848d49a043e98af0bf5658aa5877a7c4d16e6f05
[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 /*
259  * Update total_time_enabled and total_time_running for all events in a group.
260  */
261 static void update_group_times(struct perf_event *leader)
262 {
263         struct perf_event *event;
264
265         update_event_times(leader);
266         list_for_each_entry(event, &leader->sibling_list, group_entry)
267                 update_event_times(event);
268 }
269
270 static struct list_head *
271 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
272 {
273         if (event->attr.pinned)
274                 return &ctx->pinned_groups;
275         else
276                 return &ctx->flexible_groups;
277 }
278
279 /*
280  * Add a event from the lists for its context.
281  * Must be called with ctx->mutex and ctx->lock held.
282  */
283 static void
284 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
285 {
286         struct perf_event *group_leader = event->group_leader;
287
288         /*
289          * Depending on whether it is a standalone or sibling event,
290          * add it straight to the context's event list, or to the group
291          * leader's sibling list:
292          */
293         if (group_leader == event) {
294                 struct list_head *list;
295
296                 if (is_software_event(event))
297                         event->group_flags |= PERF_GROUP_SOFTWARE;
298
299                 list = ctx_group_list(event, ctx);
300                 list_add_tail(&event->group_entry, list);
301         } else {
302                 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
303                     !is_software_event(event))
304                         group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
305
306                 list_add_tail(&event->group_entry, &group_leader->sibling_list);
307                 group_leader->nr_siblings++;
308         }
309
310         list_add_rcu(&event->event_entry, &ctx->event_list);
311         ctx->nr_events++;
312         if (event->attr.inherit_stat)
313                 ctx->nr_stat++;
314 }
315
316 /*
317  * Remove a event from the lists for its context.
318  * Must be called with ctx->mutex and ctx->lock held.
319  */
320 static void
321 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
322 {
323         if (list_empty(&event->group_entry))
324                 return;
325         ctx->nr_events--;
326         if (event->attr.inherit_stat)
327                 ctx->nr_stat--;
328
329         list_del_init(&event->group_entry);
330         list_del_rcu(&event->event_entry);
331
332         if (event->group_leader != event)
333                 event->group_leader->nr_siblings--;
334
335         update_group_times(event);
336
337         /*
338          * If event was in error state, then keep it
339          * that way, otherwise bogus counts will be
340          * returned on read(). The only way to get out
341          * of error state is by explicit re-enabling
342          * of the event
343          */
344         if (event->state > PERF_EVENT_STATE_OFF)
345                 event->state = PERF_EVENT_STATE_OFF;
346 }
347
348 static void
349 perf_destroy_group(struct perf_event *event, struct perf_event_context *ctx)
350 {
351         struct perf_event *sibling, *tmp;
352
353         /*
354          * If this was a group event with sibling events then
355          * upgrade the siblings to singleton events by adding them
356          * to the context list directly:
357          */
358         list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
359                 struct list_head *list;
360
361                 list = ctx_group_list(event, ctx);
362                 list_move_tail(&sibling->group_entry, list);
363                 sibling->group_leader = sibling;
364
365                 /* Inherit group flags from the previous leader */
366                 sibling->group_flags = event->group_flags;
367         }
368 }
369
370 static void
371 event_sched_out(struct perf_event *event,
372                   struct perf_cpu_context *cpuctx,
373                   struct perf_event_context *ctx)
374 {
375         if (event->state != PERF_EVENT_STATE_ACTIVE)
376                 return;
377
378         event->state = PERF_EVENT_STATE_INACTIVE;
379         if (event->pending_disable) {
380                 event->pending_disable = 0;
381                 event->state = PERF_EVENT_STATE_OFF;
382         }
383         event->tstamp_stopped = ctx->time;
384         event->pmu->disable(event);
385         event->oncpu = -1;
386
387         if (!is_software_event(event))
388                 cpuctx->active_oncpu--;
389         ctx->nr_active--;
390         if (event->attr.exclusive || !cpuctx->active_oncpu)
391                 cpuctx->exclusive = 0;
392 }
393
394 static void
395 group_sched_out(struct perf_event *group_event,
396                 struct perf_cpu_context *cpuctx,
397                 struct perf_event_context *ctx)
398 {
399         struct perf_event *event;
400
401         if (group_event->state != PERF_EVENT_STATE_ACTIVE)
402                 return;
403
404         event_sched_out(group_event, cpuctx, ctx);
405
406         /*
407          * Schedule out siblings (if any):
408          */
409         list_for_each_entry(event, &group_event->sibling_list, group_entry)
410                 event_sched_out(event, cpuctx, ctx);
411
412         if (group_event->attr.exclusive)
413                 cpuctx->exclusive = 0;
414 }
415
416 /*
417  * Cross CPU call to remove a performance event
418  *
419  * We disable the event on the hardware level first. After that we
420  * remove it from the context list.
421  */
422 static void __perf_event_remove_from_context(void *info)
423 {
424         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
425         struct perf_event *event = info;
426         struct perf_event_context *ctx = event->ctx;
427
428         /*
429          * If this is a task context, we need to check whether it is
430          * the current task context of this cpu. If not it has been
431          * scheduled out before the smp call arrived.
432          */
433         if (ctx->task && cpuctx->task_ctx != ctx)
434                 return;
435
436         raw_spin_lock(&ctx->lock);
437         /*
438          * Protect the list operation against NMI by disabling the
439          * events on a global level.
440          */
441         perf_disable();
442
443         event_sched_out(event, cpuctx, ctx);
444
445         list_del_event(event, ctx);
446
447         if (!ctx->task) {
448                 /*
449                  * Allow more per task events with respect to the
450                  * reservation:
451                  */
452                 cpuctx->max_pertask =
453                         min(perf_max_events - ctx->nr_events,
454                             perf_max_events - perf_reserved_percpu);
455         }
456
457         perf_enable();
458         raw_spin_unlock(&ctx->lock);
459 }
460
461
462 /*
463  * Remove the event from a task's (or a CPU's) list of events.
464  *
465  * Must be called with ctx->mutex held.
466  *
467  * CPU events are removed with a smp call. For task events we only
468  * call when the task is on a CPU.
469  *
470  * If event->ctx is a cloned context, callers must make sure that
471  * every task struct that event->ctx->task could possibly point to
472  * remains valid.  This is OK when called from perf_release since
473  * that only calls us on the top-level context, which can't be a clone.
474  * When called from perf_event_exit_task, it's OK because the
475  * context has been detached from its task.
476  */
477 static void perf_event_remove_from_context(struct perf_event *event)
478 {
479         struct perf_event_context *ctx = event->ctx;
480         struct task_struct *task = ctx->task;
481
482         if (!task) {
483                 /*
484                  * Per cpu events are removed via an smp call and
485                  * the removal is always successful.
486                  */
487                 smp_call_function_single(event->cpu,
488                                          __perf_event_remove_from_context,
489                                          event, 1);
490                 return;
491         }
492
493 retry:
494         task_oncpu_function_call(task, __perf_event_remove_from_context,
495                                  event);
496
497         raw_spin_lock_irq(&ctx->lock);
498         /*
499          * If the context is active we need to retry the smp call.
500          */
501         if (ctx->nr_active && !list_empty(&event->group_entry)) {
502                 raw_spin_unlock_irq(&ctx->lock);
503                 goto retry;
504         }
505
506         /*
507          * The lock prevents that this context is scheduled in so we
508          * can remove the event safely, if the call above did not
509          * succeed.
510          */
511         if (!list_empty(&event->group_entry))
512                 list_del_event(event, ctx);
513         raw_spin_unlock_irq(&ctx->lock);
514 }
515
516 /*
517  * Cross CPU call to disable a performance event
518  */
519 static void __perf_event_disable(void *info)
520 {
521         struct perf_event *event = info;
522         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
523         struct perf_event_context *ctx = event->ctx;
524
525         /*
526          * If this is a per-task event, need to check whether this
527          * event's task is the current task on this cpu.
528          */
529         if (ctx->task && cpuctx->task_ctx != ctx)
530                 return;
531
532         raw_spin_lock(&ctx->lock);
533
534         /*
535          * If the event is on, turn it off.
536          * If it is in error state, leave it in error state.
537          */
538         if (event->state >= PERF_EVENT_STATE_INACTIVE) {
539                 update_context_time(ctx);
540                 update_group_times(event);
541                 if (event == event->group_leader)
542                         group_sched_out(event, cpuctx, ctx);
543                 else
544                         event_sched_out(event, cpuctx, ctx);
545                 event->state = PERF_EVENT_STATE_OFF;
546         }
547
548         raw_spin_unlock(&ctx->lock);
549 }
550
551 /*
552  * Disable a event.
553  *
554  * If event->ctx is a cloned context, callers must make sure that
555  * every task struct that event->ctx->task could possibly point to
556  * remains valid.  This condition is satisifed when called through
557  * perf_event_for_each_child or perf_event_for_each because they
558  * hold the top-level event's child_mutex, so any descendant that
559  * goes to exit will block in sync_child_event.
560  * When called from perf_pending_event it's OK because event->ctx
561  * is the current context on this CPU and preemption is disabled,
562  * hence we can't get into perf_event_task_sched_out for this context.
563  */
564 void perf_event_disable(struct perf_event *event)
565 {
566         struct perf_event_context *ctx = event->ctx;
567         struct task_struct *task = ctx->task;
568
569         if (!task) {
570                 /*
571                  * Disable the event on the cpu that it's on
572                  */
573                 smp_call_function_single(event->cpu, __perf_event_disable,
574                                          event, 1);
575                 return;
576         }
577
578  retry:
579         task_oncpu_function_call(task, __perf_event_disable, event);
580
581         raw_spin_lock_irq(&ctx->lock);
582         /*
583          * If the event is still active, we need to retry the cross-call.
584          */
585         if (event->state == PERF_EVENT_STATE_ACTIVE) {
586                 raw_spin_unlock_irq(&ctx->lock);
587                 goto retry;
588         }
589
590         /*
591          * Since we have the lock this context can't be scheduled
592          * in, so we can change the state safely.
593          */
594         if (event->state == PERF_EVENT_STATE_INACTIVE) {
595                 update_group_times(event);
596                 event->state = PERF_EVENT_STATE_OFF;
597         }
598
599         raw_spin_unlock_irq(&ctx->lock);
600 }
601
602 static int
603 event_sched_in(struct perf_event *event,
604                  struct perf_cpu_context *cpuctx,
605                  struct perf_event_context *ctx)
606 {
607         if (event->state <= PERF_EVENT_STATE_OFF)
608                 return 0;
609
610         event->state = PERF_EVENT_STATE_ACTIVE;
611         event->oncpu = smp_processor_id();
612         /*
613          * The new state must be visible before we turn it on in the hardware:
614          */
615         smp_wmb();
616
617         if (event->pmu->enable(event)) {
618                 event->state = PERF_EVENT_STATE_INACTIVE;
619                 event->oncpu = -1;
620                 return -EAGAIN;
621         }
622
623         event->tstamp_running += ctx->time - event->tstamp_stopped;
624
625         if (!is_software_event(event))
626                 cpuctx->active_oncpu++;
627         ctx->nr_active++;
628
629         if (event->attr.exclusive)
630                 cpuctx->exclusive = 1;
631
632         return 0;
633 }
634
635 static int
636 group_sched_in(struct perf_event *group_event,
637                struct perf_cpu_context *cpuctx,
638                struct perf_event_context *ctx)
639 {
640         struct perf_event *event, *partial_group = NULL;
641         const struct pmu *pmu = group_event->pmu;
642         bool txn = false;
643         int ret;
644
645         if (group_event->state == PERF_EVENT_STATE_OFF)
646                 return 0;
647
648         /* Check if group transaction availabe */
649         if (pmu->start_txn)
650                 txn = true;
651
652         if (txn)
653                 pmu->start_txn(pmu);
654
655         if (event_sched_in(group_event, cpuctx, ctx))
656                 return -EAGAIN;
657
658         /*
659          * Schedule in siblings as one group (if any):
660          */
661         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
662                 if (event_sched_in(event, cpuctx, ctx)) {
663                         partial_group = event;
664                         goto group_error;
665                 }
666         }
667
668         if (!txn)
669                 return 0;
670
671         ret = pmu->commit_txn(pmu);
672         if (!ret) {
673                 pmu->cancel_txn(pmu);
674                 return 0;
675         }
676
677 group_error:
678         if (txn)
679                 pmu->cancel_txn(pmu);
680
681         /*
682          * Groups can be scheduled in as one unit only, so undo any
683          * partial group before returning:
684          */
685         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
686                 if (event == partial_group)
687                         break;
688                 event_sched_out(event, cpuctx, ctx);
689         }
690         event_sched_out(group_event, cpuctx, ctx);
691
692         return -EAGAIN;
693 }
694
695 /*
696  * Work out whether we can put this event group on the CPU now.
697  */
698 static int group_can_go_on(struct perf_event *event,
699                            struct perf_cpu_context *cpuctx,
700                            int can_add_hw)
701 {
702         /*
703          * Groups consisting entirely of software events can always go on.
704          */
705         if (event->group_flags & PERF_GROUP_SOFTWARE)
706                 return 1;
707         /*
708          * If an exclusive group is already on, no other hardware
709          * events can go on.
710          */
711         if (cpuctx->exclusive)
712                 return 0;
713         /*
714          * If this group is exclusive and there are already
715          * events on the CPU, it can't go on.
716          */
717         if (event->attr.exclusive && cpuctx->active_oncpu)
718                 return 0;
719         /*
720          * Otherwise, try to add it if all previous groups were able
721          * to go on.
722          */
723         return can_add_hw;
724 }
725
726 static void add_event_to_ctx(struct perf_event *event,
727                                struct perf_event_context *ctx)
728 {
729         list_add_event(event, ctx);
730         event->tstamp_enabled = ctx->time;
731         event->tstamp_running = ctx->time;
732         event->tstamp_stopped = ctx->time;
733 }
734
735 /*
736  * Cross CPU call to install and enable a performance event
737  *
738  * Must be called with ctx->mutex held
739  */
740 static void __perf_install_in_context(void *info)
741 {
742         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
743         struct perf_event *event = info;
744         struct perf_event_context *ctx = event->ctx;
745         struct perf_event *leader = event->group_leader;
746         int err;
747
748         /*
749          * If this is a task context, we need to check whether it is
750          * the current task context of this cpu. If not it has been
751          * scheduled out before the smp call arrived.
752          * Or possibly this is the right context but it isn't
753          * on this cpu because it had no events.
754          */
755         if (ctx->task && cpuctx->task_ctx != ctx) {
756                 if (cpuctx->task_ctx || ctx->task != current)
757                         return;
758                 cpuctx->task_ctx = ctx;
759         }
760
761         raw_spin_lock(&ctx->lock);
762         ctx->is_active = 1;
763         update_context_time(ctx);
764
765         /*
766          * Protect the list operation against NMI by disabling the
767          * events on a global level. NOP for non NMI based events.
768          */
769         perf_disable();
770
771         add_event_to_ctx(event, ctx);
772
773         if (event->cpu != -1 && event->cpu != smp_processor_id())
774                 goto unlock;
775
776         /*
777          * Don't put the event on if it is disabled or if
778          * it is in a group and the group isn't on.
779          */
780         if (event->state != PERF_EVENT_STATE_INACTIVE ||
781             (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
782                 goto unlock;
783
784         /*
785          * An exclusive event can't go on if there are already active
786          * hardware events, and no hardware event can go on if there
787          * is already an exclusive event on.
788          */
789         if (!group_can_go_on(event, cpuctx, 1))
790                 err = -EEXIST;
791         else
792                 err = event_sched_in(event, cpuctx, ctx);
793
794         if (err) {
795                 /*
796                  * This event couldn't go on.  If it is in a group
797                  * then we have to pull the whole group off.
798                  * If the event group is pinned then put it in error state.
799                  */
800                 if (leader != event)
801                         group_sched_out(leader, cpuctx, ctx);
802                 if (leader->attr.pinned) {
803                         update_group_times(leader);
804                         leader->state = PERF_EVENT_STATE_ERROR;
805                 }
806         }
807
808         if (!err && !ctx->task && cpuctx->max_pertask)
809                 cpuctx->max_pertask--;
810
811  unlock:
812         perf_enable();
813
814         raw_spin_unlock(&ctx->lock);
815 }
816
817 /*
818  * Attach a performance event to a context
819  *
820  * First we add the event to the list with the hardware enable bit
821  * in event->hw_config cleared.
822  *
823  * If the event is attached to a task which is on a CPU we use a smp
824  * call to enable it in the task context. The task might have been
825  * scheduled away, but we check this in the smp call again.
826  *
827  * Must be called with ctx->mutex held.
828  */
829 static void
830 perf_install_in_context(struct perf_event_context *ctx,
831                         struct perf_event *event,
832                         int cpu)
833 {
834         struct task_struct *task = ctx->task;
835
836         if (!task) {
837                 /*
838                  * Per cpu events are installed via an smp call and
839                  * the install is always successful.
840                  */
841                 smp_call_function_single(cpu, __perf_install_in_context,
842                                          event, 1);
843                 return;
844         }
845
846 retry:
847         task_oncpu_function_call(task, __perf_install_in_context,
848                                  event);
849
850         raw_spin_lock_irq(&ctx->lock);
851         /*
852          * we need to retry the smp call.
853          */
854         if (ctx->is_active && list_empty(&event->group_entry)) {
855                 raw_spin_unlock_irq(&ctx->lock);
856                 goto retry;
857         }
858
859         /*
860          * The lock prevents that this context is scheduled in so we
861          * can add the event safely, if it the call above did not
862          * succeed.
863          */
864         if (list_empty(&event->group_entry))
865                 add_event_to_ctx(event, ctx);
866         raw_spin_unlock_irq(&ctx->lock);
867 }
868
869 /*
870  * Put a event into inactive state and update time fields.
871  * Enabling the leader of a group effectively enables all
872  * the group members that aren't explicitly disabled, so we
873  * have to update their ->tstamp_enabled also.
874  * Note: this works for group members as well as group leaders
875  * since the non-leader members' sibling_lists will be empty.
876  */
877 static void __perf_event_mark_enabled(struct perf_event *event,
878                                         struct perf_event_context *ctx)
879 {
880         struct perf_event *sub;
881
882         event->state = PERF_EVENT_STATE_INACTIVE;
883         event->tstamp_enabled = ctx->time - event->total_time_enabled;
884         list_for_each_entry(sub, &event->sibling_list, group_entry)
885                 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
886                         sub->tstamp_enabled =
887                                 ctx->time - sub->total_time_enabled;
888 }
889
890 /*
891  * Cross CPU call to enable a performance event
892  */
893 static void __perf_event_enable(void *info)
894 {
895         struct perf_event *event = info;
896         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
897         struct perf_event_context *ctx = event->ctx;
898         struct perf_event *leader = event->group_leader;
899         int err;
900
901         /*
902          * If this is a per-task event, need to check whether this
903          * event's task is the current task on this cpu.
904          */
905         if (ctx->task && cpuctx->task_ctx != ctx) {
906                 if (cpuctx->task_ctx || ctx->task != current)
907                         return;
908                 cpuctx->task_ctx = ctx;
909         }
910
911         raw_spin_lock(&ctx->lock);
912         ctx->is_active = 1;
913         update_context_time(ctx);
914
915         if (event->state >= PERF_EVENT_STATE_INACTIVE)
916                 goto unlock;
917         __perf_event_mark_enabled(event, ctx);
918
919         if (event->cpu != -1 && event->cpu != smp_processor_id())
920                 goto unlock;
921
922         /*
923          * If the event is in a group and isn't the group leader,
924          * then don't put it on unless the group is on.
925          */
926         if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
927                 goto unlock;
928
929         if (!group_can_go_on(event, cpuctx, 1)) {
930                 err = -EEXIST;
931         } else {
932                 perf_disable();
933                 if (event == leader)
934                         err = group_sched_in(event, cpuctx, ctx);
935                 else
936                         err = event_sched_in(event, cpuctx, ctx);
937                 perf_enable();
938         }
939
940         if (err) {
941                 /*
942                  * If this event can't go on and it's part of a
943                  * group, then the whole group has to come off.
944                  */
945                 if (leader != event)
946                         group_sched_out(leader, cpuctx, ctx);
947                 if (leader->attr.pinned) {
948                         update_group_times(leader);
949                         leader->state = PERF_EVENT_STATE_ERROR;
950                 }
951         }
952
953  unlock:
954         raw_spin_unlock(&ctx->lock);
955 }
956
957 /*
958  * Enable a event.
959  *
960  * If event->ctx is a cloned context, callers must make sure that
961  * every task struct that event->ctx->task could possibly point to
962  * remains valid.  This condition is satisfied when called through
963  * perf_event_for_each_child or perf_event_for_each as described
964  * for perf_event_disable.
965  */
966 void perf_event_enable(struct perf_event *event)
967 {
968         struct perf_event_context *ctx = event->ctx;
969         struct task_struct *task = ctx->task;
970
971         if (!task) {
972                 /*
973                  * Enable the event on the cpu that it's on
974                  */
975                 smp_call_function_single(event->cpu, __perf_event_enable,
976                                          event, 1);
977                 return;
978         }
979
980         raw_spin_lock_irq(&ctx->lock);
981         if (event->state >= PERF_EVENT_STATE_INACTIVE)
982                 goto out;
983
984         /*
985          * If the event is in error state, clear that first.
986          * That way, if we see the event in error state below, we
987          * know that it has gone back into error state, as distinct
988          * from the task having been scheduled away before the
989          * cross-call arrived.
990          */
991         if (event->state == PERF_EVENT_STATE_ERROR)
992                 event->state = PERF_EVENT_STATE_OFF;
993
994  retry:
995         raw_spin_unlock_irq(&ctx->lock);
996         task_oncpu_function_call(task, __perf_event_enable, event);
997
998         raw_spin_lock_irq(&ctx->lock);
999
1000         /*
1001          * If the context is active and the event is still off,
1002          * we need to retry the cross-call.
1003          */
1004         if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1005                 goto retry;
1006
1007         /*
1008          * Since we have the lock this context can't be scheduled
1009          * in, so we can change the state safely.
1010          */
1011         if (event->state == PERF_EVENT_STATE_OFF)
1012                 __perf_event_mark_enabled(event, ctx);
1013
1014  out:
1015         raw_spin_unlock_irq(&ctx->lock);
1016 }
1017
1018 static int perf_event_refresh(struct perf_event *event, int refresh)
1019 {
1020         /*
1021          * not supported on inherited events
1022          */
1023         if (event->attr.inherit)
1024                 return -EINVAL;
1025
1026         atomic_add(refresh, &event->event_limit);
1027         perf_event_enable(event);
1028
1029         return 0;
1030 }
1031
1032 enum event_type_t {
1033         EVENT_FLEXIBLE = 0x1,
1034         EVENT_PINNED = 0x2,
1035         EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1036 };
1037
1038 static void ctx_sched_out(struct perf_event_context *ctx,
1039                           struct perf_cpu_context *cpuctx,
1040                           enum event_type_t event_type)
1041 {
1042         struct perf_event *event;
1043
1044         raw_spin_lock(&ctx->lock);
1045         ctx->is_active = 0;
1046         if (likely(!ctx->nr_events))
1047                 goto out;
1048         update_context_time(ctx);
1049
1050         perf_disable();
1051         if (!ctx->nr_active)
1052                 goto out_enable;
1053
1054         if (event_type & EVENT_PINNED)
1055                 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1056                         group_sched_out(event, cpuctx, ctx);
1057
1058         if (event_type & EVENT_FLEXIBLE)
1059                 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1060                         group_sched_out(event, cpuctx, ctx);
1061
1062  out_enable:
1063         perf_enable();
1064  out:
1065         raw_spin_unlock(&ctx->lock);
1066 }
1067
1068 /*
1069  * Test whether two contexts are equivalent, i.e. whether they
1070  * have both been cloned from the same version of the same context
1071  * and they both have the same number of enabled events.
1072  * If the number of enabled events is the same, then the set
1073  * of enabled events should be the same, because these are both
1074  * inherited contexts, therefore we can't access individual events
1075  * in them directly with an fd; we can only enable/disable all
1076  * events via prctl, or enable/disable all events in a family
1077  * via ioctl, which will have the same effect on both contexts.
1078  */
1079 static int context_equiv(struct perf_event_context *ctx1,
1080                          struct perf_event_context *ctx2)
1081 {
1082         return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1083                 && ctx1->parent_gen == ctx2->parent_gen
1084                 && !ctx1->pin_count && !ctx2->pin_count;
1085 }
1086
1087 static void __perf_event_sync_stat(struct perf_event *event,
1088                                      struct perf_event *next_event)
1089 {
1090         u64 value;
1091
1092         if (!event->attr.inherit_stat)
1093                 return;
1094
1095         /*
1096          * Update the event value, we cannot use perf_event_read()
1097          * because we're in the middle of a context switch and have IRQs
1098          * disabled, which upsets smp_call_function_single(), however
1099          * we know the event must be on the current CPU, therefore we
1100          * don't need to use it.
1101          */
1102         switch (event->state) {
1103         case PERF_EVENT_STATE_ACTIVE:
1104                 event->pmu->read(event);
1105                 /* fall-through */
1106
1107         case PERF_EVENT_STATE_INACTIVE:
1108                 update_event_times(event);
1109                 break;
1110
1111         default:
1112                 break;
1113         }
1114
1115         /*
1116          * In order to keep per-task stats reliable we need to flip the event
1117          * values when we flip the contexts.
1118          */
1119         value = atomic64_read(&next_event->count);
1120         value = atomic64_xchg(&event->count, value);
1121         atomic64_set(&next_event->count, value);
1122
1123         swap(event->total_time_enabled, next_event->total_time_enabled);
1124         swap(event->total_time_running, next_event->total_time_running);
1125
1126         /*
1127          * Since we swizzled the values, update the user visible data too.
1128          */
1129         perf_event_update_userpage(event);
1130         perf_event_update_userpage(next_event);
1131 }
1132
1133 #define list_next_entry(pos, member) \
1134         list_entry(pos->member.next, typeof(*pos), member)
1135
1136 static void perf_event_sync_stat(struct perf_event_context *ctx,
1137                                    struct perf_event_context *next_ctx)
1138 {
1139         struct perf_event *event, *next_event;
1140
1141         if (!ctx->nr_stat)
1142                 return;
1143
1144         update_context_time(ctx);
1145
1146         event = list_first_entry(&ctx->event_list,
1147                                    struct perf_event, event_entry);
1148
1149         next_event = list_first_entry(&next_ctx->event_list,
1150                                         struct perf_event, event_entry);
1151
1152         while (&event->event_entry != &ctx->event_list &&
1153                &next_event->event_entry != &next_ctx->event_list) {
1154
1155                 __perf_event_sync_stat(event, next_event);
1156
1157                 event = list_next_entry(event, event_entry);
1158                 next_event = list_next_entry(next_event, event_entry);
1159         }
1160 }
1161
1162 /*
1163  * Called from scheduler to remove the events of the current task,
1164  * with interrupts disabled.
1165  *
1166  * We stop each event and update the event value in event->count.
1167  *
1168  * This does not protect us against NMI, but disable()
1169  * sets the disabled bit in the control field of event _before_
1170  * accessing the event control register. If a NMI hits, then it will
1171  * not restart the event.
1172  */
1173 void perf_event_task_sched_out(struct task_struct *task,
1174                                  struct task_struct *next)
1175 {
1176         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1177         struct perf_event_context *ctx = task->perf_event_ctxp;
1178         struct perf_event_context *next_ctx;
1179         struct perf_event_context *parent;
1180         int do_switch = 1;
1181
1182         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0);
1183
1184         if (likely(!ctx || !cpuctx->task_ctx))
1185                 return;
1186
1187         rcu_read_lock();
1188         parent = rcu_dereference(ctx->parent_ctx);
1189         next_ctx = next->perf_event_ctxp;
1190         if (parent && next_ctx &&
1191             rcu_dereference(next_ctx->parent_ctx) == parent) {
1192                 /*
1193                  * Looks like the two contexts are clones, so we might be
1194                  * able to optimize the context switch.  We lock both
1195                  * contexts and check that they are clones under the
1196                  * lock (including re-checking that neither has been
1197                  * uncloned in the meantime).  It doesn't matter which
1198                  * order we take the locks because no other cpu could
1199                  * be trying to lock both of these tasks.
1200                  */
1201                 raw_spin_lock(&ctx->lock);
1202                 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1203                 if (context_equiv(ctx, next_ctx)) {
1204                         /*
1205                          * XXX do we need a memory barrier of sorts
1206                          * wrt to rcu_dereference() of perf_event_ctxp
1207                          */
1208                         task->perf_event_ctxp = next_ctx;
1209                         next->perf_event_ctxp = ctx;
1210                         ctx->task = next;
1211                         next_ctx->task = task;
1212                         do_switch = 0;
1213
1214                         perf_event_sync_stat(ctx, next_ctx);
1215                 }
1216                 raw_spin_unlock(&next_ctx->lock);
1217                 raw_spin_unlock(&ctx->lock);
1218         }
1219         rcu_read_unlock();
1220
1221         if (do_switch) {
1222                 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1223                 cpuctx->task_ctx = NULL;
1224         }
1225 }
1226
1227 static void task_ctx_sched_out(struct perf_event_context *ctx,
1228                                enum event_type_t event_type)
1229 {
1230         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1231
1232         if (!cpuctx->task_ctx)
1233                 return;
1234
1235         if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1236                 return;
1237
1238         ctx_sched_out(ctx, cpuctx, event_type);
1239         cpuctx->task_ctx = NULL;
1240 }
1241
1242 /*
1243  * Called with IRQs disabled
1244  */
1245 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1246 {
1247         task_ctx_sched_out(ctx, EVENT_ALL);
1248 }
1249
1250 /*
1251  * Called with IRQs disabled
1252  */
1253 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1254                               enum event_type_t event_type)
1255 {
1256         ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1257 }
1258
1259 static void
1260 ctx_pinned_sched_in(struct perf_event_context *ctx,
1261                     struct perf_cpu_context *cpuctx)
1262 {
1263         struct perf_event *event;
1264
1265         list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1266                 if (event->state <= PERF_EVENT_STATE_OFF)
1267                         continue;
1268                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1269                         continue;
1270
1271                 if (group_can_go_on(event, cpuctx, 1))
1272                         group_sched_in(event, cpuctx, ctx);
1273
1274                 /*
1275                  * If this pinned group hasn't been scheduled,
1276                  * put it in error state.
1277                  */
1278                 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1279                         update_group_times(event);
1280                         event->state = PERF_EVENT_STATE_ERROR;
1281                 }
1282         }
1283 }
1284
1285 static void
1286 ctx_flexible_sched_in(struct perf_event_context *ctx,
1287                       struct perf_cpu_context *cpuctx)
1288 {
1289         struct perf_event *event;
1290         int can_add_hw = 1;
1291
1292         list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1293                 /* Ignore events in OFF or ERROR state */
1294                 if (event->state <= PERF_EVENT_STATE_OFF)
1295                         continue;
1296                 /*
1297                  * Listen to the 'cpu' scheduling filter constraint
1298                  * of events:
1299                  */
1300                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1301                         continue;
1302
1303                 if (group_can_go_on(event, cpuctx, can_add_hw))
1304                         if (group_sched_in(event, cpuctx, ctx))
1305                                 can_add_hw = 0;
1306         }
1307 }
1308
1309 static void
1310 ctx_sched_in(struct perf_event_context *ctx,
1311              struct perf_cpu_context *cpuctx,
1312              enum event_type_t event_type)
1313 {
1314         raw_spin_lock(&ctx->lock);
1315         ctx->is_active = 1;
1316         if (likely(!ctx->nr_events))
1317                 goto out;
1318
1319         ctx->timestamp = perf_clock();
1320
1321         perf_disable();
1322
1323         /*
1324          * First go through the list and put on any pinned groups
1325          * in order to give them the best chance of going on.
1326          */
1327         if (event_type & EVENT_PINNED)
1328                 ctx_pinned_sched_in(ctx, cpuctx);
1329
1330         /* Then walk through the lower prio flexible groups */
1331         if (event_type & EVENT_FLEXIBLE)
1332                 ctx_flexible_sched_in(ctx, cpuctx);
1333
1334         perf_enable();
1335  out:
1336         raw_spin_unlock(&ctx->lock);
1337 }
1338
1339 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1340                              enum event_type_t event_type)
1341 {
1342         struct perf_event_context *ctx = &cpuctx->ctx;
1343
1344         ctx_sched_in(ctx, cpuctx, event_type);
1345 }
1346
1347 static void task_ctx_sched_in(struct task_struct *task,
1348                               enum event_type_t event_type)
1349 {
1350         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1351         struct perf_event_context *ctx = task->perf_event_ctxp;
1352
1353         if (likely(!ctx))
1354                 return;
1355         if (cpuctx->task_ctx == ctx)
1356                 return;
1357         ctx_sched_in(ctx, cpuctx, event_type);
1358         cpuctx->task_ctx = ctx;
1359 }
1360 /*
1361  * Called from scheduler to add the events of the current task
1362  * with interrupts disabled.
1363  *
1364  * We restore the event value and then enable it.
1365  *
1366  * This does not protect us against NMI, but enable()
1367  * sets the enabled bit in the control field of event _before_
1368  * accessing the event control register. If a NMI hits, then it will
1369  * keep the event running.
1370  */
1371 void perf_event_task_sched_in(struct task_struct *task)
1372 {
1373         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1374         struct perf_event_context *ctx = task->perf_event_ctxp;
1375
1376         if (likely(!ctx))
1377                 return;
1378
1379         if (cpuctx->task_ctx == ctx)
1380                 return;
1381
1382         perf_disable();
1383
1384         /*
1385          * We want to keep the following priority order:
1386          * cpu pinned (that don't need to move), task pinned,
1387          * cpu flexible, task flexible.
1388          */
1389         cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1390
1391         ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1392         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1393         ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1394
1395         cpuctx->task_ctx = ctx;
1396
1397         perf_enable();
1398 }
1399
1400 #define MAX_INTERRUPTS (~0ULL)
1401
1402 static void perf_log_throttle(struct perf_event *event, int enable);
1403
1404 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1405 {
1406         u64 frequency = event->attr.sample_freq;
1407         u64 sec = NSEC_PER_SEC;
1408         u64 divisor, dividend;
1409
1410         int count_fls, nsec_fls, frequency_fls, sec_fls;
1411
1412         count_fls = fls64(count);
1413         nsec_fls = fls64(nsec);
1414         frequency_fls = fls64(frequency);
1415         sec_fls = 30;
1416
1417         /*
1418          * We got @count in @nsec, with a target of sample_freq HZ
1419          * the target period becomes:
1420          *
1421          *             @count * 10^9
1422          * period = -------------------
1423          *          @nsec * sample_freq
1424          *
1425          */
1426
1427         /*
1428          * Reduce accuracy by one bit such that @a and @b converge
1429          * to a similar magnitude.
1430          */
1431 #define REDUCE_FLS(a, b)                \
1432 do {                                    \
1433         if (a##_fls > b##_fls) {        \
1434                 a >>= 1;                \
1435                 a##_fls--;              \
1436         } else {                        \
1437                 b >>= 1;                \
1438                 b##_fls--;              \
1439         }                               \
1440 } while (0)
1441
1442         /*
1443          * Reduce accuracy until either term fits in a u64, then proceed with
1444          * the other, so that finally we can do a u64/u64 division.
1445          */
1446         while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1447                 REDUCE_FLS(nsec, frequency);
1448                 REDUCE_FLS(sec, count);
1449         }
1450
1451         if (count_fls + sec_fls > 64) {
1452                 divisor = nsec * frequency;
1453
1454                 while (count_fls + sec_fls > 64) {
1455                         REDUCE_FLS(count, sec);
1456                         divisor >>= 1;
1457                 }
1458
1459                 dividend = count * sec;
1460         } else {
1461                 dividend = count * sec;
1462
1463                 while (nsec_fls + frequency_fls > 64) {
1464                         REDUCE_FLS(nsec, frequency);
1465                         dividend >>= 1;
1466                 }
1467
1468                 divisor = nsec * frequency;
1469         }
1470
1471         return div64_u64(dividend, divisor);
1472 }
1473
1474 static void perf_event_stop(struct perf_event *event)
1475 {
1476         if (!event->pmu->stop)
1477                 return event->pmu->disable(event);
1478
1479         return event->pmu->stop(event);
1480 }
1481
1482 static int perf_event_start(struct perf_event *event)
1483 {
1484         if (!event->pmu->start)
1485                 return event->pmu->enable(event);
1486
1487         return event->pmu->start(event);
1488 }
1489
1490 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1491 {
1492         struct hw_perf_event *hwc = &event->hw;
1493         u64 period, sample_period;
1494         s64 delta;
1495
1496         period = perf_calculate_period(event, nsec, count);
1497
1498         delta = (s64)(period - hwc->sample_period);
1499         delta = (delta + 7) / 8; /* low pass filter */
1500
1501         sample_period = hwc->sample_period + delta;
1502
1503         if (!sample_period)
1504                 sample_period = 1;
1505
1506         hwc->sample_period = sample_period;
1507
1508         if (atomic64_read(&hwc->period_left) > 8*sample_period) {
1509                 perf_disable();
1510                 perf_event_stop(event);
1511                 atomic64_set(&hwc->period_left, 0);
1512                 perf_event_start(event);
1513                 perf_enable();
1514         }
1515 }
1516
1517 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1518 {
1519         struct perf_event *event;
1520         struct hw_perf_event *hwc;
1521         u64 interrupts, now;
1522         s64 delta;
1523
1524         raw_spin_lock(&ctx->lock);
1525         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1526                 if (event->state != PERF_EVENT_STATE_ACTIVE)
1527                         continue;
1528
1529                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1530                         continue;
1531
1532                 hwc = &event->hw;
1533
1534                 interrupts = hwc->interrupts;
1535                 hwc->interrupts = 0;
1536
1537                 /*
1538                  * unthrottle events on the tick
1539                  */
1540                 if (interrupts == MAX_INTERRUPTS) {
1541                         perf_log_throttle(event, 1);
1542                         perf_disable();
1543                         event->pmu->unthrottle(event);
1544                         perf_enable();
1545                 }
1546
1547                 if (!event->attr.freq || !event->attr.sample_freq)
1548                         continue;
1549
1550                 perf_disable();
1551                 event->pmu->read(event);
1552                 now = atomic64_read(&event->count);
1553                 delta = now - hwc->freq_count_stamp;
1554                 hwc->freq_count_stamp = now;
1555
1556                 if (delta > 0)
1557                         perf_adjust_period(event, TICK_NSEC, delta);
1558                 perf_enable();
1559         }
1560         raw_spin_unlock(&ctx->lock);
1561 }
1562
1563 /*
1564  * Round-robin a context's events:
1565  */
1566 static void rotate_ctx(struct perf_event_context *ctx)
1567 {
1568         raw_spin_lock(&ctx->lock);
1569
1570         /* Rotate the first entry last of non-pinned groups */
1571         list_rotate_left(&ctx->flexible_groups);
1572
1573         raw_spin_unlock(&ctx->lock);
1574 }
1575
1576 void perf_event_task_tick(struct task_struct *curr)
1577 {
1578         struct perf_cpu_context *cpuctx;
1579         struct perf_event_context *ctx;
1580         int rotate = 0;
1581
1582         if (!atomic_read(&nr_events))
1583                 return;
1584
1585         cpuctx = &__get_cpu_var(perf_cpu_context);
1586         if (cpuctx->ctx.nr_events &&
1587             cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1588                 rotate = 1;
1589
1590         ctx = curr->perf_event_ctxp;
1591         if (ctx && ctx->nr_events && ctx->nr_events != ctx->nr_active)
1592                 rotate = 1;
1593
1594         perf_ctx_adjust_freq(&cpuctx->ctx);
1595         if (ctx)
1596                 perf_ctx_adjust_freq(ctx);
1597
1598         if (!rotate)
1599                 return;
1600
1601         perf_disable();
1602         cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1603         if (ctx)
1604                 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1605
1606         rotate_ctx(&cpuctx->ctx);
1607         if (ctx)
1608                 rotate_ctx(ctx);
1609
1610         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1611         if (ctx)
1612                 task_ctx_sched_in(curr, EVENT_FLEXIBLE);
1613         perf_enable();
1614 }
1615
1616 static int event_enable_on_exec(struct perf_event *event,
1617                                 struct perf_event_context *ctx)
1618 {
1619         if (!event->attr.enable_on_exec)
1620                 return 0;
1621
1622         event->attr.enable_on_exec = 0;
1623         if (event->state >= PERF_EVENT_STATE_INACTIVE)
1624                 return 0;
1625
1626         __perf_event_mark_enabled(event, ctx);
1627
1628         return 1;
1629 }
1630
1631 /*
1632  * Enable all of a task's events that have been marked enable-on-exec.
1633  * This expects task == current.
1634  */
1635 static void perf_event_enable_on_exec(struct task_struct *task)
1636 {
1637         struct perf_event_context *ctx;
1638         struct perf_event *event;
1639         unsigned long flags;
1640         int enabled = 0;
1641         int ret;
1642
1643         local_irq_save(flags);
1644         ctx = task->perf_event_ctxp;
1645         if (!ctx || !ctx->nr_events)
1646                 goto out;
1647
1648         __perf_event_task_sched_out(ctx);
1649
1650         raw_spin_lock(&ctx->lock);
1651
1652         list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1653                 ret = event_enable_on_exec(event, ctx);
1654                 if (ret)
1655                         enabled = 1;
1656         }
1657
1658         list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1659                 ret = event_enable_on_exec(event, ctx);
1660                 if (ret)
1661                         enabled = 1;
1662         }
1663
1664         /*
1665          * Unclone this context if we enabled any event.
1666          */
1667         if (enabled)
1668                 unclone_ctx(ctx);
1669
1670         raw_spin_unlock(&ctx->lock);
1671
1672         perf_event_task_sched_in(task);
1673  out:
1674         local_irq_restore(flags);
1675 }
1676
1677 /*
1678  * Cross CPU call to read the hardware event
1679  */
1680 static void __perf_event_read(void *info)
1681 {
1682         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1683         struct perf_event *event = info;
1684         struct perf_event_context *ctx = event->ctx;
1685
1686         /*
1687          * If this is a task context, we need to check whether it is
1688          * the current task context of this cpu.  If not it has been
1689          * scheduled out before the smp call arrived.  In that case
1690          * event->count would have been updated to a recent sample
1691          * when the event was scheduled out.
1692          */
1693         if (ctx->task && cpuctx->task_ctx != ctx)
1694                 return;
1695
1696         raw_spin_lock(&ctx->lock);
1697         update_context_time(ctx);
1698         update_event_times(event);
1699         raw_spin_unlock(&ctx->lock);
1700
1701         event->pmu->read(event);
1702 }
1703
1704 static u64 perf_event_read(struct perf_event *event)
1705 {
1706         /*
1707          * If event is enabled and currently active on a CPU, update the
1708          * value in the event structure:
1709          */
1710         if (event->state == PERF_EVENT_STATE_ACTIVE) {
1711                 smp_call_function_single(event->oncpu,
1712                                          __perf_event_read, event, 1);
1713         } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1714                 struct perf_event_context *ctx = event->ctx;
1715                 unsigned long flags;
1716
1717                 raw_spin_lock_irqsave(&ctx->lock, flags);
1718                 update_context_time(ctx);
1719                 update_event_times(event);
1720                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1721         }
1722
1723         return atomic64_read(&event->count);
1724 }
1725
1726 /*
1727  * Initialize the perf_event context in a task_struct:
1728  */
1729 static void
1730 __perf_event_init_context(struct perf_event_context *ctx,
1731                             struct task_struct *task)
1732 {
1733         raw_spin_lock_init(&ctx->lock);
1734         mutex_init(&ctx->mutex);
1735         INIT_LIST_HEAD(&ctx->pinned_groups);
1736         INIT_LIST_HEAD(&ctx->flexible_groups);
1737         INIT_LIST_HEAD(&ctx->event_list);
1738         atomic_set(&ctx->refcount, 1);
1739         ctx->task = task;
1740 }
1741
1742 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1743 {
1744         struct perf_event_context *ctx;
1745         struct perf_cpu_context *cpuctx;
1746         struct task_struct *task;
1747         unsigned long flags;
1748         int err;
1749
1750         if (pid == -1 && cpu != -1) {
1751                 /* Must be root to operate on a CPU event: */
1752                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1753                         return ERR_PTR(-EACCES);
1754
1755                 if (cpu < 0 || cpu >= nr_cpumask_bits)
1756                         return ERR_PTR(-EINVAL);
1757
1758                 /*
1759                  * We could be clever and allow to attach a event to an
1760                  * offline CPU and activate it when the CPU comes up, but
1761                  * that's for later.
1762                  */
1763                 if (!cpu_online(cpu))
1764                         return ERR_PTR(-ENODEV);
1765
1766                 cpuctx = &per_cpu(perf_cpu_context, cpu);
1767                 ctx = &cpuctx->ctx;
1768                 get_ctx(ctx);
1769
1770                 return ctx;
1771         }
1772
1773         rcu_read_lock();
1774         if (!pid)
1775                 task = current;
1776         else
1777                 task = find_task_by_vpid(pid);
1778         if (task)
1779                 get_task_struct(task);
1780         rcu_read_unlock();
1781
1782         if (!task)
1783                 return ERR_PTR(-ESRCH);
1784
1785         /*
1786          * Can't attach events to a dying task.
1787          */
1788         err = -ESRCH;
1789         if (task->flags & PF_EXITING)
1790                 goto errout;
1791
1792         /* Reuse ptrace permission checks for now. */
1793         err = -EACCES;
1794         if (!ptrace_may_access(task, PTRACE_MODE_READ))
1795                 goto errout;
1796
1797  retry:
1798         ctx = perf_lock_task_context(task, &flags);
1799         if (ctx) {
1800                 unclone_ctx(ctx);
1801                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1802         }
1803
1804         if (!ctx) {
1805                 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1806                 err = -ENOMEM;
1807                 if (!ctx)
1808                         goto errout;
1809                 __perf_event_init_context(ctx, task);
1810                 get_ctx(ctx);
1811                 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1812                         /*
1813                          * We raced with some other task; use
1814                          * the context they set.
1815                          */
1816                         kfree(ctx);
1817                         goto retry;
1818                 }
1819                 get_task_struct(task);
1820         }
1821
1822         put_task_struct(task);
1823         return ctx;
1824
1825  errout:
1826         put_task_struct(task);
1827         return ERR_PTR(err);
1828 }
1829
1830 static void perf_event_free_filter(struct perf_event *event);
1831
1832 static void free_event_rcu(struct rcu_head *head)
1833 {
1834         struct perf_event *event;
1835
1836         event = container_of(head, struct perf_event, rcu_head);
1837         if (event->ns)
1838                 put_pid_ns(event->ns);
1839         perf_event_free_filter(event);
1840         kfree(event);
1841 }
1842
1843 static void perf_pending_sync(struct perf_event *event);
1844 static void perf_mmap_data_put(struct perf_mmap_data *data);
1845
1846 static void free_event(struct perf_event *event)
1847 {
1848         perf_pending_sync(event);
1849
1850         if (!event->parent) {
1851                 atomic_dec(&nr_events);
1852                 if (event->attr.mmap)
1853                         atomic_dec(&nr_mmap_events);
1854                 if (event->attr.comm)
1855                         atomic_dec(&nr_comm_events);
1856                 if (event->attr.task)
1857                         atomic_dec(&nr_task_events);
1858         }
1859
1860         if (event->data) {
1861                 perf_mmap_data_put(event->data);
1862                 event->data = NULL;
1863         }
1864
1865         if (event->destroy)
1866                 event->destroy(event);
1867
1868         put_ctx(event->ctx);
1869         call_rcu(&event->rcu_head, free_event_rcu);
1870 }
1871
1872 int perf_event_release_kernel(struct perf_event *event)
1873 {
1874         struct perf_event_context *ctx = event->ctx;
1875
1876         /*
1877          * Remove from the PMU, can't get re-enabled since we got
1878          * here because the last ref went.
1879          */
1880         perf_event_disable(event);
1881
1882         WARN_ON_ONCE(ctx->parent_ctx);
1883         /*
1884          * There are two ways this annotation is useful:
1885          *
1886          *  1) there is a lock recursion from perf_event_exit_task
1887          *     see the comment there.
1888          *
1889          *  2) there is a lock-inversion with mmap_sem through
1890          *     perf_event_read_group(), which takes faults while
1891          *     holding ctx->mutex, however this is called after
1892          *     the last filedesc died, so there is no possibility
1893          *     to trigger the AB-BA case.
1894          */
1895         mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
1896         raw_spin_lock_irq(&ctx->lock);
1897         list_del_event(event, ctx);
1898         perf_destroy_group(event, ctx);
1899         raw_spin_unlock_irq(&ctx->lock);
1900         mutex_unlock(&ctx->mutex);
1901
1902         mutex_lock(&event->owner->perf_event_mutex);
1903         list_del_init(&event->owner_entry);
1904         mutex_unlock(&event->owner->perf_event_mutex);
1905         put_task_struct(event->owner);
1906
1907         free_event(event);
1908
1909         return 0;
1910 }
1911 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1912
1913 /*
1914  * Called when the last reference to the file is gone.
1915  */
1916 static int perf_release(struct inode *inode, struct file *file)
1917 {
1918         struct perf_event *event = file->private_data;
1919
1920         file->private_data = NULL;
1921
1922         return perf_event_release_kernel(event);
1923 }
1924
1925 static int perf_event_read_size(struct perf_event *event)
1926 {
1927         int entry = sizeof(u64); /* value */
1928         int size = 0;
1929         int nr = 1;
1930
1931         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1932                 size += sizeof(u64);
1933
1934         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1935                 size += sizeof(u64);
1936
1937         if (event->attr.read_format & PERF_FORMAT_ID)
1938                 entry += sizeof(u64);
1939
1940         if (event->attr.read_format & PERF_FORMAT_GROUP) {
1941                 nr += event->group_leader->nr_siblings;
1942                 size += sizeof(u64);
1943         }
1944
1945         size += entry * nr;
1946
1947         return size;
1948 }
1949
1950 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1951 {
1952         struct perf_event *child;
1953         u64 total = 0;
1954
1955         *enabled = 0;
1956         *running = 0;
1957
1958         mutex_lock(&event->child_mutex);
1959         total += perf_event_read(event);
1960         *enabled += event->total_time_enabled +
1961                         atomic64_read(&event->child_total_time_enabled);
1962         *running += event->total_time_running +
1963                         atomic64_read(&event->child_total_time_running);
1964
1965         list_for_each_entry(child, &event->child_list, child_list) {
1966                 total += perf_event_read(child);
1967                 *enabled += child->total_time_enabled;
1968                 *running += child->total_time_running;
1969         }
1970         mutex_unlock(&event->child_mutex);
1971
1972         return total;
1973 }
1974 EXPORT_SYMBOL_GPL(perf_event_read_value);
1975
1976 static int perf_event_read_group(struct perf_event *event,
1977                                    u64 read_format, char __user *buf)
1978 {
1979         struct perf_event *leader = event->group_leader, *sub;
1980         int n = 0, size = 0, ret = -EFAULT;
1981         struct perf_event_context *ctx = leader->ctx;
1982         u64 values[5];
1983         u64 count, enabled, running;
1984
1985         mutex_lock(&ctx->mutex);
1986         count = perf_event_read_value(leader, &enabled, &running);
1987
1988         values[n++] = 1 + leader->nr_siblings;
1989         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1990                 values[n++] = enabled;
1991         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1992                 values[n++] = running;
1993         values[n++] = count;
1994         if (read_format & PERF_FORMAT_ID)
1995                 values[n++] = primary_event_id(leader);
1996
1997         size = n * sizeof(u64);
1998
1999         if (copy_to_user(buf, values, size))
2000                 goto unlock;
2001
2002         ret = size;
2003
2004         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2005                 n = 0;
2006
2007                 values[n++] = perf_event_read_value(sub, &enabled, &running);
2008                 if (read_format & PERF_FORMAT_ID)
2009                         values[n++] = primary_event_id(sub);
2010
2011                 size = n * sizeof(u64);
2012
2013                 if (copy_to_user(buf + ret, values, size)) {
2014                         ret = -EFAULT;
2015                         goto unlock;
2016                 }
2017
2018                 ret += size;
2019         }
2020 unlock:
2021         mutex_unlock(&ctx->mutex);
2022
2023         return ret;
2024 }
2025
2026 static int perf_event_read_one(struct perf_event *event,
2027                                  u64 read_format, char __user *buf)
2028 {
2029         u64 enabled, running;
2030         u64 values[4];
2031         int n = 0;
2032
2033         values[n++] = perf_event_read_value(event, &enabled, &running);
2034         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2035                 values[n++] = enabled;
2036         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2037                 values[n++] = running;
2038         if (read_format & PERF_FORMAT_ID)
2039                 values[n++] = primary_event_id(event);
2040
2041         if (copy_to_user(buf, values, n * sizeof(u64)))
2042                 return -EFAULT;
2043
2044         return n * sizeof(u64);
2045 }
2046
2047 /*
2048  * Read the performance event - simple non blocking version for now
2049  */
2050 static ssize_t
2051 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2052 {
2053         u64 read_format = event->attr.read_format;
2054         int ret;
2055
2056         /*
2057          * Return end-of-file for a read on a event that is in
2058          * error state (i.e. because it was pinned but it couldn't be
2059          * scheduled on to the CPU at some point).
2060          */
2061         if (event->state == PERF_EVENT_STATE_ERROR)
2062                 return 0;
2063
2064         if (count < perf_event_read_size(event))
2065                 return -ENOSPC;
2066
2067         WARN_ON_ONCE(event->ctx->parent_ctx);
2068         if (read_format & PERF_FORMAT_GROUP)
2069                 ret = perf_event_read_group(event, read_format, buf);
2070         else
2071                 ret = perf_event_read_one(event, read_format, buf);
2072
2073         return ret;
2074 }
2075
2076 static ssize_t
2077 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2078 {
2079         struct perf_event *event = file->private_data;
2080
2081         return perf_read_hw(event, buf, count);
2082 }
2083
2084 static unsigned int perf_poll(struct file *file, poll_table *wait)
2085 {
2086         struct perf_event *event = file->private_data;
2087         struct perf_mmap_data *data;
2088         unsigned int events = POLL_HUP;
2089
2090         rcu_read_lock();
2091         data = rcu_dereference(event->data);
2092         if (data)
2093                 events = atomic_xchg(&data->poll, 0);
2094         rcu_read_unlock();
2095
2096         poll_wait(file, &event->waitq, wait);
2097
2098         return events;
2099 }
2100
2101 static void perf_event_reset(struct perf_event *event)
2102 {
2103         (void)perf_event_read(event);
2104         atomic64_set(&event->count, 0);
2105         perf_event_update_userpage(event);
2106 }
2107
2108 /*
2109  * Holding the top-level event's child_mutex means that any
2110  * descendant process that has inherited this event will block
2111  * in sync_child_event if it goes to exit, thus satisfying the
2112  * task existence requirements of perf_event_enable/disable.
2113  */
2114 static void perf_event_for_each_child(struct perf_event *event,
2115                                         void (*func)(struct perf_event *))
2116 {
2117         struct perf_event *child;
2118
2119         WARN_ON_ONCE(event->ctx->parent_ctx);
2120         mutex_lock(&event->child_mutex);
2121         func(event);
2122         list_for_each_entry(child, &event->child_list, child_list)
2123                 func(child);
2124         mutex_unlock(&event->child_mutex);
2125 }
2126
2127 static void perf_event_for_each(struct perf_event *event,
2128                                   void (*func)(struct perf_event *))
2129 {
2130         struct perf_event_context *ctx = event->ctx;
2131         struct perf_event *sibling;
2132
2133         WARN_ON_ONCE(ctx->parent_ctx);
2134         mutex_lock(&ctx->mutex);
2135         event = event->group_leader;
2136
2137         perf_event_for_each_child(event, func);
2138         func(event);
2139         list_for_each_entry(sibling, &event->sibling_list, group_entry)
2140                 perf_event_for_each_child(event, func);
2141         mutex_unlock(&ctx->mutex);
2142 }
2143
2144 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2145 {
2146         struct perf_event_context *ctx = event->ctx;
2147         unsigned long size;
2148         int ret = 0;
2149         u64 value;
2150
2151         if (!event->attr.sample_period)
2152                 return -EINVAL;
2153
2154         size = copy_from_user(&value, arg, sizeof(value));
2155         if (size != sizeof(value))
2156                 return -EFAULT;
2157
2158         if (!value)
2159                 return -EINVAL;
2160
2161         raw_spin_lock_irq(&ctx->lock);
2162         if (event->attr.freq) {
2163                 if (value > sysctl_perf_event_sample_rate) {
2164                         ret = -EINVAL;
2165                         goto unlock;
2166                 }
2167
2168                 event->attr.sample_freq = value;
2169         } else {
2170                 event->attr.sample_period = value;
2171                 event->hw.sample_period = value;
2172         }
2173 unlock:
2174         raw_spin_unlock_irq(&ctx->lock);
2175
2176         return ret;
2177 }
2178
2179 static const struct file_operations perf_fops;
2180
2181 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
2182 {
2183         struct file *file;
2184
2185         file = fget_light(fd, fput_needed);
2186         if (!file)
2187                 return ERR_PTR(-EBADF);
2188
2189         if (file->f_op != &perf_fops) {
2190                 fput_light(file, *fput_needed);
2191                 *fput_needed = 0;
2192                 return ERR_PTR(-EBADF);
2193         }
2194
2195         return file->private_data;
2196 }
2197
2198 static int perf_event_set_output(struct perf_event *event,
2199                                  struct perf_event *output_event);
2200 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2201
2202 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2203 {
2204         struct perf_event *event = file->private_data;
2205         void (*func)(struct perf_event *);
2206         u32 flags = arg;
2207
2208         switch (cmd) {
2209         case PERF_EVENT_IOC_ENABLE:
2210                 func = perf_event_enable;
2211                 break;
2212         case PERF_EVENT_IOC_DISABLE:
2213                 func = perf_event_disable;
2214                 break;
2215         case PERF_EVENT_IOC_RESET:
2216                 func = perf_event_reset;
2217                 break;
2218
2219         case PERF_EVENT_IOC_REFRESH:
2220                 return perf_event_refresh(event, arg);
2221
2222         case PERF_EVENT_IOC_PERIOD:
2223                 return perf_event_period(event, (u64 __user *)arg);
2224
2225         case PERF_EVENT_IOC_SET_OUTPUT:
2226         {
2227                 struct perf_event *output_event = NULL;
2228                 int fput_needed = 0;
2229                 int ret;
2230
2231                 if (arg != -1) {
2232                         output_event = perf_fget_light(arg, &fput_needed);
2233                         if (IS_ERR(output_event))
2234                                 return PTR_ERR(output_event);
2235                 }
2236
2237                 ret = perf_event_set_output(event, output_event);
2238                 if (output_event)
2239                         fput_light(output_event->filp, fput_needed);
2240
2241                 return ret;
2242         }
2243
2244         case PERF_EVENT_IOC_SET_FILTER:
2245                 return perf_event_set_filter(event, (void __user *)arg);
2246
2247         default:
2248                 return -ENOTTY;
2249         }
2250
2251         if (flags & PERF_IOC_FLAG_GROUP)
2252                 perf_event_for_each(event, func);
2253         else
2254                 perf_event_for_each_child(event, func);
2255
2256         return 0;
2257 }
2258
2259 int perf_event_task_enable(void)
2260 {
2261         struct perf_event *event;
2262
2263         mutex_lock(&current->perf_event_mutex);
2264         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2265                 perf_event_for_each_child(event, perf_event_enable);
2266         mutex_unlock(&current->perf_event_mutex);
2267
2268         return 0;
2269 }
2270
2271 int perf_event_task_disable(void)
2272 {
2273         struct perf_event *event;
2274
2275         mutex_lock(&current->perf_event_mutex);
2276         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2277                 perf_event_for_each_child(event, perf_event_disable);
2278         mutex_unlock(&current->perf_event_mutex);
2279
2280         return 0;
2281 }
2282
2283 #ifndef PERF_EVENT_INDEX_OFFSET
2284 # define PERF_EVENT_INDEX_OFFSET 0
2285 #endif
2286
2287 static int perf_event_index(struct perf_event *event)
2288 {
2289         if (event->state != PERF_EVENT_STATE_ACTIVE)
2290                 return 0;
2291
2292         return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2293 }
2294
2295 /*
2296  * Callers need to ensure there can be no nesting of this function, otherwise
2297  * the seqlock logic goes bad. We can not serialize this because the arch
2298  * code calls this from NMI context.
2299  */
2300 void perf_event_update_userpage(struct perf_event *event)
2301 {
2302         struct perf_event_mmap_page *userpg;
2303         struct perf_mmap_data *data;
2304
2305         rcu_read_lock();
2306         data = rcu_dereference(event->data);
2307         if (!data)
2308                 goto unlock;
2309
2310         userpg = data->user_page;
2311
2312         /*
2313          * Disable preemption so as to not let the corresponding user-space
2314          * spin too long if we get preempted.
2315          */
2316         preempt_disable();
2317         ++userpg->lock;
2318         barrier();
2319         userpg->index = perf_event_index(event);
2320         userpg->offset = atomic64_read(&event->count);
2321         if (event->state == PERF_EVENT_STATE_ACTIVE)
2322                 userpg->offset -= atomic64_read(&event->hw.prev_count);
2323
2324         userpg->time_enabled = event->total_time_enabled +
2325                         atomic64_read(&event->child_total_time_enabled);
2326
2327         userpg->time_running = event->total_time_running +
2328                         atomic64_read(&event->child_total_time_running);
2329
2330         barrier();
2331         ++userpg->lock;
2332         preempt_enable();
2333 unlock:
2334         rcu_read_unlock();
2335 }
2336
2337 #ifndef CONFIG_PERF_USE_VMALLOC
2338
2339 /*
2340  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2341  */
2342
2343 static struct page *
2344 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2345 {
2346         if (pgoff > data->nr_pages)
2347                 return NULL;
2348
2349         if (pgoff == 0)
2350                 return virt_to_page(data->user_page);
2351
2352         return virt_to_page(data->data_pages[pgoff - 1]);
2353 }
2354
2355 static void *perf_mmap_alloc_page(int cpu)
2356 {
2357         struct page *page;
2358         int node;
2359
2360         node = (cpu == -1) ? cpu : cpu_to_node(cpu);
2361         page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
2362         if (!page)
2363                 return NULL;
2364
2365         return page_address(page);
2366 }
2367
2368 static struct perf_mmap_data *
2369 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2370 {
2371         struct perf_mmap_data *data;
2372         unsigned long size;
2373         int i;
2374
2375         size = sizeof(struct perf_mmap_data);
2376         size += nr_pages * sizeof(void *);
2377
2378         data = kzalloc(size, GFP_KERNEL);
2379         if (!data)
2380                 goto fail;
2381
2382         data->user_page = perf_mmap_alloc_page(event->cpu);
2383         if (!data->user_page)
2384                 goto fail_user_page;
2385
2386         for (i = 0; i < nr_pages; i++) {
2387                 data->data_pages[i] = perf_mmap_alloc_page(event->cpu);
2388                 if (!data->data_pages[i])
2389                         goto fail_data_pages;
2390         }
2391
2392         data->nr_pages = nr_pages;
2393
2394         return data;
2395
2396 fail_data_pages:
2397         for (i--; i >= 0; i--)
2398                 free_page((unsigned long)data->data_pages[i]);
2399
2400         free_page((unsigned long)data->user_page);
2401
2402 fail_user_page:
2403         kfree(data);
2404
2405 fail:
2406         return NULL;
2407 }
2408
2409 static void perf_mmap_free_page(unsigned long addr)
2410 {
2411         struct page *page = virt_to_page((void *)addr);
2412
2413         page->mapping = NULL;
2414         __free_page(page);
2415 }
2416
2417 static void perf_mmap_data_free(struct perf_mmap_data *data)
2418 {
2419         int i;
2420
2421         perf_mmap_free_page((unsigned long)data->user_page);
2422         for (i = 0; i < data->nr_pages; i++)
2423                 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2424         kfree(data);
2425 }
2426
2427 static inline int page_order(struct perf_mmap_data *data)
2428 {
2429         return 0;
2430 }
2431
2432 #else
2433
2434 /*
2435  * Back perf_mmap() with vmalloc memory.
2436  *
2437  * Required for architectures that have d-cache aliasing issues.
2438  */
2439
2440 static inline int page_order(struct perf_mmap_data *data)
2441 {
2442         return data->page_order;
2443 }
2444
2445 static struct page *
2446 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2447 {
2448         if (pgoff > (1UL << page_order(data)))
2449                 return NULL;
2450
2451         return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2452 }
2453
2454 static void perf_mmap_unmark_page(void *addr)
2455 {
2456         struct page *page = vmalloc_to_page(addr);
2457
2458         page->mapping = NULL;
2459 }
2460
2461 static void perf_mmap_data_free_work(struct work_struct *work)
2462 {
2463         struct perf_mmap_data *data;
2464         void *base;
2465         int i, nr;
2466
2467         data = container_of(work, struct perf_mmap_data, work);
2468         nr = 1 << page_order(data);
2469
2470         base = data->user_page;
2471         for (i = 0; i < nr + 1; i++)
2472                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2473
2474         vfree(base);
2475         kfree(data);
2476 }
2477
2478 static void perf_mmap_data_free(struct perf_mmap_data *data)
2479 {
2480         schedule_work(&data->work);
2481 }
2482
2483 static struct perf_mmap_data *
2484 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2485 {
2486         struct perf_mmap_data *data;
2487         unsigned long size;
2488         void *all_buf;
2489
2490         size = sizeof(struct perf_mmap_data);
2491         size += sizeof(void *);
2492
2493         data = kzalloc(size, GFP_KERNEL);
2494         if (!data)
2495                 goto fail;
2496
2497         INIT_WORK(&data->work, perf_mmap_data_free_work);
2498
2499         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2500         if (!all_buf)
2501                 goto fail_all_buf;
2502
2503         data->user_page = all_buf;
2504         data->data_pages[0] = all_buf + PAGE_SIZE;
2505         data->page_order = ilog2(nr_pages);
2506         data->nr_pages = 1;
2507
2508         return data;
2509
2510 fail_all_buf:
2511         kfree(data);
2512
2513 fail:
2514         return NULL;
2515 }
2516
2517 #endif
2518
2519 static unsigned long perf_data_size(struct perf_mmap_data *data)
2520 {
2521         return data->nr_pages << (PAGE_SHIFT + page_order(data));
2522 }
2523
2524 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2525 {
2526         struct perf_event *event = vma->vm_file->private_data;
2527         struct perf_mmap_data *data;
2528         int ret = VM_FAULT_SIGBUS;
2529
2530         if (vmf->flags & FAULT_FLAG_MKWRITE) {
2531                 if (vmf->pgoff == 0)
2532                         ret = 0;
2533                 return ret;
2534         }
2535
2536         rcu_read_lock();
2537         data = rcu_dereference(event->data);
2538         if (!data)
2539                 goto unlock;
2540
2541         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2542                 goto unlock;
2543
2544         vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2545         if (!vmf->page)
2546                 goto unlock;
2547
2548         get_page(vmf->page);
2549         vmf->page->mapping = vma->vm_file->f_mapping;
2550         vmf->page->index   = vmf->pgoff;
2551
2552         ret = 0;
2553 unlock:
2554         rcu_read_unlock();
2555
2556         return ret;
2557 }
2558
2559 static void
2560 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2561 {
2562         long max_size = perf_data_size(data);
2563
2564         if (event->attr.watermark) {
2565                 data->watermark = min_t(long, max_size,
2566                                         event->attr.wakeup_watermark);
2567         }
2568
2569         if (!data->watermark)
2570                 data->watermark = max_size / 2;
2571
2572         atomic_set(&data->refcount, 1);
2573         rcu_assign_pointer(event->data, data);
2574 }
2575
2576 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2577 {
2578         struct perf_mmap_data *data;
2579
2580         data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2581         perf_mmap_data_free(data);
2582 }
2583
2584 static struct perf_mmap_data *perf_mmap_data_get(struct perf_event *event)
2585 {
2586         struct perf_mmap_data *data;
2587
2588         rcu_read_lock();
2589         data = rcu_dereference(event->data);
2590         if (data) {
2591                 if (!atomic_inc_not_zero(&data->refcount))
2592                         data = NULL;
2593         }
2594         rcu_read_unlock();
2595
2596         return data;
2597 }
2598
2599 static void perf_mmap_data_put(struct perf_mmap_data *data)
2600 {
2601         if (!atomic_dec_and_test(&data->refcount))
2602                 return;
2603
2604         call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2605 }
2606
2607 static void perf_mmap_open(struct vm_area_struct *vma)
2608 {
2609         struct perf_event *event = vma->vm_file->private_data;
2610
2611         atomic_inc(&event->mmap_count);
2612 }
2613
2614 static void perf_mmap_close(struct vm_area_struct *vma)
2615 {
2616         struct perf_event *event = vma->vm_file->private_data;
2617
2618         if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2619                 unsigned long size = perf_data_size(event->data);
2620                 struct user_struct *user = event->mmap_user;
2621                 struct perf_mmap_data *data = event->data;
2622
2623                 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2624                 vma->vm_mm->locked_vm -= event->mmap_locked;
2625                 rcu_assign_pointer(event->data, NULL);
2626                 mutex_unlock(&event->mmap_mutex);
2627
2628                 perf_mmap_data_put(data);
2629                 free_uid(user);
2630         }
2631 }
2632
2633 static const struct vm_operations_struct perf_mmap_vmops = {
2634         .open           = perf_mmap_open,
2635         .close          = perf_mmap_close,
2636         .fault          = perf_mmap_fault,
2637         .page_mkwrite   = perf_mmap_fault,
2638 };
2639
2640 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2641 {
2642         struct perf_event *event = file->private_data;
2643         unsigned long user_locked, user_lock_limit;
2644         struct user_struct *user = current_user();
2645         unsigned long locked, lock_limit;
2646         struct perf_mmap_data *data;
2647         unsigned long vma_size;
2648         unsigned long nr_pages;
2649         long user_extra, extra;
2650         int ret = 0;
2651
2652         /*
2653          * Don't allow mmap() of inherited per-task counters. This would
2654          * create a performance issue due to all children writing to the
2655          * same buffer.
2656          */
2657         if (event->cpu == -1 && event->attr.inherit)
2658                 return -EINVAL;
2659
2660         if (!(vma->vm_flags & VM_SHARED))
2661                 return -EINVAL;
2662
2663         vma_size = vma->vm_end - vma->vm_start;
2664         nr_pages = (vma_size / PAGE_SIZE) - 1;
2665
2666         /*
2667          * If we have data pages ensure they're a power-of-two number, so we
2668          * can do bitmasks instead of modulo.
2669          */
2670         if (nr_pages != 0 && !is_power_of_2(nr_pages))
2671                 return -EINVAL;
2672
2673         if (vma_size != PAGE_SIZE * (1 + nr_pages))
2674                 return -EINVAL;
2675
2676         if (vma->vm_pgoff != 0)
2677                 return -EINVAL;
2678
2679         WARN_ON_ONCE(event->ctx->parent_ctx);
2680         mutex_lock(&event->mmap_mutex);
2681         if (event->data) {
2682                 if (event->data->nr_pages == nr_pages)
2683                         atomic_inc(&event->data->refcount);
2684                 else
2685                         ret = -EINVAL;
2686                 goto unlock;
2687         }
2688
2689         user_extra = nr_pages + 1;
2690         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2691
2692         /*
2693          * Increase the limit linearly with more CPUs:
2694          */
2695         user_lock_limit *= num_online_cpus();
2696
2697         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2698
2699         extra = 0;
2700         if (user_locked > user_lock_limit)
2701                 extra = user_locked - user_lock_limit;
2702
2703         lock_limit = rlimit(RLIMIT_MEMLOCK);
2704         lock_limit >>= PAGE_SHIFT;
2705         locked = vma->vm_mm->locked_vm + extra;
2706
2707         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2708                 !capable(CAP_IPC_LOCK)) {
2709                 ret = -EPERM;
2710                 goto unlock;
2711         }
2712
2713         WARN_ON(event->data);
2714
2715         data = perf_mmap_data_alloc(event, nr_pages);
2716         if (!data) {
2717                 ret = -ENOMEM;
2718                 goto unlock;
2719         }
2720
2721         perf_mmap_data_init(event, data);
2722         if (vma->vm_flags & VM_WRITE)
2723                 event->data->writable = 1;
2724
2725         atomic_long_add(user_extra, &user->locked_vm);
2726         event->mmap_locked = extra;
2727         event->mmap_user = get_current_user();
2728         vma->vm_mm->locked_vm += event->mmap_locked;
2729
2730 unlock:
2731         if (!ret)
2732                 atomic_inc(&event->mmap_count);
2733         mutex_unlock(&event->mmap_mutex);
2734
2735         vma->vm_flags |= VM_RESERVED;
2736         vma->vm_ops = &perf_mmap_vmops;
2737
2738         return ret;
2739 }
2740
2741 static int perf_fasync(int fd, struct file *filp, int on)
2742 {
2743         struct inode *inode = filp->f_path.dentry->d_inode;
2744         struct perf_event *event = filp->private_data;
2745         int retval;
2746
2747         mutex_lock(&inode->i_mutex);
2748         retval = fasync_helper(fd, filp, on, &event->fasync);
2749         mutex_unlock(&inode->i_mutex);
2750
2751         if (retval < 0)
2752                 return retval;
2753
2754         return 0;
2755 }
2756
2757 static const struct file_operations perf_fops = {
2758         .llseek                 = no_llseek,
2759         .release                = perf_release,
2760         .read                   = perf_read,
2761         .poll                   = perf_poll,
2762         .unlocked_ioctl         = perf_ioctl,
2763         .compat_ioctl           = perf_ioctl,
2764         .mmap                   = perf_mmap,
2765         .fasync                 = perf_fasync,
2766 };
2767
2768 /*
2769  * Perf event wakeup
2770  *
2771  * If there's data, ensure we set the poll() state and publish everything
2772  * to user-space before waking everybody up.
2773  */
2774
2775 void perf_event_wakeup(struct perf_event *event)
2776 {
2777         wake_up_all(&event->waitq);
2778
2779         if (event->pending_kill) {
2780                 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2781                 event->pending_kill = 0;
2782         }
2783 }
2784
2785 /*
2786  * Pending wakeups
2787  *
2788  * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2789  *
2790  * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2791  * single linked list and use cmpxchg() to add entries lockless.
2792  */
2793
2794 static void perf_pending_event(struct perf_pending_entry *entry)
2795 {
2796         struct perf_event *event = container_of(entry,
2797                         struct perf_event, pending);
2798
2799         if (event->pending_disable) {
2800                 event->pending_disable = 0;
2801                 __perf_event_disable(event);
2802         }
2803
2804         if (event->pending_wakeup) {
2805                 event->pending_wakeup = 0;
2806                 perf_event_wakeup(event);
2807         }
2808 }
2809
2810 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2811
2812 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2813         PENDING_TAIL,
2814 };
2815
2816 static void perf_pending_queue(struct perf_pending_entry *entry,
2817                                void (*func)(struct perf_pending_entry *))
2818 {
2819         struct perf_pending_entry **head;
2820
2821         if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2822                 return;
2823
2824         entry->func = func;
2825
2826         head = &get_cpu_var(perf_pending_head);
2827
2828         do {
2829                 entry->next = *head;
2830         } while (cmpxchg(head, entry->next, entry) != entry->next);
2831
2832         set_perf_event_pending();
2833
2834         put_cpu_var(perf_pending_head);
2835 }
2836
2837 static int __perf_pending_run(void)
2838 {
2839         struct perf_pending_entry *list;
2840         int nr = 0;
2841
2842         list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2843         while (list != PENDING_TAIL) {
2844                 void (*func)(struct perf_pending_entry *);
2845                 struct perf_pending_entry *entry = list;
2846
2847                 list = list->next;
2848
2849                 func = entry->func;
2850                 entry->next = NULL;
2851                 /*
2852                  * Ensure we observe the unqueue before we issue the wakeup,
2853                  * so that we won't be waiting forever.
2854                  * -- see perf_not_pending().
2855                  */
2856                 smp_wmb();
2857
2858                 func(entry);
2859                 nr++;
2860         }
2861
2862         return nr;
2863 }
2864
2865 static inline int perf_not_pending(struct perf_event *event)
2866 {
2867         /*
2868          * If we flush on whatever cpu we run, there is a chance we don't
2869          * need to wait.
2870          */
2871         get_cpu();
2872         __perf_pending_run();
2873         put_cpu();
2874
2875         /*
2876          * Ensure we see the proper queue state before going to sleep
2877          * so that we do not miss the wakeup. -- see perf_pending_handle()
2878          */
2879         smp_rmb();
2880         return event->pending.next == NULL;
2881 }
2882
2883 static void perf_pending_sync(struct perf_event *event)
2884 {
2885         wait_event(event->waitq, perf_not_pending(event));
2886 }
2887
2888 void perf_event_do_pending(void)
2889 {
2890         __perf_pending_run();
2891 }
2892
2893 /*
2894  * Callchain support -- arch specific
2895  */
2896
2897 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2898 {
2899         return NULL;
2900 }
2901
2902 __weak
2903 void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip, int skip)
2904 {
2905 }
2906
2907
2908 /*
2909  * We assume there is only KVM supporting the callbacks.
2910  * Later on, we might change it to a list if there is
2911  * another virtualization implementation supporting the callbacks.
2912  */
2913 struct perf_guest_info_callbacks *perf_guest_cbs;
2914
2915 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2916 {
2917         perf_guest_cbs = cbs;
2918         return 0;
2919 }
2920 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
2921
2922 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2923 {
2924         perf_guest_cbs = NULL;
2925         return 0;
2926 }
2927 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
2928
2929 /*
2930  * Output
2931  */
2932 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2933                               unsigned long offset, unsigned long head)
2934 {
2935         unsigned long mask;
2936
2937         if (!data->writable)
2938                 return true;
2939
2940         mask = perf_data_size(data) - 1;
2941
2942         offset = (offset - tail) & mask;
2943         head   = (head   - tail) & mask;
2944
2945         if ((int)(head - offset) < 0)
2946                 return false;
2947
2948         return true;
2949 }
2950
2951 static void perf_output_wakeup(struct perf_output_handle *handle)
2952 {
2953         atomic_set(&handle->data->poll, POLL_IN);
2954
2955         if (handle->nmi) {
2956                 handle->event->pending_wakeup = 1;
2957                 perf_pending_queue(&handle->event->pending,
2958                                    perf_pending_event);
2959         } else
2960                 perf_event_wakeup(handle->event);
2961 }
2962
2963 /*
2964  * We need to ensure a later event_id doesn't publish a head when a former
2965  * event isn't done writing. However since we need to deal with NMIs we
2966  * cannot fully serialize things.
2967  *
2968  * We only publish the head (and generate a wakeup) when the outer-most
2969  * event completes.
2970  */
2971 static void perf_output_get_handle(struct perf_output_handle *handle)
2972 {
2973         struct perf_mmap_data *data = handle->data;
2974
2975         preempt_disable();
2976         local_inc(&data->nest);
2977         handle->wakeup = local_read(&data->wakeup);
2978 }
2979
2980 static void perf_output_put_handle(struct perf_output_handle *handle)
2981 {
2982         struct perf_mmap_data *data = handle->data;
2983         unsigned long head;
2984
2985 again:
2986         head = local_read(&data->head);
2987
2988         /*
2989          * IRQ/NMI can happen here, which means we can miss a head update.
2990          */
2991
2992         if (!local_dec_and_test(&data->nest))
2993                 goto out;
2994
2995         /*
2996          * Publish the known good head. Rely on the full barrier implied
2997          * by atomic_dec_and_test() order the data->head read and this
2998          * write.
2999          */
3000         data->user_page->data_head = head;
3001
3002         /*
3003          * Now check if we missed an update, rely on the (compiler)
3004          * barrier in atomic_dec_and_test() to re-read data->head.
3005          */
3006         if (unlikely(head != local_read(&data->head))) {
3007                 local_inc(&data->nest);
3008                 goto again;
3009         }
3010
3011         if (handle->wakeup != local_read(&data->wakeup))
3012                 perf_output_wakeup(handle);
3013
3014  out:
3015         preempt_enable();
3016 }
3017
3018 __always_inline void perf_output_copy(struct perf_output_handle *handle,
3019                       const void *buf, unsigned int len)
3020 {
3021         do {
3022                 unsigned long size = min_t(unsigned long, handle->size, len);
3023
3024                 memcpy(handle->addr, buf, size);
3025
3026                 len -= size;
3027                 handle->addr += size;
3028                 handle->size -= size;
3029                 if (!handle->size) {
3030                         struct perf_mmap_data *data = handle->data;
3031
3032                         handle->page++;
3033                         handle->page &= data->nr_pages - 1;
3034                         handle->addr = data->data_pages[handle->page];
3035                         handle->size = PAGE_SIZE << page_order(data);
3036                 }
3037         } while (len);
3038 }
3039
3040 int perf_output_begin(struct perf_output_handle *handle,
3041                       struct perf_event *event, unsigned int size,
3042                       int nmi, int sample)
3043 {
3044         struct perf_mmap_data *data;
3045         unsigned long tail, offset, head;
3046         int have_lost;
3047         struct {
3048                 struct perf_event_header header;
3049                 u64                      id;
3050                 u64                      lost;
3051         } lost_event;
3052
3053         rcu_read_lock();
3054         /*
3055          * For inherited events we send all the output towards the parent.
3056          */
3057         if (event->parent)
3058                 event = event->parent;
3059
3060         data = rcu_dereference(event->data);
3061         if (!data)
3062                 goto out;
3063
3064         handle->data    = data;
3065         handle->event   = event;
3066         handle->nmi     = nmi;
3067         handle->sample  = sample;
3068
3069         if (!data->nr_pages)
3070                 goto out;
3071
3072         have_lost = local_read(&data->lost);
3073         if (have_lost)
3074                 size += sizeof(lost_event);
3075
3076         perf_output_get_handle(handle);
3077
3078         do {
3079                 /*
3080                  * Userspace could choose to issue a mb() before updating the
3081                  * tail pointer. So that all reads will be completed before the
3082                  * write is issued.
3083                  */
3084                 tail = ACCESS_ONCE(data->user_page->data_tail);
3085                 smp_rmb();
3086                 offset = head = local_read(&data->head);
3087                 head += size;
3088                 if (unlikely(!perf_output_space(data, tail, offset, head)))
3089                         goto fail;
3090         } while (local_cmpxchg(&data->head, offset, head) != offset);
3091
3092         if (head - local_read(&data->wakeup) > data->watermark)
3093                 local_add(data->watermark, &data->wakeup);
3094
3095         handle->page = offset >> (PAGE_SHIFT + page_order(data));
3096         handle->page &= data->nr_pages - 1;
3097         handle->size = offset & ((PAGE_SIZE << page_order(data)) - 1);
3098         handle->addr = data->data_pages[handle->page];
3099         handle->addr += handle->size;
3100         handle->size = (PAGE_SIZE << page_order(data)) - handle->size;
3101
3102         if (have_lost) {
3103                 lost_event.header.type = PERF_RECORD_LOST;
3104                 lost_event.header.misc = 0;
3105                 lost_event.header.size = sizeof(lost_event);
3106                 lost_event.id          = event->id;
3107                 lost_event.lost        = local_xchg(&data->lost, 0);
3108
3109                 perf_output_put(handle, lost_event);
3110         }
3111
3112         return 0;
3113
3114 fail:
3115         local_inc(&data->lost);
3116         perf_output_put_handle(handle);
3117 out:
3118         rcu_read_unlock();
3119
3120         return -ENOSPC;
3121 }
3122
3123 void perf_output_end(struct perf_output_handle *handle)
3124 {
3125         struct perf_event *event = handle->event;
3126         struct perf_mmap_data *data = handle->data;
3127
3128         int wakeup_events = event->attr.wakeup_events;
3129
3130         if (handle->sample && wakeup_events) {
3131                 int events = local_inc_return(&data->events);
3132                 if (events >= wakeup_events) {
3133                         local_sub(wakeup_events, &data->events);
3134                         local_inc(&data->wakeup);
3135                 }
3136         }
3137
3138         perf_output_put_handle(handle);
3139         rcu_read_unlock();
3140 }
3141
3142 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3143 {
3144         /*
3145          * only top level events have the pid namespace they were created in
3146          */
3147         if (event->parent)
3148                 event = event->parent;
3149
3150         return task_tgid_nr_ns(p, event->ns);
3151 }
3152
3153 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3154 {
3155         /*
3156          * only top level events have the pid namespace they were created in
3157          */
3158         if (event->parent)
3159                 event = event->parent;
3160
3161         return task_pid_nr_ns(p, event->ns);
3162 }
3163
3164 static void perf_output_read_one(struct perf_output_handle *handle,
3165                                  struct perf_event *event)
3166 {
3167         u64 read_format = event->attr.read_format;
3168         u64 values[4];
3169         int n = 0;
3170
3171         values[n++] = atomic64_read(&event->count);
3172         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3173                 values[n++] = event->total_time_enabled +
3174                         atomic64_read(&event->child_total_time_enabled);
3175         }
3176         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3177                 values[n++] = event->total_time_running +
3178                         atomic64_read(&event->child_total_time_running);
3179         }
3180         if (read_format & PERF_FORMAT_ID)
3181                 values[n++] = primary_event_id(event);
3182
3183         perf_output_copy(handle, values, n * sizeof(u64));
3184 }
3185
3186 /*
3187  * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3188  */
3189 static void perf_output_read_group(struct perf_output_handle *handle,
3190                             struct perf_event *event)
3191 {
3192         struct perf_event *leader = event->group_leader, *sub;
3193         u64 read_format = event->attr.read_format;
3194         u64 values[5];
3195         int n = 0;
3196
3197         values[n++] = 1 + leader->nr_siblings;
3198
3199         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3200                 values[n++] = leader->total_time_enabled;
3201
3202         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3203                 values[n++] = leader->total_time_running;
3204
3205         if (leader != event)
3206                 leader->pmu->read(leader);
3207
3208         values[n++] = atomic64_read(&leader->count);
3209         if (read_format & PERF_FORMAT_ID)
3210                 values[n++] = primary_event_id(leader);
3211
3212         perf_output_copy(handle, values, n * sizeof(u64));
3213
3214         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3215                 n = 0;
3216
3217                 if (sub != event)
3218                         sub->pmu->read(sub);
3219
3220                 values[n++] = atomic64_read(&sub->count);
3221                 if (read_format & PERF_FORMAT_ID)
3222                         values[n++] = primary_event_id(sub);
3223
3224                 perf_output_copy(handle, values, n * sizeof(u64));
3225         }
3226 }
3227
3228 static void perf_output_read(struct perf_output_handle *handle,
3229                              struct perf_event *event)
3230 {
3231         if (event->attr.read_format & PERF_FORMAT_GROUP)
3232                 perf_output_read_group(handle, event);
3233         else
3234                 perf_output_read_one(handle, event);
3235 }
3236
3237 void perf_output_sample(struct perf_output_handle *handle,
3238                         struct perf_event_header *header,
3239                         struct perf_sample_data *data,
3240                         struct perf_event *event)
3241 {
3242         u64 sample_type = data->type;
3243
3244         perf_output_put(handle, *header);
3245
3246         if (sample_type & PERF_SAMPLE_IP)
3247                 perf_output_put(handle, data->ip);
3248
3249         if (sample_type & PERF_SAMPLE_TID)
3250                 perf_output_put(handle, data->tid_entry);
3251
3252         if (sample_type & PERF_SAMPLE_TIME)
3253                 perf_output_put(handle, data->time);
3254
3255         if (sample_type & PERF_SAMPLE_ADDR)
3256                 perf_output_put(handle, data->addr);
3257
3258         if (sample_type & PERF_SAMPLE_ID)
3259                 perf_output_put(handle, data->id);
3260
3261         if (sample_type & PERF_SAMPLE_STREAM_ID)
3262                 perf_output_put(handle, data->stream_id);
3263
3264         if (sample_type & PERF_SAMPLE_CPU)
3265                 perf_output_put(handle, data->cpu_entry);
3266
3267         if (sample_type & PERF_SAMPLE_PERIOD)
3268                 perf_output_put(handle, data->period);
3269
3270         if (sample_type & PERF_SAMPLE_READ)
3271                 perf_output_read(handle, event);
3272
3273         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3274                 if (data->callchain) {
3275                         int size = 1;
3276
3277                         if (data->callchain)
3278                                 size += data->callchain->nr;
3279
3280                         size *= sizeof(u64);
3281
3282                         perf_output_copy(handle, data->callchain, size);
3283                 } else {
3284                         u64 nr = 0;
3285                         perf_output_put(handle, nr);
3286                 }
3287         }
3288
3289         if (sample_type & PERF_SAMPLE_RAW) {
3290                 if (data->raw) {
3291                         perf_output_put(handle, data->raw->size);
3292                         perf_output_copy(handle, data->raw->data,
3293                                          data->raw->size);
3294                 } else {
3295                         struct {
3296                                 u32     size;
3297                                 u32     data;
3298                         } raw = {
3299                                 .size = sizeof(u32),
3300                                 .data = 0,
3301                         };
3302                         perf_output_put(handle, raw);
3303                 }
3304         }
3305 }
3306
3307 void perf_prepare_sample(struct perf_event_header *header,
3308                          struct perf_sample_data *data,
3309                          struct perf_event *event,
3310                          struct pt_regs *regs)
3311 {
3312         u64 sample_type = event->attr.sample_type;
3313
3314         data->type = sample_type;
3315
3316         header->type = PERF_RECORD_SAMPLE;
3317         header->size = sizeof(*header);
3318
3319         header->misc = 0;
3320         header->misc |= perf_misc_flags(regs);
3321
3322         if (sample_type & PERF_SAMPLE_IP) {
3323                 data->ip = perf_instruction_pointer(regs);
3324
3325                 header->size += sizeof(data->ip);
3326         }
3327
3328         if (sample_type & PERF_SAMPLE_TID) {
3329                 /* namespace issues */
3330                 data->tid_entry.pid = perf_event_pid(event, current);
3331                 data->tid_entry.tid = perf_event_tid(event, current);
3332
3333                 header->size += sizeof(data->tid_entry);
3334         }
3335
3336         if (sample_type & PERF_SAMPLE_TIME) {
3337                 data->time = perf_clock();
3338
3339                 header->size += sizeof(data->time);
3340         }
3341
3342         if (sample_type & PERF_SAMPLE_ADDR)
3343                 header->size += sizeof(data->addr);
3344
3345         if (sample_type & PERF_SAMPLE_ID) {
3346                 data->id = primary_event_id(event);
3347
3348                 header->size += sizeof(data->id);
3349         }
3350
3351         if (sample_type & PERF_SAMPLE_STREAM_ID) {
3352                 data->stream_id = event->id;
3353
3354                 header->size += sizeof(data->stream_id);
3355         }
3356
3357         if (sample_type & PERF_SAMPLE_CPU) {
3358                 data->cpu_entry.cpu             = raw_smp_processor_id();
3359                 data->cpu_entry.reserved        = 0;
3360
3361                 header->size += sizeof(data->cpu_entry);
3362         }
3363
3364         if (sample_type & PERF_SAMPLE_PERIOD)
3365                 header->size += sizeof(data->period);
3366
3367         if (sample_type & PERF_SAMPLE_READ)
3368                 header->size += perf_event_read_size(event);
3369
3370         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3371                 int size = 1;
3372
3373                 data->callchain = perf_callchain(regs);
3374
3375                 if (data->callchain)
3376                         size += data->callchain->nr;
3377
3378                 header->size += size * sizeof(u64);
3379         }
3380
3381         if (sample_type & PERF_SAMPLE_RAW) {
3382                 int size = sizeof(u32);
3383
3384                 if (data->raw)
3385                         size += data->raw->size;
3386                 else
3387                         size += sizeof(u32);
3388
3389                 WARN_ON_ONCE(size & (sizeof(u64)-1));
3390                 header->size += size;
3391         }
3392 }
3393
3394 static void perf_event_output(struct perf_event *event, int nmi,
3395                                 struct perf_sample_data *data,
3396                                 struct pt_regs *regs)
3397 {
3398         struct perf_output_handle handle;
3399         struct perf_event_header header;
3400
3401         perf_prepare_sample(&header, data, event, regs);
3402
3403         if (perf_output_begin(&handle, event, header.size, nmi, 1))
3404                 return;
3405
3406         perf_output_sample(&handle, &header, data, event);
3407
3408         perf_output_end(&handle);
3409 }
3410
3411 /*
3412  * read event_id
3413  */
3414
3415 struct perf_read_event {
3416         struct perf_event_header        header;
3417
3418         u32                             pid;
3419         u32                             tid;
3420 };
3421
3422 static void
3423 perf_event_read_event(struct perf_event *event,
3424                         struct task_struct *task)
3425 {
3426         struct perf_output_handle handle;
3427         struct perf_read_event read_event = {
3428                 .header = {
3429                         .type = PERF_RECORD_READ,
3430                         .misc = 0,
3431                         .size = sizeof(read_event) + perf_event_read_size(event),
3432                 },
3433                 .pid = perf_event_pid(event, task),
3434                 .tid = perf_event_tid(event, task),
3435         };
3436         int ret;
3437
3438         ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3439         if (ret)
3440                 return;
3441
3442         perf_output_put(&handle, read_event);
3443         perf_output_read(&handle, event);
3444
3445         perf_output_end(&handle);
3446 }
3447
3448 /*
3449  * task tracking -- fork/exit
3450  *
3451  * enabled by: attr.comm | attr.mmap | attr.task
3452  */
3453
3454 struct perf_task_event {
3455         struct task_struct              *task;
3456         struct perf_event_context       *task_ctx;
3457
3458         struct {
3459                 struct perf_event_header        header;
3460
3461                 u32                             pid;
3462                 u32                             ppid;
3463                 u32                             tid;
3464                 u32                             ptid;
3465                 u64                             time;
3466         } event_id;
3467 };
3468
3469 static void perf_event_task_output(struct perf_event *event,
3470                                      struct perf_task_event *task_event)
3471 {
3472         struct perf_output_handle handle;
3473         struct task_struct *task = task_event->task;
3474         int size, ret;
3475
3476         size  = task_event->event_id.header.size;
3477         ret = perf_output_begin(&handle, event, size, 0, 0);
3478
3479         if (ret)
3480                 return;
3481
3482         task_event->event_id.pid = perf_event_pid(event, task);
3483         task_event->event_id.ppid = perf_event_pid(event, current);
3484
3485         task_event->event_id.tid = perf_event_tid(event, task);
3486         task_event->event_id.ptid = perf_event_tid(event, current);
3487
3488         perf_output_put(&handle, task_event->event_id);
3489
3490         perf_output_end(&handle);
3491 }
3492
3493 static int perf_event_task_match(struct perf_event *event)
3494 {
3495         if (event->state < PERF_EVENT_STATE_INACTIVE)
3496                 return 0;
3497
3498         if (event->cpu != -1 && event->cpu != smp_processor_id())
3499                 return 0;
3500
3501         if (event->attr.comm || event->attr.mmap || event->attr.task)
3502                 return 1;
3503
3504         return 0;
3505 }
3506
3507 static void perf_event_task_ctx(struct perf_event_context *ctx,
3508                                   struct perf_task_event *task_event)
3509 {
3510         struct perf_event *event;
3511
3512         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3513                 if (perf_event_task_match(event))
3514                         perf_event_task_output(event, task_event);
3515         }
3516 }
3517
3518 static void perf_event_task_event(struct perf_task_event *task_event)
3519 {
3520         struct perf_cpu_context *cpuctx;
3521         struct perf_event_context *ctx = task_event->task_ctx;
3522
3523         rcu_read_lock();
3524         cpuctx = &get_cpu_var(perf_cpu_context);
3525         perf_event_task_ctx(&cpuctx->ctx, task_event);
3526         if (!ctx)
3527                 ctx = rcu_dereference(current->perf_event_ctxp);
3528         if (ctx)
3529                 perf_event_task_ctx(ctx, task_event);
3530         put_cpu_var(perf_cpu_context);
3531         rcu_read_unlock();
3532 }
3533
3534 static void perf_event_task(struct task_struct *task,
3535                               struct perf_event_context *task_ctx,
3536                               int new)
3537 {
3538         struct perf_task_event task_event;
3539
3540         if (!atomic_read(&nr_comm_events) &&
3541             !atomic_read(&nr_mmap_events) &&
3542             !atomic_read(&nr_task_events))
3543                 return;
3544
3545         task_event = (struct perf_task_event){
3546                 .task     = task,
3547                 .task_ctx = task_ctx,
3548                 .event_id    = {
3549                         .header = {
3550                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3551                                 .misc = 0,
3552                                 .size = sizeof(task_event.event_id),
3553                         },
3554                         /* .pid  */
3555                         /* .ppid */
3556                         /* .tid  */
3557                         /* .ptid */
3558                         .time = perf_clock(),
3559                 },
3560         };
3561
3562         perf_event_task_event(&task_event);
3563 }
3564
3565 void perf_event_fork(struct task_struct *task)
3566 {
3567         perf_event_task(task, NULL, 1);
3568 }
3569
3570 /*
3571  * comm tracking
3572  */
3573
3574 struct perf_comm_event {
3575         struct task_struct      *task;
3576         char                    *comm;
3577         int                     comm_size;
3578
3579         struct {
3580                 struct perf_event_header        header;
3581
3582                 u32                             pid;
3583                 u32                             tid;
3584         } event_id;
3585 };
3586
3587 static void perf_event_comm_output(struct perf_event *event,
3588                                      struct perf_comm_event *comm_event)
3589 {
3590         struct perf_output_handle handle;
3591         int size = comm_event->event_id.header.size;
3592         int ret = perf_output_begin(&handle, event, size, 0, 0);
3593
3594         if (ret)
3595                 return;
3596
3597         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3598         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3599
3600         perf_output_put(&handle, comm_event->event_id);
3601         perf_output_copy(&handle, comm_event->comm,
3602                                    comm_event->comm_size);
3603         perf_output_end(&handle);
3604 }
3605
3606 static int perf_event_comm_match(struct perf_event *event)
3607 {
3608         if (event->state < PERF_EVENT_STATE_INACTIVE)
3609                 return 0;
3610
3611         if (event->cpu != -1 && event->cpu != smp_processor_id())
3612                 return 0;
3613
3614         if (event->attr.comm)
3615                 return 1;
3616
3617         return 0;
3618 }
3619
3620 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3621                                   struct perf_comm_event *comm_event)
3622 {
3623         struct perf_event *event;
3624
3625         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3626                 if (perf_event_comm_match(event))
3627                         perf_event_comm_output(event, comm_event);
3628         }
3629 }
3630
3631 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3632 {
3633         struct perf_cpu_context *cpuctx;
3634         struct perf_event_context *ctx;
3635         unsigned int size;
3636         char comm[TASK_COMM_LEN];
3637
3638         memset(comm, 0, sizeof(comm));
3639         strlcpy(comm, comm_event->task->comm, sizeof(comm));
3640         size = ALIGN(strlen(comm)+1, sizeof(u64));
3641
3642         comm_event->comm = comm;
3643         comm_event->comm_size = size;
3644
3645         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3646
3647         rcu_read_lock();
3648         cpuctx = &get_cpu_var(perf_cpu_context);
3649         perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3650         ctx = rcu_dereference(current->perf_event_ctxp);
3651         if (ctx)
3652                 perf_event_comm_ctx(ctx, comm_event);
3653         put_cpu_var(perf_cpu_context);
3654         rcu_read_unlock();
3655 }
3656
3657 void perf_event_comm(struct task_struct *task)
3658 {
3659         struct perf_comm_event comm_event;
3660
3661         if (task->perf_event_ctxp)
3662                 perf_event_enable_on_exec(task);
3663
3664         if (!atomic_read(&nr_comm_events))
3665                 return;
3666
3667         comm_event = (struct perf_comm_event){
3668                 .task   = task,
3669                 /* .comm      */
3670                 /* .comm_size */
3671                 .event_id  = {
3672                         .header = {
3673                                 .type = PERF_RECORD_COMM,
3674                                 .misc = 0,
3675                                 /* .size */
3676                         },
3677                         /* .pid */
3678                         /* .tid */
3679                 },
3680         };
3681
3682         perf_event_comm_event(&comm_event);
3683 }
3684
3685 /*
3686  * mmap tracking
3687  */
3688
3689 struct perf_mmap_event {
3690         struct vm_area_struct   *vma;
3691
3692         const char              *file_name;
3693         int                     file_size;
3694
3695         struct {
3696                 struct perf_event_header        header;
3697
3698                 u32                             pid;
3699                 u32                             tid;
3700                 u64                             start;
3701                 u64                             len;
3702                 u64                             pgoff;
3703         } event_id;
3704 };
3705
3706 static void perf_event_mmap_output(struct perf_event *event,
3707                                      struct perf_mmap_event *mmap_event)
3708 {
3709         struct perf_output_handle handle;
3710         int size = mmap_event->event_id.header.size;
3711         int ret = perf_output_begin(&handle, event, size, 0, 0);
3712
3713         if (ret)
3714                 return;
3715
3716         mmap_event->event_id.pid = perf_event_pid(event, current);
3717         mmap_event->event_id.tid = perf_event_tid(event, current);
3718
3719         perf_output_put(&handle, mmap_event->event_id);
3720         perf_output_copy(&handle, mmap_event->file_name,
3721                                    mmap_event->file_size);
3722         perf_output_end(&handle);
3723 }
3724
3725 static int perf_event_mmap_match(struct perf_event *event,
3726                                    struct perf_mmap_event *mmap_event)
3727 {
3728         if (event->state < PERF_EVENT_STATE_INACTIVE)
3729                 return 0;
3730
3731         if (event->cpu != -1 && event->cpu != smp_processor_id())
3732                 return 0;
3733
3734         if (event->attr.mmap)
3735                 return 1;
3736
3737         return 0;
3738 }
3739
3740 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3741                                   struct perf_mmap_event *mmap_event)
3742 {
3743         struct perf_event *event;
3744
3745         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3746                 if (perf_event_mmap_match(event, mmap_event))
3747                         perf_event_mmap_output(event, mmap_event);
3748         }
3749 }
3750
3751 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3752 {
3753         struct perf_cpu_context *cpuctx;
3754         struct perf_event_context *ctx;
3755         struct vm_area_struct *vma = mmap_event->vma;
3756         struct file *file = vma->vm_file;
3757         unsigned int size;
3758         char tmp[16];
3759         char *buf = NULL;
3760         const char *name;
3761
3762         memset(tmp, 0, sizeof(tmp));
3763
3764         if (file) {
3765                 /*
3766                  * d_path works from the end of the buffer backwards, so we
3767                  * need to add enough zero bytes after the string to handle
3768                  * the 64bit alignment we do later.
3769                  */
3770                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3771                 if (!buf) {
3772                         name = strncpy(tmp, "//enomem", sizeof(tmp));
3773                         goto got_name;
3774                 }
3775                 name = d_path(&file->f_path, buf, PATH_MAX);
3776                 if (IS_ERR(name)) {
3777                         name = strncpy(tmp, "//toolong", sizeof(tmp));
3778                         goto got_name;
3779                 }
3780         } else {
3781                 if (arch_vma_name(mmap_event->vma)) {
3782                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3783                                        sizeof(tmp));
3784                         goto got_name;
3785                 }
3786
3787                 if (!vma->vm_mm) {
3788                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
3789                         goto got_name;
3790                 }
3791
3792                 name = strncpy(tmp, "//anon", sizeof(tmp));
3793                 goto got_name;
3794         }
3795
3796 got_name:
3797         size = ALIGN(strlen(name)+1, sizeof(u64));
3798
3799         mmap_event->file_name = name;
3800         mmap_event->file_size = size;
3801
3802         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3803
3804         rcu_read_lock();
3805         cpuctx = &get_cpu_var(perf_cpu_context);
3806         perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3807         ctx = rcu_dereference(current->perf_event_ctxp);
3808         if (ctx)
3809                 perf_event_mmap_ctx(ctx, mmap_event);
3810         put_cpu_var(perf_cpu_context);
3811         rcu_read_unlock();
3812
3813         kfree(buf);
3814 }
3815
3816 void __perf_event_mmap(struct vm_area_struct *vma)
3817 {
3818         struct perf_mmap_event mmap_event;
3819
3820         if (!atomic_read(&nr_mmap_events))
3821                 return;
3822
3823         mmap_event = (struct perf_mmap_event){
3824                 .vma    = vma,
3825                 /* .file_name */
3826                 /* .file_size */
3827                 .event_id  = {
3828                         .header = {
3829                                 .type = PERF_RECORD_MMAP,
3830                                 .misc = PERF_RECORD_MISC_USER,
3831                                 /* .size */
3832                         },
3833                         /* .pid */
3834                         /* .tid */
3835                         .start  = vma->vm_start,
3836                         .len    = vma->vm_end - vma->vm_start,
3837                         .pgoff  = (u64)vma->vm_pgoff << PAGE_SHIFT,
3838                 },
3839         };
3840
3841         perf_event_mmap_event(&mmap_event);
3842 }
3843
3844 /*
3845  * IRQ throttle logging
3846  */
3847
3848 static void perf_log_throttle(struct perf_event *event, int enable)
3849 {
3850         struct perf_output_handle handle;
3851         int ret;
3852
3853         struct {
3854                 struct perf_event_header        header;
3855                 u64                             time;
3856                 u64                             id;
3857                 u64                             stream_id;
3858         } throttle_event = {
3859                 .header = {
3860                         .type = PERF_RECORD_THROTTLE,
3861                         .misc = 0,
3862                         .size = sizeof(throttle_event),
3863                 },
3864                 .time           = perf_clock(),
3865                 .id             = primary_event_id(event),
3866                 .stream_id      = event->id,
3867         };
3868
3869         if (enable)
3870                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3871
3872         ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3873         if (ret)
3874                 return;
3875
3876         perf_output_put(&handle, throttle_event);
3877         perf_output_end(&handle);
3878 }
3879
3880 /*
3881  * Generic event overflow handling, sampling.
3882  */
3883
3884 static int __perf_event_overflow(struct perf_event *event, int nmi,
3885                                    int throttle, struct perf_sample_data *data,
3886                                    struct pt_regs *regs)
3887 {
3888         int events = atomic_read(&event->event_limit);
3889         struct hw_perf_event *hwc = &event->hw;
3890         int ret = 0;
3891
3892         throttle = (throttle && event->pmu->unthrottle != NULL);
3893
3894         if (!throttle) {
3895                 hwc->interrupts++;
3896         } else {
3897                 if (hwc->interrupts != MAX_INTERRUPTS) {
3898                         hwc->interrupts++;
3899                         if (HZ * hwc->interrupts >
3900                                         (u64)sysctl_perf_event_sample_rate) {
3901                                 hwc->interrupts = MAX_INTERRUPTS;
3902                                 perf_log_throttle(event, 0);
3903                                 ret = 1;
3904                         }
3905                 } else {
3906                         /*
3907                          * Keep re-disabling events even though on the previous
3908                          * pass we disabled it - just in case we raced with a
3909                          * sched-in and the event got enabled again:
3910                          */
3911                         ret = 1;
3912                 }
3913         }
3914
3915         if (event->attr.freq) {
3916                 u64 now = perf_clock();
3917                 s64 delta = now - hwc->freq_time_stamp;
3918
3919                 hwc->freq_time_stamp = now;
3920
3921                 if (delta > 0 && delta < 2*TICK_NSEC)
3922                         perf_adjust_period(event, delta, hwc->last_period);
3923         }
3924
3925         /*
3926          * XXX event_limit might not quite work as expected on inherited
3927          * events
3928          */
3929
3930         event->pending_kill = POLL_IN;
3931         if (events && atomic_dec_and_test(&event->event_limit)) {
3932                 ret = 1;
3933                 event->pending_kill = POLL_HUP;
3934                 if (nmi) {
3935                         event->pending_disable = 1;
3936                         perf_pending_queue(&event->pending,
3937                                            perf_pending_event);
3938                 } else
3939                         perf_event_disable(event);
3940         }
3941
3942         if (event->overflow_handler)
3943                 event->overflow_handler(event, nmi, data, regs);
3944         else
3945                 perf_event_output(event, nmi, data, regs);
3946
3947         return ret;
3948 }
3949
3950 int perf_event_overflow(struct perf_event *event, int nmi,
3951                           struct perf_sample_data *data,
3952                           struct pt_regs *regs)
3953 {
3954         return __perf_event_overflow(event, nmi, 1, data, regs);
3955 }
3956
3957 /*
3958  * Generic software event infrastructure
3959  */
3960
3961 /*
3962  * We directly increment event->count and keep a second value in
3963  * event->hw.period_left to count intervals. This period event
3964  * is kept in the range [-sample_period, 0] so that we can use the
3965  * sign as trigger.
3966  */
3967
3968 static u64 perf_swevent_set_period(struct perf_event *event)
3969 {
3970         struct hw_perf_event *hwc = &event->hw;
3971         u64 period = hwc->last_period;
3972         u64 nr, offset;
3973         s64 old, val;
3974
3975         hwc->last_period = hwc->sample_period;
3976
3977 again:
3978         old = val = atomic64_read(&hwc->period_left);
3979         if (val < 0)
3980                 return 0;
3981
3982         nr = div64_u64(period + val, period);
3983         offset = nr * period;
3984         val -= offset;
3985         if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3986                 goto again;
3987
3988         return nr;
3989 }
3990
3991 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3992                                     int nmi, struct perf_sample_data *data,
3993                                     struct pt_regs *regs)
3994 {
3995         struct hw_perf_event *hwc = &event->hw;
3996         int throttle = 0;
3997
3998         data->period = event->hw.last_period;
3999         if (!overflow)
4000                 overflow = perf_swevent_set_period(event);
4001
4002         if (hwc->interrupts == MAX_INTERRUPTS)
4003                 return;
4004
4005         for (; overflow; overflow--) {
4006                 if (__perf_event_overflow(event, nmi, throttle,
4007                                             data, regs)) {
4008                         /*
4009                          * We inhibit the overflow from happening when
4010                          * hwc->interrupts == MAX_INTERRUPTS.
4011                          */
4012                         break;
4013                 }
4014                 throttle = 1;
4015         }
4016 }
4017
4018 static void perf_swevent_unthrottle(struct perf_event *event)
4019 {
4020         /*
4021          * Nothing to do, we already reset hwc->interrupts.
4022          */
4023 }
4024
4025 static void perf_swevent_add(struct perf_event *event, u64 nr,
4026                                int nmi, struct perf_sample_data *data,
4027                                struct pt_regs *regs)
4028 {
4029         struct hw_perf_event *hwc = &event->hw;
4030
4031         atomic64_add(nr, &event->count);
4032
4033         if (!regs)
4034                 return;
4035
4036         if (!hwc->sample_period)
4037                 return;
4038
4039         if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4040                 return perf_swevent_overflow(event, 1, nmi, data, regs);
4041
4042         if (atomic64_add_negative(nr, &hwc->period_left))
4043                 return;
4044
4045         perf_swevent_overflow(event, 0, nmi, data, regs);
4046 }
4047
4048 static int perf_exclude_event(struct perf_event *event,
4049                               struct pt_regs *regs)
4050 {
4051         if (regs) {
4052                 if (event->attr.exclude_user && user_mode(regs))
4053                         return 1;
4054
4055                 if (event->attr.exclude_kernel && !user_mode(regs))
4056                         return 1;
4057         }
4058
4059         return 0;
4060 }
4061
4062 static int perf_swevent_match(struct perf_event *event,
4063                                 enum perf_type_id type,
4064                                 u32 event_id,
4065                                 struct perf_sample_data *data,
4066                                 struct pt_regs *regs)
4067 {
4068         if (event->attr.type != type)
4069                 return 0;
4070
4071         if (event->attr.config != event_id)
4072                 return 0;
4073
4074         if (perf_exclude_event(event, regs))
4075                 return 0;
4076
4077         return 1;
4078 }
4079
4080 static inline u64 swevent_hash(u64 type, u32 event_id)
4081 {
4082         u64 val = event_id | (type << 32);
4083
4084         return hash_64(val, SWEVENT_HLIST_BITS);
4085 }
4086
4087 static inline struct hlist_head *
4088 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4089 {
4090         u64 hash = swevent_hash(type, event_id);
4091
4092         return &hlist->heads[hash];
4093 }
4094
4095 /* For the read side: events when they trigger */
4096 static inline struct hlist_head *
4097 find_swevent_head_rcu(struct perf_cpu_context *ctx, u64 type, u32 event_id)
4098 {
4099         struct swevent_hlist *hlist;
4100
4101         hlist = rcu_dereference(ctx->swevent_hlist);
4102         if (!hlist)
4103                 return NULL;
4104
4105         return __find_swevent_head(hlist, type, event_id);
4106 }
4107
4108 /* For the event head insertion and removal in the hlist */
4109 static inline struct hlist_head *
4110 find_swevent_head(struct perf_cpu_context *ctx, struct perf_event *event)
4111 {
4112         struct swevent_hlist *hlist;
4113         u32 event_id = event->attr.config;
4114         u64 type = event->attr.type;
4115
4116         /*
4117          * Event scheduling is always serialized against hlist allocation
4118          * and release. Which makes the protected version suitable here.
4119          * The context lock guarantees that.
4120          */
4121         hlist = rcu_dereference_protected(ctx->swevent_hlist,
4122                                           lockdep_is_held(&event->ctx->lock));
4123         if (!hlist)
4124                 return NULL;
4125
4126         return __find_swevent_head(hlist, type, event_id);
4127 }
4128
4129 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4130                                     u64 nr, int nmi,
4131                                     struct perf_sample_data *data,
4132                                     struct pt_regs *regs)
4133 {
4134         struct perf_cpu_context *cpuctx;
4135         struct perf_event *event;
4136         struct hlist_node *node;
4137         struct hlist_head *head;
4138
4139         cpuctx = &__get_cpu_var(perf_cpu_context);
4140
4141         rcu_read_lock();
4142
4143         head = find_swevent_head_rcu(cpuctx, type, event_id);
4144
4145         if (!head)
4146                 goto end;
4147
4148         hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4149                 if (perf_swevent_match(event, type, event_id, data, regs))
4150                         perf_swevent_add(event, nr, nmi, data, regs);
4151         }
4152 end:
4153         rcu_read_unlock();
4154 }
4155
4156 int perf_swevent_get_recursion_context(void)
4157 {
4158         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4159         int rctx;
4160
4161         if (in_nmi())
4162                 rctx = 3;
4163         else if (in_irq())
4164                 rctx = 2;
4165         else if (in_softirq())
4166                 rctx = 1;
4167         else
4168                 rctx = 0;
4169
4170         if (cpuctx->recursion[rctx])
4171                 return -1;
4172
4173         cpuctx->recursion[rctx]++;
4174         barrier();
4175
4176         return rctx;
4177 }
4178 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4179
4180 void perf_swevent_put_recursion_context(int rctx)
4181 {
4182         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4183         barrier();
4184         cpuctx->recursion[rctx]--;
4185 }
4186 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
4187
4188
4189 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4190                             struct pt_regs *regs, u64 addr)
4191 {
4192         struct perf_sample_data data;
4193         int rctx;
4194
4195         preempt_disable_notrace();
4196         rctx = perf_swevent_get_recursion_context();
4197         if (rctx < 0)
4198                 return;
4199
4200         perf_sample_data_init(&data, addr);
4201
4202         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4203
4204         perf_swevent_put_recursion_context(rctx);
4205         preempt_enable_notrace();
4206 }
4207
4208 static void perf_swevent_read(struct perf_event *event)
4209 {
4210 }
4211
4212 static int perf_swevent_enable(struct perf_event *event)
4213 {
4214         struct hw_perf_event *hwc = &event->hw;
4215         struct perf_cpu_context *cpuctx;
4216         struct hlist_head *head;
4217
4218         cpuctx = &__get_cpu_var(perf_cpu_context);
4219
4220         if (hwc->sample_period) {
4221                 hwc->last_period = hwc->sample_period;
4222                 perf_swevent_set_period(event);
4223         }
4224
4225         head = find_swevent_head(cpuctx, event);
4226         if (WARN_ON_ONCE(!head))
4227                 return -EINVAL;
4228
4229         hlist_add_head_rcu(&event->hlist_entry, head);
4230
4231         return 0;
4232 }
4233
4234 static void perf_swevent_disable(struct perf_event *event)
4235 {
4236         hlist_del_rcu(&event->hlist_entry);
4237 }
4238
4239 static const struct pmu perf_ops_generic = {
4240         .enable         = perf_swevent_enable,
4241         .disable        = perf_swevent_disable,
4242         .read           = perf_swevent_read,
4243         .unthrottle     = perf_swevent_unthrottle,
4244 };
4245
4246 /*
4247  * hrtimer based swevent callback
4248  */
4249
4250 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4251 {
4252         enum hrtimer_restart ret = HRTIMER_RESTART;
4253         struct perf_sample_data data;
4254         struct pt_regs *regs;
4255         struct perf_event *event;
4256         u64 period;
4257
4258         event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4259         event->pmu->read(event);
4260
4261         perf_sample_data_init(&data, 0);
4262         data.period = event->hw.last_period;
4263         regs = get_irq_regs();
4264
4265         if (regs && !perf_exclude_event(event, regs)) {
4266                 if (!(event->attr.exclude_idle && current->pid == 0))
4267                         if (perf_event_overflow(event, 0, &data, regs))
4268                                 ret = HRTIMER_NORESTART;
4269         }
4270
4271         period = max_t(u64, 10000, event->hw.sample_period);
4272         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4273
4274         return ret;
4275 }
4276
4277 static void perf_swevent_start_hrtimer(struct perf_event *event)
4278 {
4279         struct hw_perf_event *hwc = &event->hw;
4280
4281         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4282         hwc->hrtimer.function = perf_swevent_hrtimer;
4283         if (hwc->sample_period) {
4284                 u64 period;
4285
4286                 if (hwc->remaining) {
4287                         if (hwc->remaining < 0)
4288                                 period = 10000;
4289                         else
4290                                 period = hwc->remaining;
4291                         hwc->remaining = 0;
4292                 } else {
4293                         period = max_t(u64, 10000, hwc->sample_period);
4294                 }
4295                 __hrtimer_start_range_ns(&hwc->hrtimer,
4296                                 ns_to_ktime(period), 0,
4297                                 HRTIMER_MODE_REL, 0);
4298         }
4299 }
4300
4301 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4302 {
4303         struct hw_perf_event *hwc = &event->hw;
4304
4305         if (hwc->sample_period) {
4306                 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4307                 hwc->remaining = ktime_to_ns(remaining);
4308
4309                 hrtimer_cancel(&hwc->hrtimer);
4310         }
4311 }
4312
4313 /*
4314  * Software event: cpu wall time clock
4315  */
4316
4317 static void cpu_clock_perf_event_update(struct perf_event *event)
4318 {
4319         int cpu = raw_smp_processor_id();
4320         s64 prev;
4321         u64 now;
4322
4323         now = cpu_clock(cpu);
4324         prev = atomic64_xchg(&event->hw.prev_count, now);
4325         atomic64_add(now - prev, &event->count);
4326 }
4327
4328 static int cpu_clock_perf_event_enable(struct perf_event *event)
4329 {
4330         struct hw_perf_event *hwc = &event->hw;
4331         int cpu = raw_smp_processor_id();
4332
4333         atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4334         perf_swevent_start_hrtimer(event);
4335
4336         return 0;
4337 }
4338
4339 static void cpu_clock_perf_event_disable(struct perf_event *event)
4340 {
4341         perf_swevent_cancel_hrtimer(event);
4342         cpu_clock_perf_event_update(event);
4343 }
4344
4345 static void cpu_clock_perf_event_read(struct perf_event *event)
4346 {
4347         cpu_clock_perf_event_update(event);
4348 }
4349
4350 static const struct pmu perf_ops_cpu_clock = {
4351         .enable         = cpu_clock_perf_event_enable,
4352         .disable        = cpu_clock_perf_event_disable,
4353         .read           = cpu_clock_perf_event_read,
4354 };
4355
4356 /*
4357  * Software event: task time clock
4358  */
4359
4360 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4361 {
4362         u64 prev;
4363         s64 delta;
4364
4365         prev = atomic64_xchg(&event->hw.prev_count, now);
4366         delta = now - prev;
4367         atomic64_add(delta, &event->count);
4368 }
4369
4370 static int task_clock_perf_event_enable(struct perf_event *event)
4371 {
4372         struct hw_perf_event *hwc = &event->hw;
4373         u64 now;
4374
4375         now = event->ctx->time;
4376
4377         atomic64_set(&hwc->prev_count, now);
4378
4379         perf_swevent_start_hrtimer(event);
4380
4381         return 0;
4382 }
4383
4384 static void task_clock_perf_event_disable(struct perf_event *event)
4385 {
4386         perf_swevent_cancel_hrtimer(event);
4387         task_clock_perf_event_update(event, event->ctx->time);
4388
4389 }
4390
4391 static void task_clock_perf_event_read(struct perf_event *event)
4392 {
4393         u64 time;
4394
4395         if (!in_nmi()) {
4396                 update_context_time(event->ctx);
4397                 time = event->ctx->time;
4398         } else {
4399                 u64 now = perf_clock();
4400                 u64 delta = now - event->ctx->timestamp;
4401                 time = event->ctx->time + delta;
4402         }
4403
4404         task_clock_perf_event_update(event, time);
4405 }
4406
4407 static const struct pmu perf_ops_task_clock = {
4408         .enable         = task_clock_perf_event_enable,
4409         .disable        = task_clock_perf_event_disable,
4410         .read           = task_clock_perf_event_read,
4411 };
4412
4413 /* Deref the hlist from the update side */
4414 static inline struct swevent_hlist *
4415 swevent_hlist_deref(struct perf_cpu_context *cpuctx)
4416 {
4417         return rcu_dereference_protected(cpuctx->swevent_hlist,
4418                                          lockdep_is_held(&cpuctx->hlist_mutex));
4419 }
4420
4421 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4422 {
4423         struct swevent_hlist *hlist;
4424
4425         hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4426         kfree(hlist);
4427 }
4428
4429 static void swevent_hlist_release(struct perf_cpu_context *cpuctx)
4430 {
4431         struct swevent_hlist *hlist = swevent_hlist_deref(cpuctx);
4432
4433         if (!hlist)
4434                 return;
4435
4436         rcu_assign_pointer(cpuctx->swevent_hlist, NULL);
4437         call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4438 }
4439
4440 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4441 {
4442         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4443
4444         mutex_lock(&cpuctx->hlist_mutex);
4445
4446         if (!--cpuctx->hlist_refcount)
4447                 swevent_hlist_release(cpuctx);
4448
4449         mutex_unlock(&cpuctx->hlist_mutex);
4450 }
4451
4452 static void swevent_hlist_put(struct perf_event *event)
4453 {
4454         int cpu;
4455
4456         if (event->cpu != -1) {
4457                 swevent_hlist_put_cpu(event, event->cpu);
4458                 return;
4459         }
4460
4461         for_each_possible_cpu(cpu)
4462                 swevent_hlist_put_cpu(event, cpu);
4463 }
4464
4465 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4466 {
4467         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4468         int err = 0;
4469
4470         mutex_lock(&cpuctx->hlist_mutex);
4471
4472         if (!swevent_hlist_deref(cpuctx) && cpu_online(cpu)) {
4473                 struct swevent_hlist *hlist;
4474
4475                 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4476                 if (!hlist) {
4477                         err = -ENOMEM;
4478                         goto exit;
4479                 }
4480                 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
4481         }
4482         cpuctx->hlist_refcount++;
4483  exit:
4484         mutex_unlock(&cpuctx->hlist_mutex);
4485
4486         return err;
4487 }
4488
4489 static int swevent_hlist_get(struct perf_event *event)
4490 {
4491         int err;
4492         int cpu, failed_cpu;
4493
4494         if (event->cpu != -1)
4495                 return swevent_hlist_get_cpu(event, event->cpu);
4496
4497         get_online_cpus();
4498         for_each_possible_cpu(cpu) {
4499                 err = swevent_hlist_get_cpu(event, cpu);
4500                 if (err) {
4501                         failed_cpu = cpu;
4502                         goto fail;
4503                 }
4504         }
4505         put_online_cpus();
4506
4507         return 0;
4508  fail:
4509         for_each_possible_cpu(cpu) {
4510                 if (cpu == failed_cpu)
4511                         break;
4512                 swevent_hlist_put_cpu(event, cpu);
4513         }
4514
4515         put_online_cpus();
4516         return err;
4517 }
4518
4519 #ifdef CONFIG_EVENT_TRACING
4520
4521 static const struct pmu perf_ops_tracepoint = {
4522         .enable         = perf_trace_enable,
4523         .disable        = perf_trace_disable,
4524         .read           = perf_swevent_read,
4525         .unthrottle     = perf_swevent_unthrottle,
4526 };
4527
4528 static int perf_tp_filter_match(struct perf_event *event,
4529                                 struct perf_sample_data *data)
4530 {
4531         void *record = data->raw->data;
4532
4533         if (likely(!event->filter) || filter_match_preds(event->filter, record))
4534                 return 1;
4535         return 0;
4536 }
4537
4538 static int perf_tp_event_match(struct perf_event *event,
4539                                 struct perf_sample_data *data,
4540                                 struct pt_regs *regs)
4541 {
4542         /*
4543          * All tracepoints are from kernel-space.
4544          */
4545         if (event->attr.exclude_kernel)
4546                 return 0;
4547
4548         if (!perf_tp_filter_match(event, data))
4549                 return 0;
4550
4551         return 1;
4552 }
4553
4554 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
4555                    struct pt_regs *regs, struct hlist_head *head)
4556 {
4557         struct perf_sample_data data;
4558         struct perf_event *event;
4559         struct hlist_node *node;
4560
4561         struct perf_raw_record raw = {
4562                 .size = entry_size,
4563                 .data = record,
4564         };
4565
4566         perf_sample_data_init(&data, addr);
4567         data.raw = &raw;
4568
4569         rcu_read_lock();
4570         hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4571                 if (perf_tp_event_match(event, &data, regs))
4572                         perf_swevent_add(event, count, 1, &data, regs);
4573         }
4574         rcu_read_unlock();
4575 }
4576 EXPORT_SYMBOL_GPL(perf_tp_event);
4577
4578 static void tp_perf_event_destroy(struct perf_event *event)
4579 {
4580         perf_trace_destroy(event);
4581 }
4582
4583 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4584 {
4585         int err;
4586
4587         /*
4588          * Raw tracepoint data is a severe data leak, only allow root to
4589          * have these.
4590          */
4591         if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4592                         perf_paranoid_tracepoint_raw() &&
4593                         !capable(CAP_SYS_ADMIN))
4594                 return ERR_PTR(-EPERM);
4595
4596         err = perf_trace_init(event);
4597         if (err)
4598                 return NULL;
4599
4600         event->destroy = tp_perf_event_destroy;
4601
4602         return &perf_ops_tracepoint;
4603 }
4604
4605 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4606 {
4607         char *filter_str;
4608         int ret;
4609
4610         if (event->attr.type != PERF_TYPE_TRACEPOINT)
4611                 return -EINVAL;
4612
4613         filter_str = strndup_user(arg, PAGE_SIZE);
4614         if (IS_ERR(filter_str))
4615                 return PTR_ERR(filter_str);
4616
4617         ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4618
4619         kfree(filter_str);
4620         return ret;
4621 }
4622
4623 static void perf_event_free_filter(struct perf_event *event)
4624 {
4625         ftrace_profile_free_filter(event);
4626 }
4627
4628 #else
4629
4630 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4631 {
4632         return NULL;
4633 }
4634
4635 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4636 {
4637         return -ENOENT;
4638 }
4639
4640 static void perf_event_free_filter(struct perf_event *event)
4641 {
4642 }
4643
4644 #endif /* CONFIG_EVENT_TRACING */
4645
4646 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4647 static void bp_perf_event_destroy(struct perf_event *event)
4648 {
4649         release_bp_slot(event);
4650 }
4651
4652 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4653 {
4654         int err;
4655
4656         err = register_perf_hw_breakpoint(bp);
4657         if (err)
4658                 return ERR_PTR(err);
4659
4660         bp->destroy = bp_perf_event_destroy;
4661
4662         return &perf_ops_bp;
4663 }
4664
4665 void perf_bp_event(struct perf_event *bp, void *data)
4666 {
4667         struct perf_sample_data sample;
4668         struct pt_regs *regs = data;
4669
4670         perf_sample_data_init(&sample, bp->attr.bp_addr);
4671
4672         if (!perf_exclude_event(bp, regs))
4673                 perf_swevent_add(bp, 1, 1, &sample, regs);
4674 }
4675 #else
4676 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4677 {
4678         return NULL;
4679 }
4680
4681 void perf_bp_event(struct perf_event *bp, void *regs)
4682 {
4683 }
4684 #endif
4685
4686 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4687
4688 static void sw_perf_event_destroy(struct perf_event *event)
4689 {
4690         u64 event_id = event->attr.config;
4691
4692         WARN_ON(event->parent);
4693
4694         atomic_dec(&perf_swevent_enabled[event_id]);
4695         swevent_hlist_put(event);
4696 }
4697
4698 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4699 {
4700         const struct pmu *pmu = NULL;
4701         u64 event_id = event->attr.config;
4702
4703         /*
4704          * Software events (currently) can't in general distinguish
4705          * between user, kernel and hypervisor events.
4706          * However, context switches and cpu migrations are considered
4707          * to be kernel events, and page faults are never hypervisor
4708          * events.
4709          */
4710         switch (event_id) {
4711         case PERF_COUNT_SW_CPU_CLOCK:
4712                 pmu = &perf_ops_cpu_clock;
4713
4714                 break;
4715         case PERF_COUNT_SW_TASK_CLOCK:
4716                 /*
4717                  * If the user instantiates this as a per-cpu event,
4718                  * use the cpu_clock event instead.
4719                  */
4720                 if (event->ctx->task)
4721                         pmu = &perf_ops_task_clock;
4722                 else
4723                         pmu = &perf_ops_cpu_clock;
4724
4725                 break;
4726         case PERF_COUNT_SW_PAGE_FAULTS:
4727         case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4728         case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4729         case PERF_COUNT_SW_CONTEXT_SWITCHES:
4730         case PERF_COUNT_SW_CPU_MIGRATIONS:
4731         case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4732         case PERF_COUNT_SW_EMULATION_FAULTS:
4733                 if (!event->parent) {
4734                         int err;
4735
4736                         err = swevent_hlist_get(event);
4737                         if (err)
4738                                 return ERR_PTR(err);
4739
4740                         atomic_inc(&perf_swevent_enabled[event_id]);
4741                         event->destroy = sw_perf_event_destroy;
4742                 }
4743                 pmu = &perf_ops_generic;
4744                 break;
4745         }
4746
4747         return pmu;
4748 }
4749
4750 /*
4751  * Allocate and initialize a event structure
4752  */
4753 static struct perf_event *
4754 perf_event_alloc(struct perf_event_attr *attr,
4755                    int cpu,
4756                    struct perf_event_context *ctx,
4757                    struct perf_event *group_leader,
4758                    struct perf_event *parent_event,
4759                    perf_overflow_handler_t overflow_handler,
4760                    gfp_t gfpflags)
4761 {
4762         const struct pmu *pmu;
4763         struct perf_event *event;
4764         struct hw_perf_event *hwc;
4765         long err;
4766
4767         event = kzalloc(sizeof(*event), gfpflags);
4768         if (!event)
4769                 return ERR_PTR(-ENOMEM);
4770
4771         /*
4772          * Single events are their own group leaders, with an
4773          * empty sibling list:
4774          */
4775         if (!group_leader)
4776                 group_leader = event;
4777
4778         mutex_init(&event->child_mutex);
4779         INIT_LIST_HEAD(&event->child_list);
4780
4781         INIT_LIST_HEAD(&event->group_entry);
4782         INIT_LIST_HEAD(&event->event_entry);
4783         INIT_LIST_HEAD(&event->sibling_list);
4784         init_waitqueue_head(&event->waitq);
4785
4786         mutex_init(&event->mmap_mutex);
4787
4788         event->cpu              = cpu;
4789         event->attr             = *attr;
4790         event->group_leader     = group_leader;
4791         event->pmu              = NULL;
4792         event->ctx              = ctx;
4793         event->oncpu            = -1;
4794
4795         event->parent           = parent_event;
4796
4797         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
4798         event->id               = atomic64_inc_return(&perf_event_id);
4799
4800         event->state            = PERF_EVENT_STATE_INACTIVE;
4801
4802         if (!overflow_handler && parent_event)
4803                 overflow_handler = parent_event->overflow_handler;
4804         
4805         event->overflow_handler = overflow_handler;
4806
4807         if (attr->disabled)
4808                 event->state = PERF_EVENT_STATE_OFF;
4809
4810         pmu = NULL;
4811
4812         hwc = &event->hw;
4813         hwc->sample_period = attr->sample_period;
4814         if (attr->freq && attr->sample_freq)
4815                 hwc->sample_period = 1;
4816         hwc->last_period = hwc->sample_period;
4817
4818         atomic64_set(&hwc->period_left, hwc->sample_period);
4819
4820         /*
4821          * we currently do not support PERF_FORMAT_GROUP on inherited events
4822          */
4823         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4824                 goto done;
4825
4826         switch (attr->type) {
4827         case PERF_TYPE_RAW:
4828         case PERF_TYPE_HARDWARE:
4829         case PERF_TYPE_HW_CACHE:
4830                 pmu = hw_perf_event_init(event);
4831                 break;
4832
4833         case PERF_TYPE_SOFTWARE:
4834                 pmu = sw_perf_event_init(event);
4835                 break;
4836
4837         case PERF_TYPE_TRACEPOINT:
4838                 pmu = tp_perf_event_init(event);
4839                 break;
4840
4841         case PERF_TYPE_BREAKPOINT:
4842                 pmu = bp_perf_event_init(event);
4843                 break;
4844
4845
4846         default:
4847                 break;
4848         }
4849 done:
4850         err = 0;
4851         if (!pmu)
4852                 err = -EINVAL;
4853         else if (IS_ERR(pmu))
4854                 err = PTR_ERR(pmu);
4855
4856         if (err) {
4857                 if (event->ns)
4858                         put_pid_ns(event->ns);
4859                 kfree(event);
4860                 return ERR_PTR(err);
4861         }
4862
4863         event->pmu = pmu;
4864
4865         if (!event->parent) {
4866                 atomic_inc(&nr_events);
4867                 if (event->attr.mmap)
4868                         atomic_inc(&nr_mmap_events);
4869                 if (event->attr.comm)
4870                         atomic_inc(&nr_comm_events);
4871                 if (event->attr.task)
4872                         atomic_inc(&nr_task_events);
4873         }
4874
4875         return event;
4876 }
4877
4878 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4879                           struct perf_event_attr *attr)
4880 {
4881         u32 size;
4882         int ret;
4883
4884         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4885                 return -EFAULT;
4886
4887         /*
4888          * zero the full structure, so that a short copy will be nice.
4889          */
4890         memset(attr, 0, sizeof(*attr));
4891
4892         ret = get_user(size, &uattr->size);
4893         if (ret)
4894                 return ret;
4895
4896         if (size > PAGE_SIZE)   /* silly large */
4897                 goto err_size;
4898
4899         if (!size)              /* abi compat */
4900                 size = PERF_ATTR_SIZE_VER0;
4901
4902         if (size < PERF_ATTR_SIZE_VER0)
4903                 goto err_size;
4904
4905         /*
4906          * If we're handed a bigger struct than we know of,
4907          * ensure all the unknown bits are 0 - i.e. new
4908          * user-space does not rely on any kernel feature
4909          * extensions we dont know about yet.
4910          */
4911         if (size > sizeof(*attr)) {
4912                 unsigned char __user *addr;
4913                 unsigned char __user *end;
4914                 unsigned char val;
4915
4916                 addr = (void __user *)uattr + sizeof(*attr);
4917                 end  = (void __user *)uattr + size;
4918
4919                 for (; addr < end; addr++) {
4920                         ret = get_user(val, addr);
4921                         if (ret)
4922                                 return ret;
4923                         if (val)
4924                                 goto err_size;
4925                 }
4926                 size = sizeof(*attr);
4927         }
4928
4929         ret = copy_from_user(attr, uattr, size);
4930         if (ret)
4931                 return -EFAULT;
4932
4933         /*
4934          * If the type exists, the corresponding creation will verify
4935          * the attr->config.
4936          */
4937         if (attr->type >= PERF_TYPE_MAX)
4938                 return -EINVAL;
4939
4940         if (attr->__reserved_1)
4941                 return -EINVAL;
4942
4943         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4944                 return -EINVAL;
4945
4946         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4947                 return -EINVAL;
4948
4949 out:
4950         return ret;
4951
4952 err_size:
4953         put_user(sizeof(*attr), &uattr->size);
4954         ret = -E2BIG;
4955         goto out;
4956 }
4957
4958 static int
4959 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
4960 {
4961         struct perf_mmap_data *data = NULL, *old_data = NULL;
4962         int ret = -EINVAL;
4963
4964         if (!output_event)
4965                 goto set;
4966
4967         /* don't allow circular references */
4968         if (event == output_event)
4969                 goto out;
4970
4971         /*
4972          * Don't allow cross-cpu buffers
4973          */
4974         if (output_event->cpu != event->cpu)
4975                 goto out;
4976
4977         /*
4978          * If its not a per-cpu buffer, it must be the same task.
4979          */
4980         if (output_event->cpu == -1 && output_event->ctx != event->ctx)
4981                 goto out;
4982
4983 set:
4984         mutex_lock(&event->mmap_mutex);
4985         /* Can't redirect output if we've got an active mmap() */
4986         if (atomic_read(&event->mmap_count))
4987                 goto unlock;
4988
4989         if (output_event) {
4990                 /* get the buffer we want to redirect to */
4991                 data = perf_mmap_data_get(output_event);
4992                 if (!data)
4993                         goto unlock;
4994         }
4995
4996         old_data = event->data;
4997         rcu_assign_pointer(event->data, data);
4998         ret = 0;
4999 unlock:
5000         mutex_unlock(&event->mmap_mutex);
5001
5002         if (old_data)
5003                 perf_mmap_data_put(old_data);
5004 out:
5005         return ret;
5006 }
5007
5008 /**
5009  * sys_perf_event_open - open a performance event, associate it to a task/cpu
5010  *
5011  * @attr_uptr:  event_id type attributes for monitoring/sampling
5012  * @pid:                target pid
5013  * @cpu:                target cpu
5014  * @group_fd:           group leader event fd
5015  */
5016 SYSCALL_DEFINE5(perf_event_open,
5017                 struct perf_event_attr __user *, attr_uptr,
5018                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
5019 {
5020         struct perf_event *event, *group_leader = NULL, *output_event = NULL;
5021         struct perf_event_attr attr;
5022         struct perf_event_context *ctx;
5023         struct file *event_file = NULL;
5024         struct file *group_file = NULL;
5025         int event_fd;
5026         int fput_needed = 0;
5027         int err;
5028
5029         /* for future expandability... */
5030         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
5031                 return -EINVAL;
5032
5033         err = perf_copy_attr(attr_uptr, &attr);
5034         if (err)
5035                 return err;
5036
5037         if (!attr.exclude_kernel) {
5038                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
5039                         return -EACCES;
5040         }
5041
5042         if (attr.freq) {
5043                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
5044                         return -EINVAL;
5045         }
5046
5047         event_fd = get_unused_fd_flags(O_RDWR);
5048         if (event_fd < 0)
5049                 return event_fd;
5050
5051         /*
5052          * Get the target context (task or percpu):
5053          */
5054         ctx = find_get_context(pid, cpu);
5055         if (IS_ERR(ctx)) {
5056                 err = PTR_ERR(ctx);
5057                 goto err_fd;
5058         }
5059
5060         if (group_fd != -1) {
5061                 group_leader = perf_fget_light(group_fd, &fput_needed);
5062                 if (IS_ERR(group_leader)) {
5063                         err = PTR_ERR(group_leader);
5064                         goto err_put_context;
5065                 }
5066                 group_file = group_leader->filp;
5067                 if (flags & PERF_FLAG_FD_OUTPUT)
5068                         output_event = group_leader;
5069                 if (flags & PERF_FLAG_FD_NO_GROUP)
5070                         group_leader = NULL;
5071         }
5072
5073         /*
5074          * Look up the group leader (we will attach this event to it):
5075          */
5076         if (group_leader) {
5077                 err = -EINVAL;
5078
5079                 /*
5080                  * Do not allow a recursive hierarchy (this new sibling
5081                  * becoming part of another group-sibling):
5082                  */
5083                 if (group_leader->group_leader != group_leader)
5084                         goto err_put_context;
5085                 /*
5086                  * Do not allow to attach to a group in a different
5087                  * task or CPU context:
5088                  */
5089                 if (group_leader->ctx != ctx)
5090                         goto err_put_context;
5091                 /*
5092                  * Only a group leader can be exclusive or pinned
5093                  */
5094                 if (attr.exclusive || attr.pinned)
5095                         goto err_put_context;
5096         }
5097
5098         event = perf_event_alloc(&attr, cpu, ctx, group_leader,
5099                                      NULL, NULL, GFP_KERNEL);
5100         if (IS_ERR(event)) {
5101                 err = PTR_ERR(event);
5102                 goto err_put_context;
5103         }
5104
5105         if (output_event) {
5106                 err = perf_event_set_output(event, output_event);
5107                 if (err)
5108                         goto err_free_put_context;
5109         }
5110
5111         event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
5112         if (IS_ERR(event_file)) {
5113                 err = PTR_ERR(event_file);
5114                 goto err_free_put_context;
5115         }
5116
5117         event->filp = event_file;
5118         WARN_ON_ONCE(ctx->parent_ctx);
5119         mutex_lock(&ctx->mutex);
5120         perf_install_in_context(ctx, event, cpu);
5121         ++ctx->generation;
5122         mutex_unlock(&ctx->mutex);
5123
5124         event->owner = current;
5125         get_task_struct(current);
5126         mutex_lock(&current->perf_event_mutex);
5127         list_add_tail(&event->owner_entry, &current->perf_event_list);
5128         mutex_unlock(&current->perf_event_mutex);
5129
5130         fput_light(group_file, fput_needed);
5131         fd_install(event_fd, event_file);
5132         return event_fd;
5133
5134 err_free_put_context:
5135         free_event(event);
5136 err_put_context:
5137         fput_light(group_file, fput_needed);
5138         put_ctx(ctx);
5139 err_fd:
5140         put_unused_fd(event_fd);
5141         return err;
5142 }
5143
5144 /**
5145  * perf_event_create_kernel_counter
5146  *
5147  * @attr: attributes of the counter to create
5148  * @cpu: cpu in which the counter is bound
5149  * @pid: task to profile
5150  */
5151 struct perf_event *
5152 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
5153                                  pid_t pid,
5154                                  perf_overflow_handler_t overflow_handler)
5155 {
5156         struct perf_event *event;
5157         struct perf_event_context *ctx;
5158         int err;
5159
5160         /*
5161          * Get the target context (task or percpu):
5162          */
5163
5164         ctx = find_get_context(pid, cpu);
5165         if (IS_ERR(ctx)) {
5166                 err = PTR_ERR(ctx);
5167                 goto err_exit;
5168         }
5169
5170         event = perf_event_alloc(attr, cpu, ctx, NULL,
5171                                  NULL, overflow_handler, GFP_KERNEL);
5172         if (IS_ERR(event)) {
5173                 err = PTR_ERR(event);
5174                 goto err_put_context;
5175         }
5176
5177         event->filp = NULL;
5178         WARN_ON_ONCE(ctx->parent_ctx);
5179         mutex_lock(&ctx->mutex);
5180         perf_install_in_context(ctx, event, cpu);
5181         ++ctx->generation;
5182         mutex_unlock(&ctx->mutex);
5183
5184         event->owner = current;
5185         get_task_struct(current);
5186         mutex_lock(&current->perf_event_mutex);
5187         list_add_tail(&event->owner_entry, &current->perf_event_list);
5188         mutex_unlock(&current->perf_event_mutex);
5189
5190         return event;
5191
5192  err_put_context:
5193         put_ctx(ctx);
5194  err_exit:
5195         return ERR_PTR(err);
5196 }
5197 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
5198
5199 /*
5200  * inherit a event from parent task to child task:
5201  */
5202 static struct perf_event *
5203 inherit_event(struct perf_event *parent_event,
5204               struct task_struct *parent,
5205               struct perf_event_context *parent_ctx,
5206               struct task_struct *child,
5207               struct perf_event *group_leader,
5208               struct perf_event_context *child_ctx)
5209 {
5210         struct perf_event *child_event;
5211
5212         /*
5213          * Instead of creating recursive hierarchies of events,
5214          * we link inherited events back to the original parent,
5215          * which has a filp for sure, which we use as the reference
5216          * count:
5217          */
5218         if (parent_event->parent)
5219                 parent_event = parent_event->parent;
5220
5221         child_event = perf_event_alloc(&parent_event->attr,
5222                                            parent_event->cpu, child_ctx,
5223                                            group_leader, parent_event,
5224                                            NULL, GFP_KERNEL);
5225         if (IS_ERR(child_event))
5226                 return child_event;
5227         get_ctx(child_ctx);
5228
5229         /*
5230          * Make the child state follow the state of the parent event,
5231          * not its attr.disabled bit.  We hold the parent's mutex,
5232          * so we won't race with perf_event_{en, dis}able_family.
5233          */
5234         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
5235                 child_event->state = PERF_EVENT_STATE_INACTIVE;
5236         else
5237                 child_event->state = PERF_EVENT_STATE_OFF;
5238
5239         if (parent_event->attr.freq) {
5240                 u64 sample_period = parent_event->hw.sample_period;
5241                 struct hw_perf_event *hwc = &child_event->hw;
5242
5243                 hwc->sample_period = sample_period;
5244                 hwc->last_period   = sample_period;
5245
5246                 atomic64_set(&hwc->period_left, sample_period);
5247         }
5248
5249         child_event->overflow_handler = parent_event->overflow_handler;
5250
5251         /*
5252          * Link it up in the child's context:
5253          */
5254         add_event_to_ctx(child_event, child_ctx);
5255
5256         /*
5257          * Get a reference to the parent filp - we will fput it
5258          * when the child event exits. This is safe to do because
5259          * we are in the parent and we know that the filp still
5260          * exists and has a nonzero count:
5261          */
5262         atomic_long_inc(&parent_event->filp->f_count);
5263
5264         /*
5265          * Link this into the parent event's child list
5266          */
5267         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5268         mutex_lock(&parent_event->child_mutex);
5269         list_add_tail(&child_event->child_list, &parent_event->child_list);
5270         mutex_unlock(&parent_event->child_mutex);
5271
5272         return child_event;
5273 }
5274
5275 static int inherit_group(struct perf_event *parent_event,
5276               struct task_struct *parent,
5277               struct perf_event_context *parent_ctx,
5278               struct task_struct *child,
5279               struct perf_event_context *child_ctx)
5280 {
5281         struct perf_event *leader;
5282         struct perf_event *sub;
5283         struct perf_event *child_ctr;
5284
5285         leader = inherit_event(parent_event, parent, parent_ctx,
5286                                  child, NULL, child_ctx);
5287         if (IS_ERR(leader))
5288                 return PTR_ERR(leader);
5289         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5290                 child_ctr = inherit_event(sub, parent, parent_ctx,
5291                                             child, leader, child_ctx);
5292                 if (IS_ERR(child_ctr))
5293                         return PTR_ERR(child_ctr);
5294         }
5295         return 0;
5296 }
5297
5298 static void sync_child_event(struct perf_event *child_event,
5299                                struct task_struct *child)
5300 {
5301         struct perf_event *parent_event = child_event->parent;
5302         u64 child_val;
5303
5304         if (child_event->attr.inherit_stat)
5305                 perf_event_read_event(child_event, child);
5306
5307         child_val = atomic64_read(&child_event->count);
5308
5309         /*
5310          * Add back the child's count to the parent's count:
5311          */
5312         atomic64_add(child_val, &parent_event->count);
5313         atomic64_add(child_event->total_time_enabled,
5314                      &parent_event->child_total_time_enabled);
5315         atomic64_add(child_event->total_time_running,
5316                      &parent_event->child_total_time_running);
5317
5318         /*
5319          * Remove this event from the parent's list
5320          */
5321         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5322         mutex_lock(&parent_event->child_mutex);
5323         list_del_init(&child_event->child_list);
5324         mutex_unlock(&parent_event->child_mutex);
5325
5326         /*
5327          * Release the parent event, if this was the last
5328          * reference to it.
5329          */
5330         fput(parent_event->filp);
5331 }
5332
5333 static void
5334 __perf_event_exit_task(struct perf_event *child_event,
5335                          struct perf_event_context *child_ctx,
5336                          struct task_struct *child)
5337 {
5338         struct perf_event *parent_event;
5339
5340         perf_event_remove_from_context(child_event);
5341
5342         parent_event = child_event->parent;
5343         /*
5344          * It can happen that parent exits first, and has events
5345          * that are still around due to the child reference. These
5346          * events need to be zapped - but otherwise linger.
5347          */
5348         if (parent_event) {
5349                 sync_child_event(child_event, child);
5350                 free_event(child_event);
5351         }
5352 }
5353
5354 /*
5355  * When a child task exits, feed back event values to parent events.
5356  */
5357 void perf_event_exit_task(struct task_struct *child)
5358 {
5359         struct perf_event *child_event, *tmp;
5360         struct perf_event_context *child_ctx;
5361         unsigned long flags;
5362
5363         if (likely(!child->perf_event_ctxp)) {
5364                 perf_event_task(child, NULL, 0);
5365                 return;
5366         }
5367
5368         local_irq_save(flags);
5369         /*
5370          * We can't reschedule here because interrupts are disabled,
5371          * and either child is current or it is a task that can't be
5372          * scheduled, so we are now safe from rescheduling changing
5373          * our context.
5374          */
5375         child_ctx = child->perf_event_ctxp;
5376         __perf_event_task_sched_out(child_ctx);
5377
5378         /*
5379          * Take the context lock here so that if find_get_context is
5380          * reading child->perf_event_ctxp, we wait until it has
5381          * incremented the context's refcount before we do put_ctx below.
5382          */
5383         raw_spin_lock(&child_ctx->lock);
5384         child->perf_event_ctxp = NULL;
5385         /*
5386          * If this context is a clone; unclone it so it can't get
5387          * swapped to another process while we're removing all
5388          * the events from it.
5389          */
5390         unclone_ctx(child_ctx);
5391         update_context_time(child_ctx);
5392         raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5393
5394         /*
5395          * Report the task dead after unscheduling the events so that we
5396          * won't get any samples after PERF_RECORD_EXIT. We can however still
5397          * get a few PERF_RECORD_READ events.
5398          */
5399         perf_event_task(child, child_ctx, 0);
5400
5401         /*
5402          * We can recurse on the same lock type through:
5403          *
5404          *   __perf_event_exit_task()
5405          *     sync_child_event()
5406          *       fput(parent_event->filp)
5407          *         perf_release()
5408          *           mutex_lock(&ctx->mutex)
5409          *
5410          * But since its the parent context it won't be the same instance.
5411          */
5412         mutex_lock(&child_ctx->mutex);
5413
5414 again:
5415         list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5416                                  group_entry)
5417                 __perf_event_exit_task(child_event, child_ctx, child);
5418
5419         list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5420                                  group_entry)
5421                 __perf_event_exit_task(child_event, child_ctx, child);
5422
5423         /*
5424          * If the last event was a group event, it will have appended all
5425          * its siblings to the list, but we obtained 'tmp' before that which
5426          * will still point to the list head terminating the iteration.
5427          */
5428         if (!list_empty(&child_ctx->pinned_groups) ||
5429             !list_empty(&child_ctx->flexible_groups))
5430                 goto again;
5431
5432         mutex_unlock(&child_ctx->mutex);
5433
5434         put_ctx(child_ctx);
5435 }
5436
5437 static void perf_free_event(struct perf_event *event,
5438                             struct perf_event_context *ctx)
5439 {
5440         struct perf_event *parent = event->parent;
5441
5442         if (WARN_ON_ONCE(!parent))
5443                 return;
5444
5445         mutex_lock(&parent->child_mutex);
5446         list_del_init(&event->child_list);
5447         mutex_unlock(&parent->child_mutex);
5448
5449         fput(parent->filp);
5450
5451         list_del_event(event, ctx);
5452         free_event(event);
5453 }
5454
5455 /*
5456  * free an unexposed, unused context as created by inheritance by
5457  * init_task below, used by fork() in case of fail.
5458  */
5459 void perf_event_free_task(struct task_struct *task)
5460 {
5461         struct perf_event_context *ctx = task->perf_event_ctxp;
5462         struct perf_event *event, *tmp;
5463
5464         if (!ctx)
5465                 return;
5466
5467         mutex_lock(&ctx->mutex);
5468 again:
5469         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5470                 perf_free_event(event, ctx);
5471
5472         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5473                                  group_entry)
5474                 perf_free_event(event, ctx);
5475
5476         if (!list_empty(&ctx->pinned_groups) ||
5477             !list_empty(&ctx->flexible_groups))
5478                 goto again;
5479
5480         mutex_unlock(&ctx->mutex);
5481
5482         put_ctx(ctx);
5483 }
5484
5485 static int
5486 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5487                    struct perf_event_context *parent_ctx,
5488                    struct task_struct *child,
5489                    int *inherited_all)
5490 {
5491         int ret;
5492         struct perf_event_context *child_ctx = child->perf_event_ctxp;
5493
5494         if (!event->attr.inherit) {
5495                 *inherited_all = 0;
5496                 return 0;
5497         }
5498
5499         if (!child_ctx) {
5500                 /*
5501                  * This is executed from the parent task context, so
5502                  * inherit events that have been marked for cloning.
5503                  * First allocate and initialize a context for the
5504                  * child.
5505                  */
5506
5507                 child_ctx = kzalloc(sizeof(struct perf_event_context),
5508                                     GFP_KERNEL);
5509                 if (!child_ctx)
5510                         return -ENOMEM;
5511
5512                 __perf_event_init_context(child_ctx, child);
5513                 child->perf_event_ctxp = child_ctx;
5514                 get_task_struct(child);
5515         }
5516
5517         ret = inherit_group(event, parent, parent_ctx,
5518                             child, child_ctx);
5519
5520         if (ret)
5521                 *inherited_all = 0;
5522
5523         return ret;
5524 }
5525
5526
5527 /*
5528  * Initialize the perf_event context in task_struct
5529  */
5530 int perf_event_init_task(struct task_struct *child)
5531 {
5532         struct perf_event_context *child_ctx, *parent_ctx;
5533         struct perf_event_context *cloned_ctx;
5534         struct perf_event *event;
5535         struct task_struct *parent = current;
5536         int inherited_all = 1;
5537         int ret = 0;
5538
5539         child->perf_event_ctxp = NULL;
5540
5541         mutex_init(&child->perf_event_mutex);
5542         INIT_LIST_HEAD(&child->perf_event_list);
5543
5544         if (likely(!parent->perf_event_ctxp))
5545                 return 0;
5546
5547         /*
5548          * If the parent's context is a clone, pin it so it won't get
5549          * swapped under us.
5550          */
5551         parent_ctx = perf_pin_task_context(parent);
5552
5553         /*
5554          * No need to check if parent_ctx != NULL here; since we saw
5555          * it non-NULL earlier, the only reason for it to become NULL
5556          * is if we exit, and since we're currently in the middle of
5557          * a fork we can't be exiting at the same time.
5558          */
5559
5560         /*
5561          * Lock the parent list. No need to lock the child - not PID
5562          * hashed yet and not running, so nobody can access it.
5563          */
5564         mutex_lock(&parent_ctx->mutex);
5565
5566         /*
5567          * We dont have to disable NMIs - we are only looking at
5568          * the list, not manipulating it:
5569          */
5570         list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5571                 ret = inherit_task_group(event, parent, parent_ctx, child,
5572                                          &inherited_all);
5573                 if (ret)
5574                         break;
5575         }
5576
5577         list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5578                 ret = inherit_task_group(event, parent, parent_ctx, child,
5579                                          &inherited_all);
5580                 if (ret)
5581                         break;
5582         }
5583
5584         child_ctx = child->perf_event_ctxp;
5585
5586         if (child_ctx && inherited_all) {
5587                 /*
5588                  * Mark the child context as a clone of the parent
5589                  * context, or of whatever the parent is a clone of.
5590                  * Note that if the parent is a clone, it could get
5591                  * uncloned at any point, but that doesn't matter
5592                  * because the list of events and the generation
5593                  * count can't have changed since we took the mutex.
5594                  */
5595                 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5596                 if (cloned_ctx) {
5597                         child_ctx->parent_ctx = cloned_ctx;
5598                         child_ctx->parent_gen = parent_ctx->parent_gen;
5599                 } else {
5600                         child_ctx->parent_ctx = parent_ctx;
5601                         child_ctx->parent_gen = parent_ctx->generation;
5602                 }
5603                 get_ctx(child_ctx->parent_ctx);
5604         }
5605
5606         mutex_unlock(&parent_ctx->mutex);
5607
5608         perf_unpin_context(parent_ctx);
5609
5610         return ret;
5611 }
5612
5613 static void __init perf_event_init_all_cpus(void)
5614 {
5615         int cpu;
5616         struct perf_cpu_context *cpuctx;
5617
5618         for_each_possible_cpu(cpu) {
5619                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5620                 mutex_init(&cpuctx->hlist_mutex);
5621                 __perf_event_init_context(&cpuctx->ctx, NULL);
5622         }
5623 }
5624
5625 static void __cpuinit perf_event_init_cpu(int cpu)
5626 {
5627         struct perf_cpu_context *cpuctx;
5628
5629         cpuctx = &per_cpu(perf_cpu_context, cpu);
5630
5631         spin_lock(&perf_resource_lock);
5632         cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5633         spin_unlock(&perf_resource_lock);
5634
5635         mutex_lock(&cpuctx->hlist_mutex);
5636         if (cpuctx->hlist_refcount > 0) {
5637                 struct swevent_hlist *hlist;
5638
5639                 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5640                 WARN_ON_ONCE(!hlist);
5641                 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
5642         }
5643         mutex_unlock(&cpuctx->hlist_mutex);
5644 }
5645
5646 #ifdef CONFIG_HOTPLUG_CPU
5647 static void __perf_event_exit_cpu(void *info)
5648 {
5649         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5650         struct perf_event_context *ctx = &cpuctx->ctx;
5651         struct perf_event *event, *tmp;
5652
5653         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5654                 __perf_event_remove_from_context(event);
5655         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5656                 __perf_event_remove_from_context(event);
5657 }
5658 static void perf_event_exit_cpu(int cpu)
5659 {
5660         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5661         struct perf_event_context *ctx = &cpuctx->ctx;
5662
5663         mutex_lock(&cpuctx->hlist_mutex);
5664         swevent_hlist_release(cpuctx);
5665         mutex_unlock(&cpuctx->hlist_mutex);
5666
5667         mutex_lock(&ctx->mutex);
5668         smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5669         mutex_unlock(&ctx->mutex);
5670 }
5671 #else
5672 static inline void perf_event_exit_cpu(int cpu) { }
5673 #endif
5674
5675 static int __cpuinit
5676 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5677 {
5678         unsigned int cpu = (long)hcpu;
5679
5680         switch (action) {
5681
5682         case CPU_UP_PREPARE:
5683         case CPU_UP_PREPARE_FROZEN:
5684                 perf_event_init_cpu(cpu);
5685                 break;
5686
5687         case CPU_DOWN_PREPARE:
5688         case CPU_DOWN_PREPARE_FROZEN:
5689                 perf_event_exit_cpu(cpu);
5690                 break;
5691
5692         default:
5693                 break;
5694         }
5695
5696         return NOTIFY_OK;
5697 }
5698
5699 /*
5700  * This has to have a higher priority than migration_notifier in sched.c.
5701  */
5702 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5703         .notifier_call          = perf_cpu_notify,
5704         .priority               = 20,
5705 };
5706
5707 void __init perf_event_init(void)
5708 {
5709         perf_event_init_all_cpus();
5710         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5711                         (void *)(long)smp_processor_id());
5712         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5713                         (void *)(long)smp_processor_id());
5714         register_cpu_notifier(&perf_cpu_nb);
5715 }
5716
5717 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class,
5718                                         struct sysdev_class_attribute *attr,
5719                                         char *buf)
5720 {
5721         return sprintf(buf, "%d\n", perf_reserved_percpu);
5722 }
5723
5724 static ssize_t
5725 perf_set_reserve_percpu(struct sysdev_class *class,
5726                         struct sysdev_class_attribute *attr,
5727                         const char *buf,
5728                         size_t count)
5729 {
5730         struct perf_cpu_context *cpuctx;
5731         unsigned long val;
5732         int err, cpu, mpt;
5733
5734         err = strict_strtoul(buf, 10, &val);
5735         if (err)
5736                 return err;
5737         if (val > perf_max_events)
5738                 return -EINVAL;
5739
5740         spin_lock(&perf_resource_lock);
5741         perf_reserved_percpu = val;
5742         for_each_online_cpu(cpu) {
5743                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5744                 raw_spin_lock_irq(&cpuctx->ctx.lock);
5745                 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5746                           perf_max_events - perf_reserved_percpu);
5747                 cpuctx->max_pertask = mpt;
5748                 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5749         }
5750         spin_unlock(&perf_resource_lock);
5751
5752         return count;
5753 }
5754
5755 static ssize_t perf_show_overcommit(struct sysdev_class *class,
5756                                     struct sysdev_class_attribute *attr,
5757                                     char *buf)
5758 {
5759         return sprintf(buf, "%d\n", perf_overcommit);
5760 }
5761
5762 static ssize_t
5763 perf_set_overcommit(struct sysdev_class *class,
5764                     struct sysdev_class_attribute *attr,
5765                     const char *buf, size_t count)
5766 {
5767         unsigned long val;
5768         int err;
5769
5770         err = strict_strtoul(buf, 10, &val);
5771         if (err)
5772                 return err;
5773         if (val > 1)
5774                 return -EINVAL;
5775
5776         spin_lock(&perf_resource_lock);
5777         perf_overcommit = val;
5778         spin_unlock(&perf_resource_lock);
5779
5780         return count;
5781 }
5782
5783 static SYSDEV_CLASS_ATTR(
5784                                 reserve_percpu,
5785                                 0644,
5786                                 perf_show_reserve_percpu,
5787                                 perf_set_reserve_percpu
5788                         );
5789
5790 static SYSDEV_CLASS_ATTR(
5791                                 overcommit,
5792                                 0644,
5793                                 perf_show_overcommit,
5794                                 perf_set_overcommit
5795                         );
5796
5797 static struct attribute *perfclass_attrs[] = {
5798         &attr_reserve_percpu.attr,
5799         &attr_overcommit.attr,
5800         NULL
5801 };
5802
5803 static struct attribute_group perfclass_attr_group = {
5804         .attrs                  = perfclass_attrs,
5805         .name                   = "perf_events",
5806 };
5807
5808 static int __init perf_event_sysfs_init(void)
5809 {
5810         return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5811                                   &perfclass_attr_group);
5812 }
5813 device_initcall(perf_event_sysfs_init);