_read(&event->child_total_time_running);
1950
1951         list_for_each_entry(child, &event->child_list, child_list) {
1952                 total += perf_event_read(child);
1953                 *enabled += child->total_time_enabled;
1954                 *running += child->total_time_running;
1955         }
1956         mutex_unlock(&event->child_mutex);
1957
1958         return total;
1959 }
1960 EXPORT_SYMBOL_GPL(perf_event_read_value);
1961
1962 static int perf_event_read_group(struct perf_event *event,
1963                                    u64 read_format, char __user *buf)
1964 {
1965         struct perf_event *leader = event->group_leader, *sub;
1966         int n = 0, size = 0, ret = -EFAULT;
1967         struct perf_event_context *ctx = leader->ctx;
1968         u64 values[5];
1969         u64 count, enabled, running;
1970
1971         mutex_lock(&ctx->mutex);
1972         count = perf_event_read_value(leader, &enabled, &running);
1973
1974         values[n++] = 1 + leader->nr_siblings;
1975         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1976                 values[n++] = enabled;
1977         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1978                 values[n++] = running;
1979         values[n++] = count;
1980         if (read_format & PERF_FORMAT_ID)
1981                 values[n++] = primary_event_id(leader);
1982
1983         size = n * sizeof(u64);
1984
1985         if (copy_to_user(buf, values, size))
1986                 goto unlock;
1987
1988         ret = size;
1989
1990         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1991                 n = 0;
1992
1993                 values[n++] = perf_event_read_value(sub, &enabled, &running);
1994                 if (read_format & PERF_FORMAT_ID)
1995                         values[n++] = primary_event_id(sub);
1996
1997                 size = n * sizeof(u64);
1998
1999                 if (copy_to_user(buf + ret, values, size)) {
2000                         ret = -EFAULT;
2001                         goto unlock;
2002                 }
2003
2004                 ret += size;
2005         }
2006 unlock:
2007         mutex_unlock(&ctx->mutex);
2008
2009         return ret;
2010 }
2011
2012 static int perf_event_read_one(struct perf_event *event,
2013                                  u64 read_format, char __user *buf)
2014 {
2015         u64 enabled, running;
2016         u64 values[4];
2017         int n = 0;
2018
2019         values[n++] = perf_event_read_value(event, &enabled, &running);
2020         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2021                 values[n++] = enabled;
2022         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2023                 values[n++] = running;
2024         if (read_format & PERF_FORMAT_ID)
2025                 values[n++] = primary_event_id(event);
2026
2027         if (copy_to_user(buf, values, n * sizeof(u64)))
2028                 return -EFAULT;
2029
2030         return n * sizeof(u64);
2031 }
2032
2033 /*
2034  * Read the performance event - simple non blocking version for now
2035  */
2036 static ssize_t
2037 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2038 {
2039         u64 read_format = event->attr.read_format;
2040         int ret;
2041
2042         /*
2043          * Return end-of-file for a read on a event that is in
2044          * error state (i.e. because it was pinned but it couldn't be
2045          * scheduled on to the CPU at some point).
2046          */
2047         if (event->state == PERF_EVENT_STATE_ERROR)
2048                 return 0;
2049
2050         if (count < perf_event_read_size(event))
2051                 return -ENOSPC;
2052
2053         WARN_ON_ONCE(event->ctx->parent_ctx);
2054         if (read_format & PERF_FORMAT_GROUP)
2055                 ret = perf_event_read_group(event, read_format, buf);
2056         else
2057                 ret = perf_event_read_one(event, read_format, buf);
2058
2059         return ret;
2060 }
2061
2062 static ssize_t
2063 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2064 {
2065         struct perf_event *event = file->private_data;
2066
2067         return perf_read_hw(event, buf, count);
2068 }
2069
2070 static unsigned int perf_poll(struct file *file, poll_table *wait)
2071 {
2072         struct perf_event *event = file->private_data;
2073         struct perf_mmap_data *data;
2074         unsigned int events = POLL_HUP;
2075
2076         rcu_read_lock();
2077         data = rcu_dereference(event->data);
2078         if (data)
2079                 events = atomic_xchg(&data->poll, 0);
2080         rcu_read_unlock();
2081
2082         poll_wait(file, &event->waitq, wait);
2083
2084         return events;
2085 }
2086
2087 static void perf_event_reset(struct perf_event *event)
2088 {
2089         (void)perf_event_read(event);
2090         atomic64_set(&event->count, 0);
2091         perf_event_update_userpage(event);
2092 }
2093
2094 /*
2095  * Holding the top-level event's child_mutex means that any
2096  * descendant process that has inherited this event will block
2097  * in sync_child_event if it goes to exit, thus satisfying the
2098  * task existence requirements of perf_event_enable/disable.
2099  */
2100 static void perf_event_for_each_child(struct perf_event *event,
2101                                         void (*func)(struct perf_event *))
2102 {
2103         struct perf_event *child;
2104
2105         WARN_ON_ONCE(event->ctx->parent_ctx);
2106         mutex_lock(&event->child_mutex);
2107         func(event);
2108         list_for_each_entry(child, &event->child_list, child_list)
2109                 func(child);
2110         mutex_unlock(&event->child_mutex);
2111 }
2112
2113 static void perf_event_for_each(struct perf_event *event,
2114                                   void (*func)(struct perf_event *))
2115 {
2116         struct perf_event_context *ctx = event->ctx;
2117         struct perf_event *sibling;
2118
2119         WARN_ON_ONCE(ctx->parent_ctx);
2120         mutex_lock(&ctx->mutex);
2121         event = event->group_leader;
2122
2123         perf_event_for_each_child(event, func);
2124         func(event);
2125         list_for_each_entry(sibling, &event->sibling_list, group_entry)
2126                 perf_event_for_each_child(event, func);
2127         mutex_unlock(&ctx->mutex);
2128 }
2129
2130 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2131 {
2132         struct perf_event_context *ctx = event->ctx;
2133         unsigned long size;
2134         int ret = 0;
2135         u64 value;
2136
2137         if (!event->attr.sample_period)
2138                 return -EINVAL;
2139
2140         size = copy_from_user(&value, arg, sizeof(value));
2141         if (size != sizeof(value))
2142                 return -EFAULT;
2143
2144         if (!value)
2145                 return -EINVAL;
2146
2147         raw_spin_lock_irq(&ctx->lock);
2148         if (event->attr.freq) {
2149                 if (value > sysctl_perf_event_sample_rate) {
2150                         ret = -EINVAL;
2151                         goto unlock;
2152                 }
2153
2154                 event->attr.sample_freq = value;
2155         } else {
2156                 event->attr.sample_period = value;
2157                 event->hw.sample_period = value;
2158         }
2159 unlock:
2160         raw_spin_unlock_irq(&ctx->lock);
2161
2162         return ret;
2163 }
2164
2165 static int perf_event_set_output(struct perf_event *event, int output_fd);
2166 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2167
2168 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2169 {
2170         struct perf_event *event = file->private_data;
2171         void (*func)(struct perf_event *);
2172         u32 flags = arg;
2173
2174         switch (cmd) {
2175         case PERF_EVENT_IOC_ENABLE:
2176                 func = perf_event_enable;
2177                 break;
2178         case PERF_EVENT_IOC_DISABLE:
2179                 func = perf_event_disable;
2180                 break;
2181         case PERF_EVENT_IOC_RESET:
2182                 func = perf_event_reset;
2183                 break;
2184
2185         case PERF_EVENT_IOC_REFRESH:
2186                 return perf_event_refresh(event, arg);
2187
2188         case PERF_EVENT_IOC_PERIOD:
2189                 return perf_event_period(event, (u64 __user *)arg);
2190
2191         case PERF_EVENT_IOC_SET_OUTPUT:
2192                 return perf_event_set_output(event, arg);
2193
2194         case PERF_EVENT_IOC_SET_FILTER:
2195                 return perf_event_set_filter(event, (void __user *)arg);
2196
2197         default:
2198                 return -ENOTTY;
2199         }
2200
2201         if (flags & PERF_IOC_FLAG_GROUP)
2202                 perf_event_for_each(event, func);
2203         else
2204                 perf_event_for_each_child(event, func);
2205
2206         return 0;
2207 }
2208
2209 int perf_event_task_enable(void)
2210 {
2211         struct perf_event *event;
2212
2213         mutex_lock(&current->perf_event_mutex);
2214         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2215                 perf_event_for_each_child(event, perf_event_enable);
2216         mutex_unlock(&current->perf_event_mutex);
2217
2218         return 0;
2219 }
2220
2221 int perf_event_task_disable(void)
2222 {
2223         struct perf_event *event;
2224
2225         mutex_lock(&current->perf_event_mutex);
2226         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2227                 perf_event_for_each_child(event, perf_event_disable);
2228         mutex_unlock(&current->perf_event_mutex);
2229
2230         return 0;
2231 }
2232
2233 #ifndef PERF_EVENT_INDEX_OFFSET
2234 # define PERF_EVENT_INDEX_OFFSET 0
2235 #endif
2236
2237 static int perf_event_index(struct perf_event *event)
2238 {
2239         if (event->state != PERF_EVENT_STATE_ACTIVE)
2240                 return 0;
2241
2242         return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2243 }
2244
2245 /*
2246  * Callers need to ensure there can be no nesting of this function, otherwise
2247  * the seqlock logic goes bad. We can not serialize this because the arch
2248  * code calls this from NMI context.
2249  */
2250 void perf_event_update_userpage(struct perf_event *event)
2251 {
2252         struct perf_event_mmap_page *userpg;
2253         struct perf_mmap_data *data;
2254
2255         rcu_read_lock();
2256         data = rcu_dereference(event->data);
2257         if (!data)
2258                 goto unlock;
2259
2260         userpg = data->user_page;
2261
2262         /*
2263          * Disable preemption so as to not let the corresponding user-space
2264          * spin too long if we get preempted.
2265          */
2266         preempt_disable();
2267         ++userpg->lock;
2268         barrier();
2269         userpg->index = perf_event_index(event);
2270         userpg->offset = atomic64_read(&event->count);
2271         if (event->state == PERF_EVENT_STATE_ACTIVE)
2272                 userpg->offset -= atomic64_read(&event->hw.prev_count);
2273
2274         userpg->time_enabled = event->total_time_enabled +
2275                         atomic64_read(&event->child_total_time_enabled);
2276
2277         userpg->time_running = event->total_time_running +
2278                         atomic64_read(&event->child_total_time_running);
2279
2280         barrier();
2281         ++userpg->lock;
2282         preempt_enable();
2283 unlock:
2284         rcu_read_unlock();
2285 }
2286
2287 static unsigned long perf_data_size(struct perf_mmap_data *data)
2288 {
2289         return data->nr_pages << (PAGE_SHIFT + data->data_order);
2290 }
2291
2292 #ifndef CONFIG_PERF_USE_VMALLOC
2293
2294 /*
2295  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2296  */
2297
2298 static struct page *
2299 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2300 {
2301         if (pgoff > data->nr_pages)
2302                 return NULL;
2303
2304         if (pgoff == 0)
2305                 return virt_to_page(data->user_page);
2306
2307         return virt_to_page(data->data_pages[pgoff - 1]);
2308 }
2309
2310 static struct perf_mmap_data *
2311 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2312 {
2313         struct perf_mmap_data *data;
2314         unsigned long size;
2315         int i;
2316
2317         WARN_ON(atomic_read(&event->mmap_count));
2318
2319         size = sizeof(struct perf_mmap_data);
2320         size += nr_pages * sizeof(void *);
2321
2322         data = kzalloc(size, GFP_KERNEL);
2323         if (!data)
2324                 goto fail;
2325
2326         data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2327         if (!data->user_page)
2328                 goto fail_user_page;
2329
2330         for (i = 0; i < nr_pages; i++) {
2331                 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2332                 if (!data->data_pages[i])
2333                         goto fail_data_pages;
2334         }
2335
2336         data->data_order = 0;
2337         data->nr_pages = nr_pages;
2338
2339         return data;
2340
2341 fail_data_pages:
2342         for (i--; i >= 0; i--)
2343                 free_page((unsigned long)data->data_pages[i]);
2344
2345         free_page((unsigned long)data->user_page);
2346
2347 fail_user_page:
2348         kfree(data);
2349
2350 fail:
2351         return NULL;
2352 }
2353
2354 static void perf_mmap_free_page(unsigned long addr)
2355 {
2356         struct page *page = virt_to_page((void *)addr);
2357
2358         page->mapping = NULL;
2359         __free_page(page);
2360 }
2361
2362 static void perf_mmap_data_free(struct perf_mmap_data *data)
2363 {
2364         int i;
2365
2366         perf_mmap_free_page((unsigned long)data->user_page);
2367         for (i = 0; i < data->nr_pages; i++)
2368                 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2369         kfree(data);
2370 }
2371
2372 #else
2373
2374 /*
2375  * Back perf_mmap() with vmalloc memory.
2376  *
2377  * Required for architectures that have d-cache aliasing issues.
2378  */
2379
2380 static struct page *
2381 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2382 {
2383         if (pgoff > (1UL << data->data_order))
2384                 return NULL;
2385
2386         return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2387 }
2388
2389 static void perf_mmap_unmark_page(void *addr)
2390 {
2391         struct page *page = vmalloc_to_page(addr);
2392
2393         page->mapping = NULL;
2394 }
2395
2396 static void perf_mmap_data_free_work(struct work_struct *work)
2397 {
2398         struct perf_mmap_data *data;
2399         void *base;
2400         int i, nr;
2401
2402         data = container_of(work, struct perf_mmap_data, work);
2403         nr = 1 << data->data_order;
2404
2405         base = data->user_page;
2406         for (i = 0; i < nr + 1; i++)
2407                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2408
2409         vfree(base);
2410         kfree(data);
2411 }
2412
2413 static void perf_mmap_data_free(struct perf_mmap_data *data)
2414 {
2415         schedule_work(&data->work);
2416 }
2417
2418 static struct perf_mmap_data *
2419 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2420 {
2421         struct perf_mmap_data *data;
2422         unsigned long size;
2423         void *all_buf;
2424
2425         WARN_ON(atomic_read(&event->mmap_count));
2426
2427         size = sizeof(struct perf_mmap_data);
2428         size += sizeof(void *);
2429
2430         data = kzalloc(size, GFP_KERNEL);
2431         if (!data)
2432                 goto fail;
2433
2434         INIT_WORK(&data->work, perf_mmap_data_free_work);
2435
2436         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2437         if (!all_buf)
2438                 goto fail_all_buf;
2439
2440         data->user_page = all_buf;
2441         data->data_pages[0] = all_buf + PAGE_SIZE;
2442         data->data_order = ilog2(nr_pages);
2443         data->nr_pages = 1;
2444
2445         return data;
2446
2447 fail_all_buf:
2448         kfree(data);
2449
2450 fail:
2451         return NULL;
2452 }
2453
2454 #endif
2455
2456 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2457 {
2458         struct perf_event *event = vma->vm_file->private_data;
2459         struct perf_mmap_data *data;
2460         int ret = VM_FAULT_SIGBUS;
2461
2462         if (vmf->flags & FAULT_FLAG_MKWRITE) {
2463                 if (vmf->pgoff == 0)
2464                         ret = 0;
2465                 return ret;
2466         }
2467
2468         rcu_read_lock();
2469         data = rcu_dereference(event->data);
2470         if (!data)
2471                 goto unlock;
2472
2473         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2474                 goto unlock;
2475
2476         vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2477         if (!vmf->page)
2478                 goto unlock;
2479
2480         get_page(vmf->page);
2481         vmf->page->mapping = vma->vm_file->f_mapping;
2482         vmf->page->index   = vmf->pgoff;
2483
2484         ret = 0;
2485 unlock:
2486         rcu_read_unlock();
2487
2488         return ret;
2489 }
2490
2491 static void
2492 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2493 {
2494         long max_size = perf_data_size(data);
2495
2496         atomic_set(&data->lock, -1);
2497
2498         if (event->attr.watermark) {
2499                 data->watermark = min_t(long, max_size,
2500                                         event->attr.wakeup_watermark);
2501         }
2502
2503         if (!data->watermark)
2504                 data->watermark = max_size / 2;
2505
2506
2507         rcu_assign_pointer(event->data, data);
2508 }
2509
2510 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2511 {
2512         struct perf_mmap_data *data;
2513
2514         data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2515         perf_mmap_data_free(data);
2516 }
2517
2518 static void perf_mmap_data_release(struct perf_event *event)
2519 {
2520         struct perf_mmap_data *data = event->data;
2521
2522         WARN_ON(atomic_read(&event->mmap_count));
2523
2524         rcu_assign_pointer(event->data, NULL);
2525         call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2526 }
2527
2528 static void perf_mmap_open(struct vm_area_struct *vma)
2529 {
2530         struct perf_event *event = vma->vm_file->private_data;
2531
2532         atomic_inc(&event->mmap_count);
2533 }
2534
2535 static void perf_mmap_close(struct vm_area_struct *vma)
2536 {
2537         struct perf_event *event = vma->vm_file->private_data;
2538
2539         WARN_ON_ONCE(event->ctx->parent_ctx);
2540         if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2541                 unsigned long size = perf_data_size(event->data);
2542                 struct user_struct *user = current_user();
2543
2544                 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2545                 vma->vm_mm->locked_vm -= event->data->nr_locked;
2546                 perf_mmap_data_release(event);
2547                 mutex_unlock(&event->mmap_mutex);
2548         }
2549 }
2550
2551 static const struct vm_operations_struct perf_mmap_vmops = {
2552         .open           = perf_mmap_open,
2553         .close          = perf_mmap_close,
2554         .fault          = perf_mmap_fault,
2555         .page_mkwrite   = perf_mmap_fault,
2556 };
2557
2558 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2559 {
2560         struct perf_event *event = file->private_data;
2561         unsigned long user_locked, user_lock_limit;
2562         struct user_struct *user = current_user();
2563         unsigned long locked, lock_limit;
2564         struct perf_mmap_data *data;
2565         unsigned long vma_size;
2566         unsigned long nr_pages;
2567         long user_extra, extra;
2568         int ret = 0;
2569
2570         if (!(vma->vm_flags & VM_SHARED))
2571                 return -EINVAL;
2572
2573         vma_size = vma->vm_end - vma->vm_start;
2574         nr_pages = (vma_size / PAGE_SIZE) - 1;
2575
2576         /*
2577          * If we have data pages ensure they're a power-of-two number, so we
2578          * can do bitmasks instead of modulo.
2579          */
2580         if (nr_pages != 0 && !is_power_of_2(nr_pages))
2581                 return -EINVAL;
2582
2583         if (vma_size != PAGE_SIZE * (1 + nr_pages))
2584                 return -EINVAL;
2585
2586         if (vma->vm_pgoff != 0)
2587                 return -EINVAL;
2588
2589         WARN_ON_ONCE(event->ctx->parent_ctx);
2590         mutex_lock(&event->mmap_mutex);
2591         if (event->output) {
2592                 ret = -EINVAL;
2593                 goto unlock;
2594         }
2595
2596         if (atomic_inc_not_zero(&event->mmap_count)) {
2597                 if (nr_pages != event->data->nr_pages)
2598                         ret = -EINVAL;
2599                 goto unlock;
2600         }
2601
2602         user_extra = nr_pages + 1;
2603         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2604
2605         /*
2606          * Increase the limit linearly with more CPUs:
2607          */
2608         user_lock_limit *= num_online_cpus();
2609
2610         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2611
2612         extra = 0;
2613         if (user_locked > user_lock_limit)
2614                 extra = user_locked - user_lock_limit;
2615
2616         lock_limit = rlimit(RLIMIT_MEMLOCK);
2617         lock_limit >>= PAGE_SHIFT;
2618         locked = vma->vm_mm->locked_vm + extra;
2619
2620         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2621                 !capable(CAP_IPC_LOCK)) {
2622                 ret = -EPERM;
2623                 goto unlock;
2624         }
2625
2626         WARN_ON(event->data);
2627
2628         data = perf_mmap_data_alloc(event, nr_pages);
2629         ret = -ENOMEM;
2630         if (!data)
2631                 goto unlock;
2632
2633         ret = 0;
2634         perf_mmap_data_init(event, data);
2635
2636         atomic_set(&event->mmap_count, 1);
2637         atomic_long_add(user_extra, &user->locked_vm);
2638         vma->vm_mm->locked_vm += extra;
2639         event->data->nr_locked = extra;
2640         if (vma->vm_flags & VM_WRITE)
2641                 event->data->writable = 1;
2642
2643 unlock:
2644         mutex_unlock(&event->mmap_mutex);
2645
2646         vma->vm_flags |= VM_RESERVED;
2647         vma->vm_ops = &perf_mmap_vmops;
2648
2649         return ret;
2650 }
2651
2652 static int perf_fasync(int fd, struct file *filp, int on)
2653 {
2654         struct inode *inode = filp->f_path.dentry->d_inode;
2655         struct perf_event *event = filp->private_data;
2656         int retval;
2657
2658         mutex_lock(&inode->i_mutex);
2659         retval = fasync_helper(fd, filp, on, &event->fasync);
2660         mutex_unlock(&inode->i_mutex);
2661
2662         if (retval < 0)
2663                 return retval;
2664
2665         return 0;
2666 }
2667
2668 static const struct file_operations perf_fops = {
2669         .llseek                 = no_llseek,
2670         .release                = perf_release,
2671         .read                   = perf_read,
2672         .poll                   = perf_poll,
2673         .unlocked_ioctl         = perf_ioctl,
2674         .compat_ioctl           = perf_ioctl,
2675         .mmap                   = perf_mmap,
2676         .fasync                 = perf_fasync,
2677 };
2678
2679 /*
2680  * Perf event wakeup
2681  *
2682  * If there's data, ensure we set the poll() state and publish everything
2683  * to user-space before waking everybody up.
2684  */
2685
2686 void perf_event_wakeup(struct perf_event *event)
2687 {
2688         wake_up_all(&event->waitq);
2689
2690         if (event->pending_kill) {
2691                 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2692                 event->pending_kill = 0;
2693         }
2694 }
2695
2696 /*
2697  * Pending wakeups
2698  *
2699  * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2700  *
2701  * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2702  * single linked list and use cmpxchg() to add entries lockless.
2703  */
2704
2705 static void perf_pending_event(struct perf_pending_entry *entry)
2706 {
2707         struct perf_event *event = container_of(entry,
2708                         struct perf_event, pending);
2709
2710         if (event->pending_disable) {
2711                 event->pending_disable = 0;
2712                 __perf_event_disable(event);
2713         }
2714
2715         if (event->pending_wakeup) {
2716                 event->pending_wakeup = 0;
2717                 perf_event_wakeup(event);
2718         }
2719 }
2720
2721 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2722
2723 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2724         PENDING_TAIL,
2725 };
2726
2727 static void perf_pending_queue(struct perf_pending_entry *entry,
2728                                void (*func)(struct perf_pending_entry *))
2729 {
2730         struct perf_pending_entry **head;
2731
2732         if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2733                 return;
2734
2735         entry->func = func;
2736
2737         head = &get_cpu_var(perf_pending_head);
2738
2739         do {
2740                 entry->next = *head;
2741         } while (cmpxchg(head, entry->next, entry) != entry->next);
2742
2743         set_perf_event_pending();
2744
2745         put_cpu_var(perf_pending_head);
2746 }
2747
2748 static int __perf_pending_run(void)
2749 {
2750         struct perf_pending_entry *list;
2751         int nr = 0;
2752
2753         list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2754         while (list != PENDING_TAIL) {
2755                 void (*func)(struct perf_pending_entry *);
2756                 struct perf_pending_entry *entry = list;
2757
2758                 list = list->next;
2759
2760                 func = entry->func;
2761                 entry->next = NULL;
2762                 /*
2763                  * Ensure we observe the unqueue before we issue the wakeup,
2764                  * so that we won't be waiting forever.
2765                  * -- see perf_not_pending().
2766                  */
2767                 smp_wmb();
2768
2769                 func(entry);
2770                 nr++;
2771         }
2772
2773         return nr;
2774 }
2775
2776 static inline int perf_not_pending(struct perf_event *event)
2777 {
2778         /*
2779          * If we flush on whatever cpu we run, there is a chance we don't
2780          * need to wait.
2781          */
2782         get_cpu();
2783         __perf_pending_run();
2784         put_cpu();
2785
2786         /*
2787          * Ensure we see the proper queue state before going to sleep
2788          * so that we do not miss the wakeup. -- see perf_pending_handle()
2789          */
2790         smp_rmb();
2791         return event->pending.next == NULL;
2792 }
2793
2794 static void perf_pending_sync(struct perf_event *event)
2795 {
2796         wait_event(event->waitq, perf_not_pending(event));
2797 }
2798
2799 void perf_event_do_pending(void)
2800 {
2801         __perf_pending_run();
2802 }
2803
2804 /*
2805  * Callchain support -- arch specific
2806  */
2807
2808 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2809 {
2810         return NULL;
2811 }
2812
2813 __weak
2814 void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip, int skip)
2815 {
2816 }
2817
2818
2819 /*
2820  * We assume there is only KVM supporting the callbacks.
2821  * Later on, we might change it to a list if there is
2822  * another virtualization implementation supporting the callbacks.
2823  */
2824 struct perf_guest_info_callbacks *perf_guest_cbs;
2825
2826 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2827 {
2828         perf_guest_cbs = cbs;
2829         return 0;
2830 }
2831 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
2832
2833 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2834 {
2835         perf_guest_cbs = NULL;
2836         return 0;
2837 }
2838 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
2839
2840 /*
2841  * Output
2842  */
2843 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2844                               unsigned long offset, unsigned long head)
2845 {
2846         unsigned long mask;
2847
2848         if (!data->writable)
2849                 return true;
2850
2851         mask = perf_data_size(data) - 1;
2852
2853         offset = (offset - tail) & mask;
2854         head   = (head   - tail) & mask;
2855
2856         if ((int)(head - offset) < 0)
2857                 return false;
2858
2859         return true;
2860 }
2861
2862 static void perf_output_wakeup(struct perf_output_handle *handle)
2863 {
2864         atomic_set(&handle->data->poll, POLL_IN);
2865
2866         if (handle->nmi) {
2867                 handle->event->pending_wakeup = 1;
2868                 perf_pending_queue(&handle->event->pending,
2869                                    perf_pending_event);
2870         } else
2871                 perf_event_wakeup(handle->event);
2872 }
2873
2874 /*
2875  * Curious locking construct.
2876  *
2877  * We need to ensure a later event_id doesn't publish a head when a former
2878  * event_id isn't done writing. However since we need to deal with NMIs we
2879  * cannot fully serialize things.
2880  *
2881  * What we do is serialize between CPUs so we only have to deal with NMI
2882  * nesting on a single CPU.
2883  *
2884  * We only publish the head (and generate a wakeup) when the outer-most
2885  * event_id completes.
2886  */
2887 static void perf_output_lock(struct perf_output_handle *handle)
2888 {
2889         struct perf_mmap_data *data = handle->data;
2890         int cur, cpu = get_cpu();
2891
2892         handle->locked = 0;
2893
2894         for (;;) {
2895                 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2896                 if (cur == -1) {
2897                         handle->locked = 1;
2898                         break;
2899                 }
2900                 if (cur == cpu)
2901                         break;
2902
2903                 cpu_relax();
2904         }
2905 }
2906
2907 static void perf_output_unlock(struct perf_output_handle *handle)
2908 {
2909         struct perf_mmap_data *data = handle->data;
2910         unsigned long head;
2911         int cpu;
2912
2913         data->done_head = data->head;
2914
2915         if (!handle->locked)
2916                 goto out;
2917
2918 again:
2919         /*
2920          * The xchg implies a full barrier that ensures all writes are done
2921          * before we publish the new head, matched by a rmb() in userspace when
2922          * reading this position.
2923          */
2924         while ((head = atomic_long_xchg(&data->done_head, 0)))
2925                 data->user_page->data_head = head;
2926
2927         /*
2928          * NMI can happen here, which means we can miss a done_head update.
2929          */
2930
2931         cpu = atomic_xchg(&data->lock, -1);
2932         WARN_ON_ONCE(cpu != smp_processor_id());
2933
2934         /*
2935          * Therefore we have to validate we did not indeed do so.
2936          */
2937         if (unlikely(atomic_long_read(&data->done_head))) {
2938                 /*
2939                  * Since we had it locked, we can lock it again.
2940                  */
2941                 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2942                         cpu_relax();
2943
2944                 goto again;
2945         }
2946
2947         if (atomic_xchg(&data->wakeup, 0))
2948                 perf_output_wakeup(handle);
2949 out:
2950         put_cpu();
2951 }
2952
2953 void perf_output_copy(struct perf_output_handle *handle,
2954                       const void *buf, unsigned int len)
2955 {
2956         unsigned int pages_mask;
2957         unsigned long offset;
2958         unsigned int size;
2959         void **pages;
2960
2961         offset          = handle->offset;
2962         pages_mask      = handle->data->nr_pages - 1;
2963         pages           = handle->data->data_pages;
2964
2965         do {
2966                 unsigned long page_offset;
2967                 unsigned long page_size;
2968                 int nr;
2969
2970                 nr          = (offset >> PAGE_SHIFT) & pages_mask;
2971                 page_size   = 1UL << (handle->data->data_order + PAGE_SHIFT);
2972                 page_offset = offset & (page_size - 1);
2973                 size        = min_t(unsigned int, page_size - page_offset, len);
2974
2975                 memcpy(pages[nr] + page_offset, buf, size);
2976
2977                 len         -= size;
2978                 buf         += size;
2979                 offset      += size;
2980         } while (len);
2981
2982         handle->offset = offset;
2983
2984         /*
2985          * Check we didn't copy past our reservation window, taking the
2986          * possible unsigned int wrap into account.
2987          */
2988         WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2989 }
2990
2991 int perf_output_begin(struct perf_output_handle *handle,
2992                       struct perf_event *event, unsigned int size,
2993                       int nmi, int sample)
2994 {
2995         struct perf_event *output_event;
2996         struct perf_mmap_data *data;
2997         unsigned long tail, offset, head;
2998         int have_lost;
2999         struct {
3000                 struct perf_event_header header;
3001                 u64                      id;
3002                 u64                      lost;
3003         } lost_event;
3004
3005         rcu_read_lock();
3006         /*
3007          * For inherited events we send all the output towards the parent.
3008          */
3009         if (event->parent)
3010                 event = event->parent;
3011
3012         output_event = rcu_dereference(event->output);
3013         if (output_event)
3014                 event = output_event;
3015
3016         data = rcu_dereference(event->data);
3017         if (!data)
3018                 goto out;
3019
3020         handle->data    = data;
3021         handle->event   = event;
3022         handle->nmi     = nmi;
3023         handle->sample  = sample;
3024
3025         if (!data->nr_pages)
3026                 goto fail;
3027
3028         have_lost = atomic_read(&data->lost);
3029         if (have_lost)
3030                 size += sizeof(lost_event);
3031
3032         perf_output_lock(handle);
3033
3034         do {
3035                 /*
3036                  * Userspace could choose to issue a mb() before updating the
3037                  * tail pointer. So that all reads will be completed before the
3038                  * write is issued.
3039                  */
3040                 tail = ACCESS_ONCE(data->user_page->data_tail);
3041                 smp_rmb();
3042                 offset = head = atomic_long_read(&data->head);
3043                 head += size;
3044                 if (unlikely(!perf_output_space(data, tail, offset, head)))
3045                         goto fail;
3046         } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
3047
3048         handle->offset  = offset;
3049         handle->head    = head;
3050
3051         if (head - tail > data->watermark)
3052                 atomic_set(&data->wakeup, 1);
3053
3054         if (have_lost) {
3055                 lost_event.header.type = PERF_RECORD_LOST;
3056                 lost_event.header.misc = 0;
3057                 lost_event.header.size = sizeof(lost_event);
3058                 lost_event.id          = event->id;
3059                 lost_event.lost        = atomic_xchg(&data->lost, 0);
3060
3061                 perf_output_put(handle, lost_event);
3062         }
3063
3064         return 0;
3065
3066 fail:
3067         atomic_inc(&data->lost);
3068         perf_output_unlock(handle);
3069 out:
3070         rcu_read_unlock();
3071
3072         return -ENOSPC;
3073 }
3074
3075 void perf_output_end(struct perf_output_handle *handle)
3076 {
3077         struct perf_event *event = handle->event;
3078         struct perf_mmap_data *data = handle->data;
3079
3080         int wakeup_events = event->attr.wakeup_events;
3081
3082         if (handle->sample && wakeup_events) {
3083                 int events = atomic_inc_return(&data->events);
3084                 if (events >= wakeup_events) {
3085                         atomic_sub(wakeup_events, &data->events);
3086                         atomic_set(&data->wakeup, 1);
3087                 }
3088         }
3089
3090         perf_output_unlock(handle);
3091         rcu_read_unlock();
3092 }
3093
3094 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3095 {
3096         /*
3097          * only top level events have the pid namespace they were created in
3098          */
3099         if (event->parent)
3100                 event = event->parent;
3101
3102         return task_tgid_nr_ns(p, event->ns);
3103 }
3104
3105 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3106 {
3107         /*
3108          * only top level events have the pid namespace they were created in
3109          */
3110         if (event->parent)
3111                 event = event->parent;
3112
3113         return task_pid_nr_ns(p, event->ns);
3114 }
3115
3116 static void perf_output_read_one(struct perf_output_handle *handle,
3117                                  struct perf_event *event)
3118 {
3119         u64 read_format = event->attr.read_format;
3120         u64 values[4];
3121         int n = 0;
3122
3123         values[n++] = atomic64_read(&event->count);
3124         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3125                 values[n++] = event->total_time_enabled +
3126                         atomic64_read(&event->child_total_time_enabled);
3127         }
3128         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3129                 values[n++] = event->total_time_running +
3130                         atomic64_read(&event->child_total_time_running);
3131         }
3132         if (read_format & PERF_FORMAT_ID)
3133                 values[n++] = primary_event_id(event);
3134
3135         perf_output_copy(handle, values, n * sizeof(u64));
3136 }
3137
3138 /*
3139  * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3140  */
3141 static void perf_output_read_group(struct perf_output_handle *handle,
3142                             struct perf_event *event)
3143 {
3144         struct perf_event *leader = event->group_leader, *sub;
3145         u64 read_format = event->attr.read_format;
3146         u64 values[5];
3147         int n = 0;
3148
3149         values[n++] = 1 + leader->nr_siblings;
3150
3151         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3152                 values[n++] = leader->total_time_enabled;
3153
3154         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3155                 values[n++] = leader->total_time_running;
3156
3157         if (leader != event)
3158                 leader->pmu->read(leader);
3159
3160         values[n++] = atomic64_read(&leader->count);
3161         if (read_format & PERF_FORMAT_ID)
3162                 values[n++] = primary_event_id(leader);
3163
3164         perf_output_copy(handle, values, n * sizeof(u64));
3165
3166         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3167                 n = 0;
3168
3169                 if (sub != event)
3170                         sub->pmu->read(sub);
3171
3172                 values[n++] = atomic64_read(&sub->count);
3173                 if (read_format & PERF_FORMAT_ID)
3174                         values[n++] = primary_event_id(sub);
3175
3176                 perf_output_copy(handle, values, n * sizeof(u64));
3177         }
3178 }
3179
3180 static void perf_output_read(struct perf_output_handle *handle,
3181                              struct perf_event *event)
3182 {
3183         if (event->attr.read_format & PERF_FORMAT_GROUP)
3184                 perf_output_read_group(handle, event);
3185         else
3186                 perf_output_read_one(handle, event);
3187 }
3188
3189 void perf_output_sample(struct perf_output_handle *handle,
3190                         struct perf_event_header *header,
3191                         struct perf_sample_data *data,
3192                         struct perf_event *event)
3193 {
3194         u64 sample_type = data->type;
3195
3196         perf_output_put(handle, *header);
3197
3198         if (sample_type & PERF_SAMPLE_IP)
3199                 perf_output_put(handle, data->ip);
3200
3201         if (sample_type & PERF_SAMPLE_TID)
3202                 perf_output_put(handle, data->tid_entry);
3203
3204         if (sample_type & PERF_SAMPLE_TIME)
3205                 perf_output_put(handle, data->time);
3206
3207         if (sample_type & PERF_SAMPLE_ADDR)
3208                 perf_output_put(handle, data->addr);
3209
3210         if (sample_type & PERF_SAMPLE_ID)
3211                 perf_output_put(handle, data->id);
3212
3213         if (sample_type & PERF_SAMPLE_STREAM_ID)
3214                 perf_output_put(handle, data->stream_id);
3215
3216         if (sample_type & PERF_SAMPLE_CPU)
3217                 perf_output_put(handle, data->cpu_entry);
3218
3219         if (sample_type & PERF_SAMPLE_PERIOD)
3220                 perf_output_put(handle, data->period);
3221
3222         if (sample_type & PERF_SAMPLE_READ)
3223                 perf_output_read(handle, event);
3224
3225         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3226                 if (data->callchain) {
3227                         int size = 1;
3228
3229                         if (data->callchain)
3230                                 size += data->callchain->nr;
3231
3232                         size *= sizeof(u64);
3233
3234                         perf_output_copy(handle, data->callchain, size);
3235                 } else {
3236                         u64 nr = 0;
3237                         perf_output_put(handle, nr);
3238                 }
3239         }
3240
3241         if (sample_type & PERF_SAMPLE_RAW) {
3242                 if (data->raw) {
3243                         perf_output_put(handle, data->raw->size);
3244                         perf_output_copy(handle, data->raw->data,
3245                                          data->raw->size);
3246                 } else {
3247                         struct {
3248                                 u32     size;
3249                                 u32     data;
3250                         } raw = {
3251                                 .size = sizeof(u32),
3252                                 .data = 0,
3253                         };
3254                         perf_output_put(handle, raw);
3255                 }
3256         }
3257 }
3258
3259 void perf_prepare_sample(struct perf_event_header *header,
3260                          struct perf_sample_data *data,
3261                          struct perf_event *event,
3262                          struct pt_regs *regs)
3263 {
3264         u64 sample_type = event->attr.sample_type;
3265
3266         data->type = sample_type;
3267
3268         header->type = PERF_RECORD_SAMPLE;
3269         header->size = sizeof(*header);
3270
3271         header->misc = 0;
3272         header->misc |= perf_misc_flags(regs);
3273
3274         if (sample_type & PERF_SAMPLE_IP) {
3275                 data->ip = perf_instruction_pointer(regs);
3276
3277                 header->size += sizeof(data->ip);
3278         }
3279
3280         if (sample_type & PERF_SAMPLE_TID) {
3281                 /* namespace issues */
3282                 data->tid_entry.pid = perf_event_pid(event, current);
3283                 data->tid_entry.tid = perf_event_tid(event, current);
3284
3285                 header->size += sizeof(data->tid_entry);
3286         }
3287
3288         if (sample_type & PERF_SAMPLE_TIME) {
3289                 data->time = perf_clock();
3290
3291                 header->size += sizeof(data->time);
3292         }
3293
3294         if (sample_type & PERF_SAMPLE_ADDR)
3295                 header->size += sizeof(data->addr);
3296
3297         if (sample_type & PERF_SAMPLE_ID) {
3298                 data->id = primary_event_id(event);
3299
3300                 header->size += sizeof(data->id);
3301         }
3302
3303         if (sample_type & PERF_SAMPLE_STREAM_ID) {
3304                 data->stream_id = event->id;
3305
3306                 header->size += sizeof(data->stream_id);
3307         }
3308
3309         if (sample_type & PERF_SAMPLE_CPU) {
3310                 data->cpu_entry.cpu             = raw_smp_processor_id();
3311                 data->cpu_entry.reserved        = 0;
3312
3313                 header->size += sizeof(data->cpu_entry);
3314         }
3315
3316         if (sample_type & PERF_SAMPLE_PERIOD)
3317                 header->size += sizeof(data->period);
3318
3319         if (sample_type & PERF_SAMPLE_READ)
3320                 header->size += perf_event_read_size(event);
3321
3322         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3323                 int size = 1;
3324
3325                 data->callchain = perf_callchain(regs);
3326
3327                 if (data->callchain)
3328                         size += data->callchain->nr;
3329
3330                 header->size += size * sizeof(u64);
3331         }
3332
3333         if (sample_type & PERF_SAMPLE_RAW) {
3334                 int size = sizeof(u32);
3335
3336                 if (data->raw)
3337                         size += data->raw->size;
3338                 else
3339                         size += sizeof(u32);
3340
3341                 WARN_ON_ONCE(size & (sizeof(u64)-1));
3342                 header->size += size;
3343         }
3344 }
3345
3346 static void perf_event_output(struct perf_event *event, int nmi,
3347                                 struct perf_sample_data *data,
3348                                 struct pt_regs *regs)
3349 {
3350         struct perf_output_handle handle;
3351         struct perf_event_header header;
3352
3353         perf_prepare_sample(&header, data, event, regs);
3354
3355         if (perf_output_begin(&handle, event, header.size, nmi, 1))
3356                 return;
3357
3358         perf_output_sample(&handle, &header, data, event);
3359
3360         perf_output_end(&handle);
3361 }
3362
3363 /*
3364  * read event_id
3365  */
3366
3367 struct perf_read_event {
3368         struct perf_event_header        header;
3369
3370         u32                             pid;
3371         u32                             tid;
3372 };
3373
3374 static void
3375 perf_event_read_event(struct perf_event *event,
3376                         struct task_struct *task)
3377 {
3378         struct perf_output_handle handle;
3379         struct perf_read_event read_event = {
3380                 .header = {
3381                         .type = PERF_RECORD_READ,
3382                         .misc = 0,
3383                         .size = sizeof(read_event) + perf_event_read_size(event),
3384                 },
3385                 .pid = perf_event_pid(event, task),
3386                 .tid = perf_event_tid(event, task),
3387         };
3388         int ret;
3389
3390         ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3391         if (ret)
3392                 return;
3393
3394         perf_output_put(&handle, read_event);
3395         perf_output_read(&handle, event);
3396
3397         perf_output_end(&handle);
3398 }
3399
3400 /*
3401  * task tracking -- fork/exit
3402  *
3403  * enabled by: attr.comm | attr.mmap | attr.task
3404  */
3405
3406 struct perf_task_event {
3407         struct task_struct              *task;
3408         struct perf_event_context       *task_ctx;
3409
3410         struct {
3411                 struct perf_event_header        header;
3412
3413                 u32                             pid;
3414                 u32                             ppid;
3415                 u32                             tid;
3416                 u32                             ptid;
3417                 u64                             time;
3418         } event_id;
3419 };
3420
3421 static void perf_event_task_output(struct perf_event *event,
3422                                      struct perf_task_event *task_event)
3423 {
3424         struct perf_output_handle handle;
3425         struct task_struct *task = task_event->task;
3426         unsigned long flags;
3427         int size, ret;
3428
3429         /*
3430          * If this CPU attempts to acquire an rq lock held by a CPU spinning
3431          * in perf_output_lock() from interrupt context, it's game over.
3432          */
3433         local_irq_save(flags);
3434
3435         size  = task_event->event_id.header.size;
3436         ret = perf_output_begin(&handle, event, size, 0, 0);
3437
3438         if (ret) {
3439                 local_irq_restore(flags);
3440                 return;
3441         }
3442
3443         task_event->event_id.pid = perf_event_pid(event, task);
3444         task_event->event_id.ppid = perf_event_pid(event, current);
3445
3446         task_event->event_id.tid = perf_event_tid(event, task);
3447         task_event->event_id.ptid = perf_event_tid(event, current);
3448
3449         perf_output_put(&handle, task_event->event_id);
3450
3451         perf_output_end(&handle);
3452         local_irq_restore(flags);
3453 }
3454
3455 static int perf_event_task_match(struct perf_event *event)
3456 {
3457         if (event->state < PERF_EVENT_STATE_INACTIVE)
3458                 return 0;
3459
3460         if (event->cpu != -1 && event->cpu != smp_processor_id())
3461                 return 0;
3462
3463         if (event->attr.comm || event->attr.mmap || event->attr.task)
3464                 return 1;
3465
3466         return 0;
3467 }
3468
3469 static void perf_event_task_ctx(struct perf_event_context *ctx,
3470                                   struct perf_task_event *task_event)
3471 {
3472         struct perf_event *event;
3473
3474         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3475                 if (perf_event_task_match(event))
3476                         perf_event_task_output(event, task_event);
3477         }
3478 }
3479
3480 static void perf_event_task_event(struct perf_task_event *task_event)
3481 {
3482         struct perf_cpu_context *cpuctx;
3483         struct perf_event_context *ctx = task_event->task_ctx;
3484
3485         rcu_read_lock();
3486         cpuctx = &get_cpu_var(perf_cpu_context);
3487         perf_event_task_ctx(&cpuctx->ctx, task_event);
3488         if (!ctx)
3489                 ctx = rcu_dereference(current->perf_event_ctxp);
3490         if (ctx)
3491                 perf_event_task_ctx(ctx, task_event);
3492         put_cpu_var(perf_cpu_context);
3493         rcu_read_unlock();
3494 }
3495
3496 static void perf_event_task(struct task_struct *task,
3497                               struct perf_event_context *task_ctx,
3498                               int new)
3499 {
3500         struct perf_task_event task_event;
3501
3502         if (!atomic_read(&nr_comm_events) &&
3503             !atomic_read(&nr_mmap_events) &&
3504             !atomic_read(&nr_task_events))
3505                 return;
3506
3507         task_event = (struct perf_task_event){
3508                 .task     = task,
3509                 .task_ctx = task_ctx,
3510                 .event_id    = {
3511                         .header = {
3512                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3513                                 .misc = 0,
3514                                 .size = sizeof(task_event.event_id),
3515                         },
3516                         /* .pid  */
3517                         /* .ppid */
3518                         /* .tid  */
3519                         /* .ptid */
3520                         .time = perf_clock(),
3521                 },
3522         };
3523
3524         perf_event_task_event(&task_event);
3525 }
3526
3527 void perf_event_fork(struct task_struct *task)
3528 {
3529         perf_event_task(task, NULL, 1);
3530 }
3531
3532 /*
3533  * comm tracking
3534  */
3535
3536 struct perf_comm_event {
3537         struct task_struct      *task;
3538         char                    *comm;
3539         int                     comm_size;
3540
3541         struct {
3542                 struct perf_event_header        header;
3543
3544                 u32                             pid;
3545                 u32                             tid;
3546         } event_id;
3547 };
3548
3549 static void perf_event_comm_output(struct perf_event *event,
3550                                      struct perf_comm_event *comm_event)
3551 {
3552         struct perf_output_handle handle;
3553         int size = comm_event->event_id.header.size;
3554         int ret = perf_output_begin(&handle, event, size, 0, 0);
3555
3556         if (ret)
3557                 return;
3558
3559         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3560         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3561
3562         perf_output_put(&handle, comm_event->event_id);
3563         perf_output_copy(&handle, comm_event->comm,
3564                                    comm_event->comm_size);
3565         perf_output_end(&handle);
3566 }
3567
3568 static int perf_event_comm_match(struct perf_event *event)
3569 {
3570         if (event->state < PERF_EVENT_STATE_INACTIVE)
3571                 return 0;
3572
3573         if (event->cpu != -1 && event->cpu != smp_processor_id())
3574                 return 0;
3575
3576         if (event->attr.comm)
3577                 return 1;
3578
3579         return 0;
3580 }
3581
3582 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3583                                   struct perf_comm_event *comm_event)
3584 {
3585         struct perf_event *event;
3586
3587         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3588                 if (perf_event_comm_match(event))
3589                         perf_event_comm_output(event, comm_event);
3590         }
3591 }
3592
3593 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3594 {
3595         struct perf_cpu_context *cpuctx;
3596         struct perf_event_context *ctx;
3597         unsigned int size;
3598         char comm[TASK_COMM_LEN];
3599
3600         memset(comm, 0, sizeof(comm));
3601         strlcpy(comm, comm_event->task->comm, sizeof(comm));
3602         size = ALIGN(strlen(comm)+1, sizeof(u64));
3603
3604         comm_event->comm = comm;
3605         comm_event->comm_size = size;
3606
3607         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3608
3609         rcu_read_lock();
3610         cpuctx = &get_cpu_var(perf_cpu_context);
3611         perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3612         ctx = rcu_dereference(current->perf_event_ctxp);
3613         if (ctx)
3614                 perf_event_comm_ctx(ctx, comm_event);
3615         put_cpu_var(perf_cpu_context);
3616         rcu_read_unlock();
3617 }
3618
3619 void perf_event_comm(struct task_struct *task)
3620 {
3621         struct perf_comm_event comm_event;
3622
3623         if (task->perf_event_ctxp)
3624                 perf_event_enable_on_exec(task);
3625
3626         if (!atomic_read(&nr_comm_events))
3627                 return;
3628
3629         comm_event = (struct perf_comm_event){
3630                 .task   = task,
3631                 /* .comm      */
3632                 /* .comm_size */
3633                 .event_id  = {
3634                         .header = {
3635                                 .type = PERF_RECORD_COMM,
3636                                 .misc = 0,
3637                                 /* .size */
3638                         },
3639                         /* .pid */
3640                         /* .tid */
3641                 },
3642         };
3643
3644         perf_event_comm_event(&comm_event);
3645 }
3646
3647 /*
3648  * mmap tracking
3649  */
3650
3651 struct perf_mmap_event {
3652         struct vm_area_struct   *vma;
3653
3654         const char              *file_name;
3655         int                     file_size;
3656
3657         struct {
3658                 struct perf_event_header        header;
3659
3660                 u32                             pid;
3661                 u32                             tid;
3662                 u64                             start;
3663                 u64                             len;
3664                 u64                             pgoff;
3665         } event_id;
3666 };
3667
3668 static void perf_event_mmap_output(struct perf_event *event,
3669                                      struct perf_mmap_event *mmap_event)
3670 {
3671         struct perf_output_handle handle;
3672         int size = mmap_event->event_id.header.size;
3673         int ret = perf_output_begin(&handle, event, size, 0, 0);
3674
3675         if (ret)
3676                 return;
3677
3678         mmap_event->event_id.pid = perf_event_pid(event, current);
3679         mmap_event->event_id.tid = perf_event_tid(event, current);
3680
3681         perf_output_put(&handle, mmap_event->event_id);
3682         perf_output_copy(&handle, mmap_event->file_name,
3683                                    mmap_event->file_size);
3684         perf_output_end(&handle);
3685 }
3686
3687 static int perf_event_mmap_match(struct perf_event *event,
3688                                    struct perf_mmap_event *mmap_event)
3689 {
3690         if (event->state < PERF_EVENT_STATE_INACTIVE)
3691                 return 0;
3692
3693         if (event->cpu != -1 && event->cpu != smp_processor_id())
3694                 return 0;
3695
3696         if (event->attr.mmap)
3697                 return 1;
3698
3699         return 0;
3700 }
3701
3702 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3703                                   struct perf_mmap_event *mmap_event)
3704 {
3705         struct perf_event *event;
3706
3707         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3708                 if (perf_event_mmap_match(event, mmap_event))
3709                         perf_event_mmap_output(event, mmap_event);
3710         }
3711 }
3712
3713 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3714 {
3715         struct perf_cpu_context *cpuctx;
3716         struct perf_event_context *ctx;
3717         struct vm_area_struct *vma = mmap_event->vma;
3718         struct file *file = vma->vm_file;
3719         unsigned int size;
3720         char tmp[16];
3721         char *buf = NULL;
3722         const char *name;
3723
3724         memset(tmp, 0, sizeof(tmp));
3725
3726         if (file) {
3727                 /*
3728                  * d_path works from the end of the buffer backwards, so we
3729                  * need to add enough zero bytes after the string to handle
3730                  * the 64bit alignment we do later.
3731                  */
3732                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3733                 if (!buf) {
3734                         name = strncpy(tmp, "//enomem", sizeof(tmp));
3735                         goto got_name;
3736                 }
3737                 name = d_path(&file->f_path, buf, PATH_MAX);
3738                 if (IS_ERR(name)) {
3739                         name = strncpy(tmp, "//toolong", sizeof(tmp));
3740                         goto got_name;
3741                 }
3742         } else {
3743                 if (arch_vma_name(mmap_event->vma)) {
3744                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3745                                        sizeof(tmp));
3746                         goto got_name;
3747                 }
3748
3749                 if (!vma->vm_mm) {
3750                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
3751                         goto got_name;
3752                 }
3753
3754                 name = strncpy(tmp, "//anon", sizeof(tmp));
3755                 goto got_name;
3756         }
3757
3758 got_name:
3759         size = ALIGN(strlen(name)+1, sizeof(u64));
3760
3761         mmap_event->file_name = name;
3762         mmap_event->file_size = size;
3763
3764         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3765
3766         rcu_read_lock();
3767         cpuctx = &get_cpu_var(perf_cpu_context);
3768         perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3769         ctx = rcu_dereference(current->perf_event_ctxp);
3770         if (ctx)
3771                 perf_event_mmap_ctx(ctx, mmap_event);
3772         put_cpu_var(perf_cpu_context);
3773         rcu_read_unlock();
3774
3775         kfree(buf);
3776 }
3777
3778 void __perf_event_mmap(struct vm_area_struct *vma)
3779 {
3780         struct perf_mmap_event mmap_event;
3781
3782         if (!atomic_read(&nr_mmap_events))
3783                 return;
3784
3785         mmap_event = (struct perf_mmap_event){
3786                 .vma    = vma,
3787                 /* .file_name */
3788                 /* .file_size */
3789                 .event_id  = {
3790                         .header = {
3791                                 .type = PERF_RECORD_MMAP,
3792                                 .misc = PERF_RECORD_MISC_USER,
3793                                 /* .size */
3794                         },
3795                         /* .pid */
3796                         /* .tid */
3797                         .start  = vma->vm_start,
3798                         .len    = vma->vm_end - vma->vm_start,
3799                         .pgoff  = (u64)vma->vm_pgoff << PAGE_SHIFT,
3800                 },
3801         };
3802
3803         perf_event_mmap_event(&mmap_event);
3804 }
3805
3806 /*
3807  * IRQ throttle logging
3808  */
3809
3810 static void perf_log_throttle(struct perf_event *event, int enable)
3811 {
3812         struct perf_output_handle handle;
3813         int ret;
3814
3815         struct {
3816                 struct perf_event_header        header;
3817                 u64                             time;
3818                 u64                             id;
3819                 u64                             stream_id;
3820         } throttle_event = {
3821                 .header = {
3822                         .type = PERF_RECORD_THROTTLE,
3823                         .misc = 0,
3824                         .size = sizeof(throttle_event),
3825                 },
3826                 .time           = perf_clock(),
3827                 .id             = primary_event_id(event),
3828                 .stream_id      = event->id,
3829         };
3830
3831         if (enable)
3832                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3833
3834         ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3835         if (ret)
3836                 return;
3837
3838         perf_output_put(&handle, throttle_event);
3839         perf_output_end(&handle);
3840 }
3841
3842 /*
3843  * Generic event overflow handling, sampling.
3844  */
3845
3846 static int __perf_event_overflow(struct perf_event *event, int nmi,
3847                                    int throttle, struct perf_sample_data *data,
3848                                    struct pt_regs *regs)
3849 {
3850         int events = atomic_read(&event->event_limit);
3851         struct hw_perf_event *hwc = &event->hw;
3852         int ret = 0;
3853
3854         throttle = (throttle && event->pmu->unthrottle != NULL);
3855
3856         if (!throttle) {
3857                 hwc->interrupts++;
3858         } else {
3859                 if (hwc->interrupts != MAX_INTERRUPTS) {
3860                         hwc->interrupts++;
3861                         if (HZ * hwc->interrupts >
3862                                         (u64)sysctl_perf_event_sample_rate) {
3863                                 hwc->interrupts = MAX_INTERRUPTS;
3864                                 perf_log_throttle(event, 0);
3865                                 ret = 1;
3866                         }
3867                 } else {
3868                         /*
3869                          * Keep re-disabling events even though on the previous
3870                          * pass we disabled it - just in case we raced with a
3871                          * sched-in and the event got enabled again:
3872                          */
3873                         ret = 1;
3874                 }
3875         }
3876
3877         if (event->attr.freq) {
3878                 u64 now = perf_clock();
3879                 s64 delta = now - hwc->freq_time_stamp;
3880
3881                 hwc->freq_time_stamp = now;
3882
3883                 if (delta > 0 && delta < 2*TICK_NSEC)
3884                         perf_adjust_period(event, delta, hwc->last_period);
3885         }
3886
3887         /*
3888          * XXX event_limit might not quite work as expected on inherited
3889          * events
3890          */
3891
3892         event->pending_kill = POLL_IN;
3893         if (events && atomic_dec_and_test(&event->event_limit)) {
3894                 ret = 1;
3895                 event->pending_kill = POLL_HUP;
3896                 if (nmi) {
3897                         event->pending_disable = 1;
3898                         perf_pending_queue(&event->pending,
3899                                            perf_pending_event);
3900                 } else
3901                         perf_event_disable(event);
3902         }
3903
3904         if (event->overflow_handler)
3905                 event->overflow_handler(event, nmi, data, regs);
3906         else
3907                 perf_event_output(event, nmi, data, regs);
3908
3909         return ret;
3910 }
3911
3912 int perf_event_overflow(struct perf_event *event, int nmi,
3913                           struct perf_sample_data *data,
3914                           struct pt_regs *regs)
3915 {
3916         return __perf_event_overflow(event, nmi, 1, data, regs);
3917 }
3918
3919 /*
3920  * Generic software event infrastructure
3921  */
3922
3923 /*
3924  * We directly increment event->count and keep a second value in
3925  * event->hw.period_left to count intervals. This period event
3926  * is kept in the range [-sample_period, 0] so that we can use the
3927  * sign as trigger.
3928  */
3929
3930 static u64 perf_swevent_set_period(struct perf_event *event)
3931 {
3932         struct hw_perf_event *hwc = &event->hw;
3933         u64 period = hwc->last_period;
3934         u64 nr, offset;
3935         s64 old, val;
3936
3937         hwc->last_period = hwc->sample_period;
3938
3939 again:
3940         old = val = atomic64_read(&hwc->period_left);
3941         if (val < 0)
3942                 return 0;
3943
3944         nr = div64_u64(period + val, period);
3945         offset = nr * period;
3946         val -= offset;
3947         if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3948                 goto again;
3949
3950         return nr;
3951 }
3952
3953 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3954                                     int nmi, struct perf_sample_data *data,
3955                                     struct pt_regs *regs)
3956 {
3957         struct hw_perf_event *hwc = &event->hw;
3958         int throttle = 0;
3959
3960         data->period = event->hw.last_period;
3961         if (!overflow)
3962                 overflow = perf_swevent_set_period(event);
3963
3964         if (hwc->interrupts == MAX_INTERRUPTS)
3965                 return;
3966
3967         for (; overflow; overflow--) {
3968                 if (__perf_event_overflow(event, nmi, throttle,
3969                                             data, regs)) {
3970                         /*
3971                          * We inhibit the overflow from happening when
3972                          * hwc->interrupts == MAX_INTERRUPTS.
3973                          */
3974                         break;
3975                 }
3976                 throttle = 1;
3977         }
3978 }
3979
3980 static void perf_swevent_unthrottle(struct perf_event *event)
3981 {
3982         /*
3983          * Nothing to do, we already reset hwc->interrupts.
3984          */
3985 }
3986
3987 static void perf_swevent_add(struct perf_event *event, u64 nr,
3988                                int nmi, struct perf_sample_data *data,
3989                                struct pt_regs *regs)
3990 {
3991         struct hw_perf_event *hwc = &event->hw;
3992
3993         atomic64_add(nr, &event->count);
3994
3995         if (!regs)
3996                 return;
3997
3998         if (!hwc->sample_period)
3999                 return;
4000
4001         if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4002                 return perf_swevent_overflow(event, 1, nmi, data, regs);
4003
4004         if (atomic64_add_negative(nr, &hwc->period_left))
4005                 return;
4006
4007         perf_swevent_overflow(event, 0, nmi, data, regs);
4008 }
4009
4010 static int perf_tp_event_match(struct perf_event *event,
4011                                 struct perf_sample_data *data);
4012
4013 static int perf_exclude_event(struct perf_event *event,
4014                               struct pt_regs *regs)
4015 {
4016         if (regs) {
4017                 if (event->attr.exclude_user && user_mode(regs))
4018                         return 1;
4019
4020                 if (event->attr.exclude_kernel && !user_mode(regs))
4021                         return 1;
4022         }
4023
4024         return 0;
4025 }
4026
4027 static int perf_swevent_match(struct perf_event *event,
4028                                 enum perf_type_id type,
4029                                 u32 event_id,
4030                                 struct perf_sample_data *data,
4031                                 struct pt_regs *regs)
4032 {
4033         if (event->attr.type != type)
4034                 return 0;
4035
4036         if (event->attr.config != event_id)
4037                 return 0;
4038
4039         if (perf_exclude_event(event, regs))
4040                 return 0;
4041
4042         if (event->attr.type == PERF_TYPE_TRACEPOINT &&
4043             !perf_tp_event_match(event, data))
4044                 return 0;
4045
4046         return 1;
4047 }
4048
4049 static inline u64 swevent_hash(u64 type, u32 event_id)
4050 {
4051         u64 val = event_id | (type << 32);
4052
4053         return hash_64(val, SWEVENT_HLIST_BITS);
4054 }
4055
4056 static struct hlist_head *
4057 find_swevent_head(struct perf_cpu_context *ctx, u64 type, u32 event_id)
4058 {
4059         u64 hash;
4060         struct swevent_hlist *hlist;
4061
4062         hash = swevent_hash(type, event_id);
4063
4064         hlist = rcu_dereference(ctx->swevent_hlist);
4065         if (!hlist)
4066                 return NULL;
4067
4068         return &hlist->heads[hash];
4069 }
4070
4071 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4072                                     u64 nr, int nmi,
4073                                     struct perf_sample_data *data,
4074                                     struct pt_regs *regs)
4075 {
4076         struct perf_cpu_context *cpuctx;
4077         struct perf_event *event;
4078         struct hlist_node *node;
4079         struct hlist_head *head;
4080
4081         cpuctx = &__get_cpu_var(perf_cpu_context);
4082
4083         rcu_read_lock();
4084
4085         head = find_swevent_head(cpuctx, type, event_id);
4086
4087         if (!head)
4088                 goto end;
4089
4090         hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4091                 if (perf_swevent_match(event, type, event_id, data, regs))
4092                         perf_swevent_add(event, nr, nmi, data, regs);
4093         }
4094 end:
4095         rcu_read_unlock();
4096 }
4097
4098 int perf_swevent_get_recursion_context(void)
4099 {
4100         struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
4101         int rctx;
4102
4103         if (in_nmi())
4104                 rctx = 3;
4105         else if (in_irq())
4106                 rctx = 2;
4107         else if (in_softirq())
4108                 rctx = 1;
4109         else
4110                 rctx = 0;
4111
4112         if (cpuctx->recursion[rctx]) {
4113                 put_cpu_var(perf_cpu_context);
4114                 return -1;
4115         }
4116
4117         cpuctx->recursion[rctx]++;
4118         barrier();
4119
4120         return rctx;
4121 }
4122 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4123
4124 void perf_swevent_put_recursion_context(int rctx)
4125 {
4126         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4127         barrier();
4128         cpuctx->recursion[rctx]--;
4129         put_cpu_var(perf_cpu_context);
4130 }
4131 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
4132
4133
4134 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4135                             struct pt_regs *regs, u64 addr)
4136 {
4137         struct perf_sample_data data;
4138         int rctx;
4139
4140         rctx = perf_swevent_get_recursion_context();
4141         if (rctx < 0)
4142                 return;
4143
4144         perf_sample_data_init(&data, addr);
4145
4146         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4147
4148         perf_swevent_put_recursion_context(rctx);
4149 }
4150
4151 static void perf_swevent_read(struct perf_event *event)
4152 {
4153 }
4154
4155 static int perf_swevent_enable(struct perf_event *event)
4156 {
4157         struct hw_perf_event *hwc = &event->hw;
4158         struct perf_cpu_context *cpuctx;
4159         struct hlist_head *head;
4160
4161         cpuctx = &__get_cpu_var(perf_cpu_context);
4162
4163         if (hwc->sample_period) {
4164                 hwc->last_period = hwc->sample_period;
4165                 perf_swevent_set_period(event);
4166         }
4167
4168         head = find_swevent_head(cpuctx, event->attr.type, event->attr.config);
4169         if (WARN_ON_ONCE(!head))
4170                 return -EINVAL;
4171
4172         hlist_add_head_rcu(&event->hlist_entry, head);
4173
4174         return 0;
4175 }
4176
4177 static void perf_swevent_disable(struct perf_event *event)
4178 {
4179         hlist_del_rcu(&event->hlist_entry);
4180 }
4181
4182 static const struct pmu perf_ops_generic = {
4183         .enable         = perf_swevent_enable,
4184         .disable        = perf_swevent_disable,
4185         .read           = perf_swevent_read,
4186         .unthrottle     = perf_swevent_unthrottle,
4187 };
4188
4189 /*
4190  * hrtimer based swevent callback
4191  */
4192
4193 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4194 {
4195         enum hrtimer_restart ret = HRTIMER_RESTART;
4196         struct perf_sample_data data;
4197         struct pt_regs *regs;
4198         struct perf_event *event;
4199         u64 period;
4200
4201         event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4202         event->pmu->read(event);
4203
4204         perf_sample_data_init(&data, 0);
4205         data.period = event->hw.last_period;
4206         regs = get_irq_regs();
4207
4208         if (regs && !perf_exclude_event(event, regs)) {
4209                 if (!(event->attr.exclude_idle && current->pid == 0))
4210                         if (perf_event_overflow(event, 0, &data, regs))
4211                                 ret = HRTIMER_NORESTART;
4212         }
4213
4214         period = max_t(u64, 10000, event->hw.sample_period);
4215         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4216
4217         return ret;
4218 }
4219
4220 static void perf_swevent_start_hrtimer(struct perf_event *event)
4221 {
4222         struct hw_perf_event *hwc = &event->hw;
4223
4224         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4225         hwc->hrtimer.function = perf_swevent_hrtimer;
4226         if (hwc->sample_period) {
4227                 u64 period;
4228
4229                 if (hwc->remaining) {
4230                         if (hwc->remaining < 0)
4231                                 period = 10000;
4232                         else
4233                                 period = hwc->remaining;
4234                         hwc->remaining = 0;
4235                 } else {
4236                         period = max_t(u64, 10000, hwc->sample_period);
4237                 }
4238                 __hrtimer_start_range_ns(&hwc->hrtimer,
4239                                 ns_to_ktime(period), 0,
4240                                 HRTIMER_MODE_REL, 0);
4241         }
4242 }
4243
4244 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4245 {
4246         struct hw_perf_event *hwc = &event->hw;
4247
4248         if (hwc->sample_period) {
4249                 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4250                 hwc->remaining = ktime_to_ns(remaining);
4251
4252                 hrtimer_cancel(&hwc->hrtimer);
4253         }
4254 }
4255
4256 /*
4257  * Software event: cpu wall time clock
4258  */
4259
4260 static void cpu_clock_perf_event_update(struct perf_event *event)
4261 {
4262         int cpu = raw_smp_processor_id();
4263         s64 prev;
4264         u64 now;
4265
4266         now = cpu_clock(cpu);
4267         prev = atomic64_xchg(&event->hw.prev_count, now);
4268         atomic64_add(now - prev, &event->count);
4269 }
4270
4271 static int cpu_clock_perf_event_enable(struct perf_event *event)
4272 {
4273         struct hw_perf_event *hwc = &event->hw;
4274         int cpu = raw_smp_processor_id();
4275
4276         atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4277         perf_swevent_start_hrtimer(event);
4278
4279         return 0;
4280 }
4281
4282 static void cpu_clock_perf_event_disable(struct perf_event *event)
4283 {
4284         perf_swevent_cancel_hrtimer(event);
4285         cpu_clock_perf_event_update(event);
4286 }
4287
4288 static void cpu_clock_perf_event_read(struct perf_event *event)
4289 {
4290         cpu_clock_perf_event_update(event);
4291 }
4292
4293 static const struct pmu perf_ops_cpu_clock = {
4294         .enable         = cpu_clock_perf_event_enable,
4295         .disable        = cpu_clock_perf_event_disable,
4296         .read           = cpu_clock_perf_event_read,
4297 };
4298
4299 /*
4300  * Software event: task time clock
4301  */
4302
4303 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4304 {
4305         u64 prev;
4306         s64 delta;
4307
4308         prev = atomic64_xchg(&event->hw.prev_count, now);
4309         delta = now - prev;
4310         atomic64_add(delta, &event->count);
4311 }
4312
4313 static int task_clock_perf_event_enable(struct perf_event *event)
4314 {
4315         struct hw_perf_event *hwc = &event->hw;
4316         u64 now;
4317
4318         now = event->ctx->time;
4319
4320         atomic64_set(&hwc->prev_count, now);
4321
4322         perf_swevent_start_hrtimer(event);
4323
4324         return 0;
4325 }
4326
4327 static void task_clock_perf_event_disable(struct perf_event *event)
4328 {
4329         perf_swevent_cancel_hrtimer(event);
4330         task_clock_perf_event_update(event, event->ctx->time);
4331
4332 }
4333
4334 static void task_clock_perf_event_read(struct perf_event *event)
4335 {
4336         u64 time;
4337
4338         if (!in_nmi()) {
4339                 update_context_time(event->ctx);
4340                 time = event->ctx->time;
4341         } else {
4342                 u64 now = perf_clock();
4343                 u64 delta = now - event->ctx->timestamp;
4344                 time = event->ctx->time + delta;
4345         }
4346
4347         task_clock_perf_event_update(event, time);
4348 }
4349
4350 static const struct pmu perf_ops_task_clock = {
4351         .enable         = task_clock_perf_event_enable,
4352         .disable        = task_clock_perf_event_disable,
4353         .read           = task_clock_perf_event_read,
4354 };
4355
4356 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4357 {
4358         struct swevent_hlist *hlist;
4359
4360         hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4361         kfree(hlist);
4362 }
4363
4364 static void swevent_hlist_release(struct perf_cpu_context *cpuctx)
4365 {
4366         struct swevent_hlist *hlist;
4367
4368         if (!cpuctx->swevent_hlist)
4369                 return;
4370
4371         hlist = cpuctx->swevent_hlist;
4372         rcu_assign_pointer(cpuctx->swevent_hlist, NULL);
4373         call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4374 }
4375
4376 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4377 {
4378         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4379
4380         mutex_lock(&cpuctx->hlist_mutex);
4381
4382         if (!--cpuctx->hlist_refcount)
4383                 swevent_hlist_release(cpuctx);
4384
4385         mutex_unlock(&cpuctx->hlist_mutex);
4386 }
4387
4388 static void swevent_hlist_put(struct perf_event *event)
4389 {
4390         int cpu;
4391
4392         if (event->cpu != -1) {
4393                 swevent_hlist_put_cpu(event, event->cpu);
4394                 return;
4395         }
4396
4397         for_each_possible_cpu(cpu)
4398                 swevent_hlist_put_cpu(event, cpu);
4399 }
4400
4401 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4402 {
4403         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4404         int err = 0;
4405
4406         mutex_lock(&cpuctx->hlist_mutex);
4407
4408         if (!cpuctx->swevent_hlist && cpu_online(cpu)) {
4409                 struct swevent_hlist *hlist;
4410
4411                 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4412                 if (!hlist) {
4413                         err = -ENOMEM;
4414                         goto exit;
4415                 }
4416                 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
4417         }
4418         cpuctx->hlist_refcount++;
4419  exit:
4420         mutex_unlock(&cpuctx->hlist_mutex);
4421
4422         return err;
4423 }
4424
4425 static int swevent_hlist_get(struct perf_event *event)
4426 {
4427         int err;
4428         int cpu, failed_cpu;
4429
4430         if (event->cpu != -1)
4431                 return swevent_hlist_get_cpu(event, event->cpu);
4432
4433         get_online_cpus();
4434         for_each_possible_cpu(cpu) {
4435                 err = swevent_hlist_get_cpu(event, cpu);
4436                 if (err) {
4437                         failed_cpu = cpu;
4438                         goto fail;
4439                 }
4440         }
4441         put_online_cpus();
4442
4443         return 0;
4444  fail:
4445         for_each_possible_cpu(cpu) {
4446                 if (cpu == failed_cpu)
4447                         break;
4448                 swevent_hlist_put_cpu(event, cpu);
4449         }
4450
4451         put_online_cpus();
4452         return err;
4453 }
4454
4455 #ifdef CONFIG_EVENT_TRACING
4456
4457 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4458                    int entry_size, struct pt_regs *regs)
4459 {
4460         struct perf_sample_data data;
4461         struct perf_raw_record raw = {
4462                 .size = entry_size,
4463                 .data = record,
4464         };
4465
4466         perf_sample_data_init(&data, addr);
4467         data.raw = &raw;
4468
4469         /* Trace events already protected against recursion */
4470         do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4471                          &data, regs);
4472 }
4473 EXPORT_SYMBOL_GPL(perf_tp_event);
4474
4475 static int perf_tp_event_match(struct perf_event *event,
4476                                 struct perf_sample_data *data)
4477 {
4478         void *record = data->raw->data;
4479
4480         if (likely(!event->filter) || filter_match_preds(event->filter, record))
4481                 return 1;
4482         return 0;
4483 }
4484
4485 static void tp_perf_event_destroy(struct perf_event *event)
4486 {
4487         perf_trace_disable(event->attr.config);
4488         swevent_hlist_put(event);
4489 }
4490
4491 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4492 {
4493         int err;
4494
4495         /*
4496          * Raw tracepoint data is a severe data leak, only allow root to
4497          * have these.
4498          */
4499         if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4500                         perf_paranoid_tracepoint_raw() &&
4501                         !capable(CAP_SYS_ADMIN))
4502                 return ERR_PTR(-EPERM);
4503
4504         if (perf_trace_enable(event->attr.config))
4505                 return NULL;
4506
4507         event->destroy = tp_perf_event_destroy;
4508         err = swevent_hlist_get(event);
4509         if (err) {
4510                 perf_trace_disable(event->attr.config);
4511                 return ERR_PTR(err);
4512         }
4513
4514         return &perf_ops_generic;
4515 }
4516
4517 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4518 {
4519         char *filter_str;
4520         int ret;
4521
4522         if (event->attr.type != PERF_TYPE_TRACEPOINT)
4523                 return -EINVAL;
4524
4525         filter_str = strndup_user(arg, PAGE_SIZE);
4526         if (IS_ERR(filter_str))
4527                 return PTR_ERR(filter_str);
4528
4529         ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4530
4531         kfree(filter_str);
4532         return ret;
4533 }
4534
4535 static void perf_event_free_filter(struct perf_event *event)
4536 {
4537         ftrace_profile_free_filter(event);
4538 }
4539
4540 #else
4541
4542 static int perf_tp_event_match(struct perf_event *event,
4543                                 struct perf_sample_data *data)
4544 {
4545         return 1;
4546 }
4547
4548 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4549 {
4550         return NULL;
4551 }
4552
4553 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4554 {
4555         return -ENOENT;
4556 }
4557
4558 static void perf_event_free_filter(struct perf_event *event)
4559 {
4560 }
4561
4562 #endif /* CONFIG_EVENT_TRACING */
4563
4564 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4565 static void bp_perf_event_destroy(struct perf_event *event)
4566 {
4567         release_bp_slot(event);
4568 }
4569
4570 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4571 {
4572         int err;
4573
4574         err = register_perf_hw_breakpoint(bp);
4575         if (err)
4576                 return ERR_PTR(err);
4577
4578         bp->destroy = bp_perf_event_destroy;
4579
4580         return &perf_ops_bp;
4581 }
4582
4583 void perf_bp_event(struct perf_event *bp, void *data)
4584 {
4585         struct perf_sample_data sample;
4586         struct pt_regs *regs = data;
4587
4588         perf_sample_data_init(&sample, bp->attr.bp_addr);
4589
4590         if (!perf_exclude_event(bp, regs))
4591                 perf_swevent_add(bp, 1, 1, &sample, regs);
4592 }
4593 #else
4594 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4595 {
4596         return NULL;
4597 }
4598
4599 void perf_bp_event(struct perf_event *bp, void *regs)
4600 {
4601 }
4602 #endif
4603
4604 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4605
4606 static void sw_perf_event_destroy(struct perf_event *event)
4607 {
4608         u64 event_id = event->attr.config;
4609
4610         WARN_ON(event->parent);
4611
4612         atomic_dec(&perf_swevent_enabled[event_id]);
4613         swevent_hlist_put(event);
4614 }
4615
4616 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4617 {
4618         const struct pmu *pmu = NULL;
4619         u64 event_id = event->attr.config;
4620
4621         /*
4622          * Software events (currently) can't in general distinguish
4623          * between user, kernel and hypervisor events.
4624          * However, context switches and cpu migrations are considered
4625          * to be kernel events, and page faults are never hypervisor
4626          * events.
4627          */
4628         switch (event_id) {
4629         case PERF_COUNT_SW_CPU_CLOCK:
4630                 pmu = &perf_ops_cpu_clock;
4631
4632                 break;
4633         case PERF_COUNT_SW_TASK_CLOCK:
4634                 /*
4635                  * If the user instantiates this as a per-cpu event,
4636                  * use the cpu_clock event instead.
4637                  */
4638                 if (event->ctx->task)
4639                         pmu = &perf_ops_task_clock;
4640                 else
4641                         pmu = &perf_ops_cpu_clock;
4642
4643                 break;
4644         case PERF_COUNT_SW_PAGE_FAULTS:
4645         case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4646         case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4647         case PERF_COUNT_SW_CONTEXT_SWITCHES:
4648         case PERF_COUNT_SW_CPU_MIGRATIONS:
4649         case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4650         case PERF_COUNT_SW_EMULATION_FAULTS:
4651                 if (!event->parent) {
4652                         int err;
4653
4654                         err = swevent_hlist_get(event);
4655                         if (err)
4656                                 return ERR_PTR(err);
4657
4658                         atomic_inc(&perf_swevent_enabled[event_id]);
4659                         event->destroy = sw_perf_event_destroy;
4660                 }
4661                 pmu = &perf_ops_generic;
4662                 break;
4663         }
4664
4665         return pmu;
4666 }
4667
4668 /*
4669  * Allocate and initialize a event structure
4670  */
4671 static struct perf_event *
4672 perf_event_alloc(struct perf_event_attr *attr,
4673                    int cpu,
4674                    struct perf_event_context *ctx,
4675                    struct perf_event *group_leader,
4676                    struct perf_event *parent_event,
4677                    perf_overflow_handler_t overflow_handler,
4678                    gfp_t gfpflags)
4679 {
4680         const struct pmu *pmu;
4681         struct perf_event *event;
4682         struct hw_perf_event *hwc;
4683         long err;
4684
4685         event = kzalloc(sizeof(*event), gfpflags);
4686         if (!event)
4687                 return ERR_PTR(-ENOMEM);
4688
4689         /*
4690          * Single events are their own group leaders, with an
4691          * empty sibling list:
4692          */
4693         if (!group_leader)
4694                 group_leader = event;
4695
4696         mutex_init(&event->child_mutex);
4697         INIT_LIST_HEAD(&event->child_list);
4698
4699         INIT_LIST_HEAD(&event->group_entry);
4700         INIT_LIST_HEAD(&event->event_entry);
4701         INIT_LIST_HEAD(&event->sibling_list);
4702         init_waitqueue_head(&event->waitq);
4703
4704         mutex_init(&event->mmap_mutex);
4705
4706         event->cpu              = cpu;
4707         event->attr             = *attr;
4708         event->group_leader     = group_leader;
4709         event->pmu              = NULL;
4710         event->ctx              = ctx;
4711         event->oncpu            = -1;
4712
4713         event->parent           = parent_event;
4714
4715         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
4716         event->id               = atomic64_inc_return(&perf_event_id);
4717
4718         event->state            = PERF_EVENT_STATE_INACTIVE;
4719
4720         if (!overflow_handler && parent_event)
4721                 overflow_handler = parent_event->overflow_handler;
4722         
4723         event->overflow_handler = overflow_handler;
4724
4725         if (attr->disabled)
4726                 event->state = PERF_EVENT_STATE_OFF;
4727
4728         pmu = NULL;
4729
4730         hwc = &event->hw;
4731         hwc->sample_period = attr->sample_period;
4732         if (attr->freq && attr->sample_freq)
4733                 hwc->sample_period = 1;
4734         hwc->last_period = hwc->sample_period;
4735
4736         atomic64_set(&hwc->period_left, hwc->sample_period);
4737
4738         /*
4739          * we currently do not support PERF_FORMAT_GROUP on inherited events
4740          */
4741         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4742                 goto done;
4743
4744         switch (attr->type) {
4745         case PERF_TYPE_RAW:
4746         case PERF_TYPE_HARDWARE:
4747         case PERF_TYPE_HW_CACHE:
4748                 pmu = hw_perf_event_init(event);
4749                 break;
4750
4751         case PERF_TYPE_SOFTWARE:
4752                 pmu = sw_perf_event_init(event);
4753                 break;
4754
4755         case PERF_TYPE_TRACEPOINT:
4756                 pmu = tp_perf_event_init(event);
4757                 break;
4758
4759         case PERF_TYPE_BREAKPOINT:
4760                 pmu = bp_perf_event_init(event);
4761                 break;
4762
4763
4764         default:
4765                 break;
4766         }
4767 done:
4768         err = 0;
4769         if (!pmu)
4770                 err = -EINVAL;
4771         else if (IS_ERR(pmu))
4772                 err = PTR_ERR(pmu);
4773
4774         if (err) {
4775                 if (event->ns)
4776                         put_pid_ns(event->ns);
4777                 kfree(event);
4778                 return ERR_PTR(err);
4779         }
4780
4781         event->pmu = pmu;
4782
4783         if (!event->parent) {
4784                 atomic_inc(&nr_events);
4785                 if (event->attr.mmap)
4786                         atomic_inc(&nr_mmap_events);
4787                 if (event->attr.comm)
4788                         atomic_inc(&nr_comm_events);
4789                 if (event->attr.task)
4790                         atomic_inc(&nr_task_events);
4791         }
4792
4793         return event;
4794 }
4795
4796 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4797                           struct perf_event_attr *attr)
4798 {
4799         u32 size;
4800         int ret;
4801
4802         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4803                 return -EFAULT;
4804
4805         /*
4806          * zero the full structure, so that a short copy will be nice.
4807          */
4808         memset(attr, 0, sizeof(*attr));
4809
4810         ret = get_user(size, &uattr->size);
4811         if (ret)
4812                 return ret;
4813
4814         if (size > PAGE_SIZE)   /* silly large */
4815                 goto err_size;
4816
4817         if (!size)              /* abi compat */
4818                 size = PERF_ATTR_SIZE_VER0;
4819
4820         if (size < PERF_ATTR_SIZE_VER0)
4821                 goto err_size;
4822
4823         /*
4824          * If we're handed a bigger struct than we know of,
4825          * ensure all the unknown bits are 0 - i.e. new
4826          * user-space does not rely on any kernel feature
4827          * extensions we dont know about yet.
4828          */
4829         if (size > sizeof(*attr)) {
4830                 unsigned char __user *addr;
4831                 unsigned char __user *end;
4832                 unsigned char val;
4833
4834                 addr = (void __user *)uattr + sizeof(*attr);
4835                 end  = (void __user *)uattr + size;
4836
4837                 for (; addr < end; addr++) {
4838                         ret = get_user(val, addr);
4839                         if (ret)
4840                                 return ret;
4841                         if (val)
4842                                 goto err_size;
4843                 }
4844                 size = sizeof(*attr);
4845         }
4846
4847         ret = copy_from_user(attr, uattr, size);
4848         if (ret)
4849                 return -EFAULT;
4850
4851         /*
4852          * If the type exists, the corresponding creation will verify
4853          * the attr->config.
4854          */
4855         if (attr->type >= PERF_TYPE_MAX)
4856                 return -EINVAL;
4857
4858         if (attr->__reserved_1)
4859                 return -EINVAL;
4860
4861         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4862                 return -EINVAL;
4863
4864         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4865                 return -EINVAL;
4866
4867 out:
4868         return ret;
4869
4870 err_size:
4871         put_user(sizeof(*attr), &uattr->size);
4872         ret = -E2BIG;
4873         goto out;
4874 }
4875
4876 static int perf_event_set_output(struct perf_event *event, int output_fd)
4877 {
4878         struct perf_event *output_event = NULL;
4879         struct file *output_file = NULL;
4880         struct perf_event *old_output;
4881         int fput_needed = 0;
4882         int ret = -EINVAL;
4883
4884         if (!output_fd)
4885                 goto set;
4886
4887         output_file = fget_light(output_fd, &fput_needed);
4888         if (!output_file)
4889                 return -EBADF;
4890
4891         if (output_file->f_op != &perf_fops)
4892                 goto out;
4893
4894         output_event = output_file->private_data;
4895
4896         /* Don't chain output fds */
4897         if (output_event->output)
4898                 goto out;
4899
4900         /* Don't set an output fd when we already have an output channel */
4901         if (event->data)
4902                 goto out;
4903
4904         atomic_long_inc(&output_file->f_count);
4905
4906 set:
4907         mutex_lock(&event->mmap_mutex);
4908         old_output = event->output;
4909         rcu_assign_pointer(event->output, output_event);
4910         mutex_unlock(&event->mmap_mutex);
4911
4912         if (old_output) {
4913                 /*
4914                  * we need to make sure no existing perf_output_*()
4915                  * is still referencing this event.
4916                  */
4917                 synchronize_rcu();
4918                 fput(old_output->filp);
4919         }
4920
4921         ret = 0;
4922 out:
4923         fput_light(output_file, fput_needed);
4924         return ret;
4925 }
4926
4927 /**
4928  * sys_perf_event_open - open a performance event, associate it to a task/cpu
4929  *
4930  * @attr_uptr:  event_id type attributes for monitoring/sampling
4931  * @pid:                target pid
4932  * @cpu:                target cpu
4933  * @group_fd:           group leader event fd
4934  */
4935 SYSCALL_DEFINE5(perf_event_open,
4936                 struct perf_event_attr __user *, attr_uptr,
4937                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4938 {
4939         struct perf_event *event, *group_leader;
4940         struct perf_event_attr attr;
4941         struct perf_event_context *ctx;
4942         struct file *event_file = NULL;
4943         struct file *group_file = NULL;
4944         int fput_needed = 0;
4945         int fput_needed2 = 0;
4946         int err;
4947
4948         /* for future expandability... */
4949         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4950                 return -EINVAL;
4951
4952         err = perf_copy_attr(attr_uptr, &attr);
4953         if (err)
4954                 return err;
4955
4956         if (!attr.exclude_kernel) {
4957                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4958                         return -EACCES;
4959         }
4960
4961         if (attr.freq) {
4962                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4963                         return -EINVAL;
4964         }
4965
4966         /*
4967          * Get the target context (task or percpu):
4968          */
4969         ctx = find_get_context(pid, cpu);
4970         if (IS_ERR(ctx))
4971                 return PTR_ERR(ctx);
4972
4973         /*
4974          * Look up the group leader (we will attach this event to it):
4975          */
4976         group_leader = NULL;
4977         if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4978                 err = -EINVAL;
4979                 group_file = fget_light(group_fd, &fput_needed);
4980                 if (!group_file)
4981                         goto err_put_context;
4982                 if (group_file->f_op != &perf_fops)
4983                         goto err_put_context;
4984
4985                 group_leader = group_file->private_data;
4986                 /*
4987                  * Do not allow a recursive hierarchy (this new sibling
4988                  * becoming part of another group-sibling):
4989                  */
4990                 if (group_leader->group_leader != group_leader)
4991                         goto err_put_context;
4992                 /*
4993                  * Do not allow to attach to a group in a different
4994                  * task or CPU context:
4995                  */
4996                 if (group_leader->ctx != ctx)
4997                         goto err_put_context;
4998                 /*
4999                  * Only a group leader can be exclusive or pinned
5000                  */
5001                 if (attr.exclusive || attr.pinned)
5002                         goto err_put_context;
5003         }
5004
5005         event = perf_event_alloc(&attr, cpu, ctx, group_leader,
5006                                      NULL, NULL, GFP_KERNEL);
5007         err = PTR_ERR(event);
5008         if (IS_ERR(event))
5009                 goto err_put_context;
5010
5011         err = anon_inode_getfd("[perf_event]", &perf_fops, event, O_RDWR);
5012         if (err < 0)
5013                 goto err_free_put_context;
5014
5015         event_file = fget_light(err, &fput_needed2);
5016         if (!event_file)
5017                 goto err_free_put_context;
5018
5019         if (flags & PERF_FLAG_FD_OUTPUT) {
5020                 err = perf_event_set_output(event, group_fd);
5021                 if (err)
5022                         goto err_fput_free_put_context;
5023         }
5024
5025         event->filp = event_file;
5026         WARN_ON_ONCE(ctx->parent_ctx);
5027         mutex_lock(&ctx->mutex);
5028         perf_install_in_context(ctx, event, cpu);
5029         ++ctx->generation;
5030         mutex_unlock(&ctx->mutex);
5031
5032         event->owner = current;
5033         get_task_struct(current);
5034         mutex_lock(&current->perf_event_mutex);
5035         list_add_tail(&event->owner_entry, &current->perf_event_list);
5036         mutex_unlock(&current->perf_event_mutex);
5037
5038 err_fput_free_put_context:
5039         fput_light(event_file, fput_needed2);
5040
5041 err_free_put_context:
5042         if (err < 0)
5043                 free_event(event);
5044
5045 err_put_context:
5046         if (err < 0)
5047                 put_ctx(ctx);
5048
5049         fput_light(group_file, fput_needed);
5050
5051         return err;
5052 }
5053
5054 /**
5055  * perf_event_create_kernel_counter
5056  *
5057  * @attr: attributes of the counter to create
5058  * @cpu: cpu in which the counter is bound
5059  * @pid: task to profile
5060  */
5061 struct perf_event *
5062 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
5063                                  pid_t pid,
5064                                  perf_overflow_handler_t overflow_handler)
5065 {
5066         struct perf_event *event;
5067         struct perf_event_context *ctx;
5068         int err;
5069
5070         /*
5071          * Get the target context (task or percpu):
5072          */
5073
5074         ctx = find_get_context(pid, cpu);
5075         if (IS_ERR(ctx)) {
5076                 err = PTR_ERR(ctx);
5077                 goto err_exit;
5078         }
5079
5080         event = perf_event_alloc(attr, cpu, ctx, NULL,
5081                                  NULL, overflow_handler, GFP_KERNEL);
5082         if (IS_ERR(event)) {
5083                 err = PTR_ERR(event);
5084                 goto err_put_context;
5085         }
5086
5087         event->filp = NULL;
5088         WARN_ON_ONCE(ctx->parent_ctx);
5089         mutex_lock(&ctx->mutex);
5090         perf_install_in_context(ctx, event, cpu);
5091         ++ctx->generation;
5092         mutex_unlock(&ctx->mutex);
5093
5094         event->owner = current;
5095         get_task_struct(current);
5096         mutex_lock(&current->perf_event_mutex);
5097         list_add_tail(&event->owner_entry, &current->perf_event_list);
5098         mutex_unlock(&current->perf_event_mutex);
5099
5100         return event;
5101
5102  err_put_context:
5103         put_ctx(ctx);
5104  err_exit:
5105         return ERR_PTR(err);
5106 }
5107 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
5108
5109 /*
5110  * inherit a event from parent task to child task:
5111  */
5112 static struct perf_event *
5113 inherit_event(struct perf_event *parent_event,
5114               struct task_struct *parent,
5115               struct perf_event_context *parent_ctx,
5116               struct task_struct *child,
5117               struct perf_event *group_leader,
5118               struct perf_event_context *child_ctx)
5119 {
5120         struct perf_event *child_event;
5121
5122         /*
5123          * Instead of creating recursive hierarchies of events,
5124          * we link inherited events back to the original parent,
5125          * which has a filp for sure, which we use as the reference
5126          * count:
5127          */
5128         if (parent_event->parent)
5129                 parent_event = parent_event->parent;
5130
5131         child_event = perf_event_alloc(&parent_event->attr,
5132                                            parent_event->cpu, child_ctx,
5133                                            group_leader, parent_event,
5134                                            NULL, GFP_KERNEL);
5135         if (IS_ERR(child_event))
5136                 return child_event;
5137         get_ctx(child_ctx);
5138
5139         /*
5140          * Make the child state follow the state of the parent event,
5141          * not its attr.disabled bit.  We hold the parent's mutex,
5142          * so we won't race with perf_event_{en, dis}able_family.
5143          */
5144         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
5145                 child_event->state = PERF_EVENT_STATE_INACTIVE;
5146         else
5147                 child_event->state = PERF_EVENT_STATE_OFF;
5148
5149         if (parent_event->attr.freq) {
5150                 u64 sample_period = parent_event->hw.sample_period;
5151                 struct hw_perf_event *hwc = &child_event->hw;
5152
5153                 hwc->sample_period = sample_period;
5154                 hwc->last_period   = sample_period;
5155
5156                 atomic64_set(&hwc->period_left, sample_period);
5157         }
5158
5159         child_event->overflow_handler = parent_event->overflow_handler;
5160
5161         /*
5162          * Link it up in the child's context:
5163          */
5164         add_event_to_ctx(child_event, child_ctx);
5165
5166         /*
5167          * Get a reference to the parent filp - we will fput it
5168          * when the child event exits. This is safe to do because
5169          * we are in the parent and we know that the filp still
5170          * exists and has a nonzero count:
5171          */
5172         atomic_long_inc(&parent_event->filp->f_count);
5173
5174         /*
5175          * Link this into the parent event's child list
5176          */
5177         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5178         mutex_lock(&parent_event->child_mutex);
5179         list_add_tail(&child_event->child_list, &parent_event->child_list);
5180         mutex_unlock(&parent_event->child_mutex);
5181
5182         return child_event;
5183 }
5184
5185 static int inherit_group(struct perf_event *parent_event,
5186               struct task_struct *parent,
5187               struct perf_event_context *parent_ctx,
5188               struct task_struct *child,
5189               struct perf_event_context *child_ctx)
5190 {
5191         struct perf_event *leader;
5192         struct perf_event *sub;
5193         struct perf_event *child_ctr;
5194
5195         leader = inherit_event(parent_event, parent, parent_ctx,
5196                                  child, NULL, child_ctx);
5197         if (IS_ERR(leader))
5198                 return PTR_ERR(leader);
5199         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5200                 child_ctr = inherit_event(sub, parent, parent_ctx,
5201                                             child, leader, child_ctx);
5202                 if (IS_ERR(child_ctr))
5203                         return PTR_ERR(child_ctr);
5204         }
5205         return 0;
5206 }
5207
5208 static void sync_child_event(struct perf_event *child_event,
5209                                struct task_struct *child)
5210 {
5211         struct perf_event *parent_event = child_event->parent;
5212         u64 child_val;
5213
5214         if (child_event->attr.inherit_stat)
5215                 perf_event_read_event(child_event, child);
5216
5217         child_val = atomic64_read(&child_event->count);
5218
5219         /*
5220          * Add back the child's count to the parent's count:
5221          */
5222         atomic64_add(child_val, &parent_event->count);
5223         atomic64_add(child_event->total_time_enabled,
5224                      &parent_event->child_total_time_enabled);
5225         atomic64_add(child_event->total_time_running,
5226                      &parent_event->child_total_time_running);
5227
5228         /*
5229          * Remove this event from the parent's list
5230          */
5231         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5232         mutex_lock(&parent_event->child_mutex);
5233         list_del_init(&child_event->child_list);
5234         mutex_unlock(&parent_event->child_mutex);
5235
5236         /*
5237          * Release the parent event, if this was the last
5238          * reference to it.
5239          */
5240         fput(parent_event->filp);
5241 }
5242
5243 static void
5244 __perf_event_exit_task(struct perf_event *child_event,
5245                          struct perf_event_context *child_ctx,
5246                          struct task_struct *child)
5247 {
5248         struct perf_event *parent_event;
5249
5250         perf_event_remove_from_context(child_event);
5251
5252         parent_event = child_event->parent;
5253         /*
5254          * It can happen that parent exits first, and has events
5255          * that are still around due to the child reference. These
5256          * events need to be zapped - but otherwise linger.
5257          */
5258         if (parent_event) {
5259                 sync_child_event(child_event, child);
5260                 free_event(child_event);
5261         }
5262 }
5263
5264 /*
5265  * When a child task exits, feed back event values to parent events.
5266  */
5267 void perf_event_exit_task(struct task_struct *child)
5268 {
5269         struct perf_event *child_event, *tmp;
5270         struct perf_event_context *child_ctx;
5271         unsigned long flags;
5272
5273         if (likely(!child->perf_event_ctxp)) {
5274                 perf_event_task(child, NULL, 0);
5275                 return;
5276         }
5277
5278         local_irq_save(flags);
5279         /*
5280          * We can't reschedule here because interrupts are disabled,
5281          * and either child is current or it is a task that can't be
5282          * scheduled, so we are now safe from rescheduling changing
5283          * our context.
5284          */
5285         child_ctx = child->perf_event_ctxp;
5286         __perf_event_task_sched_out(child_ctx);
5287
5288         /*
5289          * Take the context lock here so that if find_get_context is
5290          * reading child->perf_event_ctxp, we wait until it has
5291          * incremented the context's refcount before we do put_ctx below.
5292          */
5293         raw_spin_lock(&child_ctx->lock);
5294         child->perf_event_ctxp = NULL;
5295         /*
5296          * If this context is a clone; unclone it so it can't get
5297          * swapped to another process while we're removing all
5298          * the events from it.
5299          */
5300         unclone_ctx(child_ctx);
5301         update_context_time(child_ctx);
5302         raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5303
5304         /*
5305          * Report the task dead after unscheduling the events so that we
5306          * won't get any samples after PERF_RECORD_EXIT. We can however still
5307          * get a few PERF_RECORD_READ events.
5308          */
5309         perf_event_task(child, child_ctx, 0);
5310
5311         /*
5312          * We can recurse on the same lock type through:
5313          *
5314          *   __perf_event_exit_task()
5315          *     sync_child_event()
5316          *       fput(parent_event->filp)
5317          *         perf_release()
5318          *           mutex_lock(&ctx->mutex)
5319          *
5320          * But since its the parent context it won't be the same instance.
5321          */
5322         mutex_lock(&child_ctx->mutex);
5323
5324 again:
5325         list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5326                                  group_entry)
5327                 __perf_event_exit_task(child_event, child_ctx, child);
5328
5329         list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5330                                  group_entry)
5331                 __perf_event_exit_task(child_event, child_ctx, child);
5332
5333         /*
5334          * If the last event was a group event, it will have appended all
5335          * its siblings to the list, but we obtained 'tmp' before that which
5336          * will still point to the list head terminating the iteration.
5337          */
5338         if (!list_empty(&child_ctx->pinned_groups) ||
5339             !list_empty(&child_ctx->flexible_groups))
5340                 goto again;
5341
5342         mutex_unlock(&child_ctx->mutex);
5343
5344         put_ctx(child_ctx);
5345 }
5346
5347 static void perf_free_event(struct perf_event *event,
5348                             struct perf_event_context *ctx)
5349 {
5350         struct perf_event *parent = event->parent;
5351
5352         if (WARN_ON_ONCE(!parent))
5353                 return;
5354
5355         mutex_lock(&parent->child_mutex);
5356         list_del_init(&event->child_list);
5357         mutex_unlock(&parent->child_mutex);
5358
5359         fput(parent->filp);
5360
5361         list_del_event(event, ctx);
5362         free_event(event);
5363 }
5364
5365 /*
5366  * free an unexposed, unused context as created by inheritance by
5367  * init_task below, used by fork() in case of fail.
5368  */
5369 void perf_event_free_task(struct task_struct *task)
5370 {
5371         struct perf_event_context *ctx = task->perf_event_ctxp;
5372         struct perf_event *event, *tmp;
5373
5374         if (!ctx)
5375                 return;
5376
5377         mutex_lock(&ctx->mutex);
5378 again:
5379         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5380                 perf_free_event(event, ctx);
5381
5382         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5383                                  group_entry)
5384                 perf_free_event(event, ctx);
5385
5386         if (!list_empty(&ctx->pinned_groups) ||
5387             !list_empty(&ctx->flexible_groups))
5388                 goto again;
5389
5390         mutex_unlock(&ctx->mutex);
5391
5392         put_ctx(ctx);
5393 }
5394
5395 static int
5396 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5397                    struct perf_event_context *parent_ctx,
5398                    struct task_struct *child,
5399                    int *inherited_all)
5400 {
5401         int ret;
5402         struct perf_event_context *child_ctx = child->perf_event_ctxp;
5403
5404         if (!event->attr.inherit) {
5405                 *inherited_all = 0;
5406                 return 0;
5407         }
5408
5409         if (!child_ctx) {
5410                 /*
5411                  * This is executed from the parent task context, so
5412                  * inherit events that have been marked for cloning.
5413                  * First allocate and initialize a context for the
5414                  * child.
5415                  */
5416
5417                 child_ctx = kzalloc(sizeof(struct perf_event_context),
5418                                     GFP_KERNEL);
5419                 if (!child_ctx)
5420                         return -ENOMEM;
5421
5422                 __perf_event_init_context(child_ctx, child);
5423                 child->perf_event_ctxp = child_ctx;
5424                 get_task_struct(child);
5425         }
5426
5427         ret = inherit_group(event, parent, parent_ctx,
5428                             child, child_ctx);
5429
5430         if (ret)
5431                 *inherited_all = 0;
5432
5433         return ret;
5434 }
5435
5436
5437 /*
5438  * Initialize the perf_event context in task_struct
5439  */
5440 int perf_event_init_task(struct task_struct *child)
5441 {
5442         struct perf_event_context *child_ctx, *parent_ctx;
5443         struct perf_event_context *cloned_ctx;
5444         struct perf_event *event;
5445         struct task_struct *parent = current;
5446         int inherited_all = 1;
5447         int ret = 0;
5448
5449         child->perf_event_ctxp = NULL;
5450
5451         mutex_init(&child->perf_event_mutex);
5452         INIT_LIST_HEAD(&child->perf_event_list);
5453
5454         if (likely(!parent->perf_event_ctxp))
5455                 return 0;
5456
5457         /*
5458          * If the parent's context is a clone, pin it so it won't get
5459          * swapped under us.
5460          */
5461         parent_ctx = perf_pin_task_context(parent);
5462
5463         /*
5464          * No need to check if parent_ctx != NULL here; since we saw
5465          * it non-NULL earlier, the only reason for it to become NULL
5466          * is if we exit, and since we're currently in the middle of
5467          * a fork we can't be exiting at the same time.
5468          */
5469
5470         /*
5471          * Lock the parent list. No need to lock the child - not PID
5472          * hashed yet and not running, so nobody can access it.
5473          */
5474         mutex_lock(&parent_ctx->mutex);
5475
5476         /*
5477          * We dont have to disable NMIs - we are only looking at
5478          * the list, not manipulating it:
5479          */
5480         list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5481                 ret = inherit_task_group(event, parent, parent_ctx, child,
5482                                          &inherited_all);
5483                 if (ret)
5484                         break;
5485         }
5486
5487         list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5488                 ret = inherit_task_group(event, parent, parent_ctx, child,
5489                                          &inherited_all);
5490                 if (ret)
5491                         break;
5492         }
5493
5494         child_ctx = child->perf_event_ctxp;
5495
5496         if (child_ctx && inherited_all) {
5497                 /*
5498                  * Mark the child context as a clone of the parent
5499                  * context, or of whatever the parent is a clone of.
5500                  * Note that if the parent is a clone, it could get
5501                  * uncloned at any point, but that doesn't matter
5502                  * because the list of events and the generation
5503                  * count can't have changed since we took the mutex.
5504                  */
5505                 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5506                 if (cloned_ctx) {
5507                         child_ctx->parent_ctx = cloned_ctx;
5508                         child_ctx->parent_gen = parent_ctx->parent_gen;
5509                 } else {
5510                         child_ctx->parent_ctx = parent_ctx;
5511                         child_ctx->parent_gen = parent_ctx->generation;
5512                 }
5513                 get_ctx(child_ctx->parent_ctx);
5514         }
5515
5516         mutex_unlock(&parent_ctx->mutex);
5517
5518         perf_unpin_context(parent_ctx);
5519
5520         return ret;
5521 }
5522
5523 static void __init perf_event_init_all_cpus(void)
5524 {
5525         int cpu;
5526         struct perf_cpu_context *cpuctx;
5527
5528         for_each_possible_cpu(cpu) {
5529                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5530                 mutex_init(&cpuctx->hlist_mutex);
5531                 __perf_event_init_context(&cpuctx->ctx, NULL);
5532         }
5533 }
5534
5535 static void __cpuinit perf_event_init_cpu(int cpu)
5536 {
5537         struct perf_cpu_context *cpuctx;
5538
5539         cpuctx = &per_cpu(perf_cpu_context, cpu);
5540
5541         spin_lock(&perf_resource_lock);
5542         cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5543         spin_unlock(&perf_resource_lock);
5544
5545         mutex_lock(&cpuctx->hlist_mutex);
5546         if (cpuctx->hlist_refcount > 0) {
5547                 struct swevent_hlist *hlist;
5548
5549                 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5550                 WARN_ON_ONCE(!hlist);
5551                 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
5552         }
5553         mutex_unlock(&cpuctx->hlist_mutex);
5554 }
5555
5556 #ifdef CONFIG_HOTPLUG_CPU
5557 static void __perf_event_exit_cpu(void *info)
5558 {
5559         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5560         struct perf_event_context *ctx = &cpuctx->ctx;
5561         struct perf_event *event, *tmp;
5562
5563         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5564                 __perf_event_remove_from_context(event);
5565         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5566                 __perf_event_remove_from_context(event);
5567 }
5568 static void perf_event_exit_cpu(int cpu)
5569 {
5570         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5571         struct perf_event_context *ctx = &cpuctx->ctx;
5572
5573         mutex_lock(&cpuctx->hlist_mutex);
5574         swevent_hlist_release(cpuctx);
5575         mutex_unlock(&cpuctx->hlist_mutex);
5576
5577         mutex_lock(&ctx->mutex);
5578         smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5579         mutex_unlock(&ctx->mutex);
5580 }
5581 #else
5582 static inline void perf_event_exit_cpu(int cpu) { }
5583 #endif
5584
5585 static int __cpuinit
5586 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5587 {
5588         unsigned int cpu = (long)hcpu;
5589
5590         switch (action) {
5591
5592         case CPU_UP_PREPARE:
5593         case CPU_UP_PREPARE_FROZEN:
5594                 perf_event_init_cpu(cpu);
5595                 break;
5596
5597         case CPU_DOWN_PREPARE:
5598         case CPU_DOWN_PREPARE_FROZEN:
5599                 perf_event_exit_cpu(cpu);
5600                 break;
5601
5602         default:
5603                 break;
5604         }
5605
5606         return NOTIFY_OK;
5607 }
5608
5609 /*
5610  * This has to have a higher priority than migration_notifier in sched.c.
5611  */
5612 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5613         .notifier_call          = perf_cpu_notify,
5614         .priority               = 20,
5615 };
5616
5617 void __init perf_event_init(void)
5618 {
5619         perf_event_init_all_cpus();
5620         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5621                         (void *)(long)smp_processor_id());
5622         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5623                         (void *)(long)smp_processor_id());
5624         register_cpu_notifier(&perf_cpu_nb);
5625 }
5626
5627 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class,
5628                                         struct sysdev_class_attribute *attr,
5629                                         char *buf)
5630 {
5631         return sprintf(buf, "%d\n", perf_reserved_percpu);
5632 }
5633
5634 static ssize_t
5635 perf_set_reserve_percpu(struct sysdev_class *class,
5636                         struct sysdev_class_attribute *attr,
5637                         const char *buf,
5638                         size_t count)
5639 {
5640         struct perf_cpu_context *cpuctx;
5641         unsigned long val;
5642         int err, cpu, mpt;
5643
5644         err = strict_strtoul(buf, 10, &val);
5645         if (err)
5646                 return err;
5647         if (val > perf_max_events)
5648                 return -EINVAL;
5649
5650         spin_lock(&perf_resource_lock);
5651         perf_reserved_percpu = val;
5652         for_each_online_cpu(cpu) {
5653                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5654                 raw_spin_lock_irq(&cpuctx->ctx.lock);
5655                 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5656                           perf_max_events - perf_reserved_percpu);
5657                 cpuctx->max_pertask = mpt;
5658                 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5659         }
5660         spin_unlock(&perf_resource_lock);
5661
5662         return count;
5663 }
5664
5665 static ssize_t perf_show_overcommit(struct sysdev_class *class,
5666                                     struct sysdev_class_attribute *attr,
5667                                     char *buf)
5668 {
5669         return sprintf(buf, "%d\n", perf_overcommit);
5670 }
5671
5672 static ssize_t
5673 perf_set_overcommit(struct sysdev_class *class,
5674                     struct sysdev_class_attribute *attr,
5675                     const char *buf, size_t count)
5676 {
5677         unsigned long val;
5678         int err;
5679
5680         err = strict_strtoul(buf, 10, &val);
5681         if (err)
5682                 return err;
5683         if (val > 1)
5684                 return -EINVAL;
5685
5686         spin_lock(&perf_resource_lock);
5687         perf_overcommit = val;
5688         spin_unlock(&perf_resource_lock);
5689
5690         return count;
5691 }
5692
5693 static SYSDEV_CLASS_ATTR(
5694                                 reserve_percpu,
5695                                 0644,
5696                                 perf_show_reserve_percpu,
5697                                 perf_set_reserve_percpu
5698                         );
5699
5700 static SYSDEV_CLASS_ATTR(
5701                                 overcommit,
5702                                 0644,
5703                                 perf_show_overcommit,
5704                                 perf_set_overcommit
5705                         );
5706
5707 static struct attribute *perfclass_attrs[] = {
5708         &attr_reserve_percpu.attr,
5709         &attr_overcommit.attr,
5710         NULL
5711 };
5712
5713 static struct attribute_group perfclass_attr_group = {
5714         .attrs                  = perfclass_attrs,
5715         .name                   = "perf_events",
5716 };
5717
5718 static int __init perf_event_sysfs_init(void)
5719 {
5720         return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5721                                   &perfclass_attr_group);
5722 }
5723 device_initcall(perf_event_sysfs_init);