sched_clock: Add local_clock() API and improve documentation
[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 local_clock();
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         WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
287         event->attach_state |= PERF_ATTACH_CONTEXT;
288
289         /*
290          * If we're a stand alone event or group leader, we go to the context
291          * list, group events are kept attached to the group so that
292          * perf_group_detach can, at all times, locate all siblings.
293          */
294         if (event->group_leader == event) {
295                 struct list_head *list;
296
297                 if (is_software_event(event))
298                         event->group_flags |= PERF_GROUP_SOFTWARE;
299
300                 list = ctx_group_list(event, ctx);
301                 list_add_tail(&event->group_entry, list);
302         }
303
304         list_add_rcu(&event->event_entry, &ctx->event_list);
305         ctx->nr_events++;
306         if (event->attr.inherit_stat)
307                 ctx->nr_stat++;
308 }
309
310 static void perf_group_attach(struct perf_event *event)
311 {
312         struct perf_event *group_leader = event->group_leader;
313
314         WARN_ON_ONCE(event->attach_state & PERF_ATTACH_GROUP);
315         event->attach_state |= PERF_ATTACH_GROUP;
316
317         if (group_leader == event)
318                 return;
319
320         if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
321                         !is_software_event(event))
322                 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
323
324         list_add_tail(&event->group_entry, &group_leader->sibling_list);
325         group_leader->nr_siblings++;
326 }
327
328 /*
329  * Remove a event from the lists for its context.
330  * Must be called with ctx->mutex and ctx->lock held.
331  */
332 static void
333 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
334 {
335         /*
336          * We can have double detach due to exit/hot-unplug + close.
337          */
338         if (!(event->attach_state & PERF_ATTACH_CONTEXT))
339                 return;
340
341         event->attach_state &= ~PERF_ATTACH_CONTEXT;
342
343         ctx->nr_events--;
344         if (event->attr.inherit_stat)
345                 ctx->nr_stat--;
346
347         list_del_rcu(&event->event_entry);
348
349         if (event->group_leader == event)
350                 list_del_init(&event->group_entry);
351
352         update_group_times(event);
353
354         /*
355          * If event was in error state, then keep it
356          * that way, otherwise bogus counts will be
357          * returned on read(). The only way to get out
358          * of error state is by explicit re-enabling
359          * of the event
360          */
361         if (event->state > PERF_EVENT_STATE_OFF)
362                 event->state = PERF_EVENT_STATE_OFF;
363 }
364
365 static void perf_group_detach(struct perf_event *event)
366 {
367         struct perf_event *sibling, *tmp;
368         struct list_head *list = NULL;
369
370         /*
371          * We can have double detach due to exit/hot-unplug + close.
372          */
373         if (!(event->attach_state & PERF_ATTACH_GROUP))
374                 return;
375
376         event->attach_state &= ~PERF_ATTACH_GROUP;
377
378         /*
379          * If this is a sibling, remove it from its group.
380          */
381         if (event->group_leader != event) {
382                 list_del_init(&event->group_entry);
383                 event->group_leader->nr_siblings--;
384                 return;
385         }
386
387         if (!list_empty(&event->group_entry))
388                 list = &event->group_entry;
389
390         /*
391          * If this was a group event with sibling events then
392          * upgrade the siblings to singleton events by adding them
393          * to whatever list we are on.
394          */
395         list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
396                 if (list)
397                         list_move_tail(&sibling->group_entry, list);
398                 sibling->group_leader = sibling;
399
400                 /* Inherit group flags from the previous leader */
401                 sibling->group_flags = event->group_flags;
402         }
403 }
404
405 static void
406 event_sched_out(struct perf_event *event,
407                   struct perf_cpu_context *cpuctx,
408                   struct perf_event_context *ctx)
409 {
410         if (event->state != PERF_EVENT_STATE_ACTIVE)
411                 return;
412
413         event->state = PERF_EVENT_STATE_INACTIVE;
414         if (event->pending_disable) {
415                 event->pending_disable = 0;
416                 event->state = PERF_EVENT_STATE_OFF;
417         }
418         event->tstamp_stopped = ctx->time;
419         event->pmu->disable(event);
420         event->oncpu = -1;
421
422         if (!is_software_event(event))
423                 cpuctx->active_oncpu--;
424         ctx->nr_active--;
425         if (event->attr.exclusive || !cpuctx->active_oncpu)
426                 cpuctx->exclusive = 0;
427 }
428
429 static void
430 group_sched_out(struct perf_event *group_event,
431                 struct perf_cpu_context *cpuctx,
432                 struct perf_event_context *ctx)
433 {
434         struct perf_event *event;
435
436         if (group_event->state != PERF_EVENT_STATE_ACTIVE)
437                 return;
438
439         event_sched_out(group_event, cpuctx, ctx);
440
441         /*
442          * Schedule out siblings (if any):
443          */
444         list_for_each_entry(event, &group_event->sibling_list, group_entry)
445                 event_sched_out(event, cpuctx, ctx);
446
447         if (group_event->attr.exclusive)
448                 cpuctx->exclusive = 0;
449 }
450
451 /*
452  * Cross CPU call to remove a performance event
453  *
454  * We disable the event on the hardware level first. After that we
455  * remove it from the context list.
456  */
457 static void __perf_event_remove_from_context(void *info)
458 {
459         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
460         struct perf_event *event = info;
461         struct perf_event_context *ctx = event->ctx;
462
463         /*
464          * If this is a task context, we need to check whether it is
465          * the current task context of this cpu. If not it has been
466          * scheduled out before the smp call arrived.
467          */
468         if (ctx->task && cpuctx->task_ctx != ctx)
469                 return;
470
471         raw_spin_lock(&ctx->lock);
472         /*
473          * Protect the list operation against NMI by disabling the
474          * events on a global level.
475          */
476         perf_disable();
477
478         event_sched_out(event, cpuctx, ctx);
479
480         list_del_event(event, ctx);
481
482         if (!ctx->task) {
483                 /*
484                  * Allow more per task events with respect to the
485                  * reservation:
486                  */
487                 cpuctx->max_pertask =
488                         min(perf_max_events - ctx->nr_events,
489                             perf_max_events - perf_reserved_percpu);
490         }
491
492         perf_enable();
493         raw_spin_unlock(&ctx->lock);
494 }
495
496
497 /*
498  * Remove the event from a task's (or a CPU's) list of events.
499  *
500  * Must be called with ctx->mutex held.
501  *
502  * CPU events are removed with a smp call. For task events we only
503  * call when the task is on a CPU.
504  *
505  * If event->ctx is a cloned context, callers must make sure that
506  * every task struct that event->ctx->task could possibly point to
507  * remains valid.  This is OK when called from perf_release since
508  * that only calls us on the top-level context, which can't be a clone.
509  * When called from perf_event_exit_task, it's OK because the
510  * context has been detached from its task.
511  */
512 static void perf_event_remove_from_context(struct perf_event *event)
513 {
514         struct perf_event_context *ctx = event->ctx;
515         struct task_struct *task = ctx->task;
516
517         if (!task) {
518                 /*
519                  * Per cpu events are removed via an smp call and
520                  * the removal is always successful.
521                  */
522                 smp_call_function_single(event->cpu,
523                                          __perf_event_remove_from_context,
524                                          event, 1);
525                 return;
526         }
527
528 retry:
529         task_oncpu_function_call(task, __perf_event_remove_from_context,
530                                  event);
531
532         raw_spin_lock_irq(&ctx->lock);
533         /*
534          * If the context is active we need to retry the smp call.
535          */
536         if (ctx->nr_active && !list_empty(&event->group_entry)) {
537                 raw_spin_unlock_irq(&ctx->lock);
538                 goto retry;
539         }
540
541         /*
542          * The lock prevents that this context is scheduled in so we
543          * can remove the event safely, if the call above did not
544          * succeed.
545          */
546         if (!list_empty(&event->group_entry))
547                 list_del_event(event, ctx);
548         raw_spin_unlock_irq(&ctx->lock);
549 }
550
551 /*
552  * Cross CPU call to disable a performance event
553  */
554 static void __perf_event_disable(void *info)
555 {
556         struct perf_event *event = info;
557         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
558         struct perf_event_context *ctx = event->ctx;
559
560         /*
561          * If this is a per-task event, need to check whether this
562          * event's task is the current task on this cpu.
563          */
564         if (ctx->task && cpuctx->task_ctx != ctx)
565                 return;
566
567         raw_spin_lock(&ctx->lock);
568
569         /*
570          * If the event is on, turn it off.
571          * If it is in error state, leave it in error state.
572          */
573         if (event->state >= PERF_EVENT_STATE_INACTIVE) {
574                 update_context_time(ctx);
575                 update_group_times(event);
576                 if (event == event->group_leader)
577                         group_sched_out(event, cpuctx, ctx);
578                 else
579                         event_sched_out(event, cpuctx, ctx);
580                 event->state = PERF_EVENT_STATE_OFF;
581         }
582
583         raw_spin_unlock(&ctx->lock);
584 }
585
586 /*
587  * Disable a event.
588  *
589  * If event->ctx is a cloned context, callers must make sure that
590  * every task struct that event->ctx->task could possibly point to
591  * remains valid.  This condition is satisifed when called through
592  * perf_event_for_each_child or perf_event_for_each because they
593  * hold the top-level event's child_mutex, so any descendant that
594  * goes to exit will block in sync_child_event.
595  * When called from perf_pending_event it's OK because event->ctx
596  * is the current context on this CPU and preemption is disabled,
597  * hence we can't get into perf_event_task_sched_out for this context.
598  */
599 void perf_event_disable(struct perf_event *event)
600 {
601         struct perf_event_context *ctx = event->ctx;
602         struct task_struct *task = ctx->task;
603
604         if (!task) {
605                 /*
606                  * Disable the event on the cpu that it's on
607                  */
608                 smp_call_function_single(event->cpu, __perf_event_disable,
609                                          event, 1);
610                 return;
611         }
612
613  retry:
614         task_oncpu_function_call(task, __perf_event_disable, event);
615
616         raw_spin_lock_irq(&ctx->lock);
617         /*
618          * If the event is still active, we need to retry the cross-call.
619          */
620         if (event->state == PERF_EVENT_STATE_ACTIVE) {
621                 raw_spin_unlock_irq(&ctx->lock);
622                 goto retry;
623         }
624
625         /*
626          * Since we have the lock this context can't be scheduled
627          * in, so we can change the state safely.
628          */
629         if (event->state == PERF_EVENT_STATE_INACTIVE) {
630                 update_group_times(event);
631                 event->state = PERF_EVENT_STATE_OFF;
632         }
633
634         raw_spin_unlock_irq(&ctx->lock);
635 }
636
637 static int
638 event_sched_in(struct perf_event *event,
639                  struct perf_cpu_context *cpuctx,
640                  struct perf_event_context *ctx)
641 {
642         if (event->state <= PERF_EVENT_STATE_OFF)
643                 return 0;
644
645         event->state = PERF_EVENT_STATE_ACTIVE;
646         event->oncpu = smp_processor_id();
647         /*
648          * The new state must be visible before we turn it on in the hardware:
649          */
650         smp_wmb();
651
652         if (event->pmu->enable(event)) {
653                 event->state = PERF_EVENT_STATE_INACTIVE;
654                 event->oncpu = -1;
655                 return -EAGAIN;
656         }
657
658         event->tstamp_running += ctx->time - event->tstamp_stopped;
659
660         if (!is_software_event(event))
661                 cpuctx->active_oncpu++;
662         ctx->nr_active++;
663
664         if (event->attr.exclusive)
665                 cpuctx->exclusive = 1;
666
667         return 0;
668 }
669
670 static int
671 group_sched_in(struct perf_event *group_event,
672                struct perf_cpu_context *cpuctx,
673                struct perf_event_context *ctx)
674 {
675         struct perf_event *event, *partial_group = NULL;
676         const struct pmu *pmu = group_event->pmu;
677         bool txn = false;
678         int ret;
679
680         if (group_event->state == PERF_EVENT_STATE_OFF)
681                 return 0;
682
683         /* Check if group transaction availabe */
684         if (pmu->start_txn)
685                 txn = true;
686
687         if (txn)
688                 pmu->start_txn(pmu);
689
690         if (event_sched_in(group_event, cpuctx, ctx)) {
691                 if (txn)
692                         pmu->cancel_txn(pmu);
693                 return -EAGAIN;
694         }
695
696         /*
697          * Schedule in siblings as one group (if any):
698          */
699         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
700                 if (event_sched_in(event, cpuctx, ctx)) {
701                         partial_group = event;
702                         goto group_error;
703                 }
704         }
705
706         if (!txn)
707                 return 0;
708
709         ret = pmu->commit_txn(pmu);
710         if (!ret) {
711                 pmu->cancel_txn(pmu);
712                 return 0;
713         }
714
715 group_error:
716         /*
717          * Groups can be scheduled in as one unit only, so undo any
718          * partial group before returning:
719          */
720         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
721                 if (event == partial_group)
722                         break;
723                 event_sched_out(event, cpuctx, ctx);
724         }
725         event_sched_out(group_event, cpuctx, ctx);
726
727         if (txn)
728                 pmu->cancel_txn(pmu);
729
730         return -EAGAIN;
731 }
732
733 /*
734  * Work out whether we can put this event group on the CPU now.
735  */
736 static int group_can_go_on(struct perf_event *event,
737                            struct perf_cpu_context *cpuctx,
738                            int can_add_hw)
739 {
740         /*
741          * Groups consisting entirely of software events can always go on.
742          */
743         if (event->group_flags & PERF_GROUP_SOFTWARE)
744                 return 1;
745         /*
746          * If an exclusive group is already on, no other hardware
747          * events can go on.
748          */
749         if (cpuctx->exclusive)
750                 return 0;
751         /*
752          * If this group is exclusive and there are already
753          * events on the CPU, it can't go on.
754          */
755         if (event->attr.exclusive && cpuctx->active_oncpu)
756                 return 0;
757         /*
758          * Otherwise, try to add it if all previous groups were able
759          * to go on.
760          */
761         return can_add_hw;
762 }
763
764 static void add_event_to_ctx(struct perf_event *event,
765                                struct perf_event_context *ctx)
766 {
767         list_add_event(event, ctx);
768         perf_group_attach(event);
769         event->tstamp_enabled = ctx->time;
770         event->tstamp_running = ctx->time;
771         event->tstamp_stopped = ctx->time;
772 }
773
774 /*
775  * Cross CPU call to install and enable a performance event
776  *
777  * Must be called with ctx->mutex held
778  */
779 static void __perf_install_in_context(void *info)
780 {
781         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
782         struct perf_event *event = info;
783         struct perf_event_context *ctx = event->ctx;
784         struct perf_event *leader = event->group_leader;
785         int err;
786
787         /*
788          * If this is a task context, we need to check whether it is
789          * the current task context of this cpu. If not it has been
790          * scheduled out before the smp call arrived.
791          * Or possibly this is the right context but it isn't
792          * on this cpu because it had no events.
793          */
794         if (ctx->task && cpuctx->task_ctx != ctx) {
795                 if (cpuctx->task_ctx || ctx->task != current)
796                         return;
797                 cpuctx->task_ctx = ctx;
798         }
799
800         raw_spin_lock(&ctx->lock);
801         ctx->is_active = 1;
802         update_context_time(ctx);
803
804         /*
805          * Protect the list operation against NMI by disabling the
806          * events on a global level. NOP for non NMI based events.
807          */
808         perf_disable();
809
810         add_event_to_ctx(event, ctx);
811
812         if (event->cpu != -1 && event->cpu != smp_processor_id())
813                 goto unlock;
814
815         /*
816          * Don't put the event on if it is disabled or if
817          * it is in a group and the group isn't on.
818          */
819         if (event->state != PERF_EVENT_STATE_INACTIVE ||
820             (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
821                 goto unlock;
822
823         /*
824          * An exclusive event can't go on if there are already active
825          * hardware events, and no hardware event can go on if there
826          * is already an exclusive event on.
827          */
828         if (!group_can_go_on(event, cpuctx, 1))
829                 err = -EEXIST;
830         else
831                 err = event_sched_in(event, cpuctx, ctx);
832
833         if (err) {
834                 /*
835                  * This event couldn't go on.  If it is in a group
836                  * then we have to pull the whole group off.
837                  * If the event group is pinned then put it in error state.
838                  */
839                 if (leader != event)
840                         group_sched_out(leader, cpuctx, ctx);
841                 if (leader->attr.pinned) {
842                         update_group_times(leader);
843                         leader->state = PERF_EVENT_STATE_ERROR;
844                 }
845         }
846
847         if (!err && !ctx->task && cpuctx->max_pertask)
848                 cpuctx->max_pertask--;
849
850  unlock:
851         perf_enable();
852
853         raw_spin_unlock(&ctx->lock);
854 }
855
856 /*
857  * Attach a performance event to a context
858  *
859  * First we add the event to the list with the hardware enable bit
860  * in event->hw_config cleared.
861  *
862  * If the event is attached to a task which is on a CPU we use a smp
863  * call to enable it in the task context. The task might have been
864  * scheduled away, but we check this in the smp call again.
865  *
866  * Must be called with ctx->mutex held.
867  */
868 static void
869 perf_install_in_context(struct perf_event_context *ctx,
870                         struct perf_event *event,
871                         int cpu)
872 {
873         struct task_struct *task = ctx->task;
874
875         if (!task) {
876                 /*
877                  * Per cpu events are installed via an smp call and
878                  * the install is always successful.
879                  */
880                 smp_call_function_single(cpu, __perf_install_in_context,
881                                          event, 1);
882                 return;
883         }
884
885 retry:
886         task_oncpu_function_call(task, __perf_install_in_context,
887                                  event);
888
889         raw_spin_lock_irq(&ctx->lock);
890         /*
891          * we need to retry the smp call.
892          */
893         if (ctx->is_active && list_empty(&event->group_entry)) {
894                 raw_spin_unlock_irq(&ctx->lock);
895                 goto retry;
896         }
897
898         /*
899          * The lock prevents that this context is scheduled in so we
900          * can add the event safely, if it the call above did not
901          * succeed.
902          */
903         if (list_empty(&event->group_entry))
904                 add_event_to_ctx(event, ctx);
905         raw_spin_unlock_irq(&ctx->lock);
906 }
907
908 /*
909  * Put a event into inactive state and update time fields.
910  * Enabling the leader of a group effectively enables all
911  * the group members that aren't explicitly disabled, so we
912  * have to update their ->tstamp_enabled also.
913  * Note: this works for group members as well as group leaders
914  * since the non-leader members' sibling_lists will be empty.
915  */
916 static void __perf_event_mark_enabled(struct perf_event *event,
917                                         struct perf_event_context *ctx)
918 {
919         struct perf_event *sub;
920
921         event->state = PERF_EVENT_STATE_INACTIVE;
922         event->tstamp_enabled = ctx->time - event->total_time_enabled;
923         list_for_each_entry(sub, &event->sibling_list, group_entry)
924                 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
925                         sub->tstamp_enabled =
926                                 ctx->time - sub->total_time_enabled;
927 }
928
929 /*
930  * Cross CPU call to enable a performance event
931  */
932 static void __perf_event_enable(void *info)
933 {
934         struct perf_event *event = info;
935         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
936         struct perf_event_context *ctx = event->ctx;
937         struct perf_event *leader = event->group_leader;
938         int err;
939
940         /*
941          * If this is a per-task event, need to check whether this
942          * event's task is the current task on this cpu.
943          */
944         if (ctx->task && cpuctx->task_ctx != ctx) {
945                 if (cpuctx->task_ctx || ctx->task != current)
946                         return;
947                 cpuctx->task_ctx = ctx;
948         }
949
950         raw_spin_lock(&ctx->lock);
951         ctx->is_active = 1;
952         update_context_time(ctx);
953
954         if (event->state >= PERF_EVENT_STATE_INACTIVE)
955                 goto unlock;
956         __perf_event_mark_enabled(event, ctx);
957
958         if (event->cpu != -1 && event->cpu != smp_processor_id())
959                 goto unlock;
960
961         /*
962          * If the event is in a group and isn't the group leader,
963          * then don't put it on unless the group is on.
964          */
965         if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
966                 goto unlock;
967
968         if (!group_can_go_on(event, cpuctx, 1)) {
969                 err = -EEXIST;
970         } else {
971                 perf_disable();
972                 if (event == leader)
973                         err = group_sched_in(event, cpuctx, ctx);
974                 else
975                         err = event_sched_in(event, cpuctx, ctx);
976                 perf_enable();
977         }
978
979         if (err) {
980                 /*
981                  * If this event can't go on and it's part of a
982                  * group, then the whole group has to come off.
983                  */
984                 if (leader != event)
985                         group_sched_out(leader, cpuctx, ctx);
986                 if (leader->attr.pinned) {
987                         update_group_times(leader);
988                         leader->state = PERF_EVENT_STATE_ERROR;
989                 }
990         }
991
992  unlock:
993         raw_spin_unlock(&ctx->lock);
994 }
995
996 /*
997  * Enable a event.
998  *
999  * If event->ctx is a cloned context, callers must make sure that
1000  * every task struct that event->ctx->task could possibly point to
1001  * remains valid.  This condition is satisfied when called through
1002  * perf_event_for_each_child or perf_event_for_each as described
1003  * for perf_event_disable.
1004  */
1005 void perf_event_enable(struct perf_event *event)
1006 {
1007         struct perf_event_context *ctx = event->ctx;
1008         struct task_struct *task = ctx->task;
1009
1010         if (!task) {
1011                 /*
1012                  * Enable the event on the cpu that it's on
1013                  */
1014                 smp_call_function_single(event->cpu, __perf_event_enable,
1015                                          event, 1);
1016                 return;
1017         }
1018
1019         raw_spin_lock_irq(&ctx->lock);
1020         if (event->state >= PERF_EVENT_STATE_INACTIVE)
1021                 goto out;
1022
1023         /*
1024          * If the event is in error state, clear that first.
1025          * That way, if we see the event in error state below, we
1026          * know that it has gone back into error state, as distinct
1027          * from the task having been scheduled away before the
1028          * cross-call arrived.
1029          */
1030         if (event->state == PERF_EVENT_STATE_ERROR)
1031                 event->state = PERF_EVENT_STATE_OFF;
1032
1033  retry:
1034         raw_spin_unlock_irq(&ctx->lock);
1035         task_oncpu_function_call(task, __perf_event_enable, event);
1036
1037         raw_spin_lock_irq(&ctx->lock);
1038
1039         /*
1040          * If the context is active and the event is still off,
1041          * we need to retry the cross-call.
1042          */
1043         if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1044                 goto retry;
1045
1046         /*
1047          * Since we have the lock this context can't be scheduled
1048          * in, so we can change the state safely.
1049          */
1050         if (event->state == PERF_EVENT_STATE_OFF)
1051                 __perf_event_mark_enabled(event, ctx);
1052
1053  out:
1054         raw_spin_unlock_irq(&ctx->lock);
1055 }
1056
1057 static int perf_event_refresh(struct perf_event *event, int refresh)
1058 {
1059         /*
1060          * not supported on inherited events
1061          */
1062         if (event->attr.inherit)
1063                 return -EINVAL;
1064
1065         atomic_add(refresh, &event->event_limit);
1066         perf_event_enable(event);
1067
1068         return 0;
1069 }
1070
1071 enum event_type_t {
1072         EVENT_FLEXIBLE = 0x1,
1073         EVENT_PINNED = 0x2,
1074         EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1075 };
1076
1077 static void ctx_sched_out(struct perf_event_context *ctx,
1078                           struct perf_cpu_context *cpuctx,
1079                           enum event_type_t event_type)
1080 {
1081         struct perf_event *event;
1082
1083         raw_spin_lock(&ctx->lock);
1084         ctx->is_active = 0;
1085         if (likely(!ctx->nr_events))
1086                 goto out;
1087         update_context_time(ctx);
1088
1089         perf_disable();
1090         if (!ctx->nr_active)
1091                 goto out_enable;
1092
1093         if (event_type & EVENT_PINNED)
1094                 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1095                         group_sched_out(event, cpuctx, ctx);
1096
1097         if (event_type & EVENT_FLEXIBLE)
1098                 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1099                         group_sched_out(event, cpuctx, ctx);
1100
1101  out_enable:
1102         perf_enable();
1103  out:
1104         raw_spin_unlock(&ctx->lock);
1105 }
1106
1107 /*
1108  * Test whether two contexts are equivalent, i.e. whether they
1109  * have both been cloned from the same version of the same context
1110  * and they both have the same number of enabled events.
1111  * If the number of enabled events is the same, then the set
1112  * of enabled events should be the same, because these are both
1113  * inherited contexts, therefore we can't access individual events
1114  * in them directly with an fd; we can only enable/disable all
1115  * events via prctl, or enable/disable all events in a family
1116  * via ioctl, which will have the same effect on both contexts.
1117  */
1118 static int context_equiv(struct perf_event_context *ctx1,
1119                          struct perf_event_context *ctx2)
1120 {
1121         return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1122                 && ctx1->parent_gen == ctx2->parent_gen
1123                 && !ctx1->pin_count && !ctx2->pin_count;
1124 }
1125
1126 static void __perf_event_sync_stat(struct perf_event *event,
1127                                      struct perf_event *next_event)
1128 {
1129         u64 value;
1130
1131         if (!event->attr.inherit_stat)
1132                 return;
1133
1134         /*
1135          * Update the event value, we cannot use perf_event_read()
1136          * because we're in the middle of a context switch and have IRQs
1137          * disabled, which upsets smp_call_function_single(), however
1138          * we know the event must be on the current CPU, therefore we
1139          * don't need to use it.
1140          */
1141         switch (event->state) {
1142         case PERF_EVENT_STATE_ACTIVE:
1143                 event->pmu->read(event);
1144                 /* fall-through */
1145
1146         case PERF_EVENT_STATE_INACTIVE:
1147                 update_event_times(event);
1148                 break;
1149
1150         default:
1151                 break;
1152         }
1153
1154         /*
1155          * In order to keep per-task stats reliable we need to flip the event
1156          * values when we flip the contexts.
1157          */
1158         value = atomic64_read(&next_event->count);
1159         value = atomic64_xchg(&event->count, value);
1160         atomic64_set(&next_event->count, value);
1161
1162         swap(event->total_time_enabled, next_event->total_time_enabled);
1163         swap(event->total_time_running, next_event->total_time_running);
1164
1165         /*
1166          * Since we swizzled the values, update the user visible data too.
1167          */
1168         perf_event_update_userpage(event);
1169         perf_event_update_userpage(next_event);
1170 }
1171
1172 #define list_next_entry(pos, member) \
1173         list_entry(pos->member.next, typeof(*pos), member)
1174
1175 static void perf_event_sync_stat(struct perf_event_context *ctx,
1176                                    struct perf_event_context *next_ctx)
1177 {
1178         struct perf_event *event, *next_event;
1179
1180         if (!ctx->nr_stat)
1181                 return;
1182
1183         update_context_time(ctx);
1184
1185         event = list_first_entry(&ctx->event_list,
1186                                    struct perf_event, event_entry);
1187
1188         next_event = list_first_entry(&next_ctx->event_list,
1189                                         struct perf_event, event_entry);
1190
1191         while (&event->event_entry != &ctx->event_list &&
1192                &next_event->event_entry != &next_ctx->event_list) {
1193
1194                 __perf_event_sync_stat(event, next_event);
1195
1196                 event = list_next_entry(event, event_entry);
1197                 next_event = list_next_entry(next_event, event_entry);
1198         }
1199 }
1200
1201 /*
1202  * Called from scheduler to remove the events of the current task,
1203  * with interrupts disabled.
1204  *
1205  * We stop each event and update the event value in event->count.
1206  *
1207  * This does not protect us against NMI, but disable()
1208  * sets the disabled bit in the control field of event _before_
1209  * accessing the event control register. If a NMI hits, then it will
1210  * not restart the event.
1211  */
1212 void perf_event_task_sched_out(struct task_struct *task,
1213                                  struct task_struct *next)
1214 {
1215         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1216         struct perf_event_context *ctx = task->perf_event_ctxp;
1217         struct perf_event_context *next_ctx;
1218         struct perf_event_context *parent;
1219         int do_switch = 1;
1220
1221         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0);
1222
1223         if (likely(!ctx || !cpuctx->task_ctx))
1224                 return;
1225
1226         rcu_read_lock();
1227         parent = rcu_dereference(ctx->parent_ctx);
1228         next_ctx = next->perf_event_ctxp;
1229         if (parent && next_ctx &&
1230             rcu_dereference(next_ctx->parent_ctx) == parent) {
1231                 /*
1232                  * Looks like the two contexts are clones, so we might be
1233                  * able to optimize the context switch.  We lock both
1234                  * contexts and check that they are clones under the
1235                  * lock (including re-checking that neither has been
1236                  * uncloned in the meantime).  It doesn't matter which
1237                  * order we take the locks because no other cpu could
1238                  * be trying to lock both of these tasks.
1239                  */
1240                 raw_spin_lock(&ctx->lock);
1241                 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1242                 if (context_equiv(ctx, next_ctx)) {
1243                         /*
1244                          * XXX do we need a memory barrier of sorts
1245                          * wrt to rcu_dereference() of perf_event_ctxp
1246                          */
1247                         task->perf_event_ctxp = next_ctx;
1248                         next->perf_event_ctxp = ctx;
1249                         ctx->task = next;
1250                         next_ctx->task = task;
1251                         do_switch = 0;
1252
1253                         perf_event_sync_stat(ctx, next_ctx);
1254                 }
1255                 raw_spin_unlock(&next_ctx->lock);
1256                 raw_spin_unlock(&ctx->lock);
1257         }
1258         rcu_read_unlock();
1259
1260         if (do_switch) {
1261                 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1262                 cpuctx->task_ctx = NULL;
1263         }
1264 }
1265
1266 static void task_ctx_sched_out(struct perf_event_context *ctx,
1267                                enum event_type_t event_type)
1268 {
1269         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1270
1271         if (!cpuctx->task_ctx)
1272                 return;
1273
1274         if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1275                 return;
1276
1277         ctx_sched_out(ctx, cpuctx, event_type);
1278         cpuctx->task_ctx = NULL;
1279 }
1280
1281 /*
1282  * Called with IRQs disabled
1283  */
1284 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1285 {
1286         task_ctx_sched_out(ctx, EVENT_ALL);
1287 }
1288
1289 /*
1290  * Called with IRQs disabled
1291  */
1292 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1293                               enum event_type_t event_type)
1294 {
1295         ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1296 }
1297
1298 static void
1299 ctx_pinned_sched_in(struct perf_event_context *ctx,
1300                     struct perf_cpu_context *cpuctx)
1301 {
1302         struct perf_event *event;
1303
1304         list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1305                 if (event->state <= PERF_EVENT_STATE_OFF)
1306                         continue;
1307                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1308                         continue;
1309
1310                 if (group_can_go_on(event, cpuctx, 1))
1311                         group_sched_in(event, cpuctx, ctx);
1312
1313                 /*
1314                  * If this pinned group hasn't been scheduled,
1315                  * put it in error state.
1316                  */
1317                 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1318                         update_group_times(event);
1319                         event->state = PERF_EVENT_STATE_ERROR;
1320                 }
1321         }
1322 }
1323
1324 static void
1325 ctx_flexible_sched_in(struct perf_event_context *ctx,
1326                       struct perf_cpu_context *cpuctx)
1327 {
1328         struct perf_event *event;
1329         int can_add_hw = 1;
1330
1331         list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1332                 /* Ignore events in OFF or ERROR state */
1333                 if (event->state <= PERF_EVENT_STATE_OFF)
1334                         continue;
1335                 /*
1336                  * Listen to the 'cpu' scheduling filter constraint
1337                  * of events:
1338                  */
1339                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1340                         continue;
1341
1342                 if (group_can_go_on(event, cpuctx, can_add_hw))
1343                         if (group_sched_in(event, cpuctx, ctx))
1344                                 can_add_hw = 0;
1345         }
1346 }
1347
1348 static void
1349 ctx_sched_in(struct perf_event_context *ctx,
1350              struct perf_cpu_context *cpuctx,
1351              enum event_type_t event_type)
1352 {
1353         raw_spin_lock(&ctx->lock);
1354         ctx->is_active = 1;
1355         if (likely(!ctx->nr_events))
1356                 goto out;
1357
1358         ctx->timestamp = perf_clock();
1359
1360         perf_disable();
1361
1362         /*
1363          * First go through the list and put on any pinned groups
1364          * in order to give them the best chance of going on.
1365          */
1366         if (event_type & EVENT_PINNED)
1367                 ctx_pinned_sched_in(ctx, cpuctx);
1368
1369         /* Then walk through the lower prio flexible groups */
1370         if (event_type & EVENT_FLEXIBLE)
1371                 ctx_flexible_sched_in(ctx, cpuctx);
1372
1373         perf_enable();
1374  out:
1375         raw_spin_unlock(&ctx->lock);
1376 }
1377
1378 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1379                              enum event_type_t event_type)
1380 {
1381         struct perf_event_context *ctx = &cpuctx->ctx;
1382
1383         ctx_sched_in(ctx, cpuctx, event_type);
1384 }
1385
1386 static void task_ctx_sched_in(struct task_struct *task,
1387                               enum event_type_t event_type)
1388 {
1389         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1390         struct perf_event_context *ctx = task->perf_event_ctxp;
1391
1392         if (likely(!ctx))
1393                 return;
1394         if (cpuctx->task_ctx == ctx)
1395                 return;
1396         ctx_sched_in(ctx, cpuctx, event_type);
1397         cpuctx->task_ctx = ctx;
1398 }
1399 /*
1400  * Called from scheduler to add the events of the current task
1401  * with interrupts disabled.
1402  *
1403  * We restore the event value and then enable it.
1404  *
1405  * This does not protect us against NMI, but enable()
1406  * sets the enabled bit in the control field of event _before_
1407  * accessing the event control register. If a NMI hits, then it will
1408  * keep the event running.
1409  */
1410 void perf_event_task_sched_in(struct task_struct *task)
1411 {
1412         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1413         struct perf_event_context *ctx = task->perf_event_ctxp;
1414
1415         if (likely(!ctx))
1416                 return;
1417
1418         if (cpuctx->task_ctx == ctx)
1419                 return;
1420
1421         perf_disable();
1422
1423         /*
1424          * We want to keep the following priority order:
1425          * cpu pinned (that don't need to move), task pinned,
1426          * cpu flexible, task flexible.
1427          */
1428         cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1429
1430         ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1431         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1432         ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1433
1434         cpuctx->task_ctx = ctx;
1435
1436         perf_enable();
1437 }
1438
1439 #define MAX_INTERRUPTS (~0ULL)
1440
1441 static void perf_log_throttle(struct perf_event *event, int enable);
1442
1443 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1444 {
1445         u64 frequency = event->attr.sample_freq;
1446         u64 sec = NSEC_PER_SEC;
1447         u64 divisor, dividend;
1448
1449         int count_fls, nsec_fls, frequency_fls, sec_fls;
1450
1451         count_fls = fls64(count);
1452         nsec_fls = fls64(nsec);
1453         frequency_fls = fls64(frequency);
1454         sec_fls = 30;
1455
1456         /*
1457          * We got @count in @nsec, with a target of sample_freq HZ
1458          * the target period becomes:
1459          *
1460          *             @count * 10^9
1461          * period = -------------------
1462          *          @nsec * sample_freq
1463          *
1464          */
1465
1466         /*
1467          * Reduce accuracy by one bit such that @a and @b converge
1468          * to a similar magnitude.
1469          */
1470 #define REDUCE_FLS(a, b)                \
1471 do {                                    \
1472         if (a##_fls > b##_fls) {        \
1473                 a >>= 1;                \
1474                 a##_fls--;              \
1475         } else {                        \
1476                 b >>= 1;                \
1477                 b##_fls--;              \
1478         }                               \
1479 } while (0)
1480
1481         /*
1482          * Reduce accuracy until either term fits in a u64, then proceed with
1483          * the other, so that finally we can do a u64/u64 division.
1484          */
1485         while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1486                 REDUCE_FLS(nsec, frequency);
1487                 REDUCE_FLS(sec, count);
1488         }
1489
1490         if (count_fls + sec_fls > 64) {
1491                 divisor = nsec * frequency;
1492
1493                 while (count_fls + sec_fls > 64) {
1494                         REDUCE_FLS(count, sec);
1495                         divisor >>= 1;
1496                 }
1497
1498                 dividend = count * sec;
1499         } else {
1500                 dividend = count * sec;
1501
1502                 while (nsec_fls + frequency_fls > 64) {
1503                         REDUCE_FLS(nsec, frequency);
1504                         dividend >>= 1;
1505                 }
1506
1507                 divisor = nsec * frequency;
1508         }
1509
1510         return div64_u64(dividend, divisor);
1511 }
1512
1513 static void perf_event_stop(struct perf_event *event)
1514 {
1515         if (!event->pmu->stop)
1516                 return event->pmu->disable(event);
1517
1518         return event->pmu->stop(event);
1519 }
1520
1521 static int perf_event_start(struct perf_event *event)
1522 {
1523         if (!event->pmu->start)
1524                 return event->pmu->enable(event);
1525
1526         return event->pmu->start(event);
1527 }
1528
1529 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1530 {
1531         struct hw_perf_event *hwc = &event->hw;
1532         u64 period, sample_period;
1533         s64 delta;
1534
1535         period = perf_calculate_period(event, nsec, count);
1536
1537         delta = (s64)(period - hwc->sample_period);
1538         delta = (delta + 7) / 8; /* low pass filter */
1539
1540         sample_period = hwc->sample_period + delta;
1541
1542         if (!sample_period)
1543                 sample_period = 1;
1544
1545         hwc->sample_period = sample_period;
1546
1547         if (atomic64_read(&hwc->period_left) > 8*sample_period) {
1548                 perf_disable();
1549                 perf_event_stop(event);
1550                 atomic64_set(&hwc->period_left, 0);
1551                 perf_event_start(event);
1552                 perf_enable();
1553         }
1554 }
1555
1556 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1557 {
1558         struct perf_event *event;
1559         struct hw_perf_event *hwc;
1560         u64 interrupts, now;
1561         s64 delta;
1562
1563         raw_spin_lock(&ctx->lock);
1564         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1565                 if (event->state != PERF_EVENT_STATE_ACTIVE)
1566                         continue;
1567
1568                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1569                         continue;
1570
1571                 hwc = &event->hw;
1572
1573                 interrupts = hwc->interrupts;
1574                 hwc->interrupts = 0;
1575
1576                 /*
1577                  * unthrottle events on the tick
1578                  */
1579                 if (interrupts == MAX_INTERRUPTS) {
1580                         perf_log_throttle(event, 1);
1581                         perf_disable();
1582                         event->pmu->unthrottle(event);
1583                         perf_enable();
1584                 }
1585
1586                 if (!event->attr.freq || !event->attr.sample_freq)
1587                         continue;
1588
1589                 perf_disable();
1590                 event->pmu->read(event);
1591                 now = atomic64_read(&event->count);
1592                 delta = now - hwc->freq_count_stamp;
1593                 hwc->freq_count_stamp = now;
1594
1595                 if (delta > 0)
1596                         perf_adjust_period(event, TICK_NSEC, delta);
1597                 perf_enable();
1598         }
1599         raw_spin_unlock(&ctx->lock);
1600 }
1601
1602 /*
1603  * Round-robin a context's events:
1604  */
1605 static void rotate_ctx(struct perf_event_context *ctx)
1606 {
1607         raw_spin_lock(&ctx->lock);
1608
1609         /* Rotate the first entry last of non-pinned groups */
1610         list_rotate_left(&ctx->flexible_groups);
1611
1612         raw_spin_unlock(&ctx->lock);
1613 }
1614
1615 void perf_event_task_tick(struct task_struct *curr)
1616 {
1617         struct perf_cpu_context *cpuctx;
1618         struct perf_event_context *ctx;
1619         int rotate = 0;
1620
1621         if (!atomic_read(&nr_events))
1622                 return;
1623
1624         cpuctx = &__get_cpu_var(perf_cpu_context);
1625         if (cpuctx->ctx.nr_events &&
1626             cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1627                 rotate = 1;
1628
1629         ctx = curr->perf_event_ctxp;
1630         if (ctx && ctx->nr_events && ctx->nr_events != ctx->nr_active)
1631                 rotate = 1;
1632
1633         perf_ctx_adjust_freq(&cpuctx->ctx);
1634         if (ctx)
1635                 perf_ctx_adjust_freq(ctx);
1636
1637         if (!rotate)
1638                 return;
1639
1640         perf_disable();
1641         cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1642         if (ctx)
1643                 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1644
1645         rotate_ctx(&cpuctx->ctx);
1646         if (ctx)
1647                 rotate_ctx(ctx);
1648
1649         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1650         if (ctx)
1651                 task_ctx_sched_in(curr, EVENT_FLEXIBLE);
1652         perf_enable();
1653 }
1654
1655 static int event_enable_on_exec(struct perf_event *event,
1656                                 struct perf_event_context *ctx)
1657 {
1658         if (!event->attr.enable_on_exec)
1659                 return 0;
1660
1661         event->attr.enable_on_exec = 0;
1662         if (event->state >= PERF_EVENT_STATE_INACTIVE)
1663                 return 0;
1664
1665         __perf_event_mark_enabled(event, ctx);
1666
1667         return 1;
1668 }
1669
1670 /*
1671  * Enable all of a task's events that have been marked enable-on-exec.
1672  * This expects task == current.
1673  */
1674 static void perf_event_enable_on_exec(struct task_struct *task)
1675 {
1676         struct perf_event_context *ctx;
1677         struct perf_event *event;
1678         unsigned long flags;
1679         int enabled = 0;
1680         int ret;
1681
1682         local_irq_save(flags);
1683         ctx = task->perf_event_ctxp;
1684         if (!ctx || !ctx->nr_events)
1685                 goto out;
1686
1687         __perf_event_task_sched_out(ctx);
1688
1689         raw_spin_lock(&ctx->lock);
1690
1691         list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1692                 ret = event_enable_on_exec(event, ctx);
1693                 if (ret)
1694                         enabled = 1;
1695         }
1696
1697         list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1698                 ret = event_enable_on_exec(event, ctx);
1699                 if (ret)
1700                         enabled = 1;
1701         }
1702
1703         /*
1704          * Unclone this context if we enabled any event.
1705          */
1706         if (enabled)
1707                 unclone_ctx(ctx);
1708
1709         raw_spin_unlock(&ctx->lock);
1710
1711         perf_event_task_sched_in(task);
1712  out:
1713         local_irq_restore(flags);
1714 }
1715
1716 /*
1717  * Cross CPU call to read the hardware event
1718  */
1719 static void __perf_event_read(void *info)
1720 {
1721         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1722         struct perf_event *event = info;
1723         struct perf_event_context *ctx = event->ctx;
1724
1725         /*
1726          * If this is a task context, we need to check whether it is
1727          * the current task context of this cpu.  If not it has been
1728          * scheduled out before the smp call arrived.  In that case
1729          * event->count would have been updated to a recent sample
1730          * when the event was scheduled out.
1731          */
1732         if (ctx->task && cpuctx->task_ctx != ctx)
1733                 return;
1734
1735         raw_spin_lock(&ctx->lock);
1736         update_context_time(ctx);
1737         update_event_times(event);
1738         raw_spin_unlock(&ctx->lock);
1739
1740         event->pmu->read(event);
1741 }
1742
1743 static u64 perf_event_read(struct perf_event *event)
1744 {
1745         /*
1746          * If event is enabled and currently active on a CPU, update the
1747          * value in the event structure:
1748          */
1749         if (event->state == PERF_EVENT_STATE_ACTIVE) {
1750                 smp_call_function_single(event->oncpu,
1751                                          __perf_event_read, event, 1);
1752         } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1753                 struct perf_event_context *ctx = event->ctx;
1754                 unsigned long flags;
1755
1756                 raw_spin_lock_irqsave(&ctx->lock, flags);
1757                 update_context_time(ctx);
1758                 update_event_times(event);
1759                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1760         }
1761
1762         return atomic64_read(&event->count);
1763 }
1764
1765 /*
1766  * Initialize the perf_event context in a task_struct:
1767  */
1768 static void
1769 __perf_event_init_context(struct perf_event_context *ctx,
1770                             struct task_struct *task)
1771 {
1772         raw_spin_lock_init(&ctx->lock);
1773         mutex_init(&ctx->mutex);
1774         INIT_LIST_HEAD(&ctx->pinned_groups);
1775         INIT_LIST_HEAD(&ctx->flexible_groups);
1776         INIT_LIST_HEAD(&ctx->event_list);
1777         atomic_set(&ctx->refcount, 1);
1778         ctx->task = task;
1779 }
1780
1781 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1782 {
1783         struct perf_event_context *ctx;
1784         struct perf_cpu_context *cpuctx;
1785         struct task_struct *task;
1786         unsigned long flags;
1787         int err;
1788
1789         if (pid == -1 && cpu != -1) {
1790                 /* Must be root to operate on a CPU event: */
1791                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1792                         return ERR_PTR(-EACCES);
1793
1794                 if (cpu < 0 || cpu >= nr_cpumask_bits)
1795                         return ERR_PTR(-EINVAL);
1796
1797                 /*
1798                  * We could be clever and allow to attach a event to an
1799                  * offline CPU and activate it when the CPU comes up, but
1800                  * that's for later.
1801                  */
1802                 if (!cpu_online(cpu))
1803                         return ERR_PTR(-ENODEV);
1804
1805                 cpuctx = &per_cpu(perf_cpu_context, cpu);
1806                 ctx = &cpuctx->ctx;
1807                 get_ctx(ctx);
1808
1809                 return ctx;
1810         }
1811
1812         rcu_read_lock();
1813         if (!pid)
1814                 task = current;
1815         else
1816                 task = find_task_by_vpid(pid);
1817         if (task)
1818                 get_task_struct(task);
1819         rcu_read_unlock();
1820
1821         if (!task)
1822                 return ERR_PTR(-ESRCH);
1823
1824         /*
1825          * Can't attach events to a dying task.
1826          */
1827         err = -ESRCH;
1828         if (task->flags & PF_EXITING)
1829                 goto errout;
1830
1831         /* Reuse ptrace permission checks for now. */
1832         err = -EACCES;
1833         if (!ptrace_may_access(task, PTRACE_MODE_READ))
1834                 goto errout;
1835
1836  retry:
1837         ctx = perf_lock_task_context(task, &flags);
1838         if (ctx) {
1839                 unclone_ctx(ctx);
1840                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1841         }
1842
1843         if (!ctx) {
1844                 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1845                 err = -ENOMEM;
1846                 if (!ctx)
1847                         goto errout;
1848                 __perf_event_init_context(ctx, task);
1849                 get_ctx(ctx);
1850                 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1851                         /*
1852                          * We raced with some other task; use
1853                          * the context they set.
1854                          */
1855                         kfree(ctx);
1856                         goto retry;
1857                 }
1858                 get_task_struct(task);
1859         }
1860
1861         put_task_struct(task);
1862         return ctx;
1863
1864  errout:
1865         put_task_struct(task);
1866         return ERR_PTR(err);
1867 }
1868
1869 static void perf_event_free_filter(struct perf_event *event);
1870
1871 static void free_event_rcu(struct rcu_head *head)
1872 {
1873         struct perf_event *event;
1874
1875         event = container_of(head, struct perf_event, rcu_head);
1876         if (event->ns)
1877                 put_pid_ns(event->ns);
1878         perf_event_free_filter(event);
1879         kfree(event);
1880 }
1881
1882 static void perf_pending_sync(struct perf_event *event);
1883 static void perf_mmap_data_put(struct perf_mmap_data *data);
1884
1885 static void free_event(struct perf_event *event)
1886 {
1887         perf_pending_sync(event);
1888
1889         if (!event->parent) {
1890                 atomic_dec(&nr_events);
1891                 if (event->attr.mmap)
1892                         atomic_dec(&nr_mmap_events);
1893                 if (event->attr.comm)
1894                         atomic_dec(&nr_comm_events);
1895                 if (event->attr.task)
1896                         atomic_dec(&nr_task_events);
1897         }
1898
1899         if (event->data) {
1900                 perf_mmap_data_put(event->data);
1901                 event->data = NULL;
1902         }
1903
1904         if (event->destroy)
1905                 event->destroy(event);
1906
1907         put_ctx(event->ctx);
1908         call_rcu(&event->rcu_head, free_event_rcu);
1909 }
1910
1911 int perf_event_release_kernel(struct perf_event *event)
1912 {
1913         struct perf_event_context *ctx = event->ctx;
1914
1915         /*
1916          * Remove from the PMU, can't get re-enabled since we got
1917          * here because the last ref went.
1918          */
1919         perf_event_disable(event);
1920
1921         WARN_ON_ONCE(ctx->parent_ctx);
1922         /*
1923          * There are two ways this annotation is useful:
1924          *
1925          *  1) there is a lock recursion from perf_event_exit_task
1926          *     see the comment there.
1927          *
1928          *  2) there is a lock-inversion with mmap_sem through
1929          *     perf_event_read_group(), which takes faults while
1930          *     holding ctx->mutex, however this is called after
1931          *     the last filedesc died, so there is no possibility
1932          *     to trigger the AB-BA case.
1933          */
1934         mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
1935         raw_spin_lock_irq(&ctx->lock);
1936         perf_group_detach(event);
1937         list_del_event(event, ctx);
1938         raw_spin_unlock_irq(&ctx->lock);
1939         mutex_unlock(&ctx->mutex);
1940
1941         mutex_lock(&event->owner->perf_event_mutex);
1942         list_del_init(&event->owner_entry);
1943         mutex_unlock(&event->owner->perf_event_mutex);
1944         put_task_struct(event->owner);
1945
1946         free_event(event);
1947
1948         return 0;
1949 }
1950 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1951
1952 /*
1953  * Called when the last reference to the file is gone.
1954  */
1955 static int perf_release(struct inode *inode, struct file *file)
1956 {
1957         struct perf_event *event = file->private_data;
1958
1959         file->private_data = NULL;
1960
1961         return perf_event_release_kernel(event);
1962 }
1963
1964 static int perf_event_read_size(struct perf_event *event)
1965 {
1966         int entry = sizeof(u64); /* value */
1967         int size = 0;
1968         int nr = 1;
1969
1970         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1971                 size += sizeof(u64);
1972
1973         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1974                 size += sizeof(u64);
1975
1976         if (event->attr.read_format & PERF_FORMAT_ID)
1977                 entry += sizeof(u64);
1978
1979         if (event->attr.read_format & PERF_FORMAT_GROUP) {
1980                 nr += event->group_leader->nr_siblings;
1981                 size += sizeof(u64);
1982         }
1983
1984         size += entry * nr;
1985
1986         return size;
1987 }
1988
1989 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1990 {
1991         struct perf_event *child;
1992         u64 total = 0;
1993
1994         *enabled = 0;
1995         *running = 0;
1996
1997         mutex_lock(&event->child_mutex);
1998         total += perf_event_read(event);
1999         *enabled += event->total_time_enabled +
2000                         atomic64_read(&event->child_total_time_enabled);
2001         *running += event->total_time_running +
2002                         atomic64_read(&event->child_total_time_running);
2003
2004         list_for_each_entry(child, &event->child_list, child_list) {
2005                 total += perf_event_read(child);
2006                 *enabled += child->total_time_enabled;
2007                 *running += child->total_time_running;
2008         }
2009         mutex_unlock(&event->child_mutex);
2010
2011         return total;
2012 }
2013 EXPORT_SYMBOL_GPL(perf_event_read_value);
2014
2015 static int perf_event_read_group(struct perf_event *event,
2016                                    u64 read_format, char __user *buf)
2017 {
2018         struct perf_event *leader = event->group_leader, *sub;
2019         int n = 0, size = 0, ret = -EFAULT;
2020         struct perf_event_context *ctx = leader->ctx;
2021         u64 values[5];
2022         u64 count, enabled, running;
2023
2024         mutex_lock(&ctx->mutex);
2025         count = perf_event_read_value(leader, &enabled, &running);
2026
2027         values[n++] = 1 + leader->nr_siblings;
2028         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2029                 values[n++] = enabled;
2030         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2031                 values[n++] = running;
2032         values[n++] = count;
2033         if (read_format & PERF_FORMAT_ID)
2034                 values[n++] = primary_event_id(leader);
2035
2036         size = n * sizeof(u64);
2037
2038         if (copy_to_user(buf, values, size))
2039                 goto unlock;
2040
2041         ret = size;
2042
2043         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2044                 n = 0;
2045
2046                 values[n++] = perf_event_read_value(sub, &enabled, &running);
2047                 if (read_format & PERF_FORMAT_ID)
2048                         values[n++] = primary_event_id(sub);
2049
2050                 size = n * sizeof(u64);
2051
2052                 if (copy_to_user(buf + ret, values, size)) {
2053                         ret = -EFAULT;
2054                         goto unlock;
2055                 }
2056
2057                 ret += size;
2058         }
2059 unlock:
2060         mutex_unlock(&ctx->mutex);
2061
2062         return ret;
2063 }
2064
2065 static int perf_event_read_one(struct perf_event *event,
2066                                  u64 read_format, char __user *buf)
2067 {
2068         u64 enabled, running;
2069         u64 values[4];
2070         int n = 0;
2071
2072         values[n++] = perf_event_read_value(event, &enabled, &running);
2073         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2074                 values[n++] = enabled;
2075         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2076                 values[n++] = running;
2077         if (read_format & PERF_FORMAT_ID)
2078                 values[n++] = primary_event_id(event);
2079
2080         if (copy_to_user(buf, values, n * sizeof(u64)))
2081                 return -EFAULT;
2082
2083         return n * sizeof(u64);
2084 }
2085
2086 /*
2087  * Read the performance event - simple non blocking version for now
2088  */
2089 static ssize_t
2090 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2091 {
2092         u64 read_format = event->attr.read_format;
2093         int ret;
2094
2095         /*
2096          * Return end-of-file for a read on a event that is in
2097          * error state (i.e. because it was pinned but it couldn't be
2098          * scheduled on to the CPU at some point).
2099          */
2100         if (event->state == PERF_EVENT_STATE_ERROR)
2101                 return 0;
2102
2103         if (count < perf_event_read_size(event))
2104                 return -ENOSPC;
2105
2106         WARN_ON_ONCE(event->ctx->parent_ctx);
2107         if (read_format & PERF_FORMAT_GROUP)
2108                 ret = perf_event_read_group(event, read_format, buf);
2109         else
2110                 ret = perf_event_read_one(event, read_format, buf);
2111
2112         return ret;
2113 }
2114
2115 static ssize_t
2116 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2117 {
2118         struct perf_event *event = file->private_data;
2119
2120         return perf_read_hw(event, buf, count);
2121 }
2122
2123 static unsigned int perf_poll(struct file *file, poll_table *wait)
2124 {
2125         struct perf_event *event = file->private_data;
2126         struct perf_mmap_data *data;
2127         unsigned int events = POLL_HUP;
2128
2129         rcu_read_lock();
2130         data = rcu_dereference(event->data);
2131         if (data)
2132                 events = atomic_xchg(&data->poll, 0);
2133         rcu_read_unlock();
2134
2135         poll_wait(file, &event->waitq, wait);
2136
2137         return events;
2138 }
2139
2140 static void perf_event_reset(struct perf_event *event)
2141 {
2142         (void)perf_event_read(event);
2143         atomic64_set(&event->count, 0);
2144         perf_event_update_userpage(event);
2145 }
2146
2147 /*
2148  * Holding the top-level event's child_mutex means that any
2149  * descendant process that has inherited this event will block
2150  * in sync_child_event if it goes to exit, thus satisfying the
2151  * task existence requirements of perf_event_enable/disable.
2152  */
2153 static void perf_event_for_each_child(struct perf_event *event,
2154                                         void (*func)(struct perf_event *))
2155 {
2156         struct perf_event *child;
2157
2158         WARN_ON_ONCE(event->ctx->parent_ctx);
2159         mutex_lock(&event->child_mutex);
2160         func(event);
2161         list_for_each_entry(child, &event->child_list, child_list)
2162                 func(child);
2163         mutex_unlock(&event->child_mutex);
2164 }
2165
2166 static void perf_event_for_each(struct perf_event *event,
2167                                   void (*func)(struct perf_event *))
2168 {
2169         struct perf_event_context *ctx = event->ctx;
2170         struct perf_event *sibling;
2171
2172         WARN_ON_ONCE(ctx->parent_ctx);
2173         mutex_lock(&ctx->mutex);
2174         event = event->group_leader;
2175
2176         perf_event_for_each_child(event, func);
2177         func(event);
2178         list_for_each_entry(sibling, &event->sibling_list, group_entry)
2179                 perf_event_for_each_child(event, func);
2180         mutex_unlock(&ctx->mutex);
2181 }
2182
2183 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2184 {
2185         struct perf_event_context *ctx = event->ctx;
2186         unsigned long size;
2187         int ret = 0;
2188         u64 value;
2189
2190         if (!event->attr.sample_period)
2191                 return -EINVAL;
2192
2193         size = copy_from_user(&value, arg, sizeof(value));
2194         if (size != sizeof(value))
2195                 return -EFAULT;
2196
2197         if (!value)
2198                 return -EINVAL;
2199
2200         raw_spin_lock_irq(&ctx->lock);
2201         if (event->attr.freq) {
2202                 if (value > sysctl_perf_event_sample_rate) {
2203                         ret = -EINVAL;
2204                         goto unlock;
2205                 }
2206
2207                 event->attr.sample_freq = value;
2208         } else {
2209                 event->attr.sample_period = value;
2210                 event->hw.sample_period = value;
2211         }
2212 unlock:
2213         raw_spin_unlock_irq(&ctx->lock);
2214
2215         return ret;
2216 }
2217
2218 static const struct file_operations perf_fops;
2219
2220 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
2221 {
2222         struct file *file;
2223
2224         file = fget_light(fd, fput_needed);
2225         if (!file)
2226                 return ERR_PTR(-EBADF);
2227
2228         if (file->f_op != &perf_fops) {
2229                 fput_light(file, *fput_needed);
2230                 *fput_needed = 0;
2231                 return ERR_PTR(-EBADF);
2232         }
2233
2234         return file->private_data;
2235 }
2236
2237 static int perf_event_set_output(struct perf_event *event,
2238                                  struct perf_event *output_event);
2239 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2240
2241 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2242 {
2243         struct perf_event *event = file->private_data;
2244         void (*func)(struct perf_event *);
2245         u32 flags = arg;
2246
2247         switch (cmd) {
2248         case PERF_EVENT_IOC_ENABLE:
2249                 func = perf_event_enable;
2250                 break;
2251         case PERF_EVENT_IOC_DISABLE:
2252                 func = perf_event_disable;
2253                 break;
2254         case PERF_EVENT_IOC_RESET:
2255                 func = perf_event_reset;
2256                 break;
2257
2258         case PERF_EVENT_IOC_REFRESH:
2259                 return perf_event_refresh(event, arg);
2260
2261         case PERF_EVENT_IOC_PERIOD:
2262                 return perf_event_period(event, (u64 __user *)arg);
2263
2264         case PERF_EVENT_IOC_SET_OUTPUT:
2265         {
2266                 struct perf_event *output_event = NULL;
2267                 int fput_needed = 0;
2268                 int ret;
2269
2270                 if (arg != -1) {
2271                         output_event = perf_fget_light(arg, &fput_needed);
2272                         if (IS_ERR(output_event))
2273                                 return PTR_ERR(output_event);
2274                 }
2275
2276                 ret = perf_event_set_output(event, output_event);
2277                 if (output_event)
2278                         fput_light(output_event->filp, fput_needed);
2279
2280                 return ret;
2281         }
2282
2283         case PERF_EVENT_IOC_SET_FILTER:
2284                 return perf_event_set_filter(event, (void __user *)arg);
2285
2286         default:
2287                 return -ENOTTY;
2288         }
2289
2290         if (flags & PERF_IOC_FLAG_GROUP)
2291                 perf_event_for_each(event, func);
2292         else
2293                 perf_event_for_each_child(event, func);
2294
2295         return 0;
2296 }
2297
2298 int perf_event_task_enable(void)
2299 {
2300         struct perf_event *event;
2301
2302         mutex_lock(&current->perf_event_mutex);
2303         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2304                 perf_event_for_each_child(event, perf_event_enable);
2305         mutex_unlock(&current->perf_event_mutex);
2306
2307         return 0;
2308 }
2309
2310 int perf_event_task_disable(void)
2311 {
2312         struct perf_event *event;
2313
2314         mutex_lock(&current->perf_event_mutex);
2315         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2316                 perf_event_for_each_child(event, perf_event_disable);
2317         mutex_unlock(&current->perf_event_mutex);
2318
2319         return 0;
2320 }
2321
2322 #ifndef PERF_EVENT_INDEX_OFFSET
2323 # define PERF_EVENT_INDEX_OFFSET 0
2324 #endif
2325
2326 static int perf_event_index(struct perf_event *event)
2327 {
2328         if (event->state != PERF_EVENT_STATE_ACTIVE)
2329                 return 0;
2330
2331         return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2332 }
2333
2334 /*
2335  * Callers need to ensure there can be no nesting of this function, otherwise
2336  * the seqlock logic goes bad. We can not serialize this because the arch
2337  * code calls this from NMI context.
2338  */
2339 void perf_event_update_userpage(struct perf_event *event)
2340 {
2341         struct perf_event_mmap_page *userpg;
2342         struct perf_mmap_data *data;
2343
2344         rcu_read_lock();
2345         data = rcu_dereference(event->data);
2346         if (!data)
2347                 goto unlock;
2348
2349         userpg = data->user_page;
2350
2351         /*
2352          * Disable preemption so as to not let the corresponding user-space
2353          * spin too long if we get preempted.
2354          */
2355         preempt_disable();
2356         ++userpg->lock;
2357         barrier();
2358         userpg->index = perf_event_index(event);
2359         userpg->offset = atomic64_read(&event->count);
2360         if (event->state == PERF_EVENT_STATE_ACTIVE)
2361                 userpg->offset -= atomic64_read(&event->hw.prev_count);
2362
2363         userpg->time_enabled = event->total_time_enabled +
2364                         atomic64_read(&event->child_total_time_enabled);
2365
2366         userpg->time_running = event->total_time_running +
2367                         atomic64_read(&event->child_total_time_running);
2368
2369         barrier();
2370         ++userpg->lock;
2371         preempt_enable();
2372 unlock:
2373         rcu_read_unlock();
2374 }
2375
2376 #ifndef CONFIG_PERF_USE_VMALLOC
2377
2378 /*
2379  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2380  */
2381
2382 static struct page *
2383 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2384 {
2385         if (pgoff > data->nr_pages)
2386                 return NULL;
2387
2388         if (pgoff == 0)
2389                 return virt_to_page(data->user_page);
2390
2391         return virt_to_page(data->data_pages[pgoff - 1]);
2392 }
2393
2394 static void *perf_mmap_alloc_page(int cpu)
2395 {
2396         struct page *page;
2397         int node;
2398
2399         node = (cpu == -1) ? cpu : cpu_to_node(cpu);
2400         page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
2401         if (!page)
2402                 return NULL;
2403
2404         return page_address(page);
2405 }
2406
2407 static struct perf_mmap_data *
2408 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2409 {
2410         struct perf_mmap_data *data;
2411         unsigned long size;
2412         int i;
2413
2414         size = sizeof(struct perf_mmap_data);
2415         size += nr_pages * sizeof(void *);
2416
2417         data = kzalloc(size, GFP_KERNEL);
2418         if (!data)
2419                 goto fail;
2420
2421         data->user_page = perf_mmap_alloc_page(event->cpu);
2422         if (!data->user_page)
2423                 goto fail_user_page;
2424
2425         for (i = 0; i < nr_pages; i++) {
2426                 data->data_pages[i] = perf_mmap_alloc_page(event->cpu);
2427                 if (!data->data_pages[i])
2428                         goto fail_data_pages;
2429         }
2430
2431         data->nr_pages = nr_pages;
2432
2433         return data;
2434
2435 fail_data_pages:
2436         for (i--; i >= 0; i--)
2437                 free_page((unsigned long)data->data_pages[i]);
2438
2439         free_page((unsigned long)data->user_page);
2440
2441 fail_user_page:
2442         kfree(data);
2443
2444 fail:
2445         return NULL;
2446 }
2447
2448 static void perf_mmap_free_page(unsigned long addr)
2449 {
2450         struct page *page = virt_to_page((void *)addr);
2451
2452         page->mapping = NULL;
2453         __free_page(page);
2454 }
2455
2456 static void perf_mmap_data_free(struct perf_mmap_data *data)
2457 {
2458         int i;
2459
2460         perf_mmap_free_page((unsigned long)data->user_page);
2461         for (i = 0; i < data->nr_pages; i++)
2462                 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2463         kfree(data);
2464 }
2465
2466 static inline int page_order(struct perf_mmap_data *data)
2467 {
2468         return 0;
2469 }
2470
2471 #else
2472
2473 /*
2474  * Back perf_mmap() with vmalloc memory.
2475  *
2476  * Required for architectures that have d-cache aliasing issues.
2477  */
2478
2479 static inline int page_order(struct perf_mmap_data *data)
2480 {
2481         return data->page_order;
2482 }
2483
2484 static struct page *
2485 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2486 {
2487         if (pgoff > (1UL << page_order(data)))
2488                 return NULL;
2489
2490         return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2491 }
2492
2493 static void perf_mmap_unmark_page(void *addr)
2494 {
2495         struct page *page = vmalloc_to_page(addr);
2496
2497         page->mapping = NULL;
2498 }
2499
2500 static void perf_mmap_data_free_work(struct work_struct *work)
2501 {
2502         struct perf_mmap_data *data;
2503         void *base;
2504         int i, nr;
2505
2506         data = container_of(work, struct perf_mmap_data, work);
2507         nr = 1 << page_order(data);
2508
2509         base = data->user_page;
2510         for (i = 0; i < nr + 1; i++)
2511                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2512
2513         vfree(base);
2514         kfree(data);
2515 }
2516
2517 static void perf_mmap_data_free(struct perf_mmap_data *data)
2518 {
2519         schedule_work(&data->work);
2520 }
2521
2522 static struct perf_mmap_data *
2523 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2524 {
2525         struct perf_mmap_data *data;
2526         unsigned long size;
2527         void *all_buf;
2528
2529         size = sizeof(struct perf_mmap_data);
2530         size += sizeof(void *);
2531
2532         data = kzalloc(size, GFP_KERNEL);
2533         if (!data)
2534                 goto fail;
2535
2536         INIT_WORK(&data->work, perf_mmap_data_free_work);
2537
2538         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2539         if (!all_buf)
2540                 goto fail_all_buf;
2541
2542         data->user_page = all_buf;
2543         data->data_pages[0] = all_buf + PAGE_SIZE;
2544         data->page_order = ilog2(nr_pages);
2545         data->nr_pages = 1;
2546
2547         return data;
2548
2549 fail_all_buf:
2550         kfree(data);
2551
2552 fail:
2553         return NULL;
2554 }
2555
2556 #endif
2557
2558 static unsigned long perf_data_size(struct perf_mmap_data *data)
2559 {
2560         return data->nr_pages << (PAGE_SHIFT + page_order(data));
2561 }
2562
2563 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2564 {
2565         struct perf_event *event = vma->vm_file->private_data;
2566         struct perf_mmap_data *data;
2567         int ret = VM_FAULT_SIGBUS;
2568
2569         if (vmf->flags & FAULT_FLAG_MKWRITE) {
2570                 if (vmf->pgoff == 0)
2571                         ret = 0;
2572                 return ret;
2573         }
2574
2575         rcu_read_lock();
2576         data = rcu_dereference(event->data);
2577         if (!data)
2578                 goto unlock;
2579
2580         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2581                 goto unlock;
2582
2583         vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2584         if (!vmf->page)
2585                 goto unlock;
2586
2587         get_page(vmf->page);
2588         vmf->page->mapping = vma->vm_file->f_mapping;
2589         vmf->page->index   = vmf->pgoff;
2590
2591         ret = 0;
2592 unlock:
2593         rcu_read_unlock();
2594
2595         return ret;
2596 }
2597
2598 static void
2599 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2600 {
2601         long max_size = perf_data_size(data);
2602
2603         if (event->attr.watermark) {
2604                 data->watermark = min_t(long, max_size,
2605                                         event->attr.wakeup_watermark);
2606         }
2607
2608         if (!data->watermark)
2609                 data->watermark = max_size / 2;
2610
2611         atomic_set(&data->refcount, 1);
2612         rcu_assign_pointer(event->data, data);
2613 }
2614
2615 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2616 {
2617         struct perf_mmap_data *data;
2618
2619         data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2620         perf_mmap_data_free(data);
2621 }
2622
2623 static struct perf_mmap_data *perf_mmap_data_get(struct perf_event *event)
2624 {
2625         struct perf_mmap_data *data;
2626
2627         rcu_read_lock();
2628         data = rcu_dereference(event->data);
2629         if (data) {
2630                 if (!atomic_inc_not_zero(&data->refcount))
2631                         data = NULL;
2632         }
2633         rcu_read_unlock();
2634
2635         return data;
2636 }
2637
2638 static void perf_mmap_data_put(struct perf_mmap_data *data)
2639 {
2640         if (!atomic_dec_and_test(&data->refcount))
2641                 return;
2642
2643         call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2644 }
2645
2646 static void perf_mmap_open(struct vm_area_struct *vma)
2647 {
2648         struct perf_event *event = vma->vm_file->private_data;
2649
2650         atomic_inc(&event->mmap_count);
2651 }
2652
2653 static void perf_mmap_close(struct vm_area_struct *vma)
2654 {
2655         struct perf_event *event = vma->vm_file->private_data;
2656
2657         if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2658                 unsigned long size = perf_data_size(event->data);
2659                 struct user_struct *user = event->mmap_user;
2660                 struct perf_mmap_data *data = event->data;
2661
2662                 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2663                 vma->vm_mm->locked_vm -= event->mmap_locked;
2664                 rcu_assign_pointer(event->data, NULL);
2665                 mutex_unlock(&event->mmap_mutex);
2666
2667                 perf_mmap_data_put(data);
2668                 free_uid(user);
2669         }
2670 }
2671
2672 static const struct vm_operations_struct perf_mmap_vmops = {
2673         .open           = perf_mmap_open,
2674         .close          = perf_mmap_close,
2675         .fault          = perf_mmap_fault,
2676         .page_mkwrite   = perf_mmap_fault,
2677 };
2678
2679 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2680 {
2681         struct perf_event *event = file->private_data;
2682         unsigned long user_locked, user_lock_limit;
2683         struct user_struct *user = current_user();
2684         unsigned long locked, lock_limit;
2685         struct perf_mmap_data *data;
2686         unsigned long vma_size;
2687         unsigned long nr_pages;
2688         long user_extra, extra;
2689         int ret = 0;
2690
2691         /*
2692          * Don't allow mmap() of inherited per-task counters. This would
2693          * create a performance issue due to all children writing to the
2694          * same buffer.
2695          */
2696         if (event->cpu == -1 && event->attr.inherit)
2697                 return -EINVAL;
2698
2699         if (!(vma->vm_flags & VM_SHARED))
2700                 return -EINVAL;
2701
2702         vma_size = vma->vm_end - vma->vm_start;
2703         nr_pages = (vma_size / PAGE_SIZE) - 1;
2704
2705         /*
2706          * If we have data pages ensure they're a power-of-two number, so we
2707          * can do bitmasks instead of modulo.
2708          */
2709         if (nr_pages != 0 && !is_power_of_2(nr_pages))
2710                 return -EINVAL;
2711
2712         if (vma_size != PAGE_SIZE * (1 + nr_pages))
2713                 return -EINVAL;
2714
2715         if (vma->vm_pgoff != 0)
2716                 return -EINVAL;
2717
2718         WARN_ON_ONCE(event->ctx->parent_ctx);
2719         mutex_lock(&event->mmap_mutex);
2720         if (event->data) {
2721                 if (event->data->nr_pages == nr_pages)
2722                         atomic_inc(&event->data->refcount);
2723                 else
2724                         ret = -EINVAL;
2725                 goto unlock;
2726         }
2727
2728         user_extra = nr_pages + 1;
2729         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2730
2731         /*
2732          * Increase the limit linearly with more CPUs:
2733          */
2734         user_lock_limit *= num_online_cpus();
2735
2736         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2737
2738         extra = 0;
2739         if (user_locked > user_lock_limit)
2740                 extra = user_locked - user_lock_limit;
2741
2742         lock_limit = rlimit(RLIMIT_MEMLOCK);
2743         lock_limit >>= PAGE_SHIFT;
2744         locked = vma->vm_mm->locked_vm + extra;
2745
2746         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2747                 !capable(CAP_IPC_LOCK)) {
2748                 ret = -EPERM;
2749                 goto unlock;
2750         }
2751
2752         WARN_ON(event->data);
2753
2754         data = perf_mmap_data_alloc(event, nr_pages);
2755         if (!data) {
2756                 ret = -ENOMEM;
2757                 goto unlock;
2758         }
2759
2760         perf_mmap_data_init(event, data);
2761         if (vma->vm_flags & VM_WRITE)
2762                 event->data->writable = 1;
2763
2764         atomic_long_add(user_extra, &user->locked_vm);
2765         event->mmap_locked = extra;
2766         event->mmap_user = get_current_user();
2767         vma->vm_mm->locked_vm += event->mmap_locked;
2768
2769 unlock:
2770         if (!ret)
2771                 atomic_inc(&event->mmap_count);
2772         mutex_unlock(&event->mmap_mutex);
2773
2774         vma->vm_flags |= VM_RESERVED;
2775         vma->vm_ops = &perf_mmap_vmops;
2776
2777         return ret;
2778 }
2779
2780 static int perf_fasync(int fd, struct file *filp, int on)
2781 {
2782         struct inode *inode = filp->f_path.dentry->d_inode;
2783         struct perf_event *event = filp->private_data;
2784         int retval;
2785
2786         mutex_lock(&inode->i_mutex);
2787         retval = fasync_helper(fd, filp, on, &event->fasync);
2788         mutex_unlock(&inode->i_mutex);
2789
2790         if (retval < 0)
2791                 return retval;
2792
2793         return 0;
2794 }
2795
2796 static const struct file_operations perf_fops = {
2797         .llseek                 = no_llseek,
2798         .release                = perf_release,
2799         .read                   = perf_read,
2800         .poll                   = perf_poll,
2801         .unlocked_ioctl         = perf_ioctl,
2802         .compat_ioctl           = perf_ioctl,
2803         .mmap                   = perf_mmap,
2804         .fasync                 = perf_fasync,
2805 };
2806
2807 /*
2808  * Perf event wakeup
2809  *
2810  * If there's data, ensure we set the poll() state and publish everything
2811  * to user-space before waking everybody up.
2812  */
2813
2814 void perf_event_wakeup(struct perf_event *event)
2815 {
2816         wake_up_all(&event->waitq);
2817
2818         if (event->pending_kill) {
2819                 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2820                 event->pending_kill = 0;
2821         }
2822 }
2823
2824 /*
2825  * Pending wakeups
2826  *
2827  * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2828  *
2829  * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2830  * single linked list and use cmpxchg() to add entries lockless.
2831  */
2832
2833 static void perf_pending_event(struct perf_pending_entry *entry)
2834 {
2835         struct perf_event *event = container_of(entry,
2836                         struct perf_event, pending);
2837
2838         if (event->pending_disable) {
2839                 event->pending_disable = 0;
2840                 __perf_event_disable(event);
2841         }
2842
2843         if (event->pending_wakeup) {
2844                 event->pending_wakeup = 0;
2845                 perf_event_wakeup(event);
2846         }
2847 }
2848
2849 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2850
2851 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2852         PENDING_TAIL,
2853 };
2854
2855 static void perf_pending_queue(struct perf_pending_entry *entry,
2856                                void (*func)(struct perf_pending_entry *))
2857 {
2858         struct perf_pending_entry **head;
2859
2860         if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2861                 return;
2862
2863         entry->func = func;
2864
2865         head = &get_cpu_var(perf_pending_head);
2866
2867         do {
2868                 entry->next = *head;
2869         } while (cmpxchg(head, entry->next, entry) != entry->next);
2870
2871         set_perf_event_pending();
2872
2873         put_cpu_var(perf_pending_head);
2874 }
2875
2876 static int __perf_pending_run(void)
2877 {
2878         struct perf_pending_entry *list;
2879         int nr = 0;
2880
2881         list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2882         while (list != PENDING_TAIL) {
2883                 void (*func)(struct perf_pending_entry *);
2884                 struct perf_pending_entry *entry = list;
2885
2886                 list = list->next;
2887
2888                 func = entry->func;
2889                 entry->next = NULL;
2890                 /*
2891                  * Ensure we observe the unqueue before we issue the wakeup,
2892                  * so that we won't be waiting forever.
2893                  * -- see perf_not_pending().
2894                  */
2895                 smp_wmb();
2896
2897                 func(entry);
2898                 nr++;
2899         }
2900
2901         return nr;
2902 }
2903
2904 static inline int perf_not_pending(struct perf_event *event)
2905 {
2906         /*
2907          * If we flush on whatever cpu we run, there is a chance we don't
2908          * need to wait.
2909          */
2910         get_cpu();
2911         __perf_pending_run();
2912         put_cpu();
2913
2914         /*
2915          * Ensure we see the proper queue state before going to sleep
2916          * so that we do not miss the wakeup. -- see perf_pending_handle()
2917          */
2918         smp_rmb();
2919         return event->pending.next == NULL;
2920 }
2921
2922 static void perf_pending_sync(struct perf_event *event)
2923 {
2924         wait_event(event->waitq, perf_not_pending(event));
2925 }
2926
2927 void perf_event_do_pending(void)
2928 {
2929         __perf_pending_run();
2930 }
2931
2932 /*
2933  * Callchain support -- arch specific
2934  */
2935
2936 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2937 {
2938         return NULL;
2939 }
2940
2941 __weak
2942 void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip, int skip)
2943 {
2944 }
2945
2946
2947 /*
2948  * We assume there is only KVM supporting the callbacks.
2949  * Later on, we might change it to a list if there is
2950  * another virtualization implementation supporting the callbacks.
2951  */
2952 struct perf_guest_info_callbacks *perf_guest_cbs;
2953
2954 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2955 {
2956         perf_guest_cbs = cbs;
2957         return 0;
2958 }
2959 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
2960
2961 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2962 {
2963         perf_guest_cbs = NULL;
2964         return 0;
2965 }
2966 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
2967
2968 /*
2969  * Output
2970  */
2971 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2972                               unsigned long offset, unsigned long head)
2973 {
2974         unsigned long mask;
2975
2976         if (!data->writable)
2977                 return true;
2978
2979         mask = perf_data_size(data) - 1;
2980
2981         offset = (offset - tail) & mask;
2982         head   = (head   - tail) & mask;
2983
2984         if ((int)(head - offset) < 0)
2985                 return false;
2986
2987         return true;
2988 }
2989
2990 static void perf_output_wakeup(struct perf_output_handle *handle)
2991 {
2992         atomic_set(&handle->data->poll, POLL_IN);
2993
2994         if (handle->nmi) {
2995                 handle->event->pending_wakeup = 1;
2996                 perf_pending_queue(&handle->event->pending,
2997                                    perf_pending_event);
2998         } else
2999                 perf_event_wakeup(handle->event);
3000 }
3001
3002 /*
3003  * We need to ensure a later event_id doesn't publish a head when a former
3004  * event isn't done writing. However since we need to deal with NMIs we
3005  * cannot fully serialize things.
3006  *
3007  * We only publish the head (and generate a wakeup) when the outer-most
3008  * event completes.
3009  */
3010 static void perf_output_get_handle(struct perf_output_handle *handle)
3011 {
3012         struct perf_mmap_data *data = handle->data;
3013
3014         preempt_disable();
3015         local_inc(&data->nest);
3016         handle->wakeup = local_read(&data->wakeup);
3017 }
3018
3019 static void perf_output_put_handle(struct perf_output_handle *handle)
3020 {
3021         struct perf_mmap_data *data = handle->data;
3022         unsigned long head;
3023
3024 again:
3025         head = local_read(&data->head);
3026
3027         /*
3028          * IRQ/NMI can happen here, which means we can miss a head update.
3029          */
3030
3031         if (!local_dec_and_test(&data->nest))
3032                 goto out;
3033
3034         /*
3035          * Publish the known good head. Rely on the full barrier implied
3036          * by atomic_dec_and_test() order the data->head read and this
3037          * write.
3038          */
3039         data->user_page->data_head = head;
3040
3041         /*
3042          * Now check if we missed an update, rely on the (compiler)
3043          * barrier in atomic_dec_and_test() to re-read data->head.
3044          */
3045         if (unlikely(head != local_read(&data->head))) {
3046                 local_inc(&data->nest);
3047                 goto again;
3048         }
3049
3050         if (handle->wakeup != local_read(&data->wakeup))
3051                 perf_output_wakeup(handle);
3052
3053  out:
3054         preempt_enable();
3055 }
3056
3057 __always_inline void perf_output_copy(struct perf_output_handle *handle,
3058                       const void *buf, unsigned int len)
3059 {
3060         do {
3061                 unsigned long size = min_t(unsigned long, handle->size, len);
3062
3063                 memcpy(handle->addr, buf, size);
3064
3065                 len -= size;
3066                 handle->addr += size;
3067                 buf += size;
3068                 handle->size -= size;
3069                 if (!handle->size) {
3070                         struct perf_mmap_data *data = handle->data;
3071
3072                         handle->page++;
3073                         handle->page &= data->nr_pages - 1;
3074                         handle->addr = data->data_pages[handle->page];
3075                         handle->size = PAGE_SIZE << page_order(data);
3076                 }
3077         } while (len);
3078 }
3079
3080 int perf_output_begin(struct perf_output_handle *handle,
3081                       struct perf_event *event, unsigned int size,
3082                       int nmi, int sample)
3083 {
3084         struct perf_mmap_data *data;
3085         unsigned long tail, offset, head;
3086         int have_lost;
3087         struct {
3088                 struct perf_event_header header;
3089                 u64                      id;
3090                 u64                      lost;
3091         } lost_event;
3092
3093         rcu_read_lock();
3094         /*
3095          * For inherited events we send all the output towards the parent.
3096          */
3097         if (event->parent)
3098                 event = event->parent;
3099
3100         data = rcu_dereference(event->data);
3101         if (!data)
3102                 goto out;
3103
3104         handle->data    = data;
3105         handle->event   = event;
3106         handle->nmi     = nmi;
3107         handle->sample  = sample;
3108
3109         if (!data->nr_pages)
3110                 goto out;
3111
3112         have_lost = local_read(&data->lost);
3113         if (have_lost)
3114                 size += sizeof(lost_event);
3115
3116         perf_output_get_handle(handle);
3117
3118         do {
3119                 /*
3120                  * Userspace could choose to issue a mb() before updating the
3121                  * tail pointer. So that all reads will be completed before the
3122                  * write is issued.
3123                  */
3124                 tail = ACCESS_ONCE(data->user_page->data_tail);
3125                 smp_rmb();
3126                 offset = head = local_read(&data->head);
3127                 head += size;
3128                 if (unlikely(!perf_output_space(data, tail, offset, head)))
3129                         goto fail;
3130         } while (local_cmpxchg(&data->head, offset, head) != offset);
3131
3132         if (head - local_read(&data->wakeup) > data->watermark)
3133                 local_add(data->watermark, &data->wakeup);
3134
3135         handle->page = offset >> (PAGE_SHIFT + page_order(data));
3136         handle->page &= data->nr_pages - 1;
3137         handle->size = offset & ((PAGE_SIZE << page_order(data)) - 1);
3138         handle->addr = data->data_pages[handle->page];
3139         handle->addr += handle->size;
3140         handle->size = (PAGE_SIZE << page_order(data)) - handle->size;
3141
3142         if (have_lost) {
3143                 lost_event.header.type = PERF_RECORD_LOST;
3144                 lost_event.header.misc = 0;
3145                 lost_event.header.size = sizeof(lost_event);
3146                 lost_event.id          = event->id;
3147                 lost_event.lost        = local_xchg(&data->lost, 0);
3148
3149                 perf_output_put(handle, lost_event);
3150         }
3151
3152         return 0;
3153
3154 fail:
3155         local_inc(&data->lost);
3156         perf_output_put_handle(handle);
3157 out:
3158         rcu_read_unlock();
3159
3160         return -ENOSPC;
3161 }
3162
3163 void perf_output_end(struct perf_output_handle *handle)
3164 {
3165         struct perf_event *event = handle->event;
3166         struct perf_mmap_data *data = handle->data;
3167
3168         int wakeup_events = event->attr.wakeup_events;
3169
3170         if (handle->sample && wakeup_events) {
3171                 int events = local_inc_return(&data->events);
3172                 if (events >= wakeup_events) {
3173                         local_sub(wakeup_events, &data->events);
3174                         local_inc(&data->wakeup);
3175                 }
3176         }
3177
3178         perf_output_put_handle(handle);
3179         rcu_read_unlock();
3180 }
3181
3182 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3183 {
3184         /*
3185          * only top level events have the pid namespace they were created in
3186          */
3187         if (event->parent)
3188                 event = event->parent;
3189
3190         return task_tgid_nr_ns(p, event->ns);
3191 }
3192
3193 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3194 {
3195         /*
3196          * only top level events have the pid namespace they were created in
3197          */
3198         if (event->parent)
3199                 event = event->parent;
3200
3201         return task_pid_nr_ns(p, event->ns);
3202 }
3203
3204 static void perf_output_read_one(struct perf_output_handle *handle,
3205                                  struct perf_event *event)
3206 {
3207         u64 read_format = event->attr.read_format;
3208         u64 values[4];
3209         int n = 0;
3210
3211         values[n++] = atomic64_read(&event->count);
3212         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3213                 values[n++] = event->total_time_enabled +
3214                         atomic64_read(&event->child_total_time_enabled);
3215         }
3216         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3217                 values[n++] = event->total_time_running +
3218                         atomic64_read(&event->child_total_time_running);
3219         }
3220         if (read_format & PERF_FORMAT_ID)
3221                 values[n++] = primary_event_id(event);
3222
3223         perf_output_copy(handle, values, n * sizeof(u64));
3224 }
3225
3226 /*
3227  * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3228  */
3229 static void perf_output_read_group(struct perf_output_handle *handle,
3230                             struct perf_event *event)
3231 {
3232         struct perf_event *leader = event->group_leader, *sub;
3233         u64 read_format = event->attr.read_format;
3234         u64 values[5];
3235         int n = 0;
3236
3237         values[n++] = 1 + leader->nr_siblings;
3238
3239         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3240                 values[n++] = leader->total_time_enabled;
3241
3242         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3243                 values[n++] = leader->total_time_running;
3244
3245         if (leader != event)
3246                 leader->pmu->read(leader);
3247
3248         values[n++] = atomic64_read(&leader->count);
3249         if (read_format & PERF_FORMAT_ID)
3250                 values[n++] = primary_event_id(leader);
3251
3252         perf_output_copy(handle, values, n * sizeof(u64));
3253
3254         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3255                 n = 0;
3256
3257                 if (sub != event)
3258                         sub->pmu->read(sub);
3259
3260                 values[n++] = atomic64_read(&sub->count);
3261                 if (read_format & PERF_FORMAT_ID)
3262                         values[n++] = primary_event_id(sub);
3263
3264                 perf_output_copy(handle, values, n * sizeof(u64));
3265         }
3266 }
3267
3268 static void perf_output_read(struct perf_output_handle *handle,
3269                              struct perf_event *event)
3270 {
3271         if (event->attr.read_format & PERF_FORMAT_GROUP)
3272                 perf_output_read_group(handle, event);
3273         else
3274                 perf_output_read_one(handle, event);
3275 }
3276
3277 void perf_output_sample(struct perf_output_handle *handle,
3278                         struct perf_event_header *header,
3279                         struct perf_sample_data *data,
3280                         struct perf_event *event)
3281 {
3282         u64 sample_type = data->type;
3283
3284         perf_output_put(handle, *header);
3285
3286         if (sample_type & PERF_SAMPLE_IP)
3287                 perf_output_put(handle, data->ip);
3288
3289         if (sample_type & PERF_SAMPLE_TID)
3290                 perf_output_put(handle, data->tid_entry);
3291
3292         if (sample_type & PERF_SAMPLE_TIME)
3293                 perf_output_put(handle, data->time);
3294
3295         if (sample_type & PERF_SAMPLE_ADDR)
3296                 perf_output_put(handle, data->addr);
3297
3298         if (sample_type & PERF_SAMPLE_ID)
3299                 perf_output_put(handle, data->id);
3300
3301         if (sample_type & PERF_SAMPLE_STREAM_ID)
3302                 perf_output_put(handle, data->stream_id);
3303
3304         if (sample_type & PERF_SAMPLE_CPU)
3305                 perf_output_put(handle, data->cpu_entry);
3306
3307         if (sample_type & PERF_SAMPLE_PERIOD)
3308                 perf_output_put(handle, data->period);
3309
3310         if (sample_type & PERF_SAMPLE_READ)
3311                 perf_output_read(handle, event);
3312
3313         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3314                 if (data->callchain) {
3315                         int size = 1;
3316
3317                         if (data->callchain)
3318                                 size += data->callchain->nr;
3319
3320                         size *= sizeof(u64);
3321
3322                         perf_output_copy(handle, data->callchain, size);
3323                 } else {
3324                         u64 nr = 0;
3325                         perf_output_put(handle, nr);
3326                 }
3327         }
3328
3329         if (sample_type & PERF_SAMPLE_RAW) {
3330                 if (data->raw) {
3331                         perf_output_put(handle, data->raw->size);
3332                         perf_output_copy(handle, data->raw->data,
3333                                          data->raw->size);
3334                 } else {
3335                         struct {
3336                                 u32     size;
3337                                 u32     data;
3338                         } raw = {
3339                                 .size = sizeof(u32),
3340                                 .data = 0,
3341                         };
3342                         perf_output_put(handle, raw);
3343                 }
3344         }
3345 }
3346
3347 void perf_prepare_sample(struct perf_event_header *header,
3348                          struct perf_sample_data *data,
3349                          struct perf_event *event,
3350                          struct pt_regs *regs)
3351 {
3352         u64 sample_type = event->attr.sample_type;
3353
3354         data->type = sample_type;
3355
3356         header->type = PERF_RECORD_SAMPLE;
3357         header->size = sizeof(*header);
3358
3359         header->misc = 0;
3360         header->misc |= perf_misc_flags(regs);
3361
3362         if (sample_type & PERF_SAMPLE_IP) {
3363                 data->ip = perf_instruction_pointer(regs);
3364
3365                 header->size += sizeof(data->ip);
3366         }
3367
3368         if (sample_type & PERF_SAMPLE_TID) {
3369                 /* namespace issues */
3370                 data->tid_entry.pid = perf_event_pid(event, current);
3371                 data->tid_entry.tid = perf_event_tid(event, current);
3372
3373                 header->size += sizeof(data->tid_entry);
3374         }
3375
3376         if (sample_type & PERF_SAMPLE_TIME) {
3377                 data->time = perf_clock();
3378
3379                 header->size += sizeof(data->time);
3380         }
3381
3382         if (sample_type & PERF_SAMPLE_ADDR)
3383                 header->size += sizeof(data->addr);
3384
3385         if (sample_type & PERF_SAMPLE_ID) {
3386                 data->id = primary_event_id(event);
3387
3388                 header->size += sizeof(data->id);
3389         }
3390
3391         if (sample_type & PERF_SAMPLE_STREAM_ID) {
3392                 data->stream_id = event->id;
3393
3394                 header->size += sizeof(data->stream_id);
3395         }
3396
3397         if (sample_type & PERF_SAMPLE_CPU) {
3398                 data->cpu_entry.cpu             = raw_smp_processor_id();
3399                 data->cpu_entry.reserved        = 0;
3400
3401                 header->size += sizeof(data->cpu_entry);
3402         }
3403
3404         if (sample_type & PERF_SAMPLE_PERIOD)
3405                 header->size += sizeof(data->period);
3406
3407         if (sample_type & PERF_SAMPLE_READ)
3408                 header->size += perf_event_read_size(event);
3409
3410         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3411                 int size = 1;
3412
3413                 data->callchain = perf_callchain(regs);
3414
3415                 if (data->callchain)
3416                         size += data->callchain->nr;
3417
3418                 header->size += size * sizeof(u64);
3419         }
3420
3421         if (sample_type & PERF_SAMPLE_RAW) {
3422                 int size = sizeof(u32);
3423
3424                 if (data->raw)
3425                         size += data->raw->size;
3426                 else
3427                         size += sizeof(u32);
3428
3429                 WARN_ON_ONCE(size & (sizeof(u64)-1));
3430                 header->size += size;
3431         }
3432 }
3433
3434 static void perf_event_output(struct perf_event *event, int nmi,
3435                                 struct perf_sample_data *data,
3436                                 struct pt_regs *regs)
3437 {
3438         struct perf_output_handle handle;
3439         struct perf_event_header header;
3440
3441         perf_prepare_sample(&header, data, event, regs);
3442
3443         if (perf_output_begin(&handle, event, header.size, nmi, 1))
3444                 return;
3445
3446         perf_output_sample(&handle, &header, data, event);
3447
3448         perf_output_end(&handle);
3449 }
3450
3451 /*
3452  * read event_id
3453  */
3454
3455 struct perf_read_event {
3456         struct perf_event_header        header;
3457
3458         u32                             pid;
3459         u32                             tid;
3460 };
3461
3462 static void
3463 perf_event_read_event(struct perf_event *event,
3464                         struct task_struct *task)
3465 {
3466         struct perf_output_handle handle;
3467         struct perf_read_event read_event = {
3468                 .header = {
3469                         .type = PERF_RECORD_READ,
3470                         .misc = 0,
3471                         .size = sizeof(read_event) + perf_event_read_size(event),
3472                 },
3473                 .pid = perf_event_pid(event, task),
3474                 .tid = perf_event_tid(event, task),
3475         };
3476         int ret;
3477
3478         ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3479         if (ret)
3480                 return;
3481
3482         perf_output_put(&handle, read_event);
3483         perf_output_read(&handle, event);
3484
3485         perf_output_end(&handle);
3486 }
3487
3488 /*
3489  * task tracking -- fork/exit
3490  *
3491  * enabled by: attr.comm | attr.mmap | attr.task
3492  */
3493
3494 struct perf_task_event {
3495         struct task_struct              *task;
3496         struct perf_event_context       *task_ctx;
3497
3498         struct {
3499                 struct perf_event_header        header;
3500
3501                 u32                             pid;
3502                 u32                             ppid;
3503                 u32                             tid;
3504                 u32                             ptid;
3505                 u64                             time;
3506         } event_id;
3507 };
3508
3509 static void perf_event_task_output(struct perf_event *event,
3510                                      struct perf_task_event *task_event)
3511 {
3512         struct perf_output_handle handle;
3513         struct task_struct *task = task_event->task;
3514         int size, ret;
3515
3516         size  = task_event->event_id.header.size;
3517         ret = perf_output_begin(&handle, event, size, 0, 0);
3518
3519         if (ret)
3520                 return;
3521
3522         task_event->event_id.pid = perf_event_pid(event, task);
3523         task_event->event_id.ppid = perf_event_pid(event, current);
3524
3525         task_event->event_id.tid = perf_event_tid(event, task);
3526         task_event->event_id.ptid = perf_event_tid(event, current);
3527
3528         perf_output_put(&handle, task_event->event_id);
3529
3530         perf_output_end(&handle);
3531 }
3532
3533 static int perf_event_task_match(struct perf_event *event)
3534 {
3535         if (event->state < PERF_EVENT_STATE_INACTIVE)
3536                 return 0;
3537
3538         if (event->cpu != -1 && event->cpu != smp_processor_id())
3539                 return 0;
3540
3541         if (event->attr.comm || event->attr.mmap || event->attr.task)
3542                 return 1;
3543
3544         return 0;
3545 }
3546
3547 static void perf_event_task_ctx(struct perf_event_context *ctx,
3548                                   struct perf_task_event *task_event)
3549 {
3550         struct perf_event *event;
3551
3552         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3553                 if (perf_event_task_match(event))
3554                         perf_event_task_output(event, task_event);
3555         }
3556 }
3557
3558 static void perf_event_task_event(struct perf_task_event *task_event)
3559 {
3560         struct perf_cpu_context *cpuctx;
3561         struct perf_event_context *ctx = task_event->task_ctx;
3562
3563         rcu_read_lock();
3564         cpuctx = &get_cpu_var(perf_cpu_context);
3565         perf_event_task_ctx(&cpuctx->ctx, task_event);
3566         if (!ctx)
3567                 ctx = rcu_dereference(current->perf_event_ctxp);
3568         if (ctx)
3569                 perf_event_task_ctx(ctx, task_event);
3570         put_cpu_var(perf_cpu_context);
3571         rcu_read_unlock();
3572 }
3573
3574 static void perf_event_task(struct task_struct *task,
3575                               struct perf_event_context *task_ctx,
3576                               int new)
3577 {
3578         struct perf_task_event task_event;
3579
3580         if (!atomic_read(&nr_comm_events) &&
3581             !atomic_read(&nr_mmap_events) &&
3582             !atomic_read(&nr_task_events))
3583                 return;
3584
3585         task_event = (struct perf_task_event){
3586                 .task     = task,
3587                 .task_ctx = task_ctx,
3588                 .event_id    = {
3589                         .header = {
3590                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3591                                 .misc = 0,
3592                                 .size = sizeof(task_event.event_id),
3593                         },
3594                         /* .pid  */
3595                         /* .ppid */
3596                         /* .tid  */
3597                         /* .ptid */
3598                         .time = perf_clock(),
3599                 },
3600         };
3601
3602         perf_event_task_event(&task_event);
3603 }
3604
3605 void perf_event_fork(struct task_struct *task)
3606 {
3607         perf_event_task(task, NULL, 1);
3608 }
3609
3610 /*
3611  * comm tracking
3612  */
3613
3614 struct perf_comm_event {
3615         struct task_struct      *task;
3616         char                    *comm;
3617         int                     comm_size;
3618
3619         struct {
3620                 struct perf_event_header        header;
3621
3622                 u32                             pid;
3623                 u32                             tid;
3624         } event_id;
3625 };
3626
3627 static void perf_event_comm_output(struct perf_event *event,
3628                                      struct perf_comm_event *comm_event)
3629 {
3630         struct perf_output_handle handle;
3631         int size = comm_event->event_id.header.size;
3632         int ret = perf_output_begin(&handle, event, size, 0, 0);
3633
3634         if (ret)
3635                 return;
3636
3637         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3638         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3639
3640         perf_output_put(&handle, comm_event->event_id);
3641         perf_output_copy(&handle, comm_event->comm,
3642                                    comm_event->comm_size);
3643         perf_output_end(&handle);
3644 }
3645
3646 static int perf_event_comm_match(struct perf_event *event)
3647 {
3648         if (event->state < PERF_EVENT_STATE_INACTIVE)
3649                 return 0;
3650
3651         if (event->cpu != -1 && event->cpu != smp_processor_id())
3652                 return 0;
3653
3654         if (event->attr.comm)
3655                 return 1;
3656
3657         return 0;
3658 }
3659
3660 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3661                                   struct perf_comm_event *comm_event)
3662 {
3663         struct perf_event *event;
3664
3665         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3666                 if (perf_event_comm_match(event))
3667                         perf_event_comm_output(event, comm_event);
3668         }
3669 }
3670
3671 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3672 {
3673         struct perf_cpu_context *cpuctx;
3674         struct perf_event_context *ctx;
3675         unsigned int size;
3676         char comm[TASK_COMM_LEN];
3677
3678         memset(comm, 0, sizeof(comm));
3679         strlcpy(comm, comm_event->task->comm, sizeof(comm));
3680         size = ALIGN(strlen(comm)+1, sizeof(u64));
3681
3682         comm_event->comm = comm;
3683         comm_event->comm_size = size;
3684
3685         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3686
3687         rcu_read_lock();
3688         cpuctx = &get_cpu_var(perf_cpu_context);
3689         perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3690         ctx = rcu_dereference(current->perf_event_ctxp);
3691         if (ctx)
3692                 perf_event_comm_ctx(ctx, comm_event);
3693         put_cpu_var(perf_cpu_context);
3694         rcu_read_unlock();
3695 }
3696
3697 void perf_event_comm(struct task_struct *task)
3698 {
3699         struct perf_comm_event comm_event;
3700
3701         if (task->perf_event_ctxp)
3702                 perf_event_enable_on_exec(task);
3703
3704         if (!atomic_read(&nr_comm_events))
3705                 return;
3706
3707         comm_event = (struct perf_comm_event){
3708                 .task   = task,
3709                 /* .comm      */
3710                 /* .comm_size */
3711                 .event_id  = {
3712                         .header = {
3713                                 .type = PERF_RECORD_COMM,
3714                                 .misc = 0,
3715                                 /* .size */
3716                         },
3717                         /* .pid */
3718                         /* .tid */
3719                 },
3720         };
3721
3722         perf_event_comm_event(&comm_event);
3723 }
3724
3725 /*
3726  * mmap tracking
3727  */
3728
3729 struct perf_mmap_event {
3730         struct vm_area_struct   *vma;
3731
3732         const char              *file_name;
3733         int                     file_size;
3734
3735         struct {
3736                 struct perf_event_header        header;
3737
3738                 u32                             pid;
3739                 u32                             tid;
3740                 u64                             start;
3741                 u64                             len;
3742                 u64                             pgoff;
3743         } event_id;
3744 };
3745
3746 static void perf_event_mmap_output(struct perf_event *event,
3747                                      struct perf_mmap_event *mmap_event)
3748 {
3749         struct perf_output_handle handle;
3750         int size = mmap_event->event_id.header.size;
3751         int ret = perf_output_begin(&handle, event, size, 0, 0);
3752
3753         if (ret)
3754                 return;
3755
3756         mmap_event->event_id.pid = perf_event_pid(event, current);
3757         mmap_event->event_id.tid = perf_event_tid(event, current);
3758
3759         perf_output_put(&handle, mmap_event->event_id);
3760         perf_output_copy(&handle, mmap_event->file_name,
3761                                    mmap_event->file_size);
3762         perf_output_end(&handle);
3763 }
3764
3765 static int perf_event_mmap_match(struct perf_event *event,
3766                                    struct perf_mmap_event *mmap_event)
3767 {
3768         if (event->state < PERF_EVENT_STATE_INACTIVE)
3769                 return 0;
3770
3771         if (event->cpu != -1 && event->cpu != smp_processor_id())
3772                 return 0;
3773
3774         if (event->attr.mmap)
3775                 return 1;
3776
3777         return 0;
3778 }
3779
3780 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3781                                   struct perf_mmap_event *mmap_event)
3782 {
3783         struct perf_event *event;
3784
3785         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3786                 if (perf_event_mmap_match(event, mmap_event))
3787                         perf_event_mmap_output(event, mmap_event);
3788         }
3789 }
3790
3791 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3792 {
3793         struct perf_cpu_context *cpuctx;
3794         struct perf_event_context *ctx;
3795         struct vm_area_struct *vma = mmap_event->vma;
3796         struct file *file = vma->vm_file;
3797         unsigned int size;
3798         char tmp[16];
3799         char *buf = NULL;
3800         const char *name;
3801
3802         memset(tmp, 0, sizeof(tmp));
3803
3804         if (file) {
3805                 /*
3806                  * d_path works from the end of the buffer backwards, so we
3807                  * need to add enough zero bytes after the string to handle
3808                  * the 64bit alignment we do later.
3809                  */
3810                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3811                 if (!buf) {
3812                         name = strncpy(tmp, "//enomem", sizeof(tmp));
3813                         goto got_name;
3814                 }
3815                 name = d_path(&file->f_path, buf, PATH_MAX);
3816                 if (IS_ERR(name)) {
3817                         name = strncpy(tmp, "//toolong", sizeof(tmp));
3818                         goto got_name;
3819                 }
3820         } else {
3821                 if (arch_vma_name(mmap_event->vma)) {
3822                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3823                                        sizeof(tmp));
3824                         goto got_name;
3825                 }
3826
3827                 if (!vma->vm_mm) {
3828                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
3829                         goto got_name;
3830                 }
3831
3832                 name = strncpy(tmp, "//anon", sizeof(tmp));
3833                 goto got_name;
3834         }
3835
3836 got_name:
3837         size = ALIGN(strlen(name)+1, sizeof(u64));
3838
3839         mmap_event->file_name = name;
3840         mmap_event->file_size = size;
3841
3842         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3843
3844         rcu_read_lock();
3845         cpuctx = &get_cpu_var(perf_cpu_context);
3846         perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3847         ctx = rcu_dereference(current->perf_event_ctxp);
3848         if (ctx)
3849                 perf_event_mmap_ctx(ctx, mmap_event);
3850         put_cpu_var(perf_cpu_context);
3851         rcu_read_unlock();
3852
3853         kfree(buf);
3854 }
3855
3856 void __perf_event_mmap(struct vm_area_struct *vma)
3857 {
3858         struct perf_mmap_event mmap_event;
3859
3860         if (!atomic_read(&nr_mmap_events))
3861                 return;
3862
3863         mmap_event = (struct perf_mmap_event){
3864                 .vma    = vma,
3865                 /* .file_name */
3866                 /* .file_size */
3867                 .event_id  = {
3868                         .header = {
3869                                 .type = PERF_RECORD_MMAP,
3870                                 .misc = PERF_RECORD_MISC_USER,
3871                                 /* .size */
3872                         },
3873                         /* .pid */
3874                         /* .tid */
3875                         .start  = vma->vm_start,
3876                         .len    = vma->vm_end - vma->vm_start,
3877                         .pgoff  = (u64)vma->vm_pgoff << PAGE_SHIFT,
3878                 },
3879         };
3880
3881         perf_event_mmap_event(&mmap_event);
3882 }
3883
3884 /*
3885  * IRQ throttle logging
3886  */
3887
3888 static void perf_log_throttle(struct perf_event *event, int enable)
3889 {
3890         struct perf_output_handle handle;
3891         int ret;
3892
3893         struct {
3894                 struct perf_event_header        header;
3895                 u64                             time;
3896                 u64                             id;
3897                 u64                             stream_id;
3898         } throttle_event = {
3899                 .header = {
3900                         .type = PERF_RECORD_THROTTLE,
3901                         .misc = 0,
3902                         .size = sizeof(throttle_event),
3903                 },
3904                 .time           = perf_clock(),
3905                 .id             = primary_event_id(event),
3906                 .stream_id      = event->id,
3907         };
3908
3909         if (enable)
3910                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3911
3912         ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3913         if (ret)
3914                 return;
3915
3916         perf_output_put(&handle, throttle_event);
3917         perf_output_end(&handle);
3918 }
3919
3920 /*
3921  * Generic event overflow handling, sampling.
3922  */
3923
3924 static int __perf_event_overflow(struct perf_event *event, int nmi,
3925                                    int throttle, struct perf_sample_data *data,
3926                                    struct pt_regs *regs)
3927 {
3928         int events = atomic_read(&event->event_limit);
3929         struct hw_perf_event *hwc = &event->hw;
3930         int ret = 0;
3931
3932         throttle = (throttle && event->pmu->unthrottle != NULL);
3933
3934         if (!throttle) {
3935                 hwc->interrupts++;
3936         } else {
3937                 if (hwc->interrupts != MAX_INTERRUPTS) {
3938                         hwc->interrupts++;
3939                         if (HZ * hwc->interrupts >
3940                                         (u64)sysctl_perf_event_sample_rate) {
3941                                 hwc->interrupts = MAX_INTERRUPTS;
3942                                 perf_log_throttle(event, 0);
3943                                 ret = 1;
3944                         }
3945                 } else {
3946                         /*
3947                          * Keep re-disabling events even though on the previous
3948                          * pass we disabled it - just in case we raced with a
3949                          * sched-in and the event got enabled again:
3950                          */
3951                         ret = 1;
3952                 }
3953         }
3954
3955         if (event->attr.freq) {
3956                 u64 now = perf_clock();
3957                 s64 delta = now - hwc->freq_time_stamp;
3958
3959                 hwc->freq_time_stamp = now;
3960
3961                 if (delta > 0 && delta < 2*TICK_NSEC)
3962                         perf_adjust_period(event, delta, hwc->last_period);
3963         }
3964
3965         /*
3966          * XXX event_limit might not quite work as expected on inherited
3967          * events
3968          */
3969
3970         event->pending_kill = POLL_IN;
3971         if (events && atomic_dec_and_test(&event->event_limit)) {
3972                 ret = 1;
3973                 event->pending_kill = POLL_HUP;
3974                 if (nmi) {
3975                         event->pending_disable = 1;
3976                         perf_pending_queue(&event->pending,
3977                                            perf_pending_event);
3978                 } else
3979                         perf_event_disable(event);
3980         }
3981
3982         if (event->overflow_handler)
3983                 event->overflow_handler(event, nmi, data, regs);
3984         else
3985                 perf_event_output(event, nmi, data, regs);
3986
3987         return ret;
3988 }
3989
3990 int perf_event_overflow(struct perf_event *event, int nmi,
3991                           struct perf_sample_data *data,
3992                           struct pt_regs *regs)
3993 {
3994         return __perf_event_overflow(event, nmi, 1, data, regs);
3995 }
3996
3997 /*
3998  * Generic software event infrastructure
3999  */
4000
4001 /*
4002  * We directly increment event->count and keep a second value in
4003  * event->hw.period_left to count intervals. This period event
4004  * is kept in the range [-sample_period, 0] so that we can use the
4005  * sign as trigger.
4006  */
4007
4008 static u64 perf_swevent_set_period(struct perf_event *event)
4009 {
4010         struct hw_perf_event *hwc = &event->hw;
4011         u64 period = hwc->last_period;
4012         u64 nr, offset;
4013         s64 old, val;
4014
4015         hwc->last_period = hwc->sample_period;
4016
4017 again:
4018         old = val = atomic64_read(&hwc->period_left);
4019         if (val < 0)
4020                 return 0;
4021
4022         nr = div64_u64(period + val, period);
4023         offset = nr * period;
4024         val -= offset;
4025         if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
4026                 goto again;
4027
4028         return nr;
4029 }
4030
4031 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4032                                     int nmi, struct perf_sample_data *data,
4033                                     struct pt_regs *regs)
4034 {
4035         struct hw_perf_event *hwc = &event->hw;
4036         int throttle = 0;
4037
4038         data->period = event->hw.last_period;
4039         if (!overflow)
4040                 overflow = perf_swevent_set_period(event);
4041
4042         if (hwc->interrupts == MAX_INTERRUPTS)
4043                 return;
4044
4045         for (; overflow; overflow--) {
4046                 if (__perf_event_overflow(event, nmi, throttle,
4047                                             data, regs)) {
4048                         /*
4049                          * We inhibit the overflow from happening when
4050                          * hwc->interrupts == MAX_INTERRUPTS.
4051                          */
4052                         break;
4053                 }
4054                 throttle = 1;
4055         }
4056 }
4057
4058 static void perf_swevent_add(struct perf_event *event, u64 nr,
4059                                int nmi, struct perf_sample_data *data,
4060                                struct pt_regs *regs)
4061 {
4062         struct hw_perf_event *hwc = &event->hw;
4063
4064         atomic64_add(nr, &event->count);
4065
4066         if (!regs)
4067                 return;
4068
4069         if (!hwc->sample_period)
4070                 return;
4071
4072         if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4073                 return perf_swevent_overflow(event, 1, nmi, data, regs);
4074
4075         if (atomic64_add_negative(nr, &hwc->period_left))
4076                 return;
4077
4078         perf_swevent_overflow(event, 0, nmi, data, regs);
4079 }
4080
4081 static int perf_exclude_event(struct perf_event *event,
4082                               struct pt_regs *regs)
4083 {
4084         if (regs) {
4085                 if (event->attr.exclude_user && user_mode(regs))
4086                         return 1;
4087
4088                 if (event->attr.exclude_kernel && !user_mode(regs))
4089                         return 1;
4090         }
4091
4092         return 0;
4093 }
4094
4095 static int perf_swevent_match(struct perf_event *event,
4096                                 enum perf_type_id type,
4097                                 u32 event_id,
4098                                 struct perf_sample_data *data,
4099                                 struct pt_regs *regs)
4100 {
4101         if (event->attr.type != type)
4102                 return 0;
4103
4104         if (event->attr.config != event_id)
4105                 return 0;
4106
4107         if (perf_exclude_event(event, regs))
4108                 return 0;
4109
4110         return 1;
4111 }
4112
4113 static inline u64 swevent_hash(u64 type, u32 event_id)
4114 {
4115         u64 val = event_id | (type << 32);
4116
4117         return hash_64(val, SWEVENT_HLIST_BITS);
4118 }
4119
4120 static inline struct hlist_head *
4121 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4122 {
4123         u64 hash = swevent_hash(type, event_id);
4124
4125         return &hlist->heads[hash];
4126 }
4127
4128 /* For the read side: events when they trigger */
4129 static inline struct hlist_head *
4130 find_swevent_head_rcu(struct perf_cpu_context *ctx, u64 type, u32 event_id)
4131 {
4132         struct swevent_hlist *hlist;
4133
4134         hlist = rcu_dereference(ctx->swevent_hlist);
4135         if (!hlist)
4136                 return NULL;
4137
4138         return __find_swevent_head(hlist, type, event_id);
4139 }
4140
4141 /* For the event head insertion and removal in the hlist */
4142 static inline struct hlist_head *
4143 find_swevent_head(struct perf_cpu_context *ctx, struct perf_event *event)
4144 {
4145         struct swevent_hlist *hlist;
4146         u32 event_id = event->attr.config;
4147         u64 type = event->attr.type;
4148
4149         /*
4150          * Event scheduling is always serialized against hlist allocation
4151          * and release. Which makes the protected version suitable here.
4152          * The context lock guarantees that.
4153          */
4154         hlist = rcu_dereference_protected(ctx->swevent_hlist,
4155                                           lockdep_is_held(&event->ctx->lock));
4156         if (!hlist)
4157                 return NULL;
4158
4159         return __find_swevent_head(hlist, type, event_id);
4160 }
4161
4162 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4163                                     u64 nr, int nmi,
4164                                     struct perf_sample_data *data,
4165                                     struct pt_regs *regs)
4166 {
4167         struct perf_cpu_context *cpuctx;
4168         struct perf_event *event;
4169         struct hlist_node *node;
4170         struct hlist_head *head;
4171
4172         cpuctx = &__get_cpu_var(perf_cpu_context);
4173
4174         rcu_read_lock();
4175
4176         head = find_swevent_head_rcu(cpuctx, type, event_id);
4177
4178         if (!head)
4179                 goto end;
4180
4181         hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4182                 if (perf_swevent_match(event, type, event_id, data, regs))
4183                         perf_swevent_add(event, nr, nmi, data, regs);
4184         }
4185 end:
4186         rcu_read_unlock();
4187 }
4188
4189 int perf_swevent_get_recursion_context(void)
4190 {
4191         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4192         int rctx;
4193
4194         if (in_nmi())
4195                 rctx = 3;
4196         else if (in_irq())
4197                 rctx = 2;
4198         else if (in_softirq())
4199                 rctx = 1;
4200         else
4201                 rctx = 0;
4202
4203         if (cpuctx->recursion[rctx])
4204                 return -1;
4205
4206         cpuctx->recursion[rctx]++;
4207         barrier();
4208
4209         return rctx;
4210 }
4211 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4212
4213 void perf_swevent_put_recursion_context(int rctx)
4214 {
4215         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4216         barrier();
4217         cpuctx->recursion[rctx]--;
4218 }
4219 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
4220
4221
4222 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4223                             struct pt_regs *regs, u64 addr)
4224 {
4225         struct perf_sample_data data;
4226         int rctx;
4227
4228         preempt_disable_notrace();
4229         rctx = perf_swevent_get_recursion_context();
4230         if (rctx < 0)
4231                 return;
4232
4233         perf_sample_data_init(&data, addr);
4234
4235         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4236
4237         perf_swevent_put_recursion_context(rctx);
4238         preempt_enable_notrace();
4239 }
4240
4241 static void perf_swevent_read(struct perf_event *event)
4242 {
4243 }
4244
4245 static int perf_swevent_enable(struct perf_event *event)
4246 {
4247         struct hw_perf_event *hwc = &event->hw;
4248         struct perf_cpu_context *cpuctx;
4249         struct hlist_head *head;
4250
4251         cpuctx = &__get_cpu_var(perf_cpu_context);
4252
4253         if (hwc->sample_period) {
4254                 hwc->last_period = hwc->sample_period;
4255                 perf_swevent_set_period(event);
4256         }
4257
4258         head = find_swevent_head(cpuctx, event);
4259         if (WARN_ON_ONCE(!head))
4260                 return -EINVAL;
4261
4262         hlist_add_head_rcu(&event->hlist_entry, head);
4263
4264         return 0;
4265 }
4266
4267 static void perf_swevent_disable(struct perf_event *event)
4268 {
4269         hlist_del_rcu(&event->hlist_entry);
4270 }
4271
4272 static void perf_swevent_void(struct perf_event *event)
4273 {
4274 }
4275
4276 static int perf_swevent_int(struct perf_event *event)
4277 {
4278         return 0;
4279 }
4280
4281 static const struct pmu perf_ops_generic = {
4282         .enable         = perf_swevent_enable,
4283         .disable        = perf_swevent_disable,
4284         .start          = perf_swevent_int,
4285         .stop           = perf_swevent_void,
4286         .read           = perf_swevent_read,
4287         .unthrottle     = perf_swevent_void, /* hwc->interrupts already reset */
4288 };
4289
4290 /*
4291  * hrtimer based swevent callback
4292  */
4293
4294 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4295 {
4296         enum hrtimer_restart ret = HRTIMER_RESTART;
4297         struct perf_sample_data data;
4298         struct pt_regs *regs;
4299         struct perf_event *event;
4300         u64 period;
4301
4302         event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4303         event->pmu->read(event);
4304
4305         perf_sample_data_init(&data, 0);
4306         data.period = event->hw.last_period;
4307         regs = get_irq_regs();
4308
4309         if (regs && !perf_exclude_event(event, regs)) {
4310                 if (!(event->attr.exclude_idle && current->pid == 0))
4311                         if (perf_event_overflow(event, 0, &data, regs))
4312                                 ret = HRTIMER_NORESTART;
4313         }
4314
4315         period = max_t(u64, 10000, event->hw.sample_period);
4316         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4317
4318         return ret;
4319 }
4320
4321 static void perf_swevent_start_hrtimer(struct perf_event *event)
4322 {
4323         struct hw_perf_event *hwc = &event->hw;
4324
4325         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4326         hwc->hrtimer.function = perf_swevent_hrtimer;
4327         if (hwc->sample_period) {
4328                 u64 period;
4329
4330                 if (hwc->remaining) {
4331                         if (hwc->remaining < 0)
4332                                 period = 10000;
4333                         else
4334                                 period = hwc->remaining;
4335                         hwc->remaining = 0;
4336                 } else {
4337                         period = max_t(u64, 10000, hwc->sample_period);
4338                 }
4339                 __hrtimer_start_range_ns(&hwc->hrtimer,
4340                                 ns_to_ktime(period), 0,
4341                                 HRTIMER_MODE_REL, 0);
4342         }
4343 }
4344
4345 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4346 {
4347         struct hw_perf_event *hwc = &event->hw;
4348
4349         if (hwc->sample_period) {
4350                 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4351                 hwc->remaining = ktime_to_ns(remaining);
4352
4353                 hrtimer_cancel(&hwc->hrtimer);
4354         }
4355 }
4356
4357 /*
4358  * Software event: cpu wall time clock
4359  */
4360
4361 static void cpu_clock_perf_event_update(struct perf_event *event)
4362 {
4363         int cpu = raw_smp_processor_id();
4364         s64 prev;
4365         u64 now;
4366
4367         now = cpu_clock(cpu);
4368         prev = atomic64_xchg(&event->hw.prev_count, now);
4369         atomic64_add(now - prev, &event->count);
4370 }
4371
4372 static int cpu_clock_perf_event_enable(struct perf_event *event)
4373 {
4374         struct hw_perf_event *hwc = &event->hw;
4375         int cpu = raw_smp_processor_id();
4376
4377         atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4378         perf_swevent_start_hrtimer(event);
4379
4380         return 0;
4381 }
4382
4383 static void cpu_clock_perf_event_disable(struct perf_event *event)
4384 {
4385         perf_swevent_cancel_hrtimer(event);
4386         cpu_clock_perf_event_update(event);
4387 }
4388
4389 static void cpu_clock_perf_event_read(struct perf_event *event)
4390 {
4391         cpu_clock_perf_event_update(event);
4392 }
4393
4394 static const struct pmu perf_ops_cpu_clock = {
4395         .enable         = cpu_clock_perf_event_enable,
4396         .disable        = cpu_clock_perf_event_disable,
4397         .read           = cpu_clock_perf_event_read,
4398 };
4399
4400 /*
4401  * Software event: task time clock
4402  */
4403
4404 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4405 {
4406         u64 prev;
4407         s64 delta;
4408
4409         prev = atomic64_xchg(&event->hw.prev_count, now);
4410         delta = now - prev;
4411         atomic64_add(delta, &event->count);
4412 }
4413
4414 static int task_clock_perf_event_enable(struct perf_event *event)
4415 {
4416         struct hw_perf_event *hwc = &event->hw;
4417         u64 now;
4418
4419         now = event->ctx->time;
4420
4421         atomic64_set(&hwc->prev_count, now);
4422
4423         perf_swevent_start_hrtimer(event);
4424
4425         return 0;
4426 }
4427
4428 static void task_clock_perf_event_disable(struct perf_event *event)
4429 {
4430         perf_swevent_cancel_hrtimer(event);
4431         task_clock_perf_event_update(event, event->ctx->time);
4432
4433 }
4434
4435 static void task_clock_perf_event_read(struct perf_event *event)
4436 {
4437         u64 time;
4438
4439         if (!in_nmi()) {
4440                 update_context_time(event->ctx);
4441                 time = event->ctx->time;
4442         } else {
4443                 u64 now = perf_clock();
4444                 u64 delta = now - event->ctx->timestamp;
4445                 time = event->ctx->time + delta;
4446         }
4447
4448         task_clock_perf_event_update(event, time);
4449 }
4450
4451 static const struct pmu perf_ops_task_clock = {
4452         .enable         = task_clock_perf_event_enable,
4453         .disable        = task_clock_perf_event_disable,
4454         .read           = task_clock_perf_event_read,
4455 };
4456
4457 /* Deref the hlist from the update side */
4458 static inline struct swevent_hlist *
4459 swevent_hlist_deref(struct perf_cpu_context *cpuctx)
4460 {
4461         return rcu_dereference_protected(cpuctx->swevent_hlist,
4462                                          lockdep_is_held(&cpuctx->hlist_mutex));
4463 }
4464
4465 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4466 {
4467         struct swevent_hlist *hlist;
4468
4469         hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4470         kfree(hlist);
4471 }
4472
4473 static void swevent_hlist_release(struct perf_cpu_context *cpuctx)
4474 {
4475         struct swevent_hlist *hlist = swevent_hlist_deref(cpuctx);
4476
4477         if (!hlist)
4478                 return;
4479
4480         rcu_assign_pointer(cpuctx->swevent_hlist, NULL);
4481         call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4482 }
4483
4484 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4485 {
4486         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4487
4488         mutex_lock(&cpuctx->hlist_mutex);
4489
4490         if (!--cpuctx->hlist_refcount)
4491                 swevent_hlist_release(cpuctx);
4492
4493         mutex_unlock(&cpuctx->hlist_mutex);
4494 }
4495
4496 static void swevent_hlist_put(struct perf_event *event)
4497 {
4498         int cpu;
4499
4500         if (event->cpu != -1) {
4501                 swevent_hlist_put_cpu(event, event->cpu);
4502                 return;
4503         }
4504
4505         for_each_possible_cpu(cpu)
4506                 swevent_hlist_put_cpu(event, cpu);
4507 }
4508
4509 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4510 {
4511         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4512         int err = 0;
4513
4514         mutex_lock(&cpuctx->hlist_mutex);
4515
4516         if (!swevent_hlist_deref(cpuctx) && cpu_online(cpu)) {
4517                 struct swevent_hlist *hlist;
4518
4519                 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4520                 if (!hlist) {
4521                         err = -ENOMEM;
4522                         goto exit;
4523                 }
4524                 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
4525         }
4526         cpuctx->hlist_refcount++;
4527  exit:
4528         mutex_unlock(&cpuctx->hlist_mutex);
4529
4530         return err;
4531 }
4532
4533 static int swevent_hlist_get(struct perf_event *event)
4534 {
4535         int err;
4536         int cpu, failed_cpu;
4537
4538         if (event->cpu != -1)
4539                 return swevent_hlist_get_cpu(event, event->cpu);
4540
4541         get_online_cpus();
4542         for_each_possible_cpu(cpu) {
4543                 err = swevent_hlist_get_cpu(event, cpu);
4544                 if (err) {
4545                         failed_cpu = cpu;
4546                         goto fail;
4547                 }
4548         }
4549         put_online_cpus();
4550
4551         return 0;
4552  fail:
4553         for_each_possible_cpu(cpu) {
4554                 if (cpu == failed_cpu)
4555                         break;
4556                 swevent_hlist_put_cpu(event, cpu);
4557         }
4558
4559         put_online_cpus();
4560         return err;
4561 }
4562
4563 #ifdef CONFIG_EVENT_TRACING
4564
4565 static const struct pmu perf_ops_tracepoint = {
4566         .enable         = perf_trace_enable,
4567         .disable        = perf_trace_disable,
4568         .start          = perf_swevent_int,
4569         .stop           = perf_swevent_void,
4570         .read           = perf_swevent_read,
4571         .unthrottle     = perf_swevent_void,
4572 };
4573
4574 static int perf_tp_filter_match(struct perf_event *event,
4575                                 struct perf_sample_data *data)
4576 {
4577         void *record = data->raw->data;
4578
4579         if (likely(!event->filter) || filter_match_preds(event->filter, record))
4580                 return 1;
4581         return 0;
4582 }
4583
4584 static int perf_tp_event_match(struct perf_event *event,
4585                                 struct perf_sample_data *data,
4586                                 struct pt_regs *regs)
4587 {
4588         /*
4589          * All tracepoints are from kernel-space.
4590          */
4591         if (event->attr.exclude_kernel)
4592                 return 0;
4593
4594         if (!perf_tp_filter_match(event, data))
4595                 return 0;
4596
4597         return 1;
4598 }
4599
4600 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
4601                    struct pt_regs *regs, struct hlist_head *head)
4602 {
4603         struct perf_sample_data data;
4604         struct perf_event *event;
4605         struct hlist_node *node;
4606
4607         struct perf_raw_record raw = {
4608                 .size = entry_size,
4609                 .data = record,
4610         };
4611
4612         perf_sample_data_init(&data, addr);
4613         data.raw = &raw;
4614
4615         rcu_read_lock();
4616         hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4617                 if (perf_tp_event_match(event, &data, regs))
4618                         perf_swevent_add(event, count, 1, &data, regs);
4619         }
4620         rcu_read_unlock();
4621 }
4622 EXPORT_SYMBOL_GPL(perf_tp_event);
4623
4624 static void tp_perf_event_destroy(struct perf_event *event)
4625 {
4626         perf_trace_destroy(event);
4627 }
4628
4629 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4630 {
4631         int err;
4632
4633         /*
4634          * Raw tracepoint data is a severe data leak, only allow root to
4635          * have these.
4636          */
4637         if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4638                         perf_paranoid_tracepoint_raw() &&
4639                         !capable(CAP_SYS_ADMIN))
4640                 return ERR_PTR(-EPERM);
4641
4642         err = perf_trace_init(event);
4643         if (err)
4644                 return NULL;
4645
4646         event->destroy = tp_perf_event_destroy;
4647
4648         return &perf_ops_tracepoint;
4649 }
4650
4651 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4652 {
4653         char *filter_str;
4654         int ret;
4655
4656         if (event->attr.type != PERF_TYPE_TRACEPOINT)
4657                 return -EINVAL;
4658
4659         filter_str = strndup_user(arg, PAGE_SIZE);
4660         if (IS_ERR(filter_str))
4661                 return PTR_ERR(filter_str);
4662
4663         ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4664
4665         kfree(filter_str);
4666         return ret;
4667 }
4668
4669 static void perf_event_free_filter(struct perf_event *event)
4670 {
4671         ftrace_profile_free_filter(event);
4672 }
4673
4674 #else
4675
4676 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4677 {
4678         return NULL;
4679 }
4680
4681 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4682 {
4683         return -ENOENT;
4684 }
4685
4686 static void perf_event_free_filter(struct perf_event *event)
4687 {
4688 }
4689
4690 #endif /* CONFIG_EVENT_TRACING */
4691
4692 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4693 static void bp_perf_event_destroy(struct perf_event *event)
4694 {
4695         release_bp_slot(event);
4696 }
4697
4698 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4699 {
4700         int err;
4701
4702         err = register_perf_hw_breakpoint(bp);
4703         if (err)
4704                 return ERR_PTR(err);
4705
4706         bp->destroy = bp_perf_event_destroy;
4707
4708         return &perf_ops_bp;
4709 }
4710
4711 void perf_bp_event(struct perf_event *bp, void *data)
4712 {
4713         struct perf_sample_data sample;
4714         struct pt_regs *regs = data;
4715
4716         perf_sample_data_init(&sample, bp->attr.bp_addr);
4717
4718         if (!perf_exclude_event(bp, regs))
4719                 perf_swevent_add(bp, 1, 1, &sample, regs);
4720 }
4721 #else
4722 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4723 {
4724         return NULL;
4725 }
4726
4727 void perf_bp_event(struct perf_event *bp, void *regs)
4728 {
4729 }
4730 #endif
4731
4732 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4733
4734 static void sw_perf_event_destroy(struct perf_event *event)
4735 {
4736         u64 event_id = event->attr.config;
4737
4738         WARN_ON(event->parent);
4739
4740         atomic_dec(&perf_swevent_enabled[event_id]);
4741         swevent_hlist_put(event);
4742 }
4743
4744 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4745 {
4746         const struct pmu *pmu = NULL;
4747         u64 event_id = event->attr.config;
4748
4749         /*
4750          * Software events (currently) can't in general distinguish
4751          * between user, kernel and hypervisor events.
4752          * However, context switches and cpu migrations are considered
4753          * to be kernel events, and page faults are never hypervisor
4754          * events.
4755          */
4756         switch (event_id) {
4757         case PERF_COUNT_SW_CPU_CLOCK:
4758                 pmu = &perf_ops_cpu_clock;
4759
4760                 break;
4761         case PERF_COUNT_SW_TASK_CLOCK:
4762                 /*
4763                  * If the user instantiates this as a per-cpu event,
4764                  * use the cpu_clock event instead.
4765                  */
4766                 if (event->ctx->task)
4767                         pmu = &perf_ops_task_clock;
4768                 else
4769                         pmu = &perf_ops_cpu_clock;
4770
4771                 break;
4772         case PERF_COUNT_SW_PAGE_FAULTS:
4773         case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4774         case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4775         case PERF_COUNT_SW_CONTEXT_SWITCHES:
4776         case PERF_COUNT_SW_CPU_MIGRATIONS:
4777         case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4778         case PERF_COUNT_SW_EMULATION_FAULTS:
4779                 if (!event->parent) {
4780                         int err;
4781
4782                         err = swevent_hlist_get(event);
4783                         if (err)
4784                                 return ERR_PTR(err);
4785
4786                         atomic_inc(&perf_swevent_enabled[event_id]);
4787                         event->destroy = sw_perf_event_destroy;
4788                 }
4789                 pmu = &perf_ops_generic;
4790                 break;
4791         }
4792
4793         return pmu;
4794 }
4795
4796 /*
4797  * Allocate and initialize a event structure
4798  */
4799 static struct perf_event *
4800 perf_event_alloc(struct perf_event_attr *attr,
4801                    int cpu,
4802                    struct perf_event_context *ctx,
4803                    struct perf_event *group_leader,
4804                    struct perf_event *parent_event,
4805                    perf_overflow_handler_t overflow_handler,
4806                    gfp_t gfpflags)
4807 {
4808         const struct pmu *pmu;
4809         struct perf_event *event;
4810         struct hw_perf_event *hwc;
4811         long err;
4812
4813         event = kzalloc(sizeof(*event), gfpflags);
4814         if (!event)
4815                 return ERR_PTR(-ENOMEM);
4816
4817         /*
4818          * Single events are their own group leaders, with an
4819          * empty sibling list:
4820          */
4821         if (!group_leader)
4822                 group_leader = event;
4823
4824         mutex_init(&event->child_mutex);
4825         INIT_LIST_HEAD(&event->child_list);
4826
4827         INIT_LIST_HEAD(&event->group_entry);
4828         INIT_LIST_HEAD(&event->event_entry);
4829         INIT_LIST_HEAD(&event->sibling_list);
4830         init_waitqueue_head(&event->waitq);
4831
4832         mutex_init(&event->mmap_mutex);
4833
4834         event->cpu              = cpu;
4835         event->attr             = *attr;
4836         event->group_leader     = group_leader;
4837         event->pmu              = NULL;
4838         event->ctx              = ctx;
4839         event->oncpu            = -1;
4840
4841         event->parent           = parent_event;
4842
4843         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
4844         event->id               = atomic64_inc_return(&perf_event_id);
4845
4846         event->state            = PERF_EVENT_STATE_INACTIVE;
4847
4848         if (!overflow_handler && parent_event)
4849                 overflow_handler = parent_event->overflow_handler;
4850         
4851         event->overflow_handler = overflow_handler;
4852
4853         if (attr->disabled)
4854                 event->state = PERF_EVENT_STATE_OFF;
4855
4856         pmu = NULL;
4857
4858         hwc = &event->hw;
4859         hwc->sample_period = attr->sample_period;
4860         if (attr->freq && attr->sample_freq)
4861                 hwc->sample_period = 1;
4862         hwc->last_period = hwc->sample_period;
4863
4864         atomic64_set(&hwc->period_left, hwc->sample_period);
4865
4866         /*
4867          * we currently do not support PERF_FORMAT_GROUP on inherited events
4868          */
4869         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4870                 goto done;
4871
4872         switch (attr->type) {
4873         case PERF_TYPE_RAW:
4874         case PERF_TYPE_HARDWARE:
4875         case PERF_TYPE_HW_CACHE:
4876                 pmu = hw_perf_event_init(event);
4877                 break;
4878
4879         case PERF_TYPE_SOFTWARE:
4880                 pmu = sw_perf_event_init(event);
4881                 break;
4882
4883         case PERF_TYPE_TRACEPOINT:
4884                 pmu = tp_perf_event_init(event);
4885                 break;
4886
4887         case PERF_TYPE_BREAKPOINT:
4888                 pmu = bp_perf_event_init(event);
4889                 break;
4890
4891
4892         default:
4893                 break;
4894         }
4895 done:
4896         err = 0;
4897         if (!pmu)
4898                 err = -EINVAL;
4899         else if (IS_ERR(pmu))
4900                 err = PTR_ERR(pmu);
4901
4902         if (err) {
4903                 if (event->ns)
4904                         put_pid_ns(event->ns);
4905                 kfree(event);
4906                 return ERR_PTR(err);
4907         }
4908
4909         event->pmu = pmu;
4910
4911         if (!event->parent) {
4912                 atomic_inc(&nr_events);
4913                 if (event->attr.mmap)
4914                         atomic_inc(&nr_mmap_events);
4915                 if (event->attr.comm)
4916                         atomic_inc(&nr_comm_events);
4917                 if (event->attr.task)
4918                         atomic_inc(&nr_task_events);
4919         }
4920
4921         return event;
4922 }
4923
4924 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4925                           struct perf_event_attr *attr)
4926 {
4927         u32 size;
4928         int ret;
4929
4930         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4931                 return -EFAULT;
4932
4933         /*
4934          * zero the full structure, so that a short copy will be nice.
4935          */
4936         memset(attr, 0, sizeof(*attr));
4937
4938         ret = get_user(size, &uattr->size);
4939         if (ret)
4940                 return ret;
4941
4942         if (size > PAGE_SIZE)   /* silly large */
4943                 goto err_size;
4944
4945         if (!size)              /* abi compat */
4946                 size = PERF_ATTR_SIZE_VER0;
4947
4948         if (size < PERF_ATTR_SIZE_VER0)
4949                 goto err_size;
4950
4951         /*
4952          * If we're handed a bigger struct than we know of,
4953          * ensure all the unknown bits are 0 - i.e. new
4954          * user-space does not rely on any kernel feature
4955          * extensions we dont know about yet.
4956          */
4957         if (size > sizeof(*attr)) {
4958                 unsigned char __user *addr;
4959                 unsigned char __user *end;
4960                 unsigned char val;
4961
4962                 addr = (void __user *)uattr + sizeof(*attr);
4963                 end  = (void __user *)uattr + size;
4964
4965                 for (; addr < end; addr++) {
4966                         ret = get_user(val, addr);
4967                         if (ret)
4968                                 return ret;
4969                         if (val)
4970                                 goto err_size;
4971                 }
4972                 size = sizeof(*attr);
4973         }
4974
4975         ret = copy_from_user(attr, uattr, size);
4976         if (ret)
4977                 return -EFAULT;
4978
4979         /*
4980          * If the type exists, the corresponding creation will verify
4981          * the attr->config.
4982          */
4983         if (attr->type >= PERF_TYPE_MAX)
4984                 return -EINVAL;
4985
4986         if (attr->__reserved_1)
4987                 return -EINVAL;
4988
4989         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4990                 return -EINVAL;
4991
4992         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4993                 return -EINVAL;
4994
4995 out:
4996         return ret;
4997
4998 err_size:
4999         put_user(sizeof(*attr), &uattr->size);
5000         ret = -E2BIG;
5001         goto out;
5002 }
5003
5004 static int
5005 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
5006 {
5007         struct perf_mmap_data *data = NULL, *old_data = NULL;
5008         int ret = -EINVAL;
5009
5010         if (!output_event)
5011                 goto set;
5012
5013         /* don't allow circular references */
5014         if (event == output_event)
5015                 goto out;
5016
5017         /*
5018          * Don't allow cross-cpu buffers
5019          */
5020         if (output_event->cpu != event->cpu)
5021                 goto out;
5022
5023         /*
5024          * If its not a per-cpu buffer, it must be the same task.
5025          */
5026         if (output_event->cpu == -1 && output_event->ctx != event->ctx)
5027                 goto out;
5028
5029 set:
5030         mutex_lock(&event->mmap_mutex);
5031         /* Can't redirect output if we've got an active mmap() */
5032         if (atomic_read(&event->mmap_count))
5033                 goto unlock;
5034
5035         if (output_event) {
5036                 /* get the buffer we want to redirect to */
5037                 data = perf_mmap_data_get(output_event);
5038                 if (!data)
5039                         goto unlock;
5040         }
5041
5042         old_data = event->data;
5043         rcu_assign_pointer(event->data, data);
5044         ret = 0;
5045 unlock:
5046         mutex_unlock(&event->mmap_mutex);
5047
5048         if (old_data)
5049                 perf_mmap_data_put(old_data);
5050 out:
5051         return ret;
5052 }
5053
5054 /**
5055  * sys_perf_event_open - open a performance event, associate it to a task/cpu
5056  *
5057  * @attr_uptr:  event_id type attributes for monitoring/sampling
5058  * @pid:                target pid
5059  * @cpu:                target cpu
5060  * @group_fd:           group leader event fd
5061  */
5062 SYSCALL_DEFINE5(perf_event_open,
5063                 struct perf_event_attr __user *, attr_uptr,
5064                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
5065 {
5066         struct perf_event *event, *group_leader = NULL, *output_event = NULL;
5067         struct perf_event_attr attr;
5068         struct perf_event_context *ctx;
5069         struct file *event_file = NULL;
5070         struct file *group_file = NULL;
5071         int event_fd;
5072         int fput_needed = 0;
5073         int err;
5074
5075         /* for future expandability... */
5076         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
5077                 return -EINVAL;
5078
5079         err = perf_copy_attr(attr_uptr, &attr);
5080         if (err)
5081                 return err;
5082
5083         if (!attr.exclude_kernel) {
5084                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
5085                         return -EACCES;
5086         }
5087
5088         if (attr.freq) {
5089                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
5090                         return -EINVAL;
5091         }
5092
5093         event_fd = get_unused_fd_flags(O_RDWR);
5094         if (event_fd < 0)
5095                 return event_fd;
5096
5097         /*
5098          * Get the target context (task or percpu):
5099          */
5100         ctx = find_get_context(pid, cpu);
5101         if (IS_ERR(ctx)) {
5102                 err = PTR_ERR(ctx);
5103                 goto err_fd;
5104         }
5105
5106         if (group_fd != -1) {
5107                 group_leader = perf_fget_light(group_fd, &fput_needed);
5108                 if (IS_ERR(group_leader)) {
5109                         err = PTR_ERR(group_leader);
5110                         goto err_put_context;
5111                 }
5112                 group_file = group_leader->filp;
5113                 if (flags & PERF_FLAG_FD_OUTPUT)
5114                         output_event = group_leader;
5115                 if (flags & PERF_FLAG_FD_NO_GROUP)
5116                         group_leader = NULL;
5117         }
5118
5119         /*
5120          * Look up the group leader (we will attach this event to it):
5121          */
5122         if (group_leader) {
5123                 err = -EINVAL;
5124
5125                 /*
5126                  * Do not allow a recursive hierarchy (this new sibling
5127                  * becoming part of another group-sibling):
5128                  */
5129                 if (group_leader->group_leader != group_leader)
5130                         goto err_put_context;
5131                 /*
5132                  * Do not allow to attach to a group in a different
5133                  * task or CPU context:
5134                  */
5135                 if (group_leader->ctx != ctx)
5136                         goto err_put_context;
5137                 /*
5138                  * Only a group leader can be exclusive or pinned
5139                  */
5140                 if (attr.exclusive || attr.pinned)
5141                         goto err_put_context;
5142         }
5143
5144         event = perf_event_alloc(&attr, cpu, ctx, group_leader,
5145                                      NULL, NULL, GFP_KERNEL);
5146         if (IS_ERR(event)) {
5147                 err = PTR_ERR(event);
5148                 goto err_put_context;
5149         }
5150
5151         if (output_event) {
5152                 err = perf_event_set_output(event, output_event);
5153                 if (err)
5154                         goto err_free_put_context;
5155         }
5156
5157         event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
5158         if (IS_ERR(event_file)) {
5159                 err = PTR_ERR(event_file);
5160                 goto err_free_put_context;
5161         }
5162
5163         event->filp = event_file;
5164         WARN_ON_ONCE(ctx->parent_ctx);
5165         mutex_lock(&ctx->mutex);
5166         perf_install_in_context(ctx, event, cpu);
5167         ++ctx->generation;
5168         mutex_unlock(&ctx->mutex);
5169
5170         event->owner = current;
5171         get_task_struct(current);
5172         mutex_lock(&current->perf_event_mutex);
5173         list_add_tail(&event->owner_entry, &current->perf_event_list);
5174         mutex_unlock(&current->perf_event_mutex);
5175
5176         /*
5177          * Drop the reference on the group_event after placing the
5178          * new event on the sibling_list. This ensures destruction
5179          * of the group leader will find the pointer to itself in
5180          * perf_group_detach().
5181          */
5182         fput_light(group_file, fput_needed);
5183         fd_install(event_fd, event_file);
5184         return event_fd;
5185
5186 err_free_put_context:
5187         free_event(event);
5188 err_put_context:
5189         fput_light(group_file, fput_needed);
5190         put_ctx(ctx);
5191 err_fd:
5192         put_unused_fd(event_fd);
5193         return err;
5194 }
5195
5196 /**
5197  * perf_event_create_kernel_counter
5198  *
5199  * @attr: attributes of the counter to create
5200  * @cpu: cpu in which the counter is bound
5201  * @pid: task to profile
5202  */
5203 struct perf_event *
5204 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
5205                                  pid_t pid,
5206                                  perf_overflow_handler_t overflow_handler)
5207 {
5208         struct perf_event *event;
5209         struct perf_event_context *ctx;
5210         int err;
5211
5212         /*
5213          * Get the target context (task or percpu):
5214          */
5215
5216         ctx = find_get_context(pid, cpu);
5217         if (IS_ERR(ctx)) {
5218                 err = PTR_ERR(ctx);
5219                 goto err_exit;
5220         }
5221
5222         event = perf_event_alloc(attr, cpu, ctx, NULL,
5223                                  NULL, overflow_handler, GFP_KERNEL);
5224         if (IS_ERR(event)) {
5225                 err = PTR_ERR(event);
5226                 goto err_put_context;
5227         }
5228
5229         event->filp = NULL;
5230         WARN_ON_ONCE(ctx->parent_ctx);
5231         mutex_lock(&ctx->mutex);
5232         perf_install_in_context(ctx, event, cpu);
5233         ++ctx->generation;
5234         mutex_unlock(&ctx->mutex);
5235
5236         event->owner = current;
5237         get_task_struct(current);
5238         mutex_lock(&current->perf_event_mutex);
5239         list_add_tail(&event->owner_entry, &current->perf_event_list);
5240         mutex_unlock(&current->perf_event_mutex);
5241
5242         return event;
5243
5244  err_put_context:
5245         put_ctx(ctx);
5246  err_exit:
5247         return ERR_PTR(err);
5248 }
5249 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
5250
5251 /*
5252  * inherit a event from parent task to child task:
5253  */
5254 static struct perf_event *
5255 inherit_event(struct perf_event *parent_event,
5256               struct task_struct *parent,
5257               struct perf_event_context *parent_ctx,
5258               struct task_struct *child,
5259               struct perf_event *group_leader,
5260               struct perf_event_context *child_ctx)
5261 {
5262         struct perf_event *child_event;
5263
5264         /*
5265          * Instead of creating recursive hierarchies of events,
5266          * we link inherited events back to the original parent,
5267          * which has a filp for sure, which we use as the reference
5268          * count:
5269          */
5270         if (parent_event->parent)
5271                 parent_event = parent_event->parent;
5272
5273         child_event = perf_event_alloc(&parent_event->attr,
5274                                            parent_event->cpu, child_ctx,
5275                                            group_leader, parent_event,
5276                                            NULL, GFP_KERNEL);
5277         if (IS_ERR(child_event))
5278                 return child_event;
5279         get_ctx(child_ctx);
5280
5281         /*
5282          * Make the child state follow the state of the parent event,
5283          * not its attr.disabled bit.  We hold the parent's mutex,
5284          * so we won't race with perf_event_{en, dis}able_family.
5285          */
5286         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
5287                 child_event->state = PERF_EVENT_STATE_INACTIVE;
5288         else
5289                 child_event->state = PERF_EVENT_STATE_OFF;
5290
5291         if (parent_event->attr.freq) {
5292                 u64 sample_period = parent_event->hw.sample_period;
5293                 struct hw_perf_event *hwc = &child_event->hw;
5294
5295                 hwc->sample_period = sample_period;
5296                 hwc->last_period   = sample_period;
5297
5298                 atomic64_set(&hwc->period_left, sample_period);
5299         }
5300
5301         child_event->overflow_handler = parent_event->overflow_handler;
5302
5303         /*
5304          * Link it up in the child's context:
5305          */
5306         add_event_to_ctx(child_event, child_ctx);
5307
5308         /*
5309          * Get a reference to the parent filp - we will fput it
5310          * when the child event exits. This is safe to do because
5311          * we are in the parent and we know that the filp still
5312          * exists and has a nonzero count:
5313          */
5314         atomic_long_inc(&parent_event->filp->f_count);
5315
5316         /*
5317          * Link this into the parent event's child list
5318          */
5319         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5320         mutex_lock(&parent_event->child_mutex);
5321         list_add_tail(&child_event->child_list, &parent_event->child_list);
5322         mutex_unlock(&parent_event->child_mutex);
5323
5324         return child_event;
5325 }
5326
5327 static int inherit_group(struct perf_event *parent_event,
5328               struct task_struct *parent,
5329               struct perf_event_context *parent_ctx,
5330               struct task_struct *child,
5331               struct perf_event_context *child_ctx)
5332 {
5333         struct perf_event *leader;
5334         struct perf_event *sub;
5335         struct perf_event *child_ctr;
5336
5337         leader = inherit_event(parent_event, parent, parent_ctx,
5338                                  child, NULL, child_ctx);
5339         if (IS_ERR(leader))
5340                 return PTR_ERR(leader);
5341         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5342                 child_ctr = inherit_event(sub, parent, parent_ctx,
5343                                             child, leader, child_ctx);
5344                 if (IS_ERR(child_ctr))
5345                         return PTR_ERR(child_ctr);
5346         }
5347         return 0;
5348 }
5349
5350 static void sync_child_event(struct perf_event *child_event,
5351                                struct task_struct *child)
5352 {
5353         struct perf_event *parent_event = child_event->parent;
5354         u64 child_val;
5355
5356         if (child_event->attr.inherit_stat)
5357                 perf_event_read_event(child_event, child);
5358
5359         child_val = atomic64_read(&child_event->count);
5360
5361         /*
5362          * Add back the child's count to the parent's count:
5363          */
5364         atomic64_add(child_val, &parent_event->count);
5365         atomic64_add(child_event->total_time_enabled,
5366                      &parent_event->child_total_time_enabled);
5367         atomic64_add(child_event->total_time_running,
5368                      &parent_event->child_total_time_running);
5369
5370         /*
5371          * Remove this event from the parent's list
5372          */
5373         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5374         mutex_lock(&parent_event->child_mutex);
5375         list_del_init(&child_event->child_list);
5376         mutex_unlock(&parent_event->child_mutex);
5377
5378         /*
5379          * Release the parent event, if this was the last
5380          * reference to it.
5381          */
5382         fput(parent_event->filp);
5383 }
5384
5385 static void
5386 __perf_event_exit_task(struct perf_event *child_event,
5387                          struct perf_event_context *child_ctx,
5388                          struct task_struct *child)
5389 {
5390         struct perf_event *parent_event;
5391
5392         perf_event_remove_from_context(child_event);
5393
5394         parent_event = child_event->parent;
5395         /*
5396          * It can happen that parent exits first, and has events
5397          * that are still around due to the child reference. These
5398          * events need to be zapped - but otherwise linger.
5399          */
5400         if (parent_event) {
5401                 sync_child_event(child_event, child);
5402                 free_event(child_event);
5403         }
5404 }
5405
5406 /*
5407  * When a child task exits, feed back event values to parent events.
5408  */
5409 void perf_event_exit_task(struct task_struct *child)
5410 {
5411         struct perf_event *child_event, *tmp;
5412         struct perf_event_context *child_ctx;
5413         unsigned long flags;
5414
5415         if (likely(!child->perf_event_ctxp)) {
5416                 perf_event_task(child, NULL, 0);
5417                 return;
5418         }
5419
5420         local_irq_save(flags);
5421         /*
5422          * We can't reschedule here because interrupts are disabled,
5423          * and either child is current or it is a task that can't be
5424          * scheduled, so we are now safe from rescheduling changing
5425          * our context.
5426          */
5427         child_ctx = child->perf_event_ctxp;
5428         __perf_event_task_sched_out(child_ctx);
5429
5430         /*
5431          * Take the context lock here so that if find_get_context is
5432          * reading child->perf_event_ctxp, we wait until it has
5433          * incremented the context's refcount before we do put_ctx below.
5434          */
5435         raw_spin_lock(&child_ctx->lock);
5436         child->perf_event_ctxp = NULL;
5437         /*
5438          * If this context is a clone; unclone it so it can't get
5439          * swapped to another process while we're removing all
5440          * the events from it.
5441          */
5442         unclone_ctx(child_ctx);
5443         update_context_time(child_ctx);
5444         raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5445
5446         /*
5447          * Report the task dead after unscheduling the events so that we
5448          * won't get any samples after PERF_RECORD_EXIT. We can however still
5449          * get a few PERF_RECORD_READ events.
5450          */
5451         perf_event_task(child, child_ctx, 0);
5452
5453         /*
5454          * We can recurse on the same lock type through:
5455          *
5456          *   __perf_event_exit_task()
5457          *     sync_child_event()
5458          *       fput(parent_event->filp)
5459          *         perf_release()
5460          *           mutex_lock(&ctx->mutex)
5461          *
5462          * But since its the parent context it won't be the same instance.
5463          */
5464         mutex_lock(&child_ctx->mutex);
5465
5466 again:
5467         list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5468                                  group_entry)
5469                 __perf_event_exit_task(child_event, child_ctx, child);
5470
5471         list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5472                                  group_entry)
5473                 __perf_event_exit_task(child_event, child_ctx, child);
5474
5475         /*
5476          * If the last event was a group event, it will have appended all
5477          * its siblings to the list, but we obtained 'tmp' before that which
5478          * will still point to the list head terminating the iteration.
5479          */
5480         if (!list_empty(&child_ctx->pinned_groups) ||
5481             !list_empty(&child_ctx->flexible_groups))
5482                 goto again;
5483
5484         mutex_unlock(&child_ctx->mutex);
5485
5486         put_ctx(child_ctx);
5487 }
5488
5489 static void perf_free_event(struct perf_event *event,
5490                             struct perf_event_context *ctx)
5491 {
5492         struct perf_event *parent = event->parent;
5493
5494         if (WARN_ON_ONCE(!parent))
5495                 return;
5496
5497         mutex_lock(&parent->child_mutex);
5498         list_del_init(&event->child_list);
5499         mutex_unlock(&parent->child_mutex);
5500
5501         fput(parent->filp);
5502
5503         perf_group_detach(event);
5504         list_del_event(event, ctx);
5505         free_event(event);
5506 }
5507
5508 /*
5509  * free an unexposed, unused context as created by inheritance by
5510  * init_task below, used by fork() in case of fail.
5511  */
5512 void perf_event_free_task(struct task_struct *task)
5513 {
5514         struct perf_event_context *ctx = task->perf_event_ctxp;
5515         struct perf_event *event, *tmp;
5516
5517         if (!ctx)
5518                 return;
5519
5520         mutex_lock(&ctx->mutex);
5521 again:
5522         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5523                 perf_free_event(event, ctx);
5524
5525         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5526                                  group_entry)
5527                 perf_free_event(event, ctx);
5528
5529         if (!list_empty(&ctx->pinned_groups) ||
5530             !list_empty(&ctx->flexible_groups))
5531                 goto again;
5532
5533         mutex_unlock(&ctx->mutex);
5534
5535         put_ctx(ctx);
5536 }
5537
5538 static int
5539 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5540                    struct perf_event_context *parent_ctx,
5541                    struct task_struct *child,
5542                    int *inherited_all)
5543 {
5544         int ret;
5545         struct perf_event_context *child_ctx = child->perf_event_ctxp;
5546
5547         if (!event->attr.inherit) {
5548                 *inherited_all = 0;
5549                 return 0;
5550         }
5551
5552         if (!child_ctx) {
5553                 /*
5554                  * This is executed from the parent task context, so
5555                  * inherit events that have been marked for cloning.
5556                  * First allocate and initialize a context for the
5557                  * child.
5558                  */
5559
5560                 child_ctx = kzalloc(sizeof(struct perf_event_context),
5561                                     GFP_KERNEL);
5562                 if (!child_ctx)
5563                         return -ENOMEM;
5564
5565                 __perf_event_init_context(child_ctx, child);
5566                 child->perf_event_ctxp = child_ctx;
5567                 get_task_struct(child);
5568         }
5569
5570         ret = inherit_group(event, parent, parent_ctx,
5571                             child, child_ctx);
5572
5573         if (ret)
5574                 *inherited_all = 0;
5575
5576         return ret;
5577 }
5578
5579
5580 /*
5581  * Initialize the perf_event context in task_struct
5582  */
5583 int perf_event_init_task(struct task_struct *child)
5584 {
5585         struct perf_event_context *child_ctx, *parent_ctx;
5586         struct perf_event_context *cloned_ctx;
5587         struct perf_event *event;
5588         struct task_struct *parent = current;
5589         int inherited_all = 1;
5590         int ret = 0;
5591
5592         child->perf_event_ctxp = NULL;
5593
5594         mutex_init(&child->perf_event_mutex);
5595         INIT_LIST_HEAD(&child->perf_event_list);
5596
5597         if (likely(!parent->perf_event_ctxp))
5598                 return 0;
5599
5600         /*
5601          * If the parent's context is a clone, pin it so it won't get
5602          * swapped under us.
5603          */
5604         parent_ctx = perf_pin_task_context(parent);
5605
5606         /*
5607          * No need to check if parent_ctx != NULL here; since we saw
5608          * it non-NULL earlier, the only reason for it to become NULL
5609          * is if we exit, and since we're currently in the middle of
5610          * a fork we can't be exiting at the same time.
5611          */
5612
5613         /*
5614          * Lock the parent list. No need to lock the child - not PID
5615          * hashed yet and not running, so nobody can access it.
5616          */
5617         mutex_lock(&parent_ctx->mutex);
5618
5619         /*
5620          * We dont have to disable NMIs - we are only looking at
5621          * the list, not manipulating it:
5622          */
5623         list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5624                 ret = inherit_task_group(event, parent, parent_ctx, child,
5625                                          &inherited_all);
5626                 if (ret)
5627                         break;
5628         }
5629
5630         list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5631                 ret = inherit_task_group(event, parent, parent_ctx, child,
5632                                          &inherited_all);
5633                 if (ret)
5634                         break;
5635         }
5636
5637         child_ctx = child->perf_event_ctxp;
5638
5639         if (child_ctx && inherited_all) {
5640                 /*
5641                  * Mark the child context as a clone of the parent
5642                  * context, or of whatever the parent is a clone of.
5643                  * Note that if the parent is a clone, it could get
5644                  * uncloned at any point, but that doesn't matter
5645                  * because the list of events and the generation
5646                  * count can't have changed since we took the mutex.
5647                  */
5648                 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5649                 if (cloned_ctx) {
5650                         child_ctx->parent_ctx = cloned_ctx;
5651                         child_ctx->parent_gen = parent_ctx->parent_gen;
5652                 } else {
5653                         child_ctx->parent_ctx = parent_ctx;
5654                         child_ctx->parent_gen = parent_ctx->generation;
5655                 }
5656                 get_ctx(child_ctx->parent_ctx);
5657         }
5658
5659         mutex_unlock(&parent_ctx->mutex);
5660
5661         perf_unpin_context(parent_ctx);
5662
5663         return ret;
5664 }
5665
5666 static void __init perf_event_init_all_cpus(void)
5667 {
5668         int cpu;
5669         struct perf_cpu_context *cpuctx;
5670
5671         for_each_possible_cpu(cpu) {
5672                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5673                 mutex_init(&cpuctx->hlist_mutex);
5674                 __perf_event_init_context(&cpuctx->ctx, NULL);
5675         }
5676 }
5677
5678 static void __cpuinit perf_event_init_cpu(int cpu)
5679 {
5680         struct perf_cpu_context *cpuctx;
5681
5682         cpuctx = &per_cpu(perf_cpu_context, cpu);
5683
5684         spin_lock(&perf_resource_lock);
5685         cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5686         spin_unlock(&perf_resource_lock);
5687
5688         mutex_lock(&cpuctx->hlist_mutex);
5689         if (cpuctx->hlist_refcount > 0) {
5690                 struct swevent_hlist *hlist;
5691
5692                 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5693                 WARN_ON_ONCE(!hlist);
5694                 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
5695         }
5696         mutex_unlock(&cpuctx->hlist_mutex);
5697 }
5698
5699 #ifdef CONFIG_HOTPLUG_CPU
5700 static void __perf_event_exit_cpu(void *info)
5701 {
5702         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5703         struct perf_event_context *ctx = &cpuctx->ctx;
5704         struct perf_event *event, *tmp;
5705
5706         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5707                 __perf_event_remove_from_context(event);
5708         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5709                 __perf_event_remove_from_context(event);
5710 }
5711 static void perf_event_exit_cpu(int cpu)
5712 {
5713         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5714         struct perf_event_context *ctx = &cpuctx->ctx;
5715
5716         mutex_lock(&cpuctx->hlist_mutex);
5717         swevent_hlist_release(cpuctx);
5718         mutex_unlock(&cpuctx->hlist_mutex);
5719
5720         mutex_lock(&ctx->mutex);
5721         smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5722         mutex_unlock(&ctx->mutex);
5723 }
5724 #else
5725 static inline void perf_event_exit_cpu(int cpu) { }
5726 #endif
5727
5728 static int __cpuinit
5729 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5730 {
5731         unsigned int cpu = (long)hcpu;
5732
5733         switch (action) {
5734
5735         case CPU_UP_PREPARE:
5736         case CPU_UP_PREPARE_FROZEN:
5737                 perf_event_init_cpu(cpu);
5738                 break;
5739
5740         case CPU_DOWN_PREPARE:
5741         case CPU_DOWN_PREPARE_FROZEN:
5742                 perf_event_exit_cpu(cpu);
5743                 break;
5744
5745         default:
5746                 break;
5747         }
5748
5749         return NOTIFY_OK;
5750 }
5751
5752 /*
5753  * This has to have a higher priority than migration_notifier in sched.c.
5754  */
5755 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5756         .notifier_call          = perf_cpu_notify,
5757         .priority               = 20,
5758 };
5759
5760 void __init perf_event_init(void)
5761 {
5762         perf_event_init_all_cpus();
5763         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5764                         (void *)(long)smp_processor_id());
5765         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5766                         (void *)(long)smp_processor_id());
5767         register_cpu_notifier(&perf_cpu_nb);
5768 }
5769
5770 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class,
5771                                         struct sysdev_class_attribute *attr,
5772                                         char *buf)
5773 {
5774         return sprintf(buf, "%d\n", perf_reserved_percpu);
5775 }
5776
5777 static ssize_t
5778 perf_set_reserve_percpu(struct sysdev_class *class,
5779                         struct sysdev_class_attribute *attr,
5780                         const char *buf,
5781                         size_t count)
5782 {
5783         struct perf_cpu_context *cpuctx;
5784         unsigned long val;
5785         int err, cpu, mpt;
5786
5787         err = strict_strtoul(buf, 10, &val);
5788         if (err)
5789                 return err;
5790         if (val > perf_max_events)
5791                 return -EINVAL;
5792
5793         spin_lock(&perf_resource_lock);
5794         perf_reserved_percpu = val;
5795         for_each_online_cpu(cpu) {
5796                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5797                 raw_spin_lock_irq(&cpuctx->ctx.lock);
5798                 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5799                           perf_max_events - perf_reserved_percpu);
5800                 cpuctx->max_pertask = mpt;
5801                 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5802         }
5803         spin_unlock(&perf_resource_lock);
5804
5805         return count;
5806 }
5807
5808 static ssize_t perf_show_overcommit(struct sysdev_class *class,
5809                                     struct sysdev_class_attribute *attr,
5810                                     char *buf)
5811 {
5812         return sprintf(buf, "%d\n", perf_overcommit);
5813 }
5814
5815 static ssize_t
5816 perf_set_overcommit(struct sysdev_class *class,
5817                     struct sysdev_class_attribute *attr,
5818                     const char *buf, size_t count)
5819 {
5820         unsigned long val;
5821         int err;
5822
5823         err = strict_strtoul(buf, 10, &val);
5824         if (err)
5825                 return err;
5826         if (val > 1)
5827                 return -EINVAL;
5828
5829         spin_lock(&perf_resource_lock);
5830         perf_overcommit = val;
5831         spin_unlock(&perf_resource_lock);
5832
5833         return count;
5834 }
5835
5836 static SYSDEV_CLASS_ATTR(
5837                                 reserve_percpu,
5838                                 0644,
5839                                 perf_show_reserve_percpu,
5840                                 perf_set_reserve_percpu
5841                         );
5842
5843 static SYSDEV_CLASS_ATTR(
5844                                 overcommit,
5845                                 0644,
5846                                 perf_show_overcommit,
5847                                 perf_set_overcommit
5848                         );
5849
5850 static struct attribute *perfclass_attrs[] = {
5851         &attr_reserve_percpu.attr,
5852         &attr_overcommit.attr,
5853         NULL
5854 };
5855
5856 static struct attribute_group perfclass_attr_group = {
5857         .attrs                  = perfclass_attrs,
5858         .name                   = "perf_events",
5859 };
5860
5861 static int __init perf_event_sysfs_init(void)
5862 {
5863         return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5864                                   &perfclass_attr_group);
5865 }
5866 device_initcall(perf_event_sysfs_init);