oprofile: disable write access to oprofilefs while profiler is running
[linux-3.10.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 extern __weak const char *perf_pmu_name(void)
89 {
90         return "pmu";
91 }
92
93 static DEFINE_PER_CPU(int, perf_disable_count);
94
95 void perf_disable(void)
96 {
97         if (!__get_cpu_var(perf_disable_count)++)
98                 hw_perf_disable();
99 }
100
101 void perf_enable(void)
102 {
103         if (!--__get_cpu_var(perf_disable_count))
104                 hw_perf_enable();
105 }
106
107 static void get_ctx(struct perf_event_context *ctx)
108 {
109         WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
110 }
111
112 static void free_ctx(struct rcu_head *head)
113 {
114         struct perf_event_context *ctx;
115
116         ctx = container_of(head, struct perf_event_context, rcu_head);
117         kfree(ctx);
118 }
119
120 static void put_ctx(struct perf_event_context *ctx)
121 {
122         if (atomic_dec_and_test(&ctx->refcount)) {
123                 if (ctx->parent_ctx)
124                         put_ctx(ctx->parent_ctx);
125                 if (ctx->task)
126                         put_task_struct(ctx->task);
127                 call_rcu(&ctx->rcu_head, free_ctx);
128         }
129 }
130
131 static void unclone_ctx(struct perf_event_context *ctx)
132 {
133         if (ctx->parent_ctx) {
134                 put_ctx(ctx->parent_ctx);
135                 ctx->parent_ctx = NULL;
136         }
137 }
138
139 /*
140  * If we inherit events we want to return the parent event id
141  * to userspace.
142  */
143 static u64 primary_event_id(struct perf_event *event)
144 {
145         u64 id = event->id;
146
147         if (event->parent)
148                 id = event->parent->id;
149
150         return id;
151 }
152
153 /*
154  * Get the perf_event_context for a task and lock it.
155  * This has to cope with with the fact that until it is locked,
156  * the context could get moved to another task.
157  */
158 static struct perf_event_context *
159 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
160 {
161         struct perf_event_context *ctx;
162
163         rcu_read_lock();
164  retry:
165         ctx = rcu_dereference(task->perf_event_ctxp);
166         if (ctx) {
167                 /*
168                  * If this context is a clone of another, it might
169                  * get swapped for another underneath us by
170                  * perf_event_task_sched_out, though the
171                  * rcu_read_lock() protects us from any context
172                  * getting freed.  Lock the context and check if it
173                  * got swapped before we could get the lock, and retry
174                  * if so.  If we locked the right context, then it
175                  * can't get swapped on us any more.
176                  */
177                 raw_spin_lock_irqsave(&ctx->lock, *flags);
178                 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
179                         raw_spin_unlock_irqrestore(&ctx->lock, *flags);
180                         goto retry;
181                 }
182
183                 if (!atomic_inc_not_zero(&ctx->refcount)) {
184                         raw_spin_unlock_irqrestore(&ctx->lock, *flags);
185                         ctx = NULL;
186                 }
187         }
188         rcu_read_unlock();
189         return ctx;
190 }
191
192 /*
193  * Get the context for a task and increment its pin_count so it
194  * can't get swapped to another task.  This also increments its
195  * reference count so that the context can't get freed.
196  */
197 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
198 {
199         struct perf_event_context *ctx;
200         unsigned long flags;
201
202         ctx = perf_lock_task_context(task, &flags);
203         if (ctx) {
204                 ++ctx->pin_count;
205                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
206         }
207         return ctx;
208 }
209
210 static void perf_unpin_context(struct perf_event_context *ctx)
211 {
212         unsigned long flags;
213
214         raw_spin_lock_irqsave(&ctx->lock, flags);
215         --ctx->pin_count;
216         raw_spin_unlock_irqrestore(&ctx->lock, flags);
217         put_ctx(ctx);
218 }
219
220 static inline u64 perf_clock(void)
221 {
222         return local_clock();
223 }
224
225 /*
226  * Update the record of the current time in a context.
227  */
228 static void update_context_time(struct perf_event_context *ctx)
229 {
230         u64 now = perf_clock();
231
232         ctx->time += now - ctx->timestamp;
233         ctx->timestamp = now;
234 }
235
236 /*
237  * Update the total_time_enabled and total_time_running fields for a event.
238  */
239 static void update_event_times(struct perf_event *event)
240 {
241         struct perf_event_context *ctx = event->ctx;
242         u64 run_end;
243
244         if (event->state < PERF_EVENT_STATE_INACTIVE ||
245             event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
246                 return;
247
248         if (ctx->is_active)
249                 run_end = ctx->time;
250         else
251                 run_end = event->tstamp_stopped;
252
253         event->total_time_enabled = run_end - event->tstamp_enabled;
254
255         if (event->state == PERF_EVENT_STATE_INACTIVE)
256                 run_end = event->tstamp_stopped;
257         else
258                 run_end = ctx->time;
259
260         event->total_time_running = run_end - event->tstamp_running;
261 }
262
263 /*
264  * Update total_time_enabled and total_time_running for all events in a group.
265  */
266 static void update_group_times(struct perf_event *leader)
267 {
268         struct perf_event *event;
269
270         update_event_times(leader);
271         list_for_each_entry(event, &leader->sibling_list, group_entry)
272                 update_event_times(event);
273 }
274
275 static struct list_head *
276 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
277 {
278         if (event->attr.pinned)
279                 return &ctx->pinned_groups;
280         else
281                 return &ctx->flexible_groups;
282 }
283
284 /*
285  * Add a event from the lists for its context.
286  * Must be called with ctx->mutex and ctx->lock held.
287  */
288 static void
289 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
290 {
291         WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
292         event->attach_state |= PERF_ATTACH_CONTEXT;
293
294         /*
295          * If we're a stand alone event or group leader, we go to the context
296          * list, group events are kept attached to the group so that
297          * perf_group_detach can, at all times, locate all siblings.
298          */
299         if (event->group_leader == event) {
300                 struct list_head *list;
301
302                 if (is_software_event(event))
303                         event->group_flags |= PERF_GROUP_SOFTWARE;
304
305                 list = ctx_group_list(event, ctx);
306                 list_add_tail(&event->group_entry, list);
307         }
308
309         list_add_rcu(&event->event_entry, &ctx->event_list);
310         ctx->nr_events++;
311         if (event->attr.inherit_stat)
312                 ctx->nr_stat++;
313 }
314
315 static void perf_group_attach(struct perf_event *event)
316 {
317         struct perf_event *group_leader = event->group_leader;
318
319         WARN_ON_ONCE(event->attach_state & PERF_ATTACH_GROUP);
320         event->attach_state |= PERF_ATTACH_GROUP;
321
322         if (group_leader == event)
323                 return;
324
325         if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
326                         !is_software_event(event))
327                 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
328
329         list_add_tail(&event->group_entry, &group_leader->sibling_list);
330         group_leader->nr_siblings++;
331 }
332
333 /*
334  * Remove a event from the lists for its context.
335  * Must be called with ctx->mutex and ctx->lock held.
336  */
337 static void
338 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
339 {
340         /*
341          * We can have double detach due to exit/hot-unplug + close.
342          */
343         if (!(event->attach_state & PERF_ATTACH_CONTEXT))
344                 return;
345
346         event->attach_state &= ~PERF_ATTACH_CONTEXT;
347
348         ctx->nr_events--;
349         if (event->attr.inherit_stat)
350                 ctx->nr_stat--;
351
352         list_del_rcu(&event->event_entry);
353
354         if (event->group_leader == event)
355                 list_del_init(&event->group_entry);
356
357         update_group_times(event);
358
359         /*
360          * If event was in error state, then keep it
361          * that way, otherwise bogus counts will be
362          * returned on read(). The only way to get out
363          * of error state is by explicit re-enabling
364          * of the event
365          */
366         if (event->state > PERF_EVENT_STATE_OFF)
367                 event->state = PERF_EVENT_STATE_OFF;
368 }
369
370 static void perf_group_detach(struct perf_event *event)
371 {
372         struct perf_event *sibling, *tmp;
373         struct list_head *list = NULL;
374
375         /*
376          * We can have double detach due to exit/hot-unplug + close.
377          */
378         if (!(event->attach_state & PERF_ATTACH_GROUP))
379                 return;
380
381         event->attach_state &= ~PERF_ATTACH_GROUP;
382
383         /*
384          * If this is a sibling, remove it from its group.
385          */
386         if (event->group_leader != event) {
387                 list_del_init(&event->group_entry);
388                 event->group_leader->nr_siblings--;
389                 return;
390         }
391
392         if (!list_empty(&event->group_entry))
393                 list = &event->group_entry;
394
395         /*
396          * If this was a group event with sibling events then
397          * upgrade the siblings to singleton events by adding them
398          * to whatever list we are on.
399          */
400         list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
401                 if (list)
402                         list_move_tail(&sibling->group_entry, list);
403                 sibling->group_leader = sibling;
404
405                 /* Inherit group flags from the previous leader */
406                 sibling->group_flags = event->group_flags;
407         }
408 }
409
410 static inline int
411 event_filter_match(struct perf_event *event)
412 {
413         return event->cpu == -1 || event->cpu == smp_processor_id();
414 }
415
416 static void
417 event_sched_out(struct perf_event *event,
418                   struct perf_cpu_context *cpuctx,
419                   struct perf_event_context *ctx)
420 {
421         u64 delta;
422         /*
423          * An event which could not be activated because of
424          * filter mismatch still needs to have its timings
425          * maintained, otherwise bogus information is return
426          * via read() for time_enabled, time_running:
427          */
428         if (event->state == PERF_EVENT_STATE_INACTIVE
429             && !event_filter_match(event)) {
430                 delta = ctx->time - event->tstamp_stopped;
431                 event->tstamp_running += delta;
432                 event->tstamp_stopped = ctx->time;
433         }
434
435         if (event->state != PERF_EVENT_STATE_ACTIVE)
436                 return;
437
438         event->state = PERF_EVENT_STATE_INACTIVE;
439         if (event->pending_disable) {
440                 event->pending_disable = 0;
441                 event->state = PERF_EVENT_STATE_OFF;
442         }
443         event->tstamp_stopped = ctx->time;
444         event->pmu->disable(event);
445         event->oncpu = -1;
446
447         if (!is_software_event(event))
448                 cpuctx->active_oncpu--;
449         ctx->nr_active--;
450         if (event->attr.exclusive || !cpuctx->active_oncpu)
451                 cpuctx->exclusive = 0;
452 }
453
454 static void
455 group_sched_out(struct perf_event *group_event,
456                 struct perf_cpu_context *cpuctx,
457                 struct perf_event_context *ctx)
458 {
459         struct perf_event *event;
460         int state = group_event->state;
461
462         event_sched_out(group_event, cpuctx, ctx);
463
464         /*
465          * Schedule out siblings (if any):
466          */
467         list_for_each_entry(event, &group_event->sibling_list, group_entry)
468                 event_sched_out(event, cpuctx, ctx);
469
470         if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
471                 cpuctx->exclusive = 0;
472 }
473
474 /*
475  * Cross CPU call to remove a performance event
476  *
477  * We disable the event on the hardware level first. After that we
478  * remove it from the context list.
479  */
480 static void __perf_event_remove_from_context(void *info)
481 {
482         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
483         struct perf_event *event = info;
484         struct perf_event_context *ctx = event->ctx;
485
486         /*
487          * If this is a task context, we need to check whether it is
488          * the current task context of this cpu. If not it has been
489          * scheduled out before the smp call arrived.
490          */
491         if (ctx->task && cpuctx->task_ctx != ctx)
492                 return;
493
494         raw_spin_lock(&ctx->lock);
495         /*
496          * Protect the list operation against NMI by disabling the
497          * events on a global level.
498          */
499         perf_disable();
500
501         event_sched_out(event, cpuctx, ctx);
502
503         list_del_event(event, ctx);
504
505         if (!ctx->task) {
506                 /*
507                  * Allow more per task events with respect to the
508                  * reservation:
509                  */
510                 cpuctx->max_pertask =
511                         min(perf_max_events - ctx->nr_events,
512                             perf_max_events - perf_reserved_percpu);
513         }
514
515         perf_enable();
516         raw_spin_unlock(&ctx->lock);
517 }
518
519
520 /*
521  * Remove the event from a task's (or a CPU's) list of events.
522  *
523  * Must be called with ctx->mutex held.
524  *
525  * CPU events are removed with a smp call. For task events we only
526  * call when the task is on a CPU.
527  *
528  * If event->ctx is a cloned context, callers must make sure that
529  * every task struct that event->ctx->task could possibly point to
530  * remains valid.  This is OK when called from perf_release since
531  * that only calls us on the top-level context, which can't be a clone.
532  * When called from perf_event_exit_task, it's OK because the
533  * context has been detached from its task.
534  */
535 static void perf_event_remove_from_context(struct perf_event *event)
536 {
537         struct perf_event_context *ctx = event->ctx;
538         struct task_struct *task = ctx->task;
539
540         if (!task) {
541                 /*
542                  * Per cpu events are removed via an smp call and
543                  * the removal is always successful.
544                  */
545                 smp_call_function_single(event->cpu,
546                                          __perf_event_remove_from_context,
547                                          event, 1);
548                 return;
549         }
550
551 retry:
552         task_oncpu_function_call(task, __perf_event_remove_from_context,
553                                  event);
554
555         raw_spin_lock_irq(&ctx->lock);
556         /*
557          * If the context is active we need to retry the smp call.
558          */
559         if (ctx->nr_active && !list_empty(&event->group_entry)) {
560                 raw_spin_unlock_irq(&ctx->lock);
561                 goto retry;
562         }
563
564         /*
565          * The lock prevents that this context is scheduled in so we
566          * can remove the event safely, if the call above did not
567          * succeed.
568          */
569         if (!list_empty(&event->group_entry))
570                 list_del_event(event, ctx);
571         raw_spin_unlock_irq(&ctx->lock);
572 }
573
574 /*
575  * Cross CPU call to disable a performance event
576  */
577 static void __perf_event_disable(void *info)
578 {
579         struct perf_event *event = info;
580         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
581         struct perf_event_context *ctx = event->ctx;
582
583         /*
584          * If this is a per-task event, need to check whether this
585          * event's task is the current task on this cpu.
586          */
587         if (ctx->task && cpuctx->task_ctx != ctx)
588                 return;
589
590         raw_spin_lock(&ctx->lock);
591
592         /*
593          * If the event is on, turn it off.
594          * If it is in error state, leave it in error state.
595          */
596         if (event->state >= PERF_EVENT_STATE_INACTIVE) {
597                 update_context_time(ctx);
598                 update_group_times(event);
599                 if (event == event->group_leader)
600                         group_sched_out(event, cpuctx, ctx);
601                 else
602                         event_sched_out(event, cpuctx, ctx);
603                 event->state = PERF_EVENT_STATE_OFF;
604         }
605
606         raw_spin_unlock(&ctx->lock);
607 }
608
609 /*
610  * Disable a event.
611  *
612  * If event->ctx is a cloned context, callers must make sure that
613  * every task struct that event->ctx->task could possibly point to
614  * remains valid.  This condition is satisifed when called through
615  * perf_event_for_each_child or perf_event_for_each because they
616  * hold the top-level event's child_mutex, so any descendant that
617  * goes to exit will block in sync_child_event.
618  * When called from perf_pending_event it's OK because event->ctx
619  * is the current context on this CPU and preemption is disabled,
620  * hence we can't get into perf_event_task_sched_out for this context.
621  */
622 void perf_event_disable(struct perf_event *event)
623 {
624         struct perf_event_context *ctx = event->ctx;
625         struct task_struct *task = ctx->task;
626
627         if (!task) {
628                 /*
629                  * Disable the event on the cpu that it's on
630                  */
631                 smp_call_function_single(event->cpu, __perf_event_disable,
632                                          event, 1);
633                 return;
634         }
635
636  retry:
637         task_oncpu_function_call(task, __perf_event_disable, event);
638
639         raw_spin_lock_irq(&ctx->lock);
640         /*
641          * If the event is still active, we need to retry the cross-call.
642          */
643         if (event->state == PERF_EVENT_STATE_ACTIVE) {
644                 raw_spin_unlock_irq(&ctx->lock);
645                 goto retry;
646         }
647
648         /*
649          * Since we have the lock this context can't be scheduled
650          * in, so we can change the state safely.
651          */
652         if (event->state == PERF_EVENT_STATE_INACTIVE) {
653                 update_group_times(event);
654                 event->state = PERF_EVENT_STATE_OFF;
655         }
656
657         raw_spin_unlock_irq(&ctx->lock);
658 }
659
660 static int
661 event_sched_in(struct perf_event *event,
662                  struct perf_cpu_context *cpuctx,
663                  struct perf_event_context *ctx)
664 {
665         if (event->state <= PERF_EVENT_STATE_OFF)
666                 return 0;
667
668         event->state = PERF_EVENT_STATE_ACTIVE;
669         event->oncpu = smp_processor_id();
670         /*
671          * The new state must be visible before we turn it on in the hardware:
672          */
673         smp_wmb();
674
675         if (event->pmu->enable(event)) {
676                 event->state = PERF_EVENT_STATE_INACTIVE;
677                 event->oncpu = -1;
678                 return -EAGAIN;
679         }
680
681         event->tstamp_running += ctx->time - event->tstamp_stopped;
682
683         if (!is_software_event(event))
684                 cpuctx->active_oncpu++;
685         ctx->nr_active++;
686
687         if (event->attr.exclusive)
688                 cpuctx->exclusive = 1;
689
690         return 0;
691 }
692
693 static int
694 group_sched_in(struct perf_event *group_event,
695                struct perf_cpu_context *cpuctx,
696                struct perf_event_context *ctx)
697 {
698         struct perf_event *event, *partial_group = NULL;
699         const struct pmu *pmu = group_event->pmu;
700         bool txn = false;
701
702         if (group_event->state == PERF_EVENT_STATE_OFF)
703                 return 0;
704
705         /* Check if group transaction availabe */
706         if (pmu->start_txn)
707                 txn = true;
708
709         if (txn)
710                 pmu->start_txn(pmu);
711
712         if (event_sched_in(group_event, cpuctx, ctx)) {
713                 if (txn)
714                         pmu->cancel_txn(pmu);
715                 return -EAGAIN;
716         }
717
718         /*
719          * Schedule in siblings as one group (if any):
720          */
721         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
722                 if (event_sched_in(event, cpuctx, ctx)) {
723                         partial_group = event;
724                         goto group_error;
725                 }
726         }
727
728         if (!txn || !pmu->commit_txn(pmu))
729                 return 0;
730
731 group_error:
732         /*
733          * Groups can be scheduled in as one unit only, so undo any
734          * partial group before returning:
735          */
736         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
737                 if (event == partial_group)
738                         break;
739                 event_sched_out(event, cpuctx, ctx);
740         }
741         event_sched_out(group_event, cpuctx, ctx);
742
743         if (txn)
744                 pmu->cancel_txn(pmu);
745
746         return -EAGAIN;
747 }
748
749 /*
750  * Work out whether we can put this event group on the CPU now.
751  */
752 static int group_can_go_on(struct perf_event *event,
753                            struct perf_cpu_context *cpuctx,
754                            int can_add_hw)
755 {
756         /*
757          * Groups consisting entirely of software events can always go on.
758          */
759         if (event->group_flags & PERF_GROUP_SOFTWARE)
760                 return 1;
761         /*
762          * If an exclusive group is already on, no other hardware
763          * events can go on.
764          */
765         if (cpuctx->exclusive)
766                 return 0;
767         /*
768          * If this group is exclusive and there are already
769          * events on the CPU, it can't go on.
770          */
771         if (event->attr.exclusive && cpuctx->active_oncpu)
772                 return 0;
773         /*
774          * Otherwise, try to add it if all previous groups were able
775          * to go on.
776          */
777         return can_add_hw;
778 }
779
780 static void add_event_to_ctx(struct perf_event *event,
781                                struct perf_event_context *ctx)
782 {
783         list_add_event(event, ctx);
784         perf_group_attach(event);
785         event->tstamp_enabled = ctx->time;
786         event->tstamp_running = ctx->time;
787         event->tstamp_stopped = ctx->time;
788 }
789
790 /*
791  * Cross CPU call to install and enable a performance event
792  *
793  * Must be called with ctx->mutex held
794  */
795 static void __perf_install_in_context(void *info)
796 {
797         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
798         struct perf_event *event = info;
799         struct perf_event_context *ctx = event->ctx;
800         struct perf_event *leader = event->group_leader;
801         int err;
802
803         /*
804          * If this is a task context, we need to check whether it is
805          * the current task context of this cpu. If not it has been
806          * scheduled out before the smp call arrived.
807          * Or possibly this is the right context but it isn't
808          * on this cpu because it had no events.
809          */
810         if (ctx->task && cpuctx->task_ctx != ctx) {
811                 if (cpuctx->task_ctx || ctx->task != current)
812                         return;
813                 cpuctx->task_ctx = ctx;
814         }
815
816         raw_spin_lock(&ctx->lock);
817         ctx->is_active = 1;
818         update_context_time(ctx);
819
820         /*
821          * Protect the list operation against NMI by disabling the
822          * events on a global level. NOP for non NMI based events.
823          */
824         perf_disable();
825
826         add_event_to_ctx(event, ctx);
827
828         if (event->cpu != -1 && event->cpu != smp_processor_id())
829                 goto unlock;
830
831         /*
832          * Don't put the event on if it is disabled or if
833          * it is in a group and the group isn't on.
834          */
835         if (event->state != PERF_EVENT_STATE_INACTIVE ||
836             (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
837                 goto unlock;
838
839         /*
840          * An exclusive event can't go on if there are already active
841          * hardware events, and no hardware event can go on if there
842          * is already an exclusive event on.
843          */
844         if (!group_can_go_on(event, cpuctx, 1))
845                 err = -EEXIST;
846         else
847                 err = event_sched_in(event, cpuctx, ctx);
848
849         if (err) {
850                 /*
851                  * This event couldn't go on.  If it is in a group
852                  * then we have to pull the whole group off.
853                  * If the event group is pinned then put it in error state.
854                  */
855                 if (leader != event)
856                         group_sched_out(leader, cpuctx, ctx);
857                 if (leader->attr.pinned) {
858                         update_group_times(leader);
859                         leader->state = PERF_EVENT_STATE_ERROR;
860                 }
861         }
862
863         if (!err && !ctx->task && cpuctx->max_pertask)
864                 cpuctx->max_pertask--;
865
866  unlock:
867         perf_enable();
868
869         raw_spin_unlock(&ctx->lock);
870 }
871
872 /*
873  * Attach a performance event to a context
874  *
875  * First we add the event to the list with the hardware enable bit
876  * in event->hw_config cleared.
877  *
878  * If the event is attached to a task which is on a CPU we use a smp
879  * call to enable it in the task context. The task might have been
880  * scheduled away, but we check this in the smp call again.
881  *
882  * Must be called with ctx->mutex held.
883  */
884 static void
885 perf_install_in_context(struct perf_event_context *ctx,
886                         struct perf_event *event,
887                         int cpu)
888 {
889         struct task_struct *task = ctx->task;
890
891         if (!task) {
892                 /*
893                  * Per cpu events are installed via an smp call and
894                  * the install is always successful.
895                  */
896                 smp_call_function_single(cpu, __perf_install_in_context,
897                                          event, 1);
898                 return;
899         }
900
901 retry:
902         task_oncpu_function_call(task, __perf_install_in_context,
903                                  event);
904
905         raw_spin_lock_irq(&ctx->lock);
906         /*
907          * we need to retry the smp call.
908          */
909         if (ctx->is_active && list_empty(&event->group_entry)) {
910                 raw_spin_unlock_irq(&ctx->lock);
911                 goto retry;
912         }
913
914         /*
915          * The lock prevents that this context is scheduled in so we
916          * can add the event safely, if it the call above did not
917          * succeed.
918          */
919         if (list_empty(&event->group_entry))
920                 add_event_to_ctx(event, ctx);
921         raw_spin_unlock_irq(&ctx->lock);
922 }
923
924 /*
925  * Put a event into inactive state and update time fields.
926  * Enabling the leader of a group effectively enables all
927  * the group members that aren't explicitly disabled, so we
928  * have to update their ->tstamp_enabled also.
929  * Note: this works for group members as well as group leaders
930  * since the non-leader members' sibling_lists will be empty.
931  */
932 static void __perf_event_mark_enabled(struct perf_event *event,
933                                         struct perf_event_context *ctx)
934 {
935         struct perf_event *sub;
936
937         event->state = PERF_EVENT_STATE_INACTIVE;
938         event->tstamp_enabled = ctx->time - event->total_time_enabled;
939         list_for_each_entry(sub, &event->sibling_list, group_entry)
940                 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
941                         sub->tstamp_enabled =
942                                 ctx->time - sub->total_time_enabled;
943 }
944
945 /*
946  * Cross CPU call to enable a performance event
947  */
948 static void __perf_event_enable(void *info)
949 {
950         struct perf_event *event = info;
951         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
952         struct perf_event_context *ctx = event->ctx;
953         struct perf_event *leader = event->group_leader;
954         int err;
955
956         /*
957          * If this is a per-task event, need to check whether this
958          * event's task is the current task on this cpu.
959          */
960         if (ctx->task && cpuctx->task_ctx != ctx) {
961                 if (cpuctx->task_ctx || ctx->task != current)
962                         return;
963                 cpuctx->task_ctx = ctx;
964         }
965
966         raw_spin_lock(&ctx->lock);
967         ctx->is_active = 1;
968         update_context_time(ctx);
969
970         if (event->state >= PERF_EVENT_STATE_INACTIVE)
971                 goto unlock;
972         __perf_event_mark_enabled(event, ctx);
973
974         if (event->cpu != -1 && event->cpu != smp_processor_id())
975                 goto unlock;
976
977         /*
978          * If the event is in a group and isn't the group leader,
979          * then don't put it on unless the group is on.
980          */
981         if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
982                 goto unlock;
983
984         if (!group_can_go_on(event, cpuctx, 1)) {
985                 err = -EEXIST;
986         } else {
987                 perf_disable();
988                 if (event == leader)
989                         err = group_sched_in(event, cpuctx, ctx);
990                 else
991                         err = event_sched_in(event, cpuctx, ctx);
992                 perf_enable();
993         }
994
995         if (err) {
996                 /*
997                  * If this event can't go on and it's part of a
998                  * group, then the whole group has to come off.
999                  */
1000                 if (leader != event)
1001                         group_sched_out(leader, cpuctx, ctx);
1002                 if (leader->attr.pinned) {
1003                         update_group_times(leader);
1004                         leader->state = PERF_EVENT_STATE_ERROR;
1005                 }
1006         }
1007
1008  unlock:
1009         raw_spin_unlock(&ctx->lock);
1010 }
1011
1012 /*
1013  * Enable a event.
1014  *
1015  * If event->ctx is a cloned context, callers must make sure that
1016  * every task struct that event->ctx->task could possibly point to
1017  * remains valid.  This condition is satisfied when called through
1018  * perf_event_for_each_child or perf_event_for_each as described
1019  * for perf_event_disable.
1020  */
1021 void perf_event_enable(struct perf_event *event)
1022 {
1023         struct perf_event_context *ctx = event->ctx;
1024         struct task_struct *task = ctx->task;
1025
1026         if (!task) {
1027                 /*
1028                  * Enable the event on the cpu that it's on
1029                  */
1030                 smp_call_function_single(event->cpu, __perf_event_enable,
1031                                          event, 1);
1032                 return;
1033         }
1034
1035         raw_spin_lock_irq(&ctx->lock);
1036         if (event->state >= PERF_EVENT_STATE_INACTIVE)
1037                 goto out;
1038
1039         /*
1040          * If the event is in error state, clear that first.
1041          * That way, if we see the event in error state below, we
1042          * know that it has gone back into error state, as distinct
1043          * from the task having been scheduled away before the
1044          * cross-call arrived.
1045          */
1046         if (event->state == PERF_EVENT_STATE_ERROR)
1047                 event->state = PERF_EVENT_STATE_OFF;
1048
1049  retry:
1050         raw_spin_unlock_irq(&ctx->lock);
1051         task_oncpu_function_call(task, __perf_event_enable, event);
1052
1053         raw_spin_lock_irq(&ctx->lock);
1054
1055         /*
1056          * If the context is active and the event is still off,
1057          * we need to retry the cross-call.
1058          */
1059         if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1060                 goto retry;
1061
1062         /*
1063          * Since we have the lock this context can't be scheduled
1064          * in, so we can change the state safely.
1065          */
1066         if (event->state == PERF_EVENT_STATE_OFF)
1067                 __perf_event_mark_enabled(event, ctx);
1068
1069  out:
1070         raw_spin_unlock_irq(&ctx->lock);
1071 }
1072
1073 static int perf_event_refresh(struct perf_event *event, int refresh)
1074 {
1075         /*
1076          * not supported on inherited events
1077          */
1078         if (event->attr.inherit)
1079                 return -EINVAL;
1080
1081         atomic_add(refresh, &event->event_limit);
1082         perf_event_enable(event);
1083
1084         return 0;
1085 }
1086
1087 enum event_type_t {
1088         EVENT_FLEXIBLE = 0x1,
1089         EVENT_PINNED = 0x2,
1090         EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1091 };
1092
1093 static void ctx_sched_out(struct perf_event_context *ctx,
1094                           struct perf_cpu_context *cpuctx,
1095                           enum event_type_t event_type)
1096 {
1097         struct perf_event *event;
1098
1099         raw_spin_lock(&ctx->lock);
1100         ctx->is_active = 0;
1101         if (likely(!ctx->nr_events))
1102                 goto out;
1103         update_context_time(ctx);
1104
1105         perf_disable();
1106         if (!ctx->nr_active)
1107                 goto out_enable;
1108
1109         if (event_type & EVENT_PINNED)
1110                 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1111                         group_sched_out(event, cpuctx, ctx);
1112
1113         if (event_type & EVENT_FLEXIBLE)
1114                 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1115                         group_sched_out(event, cpuctx, ctx);
1116
1117  out_enable:
1118         perf_enable();
1119  out:
1120         raw_spin_unlock(&ctx->lock);
1121 }
1122
1123 /*
1124  * Test whether two contexts are equivalent, i.e. whether they
1125  * have both been cloned from the same version of the same context
1126  * and they both have the same number of enabled events.
1127  * If the number of enabled events is the same, then the set
1128  * of enabled events should be the same, because these are both
1129  * inherited contexts, therefore we can't access individual events
1130  * in them directly with an fd; we can only enable/disable all
1131  * events via prctl, or enable/disable all events in a family
1132  * via ioctl, which will have the same effect on both contexts.
1133  */
1134 static int context_equiv(struct perf_event_context *ctx1,
1135                          struct perf_event_context *ctx2)
1136 {
1137         return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1138                 && ctx1->parent_gen == ctx2->parent_gen
1139                 && !ctx1->pin_count && !ctx2->pin_count;
1140 }
1141
1142 static void __perf_event_sync_stat(struct perf_event *event,
1143                                      struct perf_event *next_event)
1144 {
1145         u64 value;
1146
1147         if (!event->attr.inherit_stat)
1148                 return;
1149
1150         /*
1151          * Update the event value, we cannot use perf_event_read()
1152          * because we're in the middle of a context switch and have IRQs
1153          * disabled, which upsets smp_call_function_single(), however
1154          * we know the event must be on the current CPU, therefore we
1155          * don't need to use it.
1156          */
1157         switch (event->state) {
1158         case PERF_EVENT_STATE_ACTIVE:
1159                 event->pmu->read(event);
1160                 /* fall-through */
1161
1162         case PERF_EVENT_STATE_INACTIVE:
1163                 update_event_times(event);
1164                 break;
1165
1166         default:
1167                 break;
1168         }
1169
1170         /*
1171          * In order to keep per-task stats reliable we need to flip the event
1172          * values when we flip the contexts.
1173          */
1174         value = local64_read(&next_event->count);
1175         value = local64_xchg(&event->count, value);
1176         local64_set(&next_event->count, value);
1177
1178         swap(event->total_time_enabled, next_event->total_time_enabled);
1179         swap(event->total_time_running, next_event->total_time_running);
1180
1181         /*
1182          * Since we swizzled the values, update the user visible data too.
1183          */
1184         perf_event_update_userpage(event);
1185         perf_event_update_userpage(next_event);
1186 }
1187
1188 #define list_next_entry(pos, member) \
1189         list_entry(pos->member.next, typeof(*pos), member)
1190
1191 static void perf_event_sync_stat(struct perf_event_context *ctx,
1192                                    struct perf_event_context *next_ctx)
1193 {
1194         struct perf_event *event, *next_event;
1195
1196         if (!ctx->nr_stat)
1197                 return;
1198
1199         update_context_time(ctx);
1200
1201         event = list_first_entry(&ctx->event_list,
1202                                    struct perf_event, event_entry);
1203
1204         next_event = list_first_entry(&next_ctx->event_list,
1205                                         struct perf_event, event_entry);
1206
1207         while (&event->event_entry != &ctx->event_list &&
1208                &next_event->event_entry != &next_ctx->event_list) {
1209
1210                 __perf_event_sync_stat(event, next_event);
1211
1212                 event = list_next_entry(event, event_entry);
1213                 next_event = list_next_entry(next_event, event_entry);
1214         }
1215 }
1216
1217 /*
1218  * Called from scheduler to remove the events of the current task,
1219  * with interrupts disabled.
1220  *
1221  * We stop each event and update the event value in event->count.
1222  *
1223  * This does not protect us against NMI, but disable()
1224  * sets the disabled bit in the control field of event _before_
1225  * accessing the event control register. If a NMI hits, then it will
1226  * not restart the event.
1227  */
1228 void perf_event_task_sched_out(struct task_struct *task,
1229                                  struct task_struct *next)
1230 {
1231         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1232         struct perf_event_context *ctx = task->perf_event_ctxp;
1233         struct perf_event_context *next_ctx;
1234         struct perf_event_context *parent;
1235         int do_switch = 1;
1236
1237         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0);
1238
1239         if (likely(!ctx || !cpuctx->task_ctx))
1240                 return;
1241
1242         rcu_read_lock();
1243         parent = rcu_dereference(ctx->parent_ctx);
1244         next_ctx = next->perf_event_ctxp;
1245         if (parent && next_ctx &&
1246             rcu_dereference(next_ctx->parent_ctx) == parent) {
1247                 /*
1248                  * Looks like the two contexts are clones, so we might be
1249                  * able to optimize the context switch.  We lock both
1250                  * contexts and check that they are clones under the
1251                  * lock (including re-checking that neither has been
1252                  * uncloned in the meantime).  It doesn't matter which
1253                  * order we take the locks because no other cpu could
1254                  * be trying to lock both of these tasks.
1255                  */
1256                 raw_spin_lock(&ctx->lock);
1257                 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1258                 if (context_equiv(ctx, next_ctx)) {
1259                         /*
1260                          * XXX do we need a memory barrier of sorts
1261                          * wrt to rcu_dereference() of perf_event_ctxp
1262                          */
1263                         task->perf_event_ctxp = next_ctx;
1264                         next->perf_event_ctxp = ctx;
1265                         ctx->task = next;
1266                         next_ctx->task = task;
1267                         do_switch = 0;
1268
1269                         perf_event_sync_stat(ctx, next_ctx);
1270                 }
1271                 raw_spin_unlock(&next_ctx->lock);
1272                 raw_spin_unlock(&ctx->lock);
1273         }
1274         rcu_read_unlock();
1275
1276         if (do_switch) {
1277                 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1278                 cpuctx->task_ctx = NULL;
1279         }
1280 }
1281
1282 static void task_ctx_sched_out(struct perf_event_context *ctx,
1283                                enum event_type_t event_type)
1284 {
1285         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1286
1287         if (!cpuctx->task_ctx)
1288                 return;
1289
1290         if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1291                 return;
1292
1293         ctx_sched_out(ctx, cpuctx, event_type);
1294         cpuctx->task_ctx = NULL;
1295 }
1296
1297 /*
1298  * Called with IRQs disabled
1299  */
1300 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1301 {
1302         task_ctx_sched_out(ctx, EVENT_ALL);
1303 }
1304
1305 /*
1306  * Called with IRQs disabled
1307  */
1308 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1309                               enum event_type_t event_type)
1310 {
1311         ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1312 }
1313
1314 static void
1315 ctx_pinned_sched_in(struct perf_event_context *ctx,
1316                     struct perf_cpu_context *cpuctx)
1317 {
1318         struct perf_event *event;
1319
1320         list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1321                 if (event->state <= PERF_EVENT_STATE_OFF)
1322                         continue;
1323                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1324                         continue;
1325
1326                 if (group_can_go_on(event, cpuctx, 1))
1327                         group_sched_in(event, cpuctx, ctx);
1328
1329                 /*
1330                  * If this pinned group hasn't been scheduled,
1331                  * put it in error state.
1332                  */
1333                 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1334                         update_group_times(event);
1335                         event->state = PERF_EVENT_STATE_ERROR;
1336                 }
1337         }
1338 }
1339
1340 static void
1341 ctx_flexible_sched_in(struct perf_event_context *ctx,
1342                       struct perf_cpu_context *cpuctx)
1343 {
1344         struct perf_event *event;
1345         int can_add_hw = 1;
1346
1347         list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1348                 /* Ignore events in OFF or ERROR state */
1349                 if (event->state <= PERF_EVENT_STATE_OFF)
1350                         continue;
1351                 /*
1352                  * Listen to the 'cpu' scheduling filter constraint
1353                  * of events:
1354                  */
1355                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1356                         continue;
1357
1358                 if (group_can_go_on(event, cpuctx, can_add_hw))
1359                         if (group_sched_in(event, cpuctx, ctx))
1360                                 can_add_hw = 0;
1361         }
1362 }
1363
1364 static void
1365 ctx_sched_in(struct perf_event_context *ctx,
1366              struct perf_cpu_context *cpuctx,
1367              enum event_type_t event_type)
1368 {
1369         raw_spin_lock(&ctx->lock);
1370         ctx->is_active = 1;
1371         if (likely(!ctx->nr_events))
1372                 goto out;
1373
1374         ctx->timestamp = perf_clock();
1375
1376         perf_disable();
1377
1378         /*
1379          * First go through the list and put on any pinned groups
1380          * in order to give them the best chance of going on.
1381          */
1382         if (event_type & EVENT_PINNED)
1383                 ctx_pinned_sched_in(ctx, cpuctx);
1384
1385         /* Then walk through the lower prio flexible groups */
1386         if (event_type & EVENT_FLEXIBLE)
1387                 ctx_flexible_sched_in(ctx, cpuctx);
1388
1389         perf_enable();
1390  out:
1391         raw_spin_unlock(&ctx->lock);
1392 }
1393
1394 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1395                              enum event_type_t event_type)
1396 {
1397         struct perf_event_context *ctx = &cpuctx->ctx;
1398
1399         ctx_sched_in(ctx, cpuctx, event_type);
1400 }
1401
1402 static void task_ctx_sched_in(struct task_struct *task,
1403                               enum event_type_t event_type)
1404 {
1405         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1406         struct perf_event_context *ctx = task->perf_event_ctxp;
1407
1408         if (likely(!ctx))
1409                 return;
1410         if (cpuctx->task_ctx == ctx)
1411                 return;
1412         ctx_sched_in(ctx, cpuctx, event_type);
1413         cpuctx->task_ctx = ctx;
1414 }
1415 /*
1416  * Called from scheduler to add the events of the current task
1417  * with interrupts disabled.
1418  *
1419  * We restore the event value and then enable it.
1420  *
1421  * This does not protect us against NMI, but enable()
1422  * sets the enabled bit in the control field of event _before_
1423  * accessing the event control register. If a NMI hits, then it will
1424  * keep the event running.
1425  */
1426 void perf_event_task_sched_in(struct task_struct *task)
1427 {
1428         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1429         struct perf_event_context *ctx = task->perf_event_ctxp;
1430
1431         if (likely(!ctx))
1432                 return;
1433
1434         if (cpuctx->task_ctx == ctx)
1435                 return;
1436
1437         perf_disable();
1438
1439         /*
1440          * We want to keep the following priority order:
1441          * cpu pinned (that don't need to move), task pinned,
1442          * cpu flexible, task flexible.
1443          */
1444         cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1445
1446         ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1447         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1448         ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1449
1450         cpuctx->task_ctx = ctx;
1451
1452         perf_enable();
1453 }
1454
1455 #define MAX_INTERRUPTS (~0ULL)
1456
1457 static void perf_log_throttle(struct perf_event *event, int enable);
1458
1459 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1460 {
1461         u64 frequency = event->attr.sample_freq;
1462         u64 sec = NSEC_PER_SEC;
1463         u64 divisor, dividend;
1464
1465         int count_fls, nsec_fls, frequency_fls, sec_fls;
1466
1467         count_fls = fls64(count);
1468         nsec_fls = fls64(nsec);
1469         frequency_fls = fls64(frequency);
1470         sec_fls = 30;
1471
1472         /*
1473          * We got @count in @nsec, with a target of sample_freq HZ
1474          * the target period becomes:
1475          *
1476          *             @count * 10^9
1477          * period = -------------------
1478          *          @nsec * sample_freq
1479          *
1480          */
1481
1482         /*
1483          * Reduce accuracy by one bit such that @a and @b converge
1484          * to a similar magnitude.
1485          */
1486 #define REDUCE_FLS(a, b)                \
1487 do {                                    \
1488         if (a##_fls > b##_fls) {        \
1489                 a >>= 1;                \
1490                 a##_fls--;              \
1491         } else {                        \
1492                 b >>= 1;                \
1493                 b##_fls--;              \
1494         }                               \
1495 } while (0)
1496
1497         /*
1498          * Reduce accuracy until either term fits in a u64, then proceed with
1499          * the other, so that finally we can do a u64/u64 division.
1500          */
1501         while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1502                 REDUCE_FLS(nsec, frequency);
1503                 REDUCE_FLS(sec, count);
1504         }
1505
1506         if (count_fls + sec_fls > 64) {
1507                 divisor = nsec * frequency;
1508
1509                 while (count_fls + sec_fls > 64) {
1510                         REDUCE_FLS(count, sec);
1511                         divisor >>= 1;
1512                 }
1513
1514                 dividend = count * sec;
1515         } else {
1516                 dividend = count * sec;
1517
1518                 while (nsec_fls + frequency_fls > 64) {
1519                         REDUCE_FLS(nsec, frequency);
1520                         dividend >>= 1;
1521                 }
1522
1523                 divisor = nsec * frequency;
1524         }
1525
1526         if (!divisor)
1527                 return dividend;
1528
1529         return div64_u64(dividend, divisor);
1530 }
1531
1532 static void perf_event_stop(struct perf_event *event)
1533 {
1534         if (!event->pmu->stop)
1535                 return event->pmu->disable(event);
1536
1537         return event->pmu->stop(event);
1538 }
1539
1540 static int perf_event_start(struct perf_event *event)
1541 {
1542         if (!event->pmu->start)
1543                 return event->pmu->enable(event);
1544
1545         return event->pmu->start(event);
1546 }
1547
1548 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1549 {
1550         struct hw_perf_event *hwc = &event->hw;
1551         s64 period, sample_period;
1552         s64 delta;
1553
1554         period = perf_calculate_period(event, nsec, count);
1555
1556         delta = (s64)(period - hwc->sample_period);
1557         delta = (delta + 7) / 8; /* low pass filter */
1558
1559         sample_period = hwc->sample_period + delta;
1560
1561         if (!sample_period)
1562                 sample_period = 1;
1563
1564         hwc->sample_period = sample_period;
1565
1566         if (local64_read(&hwc->period_left) > 8*sample_period) {
1567                 perf_disable();
1568                 perf_event_stop(event);
1569                 local64_set(&hwc->period_left, 0);
1570                 perf_event_start(event);
1571                 perf_enable();
1572         }
1573 }
1574
1575 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1576 {
1577         struct perf_event *event;
1578         struct hw_perf_event *hwc;
1579         u64 interrupts, now;
1580         s64 delta;
1581
1582         raw_spin_lock(&ctx->lock);
1583         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1584                 if (event->state != PERF_EVENT_STATE_ACTIVE)
1585                         continue;
1586
1587                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1588                         continue;
1589
1590                 hwc = &event->hw;
1591
1592                 interrupts = hwc->interrupts;
1593                 hwc->interrupts = 0;
1594
1595                 /*
1596                  * unthrottle events on the tick
1597                  */
1598                 if (interrupts == MAX_INTERRUPTS) {
1599                         perf_log_throttle(event, 1);
1600                         perf_disable();
1601                         event->pmu->unthrottle(event);
1602                         perf_enable();
1603                 }
1604
1605                 if (!event->attr.freq || !event->attr.sample_freq)
1606                         continue;
1607
1608                 perf_disable();
1609                 event->pmu->read(event);
1610                 now = local64_read(&event->count);
1611                 delta = now - hwc->freq_count_stamp;
1612                 hwc->freq_count_stamp = now;
1613
1614                 if (delta > 0)
1615                         perf_adjust_period(event, TICK_NSEC, delta);
1616                 perf_enable();
1617         }
1618         raw_spin_unlock(&ctx->lock);
1619 }
1620
1621 /*
1622  * Round-robin a context's events:
1623  */
1624 static void rotate_ctx(struct perf_event_context *ctx)
1625 {
1626         raw_spin_lock(&ctx->lock);
1627
1628         /* Rotate the first entry last of non-pinned groups */
1629         list_rotate_left(&ctx->flexible_groups);
1630
1631         raw_spin_unlock(&ctx->lock);
1632 }
1633
1634 void perf_event_task_tick(struct task_struct *curr)
1635 {
1636         struct perf_cpu_context *cpuctx;
1637         struct perf_event_context *ctx;
1638         int rotate = 0;
1639
1640         if (!atomic_read(&nr_events))
1641                 return;
1642
1643         cpuctx = &__get_cpu_var(perf_cpu_context);
1644         if (cpuctx->ctx.nr_events &&
1645             cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1646                 rotate = 1;
1647
1648         ctx = curr->perf_event_ctxp;
1649         if (ctx && ctx->nr_events && ctx->nr_events != ctx->nr_active)
1650                 rotate = 1;
1651
1652         perf_ctx_adjust_freq(&cpuctx->ctx);
1653         if (ctx)
1654                 perf_ctx_adjust_freq(ctx);
1655
1656         if (!rotate)
1657                 return;
1658
1659         perf_disable();
1660         cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1661         if (ctx)
1662                 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1663
1664         rotate_ctx(&cpuctx->ctx);
1665         if (ctx)
1666                 rotate_ctx(ctx);
1667
1668         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1669         if (ctx)
1670                 task_ctx_sched_in(curr, EVENT_FLEXIBLE);
1671         perf_enable();
1672 }
1673
1674 static int event_enable_on_exec(struct perf_event *event,
1675                                 struct perf_event_context *ctx)
1676 {
1677         if (!event->attr.enable_on_exec)
1678                 return 0;
1679
1680         event->attr.enable_on_exec = 0;
1681         if (event->state >= PERF_EVENT_STATE_INACTIVE)
1682                 return 0;
1683
1684         __perf_event_mark_enabled(event, ctx);
1685
1686         return 1;
1687 }
1688
1689 /*
1690  * Enable all of a task's events that have been marked enable-on-exec.
1691  * This expects task == current.
1692  */
1693 static void perf_event_enable_on_exec(struct task_struct *task)
1694 {
1695         struct perf_event_context *ctx;
1696         struct perf_event *event;
1697         unsigned long flags;
1698         int enabled = 0;
1699         int ret;
1700
1701         local_irq_save(flags);
1702         ctx = task->perf_event_ctxp;
1703         if (!ctx || !ctx->nr_events)
1704                 goto out;
1705
1706         __perf_event_task_sched_out(ctx);
1707
1708         raw_spin_lock(&ctx->lock);
1709
1710         list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1711                 ret = event_enable_on_exec(event, ctx);
1712                 if (ret)
1713                         enabled = 1;
1714         }
1715
1716         list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1717                 ret = event_enable_on_exec(event, ctx);
1718                 if (ret)
1719                         enabled = 1;
1720         }
1721
1722         /*
1723          * Unclone this context if we enabled any event.
1724          */
1725         if (enabled)
1726                 unclone_ctx(ctx);
1727
1728         raw_spin_unlock(&ctx->lock);
1729
1730         perf_event_task_sched_in(task);
1731  out:
1732         local_irq_restore(flags);
1733 }
1734
1735 /*
1736  * Cross CPU call to read the hardware event
1737  */
1738 static void __perf_event_read(void *info)
1739 {
1740         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1741         struct perf_event *event = info;
1742         struct perf_event_context *ctx = event->ctx;
1743
1744         /*
1745          * If this is a task context, we need to check whether it is
1746          * the current task context of this cpu.  If not it has been
1747          * scheduled out before the smp call arrived.  In that case
1748          * event->count would have been updated to a recent sample
1749          * when the event was scheduled out.
1750          */
1751         if (ctx->task && cpuctx->task_ctx != ctx)
1752                 return;
1753
1754         raw_spin_lock(&ctx->lock);
1755         update_context_time(ctx);
1756         update_event_times(event);
1757         raw_spin_unlock(&ctx->lock);
1758
1759         event->pmu->read(event);
1760 }
1761
1762 static inline u64 perf_event_count(struct perf_event *event)
1763 {
1764         return local64_read(&event->count) + atomic64_read(&event->child_count);
1765 }
1766
1767 static u64 perf_event_read(struct perf_event *event)
1768 {
1769         /*
1770          * If event is enabled and currently active on a CPU, update the
1771          * value in the event structure:
1772          */
1773         if (event->state == PERF_EVENT_STATE_ACTIVE) {
1774                 smp_call_function_single(event->oncpu,
1775                                          __perf_event_read, event, 1);
1776         } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1777                 struct perf_event_context *ctx = event->ctx;
1778                 unsigned long flags;
1779
1780                 raw_spin_lock_irqsave(&ctx->lock, flags);
1781                 update_context_time(ctx);
1782                 update_event_times(event);
1783                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1784         }
1785
1786         return perf_event_count(event);
1787 }
1788
1789 /*
1790  * Initialize the perf_event context in a task_struct:
1791  */
1792 static void
1793 __perf_event_init_context(struct perf_event_context *ctx,
1794                             struct task_struct *task)
1795 {
1796         raw_spin_lock_init(&ctx->lock);
1797         mutex_init(&ctx->mutex);
1798         INIT_LIST_HEAD(&ctx->pinned_groups);
1799         INIT_LIST_HEAD(&ctx->flexible_groups);
1800         INIT_LIST_HEAD(&ctx->event_list);
1801         atomic_set(&ctx->refcount, 1);
1802         ctx->task = task;
1803 }
1804
1805 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1806 {
1807         struct perf_event_context *ctx;
1808         struct perf_cpu_context *cpuctx;
1809         struct task_struct *task;
1810         unsigned long flags;
1811         int err;
1812
1813         if (pid == -1 && cpu != -1) {
1814                 /* Must be root to operate on a CPU event: */
1815                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1816                         return ERR_PTR(-EACCES);
1817
1818                 if (cpu < 0 || cpu >= nr_cpumask_bits)
1819                         return ERR_PTR(-EINVAL);
1820
1821                 /*
1822                  * We could be clever and allow to attach a event to an
1823                  * offline CPU and activate it when the CPU comes up, but
1824                  * that's for later.
1825                  */
1826                 if (!cpu_online(cpu))
1827                         return ERR_PTR(-ENODEV);
1828
1829                 cpuctx = &per_cpu(perf_cpu_context, cpu);
1830                 ctx = &cpuctx->ctx;
1831                 get_ctx(ctx);
1832
1833                 return ctx;
1834         }
1835
1836         rcu_read_lock();
1837         if (!pid)
1838                 task = current;
1839         else
1840                 task = find_task_by_vpid(pid);
1841         if (task)
1842                 get_task_struct(task);
1843         rcu_read_unlock();
1844
1845         if (!task)
1846                 return ERR_PTR(-ESRCH);
1847
1848         /*
1849          * Can't attach events to a dying task.
1850          */
1851         err = -ESRCH;
1852         if (task->flags & PF_EXITING)
1853                 goto errout;
1854
1855         /* Reuse ptrace permission checks for now. */
1856         err = -EACCES;
1857         if (!ptrace_may_access(task, PTRACE_MODE_READ))
1858                 goto errout;
1859
1860  retry:
1861         ctx = perf_lock_task_context(task, &flags);
1862         if (ctx) {
1863                 unclone_ctx(ctx);
1864                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1865         }
1866
1867         if (!ctx) {
1868                 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1869                 err = -ENOMEM;
1870                 if (!ctx)
1871                         goto errout;
1872                 __perf_event_init_context(ctx, task);
1873                 get_ctx(ctx);
1874                 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1875                         /*
1876                          * We raced with some other task; use
1877                          * the context they set.
1878                          */
1879                         kfree(ctx);
1880                         goto retry;
1881                 }
1882                 get_task_struct(task);
1883         }
1884
1885         put_task_struct(task);
1886         return ctx;
1887
1888  errout:
1889         put_task_struct(task);
1890         return ERR_PTR(err);
1891 }
1892
1893 static void perf_event_free_filter(struct perf_event *event);
1894
1895 static void free_event_rcu(struct rcu_head *head)
1896 {
1897         struct perf_event *event;
1898
1899         event = container_of(head, struct perf_event, rcu_head);
1900         if (event->ns)
1901                 put_pid_ns(event->ns);
1902         perf_event_free_filter(event);
1903         kfree(event);
1904 }
1905
1906 static void perf_pending_sync(struct perf_event *event);
1907 static void perf_buffer_put(struct perf_buffer *buffer);
1908
1909 static void free_event(struct perf_event *event)
1910 {
1911         perf_pending_sync(event);
1912
1913         if (!event->parent) {
1914                 atomic_dec(&nr_events);
1915                 if (event->attr.mmap || event->attr.mmap_data)
1916                         atomic_dec(&nr_mmap_events);
1917                 if (event->attr.comm)
1918                         atomic_dec(&nr_comm_events);
1919                 if (event->attr.task)
1920                         atomic_dec(&nr_task_events);
1921         }
1922
1923         if (event->buffer) {
1924                 perf_buffer_put(event->buffer);
1925                 event->buffer = NULL;
1926         }
1927
1928         if (event->destroy)
1929                 event->destroy(event);
1930
1931         put_ctx(event->ctx);
1932         call_rcu(&event->rcu_head, free_event_rcu);
1933 }
1934
1935 int perf_event_release_kernel(struct perf_event *event)
1936 {
1937         struct perf_event_context *ctx = event->ctx;
1938
1939         /*
1940          * Remove from the PMU, can't get re-enabled since we got
1941          * here because the last ref went.
1942          */
1943         perf_event_disable(event);
1944
1945         WARN_ON_ONCE(ctx->parent_ctx);
1946         /*
1947          * There are two ways this annotation is useful:
1948          *
1949          *  1) there is a lock recursion from perf_event_exit_task
1950          *     see the comment there.
1951          *
1952          *  2) there is a lock-inversion with mmap_sem through
1953          *     perf_event_read_group(), which takes faults while
1954          *     holding ctx->mutex, however this is called after
1955          *     the last filedesc died, so there is no possibility
1956          *     to trigger the AB-BA case.
1957          */
1958         mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
1959         raw_spin_lock_irq(&ctx->lock);
1960         perf_group_detach(event);
1961         list_del_event(event, ctx);
1962         raw_spin_unlock_irq(&ctx->lock);
1963         mutex_unlock(&ctx->mutex);
1964
1965         mutex_lock(&event->owner->perf_event_mutex);
1966         list_del_init(&event->owner_entry);
1967         mutex_unlock(&event->owner->perf_event_mutex);
1968         put_task_struct(event->owner);
1969
1970         free_event(event);
1971
1972         return 0;
1973 }
1974 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1975
1976 /*
1977  * Called when the last reference to the file is gone.
1978  */
1979 static int perf_release(struct inode *inode, struct file *file)
1980 {
1981         struct perf_event *event = file->private_data;
1982
1983         file->private_data = NULL;
1984
1985         return perf_event_release_kernel(event);
1986 }
1987
1988 static int perf_event_read_size(struct perf_event *event)
1989 {
1990         int entry = sizeof(u64); /* value */
1991         int size = 0;
1992         int nr = 1;
1993
1994         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1995                 size += sizeof(u64);
1996
1997         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1998                 size += sizeof(u64);
1999
2000         if (event->attr.read_format & PERF_FORMAT_ID)
2001                 entry += sizeof(u64);
2002
2003         if (event->attr.read_format & PERF_FORMAT_GROUP) {
2004                 nr += event->group_leader->nr_siblings;
2005                 size += sizeof(u64);
2006         }
2007
2008         size += entry * nr;
2009
2010         return size;
2011 }
2012
2013 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
2014 {
2015         struct perf_event *child;
2016         u64 total = 0;
2017
2018         *enabled = 0;
2019         *running = 0;
2020
2021         mutex_lock(&event->child_mutex);
2022         total += perf_event_read(event);
2023         *enabled += event->total_time_enabled +
2024                         atomic64_read(&event->child_total_time_enabled);
2025         *running += event->total_time_running +
2026                         atomic64_read(&event->child_total_time_running);
2027
2028         list_for_each_entry(child, &event->child_list, child_list) {
2029                 total += perf_event_read(child);
2030                 *enabled += child->total_time_enabled;
2031                 *running += child->total_time_running;
2032         }
2033         mutex_unlock(&event->child_mutex);
2034
2035         return total;
2036 }
2037 EXPORT_SYMBOL_GPL(perf_event_read_value);
2038
2039 static int perf_event_read_group(struct perf_event *event,
2040                                    u64 read_format, char __user *buf)
2041 {
2042         struct perf_event *leader = event->group_leader, *sub;
2043         int n = 0, size = 0, ret = -EFAULT;
2044         struct perf_event_context *ctx = leader->ctx;
2045         u64 values[5];
2046         u64 count, enabled, running;
2047
2048         mutex_lock(&ctx->mutex);
2049         count = perf_event_read_value(leader, &enabled, &running);
2050
2051         values[n++] = 1 + leader->nr_siblings;
2052         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2053                 values[n++] = enabled;
2054         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2055                 values[n++] = running;
2056         values[n++] = count;
2057         if (read_format & PERF_FORMAT_ID)
2058                 values[n++] = primary_event_id(leader);
2059
2060         size = n * sizeof(u64);
2061
2062         if (copy_to_user(buf, values, size))
2063                 goto unlock;
2064
2065         ret = size;
2066
2067         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2068                 n = 0;
2069
2070                 values[n++] = perf_event_read_value(sub, &enabled, &running);
2071                 if (read_format & PERF_FORMAT_ID)
2072                         values[n++] = primary_event_id(sub);
2073
2074                 size = n * sizeof(u64);
2075
2076                 if (copy_to_user(buf + ret, values, size)) {
2077                         ret = -EFAULT;
2078                         goto unlock;
2079                 }
2080
2081                 ret += size;
2082         }
2083 unlock:
2084         mutex_unlock(&ctx->mutex);
2085
2086         return ret;
2087 }
2088
2089 static int perf_event_read_one(struct perf_event *event,
2090                                  u64 read_format, char __user *buf)
2091 {
2092         u64 enabled, running;
2093         u64 values[4];
2094         int n = 0;
2095
2096         values[n++] = perf_event_read_value(event, &enabled, &running);
2097         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2098                 values[n++] = enabled;
2099         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2100                 values[n++] = running;
2101         if (read_format & PERF_FORMAT_ID)
2102                 values[n++] = primary_event_id(event);
2103
2104         if (copy_to_user(buf, values, n * sizeof(u64)))
2105                 return -EFAULT;
2106
2107         return n * sizeof(u64);
2108 }
2109
2110 /*
2111  * Read the performance event - simple non blocking version for now
2112  */
2113 static ssize_t
2114 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2115 {
2116         u64 read_format = event->attr.read_format;
2117         int ret;
2118
2119         /*
2120          * Return end-of-file for a read on a event that is in
2121          * error state (i.e. because it was pinned but it couldn't be
2122          * scheduled on to the CPU at some point).
2123          */
2124         if (event->state == PERF_EVENT_STATE_ERROR)
2125                 return 0;
2126
2127         if (count < perf_event_read_size(event))
2128                 return -ENOSPC;
2129
2130         WARN_ON_ONCE(event->ctx->parent_ctx);
2131         if (read_format & PERF_FORMAT_GROUP)
2132                 ret = perf_event_read_group(event, read_format, buf);
2133         else
2134                 ret = perf_event_read_one(event, read_format, buf);
2135
2136         return ret;
2137 }
2138
2139 static ssize_t
2140 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2141 {
2142         struct perf_event *event = file->private_data;
2143
2144         return perf_read_hw(event, buf, count);
2145 }
2146
2147 static unsigned int perf_poll(struct file *file, poll_table *wait)
2148 {
2149         struct perf_event *event = file->private_data;
2150         struct perf_buffer *buffer;
2151         unsigned int events = POLL_HUP;
2152
2153         rcu_read_lock();
2154         buffer = rcu_dereference(event->buffer);
2155         if (buffer)
2156                 events = atomic_xchg(&buffer->poll, 0);
2157         rcu_read_unlock();
2158
2159         poll_wait(file, &event->waitq, wait);
2160
2161         return events;
2162 }
2163
2164 static void perf_event_reset(struct perf_event *event)
2165 {
2166         (void)perf_event_read(event);
2167         local64_set(&event->count, 0);
2168         perf_event_update_userpage(event);
2169 }
2170
2171 /*
2172  * Holding the top-level event's child_mutex means that any
2173  * descendant process that has inherited this event will block
2174  * in sync_child_event if it goes to exit, thus satisfying the
2175  * task existence requirements of perf_event_enable/disable.
2176  */
2177 static void perf_event_for_each_child(struct perf_event *event,
2178                                         void (*func)(struct perf_event *))
2179 {
2180         struct perf_event *child;
2181
2182         WARN_ON_ONCE(event->ctx->parent_ctx);
2183         mutex_lock(&event->child_mutex);
2184         func(event);
2185         list_for_each_entry(child, &event->child_list, child_list)
2186                 func(child);
2187         mutex_unlock(&event->child_mutex);
2188 }
2189
2190 static void perf_event_for_each(struct perf_event *event,
2191                                   void (*func)(struct perf_event *))
2192 {
2193         struct perf_event_context *ctx = event->ctx;
2194         struct perf_event *sibling;
2195
2196         WARN_ON_ONCE(ctx->parent_ctx);
2197         mutex_lock(&ctx->mutex);
2198         event = event->group_leader;
2199
2200         perf_event_for_each_child(event, func);
2201         func(event);
2202         list_for_each_entry(sibling, &event->sibling_list, group_entry)
2203                 perf_event_for_each_child(event, func);
2204         mutex_unlock(&ctx->mutex);
2205 }
2206
2207 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2208 {
2209         struct perf_event_context *ctx = event->ctx;
2210         unsigned long size;
2211         int ret = 0;
2212         u64 value;
2213
2214         if (!event->attr.sample_period)
2215                 return -EINVAL;
2216
2217         size = copy_from_user(&value, arg, sizeof(value));
2218         if (size != sizeof(value))
2219                 return -EFAULT;
2220
2221         if (!value)
2222                 return -EINVAL;
2223
2224         raw_spin_lock_irq(&ctx->lock);
2225         if (event->attr.freq) {
2226                 if (value > sysctl_perf_event_sample_rate) {
2227                         ret = -EINVAL;
2228                         goto unlock;
2229                 }
2230
2231                 event->attr.sample_freq = value;
2232         } else {
2233                 event->attr.sample_period = value;
2234                 event->hw.sample_period = value;
2235         }
2236 unlock:
2237         raw_spin_unlock_irq(&ctx->lock);
2238
2239         return ret;
2240 }
2241
2242 static const struct file_operations perf_fops;
2243
2244 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
2245 {
2246         struct file *file;
2247
2248         file = fget_light(fd, fput_needed);
2249         if (!file)
2250                 return ERR_PTR(-EBADF);
2251
2252         if (file->f_op != &perf_fops) {
2253                 fput_light(file, *fput_needed);
2254                 *fput_needed = 0;
2255                 return ERR_PTR(-EBADF);
2256         }
2257
2258         return file->private_data;
2259 }
2260
2261 static int perf_event_set_output(struct perf_event *event,
2262                                  struct perf_event *output_event);
2263 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2264
2265 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2266 {
2267         struct perf_event *event = file->private_data;
2268         void (*func)(struct perf_event *);
2269         u32 flags = arg;
2270
2271         switch (cmd) {
2272         case PERF_EVENT_IOC_ENABLE:
2273                 func = perf_event_enable;
2274                 break;
2275         case PERF_EVENT_IOC_DISABLE:
2276                 func = perf_event_disable;
2277                 break;
2278         case PERF_EVENT_IOC_RESET:
2279                 func = perf_event_reset;
2280                 break;
2281
2282         case PERF_EVENT_IOC_REFRESH:
2283                 return perf_event_refresh(event, arg);
2284
2285         case PERF_EVENT_IOC_PERIOD:
2286                 return perf_event_period(event, (u64 __user *)arg);
2287
2288         case PERF_EVENT_IOC_SET_OUTPUT:
2289         {
2290                 struct perf_event *output_event = NULL;
2291                 int fput_needed = 0;
2292                 int ret;
2293
2294                 if (arg != -1) {
2295                         output_event = perf_fget_light(arg, &fput_needed);
2296                         if (IS_ERR(output_event))
2297                                 return PTR_ERR(output_event);
2298                 }
2299
2300                 ret = perf_event_set_output(event, output_event);
2301                 if (output_event)
2302                         fput_light(output_event->filp, fput_needed);
2303
2304                 return ret;
2305         }
2306
2307         case PERF_EVENT_IOC_SET_FILTER:
2308                 return perf_event_set_filter(event, (void __user *)arg);
2309
2310         default:
2311                 return -ENOTTY;
2312         }
2313
2314         if (flags & PERF_IOC_FLAG_GROUP)
2315                 perf_event_for_each(event, func);
2316         else
2317                 perf_event_for_each_child(event, func);
2318
2319         return 0;
2320 }
2321
2322 int perf_event_task_enable(void)
2323 {
2324         struct perf_event *event;
2325
2326         mutex_lock(&current->perf_event_mutex);
2327         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2328                 perf_event_for_each_child(event, perf_event_enable);
2329         mutex_unlock(&current->perf_event_mutex);
2330
2331         return 0;
2332 }
2333
2334 int perf_event_task_disable(void)
2335 {
2336         struct perf_event *event;
2337
2338         mutex_lock(&current->perf_event_mutex);
2339         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2340                 perf_event_for_each_child(event, perf_event_disable);
2341         mutex_unlock(&current->perf_event_mutex);
2342
2343         return 0;
2344 }
2345
2346 #ifndef PERF_EVENT_INDEX_OFFSET
2347 # define PERF_EVENT_INDEX_OFFSET 0
2348 #endif
2349
2350 static int perf_event_index(struct perf_event *event)
2351 {
2352         if (event->state != PERF_EVENT_STATE_ACTIVE)
2353                 return 0;
2354
2355         return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2356 }
2357
2358 /*
2359  * Callers need to ensure there can be no nesting of this function, otherwise
2360  * the seqlock logic goes bad. We can not serialize this because the arch
2361  * code calls this from NMI context.
2362  */
2363 void perf_event_update_userpage(struct perf_event *event)
2364 {
2365         struct perf_event_mmap_page *userpg;
2366         struct perf_buffer *buffer;
2367
2368         rcu_read_lock();
2369         buffer = rcu_dereference(event->buffer);
2370         if (!buffer)
2371                 goto unlock;
2372
2373         userpg = buffer->user_page;
2374
2375         /*
2376          * Disable preemption so as to not let the corresponding user-space
2377          * spin too long if we get preempted.
2378          */
2379         preempt_disable();
2380         ++userpg->lock;
2381         barrier();
2382         userpg->index = perf_event_index(event);
2383         userpg->offset = perf_event_count(event);
2384         if (event->state == PERF_EVENT_STATE_ACTIVE)
2385                 userpg->offset -= local64_read(&event->hw.prev_count);
2386
2387         userpg->time_enabled = event->total_time_enabled +
2388                         atomic64_read(&event->child_total_time_enabled);
2389
2390         userpg->time_running = event->total_time_running +
2391                         atomic64_read(&event->child_total_time_running);
2392
2393         barrier();
2394         ++userpg->lock;
2395         preempt_enable();
2396 unlock:
2397         rcu_read_unlock();
2398 }
2399
2400 static unsigned long perf_data_size(struct perf_buffer *buffer);
2401
2402 static void
2403 perf_buffer_init(struct perf_buffer *buffer, long watermark, int flags)
2404 {
2405         long max_size = perf_data_size(buffer);
2406
2407         if (watermark)
2408                 buffer->watermark = min(max_size, watermark);
2409
2410         if (!buffer->watermark)
2411                 buffer->watermark = max_size / 2;
2412
2413         if (flags & PERF_BUFFER_WRITABLE)
2414                 buffer->writable = 1;
2415
2416         atomic_set(&buffer->refcount, 1);
2417 }
2418
2419 #ifndef CONFIG_PERF_USE_VMALLOC
2420
2421 /*
2422  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2423  */
2424
2425 static struct page *
2426 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2427 {
2428         if (pgoff > buffer->nr_pages)
2429                 return NULL;
2430
2431         if (pgoff == 0)
2432                 return virt_to_page(buffer->user_page);
2433
2434         return virt_to_page(buffer->data_pages[pgoff - 1]);
2435 }
2436
2437 static void *perf_mmap_alloc_page(int cpu)
2438 {
2439         struct page *page;
2440         int node;
2441
2442         node = (cpu == -1) ? cpu : cpu_to_node(cpu);
2443         page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
2444         if (!page)
2445                 return NULL;
2446
2447         return page_address(page);
2448 }
2449
2450 static struct perf_buffer *
2451 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2452 {
2453         struct perf_buffer *buffer;
2454         unsigned long size;
2455         int i;
2456
2457         size = sizeof(struct perf_buffer);
2458         size += nr_pages * sizeof(void *);
2459
2460         buffer = kzalloc(size, GFP_KERNEL);
2461         if (!buffer)
2462                 goto fail;
2463
2464         buffer->user_page = perf_mmap_alloc_page(cpu);
2465         if (!buffer->user_page)
2466                 goto fail_user_page;
2467
2468         for (i = 0; i < nr_pages; i++) {
2469                 buffer->data_pages[i] = perf_mmap_alloc_page(cpu);
2470                 if (!buffer->data_pages[i])
2471                         goto fail_data_pages;
2472         }
2473
2474         buffer->nr_pages = nr_pages;
2475
2476         perf_buffer_init(buffer, watermark, flags);
2477
2478         return buffer;
2479
2480 fail_data_pages:
2481         for (i--; i >= 0; i--)
2482                 free_page((unsigned long)buffer->data_pages[i]);
2483
2484         free_page((unsigned long)buffer->user_page);
2485
2486 fail_user_page:
2487         kfree(buffer);
2488
2489 fail:
2490         return NULL;
2491 }
2492
2493 static void perf_mmap_free_page(unsigned long addr)
2494 {
2495         struct page *page = virt_to_page((void *)addr);
2496
2497         page->mapping = NULL;
2498         __free_page(page);
2499 }
2500
2501 static void perf_buffer_free(struct perf_buffer *buffer)
2502 {
2503         int i;
2504
2505         perf_mmap_free_page((unsigned long)buffer->user_page);
2506         for (i = 0; i < buffer->nr_pages; i++)
2507                 perf_mmap_free_page((unsigned long)buffer->data_pages[i]);
2508         kfree(buffer);
2509 }
2510
2511 static inline int page_order(struct perf_buffer *buffer)
2512 {
2513         return 0;
2514 }
2515
2516 #else
2517
2518 /*
2519  * Back perf_mmap() with vmalloc memory.
2520  *
2521  * Required for architectures that have d-cache aliasing issues.
2522  */
2523
2524 static inline int page_order(struct perf_buffer *buffer)
2525 {
2526         return buffer->page_order;
2527 }
2528
2529 static struct page *
2530 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2531 {
2532         if (pgoff > (1UL << page_order(buffer)))
2533                 return NULL;
2534
2535         return vmalloc_to_page((void *)buffer->user_page + pgoff * PAGE_SIZE);
2536 }
2537
2538 static void perf_mmap_unmark_page(void *addr)
2539 {
2540         struct page *page = vmalloc_to_page(addr);
2541
2542         page->mapping = NULL;
2543 }
2544
2545 static void perf_buffer_free_work(struct work_struct *work)
2546 {
2547         struct perf_buffer *buffer;
2548         void *base;
2549         int i, nr;
2550
2551         buffer = container_of(work, struct perf_buffer, work);
2552         nr = 1 << page_order(buffer);
2553
2554         base = buffer->user_page;
2555         for (i = 0; i < nr + 1; i++)
2556                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2557
2558         vfree(base);
2559         kfree(buffer);
2560 }
2561
2562 static void perf_buffer_free(struct perf_buffer *buffer)
2563 {
2564         schedule_work(&buffer->work);
2565 }
2566
2567 static struct perf_buffer *
2568 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2569 {
2570         struct perf_buffer *buffer;
2571         unsigned long size;
2572         void *all_buf;
2573
2574         size = sizeof(struct perf_buffer);
2575         size += sizeof(void *);
2576
2577         buffer = kzalloc(size, GFP_KERNEL);
2578         if (!buffer)
2579                 goto fail;
2580
2581         INIT_WORK(&buffer->work, perf_buffer_free_work);
2582
2583         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2584         if (!all_buf)
2585                 goto fail_all_buf;
2586
2587         buffer->user_page = all_buf;
2588         buffer->data_pages[0] = all_buf + PAGE_SIZE;
2589         buffer->page_order = ilog2(nr_pages);
2590         buffer->nr_pages = 1;
2591
2592         perf_buffer_init(buffer, watermark, flags);
2593
2594         return buffer;
2595
2596 fail_all_buf:
2597         kfree(buffer);
2598
2599 fail:
2600         return NULL;
2601 }
2602
2603 #endif
2604
2605 static unsigned long perf_data_size(struct perf_buffer *buffer)
2606 {
2607         return buffer->nr_pages << (PAGE_SHIFT + page_order(buffer));
2608 }
2609
2610 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2611 {
2612         struct perf_event *event = vma->vm_file->private_data;
2613         struct perf_buffer *buffer;
2614         int ret = VM_FAULT_SIGBUS;
2615
2616         if (vmf->flags & FAULT_FLAG_MKWRITE) {
2617                 if (vmf->pgoff == 0)
2618                         ret = 0;
2619                 return ret;
2620         }
2621
2622         rcu_read_lock();
2623         buffer = rcu_dereference(event->buffer);
2624         if (!buffer)
2625                 goto unlock;
2626
2627         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2628                 goto unlock;
2629
2630         vmf->page = perf_mmap_to_page(buffer, vmf->pgoff);
2631         if (!vmf->page)
2632                 goto unlock;
2633
2634         get_page(vmf->page);
2635         vmf->page->mapping = vma->vm_file->f_mapping;
2636         vmf->page->index   = vmf->pgoff;
2637
2638         ret = 0;
2639 unlock:
2640         rcu_read_unlock();
2641
2642         return ret;
2643 }
2644
2645 static void perf_buffer_free_rcu(struct rcu_head *rcu_head)
2646 {
2647         struct perf_buffer *buffer;
2648
2649         buffer = container_of(rcu_head, struct perf_buffer, rcu_head);
2650         perf_buffer_free(buffer);
2651 }
2652
2653 static struct perf_buffer *perf_buffer_get(struct perf_event *event)
2654 {
2655         struct perf_buffer *buffer;
2656
2657         rcu_read_lock();
2658         buffer = rcu_dereference(event->buffer);
2659         if (buffer) {
2660                 if (!atomic_inc_not_zero(&buffer->refcount))
2661                         buffer = NULL;
2662         }
2663         rcu_read_unlock();
2664
2665         return buffer;
2666 }
2667
2668 static void perf_buffer_put(struct perf_buffer *buffer)
2669 {
2670         if (!atomic_dec_and_test(&buffer->refcount))
2671                 return;
2672
2673         call_rcu(&buffer->rcu_head, perf_buffer_free_rcu);
2674 }
2675
2676 static void perf_mmap_open(struct vm_area_struct *vma)
2677 {
2678         struct perf_event *event = vma->vm_file->private_data;
2679
2680         atomic_inc(&event->mmap_count);
2681 }
2682
2683 static void perf_mmap_close(struct vm_area_struct *vma)
2684 {
2685         struct perf_event *event = vma->vm_file->private_data;
2686
2687         if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2688                 unsigned long size = perf_data_size(event->buffer);
2689                 struct user_struct *user = event->mmap_user;
2690                 struct perf_buffer *buffer = event->buffer;
2691
2692                 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2693                 vma->vm_mm->locked_vm -= event->mmap_locked;
2694                 rcu_assign_pointer(event->buffer, NULL);
2695                 mutex_unlock(&event->mmap_mutex);
2696
2697                 perf_buffer_put(buffer);
2698                 free_uid(user);
2699         }
2700 }
2701
2702 static const struct vm_operations_struct perf_mmap_vmops = {
2703         .open           = perf_mmap_open,
2704         .close          = perf_mmap_close,
2705         .fault          = perf_mmap_fault,
2706         .page_mkwrite   = perf_mmap_fault,
2707 };
2708
2709 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2710 {
2711         struct perf_event *event = file->private_data;
2712         unsigned long user_locked, user_lock_limit;
2713         struct user_struct *user = current_user();
2714         unsigned long locked, lock_limit;
2715         struct perf_buffer *buffer;
2716         unsigned long vma_size;
2717         unsigned long nr_pages;
2718         long user_extra, extra;
2719         int ret = 0, flags = 0;
2720
2721         /*
2722          * Don't allow mmap() of inherited per-task counters. This would
2723          * create a performance issue due to all children writing to the
2724          * same buffer.
2725          */
2726         if (event->cpu == -1 && event->attr.inherit)
2727                 return -EINVAL;
2728
2729         if (!(vma->vm_flags & VM_SHARED))
2730                 return -EINVAL;
2731
2732         vma_size = vma->vm_end - vma->vm_start;
2733         nr_pages = (vma_size / PAGE_SIZE) - 1;
2734
2735         /*
2736          * If we have buffer pages ensure they're a power-of-two number, so we
2737          * can do bitmasks instead of modulo.
2738          */
2739         if (nr_pages != 0 && !is_power_of_2(nr_pages))
2740                 return -EINVAL;
2741
2742         if (vma_size != PAGE_SIZE * (1 + nr_pages))
2743                 return -EINVAL;
2744
2745         if (vma->vm_pgoff != 0)
2746                 return -EINVAL;
2747
2748         WARN_ON_ONCE(event->ctx->parent_ctx);
2749         mutex_lock(&event->mmap_mutex);
2750         if (event->buffer) {
2751                 if (event->buffer->nr_pages == nr_pages)
2752                         atomic_inc(&event->buffer->refcount);
2753                 else
2754                         ret = -EINVAL;
2755                 goto unlock;
2756         }
2757
2758         user_extra = nr_pages + 1;
2759         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2760
2761         /*
2762          * Increase the limit linearly with more CPUs:
2763          */
2764         user_lock_limit *= num_online_cpus();
2765
2766         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2767
2768         extra = 0;
2769         if (user_locked > user_lock_limit)
2770                 extra = user_locked - user_lock_limit;
2771
2772         lock_limit = rlimit(RLIMIT_MEMLOCK);
2773         lock_limit >>= PAGE_SHIFT;
2774         locked = vma->vm_mm->locked_vm + extra;
2775
2776         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2777                 !capable(CAP_IPC_LOCK)) {
2778                 ret = -EPERM;
2779                 goto unlock;
2780         }
2781
2782         WARN_ON(event->buffer);
2783
2784         if (vma->vm_flags & VM_WRITE)
2785                 flags |= PERF_BUFFER_WRITABLE;
2786
2787         buffer = perf_buffer_alloc(nr_pages, event->attr.wakeup_watermark,
2788                                    event->cpu, flags);
2789         if (!buffer) {
2790                 ret = -ENOMEM;
2791                 goto unlock;
2792         }
2793         rcu_assign_pointer(event->buffer, buffer);
2794
2795         atomic_long_add(user_extra, &user->locked_vm);
2796         event->mmap_locked = extra;
2797         event->mmap_user = get_current_user();
2798         vma->vm_mm->locked_vm += event->mmap_locked;
2799
2800 unlock:
2801         if (!ret)
2802                 atomic_inc(&event->mmap_count);
2803         mutex_unlock(&event->mmap_mutex);
2804
2805         vma->vm_flags |= VM_RESERVED;
2806         vma->vm_ops = &perf_mmap_vmops;
2807
2808         return ret;
2809 }
2810
2811 static int perf_fasync(int fd, struct file *filp, int on)
2812 {
2813         struct inode *inode = filp->f_path.dentry->d_inode;
2814         struct perf_event *event = filp->private_data;
2815         int retval;
2816
2817         mutex_lock(&inode->i_mutex);
2818         retval = fasync_helper(fd, filp, on, &event->fasync);
2819         mutex_unlock(&inode->i_mutex);
2820
2821         if (retval < 0)
2822                 return retval;
2823
2824         return 0;
2825 }
2826
2827 static const struct file_operations perf_fops = {
2828         .llseek                 = no_llseek,
2829         .release                = perf_release,
2830         .read                   = perf_read,
2831         .poll                   = perf_poll,
2832         .unlocked_ioctl         = perf_ioctl,
2833         .compat_ioctl           = perf_ioctl,
2834         .mmap                   = perf_mmap,
2835         .fasync                 = perf_fasync,
2836 };
2837
2838 /*
2839  * Perf event wakeup
2840  *
2841  * If there's data, ensure we set the poll() state and publish everything
2842  * to user-space before waking everybody up.
2843  */
2844
2845 void perf_event_wakeup(struct perf_event *event)
2846 {
2847         wake_up_all(&event->waitq);
2848
2849         if (event->pending_kill) {
2850                 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2851                 event->pending_kill = 0;
2852         }
2853 }
2854
2855 /*
2856  * Pending wakeups
2857  *
2858  * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2859  *
2860  * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2861  * single linked list and use cmpxchg() to add entries lockless.
2862  */
2863
2864 static void perf_pending_event(struct perf_pending_entry *entry)
2865 {
2866         struct perf_event *event = container_of(entry,
2867                         struct perf_event, pending);
2868
2869         if (event->pending_disable) {
2870                 event->pending_disable = 0;
2871                 __perf_event_disable(event);
2872         }
2873
2874         if (event->pending_wakeup) {
2875                 event->pending_wakeup = 0;
2876                 perf_event_wakeup(event);
2877         }
2878 }
2879
2880 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2881
2882 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2883         PENDING_TAIL,
2884 };
2885
2886 static void perf_pending_queue(struct perf_pending_entry *entry,
2887                                void (*func)(struct perf_pending_entry *))
2888 {
2889         struct perf_pending_entry **head;
2890
2891         if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2892                 return;
2893
2894         entry->func = func;
2895
2896         head = &get_cpu_var(perf_pending_head);
2897
2898         do {
2899                 entry->next = *head;
2900         } while (cmpxchg(head, entry->next, entry) != entry->next);
2901
2902         set_perf_event_pending();
2903
2904         put_cpu_var(perf_pending_head);
2905 }
2906
2907 static int __perf_pending_run(void)
2908 {
2909         struct perf_pending_entry *list;
2910         int nr = 0;
2911
2912         list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2913         while (list != PENDING_TAIL) {
2914                 void (*func)(struct perf_pending_entry *);
2915                 struct perf_pending_entry *entry = list;
2916
2917                 list = list->next;
2918
2919                 func = entry->func;
2920                 entry->next = NULL;
2921                 /*
2922                  * Ensure we observe the unqueue before we issue the wakeup,
2923                  * so that we won't be waiting forever.
2924                  * -- see perf_not_pending().
2925                  */
2926                 smp_wmb();
2927
2928                 func(entry);
2929                 nr++;
2930         }
2931
2932         return nr;
2933 }
2934
2935 static inline int perf_not_pending(struct perf_event *event)
2936 {
2937         /*
2938          * If we flush on whatever cpu we run, there is a chance we don't
2939          * need to wait.
2940          */
2941         get_cpu();
2942         __perf_pending_run();
2943         put_cpu();
2944
2945         /*
2946          * Ensure we see the proper queue state before going to sleep
2947          * so that we do not miss the wakeup. -- see perf_pending_handle()
2948          */
2949         smp_rmb();
2950         return event->pending.next == NULL;
2951 }
2952
2953 static void perf_pending_sync(struct perf_event *event)
2954 {
2955         wait_event(event->waitq, perf_not_pending(event));
2956 }
2957
2958 void perf_event_do_pending(void)
2959 {
2960         __perf_pending_run();
2961 }
2962
2963 /*
2964  * Callchain support -- arch specific
2965  */
2966
2967 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2968 {
2969         return NULL;
2970 }
2971
2972
2973 /*
2974  * We assume there is only KVM supporting the callbacks.
2975  * Later on, we might change it to a list if there is
2976  * another virtualization implementation supporting the callbacks.
2977  */
2978 struct perf_guest_info_callbacks *perf_guest_cbs;
2979
2980 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2981 {
2982         perf_guest_cbs = cbs;
2983         return 0;
2984 }
2985 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
2986
2987 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2988 {
2989         perf_guest_cbs = NULL;
2990         return 0;
2991 }
2992 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
2993
2994 /*
2995  * Output
2996  */
2997 static bool perf_output_space(struct perf_buffer *buffer, unsigned long tail,
2998                               unsigned long offset, unsigned long head)
2999 {
3000         unsigned long mask;
3001
3002         if (!buffer->writable)
3003                 return true;
3004
3005         mask = perf_data_size(buffer) - 1;
3006
3007         offset = (offset - tail) & mask;
3008         head   = (head   - tail) & mask;
3009
3010         if ((int)(head - offset) < 0)
3011                 return false;
3012
3013         return true;
3014 }
3015
3016 static void perf_output_wakeup(struct perf_output_handle *handle)
3017 {
3018         atomic_set(&handle->buffer->poll, POLL_IN);
3019
3020         if (handle->nmi) {
3021                 handle->event->pending_wakeup = 1;
3022                 perf_pending_queue(&handle->event->pending,
3023                                    perf_pending_event);
3024         } else
3025                 perf_event_wakeup(handle->event);
3026 }
3027
3028 /*
3029  * We need to ensure a later event_id doesn't publish a head when a former
3030  * event isn't done writing. However since we need to deal with NMIs we
3031  * cannot fully serialize things.
3032  *
3033  * We only publish the head (and generate a wakeup) when the outer-most
3034  * event completes.
3035  */
3036 static void perf_output_get_handle(struct perf_output_handle *handle)
3037 {
3038         struct perf_buffer *buffer = handle->buffer;
3039
3040         preempt_disable();
3041         local_inc(&buffer->nest);
3042         handle->wakeup = local_read(&buffer->wakeup);
3043 }
3044
3045 static void perf_output_put_handle(struct perf_output_handle *handle)
3046 {
3047         struct perf_buffer *buffer = handle->buffer;
3048         unsigned long head;
3049
3050 again:
3051         head = local_read(&buffer->head);
3052
3053         /*
3054          * IRQ/NMI can happen here, which means we can miss a head update.
3055          */
3056
3057         if (!local_dec_and_test(&buffer->nest))
3058                 goto out;
3059
3060         /*
3061          * Publish the known good head. Rely on the full barrier implied
3062          * by atomic_dec_and_test() order the buffer->head read and this
3063          * write.
3064          */
3065         buffer->user_page->data_head = head;
3066
3067         /*
3068          * Now check if we missed an update, rely on the (compiler)
3069          * barrier in atomic_dec_and_test() to re-read buffer->head.
3070          */
3071         if (unlikely(head != local_read(&buffer->head))) {
3072                 local_inc(&buffer->nest);
3073                 goto again;
3074         }
3075
3076         if (handle->wakeup != local_read(&buffer->wakeup))
3077                 perf_output_wakeup(handle);
3078
3079  out:
3080         preempt_enable();
3081 }
3082
3083 __always_inline void perf_output_copy(struct perf_output_handle *handle,
3084                       const void *buf, unsigned int len)
3085 {
3086         do {
3087                 unsigned long size = min_t(unsigned long, handle->size, len);
3088
3089                 memcpy(handle->addr, buf, size);
3090
3091                 len -= size;
3092                 handle->addr += size;
3093                 buf += size;
3094                 handle->size -= size;
3095                 if (!handle->size) {
3096                         struct perf_buffer *buffer = handle->buffer;
3097
3098                         handle->page++;
3099                         handle->page &= buffer->nr_pages - 1;
3100                         handle->addr = buffer->data_pages[handle->page];
3101                         handle->size = PAGE_SIZE << page_order(buffer);
3102                 }
3103         } while (len);
3104 }
3105
3106 int perf_output_begin(struct perf_output_handle *handle,
3107                       struct perf_event *event, unsigned int size,
3108                       int nmi, int sample)
3109 {
3110         struct perf_buffer *buffer;
3111         unsigned long tail, offset, head;
3112         int have_lost;
3113         struct {
3114                 struct perf_event_header header;
3115                 u64                      id;
3116                 u64                      lost;
3117         } lost_event;
3118
3119         rcu_read_lock();
3120         /*
3121          * For inherited events we send all the output towards the parent.
3122          */
3123         if (event->parent)
3124                 event = event->parent;
3125
3126         buffer = rcu_dereference(event->buffer);
3127         if (!buffer)
3128                 goto out;
3129
3130         handle->buffer  = buffer;
3131         handle->event   = event;
3132         handle->nmi     = nmi;
3133         handle->sample  = sample;
3134
3135         if (!buffer->nr_pages)
3136                 goto out;
3137
3138         have_lost = local_read(&buffer->lost);
3139         if (have_lost)
3140                 size += sizeof(lost_event);
3141
3142         perf_output_get_handle(handle);
3143
3144         do {
3145                 /*
3146                  * Userspace could choose to issue a mb() before updating the
3147                  * tail pointer. So that all reads will be completed before the
3148                  * write is issued.
3149                  */
3150                 tail = ACCESS_ONCE(buffer->user_page->data_tail);
3151                 smp_rmb();
3152                 offset = head = local_read(&buffer->head);
3153                 head += size;
3154                 if (unlikely(!perf_output_space(buffer, tail, offset, head)))
3155                         goto fail;
3156         } while (local_cmpxchg(&buffer->head, offset, head) != offset);
3157
3158         if (head - local_read(&buffer->wakeup) > buffer->watermark)
3159                 local_add(buffer->watermark, &buffer->wakeup);
3160
3161         handle->page = offset >> (PAGE_SHIFT + page_order(buffer));
3162         handle->page &= buffer->nr_pages - 1;
3163         handle->size = offset & ((PAGE_SIZE << page_order(buffer)) - 1);
3164         handle->addr = buffer->data_pages[handle->page];
3165         handle->addr += handle->size;
3166         handle->size = (PAGE_SIZE << page_order(buffer)) - handle->size;
3167
3168         if (have_lost) {
3169                 lost_event.header.type = PERF_RECORD_LOST;
3170                 lost_event.header.misc = 0;
3171                 lost_event.header.size = sizeof(lost_event);
3172                 lost_event.id          = event->id;
3173                 lost_event.lost        = local_xchg(&buffer->lost, 0);
3174
3175                 perf_output_put(handle, lost_event);
3176         }
3177
3178         return 0;
3179
3180 fail:
3181         local_inc(&buffer->lost);
3182         perf_output_put_handle(handle);
3183 out:
3184         rcu_read_unlock();
3185
3186         return -ENOSPC;
3187 }
3188
3189 void perf_output_end(struct perf_output_handle *handle)
3190 {
3191         struct perf_event *event = handle->event;
3192         struct perf_buffer *buffer = handle->buffer;
3193
3194         int wakeup_events = event->attr.wakeup_events;
3195
3196         if (handle->sample && wakeup_events) {
3197                 int events = local_inc_return(&buffer->events);
3198                 if (events >= wakeup_events) {
3199                         local_sub(wakeup_events, &buffer->events);
3200                         local_inc(&buffer->wakeup);
3201                 }
3202         }
3203
3204         perf_output_put_handle(handle);
3205         rcu_read_unlock();
3206 }
3207
3208 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3209 {
3210         /*
3211          * only top level events have the pid namespace they were created in
3212          */
3213         if (event->parent)
3214                 event = event->parent;
3215
3216         return task_tgid_nr_ns(p, event->ns);
3217 }
3218
3219 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3220 {
3221         /*
3222          * only top level events have the pid namespace they were created in
3223          */
3224         if (event->parent)
3225                 event = event->parent;
3226
3227         return task_pid_nr_ns(p, event->ns);
3228 }
3229
3230 static void perf_output_read_one(struct perf_output_handle *handle,
3231                                  struct perf_event *event)
3232 {
3233         u64 read_format = event->attr.read_format;
3234         u64 values[4];
3235         int n = 0;
3236
3237         values[n++] = perf_event_count(event);
3238         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3239                 values[n++] = event->total_time_enabled +
3240                         atomic64_read(&event->child_total_time_enabled);
3241         }
3242         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3243                 values[n++] = event->total_time_running +
3244                         atomic64_read(&event->child_total_time_running);
3245         }
3246         if (read_format & PERF_FORMAT_ID)
3247                 values[n++] = primary_event_id(event);
3248
3249         perf_output_copy(handle, values, n * sizeof(u64));
3250 }
3251
3252 /*
3253  * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3254  */
3255 static void perf_output_read_group(struct perf_output_handle *handle,
3256                             struct perf_event *event)
3257 {
3258         struct perf_event *leader = event->group_leader, *sub;
3259         u64 read_format = event->attr.read_format;
3260         u64 values[5];
3261         int n = 0;
3262
3263         values[n++] = 1 + leader->nr_siblings;
3264
3265         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3266                 values[n++] = leader->total_time_enabled;
3267
3268         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3269                 values[n++] = leader->total_time_running;
3270
3271         if (leader != event)
3272                 leader->pmu->read(leader);
3273
3274         values[n++] = perf_event_count(leader);
3275         if (read_format & PERF_FORMAT_ID)
3276                 values[n++] = primary_event_id(leader);
3277
3278         perf_output_copy(handle, values, n * sizeof(u64));
3279
3280         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3281                 n = 0;
3282
3283                 if (sub != event)
3284                         sub->pmu->read(sub);
3285
3286                 values[n++] = perf_event_count(sub);
3287                 if (read_format & PERF_FORMAT_ID)
3288                         values[n++] = primary_event_id(sub);
3289
3290                 perf_output_copy(handle, values, n * sizeof(u64));
3291         }
3292 }
3293
3294 static void perf_output_read(struct perf_output_handle *handle,
3295                              struct perf_event *event)
3296 {
3297         if (event->attr.read_format & PERF_FORMAT_GROUP)
3298                 perf_output_read_group(handle, event);
3299         else
3300                 perf_output_read_one(handle, event);
3301 }
3302
3303 void perf_output_sample(struct perf_output_handle *handle,
3304                         struct perf_event_header *header,
3305                         struct perf_sample_data *data,
3306                         struct perf_event *event)
3307 {
3308         u64 sample_type = data->type;
3309
3310         perf_output_put(handle, *header);
3311
3312         if (sample_type & PERF_SAMPLE_IP)
3313                 perf_output_put(handle, data->ip);
3314
3315         if (sample_type & PERF_SAMPLE_TID)
3316                 perf_output_put(handle, data->tid_entry);
3317
3318         if (sample_type & PERF_SAMPLE_TIME)
3319                 perf_output_put(handle, data->time);
3320
3321         if (sample_type & PERF_SAMPLE_ADDR)
3322                 perf_output_put(handle, data->addr);
3323
3324         if (sample_type & PERF_SAMPLE_ID)
3325                 perf_output_put(handle, data->id);
3326
3327         if (sample_type & PERF_SAMPLE_STREAM_ID)
3328                 perf_output_put(handle, data->stream_id);
3329
3330         if (sample_type & PERF_SAMPLE_CPU)
3331                 perf_output_put(handle, data->cpu_entry);
3332
3333         if (sample_type & PERF_SAMPLE_PERIOD)
3334                 perf_output_put(handle, data->period);
3335
3336         if (sample_type & PERF_SAMPLE_READ)
3337                 perf_output_read(handle, event);
3338
3339         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3340                 if (data->callchain) {
3341                         int size = 1;
3342
3343                         if (data->callchain)
3344                                 size += data->callchain->nr;
3345
3346                         size *= sizeof(u64);
3347
3348                         perf_output_copy(handle, data->callchain, size);
3349                 } else {
3350                         u64 nr = 0;
3351                         perf_output_put(handle, nr);
3352                 }
3353         }
3354
3355         if (sample_type & PERF_SAMPLE_RAW) {
3356                 if (data->raw) {
3357                         perf_output_put(handle, data->raw->size);
3358                         perf_output_copy(handle, data->raw->data,
3359                                          data->raw->size);
3360                 } else {
3361                         struct {
3362                                 u32     size;
3363                                 u32     data;
3364                         } raw = {
3365                                 .size = sizeof(u32),
3366                                 .data = 0,
3367                         };
3368                         perf_output_put(handle, raw);
3369                 }
3370         }
3371 }
3372
3373 void perf_prepare_sample(struct perf_event_header *header,
3374                          struct perf_sample_data *data,
3375                          struct perf_event *event,
3376                          struct pt_regs *regs)
3377 {
3378         u64 sample_type = event->attr.sample_type;
3379
3380         data->type = sample_type;
3381
3382         header->type = PERF_RECORD_SAMPLE;
3383         header->size = sizeof(*header);
3384
3385         header->misc = 0;
3386         header->misc |= perf_misc_flags(regs);
3387
3388         if (sample_type & PERF_SAMPLE_IP) {
3389                 data->ip = perf_instruction_pointer(regs);
3390
3391                 header->size += sizeof(data->ip);
3392         }
3393
3394         if (sample_type & PERF_SAMPLE_TID) {
3395                 /* namespace issues */
3396                 data->tid_entry.pid = perf_event_pid(event, current);
3397                 data->tid_entry.tid = perf_event_tid(event, current);
3398
3399                 header->size += sizeof(data->tid_entry);
3400         }
3401
3402         if (sample_type & PERF_SAMPLE_TIME) {
3403                 data->time = perf_clock();
3404
3405                 header->size += sizeof(data->time);
3406         }
3407
3408         if (sample_type & PERF_SAMPLE_ADDR)
3409                 header->size += sizeof(data->addr);
3410
3411         if (sample_type & PERF_SAMPLE_ID) {
3412                 data->id = primary_event_id(event);
3413
3414                 header->size += sizeof(data->id);
3415         }
3416
3417         if (sample_type & PERF_SAMPLE_STREAM_ID) {
3418                 data->stream_id = event->id;
3419
3420                 header->size += sizeof(data->stream_id);
3421         }
3422
3423         if (sample_type & PERF_SAMPLE_CPU) {
3424                 data->cpu_entry.cpu             = raw_smp_processor_id();
3425                 data->cpu_entry.reserved        = 0;
3426
3427                 header->size += sizeof(data->cpu_entry);
3428         }
3429
3430         if (sample_type & PERF_SAMPLE_PERIOD)
3431                 header->size += sizeof(data->period);
3432
3433         if (sample_type & PERF_SAMPLE_READ)
3434                 header->size += perf_event_read_size(event);
3435
3436         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3437                 int size = 1;
3438
3439                 data->callchain = perf_callchain(regs);
3440
3441                 if (data->callchain)
3442                         size += data->callchain->nr;
3443
3444                 header->size += size * sizeof(u64);
3445         }
3446
3447         if (sample_type & PERF_SAMPLE_RAW) {
3448                 int size = sizeof(u32);
3449
3450                 if (data->raw)
3451                         size += data->raw->size;
3452                 else
3453                         size += sizeof(u32);
3454
3455                 WARN_ON_ONCE(size & (sizeof(u64)-1));
3456                 header->size += size;
3457         }
3458 }
3459
3460 static void perf_event_output(struct perf_event *event, int nmi,
3461                                 struct perf_sample_data *data,
3462                                 struct pt_regs *regs)
3463 {
3464         struct perf_output_handle handle;
3465         struct perf_event_header header;
3466
3467         perf_prepare_sample(&header, data, event, regs);
3468
3469         if (perf_output_begin(&handle, event, header.size, nmi, 1))
3470                 return;
3471
3472         perf_output_sample(&handle, &header, data, event);
3473
3474         perf_output_end(&handle);
3475 }
3476
3477 /*
3478  * read event_id
3479  */
3480
3481 struct perf_read_event {
3482         struct perf_event_header        header;
3483
3484         u32                             pid;
3485         u32                             tid;
3486 };
3487
3488 static void
3489 perf_event_read_event(struct perf_event *event,
3490                         struct task_struct *task)
3491 {
3492         struct perf_output_handle handle;
3493         struct perf_read_event read_event = {
3494                 .header = {
3495                         .type = PERF_RECORD_READ,
3496                         .misc = 0,
3497                         .size = sizeof(read_event) + perf_event_read_size(event),
3498                 },
3499                 .pid = perf_event_pid(event, task),
3500                 .tid = perf_event_tid(event, task),
3501         };
3502         int ret;
3503
3504         ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3505         if (ret)
3506                 return;
3507
3508         perf_output_put(&handle, read_event);
3509         perf_output_read(&handle, event);
3510
3511         perf_output_end(&handle);
3512 }
3513
3514 /*
3515  * task tracking -- fork/exit
3516  *
3517  * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3518  */
3519
3520 struct perf_task_event {
3521         struct task_struct              *task;
3522         struct perf_event_context       *task_ctx;
3523
3524         struct {
3525                 struct perf_event_header        header;
3526
3527                 u32                             pid;
3528                 u32                             ppid;
3529                 u32                             tid;
3530                 u32                             ptid;
3531                 u64                             time;
3532         } event_id;
3533 };
3534
3535 static void perf_event_task_output(struct perf_event *event,
3536                                      struct perf_task_event *task_event)
3537 {
3538         struct perf_output_handle handle;
3539         struct task_struct *task = task_event->task;
3540         int size, ret;
3541
3542         size  = task_event->event_id.header.size;
3543         ret = perf_output_begin(&handle, event, size, 0, 0);
3544
3545         if (ret)
3546                 return;
3547
3548         task_event->event_id.pid = perf_event_pid(event, task);
3549         task_event->event_id.ppid = perf_event_pid(event, current);
3550
3551         task_event->event_id.tid = perf_event_tid(event, task);
3552         task_event->event_id.ptid = perf_event_tid(event, current);
3553
3554         perf_output_put(&handle, task_event->event_id);
3555
3556         perf_output_end(&handle);
3557 }
3558
3559 static int perf_event_task_match(struct perf_event *event)
3560 {
3561         if (event->state < PERF_EVENT_STATE_INACTIVE)
3562                 return 0;
3563
3564         if (event->cpu != -1 && event->cpu != smp_processor_id())
3565                 return 0;
3566
3567         if (event->attr.comm || event->attr.mmap ||
3568             event->attr.mmap_data || event->attr.task)
3569                 return 1;
3570
3571         return 0;
3572 }
3573
3574 static void perf_event_task_ctx(struct perf_event_context *ctx,
3575                                   struct perf_task_event *task_event)
3576 {
3577         struct perf_event *event;
3578
3579         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3580                 if (perf_event_task_match(event))
3581                         perf_event_task_output(event, task_event);
3582         }
3583 }
3584
3585 static void perf_event_task_event(struct perf_task_event *task_event)
3586 {
3587         struct perf_cpu_context *cpuctx;
3588         struct perf_event_context *ctx = task_event->task_ctx;
3589
3590         rcu_read_lock();
3591         cpuctx = &get_cpu_var(perf_cpu_context);
3592         perf_event_task_ctx(&cpuctx->ctx, task_event);
3593         if (!ctx)
3594                 ctx = rcu_dereference(current->perf_event_ctxp);
3595         if (ctx)
3596                 perf_event_task_ctx(ctx, task_event);
3597         put_cpu_var(perf_cpu_context);
3598         rcu_read_unlock();
3599 }
3600
3601 static void perf_event_task(struct task_struct *task,
3602                               struct perf_event_context *task_ctx,
3603                               int new)
3604 {
3605         struct perf_task_event task_event;
3606
3607         if (!atomic_read(&nr_comm_events) &&
3608             !atomic_read(&nr_mmap_events) &&
3609             !atomic_read(&nr_task_events))
3610                 return;
3611
3612         task_event = (struct perf_task_event){
3613                 .task     = task,
3614                 .task_ctx = task_ctx,
3615                 .event_id    = {
3616                         .header = {
3617                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3618                                 .misc = 0,
3619                                 .size = sizeof(task_event.event_id),
3620                         },
3621                         /* .pid  */
3622                         /* .ppid */
3623                         /* .tid  */
3624                         /* .ptid */
3625                         .time = perf_clock(),
3626                 },
3627         };
3628
3629         perf_event_task_event(&task_event);
3630 }
3631
3632 void perf_event_fork(struct task_struct *task)
3633 {
3634         perf_event_task(task, NULL, 1);
3635 }
3636
3637 /*
3638  * comm tracking
3639  */
3640
3641 struct perf_comm_event {
3642         struct task_struct      *task;
3643         char                    *comm;
3644         int                     comm_size;
3645
3646         struct {
3647                 struct perf_event_header        header;
3648
3649                 u32                             pid;
3650                 u32                             tid;
3651         } event_id;
3652 };
3653
3654 static void perf_event_comm_output(struct perf_event *event,
3655                                      struct perf_comm_event *comm_event)
3656 {
3657         struct perf_output_handle handle;
3658         int size = comm_event->event_id.header.size;
3659         int ret = perf_output_begin(&handle, event, size, 0, 0);
3660
3661         if (ret)
3662                 return;
3663
3664         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3665         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3666
3667         perf_output_put(&handle, comm_event->event_id);
3668         perf_output_copy(&handle, comm_event->comm,
3669                                    comm_event->comm_size);
3670         perf_output_end(&handle);
3671 }
3672
3673 static int perf_event_comm_match(struct perf_event *event)
3674 {
3675         if (event->state < PERF_EVENT_STATE_INACTIVE)
3676                 return 0;
3677
3678         if (event->cpu != -1 && event->cpu != smp_processor_id())
3679                 return 0;
3680
3681         if (event->attr.comm)
3682                 return 1;
3683
3684         return 0;
3685 }
3686
3687 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3688                                   struct perf_comm_event *comm_event)
3689 {
3690         struct perf_event *event;
3691
3692         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3693                 if (perf_event_comm_match(event))
3694                         perf_event_comm_output(event, comm_event);
3695         }
3696 }
3697
3698 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3699 {
3700         struct perf_cpu_context *cpuctx;
3701         struct perf_event_context *ctx;
3702         unsigned int size;
3703         char comm[TASK_COMM_LEN];
3704
3705         memset(comm, 0, sizeof(comm));
3706         strlcpy(comm, comm_event->task->comm, sizeof(comm));
3707         size = ALIGN(strlen(comm)+1, sizeof(u64));
3708
3709         comm_event->comm = comm;
3710         comm_event->comm_size = size;
3711
3712         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3713
3714         rcu_read_lock();
3715         cpuctx = &get_cpu_var(perf_cpu_context);
3716         perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3717         ctx = rcu_dereference(current->perf_event_ctxp);
3718         if (ctx)
3719                 perf_event_comm_ctx(ctx, comm_event);
3720         put_cpu_var(perf_cpu_context);
3721         rcu_read_unlock();
3722 }
3723
3724 void perf_event_comm(struct task_struct *task)
3725 {
3726         struct perf_comm_event comm_event;
3727
3728         if (task->perf_event_ctxp)
3729                 perf_event_enable_on_exec(task);
3730
3731         if (!atomic_read(&nr_comm_events))
3732                 return;
3733
3734         comm_event = (struct perf_comm_event){
3735                 .task   = task,
3736                 /* .comm      */
3737                 /* .comm_size */
3738                 .event_id  = {
3739                         .header = {
3740                                 .type = PERF_RECORD_COMM,
3741                                 .misc = 0,
3742                                 /* .size */
3743                         },
3744                         /* .pid */
3745                         /* .tid */
3746                 },
3747         };
3748
3749         perf_event_comm_event(&comm_event);
3750 }
3751
3752 /*
3753  * mmap tracking
3754  */
3755
3756 struct perf_mmap_event {
3757         struct vm_area_struct   *vma;
3758
3759         const char              *file_name;
3760         int                     file_size;
3761
3762         struct {
3763                 struct perf_event_header        header;
3764
3765                 u32                             pid;
3766                 u32                             tid;
3767                 u64                             start;
3768                 u64                             len;
3769                 u64                             pgoff;
3770         } event_id;
3771 };
3772
3773 static void perf_event_mmap_output(struct perf_event *event,
3774                                      struct perf_mmap_event *mmap_event)
3775 {
3776         struct perf_output_handle handle;
3777         int size = mmap_event->event_id.header.size;
3778         int ret = perf_output_begin(&handle, event, size, 0, 0);
3779
3780         if (ret)
3781                 return;
3782
3783         mmap_event->event_id.pid = perf_event_pid(event, current);
3784         mmap_event->event_id.tid = perf_event_tid(event, current);
3785
3786         perf_output_put(&handle, mmap_event->event_id);
3787         perf_output_copy(&handle, mmap_event->file_name,
3788                                    mmap_event->file_size);
3789         perf_output_end(&handle);
3790 }
3791
3792 static int perf_event_mmap_match(struct perf_event *event,
3793                                    struct perf_mmap_event *mmap_event,
3794                                    int executable)
3795 {
3796         if (event->state < PERF_EVENT_STATE_INACTIVE)
3797                 return 0;
3798
3799         if (event->cpu != -1 && event->cpu != smp_processor_id())
3800                 return 0;
3801
3802         if ((!executable && event->attr.mmap_data) ||
3803             (executable && event->attr.mmap))
3804                 return 1;
3805
3806         return 0;
3807 }
3808
3809 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3810                                   struct perf_mmap_event *mmap_event,
3811                                   int executable)
3812 {
3813         struct perf_event *event;
3814
3815         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3816                 if (perf_event_mmap_match(event, mmap_event, executable))
3817                         perf_event_mmap_output(event, mmap_event);
3818         }
3819 }
3820
3821 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3822 {
3823         struct perf_cpu_context *cpuctx;
3824         struct perf_event_context *ctx;
3825         struct vm_area_struct *vma = mmap_event->vma;
3826         struct file *file = vma->vm_file;
3827         unsigned int size;
3828         char tmp[16];
3829         char *buf = NULL;
3830         const char *name;
3831
3832         memset(tmp, 0, sizeof(tmp));
3833
3834         if (file) {
3835                 /*
3836                  * d_path works from the end of the buffer backwards, so we
3837                  * need to add enough zero bytes after the string to handle
3838                  * the 64bit alignment we do later.
3839                  */
3840                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3841                 if (!buf) {
3842                         name = strncpy(tmp, "//enomem", sizeof(tmp));
3843                         goto got_name;
3844                 }
3845                 name = d_path(&file->f_path, buf, PATH_MAX);
3846                 if (IS_ERR(name)) {
3847                         name = strncpy(tmp, "//toolong", sizeof(tmp));
3848                         goto got_name;
3849                 }
3850         } else {
3851                 if (arch_vma_name(mmap_event->vma)) {
3852                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3853                                        sizeof(tmp));
3854                         goto got_name;
3855                 }
3856
3857                 if (!vma->vm_mm) {
3858                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
3859                         goto got_name;
3860                 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
3861                                 vma->vm_end >= vma->vm_mm->brk) {
3862                         name = strncpy(tmp, "[heap]", sizeof(tmp));
3863                         goto got_name;
3864                 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
3865                                 vma->vm_end >= vma->vm_mm->start_stack) {
3866                         name = strncpy(tmp, "[stack]", sizeof(tmp));
3867                         goto got_name;
3868                 }
3869
3870                 name = strncpy(tmp, "//anon", sizeof(tmp));
3871                 goto got_name;
3872         }
3873
3874 got_name:
3875         size = ALIGN(strlen(name)+1, sizeof(u64));
3876
3877         mmap_event->file_name = name;
3878         mmap_event->file_size = size;
3879
3880         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3881
3882         rcu_read_lock();
3883         cpuctx = &get_cpu_var(perf_cpu_context);
3884         perf_event_mmap_ctx(&cpuctx->ctx, mmap_event, vma->vm_flags & VM_EXEC);
3885         ctx = rcu_dereference(current->perf_event_ctxp);
3886         if (ctx)
3887                 perf_event_mmap_ctx(ctx, mmap_event, vma->vm_flags & VM_EXEC);
3888         put_cpu_var(perf_cpu_context);
3889         rcu_read_unlock();
3890
3891         kfree(buf);
3892 }
3893
3894 void perf_event_mmap(struct vm_area_struct *vma)
3895 {
3896         struct perf_mmap_event mmap_event;
3897
3898         if (!atomic_read(&nr_mmap_events))
3899                 return;
3900
3901         mmap_event = (struct perf_mmap_event){
3902                 .vma    = vma,
3903                 /* .file_name */
3904                 /* .file_size */
3905                 .event_id  = {
3906                         .header = {
3907                                 .type = PERF_RECORD_MMAP,
3908                                 .misc = PERF_RECORD_MISC_USER,
3909                                 /* .size */
3910                         },
3911                         /* .pid */
3912                         /* .tid */
3913                         .start  = vma->vm_start,
3914                         .len    = vma->vm_end - vma->vm_start,
3915                         .pgoff  = (u64)vma->vm_pgoff << PAGE_SHIFT,
3916                 },
3917         };
3918
3919         perf_event_mmap_event(&mmap_event);
3920 }
3921
3922 /*
3923  * IRQ throttle logging
3924  */
3925
3926 static void perf_log_throttle(struct perf_event *event, int enable)
3927 {
3928         struct perf_output_handle handle;
3929         int ret;
3930
3931         struct {
3932                 struct perf_event_header        header;
3933                 u64                             time;
3934                 u64                             id;
3935                 u64                             stream_id;
3936         } throttle_event = {
3937                 .header = {
3938                         .type = PERF_RECORD_THROTTLE,
3939                         .misc = 0,
3940                         .size = sizeof(throttle_event),
3941                 },
3942                 .time           = perf_clock(),
3943                 .id             = primary_event_id(event),
3944                 .stream_id      = event->id,
3945         };
3946
3947         if (enable)
3948                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3949
3950         ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3951         if (ret)
3952                 return;
3953
3954         perf_output_put(&handle, throttle_event);
3955         perf_output_end(&handle);
3956 }
3957
3958 /*
3959  * Generic event overflow handling, sampling.
3960  */
3961
3962 static int __perf_event_overflow(struct perf_event *event, int nmi,
3963                                    int throttle, struct perf_sample_data *data,
3964                                    struct pt_regs *regs)
3965 {
3966         int events = atomic_read(&event->event_limit);
3967         struct hw_perf_event *hwc = &event->hw;
3968         int ret = 0;
3969
3970         throttle = (throttle && event->pmu->unthrottle != NULL);
3971
3972         if (!throttle) {
3973                 hwc->interrupts++;
3974         } else {
3975                 if (hwc->interrupts != MAX_INTERRUPTS) {
3976                         hwc->interrupts++;
3977                         if (HZ * hwc->interrupts >
3978                                         (u64)sysctl_perf_event_sample_rate) {
3979                                 hwc->interrupts = MAX_INTERRUPTS;
3980                                 perf_log_throttle(event, 0);
3981                                 ret = 1;
3982                         }
3983                 } else {
3984                         /*
3985                          * Keep re-disabling events even though on the previous
3986                          * pass we disabled it - just in case we raced with a
3987                          * sched-in and the event got enabled again:
3988                          */
3989                         ret = 1;
3990                 }
3991         }
3992
3993         if (event->attr.freq) {
3994                 u64 now = perf_clock();
3995                 s64 delta = now - hwc->freq_time_stamp;
3996
3997                 hwc->freq_time_stamp = now;
3998
3999                 if (delta > 0 && delta < 2*TICK_NSEC)
4000                         perf_adjust_period(event, delta, hwc->last_period);
4001         }
4002
4003         /*
4004          * XXX event_limit might not quite work as expected on inherited
4005          * events
4006          */
4007
4008         event->pending_kill = POLL_IN;
4009         if (events && atomic_dec_and_test(&event->event_limit)) {
4010                 ret = 1;
4011                 event->pending_kill = POLL_HUP;
4012                 if (nmi) {
4013                         event->pending_disable = 1;
4014                         perf_pending_queue(&event->pending,
4015                                            perf_pending_event);
4016                 } else
4017                         perf_event_disable(event);
4018         }
4019
4020         if (event->overflow_handler)
4021                 event->overflow_handler(event, nmi, data, regs);
4022         else
4023                 perf_event_output(event, nmi, data, regs);
4024
4025         return ret;
4026 }
4027
4028 int perf_event_overflow(struct perf_event *event, int nmi,
4029                           struct perf_sample_data *data,
4030                           struct pt_regs *regs)
4031 {
4032         return __perf_event_overflow(event, nmi, 1, data, regs);
4033 }
4034
4035 /*
4036  * Generic software event infrastructure
4037  */
4038
4039 /*
4040  * We directly increment event->count and keep a second value in
4041  * event->hw.period_left to count intervals. This period event
4042  * is kept in the range [-sample_period, 0] so that we can use the
4043  * sign as trigger.
4044  */
4045
4046 static u64 perf_swevent_set_period(struct perf_event *event)
4047 {
4048         struct hw_perf_event *hwc = &event->hw;
4049         u64 period = hwc->last_period;
4050         u64 nr, offset;
4051         s64 old, val;
4052
4053         hwc->last_period = hwc->sample_period;
4054
4055 again:
4056         old = val = local64_read(&hwc->period_left);
4057         if (val < 0)
4058                 return 0;
4059
4060         nr = div64_u64(period + val, period);
4061         offset = nr * period;
4062         val -= offset;
4063         if (local64_cmpxchg(&hwc->period_left, old, val) != old)
4064                 goto again;
4065
4066         return nr;
4067 }
4068
4069 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4070                                     int nmi, struct perf_sample_data *data,
4071                                     struct pt_regs *regs)
4072 {
4073         struct hw_perf_event *hwc = &event->hw;
4074         int throttle = 0;
4075
4076         data->period = event->hw.last_period;
4077         if (!overflow)
4078                 overflow = perf_swevent_set_period(event);
4079
4080         if (hwc->interrupts == MAX_INTERRUPTS)
4081                 return;
4082
4083         for (; overflow; overflow--) {
4084                 if (__perf_event_overflow(event, nmi, throttle,
4085                                             data, regs)) {
4086                         /*
4087                          * We inhibit the overflow from happening when
4088                          * hwc->interrupts == MAX_INTERRUPTS.
4089                          */
4090                         break;
4091                 }
4092                 throttle = 1;
4093         }
4094 }
4095
4096 static void perf_swevent_add(struct perf_event *event, u64 nr,
4097                                int nmi, struct perf_sample_data *data,
4098                                struct pt_regs *regs)
4099 {
4100         struct hw_perf_event *hwc = &event->hw;
4101
4102         local64_add(nr, &event->count);
4103
4104         if (!regs)
4105                 return;
4106
4107         if (!hwc->sample_period)
4108                 return;
4109
4110         if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4111                 return perf_swevent_overflow(event, 1, nmi, data, regs);
4112
4113         if (local64_add_negative(nr, &hwc->period_left))
4114                 return;
4115
4116         perf_swevent_overflow(event, 0, nmi, data, regs);
4117 }
4118
4119 static int perf_exclude_event(struct perf_event *event,
4120                               struct pt_regs *regs)
4121 {
4122         if (regs) {
4123                 if (event->attr.exclude_user && user_mode(regs))
4124                         return 1;
4125
4126                 if (event->attr.exclude_kernel && !user_mode(regs))
4127                         return 1;
4128         }
4129
4130         return 0;
4131 }
4132
4133 static int perf_swevent_match(struct perf_event *event,
4134                                 enum perf_type_id type,
4135                                 u32 event_id,
4136                                 struct perf_sample_data *data,
4137                                 struct pt_regs *regs)
4138 {
4139         if (event->attr.type != type)
4140                 return 0;
4141
4142         if (event->attr.config != event_id)
4143                 return 0;
4144
4145         if (perf_exclude_event(event, regs))
4146                 return 0;
4147
4148         return 1;
4149 }
4150
4151 static inline u64 swevent_hash(u64 type, u32 event_id)
4152 {
4153         u64 val = event_id | (type << 32);
4154
4155         return hash_64(val, SWEVENT_HLIST_BITS);
4156 }
4157
4158 static inline struct hlist_head *
4159 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4160 {
4161         u64 hash = swevent_hash(type, event_id);
4162
4163         return &hlist->heads[hash];
4164 }
4165
4166 /* For the read side: events when they trigger */
4167 static inline struct hlist_head *
4168 find_swevent_head_rcu(struct perf_cpu_context *ctx, u64 type, u32 event_id)
4169 {
4170         struct swevent_hlist *hlist;
4171
4172         hlist = rcu_dereference(ctx->swevent_hlist);
4173         if (!hlist)
4174                 return NULL;
4175
4176         return __find_swevent_head(hlist, type, event_id);
4177 }
4178
4179 /* For the event head insertion and removal in the hlist */
4180 static inline struct hlist_head *
4181 find_swevent_head(struct perf_cpu_context *ctx, struct perf_event *event)
4182 {
4183         struct swevent_hlist *hlist;
4184         u32 event_id = event->attr.config;
4185         u64 type = event->attr.type;
4186
4187         /*
4188          * Event scheduling is always serialized against hlist allocation
4189          * and release. Which makes the protected version suitable here.
4190          * The context lock guarantees that.
4191          */
4192         hlist = rcu_dereference_protected(ctx->swevent_hlist,
4193                                           lockdep_is_held(&event->ctx->lock));
4194         if (!hlist)
4195                 return NULL;
4196
4197         return __find_swevent_head(hlist, type, event_id);
4198 }
4199
4200 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4201                                     u64 nr, int nmi,
4202                                     struct perf_sample_data *data,
4203                                     struct pt_regs *regs)
4204 {
4205         struct perf_cpu_context *cpuctx;
4206         struct perf_event *event;
4207         struct hlist_node *node;
4208         struct hlist_head *head;
4209
4210         cpuctx = &__get_cpu_var(perf_cpu_context);
4211
4212         rcu_read_lock();
4213
4214         head = find_swevent_head_rcu(cpuctx, type, event_id);
4215
4216         if (!head)
4217                 goto end;
4218
4219         hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4220                 if (perf_swevent_match(event, type, event_id, data, regs))
4221                         perf_swevent_add(event, nr, nmi, data, regs);
4222         }
4223 end:
4224         rcu_read_unlock();
4225 }
4226
4227 int perf_swevent_get_recursion_context(void)
4228 {
4229         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4230         int rctx;
4231
4232         if (in_nmi())
4233                 rctx = 3;
4234         else if (in_irq())
4235                 rctx = 2;
4236         else if (in_softirq())
4237                 rctx = 1;
4238         else
4239                 rctx = 0;
4240
4241         if (cpuctx->recursion[rctx])
4242                 return -1;
4243
4244         cpuctx->recursion[rctx]++;
4245         barrier();
4246
4247         return rctx;
4248 }
4249 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4250
4251 void inline perf_swevent_put_recursion_context(int rctx)
4252 {
4253         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4254         barrier();
4255         cpuctx->recursion[rctx]--;
4256 }
4257
4258 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4259                             struct pt_regs *regs, u64 addr)
4260 {
4261         struct perf_sample_data data;
4262         int rctx;
4263
4264         preempt_disable_notrace();
4265         rctx = perf_swevent_get_recursion_context();
4266         if (rctx < 0)
4267                 return;
4268
4269         perf_sample_data_init(&data, addr);
4270
4271         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4272
4273         perf_swevent_put_recursion_context(rctx);
4274         preempt_enable_notrace();
4275 }
4276
4277 static void perf_swevent_read(struct perf_event *event)
4278 {
4279 }
4280
4281 static int perf_swevent_enable(struct perf_event *event)
4282 {
4283         struct hw_perf_event *hwc = &event->hw;
4284         struct perf_cpu_context *cpuctx;
4285         struct hlist_head *head;
4286
4287         cpuctx = &__get_cpu_var(perf_cpu_context);
4288
4289         if (hwc->sample_period) {
4290                 hwc->last_period = hwc->sample_period;
4291                 perf_swevent_set_period(event);
4292         }
4293
4294         head = find_swevent_head(cpuctx, event);
4295         if (WARN_ON_ONCE(!head))
4296                 return -EINVAL;
4297
4298         hlist_add_head_rcu(&event->hlist_entry, head);
4299
4300         return 0;
4301 }
4302
4303 static void perf_swevent_disable(struct perf_event *event)
4304 {
4305         hlist_del_rcu(&event->hlist_entry);
4306 }
4307
4308 static void perf_swevent_void(struct perf_event *event)
4309 {
4310 }
4311
4312 static int perf_swevent_int(struct perf_event *event)
4313 {
4314         return 0;
4315 }
4316
4317 static const struct pmu perf_ops_generic = {
4318         .enable         = perf_swevent_enable,
4319         .disable        = perf_swevent_disable,
4320         .start          = perf_swevent_int,
4321         .stop           = perf_swevent_void,
4322         .read           = perf_swevent_read,
4323         .unthrottle     = perf_swevent_void, /* hwc->interrupts already reset */
4324 };
4325
4326 /*
4327  * hrtimer based swevent callback
4328  */
4329
4330 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4331 {
4332         enum hrtimer_restart ret = HRTIMER_RESTART;
4333         struct perf_sample_data data;
4334         struct pt_regs *regs;
4335         struct perf_event *event;
4336         u64 period;
4337
4338         event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4339         event->pmu->read(event);
4340
4341         perf_sample_data_init(&data, 0);
4342         data.period = event->hw.last_period;
4343         regs = get_irq_regs();
4344
4345         if (regs && !perf_exclude_event(event, regs)) {
4346                 if (!(event->attr.exclude_idle && current->pid == 0))
4347                         if (perf_event_overflow(event, 0, &data, regs))
4348                                 ret = HRTIMER_NORESTART;
4349         }
4350
4351         period = max_t(u64, 10000, event->hw.sample_period);
4352         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4353
4354         return ret;
4355 }
4356
4357 static void perf_swevent_start_hrtimer(struct perf_event *event)
4358 {
4359         struct hw_perf_event *hwc = &event->hw;
4360
4361         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4362         hwc->hrtimer.function = perf_swevent_hrtimer;
4363         if (hwc->sample_period) {
4364                 u64 period;
4365
4366                 if (hwc->remaining) {
4367                         if (hwc->remaining < 0)
4368                                 period = 10000;
4369                         else
4370                                 period = hwc->remaining;
4371                         hwc->remaining = 0;
4372                 } else {
4373                         period = max_t(u64, 10000, hwc->sample_period);
4374                 }
4375                 __hrtimer_start_range_ns(&hwc->hrtimer,
4376                                 ns_to_ktime(period), 0,
4377                                 HRTIMER_MODE_REL, 0);
4378         }
4379 }
4380
4381 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4382 {
4383         struct hw_perf_event *hwc = &event->hw;
4384
4385         if (hwc->sample_period) {
4386                 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4387                 hwc->remaining = ktime_to_ns(remaining);
4388
4389                 hrtimer_cancel(&hwc->hrtimer);
4390         }
4391 }
4392
4393 /*
4394  * Software event: cpu wall time clock
4395  */
4396
4397 static void cpu_clock_perf_event_update(struct perf_event *event)
4398 {
4399         int cpu = raw_smp_processor_id();
4400         s64 prev;
4401         u64 now;
4402
4403         now = cpu_clock(cpu);
4404         prev = local64_xchg(&event->hw.prev_count, now);
4405         local64_add(now - prev, &event->count);
4406 }
4407
4408 static int cpu_clock_perf_event_enable(struct perf_event *event)
4409 {
4410         struct hw_perf_event *hwc = &event->hw;
4411         int cpu = raw_smp_processor_id();
4412
4413         local64_set(&hwc->prev_count, cpu_clock(cpu));
4414         perf_swevent_start_hrtimer(event);
4415
4416         return 0;
4417 }
4418
4419 static void cpu_clock_perf_event_disable(struct perf_event *event)
4420 {
4421         perf_swevent_cancel_hrtimer(event);
4422         cpu_clock_perf_event_update(event);
4423 }
4424
4425 static void cpu_clock_perf_event_read(struct perf_event *event)
4426 {
4427         cpu_clock_perf_event_update(event);
4428 }
4429
4430 static const struct pmu perf_ops_cpu_clock = {
4431         .enable         = cpu_clock_perf_event_enable,
4432         .disable        = cpu_clock_perf_event_disable,
4433         .read           = cpu_clock_perf_event_read,
4434 };
4435
4436 /*
4437  * Software event: task time clock
4438  */
4439
4440 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4441 {
4442         u64 prev;
4443         s64 delta;
4444
4445         prev = local64_xchg(&event->hw.prev_count, now);
4446         delta = now - prev;
4447         local64_add(delta, &event->count);
4448 }
4449
4450 static int task_clock_perf_event_enable(struct perf_event *event)
4451 {
4452         struct hw_perf_event *hwc = &event->hw;
4453         u64 now;
4454
4455         now = event->ctx->time;
4456
4457         local64_set(&hwc->prev_count, now);
4458
4459         perf_swevent_start_hrtimer(event);
4460
4461         return 0;
4462 }
4463
4464 static void task_clock_perf_event_disable(struct perf_event *event)
4465 {
4466         perf_swevent_cancel_hrtimer(event);
4467         task_clock_perf_event_update(event, event->ctx->time);
4468
4469 }
4470
4471 static void task_clock_perf_event_read(struct perf_event *event)
4472 {
4473         u64 time;
4474
4475         if (!in_nmi()) {
4476                 update_context_time(event->ctx);
4477                 time = event->ctx->time;
4478         } else {
4479                 u64 now = perf_clock();
4480                 u64 delta = now - event->ctx->timestamp;
4481                 time = event->ctx->time + delta;
4482         }
4483
4484         task_clock_perf_event_update(event, time);
4485 }
4486
4487 static const struct pmu perf_ops_task_clock = {
4488         .enable         = task_clock_perf_event_enable,
4489         .disable        = task_clock_perf_event_disable,
4490         .read           = task_clock_perf_event_read,
4491 };
4492
4493 /* Deref the hlist from the update side */
4494 static inline struct swevent_hlist *
4495 swevent_hlist_deref(struct perf_cpu_context *cpuctx)
4496 {
4497         return rcu_dereference_protected(cpuctx->swevent_hlist,
4498                                          lockdep_is_held(&cpuctx->hlist_mutex));
4499 }
4500
4501 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4502 {
4503         struct swevent_hlist *hlist;
4504
4505         hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4506         kfree(hlist);
4507 }
4508
4509 static void swevent_hlist_release(struct perf_cpu_context *cpuctx)
4510 {
4511         struct swevent_hlist *hlist = swevent_hlist_deref(cpuctx);
4512
4513         if (!hlist)
4514                 return;
4515
4516         rcu_assign_pointer(cpuctx->swevent_hlist, NULL);
4517         call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4518 }
4519
4520 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4521 {
4522         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4523
4524         mutex_lock(&cpuctx->hlist_mutex);
4525
4526         if (!--cpuctx->hlist_refcount)
4527                 swevent_hlist_release(cpuctx);
4528
4529         mutex_unlock(&cpuctx->hlist_mutex);
4530 }
4531
4532 static void swevent_hlist_put(struct perf_event *event)
4533 {
4534         int cpu;
4535
4536         if (event->cpu != -1) {
4537                 swevent_hlist_put_cpu(event, event->cpu);
4538                 return;
4539         }
4540
4541         for_each_possible_cpu(cpu)
4542                 swevent_hlist_put_cpu(event, cpu);
4543 }
4544
4545 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4546 {
4547         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4548         int err = 0;
4549
4550         mutex_lock(&cpuctx->hlist_mutex);
4551
4552         if (!swevent_hlist_deref(cpuctx) && cpu_online(cpu)) {
4553                 struct swevent_hlist *hlist;
4554
4555                 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4556                 if (!hlist) {
4557                         err = -ENOMEM;
4558                         goto exit;
4559                 }
4560                 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
4561         }
4562         cpuctx->hlist_refcount++;
4563  exit:
4564         mutex_unlock(&cpuctx->hlist_mutex);
4565
4566         return err;
4567 }
4568
4569 static int swevent_hlist_get(struct perf_event *event)
4570 {
4571         int err;
4572         int cpu, failed_cpu;
4573
4574         if (event->cpu != -1)
4575                 return swevent_hlist_get_cpu(event, event->cpu);
4576
4577         get_online_cpus();
4578         for_each_possible_cpu(cpu) {
4579                 err = swevent_hlist_get_cpu(event, cpu);
4580                 if (err) {
4581                         failed_cpu = cpu;
4582                         goto fail;
4583                 }
4584         }
4585         put_online_cpus();
4586
4587         return 0;
4588  fail:
4589         for_each_possible_cpu(cpu) {
4590                 if (cpu == failed_cpu)
4591                         break;
4592                 swevent_hlist_put_cpu(event, cpu);
4593         }
4594
4595         put_online_cpus();
4596         return err;
4597 }
4598
4599 #ifdef CONFIG_EVENT_TRACING
4600
4601 static const struct pmu perf_ops_tracepoint = {
4602         .enable         = perf_trace_enable,
4603         .disable        = perf_trace_disable,
4604         .start          = perf_swevent_int,
4605         .stop           = perf_swevent_void,
4606         .read           = perf_swevent_read,
4607         .unthrottle     = perf_swevent_void,
4608 };
4609
4610 static int perf_tp_filter_match(struct perf_event *event,
4611                                 struct perf_sample_data *data)
4612 {
4613         void *record = data->raw->data;
4614
4615         if (likely(!event->filter) || filter_match_preds(event->filter, record))
4616                 return 1;
4617         return 0;
4618 }
4619
4620 static int perf_tp_event_match(struct perf_event *event,
4621                                 struct perf_sample_data *data,
4622                                 struct pt_regs *regs)
4623 {
4624         /*
4625          * All tracepoints are from kernel-space.
4626          */
4627         if (event->attr.exclude_kernel)
4628                 return 0;
4629
4630         if (!perf_tp_filter_match(event, data))
4631                 return 0;
4632
4633         return 1;
4634 }
4635
4636 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
4637                    struct pt_regs *regs, struct hlist_head *head, int rctx)
4638 {
4639         struct perf_sample_data data;
4640         struct perf_event *event;
4641         struct hlist_node *node;
4642
4643         struct perf_raw_record raw = {
4644                 .size = entry_size,
4645                 .data = record,
4646         };
4647
4648         perf_sample_data_init(&data, addr);
4649         data.raw = &raw;
4650
4651         hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4652                 if (perf_tp_event_match(event, &data, regs))
4653                         perf_swevent_add(event, count, 1, &data, regs);
4654         }
4655
4656         perf_swevent_put_recursion_context(rctx);
4657 }
4658 EXPORT_SYMBOL_GPL(perf_tp_event);
4659
4660 static void tp_perf_event_destroy(struct perf_event *event)
4661 {
4662         perf_trace_destroy(event);
4663 }
4664
4665 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4666 {
4667         int err;
4668
4669         /*
4670          * Raw tracepoint data is a severe data leak, only allow root to
4671          * have these.
4672          */
4673         if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4674                         perf_paranoid_tracepoint_raw() &&
4675                         !capable(CAP_SYS_ADMIN))
4676                 return ERR_PTR(-EPERM);
4677
4678         err = perf_trace_init(event);
4679         if (err)
4680                 return NULL;
4681
4682         event->destroy = tp_perf_event_destroy;
4683
4684         return &perf_ops_tracepoint;
4685 }
4686
4687 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4688 {
4689         char *filter_str;
4690         int ret;
4691
4692         if (event->attr.type != PERF_TYPE_TRACEPOINT)
4693                 return -EINVAL;
4694
4695         filter_str = strndup_user(arg, PAGE_SIZE);
4696         if (IS_ERR(filter_str))
4697                 return PTR_ERR(filter_str);
4698
4699         ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4700
4701         kfree(filter_str);
4702         return ret;
4703 }
4704
4705 static void perf_event_free_filter(struct perf_event *event)
4706 {
4707         ftrace_profile_free_filter(event);
4708 }
4709
4710 #else
4711
4712 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4713 {
4714         return NULL;
4715 }
4716
4717 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4718 {
4719         return -ENOENT;
4720 }
4721
4722 static void perf_event_free_filter(struct perf_event *event)
4723 {
4724 }
4725
4726 #endif /* CONFIG_EVENT_TRACING */
4727
4728 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4729 static void bp_perf_event_destroy(struct perf_event *event)
4730 {
4731         release_bp_slot(event);
4732 }
4733
4734 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4735 {
4736         int err;
4737
4738         err = register_perf_hw_breakpoint(bp);
4739         if (err)
4740                 return ERR_PTR(err);
4741
4742         bp->destroy = bp_perf_event_destroy;
4743
4744         return &perf_ops_bp;
4745 }
4746
4747 void perf_bp_event(struct perf_event *bp, void *data)
4748 {
4749         struct perf_sample_data sample;
4750         struct pt_regs *regs = data;
4751
4752         perf_sample_data_init(&sample, bp->attr.bp_addr);
4753
4754         if (!perf_exclude_event(bp, regs))
4755                 perf_swevent_add(bp, 1, 1, &sample, regs);
4756 }
4757 #else
4758 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4759 {
4760         return NULL;
4761 }
4762
4763 void perf_bp_event(struct perf_event *bp, void *regs)
4764 {
4765 }
4766 #endif
4767
4768 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4769
4770 static void sw_perf_event_destroy(struct perf_event *event)
4771 {
4772         u64 event_id = event->attr.config;
4773
4774         WARN_ON(event->parent);
4775
4776         atomic_dec(&perf_swevent_enabled[event_id]);
4777         swevent_hlist_put(event);
4778 }
4779
4780 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4781 {
4782         const struct pmu *pmu = NULL;
4783         u64 event_id = event->attr.config;
4784
4785         /*
4786          * Software events (currently) can't in general distinguish
4787          * between user, kernel and hypervisor events.
4788          * However, context switches and cpu migrations are considered
4789          * to be kernel events, and page faults are never hypervisor
4790          * events.
4791          */
4792         switch (event_id) {
4793         case PERF_COUNT_SW_CPU_CLOCK:
4794                 pmu = &perf_ops_cpu_clock;
4795
4796                 break;
4797         case PERF_COUNT_SW_TASK_CLOCK:
4798                 /*
4799                  * If the user instantiates this as a per-cpu event,
4800                  * use the cpu_clock event instead.
4801                  */
4802                 if (event->ctx->task)
4803                         pmu = &perf_ops_task_clock;
4804                 else
4805                         pmu = &perf_ops_cpu_clock;
4806
4807                 break;
4808         case PERF_COUNT_SW_PAGE_FAULTS:
4809         case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4810         case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4811         case PERF_COUNT_SW_CONTEXT_SWITCHES:
4812         case PERF_COUNT_SW_CPU_MIGRATIONS:
4813         case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4814         case PERF_COUNT_SW_EMULATION_FAULTS:
4815                 if (!event->parent) {
4816                         int err;
4817
4818                         err = swevent_hlist_get(event);
4819                         if (err)
4820                                 return ERR_PTR(err);
4821
4822                         atomic_inc(&perf_swevent_enabled[event_id]);
4823                         event->destroy = sw_perf_event_destroy;
4824                 }
4825                 pmu = &perf_ops_generic;
4826                 break;
4827         }
4828
4829         return pmu;
4830 }
4831
4832 /*
4833  * Allocate and initialize a event structure
4834  */
4835 static struct perf_event *
4836 perf_event_alloc(struct perf_event_attr *attr,
4837                    int cpu,
4838                    struct perf_event_context *ctx,
4839                    struct perf_event *group_leader,
4840                    struct perf_event *parent_event,
4841                    perf_overflow_handler_t overflow_handler,
4842                    gfp_t gfpflags)
4843 {
4844         const struct pmu *pmu;
4845         struct perf_event *event;
4846         struct hw_perf_event *hwc;
4847         long err;
4848
4849         event = kzalloc(sizeof(*event), gfpflags);
4850         if (!event)
4851                 return ERR_PTR(-ENOMEM);
4852
4853         /*
4854          * Single events are their own group leaders, with an
4855          * empty sibling list:
4856          */
4857         if (!group_leader)
4858                 group_leader = event;
4859
4860         mutex_init(&event->child_mutex);
4861         INIT_LIST_HEAD(&event->child_list);
4862
4863         INIT_LIST_HEAD(&event->group_entry);
4864         INIT_LIST_HEAD(&event->event_entry);
4865         INIT_LIST_HEAD(&event->sibling_list);
4866         init_waitqueue_head(&event->waitq);
4867
4868         mutex_init(&event->mmap_mutex);
4869
4870         event->cpu              = cpu;
4871         event->attr             = *attr;
4872         event->group_leader     = group_leader;
4873         event->pmu              = NULL;
4874         event->ctx              = ctx;
4875         event->oncpu            = -1;
4876
4877         event->parent           = parent_event;
4878
4879         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
4880         event->id               = atomic64_inc_return(&perf_event_id);
4881
4882         event->state            = PERF_EVENT_STATE_INACTIVE;
4883
4884         if (!overflow_handler && parent_event)
4885                 overflow_handler = parent_event->overflow_handler;
4886         
4887         event->overflow_handler = overflow_handler;
4888
4889         if (attr->disabled)
4890                 event->state = PERF_EVENT_STATE_OFF;
4891
4892         pmu = NULL;
4893
4894         hwc = &event->hw;
4895         hwc->sample_period = attr->sample_period;
4896         if (attr->freq && attr->sample_freq)
4897                 hwc->sample_period = 1;
4898         hwc->last_period = hwc->sample_period;
4899
4900         local64_set(&hwc->period_left, hwc->sample_period);
4901
4902         /*
4903          * we currently do not support PERF_FORMAT_GROUP on inherited events
4904          */
4905         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4906                 goto done;
4907
4908         switch (attr->type) {
4909         case PERF_TYPE_RAW:
4910         case PERF_TYPE_HARDWARE:
4911         case PERF_TYPE_HW_CACHE:
4912                 pmu = hw_perf_event_init(event);
4913                 break;
4914
4915         case PERF_TYPE_SOFTWARE:
4916                 pmu = sw_perf_event_init(event);
4917                 break;
4918
4919         case PERF_TYPE_TRACEPOINT:
4920                 pmu = tp_perf_event_init(event);
4921                 break;
4922
4923         case PERF_TYPE_BREAKPOINT:
4924                 pmu = bp_perf_event_init(event);
4925                 break;
4926
4927
4928         default:
4929                 break;
4930         }
4931 done:
4932         err = 0;
4933         if (!pmu)
4934                 err = -EINVAL;
4935         else if (IS_ERR(pmu))
4936                 err = PTR_ERR(pmu);
4937
4938         if (err) {
4939                 if (event->ns)
4940                         put_pid_ns(event->ns);
4941                 kfree(event);
4942                 return ERR_PTR(err);
4943         }
4944
4945         event->pmu = pmu;
4946
4947         if (!event->parent) {
4948                 atomic_inc(&nr_events);
4949                 if (event->attr.mmap || event->attr.mmap_data)
4950                         atomic_inc(&nr_mmap_events);
4951                 if (event->attr.comm)
4952                         atomic_inc(&nr_comm_events);
4953                 if (event->attr.task)
4954                         atomic_inc(&nr_task_events);
4955         }
4956
4957         return event;
4958 }
4959
4960 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4961                           struct perf_event_attr *attr)
4962 {
4963         u32 size;
4964         int ret;
4965
4966         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4967                 return -EFAULT;
4968
4969         /*
4970          * zero the full structure, so that a short copy will be nice.
4971          */
4972         memset(attr, 0, sizeof(*attr));
4973
4974         ret = get_user(size, &uattr->size);
4975         if (ret)
4976                 return ret;
4977
4978         if (size > PAGE_SIZE)   /* silly large */
4979                 goto err_size;
4980
4981         if (!size)              /* abi compat */
4982                 size = PERF_ATTR_SIZE_VER0;
4983
4984         if (size < PERF_ATTR_SIZE_VER0)
4985                 goto err_size;
4986
4987         /*
4988          * If we're handed a bigger struct than we know of,
4989          * ensure all the unknown bits are 0 - i.e. new
4990          * user-space does not rely on any kernel feature
4991          * extensions we dont know about yet.
4992          */
4993         if (size > sizeof(*attr)) {
4994                 unsigned char __user *addr;
4995                 unsigned char __user *end;
4996                 unsigned char val;
4997
4998                 addr = (void __user *)uattr + sizeof(*attr);
4999                 end  = (void __user *)uattr + size;
5000
5001                 for (; addr < end; addr++) {
5002                         ret = get_user(val, addr);
5003                         if (ret)
5004                                 return ret;
5005                         if (val)
5006                                 goto err_size;
5007                 }
5008                 size = sizeof(*attr);
5009         }
5010
5011         ret = copy_from_user(attr, uattr, size);
5012         if (ret)
5013                 return -EFAULT;
5014
5015         /*
5016          * If the type exists, the corresponding creation will verify
5017          * the attr->config.
5018          */
5019         if (attr->type >= PERF_TYPE_MAX)
5020                 return -EINVAL;
5021
5022         if (attr->__reserved_1)
5023                 return -EINVAL;
5024
5025         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
5026                 return -EINVAL;
5027
5028         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
5029                 return -EINVAL;
5030
5031 out:
5032         return ret;
5033
5034 err_size:
5035         put_user(sizeof(*attr), &uattr->size);
5036         ret = -E2BIG;
5037         goto out;
5038 }
5039
5040 static int
5041 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
5042 {
5043         struct perf_buffer *buffer = NULL, *old_buffer = NULL;
5044         int ret = -EINVAL;
5045
5046         if (!output_event)
5047                 goto set;
5048
5049         /* don't allow circular references */
5050         if (event == output_event)
5051                 goto out;
5052
5053         /*
5054          * Don't allow cross-cpu buffers
5055          */
5056         if (output_event->cpu != event->cpu)
5057                 goto out;
5058
5059         /*
5060          * If its not a per-cpu buffer, it must be the same task.
5061          */
5062         if (output_event->cpu == -1 && output_event->ctx != event->ctx)
5063                 goto out;
5064
5065 set:
5066         mutex_lock(&event->mmap_mutex);
5067         /* Can't redirect output if we've got an active mmap() */
5068         if (atomic_read(&event->mmap_count))
5069                 goto unlock;
5070
5071         if (output_event) {
5072                 /* get the buffer we want to redirect to */
5073                 buffer = perf_buffer_get(output_event);
5074                 if (!buffer)
5075                         goto unlock;
5076         }
5077
5078         old_buffer = event->buffer;
5079         rcu_assign_pointer(event->buffer, buffer);
5080         ret = 0;
5081 unlock:
5082         mutex_unlock(&event->mmap_mutex);
5083
5084         if (old_buffer)
5085                 perf_buffer_put(old_buffer);
5086 out:
5087         return ret;
5088 }
5089
5090 /**
5091  * sys_perf_event_open - open a performance event, associate it to a task/cpu
5092  *
5093  * @attr_uptr:  event_id type attributes for monitoring/sampling
5094  * @pid:                target pid
5095  * @cpu:                target cpu
5096  * @group_fd:           group leader event fd
5097  */
5098 SYSCALL_DEFINE5(perf_event_open,
5099                 struct perf_event_attr __user *, attr_uptr,
5100                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
5101 {
5102         struct perf_event *event, *group_leader = NULL, *output_event = NULL;
5103         struct perf_event_attr attr;
5104         struct perf_event_context *ctx;
5105         struct file *event_file = NULL;
5106         struct file *group_file = NULL;
5107         int event_fd;
5108         int fput_needed = 0;
5109         int err;
5110
5111         /* for future expandability... */
5112         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
5113                 return -EINVAL;
5114
5115         err = perf_copy_attr(attr_uptr, &attr);
5116         if (err)
5117                 return err;
5118
5119         if (!attr.exclude_kernel) {
5120                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
5121                         return -EACCES;
5122         }
5123
5124         if (attr.freq) {
5125                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
5126                         return -EINVAL;
5127         }
5128
5129         event_fd = get_unused_fd_flags(O_RDWR);
5130         if (event_fd < 0)
5131                 return event_fd;
5132
5133         /*
5134          * Get the target context (task or percpu):
5135          */
5136         ctx = find_get_context(pid, cpu);
5137         if (IS_ERR(ctx)) {
5138                 err = PTR_ERR(ctx);
5139                 goto err_fd;
5140         }
5141
5142         if (group_fd != -1) {
5143                 group_leader = perf_fget_light(group_fd, &fput_needed);
5144                 if (IS_ERR(group_leader)) {
5145                         err = PTR_ERR(group_leader);
5146                         goto err_put_context;
5147                 }
5148                 group_file = group_leader->filp;
5149                 if (flags & PERF_FLAG_FD_OUTPUT)
5150                         output_event = group_leader;
5151                 if (flags & PERF_FLAG_FD_NO_GROUP)
5152                         group_leader = NULL;
5153         }
5154
5155         /*
5156          * Look up the group leader (we will attach this event to it):
5157          */
5158         if (group_leader) {
5159                 err = -EINVAL;
5160
5161                 /*
5162                  * Do not allow a recursive hierarchy (this new sibling
5163                  * becoming part of another group-sibling):
5164                  */
5165                 if (group_leader->group_leader != group_leader)
5166                         goto err_put_context;
5167                 /*
5168                  * Do not allow to attach to a group in a different
5169                  * task or CPU context:
5170                  */
5171                 if (group_leader->ctx != ctx)
5172                         goto err_put_context;
5173                 /*
5174                  * Only a group leader can be exclusive or pinned
5175                  */
5176                 if (attr.exclusive || attr.pinned)
5177                         goto err_put_context;
5178         }
5179
5180         event = perf_event_alloc(&attr, cpu, ctx, group_leader,
5181                                      NULL, NULL, GFP_KERNEL);
5182         if (IS_ERR(event)) {
5183                 err = PTR_ERR(event);
5184                 goto err_put_context;
5185         }
5186
5187         if (output_event) {
5188                 err = perf_event_set_output(event, output_event);
5189                 if (err)
5190                         goto err_free_put_context;
5191         }
5192
5193         event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
5194         if (IS_ERR(event_file)) {
5195                 err = PTR_ERR(event_file);
5196                 goto err_free_put_context;
5197         }
5198
5199         event->filp = event_file;
5200         WARN_ON_ONCE(ctx->parent_ctx);
5201         mutex_lock(&ctx->mutex);
5202         perf_install_in_context(ctx, event, cpu);
5203         ++ctx->generation;
5204         mutex_unlock(&ctx->mutex);
5205
5206         event->owner = current;
5207         get_task_struct(current);
5208         mutex_lock(&current->perf_event_mutex);
5209         list_add_tail(&event->owner_entry, &current->perf_event_list);
5210         mutex_unlock(&current->perf_event_mutex);
5211
5212         /*
5213          * Drop the reference on the group_event after placing the
5214          * new event on the sibling_list. This ensures destruction
5215          * of the group leader will find the pointer to itself in
5216          * perf_group_detach().
5217          */
5218         fput_light(group_file, fput_needed);
5219         fd_install(event_fd, event_file);
5220         return event_fd;
5221
5222 err_free_put_context:
5223         free_event(event);
5224 err_put_context:
5225         fput_light(group_file, fput_needed);
5226         put_ctx(ctx);
5227 err_fd:
5228         put_unused_fd(event_fd);
5229         return err;
5230 }
5231
5232 /**
5233  * perf_event_create_kernel_counter
5234  *
5235  * @attr: attributes of the counter to create
5236  * @cpu: cpu in which the counter is bound
5237  * @pid: task to profile
5238  */
5239 struct perf_event *
5240 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
5241                                  pid_t pid,
5242                                  perf_overflow_handler_t overflow_handler)
5243 {
5244         struct perf_event *event;
5245         struct perf_event_context *ctx;
5246         int err;
5247
5248         /*
5249          * Get the target context (task or percpu):
5250          */
5251
5252         ctx = find_get_context(pid, cpu);
5253         if (IS_ERR(ctx)) {
5254                 err = PTR_ERR(ctx);
5255                 goto err_exit;
5256         }
5257
5258         event = perf_event_alloc(attr, cpu, ctx, NULL,
5259                                  NULL, overflow_handler, GFP_KERNEL);
5260         if (IS_ERR(event)) {
5261                 err = PTR_ERR(event);
5262                 goto err_put_context;
5263         }
5264
5265         event->filp = NULL;
5266         WARN_ON_ONCE(ctx->parent_ctx);
5267         mutex_lock(&ctx->mutex);
5268         perf_install_in_context(ctx, event, cpu);
5269         ++ctx->generation;
5270         mutex_unlock(&ctx->mutex);
5271
5272         event->owner = current;
5273         get_task_struct(current);
5274         mutex_lock(&current->perf_event_mutex);
5275         list_add_tail(&event->owner_entry, &current->perf_event_list);
5276         mutex_unlock(&current->perf_event_mutex);
5277
5278         return event;
5279
5280  err_put_context:
5281         put_ctx(ctx);
5282  err_exit:
5283         return ERR_PTR(err);
5284 }
5285 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
5286
5287 /*
5288  * inherit a event from parent task to child task:
5289  */
5290 static struct perf_event *
5291 inherit_event(struct perf_event *parent_event,
5292               struct task_struct *parent,
5293               struct perf_event_context *parent_ctx,
5294               struct task_struct *child,
5295               struct perf_event *group_leader,
5296               struct perf_event_context *child_ctx)
5297 {
5298         struct perf_event *child_event;
5299
5300         /*
5301          * Instead of creating recursive hierarchies of events,
5302          * we link inherited events back to the original parent,
5303          * which has a filp for sure, which we use as the reference
5304          * count:
5305          */
5306         if (parent_event->parent)
5307                 parent_event = parent_event->parent;
5308
5309         child_event = perf_event_alloc(&parent_event->attr,
5310                                            parent_event->cpu, child_ctx,
5311                                            group_leader, parent_event,
5312                                            NULL, GFP_KERNEL);
5313         if (IS_ERR(child_event))
5314                 return child_event;
5315         get_ctx(child_ctx);
5316
5317         /*
5318          * Make the child state follow the state of the parent event,
5319          * not its attr.disabled bit.  We hold the parent's mutex,
5320          * so we won't race with perf_event_{en, dis}able_family.
5321          */
5322         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
5323                 child_event->state = PERF_EVENT_STATE_INACTIVE;
5324         else
5325                 child_event->state = PERF_EVENT_STATE_OFF;
5326
5327         if (parent_event->attr.freq) {
5328                 u64 sample_period = parent_event->hw.sample_period;
5329                 struct hw_perf_event *hwc = &child_event->hw;
5330
5331                 hwc->sample_period = sample_period;
5332                 hwc->last_period   = sample_period;
5333
5334                 local64_set(&hwc->period_left, sample_period);
5335         }
5336
5337         child_event->overflow_handler = parent_event->overflow_handler;
5338
5339         /*
5340          * Link it up in the child's context:
5341          */
5342         add_event_to_ctx(child_event, child_ctx);
5343
5344         /*
5345          * Get a reference to the parent filp - we will fput it
5346          * when the child event exits. This is safe to do because
5347          * we are in the parent and we know that the filp still
5348          * exists and has a nonzero count:
5349          */
5350         atomic_long_inc(&parent_event->filp->f_count);
5351
5352         /*
5353          * Link this into the parent event's child list
5354          */
5355         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5356         mutex_lock(&parent_event->child_mutex);
5357         list_add_tail(&child_event->child_list, &parent_event->child_list);
5358         mutex_unlock(&parent_event->child_mutex);
5359
5360         return child_event;
5361 }
5362
5363 static int inherit_group(struct perf_event *parent_event,
5364               struct task_struct *parent,
5365               struct perf_event_context *parent_ctx,
5366               struct task_struct *child,
5367               struct perf_event_context *child_ctx)
5368 {
5369         struct perf_event *leader;
5370         struct perf_event *sub;
5371         struct perf_event *child_ctr;
5372
5373         leader = inherit_event(parent_event, parent, parent_ctx,
5374                                  child, NULL, child_ctx);
5375         if (IS_ERR(leader))
5376                 return PTR_ERR(leader);
5377         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5378                 child_ctr = inherit_event(sub, parent, parent_ctx,
5379                                             child, leader, child_ctx);
5380                 if (IS_ERR(child_ctr))
5381                         return PTR_ERR(child_ctr);
5382         }
5383         return 0;
5384 }
5385
5386 static void sync_child_event(struct perf_event *child_event,
5387                                struct task_struct *child)
5388 {
5389         struct perf_event *parent_event = child_event->parent;
5390         u64 child_val;
5391
5392         if (child_event->attr.inherit_stat)
5393                 perf_event_read_event(child_event, child);
5394
5395         child_val = perf_event_count(child_event);
5396
5397         /*
5398          * Add back the child's count to the parent's count:
5399          */
5400         atomic64_add(child_val, &parent_event->child_count);
5401         atomic64_add(child_event->total_time_enabled,
5402                      &parent_event->child_total_time_enabled);
5403         atomic64_add(child_event->total_time_running,
5404                      &parent_event->child_total_time_running);
5405
5406         /*
5407          * Remove this event from the parent's list
5408          */
5409         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5410         mutex_lock(&parent_event->child_mutex);
5411         list_del_init(&child_event->child_list);
5412         mutex_unlock(&parent_event->child_mutex);
5413
5414         /*
5415          * Release the parent event, if this was the last
5416          * reference to it.
5417          */
5418         fput(parent_event->filp);
5419 }
5420
5421 static void
5422 __perf_event_exit_task(struct perf_event *child_event,
5423                          struct perf_event_context *child_ctx,
5424                          struct task_struct *child)
5425 {
5426         struct perf_event *parent_event;
5427
5428         perf_event_remove_from_context(child_event);
5429
5430         parent_event = child_event->parent;
5431         /*
5432          * It can happen that parent exits first, and has events
5433          * that are still around due to the child reference. These
5434          * events need to be zapped - but otherwise linger.
5435          */
5436         if (parent_event) {
5437                 sync_child_event(child_event, child);
5438                 free_event(child_event);
5439         }
5440 }
5441
5442 /*
5443  * When a child task exits, feed back event values to parent events.
5444  */
5445 void perf_event_exit_task(struct task_struct *child)
5446 {
5447         struct perf_event *child_event, *tmp;
5448         struct perf_event_context *child_ctx;
5449         unsigned long flags;
5450
5451         if (likely(!child->perf_event_ctxp)) {
5452                 perf_event_task(child, NULL, 0);
5453                 return;
5454         }
5455
5456         local_irq_save(flags);
5457         /*
5458          * We can't reschedule here because interrupts are disabled,
5459          * and either child is current or it is a task that can't be
5460          * scheduled, so we are now safe from rescheduling changing
5461          * our context.
5462          */
5463         child_ctx = child->perf_event_ctxp;
5464         __perf_event_task_sched_out(child_ctx);
5465
5466         /*
5467          * Take the context lock here so that if find_get_context is
5468          * reading child->perf_event_ctxp, we wait until it has
5469          * incremented the context's refcount before we do put_ctx below.
5470          */
5471         raw_spin_lock(&child_ctx->lock);
5472         child->perf_event_ctxp = NULL;
5473         /*
5474          * If this context is a clone; unclone it so it can't get
5475          * swapped to another process while we're removing all
5476          * the events from it.
5477          */
5478         unclone_ctx(child_ctx);
5479         update_context_time(child_ctx);
5480         raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5481
5482         /*
5483          * Report the task dead after unscheduling the events so that we
5484          * won't get any samples after PERF_RECORD_EXIT. We can however still
5485          * get a few PERF_RECORD_READ events.
5486          */
5487         perf_event_task(child, child_ctx, 0);
5488
5489         /*
5490          * We can recurse on the same lock type through:
5491          *
5492          *   __perf_event_exit_task()
5493          *     sync_child_event()
5494          *       fput(parent_event->filp)
5495          *         perf_release()
5496          *           mutex_lock(&ctx->mutex)
5497          *
5498          * But since its the parent context it won't be the same instance.
5499          */
5500         mutex_lock(&child_ctx->mutex);
5501
5502 again:
5503         list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5504                                  group_entry)
5505                 __perf_event_exit_task(child_event, child_ctx, child);
5506
5507         list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5508                                  group_entry)
5509                 __perf_event_exit_task(child_event, child_ctx, child);
5510
5511         /*
5512          * If the last event was a group event, it will have appended all
5513          * its siblings to the list, but we obtained 'tmp' before that which
5514          * will still point to the list head terminating the iteration.
5515          */
5516         if (!list_empty(&child_ctx->pinned_groups) ||
5517             !list_empty(&child_ctx->flexible_groups))
5518                 goto again;
5519
5520         mutex_unlock(&child_ctx->mutex);
5521
5522         put_ctx(child_ctx);
5523 }
5524
5525 static void perf_free_event(struct perf_event *event,
5526                             struct perf_event_context *ctx)
5527 {
5528         struct perf_event *parent = event->parent;
5529
5530         if (WARN_ON_ONCE(!parent))
5531                 return;
5532
5533         mutex_lock(&parent->child_mutex);
5534         list_del_init(&event->child_list);
5535         mutex_unlock(&parent->child_mutex);
5536
5537         fput(parent->filp);
5538
5539         perf_group_detach(event);
5540         list_del_event(event, ctx);
5541         free_event(event);
5542 }
5543
5544 /*
5545  * free an unexposed, unused context as created by inheritance by
5546  * init_task below, used by fork() in case of fail.
5547  */
5548 void perf_event_free_task(struct task_struct *task)
5549 {
5550         struct perf_event_context *ctx = task->perf_event_ctxp;
5551         struct perf_event *event, *tmp;
5552
5553         if (!ctx)
5554                 return;
5555
5556         mutex_lock(&ctx->mutex);
5557 again:
5558         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5559                 perf_free_event(event, ctx);
5560
5561         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5562                                  group_entry)
5563                 perf_free_event(event, ctx);
5564
5565         if (!list_empty(&ctx->pinned_groups) ||
5566             !list_empty(&ctx->flexible_groups))
5567                 goto again;
5568
5569         mutex_unlock(&ctx->mutex);
5570
5571         put_ctx(ctx);
5572 }
5573
5574 static int
5575 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5576                    struct perf_event_context *parent_ctx,
5577                    struct task_struct *child,
5578                    int *inherited_all)
5579 {
5580         int ret;
5581         struct perf_event_context *child_ctx = child->perf_event_ctxp;
5582
5583         if (!event->attr.inherit) {
5584                 *inherited_all = 0;
5585                 return 0;
5586         }
5587
5588         if (!child_ctx) {
5589                 /*
5590                  * This is executed from the parent task context, so
5591                  * inherit events that have been marked for cloning.
5592                  * First allocate and initialize a context for the
5593                  * child.
5594                  */
5595
5596                 child_ctx = kzalloc(sizeof(struct perf_event_context),
5597                                     GFP_KERNEL);
5598                 if (!child_ctx)
5599                         return -ENOMEM;
5600
5601                 __perf_event_init_context(child_ctx, child);
5602                 child->perf_event_ctxp = child_ctx;
5603                 get_task_struct(child);
5604         }
5605
5606         ret = inherit_group(event, parent, parent_ctx,
5607                             child, child_ctx);
5608
5609         if (ret)
5610                 *inherited_all = 0;
5611
5612         return ret;
5613 }
5614
5615
5616 /*
5617  * Initialize the perf_event context in task_struct
5618  */
5619 int perf_event_init_task(struct task_struct *child)
5620 {
5621         struct perf_event_context *child_ctx, *parent_ctx;
5622         struct perf_event_context *cloned_ctx;
5623         struct perf_event *event;
5624         struct task_struct *parent = current;
5625         int inherited_all = 1;
5626         int ret = 0;
5627
5628         child->perf_event_ctxp = NULL;
5629
5630         mutex_init(&child->perf_event_mutex);
5631         INIT_LIST_HEAD(&child->perf_event_list);
5632
5633         if (likely(!parent->perf_event_ctxp))
5634                 return 0;
5635
5636         /*
5637          * If the parent's context is a clone, pin it so it won't get
5638          * swapped under us.
5639          */
5640         parent_ctx = perf_pin_task_context(parent);
5641
5642         /*
5643          * No need to check if parent_ctx != NULL here; since we saw
5644          * it non-NULL earlier, the only reason for it to become NULL
5645          * is if we exit, and since we're currently in the middle of
5646          * a fork we can't be exiting at the same time.
5647          */
5648
5649         /*
5650          * Lock the parent list. No need to lock the child - not PID
5651          * hashed yet and not running, so nobody can access it.
5652          */
5653         mutex_lock(&parent_ctx->mutex);
5654
5655         /*
5656          * We dont have to disable NMIs - we are only looking at
5657          * the list, not manipulating it:
5658          */
5659         list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5660                 ret = inherit_task_group(event, parent, parent_ctx, child,
5661                                          &inherited_all);
5662                 if (ret)
5663                         break;
5664         }
5665
5666         list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5667                 ret = inherit_task_group(event, parent, parent_ctx, child,
5668                                          &inherited_all);
5669                 if (ret)
5670                         break;
5671         }
5672
5673         child_ctx = child->perf_event_ctxp;
5674
5675         if (child_ctx && inherited_all) {
5676                 /*
5677                  * Mark the child context as a clone of the parent
5678                  * context, or of whatever the parent is a clone of.
5679                  * Note that if the parent is a clone, it could get
5680                  * uncloned at any point, but that doesn't matter
5681                  * because the list of events and the generation
5682                  * count can't have changed since we took the mutex.
5683                  */
5684                 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5685                 if (cloned_ctx) {
5686                         child_ctx->parent_ctx = cloned_ctx;
5687                         child_ctx->parent_gen = parent_ctx->parent_gen;
5688                 } else {
5689                         child_ctx->parent_ctx = parent_ctx;
5690                         child_ctx->parent_gen = parent_ctx->generation;
5691                 }
5692                 get_ctx(child_ctx->parent_ctx);
5693         }
5694
5695         mutex_unlock(&parent_ctx->mutex);
5696
5697         perf_unpin_context(parent_ctx);
5698
5699         return ret;
5700 }
5701
5702 static void __init perf_event_init_all_cpus(void)
5703 {
5704         int cpu;
5705         struct perf_cpu_context *cpuctx;
5706
5707         for_each_possible_cpu(cpu) {
5708                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5709                 mutex_init(&cpuctx->hlist_mutex);
5710                 __perf_event_init_context(&cpuctx->ctx, NULL);
5711         }
5712 }
5713
5714 static void __cpuinit perf_event_init_cpu(int cpu)
5715 {
5716         struct perf_cpu_context *cpuctx;
5717
5718         cpuctx = &per_cpu(perf_cpu_context, cpu);
5719
5720         spin_lock(&perf_resource_lock);
5721         cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5722         spin_unlock(&perf_resource_lock);
5723
5724         mutex_lock(&cpuctx->hlist_mutex);
5725         if (cpuctx->hlist_refcount > 0) {
5726                 struct swevent_hlist *hlist;
5727
5728                 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5729                 WARN_ON_ONCE(!hlist);
5730                 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
5731         }
5732         mutex_unlock(&cpuctx->hlist_mutex);
5733 }
5734
5735 #ifdef CONFIG_HOTPLUG_CPU
5736 static void __perf_event_exit_cpu(void *info)
5737 {
5738         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5739         struct perf_event_context *ctx = &cpuctx->ctx;
5740         struct perf_event *event, *tmp;
5741
5742         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5743                 __perf_event_remove_from_context(event);
5744         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5745                 __perf_event_remove_from_context(event);
5746 }
5747 static void perf_event_exit_cpu(int cpu)
5748 {
5749         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5750         struct perf_event_context *ctx = &cpuctx->ctx;
5751
5752         mutex_lock(&cpuctx->hlist_mutex);
5753         swevent_hlist_release(cpuctx);
5754         mutex_unlock(&cpuctx->hlist_mutex);
5755
5756         mutex_lock(&ctx->mutex);
5757         smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5758         mutex_unlock(&ctx->mutex);
5759 }
5760 #else
5761 static inline void perf_event_exit_cpu(int cpu) { }
5762 #endif
5763
5764 static int __cpuinit
5765 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5766 {
5767         unsigned int cpu = (long)hcpu;
5768
5769         switch (action & ~CPU_TASKS_FROZEN) {
5770
5771         case CPU_UP_PREPARE:
5772         case CPU_DOWN_FAILED:
5773                 perf_event_init_cpu(cpu);
5774                 break;
5775
5776         case CPU_UP_CANCELED:
5777         case CPU_DOWN_PREPARE:
5778                 perf_event_exit_cpu(cpu);
5779                 break;
5780
5781         default:
5782                 break;
5783         }
5784
5785         return NOTIFY_OK;
5786 }
5787
5788 /*
5789  * This has to have a higher priority than migration_notifier in sched.c.
5790  */
5791 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5792         .notifier_call          = perf_cpu_notify,
5793         .priority               = 20,
5794 };
5795
5796 void __init perf_event_init(void)
5797 {
5798         perf_event_init_all_cpus();
5799         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5800                         (void *)(long)smp_processor_id());
5801         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5802                         (void *)(long)smp_processor_id());
5803         register_cpu_notifier(&perf_cpu_nb);
5804 }
5805
5806 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class,
5807                                         struct sysdev_class_attribute *attr,
5808                                         char *buf)
5809 {
5810         return sprintf(buf, "%d\n", perf_reserved_percpu);
5811 }
5812
5813 static ssize_t
5814 perf_set_reserve_percpu(struct sysdev_class *class,
5815                         struct sysdev_class_attribute *attr,
5816                         const char *buf,
5817                         size_t count)
5818 {
5819         struct perf_cpu_context *cpuctx;
5820         unsigned long val;
5821         int err, cpu, mpt;
5822
5823         err = strict_strtoul(buf, 10, &val);
5824         if (err)
5825                 return err;
5826         if (val > perf_max_events)
5827                 return -EINVAL;
5828
5829         spin_lock(&perf_resource_lock);
5830         perf_reserved_percpu = val;
5831         for_each_online_cpu(cpu) {
5832                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5833                 raw_spin_lock_irq(&cpuctx->ctx.lock);
5834                 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5835                           perf_max_events - perf_reserved_percpu);
5836                 cpuctx->max_pertask = mpt;
5837                 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5838         }
5839         spin_unlock(&perf_resource_lock);
5840
5841         return count;
5842 }
5843
5844 static ssize_t perf_show_overcommit(struct sysdev_class *class,
5845                                     struct sysdev_class_attribute *attr,
5846                                     char *buf)
5847 {
5848         return sprintf(buf, "%d\n", perf_overcommit);
5849 }
5850
5851 static ssize_t
5852 perf_set_overcommit(struct sysdev_class *class,
5853                     struct sysdev_class_attribute *attr,
5854                     const char *buf, size_t count)
5855 {
5856         unsigned long val;
5857         int err;
5858
5859         err = strict_strtoul(buf, 10, &val);
5860         if (err)
5861                 return err;
5862         if (val > 1)
5863                 return -EINVAL;
5864
5865         spin_lock(&perf_resource_lock);
5866         perf_overcommit = val;
5867         spin_unlock(&perf_resource_lock);
5868
5869         return count;
5870 }
5871
5872 static SYSDEV_CLASS_ATTR(
5873                                 reserve_percpu,
5874                                 0644,
5875                                 perf_show_reserve_percpu,
5876                                 perf_set_reserve_percpu
5877                         );
5878
5879 static SYSDEV_CLASS_ATTR(
5880                                 overcommit,
5881                                 0644,
5882                                 perf_show_overcommit,
5883                                 perf_set_overcommit
5884                         );
5885
5886 static struct attribute *perfclass_attrs[] = {
5887         &attr_reserve_percpu.attr,
5888         &attr_overcommit.attr,
5889         NULL
5890 };
5891
5892 static struct attribute_group perfclass_attr_group = {
5893         .attrs                  = perfclass_attrs,
5894         .name                   = "perf_events",
5895 };
5896
5897 static int __init perf_event_sysfs_init(void)
5898 {
5899         return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5900                                   &perfclass_attr_group);
5901 }
5902 device_initcall(perf_event_sysfs_init);