perf events: Allow per-task-per-cpu counters
[linux-2.6.git] / kernel / perf_event.c
1 /*
2  * Performance events core code:
3  *
4  *  Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5  *  Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6  *  Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7  *  Copyright  ©  2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
8  *
9  * For licensing details see kernel-base/COPYING
10  */
11
12 #include <linux/fs.h>
13 #include <linux/mm.h>
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/file.h>
17 #include <linux/poll.h>
18 #include <linux/sysfs.h>
19 #include <linux/dcache.h>
20 #include <linux/percpu.h>
21 #include <linux/ptrace.h>
22 #include <linux/vmstat.h>
23 #include <linux/vmalloc.h>
24 #include <linux/hardirq.h>
25 #include <linux/rculist.h>
26 #include <linux/uaccess.h>
27 #include <linux/syscalls.h>
28 #include <linux/anon_inodes.h>
29 #include <linux/kernel_stat.h>
30 #include <linux/perf_event.h>
31 #include <linux/ftrace_event.h>
32 #include <linux/hw_breakpoint.h>
33
34 #include <asm/irq_regs.h>
35
36 /*
37  * Each CPU has a list of per CPU events:
38  */
39 static DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
40
41 int perf_max_events __read_mostly = 1;
42 static int perf_reserved_percpu __read_mostly;
43 static int perf_overcommit __read_mostly = 1;
44
45 static atomic_t nr_events __read_mostly;
46 static atomic_t nr_mmap_events __read_mostly;
47 static atomic_t nr_comm_events __read_mostly;
48 static atomic_t nr_task_events __read_mostly;
49
50 /*
51  * perf event paranoia level:
52  *  -1 - not paranoid at all
53  *   0 - disallow raw tracepoint access for unpriv
54  *   1 - disallow cpu events for unpriv
55  *   2 - disallow kernel profiling for unpriv
56  */
57 int sysctl_perf_event_paranoid __read_mostly = 1;
58
59 static inline bool perf_paranoid_tracepoint_raw(void)
60 {
61         return sysctl_perf_event_paranoid > -1;
62 }
63
64 static inline bool perf_paranoid_cpu(void)
65 {
66         return sysctl_perf_event_paranoid > 0;
67 }
68
69 static inline bool perf_paranoid_kernel(void)
70 {
71         return sysctl_perf_event_paranoid > 1;
72 }
73
74 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
75
76 /*
77  * max perf event sample rate
78  */
79 int sysctl_perf_event_sample_rate __read_mostly = 100000;
80
81 static atomic64_t perf_event_id;
82
83 /*
84  * Lock for (sysadmin-configurable) event reservations:
85  */
86 static DEFINE_SPINLOCK(perf_resource_lock);
87
88 /*
89  * Architecture provided APIs - weak aliases:
90  */
91 extern __weak const struct pmu *hw_perf_event_init(struct perf_event *event)
92 {
93         return NULL;
94 }
95
96 void __weak hw_perf_disable(void)               { barrier(); }
97 void __weak hw_perf_enable(void)                { barrier(); }
98
99 void __weak hw_perf_event_setup(int cpu)        { barrier(); }
100 void __weak hw_perf_event_setup_online(int cpu) { barrier(); }
101
102 int __weak
103 hw_perf_group_sched_in(struct perf_event *group_leader,
104                struct perf_cpu_context *cpuctx,
105                struct perf_event_context *ctx, int cpu)
106 {
107         return 0;
108 }
109
110 void __weak perf_event_print_debug(void)        { }
111
112 static DEFINE_PER_CPU(int, perf_disable_count);
113
114 void __perf_disable(void)
115 {
116         __get_cpu_var(perf_disable_count)++;
117 }
118
119 bool __perf_enable(void)
120 {
121         return !--__get_cpu_var(perf_disable_count);
122 }
123
124 void perf_disable(void)
125 {
126         __perf_disable();
127         hw_perf_disable();
128 }
129
130 void perf_enable(void)
131 {
132         if (__perf_enable())
133                 hw_perf_enable();
134 }
135
136 static void get_ctx(struct perf_event_context *ctx)
137 {
138         WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
139 }
140
141 static void free_ctx(struct rcu_head *head)
142 {
143         struct perf_event_context *ctx;
144
145         ctx = container_of(head, struct perf_event_context, rcu_head);
146         kfree(ctx);
147 }
148
149 static void put_ctx(struct perf_event_context *ctx)
150 {
151         if (atomic_dec_and_test(&ctx->refcount)) {
152                 if (ctx->parent_ctx)
153                         put_ctx(ctx->parent_ctx);
154                 if (ctx->task)
155                         put_task_struct(ctx->task);
156                 call_rcu(&ctx->rcu_head, free_ctx);
157         }
158 }
159
160 static void unclone_ctx(struct perf_event_context *ctx)
161 {
162         if (ctx->parent_ctx) {
163                 put_ctx(ctx->parent_ctx);
164                 ctx->parent_ctx = NULL;
165         }
166 }
167
168 /*
169  * If we inherit events we want to return the parent event id
170  * to userspace.
171  */
172 static u64 primary_event_id(struct perf_event *event)
173 {
174         u64 id = event->id;
175
176         if (event->parent)
177                 id = event->parent->id;
178
179         return id;
180 }
181
182 /*
183  * Get the perf_event_context for a task and lock it.
184  * This has to cope with with the fact that until it is locked,
185  * the context could get moved to another task.
186  */
187 static struct perf_event_context *
188 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
189 {
190         struct perf_event_context *ctx;
191
192         rcu_read_lock();
193  retry:
194         ctx = rcu_dereference(task->perf_event_ctxp);
195         if (ctx) {
196                 /*
197                  * If this context is a clone of another, it might
198                  * get swapped for another underneath us by
199                  * perf_event_task_sched_out, though the
200                  * rcu_read_lock() protects us from any context
201                  * getting freed.  Lock the context and check if it
202                  * got swapped before we could get the lock, and retry
203                  * if so.  If we locked the right context, then it
204                  * can't get swapped on us any more.
205                  */
206                 spin_lock_irqsave(&ctx->lock, *flags);
207                 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
208                         spin_unlock_irqrestore(&ctx->lock, *flags);
209                         goto retry;
210                 }
211
212                 if (!atomic_inc_not_zero(&ctx->refcount)) {
213                         spin_unlock_irqrestore(&ctx->lock, *flags);
214                         ctx = NULL;
215                 }
216         }
217         rcu_read_unlock();
218         return ctx;
219 }
220
221 /*
222  * Get the context for a task and increment its pin_count so it
223  * can't get swapped to another task.  This also increments its
224  * reference count so that the context can't get freed.
225  */
226 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
227 {
228         struct perf_event_context *ctx;
229         unsigned long flags;
230
231         ctx = perf_lock_task_context(task, &flags);
232         if (ctx) {
233                 ++ctx->pin_count;
234                 spin_unlock_irqrestore(&ctx->lock, flags);
235         }
236         return ctx;
237 }
238
239 static void perf_unpin_context(struct perf_event_context *ctx)
240 {
241         unsigned long flags;
242
243         spin_lock_irqsave(&ctx->lock, flags);
244         --ctx->pin_count;
245         spin_unlock_irqrestore(&ctx->lock, flags);
246         put_ctx(ctx);
247 }
248
249 static inline u64 perf_clock(void)
250 {
251         return cpu_clock(smp_processor_id());
252 }
253
254 /*
255  * Update the record of the current time in a context.
256  */
257 static void update_context_time(struct perf_event_context *ctx)
258 {
259         u64 now = perf_clock();
260
261         ctx->time += now - ctx->timestamp;
262         ctx->timestamp = now;
263 }
264
265 /*
266  * Update the total_time_enabled and total_time_running fields for a event.
267  */
268 static void update_event_times(struct perf_event *event)
269 {
270         struct perf_event_context *ctx = event->ctx;
271         u64 run_end;
272
273         if (event->state < PERF_EVENT_STATE_INACTIVE ||
274             event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
275                 return;
276
277         if (ctx->is_active)
278                 run_end = ctx->time;
279         else
280                 run_end = event->tstamp_stopped;
281
282         event->total_time_enabled = run_end - event->tstamp_enabled;
283
284         if (event->state == PERF_EVENT_STATE_INACTIVE)
285                 run_end = event->tstamp_stopped;
286         else
287                 run_end = ctx->time;
288
289         event->total_time_running = run_end - event->tstamp_running;
290 }
291
292 /*
293  * Add a event from the lists for its context.
294  * Must be called with ctx->mutex and ctx->lock held.
295  */
296 static void
297 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
298 {
299         struct perf_event *group_leader = event->group_leader;
300
301         /*
302          * Depending on whether it is a standalone or sibling event,
303          * add it straight to the context's event list, or to the group
304          * leader's sibling list:
305          */
306         if (group_leader == event)
307                 list_add_tail(&event->group_entry, &ctx->group_list);
308         else {
309                 list_add_tail(&event->group_entry, &group_leader->sibling_list);
310                 group_leader->nr_siblings++;
311         }
312
313         list_add_rcu(&event->event_entry, &ctx->event_list);
314         ctx->nr_events++;
315         if (event->attr.inherit_stat)
316                 ctx->nr_stat++;
317 }
318
319 /*
320  * Remove a event from the lists for its context.
321  * Must be called with ctx->mutex and ctx->lock held.
322  */
323 static void
324 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
325 {
326         struct perf_event *sibling, *tmp;
327
328         if (list_empty(&event->group_entry))
329                 return;
330         ctx->nr_events--;
331         if (event->attr.inherit_stat)
332                 ctx->nr_stat--;
333
334         list_del_init(&event->group_entry);
335         list_del_rcu(&event->event_entry);
336
337         if (event->group_leader != event)
338                 event->group_leader->nr_siblings--;
339
340         update_event_times(event);
341
342         /*
343          * If event was in error state, then keep it
344          * that way, otherwise bogus counts will be
345          * returned on read(). The only way to get out
346          * of error state is by explicit re-enabling
347          * of the event
348          */
349         if (event->state > PERF_EVENT_STATE_OFF)
350                 event->state = PERF_EVENT_STATE_OFF;
351
352         /*
353          * If this was a group event with sibling events then
354          * upgrade the siblings to singleton events by adding them
355          * to the context list directly:
356          */
357         list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
358
359                 list_move_tail(&sibling->group_entry, &ctx->group_list);
360                 sibling->group_leader = sibling;
361         }
362 }
363
364 static void
365 event_sched_out(struct perf_event *event,
366                   struct perf_cpu_context *cpuctx,
367                   struct perf_event_context *ctx)
368 {
369         if (event->state != PERF_EVENT_STATE_ACTIVE)
370                 return;
371
372         event->state = PERF_EVENT_STATE_INACTIVE;
373         if (event->pending_disable) {
374                 event->pending_disable = 0;
375                 event->state = PERF_EVENT_STATE_OFF;
376         }
377         event->tstamp_stopped = ctx->time;
378         event->pmu->disable(event);
379         event->oncpu = -1;
380
381         if (!is_software_event(event))
382                 cpuctx->active_oncpu--;
383         ctx->nr_active--;
384         if (event->attr.exclusive || !cpuctx->active_oncpu)
385                 cpuctx->exclusive = 0;
386 }
387
388 static void
389 group_sched_out(struct perf_event *group_event,
390                 struct perf_cpu_context *cpuctx,
391                 struct perf_event_context *ctx)
392 {
393         struct perf_event *event;
394
395         if (group_event->state != PERF_EVENT_STATE_ACTIVE)
396                 return;
397
398         event_sched_out(group_event, cpuctx, ctx);
399
400         /*
401          * Schedule out siblings (if any):
402          */
403         list_for_each_entry(event, &group_event->sibling_list, group_entry)
404                 event_sched_out(event, cpuctx, ctx);
405
406         if (group_event->attr.exclusive)
407                 cpuctx->exclusive = 0;
408 }
409
410 /*
411  * Cross CPU call to remove a performance event
412  *
413  * We disable the event on the hardware level first. After that we
414  * remove it from the context list.
415  */
416 static void __perf_event_remove_from_context(void *info)
417 {
418         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
419         struct perf_event *event = info;
420         struct perf_event_context *ctx = event->ctx;
421
422         /*
423          * If this is a task context, we need to check whether it is
424          * the current task context of this cpu. If not it has been
425          * scheduled out before the smp call arrived.
426          */
427         if (ctx->task && cpuctx->task_ctx != ctx)
428                 return;
429
430         spin_lock(&ctx->lock);
431         /*
432          * Protect the list operation against NMI by disabling the
433          * events on a global level.
434          */
435         perf_disable();
436
437         event_sched_out(event, cpuctx, ctx);
438
439         list_del_event(event, ctx);
440
441         if (!ctx->task) {
442                 /*
443                  * Allow more per task events with respect to the
444                  * reservation:
445                  */
446                 cpuctx->max_pertask =
447                         min(perf_max_events - ctx->nr_events,
448                             perf_max_events - perf_reserved_percpu);
449         }
450
451         perf_enable();
452         spin_unlock(&ctx->lock);
453 }
454
455
456 /*
457  * Remove the event from a task's (or a CPU's) list of events.
458  *
459  * Must be called with ctx->mutex held.
460  *
461  * CPU events are removed with a smp call. For task events we only
462  * call when the task is on a CPU.
463  *
464  * If event->ctx is a cloned context, callers must make sure that
465  * every task struct that event->ctx->task could possibly point to
466  * remains valid.  This is OK when called from perf_release since
467  * that only calls us on the top-level context, which can't be a clone.
468  * When called from perf_event_exit_task, it's OK because the
469  * context has been detached from its task.
470  */
471 static void perf_event_remove_from_context(struct perf_event *event)
472 {
473         struct perf_event_context *ctx = event->ctx;
474         struct task_struct *task = ctx->task;
475
476         if (!task) {
477                 /*
478                  * Per cpu events are removed via an smp call and
479                  * the removal is always sucessful.
480                  */
481                 smp_call_function_single(event->cpu,
482                                          __perf_event_remove_from_context,
483                                          event, 1);
484                 return;
485         }
486
487 retry:
488         task_oncpu_function_call(task, __perf_event_remove_from_context,
489                                  event);
490
491         spin_lock_irq(&ctx->lock);
492         /*
493          * If the context is active we need to retry the smp call.
494          */
495         if (ctx->nr_active && !list_empty(&event->group_entry)) {
496                 spin_unlock_irq(&ctx->lock);
497                 goto retry;
498         }
499
500         /*
501          * The lock prevents that this context is scheduled in so we
502          * can remove the event safely, if the call above did not
503          * succeed.
504          */
505         if (!list_empty(&event->group_entry))
506                 list_del_event(event, ctx);
507         spin_unlock_irq(&ctx->lock);
508 }
509
510 /*
511  * Update total_time_enabled and total_time_running for all events in a group.
512  */
513 static void update_group_times(struct perf_event *leader)
514 {
515         struct perf_event *event;
516
517         update_event_times(leader);
518         list_for_each_entry(event, &leader->sibling_list, group_entry)
519                 update_event_times(event);
520 }
521
522 /*
523  * Cross CPU call to disable a performance event
524  */
525 static void __perf_event_disable(void *info)
526 {
527         struct perf_event *event = info;
528         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
529         struct perf_event_context *ctx = event->ctx;
530
531         /*
532          * If this is a per-task event, need to check whether this
533          * event's task is the current task on this cpu.
534          */
535         if (ctx->task && cpuctx->task_ctx != ctx)
536                 return;
537
538         spin_lock(&ctx->lock);
539
540         /*
541          * If the event is on, turn it off.
542          * If it is in error state, leave it in error state.
543          */
544         if (event->state >= PERF_EVENT_STATE_INACTIVE) {
545                 update_context_time(ctx);
546                 update_group_times(event);
547                 if (event == event->group_leader)
548                         group_sched_out(event, cpuctx, ctx);
549                 else
550                         event_sched_out(event, cpuctx, ctx);
551                 event->state = PERF_EVENT_STATE_OFF;
552         }
553
554         spin_unlock(&ctx->lock);
555 }
556
557 /*
558  * Disable a event.
559  *
560  * If event->ctx is a cloned context, callers must make sure that
561  * every task struct that event->ctx->task could possibly point to
562  * remains valid.  This condition is satisifed when called through
563  * perf_event_for_each_child or perf_event_for_each because they
564  * hold the top-level event's child_mutex, so any descendant that
565  * goes to exit will block in sync_child_event.
566  * When called from perf_pending_event it's OK because event->ctx
567  * is the current context on this CPU and preemption is disabled,
568  * hence we can't get into perf_event_task_sched_out for this context.
569  */
570 void perf_event_disable(struct perf_event *event)
571 {
572         struct perf_event_context *ctx = event->ctx;
573         struct task_struct *task = ctx->task;
574
575         if (!task) {
576                 /*
577                  * Disable the event on the cpu that it's on
578                  */
579                 smp_call_function_single(event->cpu, __perf_event_disable,
580                                          event, 1);
581                 return;
582         }
583
584  retry:
585         task_oncpu_function_call(task, __perf_event_disable, event);
586
587         spin_lock_irq(&ctx->lock);
588         /*
589          * If the event is still active, we need to retry the cross-call.
590          */
591         if (event->state == PERF_EVENT_STATE_ACTIVE) {
592                 spin_unlock_irq(&ctx->lock);
593                 goto retry;
594         }
595
596         /*
597          * Since we have the lock this context can't be scheduled
598          * in, so we can change the state safely.
599          */
600         if (event->state == PERF_EVENT_STATE_INACTIVE) {
601                 update_group_times(event);
602                 event->state = PERF_EVENT_STATE_OFF;
603         }
604
605         spin_unlock_irq(&ctx->lock);
606 }
607
608 static int
609 event_sched_in(struct perf_event *event,
610                  struct perf_cpu_context *cpuctx,
611                  struct perf_event_context *ctx,
612                  int cpu)
613 {
614         if (event->state <= PERF_EVENT_STATE_OFF)
615                 return 0;
616
617         event->state = PERF_EVENT_STATE_ACTIVE;
618         event->oncpu = cpu;     /* TODO: put 'cpu' into cpuctx->cpu */
619         /*
620          * The new state must be visible before we turn it on in the hardware:
621          */
622         smp_wmb();
623
624         if (event->pmu->enable(event)) {
625                 event->state = PERF_EVENT_STATE_INACTIVE;
626                 event->oncpu = -1;
627                 return -EAGAIN;
628         }
629
630         event->tstamp_running += ctx->time - event->tstamp_stopped;
631
632         if (!is_software_event(event))
633                 cpuctx->active_oncpu++;
634         ctx->nr_active++;
635
636         if (event->attr.exclusive)
637                 cpuctx->exclusive = 1;
638
639         return 0;
640 }
641
642 static int
643 group_sched_in(struct perf_event *group_event,
644                struct perf_cpu_context *cpuctx,
645                struct perf_event_context *ctx,
646                int cpu)
647 {
648         struct perf_event *event, *partial_group;
649         int ret;
650
651         if (group_event->state == PERF_EVENT_STATE_OFF)
652                 return 0;
653
654         ret = hw_perf_group_sched_in(group_event, cpuctx, ctx, cpu);
655         if (ret)
656                 return ret < 0 ? ret : 0;
657
658         if (event_sched_in(group_event, cpuctx, ctx, cpu))
659                 return -EAGAIN;
660
661         /*
662          * Schedule in siblings as one group (if any):
663          */
664         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
665                 if (event_sched_in(event, cpuctx, ctx, cpu)) {
666                         partial_group = event;
667                         goto group_error;
668                 }
669         }
670
671         return 0;
672
673 group_error:
674         /*
675          * Groups can be scheduled in as one unit only, so undo any
676          * partial group before returning:
677          */
678         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
679                 if (event == partial_group)
680                         break;
681                 event_sched_out(event, cpuctx, ctx);
682         }
683         event_sched_out(group_event, cpuctx, ctx);
684
685         return -EAGAIN;
686 }
687
688 /*
689  * Return 1 for a group consisting entirely of software events,
690  * 0 if the group contains any hardware events.
691  */
692 static int is_software_only_group(struct perf_event *leader)
693 {
694         struct perf_event *event;
695
696         if (!is_software_event(leader))
697                 return 0;
698
699         list_for_each_entry(event, &leader->sibling_list, group_entry)
700                 if (!is_software_event(event))
701                         return 0;
702
703         return 1;
704 }
705
706 /*
707  * Work out whether we can put this event group on the CPU now.
708  */
709 static int group_can_go_on(struct perf_event *event,
710                            struct perf_cpu_context *cpuctx,
711                            int can_add_hw)
712 {
713         /*
714          * Groups consisting entirely of software events can always go on.
715          */
716         if (is_software_only_group(event))
717                 return 1;
718         /*
719          * If an exclusive group is already on, no other hardware
720          * events can go on.
721          */
722         if (cpuctx->exclusive)
723                 return 0;
724         /*
725          * If this group is exclusive and there are already
726          * events on the CPU, it can't go on.
727          */
728         if (event->attr.exclusive && cpuctx->active_oncpu)
729                 return 0;
730         /*
731          * Otherwise, try to add it if all previous groups were able
732          * to go on.
733          */
734         return can_add_hw;
735 }
736
737 static void add_event_to_ctx(struct perf_event *event,
738                                struct perf_event_context *ctx)
739 {
740         list_add_event(event, ctx);
741         event->tstamp_enabled = ctx->time;
742         event->tstamp_running = ctx->time;
743         event->tstamp_stopped = ctx->time;
744 }
745
746 /*
747  * Cross CPU call to install and enable a performance event
748  *
749  * Must be called with ctx->mutex held
750  */
751 static void __perf_install_in_context(void *info)
752 {
753         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
754         struct perf_event *event = info;
755         struct perf_event_context *ctx = event->ctx;
756         struct perf_event *leader = event->group_leader;
757         int cpu = smp_processor_id();
758         int err;
759
760         /*
761          * If this is a task context, we need to check whether it is
762          * the current task context of this cpu. If not it has been
763          * scheduled out before the smp call arrived.
764          * Or possibly this is the right context but it isn't
765          * on this cpu because it had no events.
766          */
767         if (ctx->task && cpuctx->task_ctx != ctx) {
768                 if (cpuctx->task_ctx || ctx->task != current)
769                         return;
770                 cpuctx->task_ctx = ctx;
771         }
772
773         spin_lock(&ctx->lock);
774         ctx->is_active = 1;
775         update_context_time(ctx);
776
777         /*
778          * Protect the list operation against NMI by disabling the
779          * events on a global level. NOP for non NMI based events.
780          */
781         perf_disable();
782
783         add_event_to_ctx(event, ctx);
784
785         if (event->cpu != -1 && event->cpu != smp_processor_id())
786                 goto unlock;
787
788         /*
789          * Don't put the event on if it is disabled or if
790          * it is in a group and the group isn't on.
791          */
792         if (event->state != PERF_EVENT_STATE_INACTIVE ||
793             (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
794                 goto unlock;
795
796         /*
797          * An exclusive event can't go on if there are already active
798          * hardware events, and no hardware event can go on if there
799          * is already an exclusive event on.
800          */
801         if (!group_can_go_on(event, cpuctx, 1))
802                 err = -EEXIST;
803         else
804                 err = event_sched_in(event, cpuctx, ctx, cpu);
805
806         if (err) {
807                 /*
808                  * This event couldn't go on.  If it is in a group
809                  * then we have to pull the whole group off.
810                  * If the event group is pinned then put it in error state.
811                  */
812                 if (leader != event)
813                         group_sched_out(leader, cpuctx, ctx);
814                 if (leader->attr.pinned) {
815                         update_group_times(leader);
816                         leader->state = PERF_EVENT_STATE_ERROR;
817                 }
818         }
819
820         if (!err && !ctx->task && cpuctx->max_pertask)
821                 cpuctx->max_pertask--;
822
823  unlock:
824         perf_enable();
825
826         spin_unlock(&ctx->lock);
827 }
828
829 /*
830  * Attach a performance event to a context
831  *
832  * First we add the event to the list with the hardware enable bit
833  * in event->hw_config cleared.
834  *
835  * If the event is attached to a task which is on a CPU we use a smp
836  * call to enable it in the task context. The task might have been
837  * scheduled away, but we check this in the smp call again.
838  *
839  * Must be called with ctx->mutex held.
840  */
841 static void
842 perf_install_in_context(struct perf_event_context *ctx,
843                         struct perf_event *event,
844                         int cpu)
845 {
846         struct task_struct *task = ctx->task;
847
848         if (!task) {
849                 /*
850                  * Per cpu events are installed via an smp call and
851                  * the install is always sucessful.
852                  */
853                 smp_call_function_single(cpu, __perf_install_in_context,
854                                          event, 1);
855                 return;
856         }
857
858 retry:
859         task_oncpu_function_call(task, __perf_install_in_context,
860                                  event);
861
862         spin_lock_irq(&ctx->lock);
863         /*
864          * we need to retry the smp call.
865          */
866         if (ctx->is_active && list_empty(&event->group_entry)) {
867                 spin_unlock_irq(&ctx->lock);
868                 goto retry;
869         }
870
871         /*
872          * The lock prevents that this context is scheduled in so we
873          * can add the event safely, if it the call above did not
874          * succeed.
875          */
876         if (list_empty(&event->group_entry))
877                 add_event_to_ctx(event, ctx);
878         spin_unlock_irq(&ctx->lock);
879 }
880
881 /*
882  * Put a event into inactive state and update time fields.
883  * Enabling the leader of a group effectively enables all
884  * the group members that aren't explicitly disabled, so we
885  * have to update their ->tstamp_enabled also.
886  * Note: this works for group members as well as group leaders
887  * since the non-leader members' sibling_lists will be empty.
888  */
889 static void __perf_event_mark_enabled(struct perf_event *event,
890                                         struct perf_event_context *ctx)
891 {
892         struct perf_event *sub;
893
894         event->state = PERF_EVENT_STATE_INACTIVE;
895         event->tstamp_enabled = ctx->time - event->total_time_enabled;
896         list_for_each_entry(sub, &event->sibling_list, group_entry)
897                 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
898                         sub->tstamp_enabled =
899                                 ctx->time - sub->total_time_enabled;
900 }
901
902 /*
903  * Cross CPU call to enable a performance event
904  */
905 static void __perf_event_enable(void *info)
906 {
907         struct perf_event *event = info;
908         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
909         struct perf_event_context *ctx = event->ctx;
910         struct perf_event *leader = event->group_leader;
911         int err;
912
913         /*
914          * If this is a per-task event, need to check whether this
915          * event's task is the current task on this cpu.
916          */
917         if (ctx->task && cpuctx->task_ctx != ctx) {
918                 if (cpuctx->task_ctx || ctx->task != current)
919                         return;
920                 cpuctx->task_ctx = ctx;
921         }
922
923         spin_lock(&ctx->lock);
924         ctx->is_active = 1;
925         update_context_time(ctx);
926
927         if (event->state >= PERF_EVENT_STATE_INACTIVE)
928                 goto unlock;
929         __perf_event_mark_enabled(event, ctx);
930
931         if (event->cpu != -1 && event->cpu != smp_processor_id())
932                 goto unlock;
933
934         /*
935          * If the event is in a group and isn't the group leader,
936          * then don't put it on unless the group is on.
937          */
938         if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
939                 goto unlock;
940
941         if (!group_can_go_on(event, cpuctx, 1)) {
942                 err = -EEXIST;
943         } else {
944                 perf_disable();
945                 if (event == leader)
946                         err = group_sched_in(event, cpuctx, ctx,
947                                              smp_processor_id());
948                 else
949                         err = event_sched_in(event, cpuctx, ctx,
950                                                smp_processor_id());
951                 perf_enable();
952         }
953
954         if (err) {
955                 /*
956                  * If this event can't go on and it's part of a
957                  * group, then the whole group has to come off.
958                  */
959                 if (leader != event)
960                         group_sched_out(leader, cpuctx, ctx);
961                 if (leader->attr.pinned) {
962                         update_group_times(leader);
963                         leader->state = PERF_EVENT_STATE_ERROR;
964                 }
965         }
966
967  unlock:
968         spin_unlock(&ctx->lock);
969 }
970
971 /*
972  * Enable a event.
973  *
974  * If event->ctx is a cloned context, callers must make sure that
975  * every task struct that event->ctx->task could possibly point to
976  * remains valid.  This condition is satisfied when called through
977  * perf_event_for_each_child or perf_event_for_each as described
978  * for perf_event_disable.
979  */
980 void perf_event_enable(struct perf_event *event)
981 {
982         struct perf_event_context *ctx = event->ctx;
983         struct task_struct *task = ctx->task;
984
985         if (!task) {
986                 /*
987                  * Enable the event on the cpu that it's on
988                  */
989                 smp_call_function_single(event->cpu, __perf_event_enable,
990                                          event, 1);
991                 return;
992         }
993
994         spin_lock_irq(&ctx->lock);
995         if (event->state >= PERF_EVENT_STATE_INACTIVE)
996                 goto out;
997
998         /*
999          * If the event is in error state, clear that first.
1000          * That way, if we see the event in error state below, we
1001          * know that it has gone back into error state, as distinct
1002          * from the task having been scheduled away before the
1003          * cross-call arrived.
1004          */
1005         if (event->state == PERF_EVENT_STATE_ERROR)
1006                 event->state = PERF_EVENT_STATE_OFF;
1007
1008  retry:
1009         spin_unlock_irq(&ctx->lock);
1010         task_oncpu_function_call(task, __perf_event_enable, event);
1011
1012         spin_lock_irq(&ctx->lock);
1013
1014         /*
1015          * If the context is active and the event is still off,
1016          * we need to retry the cross-call.
1017          */
1018         if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1019                 goto retry;
1020
1021         /*
1022          * Since we have the lock this context can't be scheduled
1023          * in, so we can change the state safely.
1024          */
1025         if (event->state == PERF_EVENT_STATE_OFF)
1026                 __perf_event_mark_enabled(event, ctx);
1027
1028  out:
1029         spin_unlock_irq(&ctx->lock);
1030 }
1031
1032 static int perf_event_refresh(struct perf_event *event, int refresh)
1033 {
1034         /*
1035          * not supported on inherited events
1036          */
1037         if (event->attr.inherit)
1038                 return -EINVAL;
1039
1040         atomic_add(refresh, &event->event_limit);
1041         perf_event_enable(event);
1042
1043         return 0;
1044 }
1045
1046 void __perf_event_sched_out(struct perf_event_context *ctx,
1047                               struct perf_cpu_context *cpuctx)
1048 {
1049         struct perf_event *event;
1050
1051         spin_lock(&ctx->lock);
1052         ctx->is_active = 0;
1053         if (likely(!ctx->nr_events))
1054                 goto out;
1055         update_context_time(ctx);
1056
1057         perf_disable();
1058         if (ctx->nr_active) {
1059                 list_for_each_entry(event, &ctx->group_list, group_entry)
1060                         group_sched_out(event, cpuctx, ctx);
1061         }
1062         perf_enable();
1063  out:
1064         spin_unlock(&ctx->lock);
1065 }
1066
1067 /*
1068  * Test whether two contexts are equivalent, i.e. whether they
1069  * have both been cloned from the same version of the same context
1070  * and they both have the same number of enabled events.
1071  * If the number of enabled events is the same, then the set
1072  * of enabled events should be the same, because these are both
1073  * inherited contexts, therefore we can't access individual events
1074  * in them directly with an fd; we can only enable/disable all
1075  * events via prctl, or enable/disable all events in a family
1076  * via ioctl, which will have the same effect on both contexts.
1077  */
1078 static int context_equiv(struct perf_event_context *ctx1,
1079                          struct perf_event_context *ctx2)
1080 {
1081         return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1082                 && ctx1->parent_gen == ctx2->parent_gen
1083                 && !ctx1->pin_count && !ctx2->pin_count;
1084 }
1085
1086 static void __perf_event_sync_stat(struct perf_event *event,
1087                                      struct perf_event *next_event)
1088 {
1089         u64 value;
1090
1091         if (!event->attr.inherit_stat)
1092                 return;
1093
1094         /*
1095          * Update the event value, we cannot use perf_event_read()
1096          * because we're in the middle of a context switch and have IRQs
1097          * disabled, which upsets smp_call_function_single(), however
1098          * we know the event must be on the current CPU, therefore we
1099          * don't need to use it.
1100          */
1101         switch (event->state) {
1102         case PERF_EVENT_STATE_ACTIVE:
1103                 event->pmu->read(event);
1104                 /* fall-through */
1105
1106         case PERF_EVENT_STATE_INACTIVE:
1107                 update_event_times(event);
1108                 break;
1109
1110         default:
1111                 break;
1112         }
1113
1114         /*
1115          * In order to keep per-task stats reliable we need to flip the event
1116          * values when we flip the contexts.
1117          */
1118         value = atomic64_read(&next_event->count);
1119         value = atomic64_xchg(&event->count, value);
1120         atomic64_set(&next_event->count, value);
1121
1122         swap(event->total_time_enabled, next_event->total_time_enabled);
1123         swap(event->total_time_running, next_event->total_time_running);
1124
1125         /*
1126          * Since we swizzled the values, update the user visible data too.
1127          */
1128         perf_event_update_userpage(event);
1129         perf_event_update_userpage(next_event);
1130 }
1131
1132 #define list_next_entry(pos, member) \
1133         list_entry(pos->member.next, typeof(*pos), member)
1134
1135 static void perf_event_sync_stat(struct perf_event_context *ctx,
1136                                    struct perf_event_context *next_ctx)
1137 {
1138         struct perf_event *event, *next_event;
1139
1140         if (!ctx->nr_stat)
1141                 return;
1142
1143         update_context_time(ctx);
1144
1145         event = list_first_entry(&ctx->event_list,
1146                                    struct perf_event, event_entry);
1147
1148         next_event = list_first_entry(&next_ctx->event_list,
1149                                         struct perf_event, event_entry);
1150
1151         while (&event->event_entry != &ctx->event_list &&
1152                &next_event->event_entry != &next_ctx->event_list) {
1153
1154                 __perf_event_sync_stat(event, next_event);
1155
1156                 event = list_next_entry(event, event_entry);
1157                 next_event = list_next_entry(next_event, event_entry);
1158         }
1159 }
1160
1161 /*
1162  * Called from scheduler to remove the events of the current task,
1163  * with interrupts disabled.
1164  *
1165  * We stop each event and update the event value in event->count.
1166  *
1167  * This does not protect us against NMI, but disable()
1168  * sets the disabled bit in the control field of event _before_
1169  * accessing the event control register. If a NMI hits, then it will
1170  * not restart the event.
1171  */
1172 void perf_event_task_sched_out(struct task_struct *task,
1173                                  struct task_struct *next, int cpu)
1174 {
1175         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1176         struct perf_event_context *ctx = task->perf_event_ctxp;
1177         struct perf_event_context *next_ctx;
1178         struct perf_event_context *parent;
1179         struct pt_regs *regs;
1180         int do_switch = 1;
1181
1182         regs = task_pt_regs(task);
1183         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1184
1185         if (likely(!ctx || !cpuctx->task_ctx))
1186                 return;
1187
1188         rcu_read_lock();
1189         parent = rcu_dereference(ctx->parent_ctx);
1190         next_ctx = next->perf_event_ctxp;
1191         if (parent && next_ctx &&
1192             rcu_dereference(next_ctx->parent_ctx) == parent) {
1193                 /*
1194                  * Looks like the two contexts are clones, so we might be
1195                  * able to optimize the context switch.  We lock both
1196                  * contexts and check that they are clones under the
1197                  * lock (including re-checking that neither has been
1198                  * uncloned in the meantime).  It doesn't matter which
1199                  * order we take the locks because no other cpu could
1200                  * be trying to lock both of these tasks.
1201                  */
1202                 spin_lock(&ctx->lock);
1203                 spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1204                 if (context_equiv(ctx, next_ctx)) {
1205                         /*
1206                          * XXX do we need a memory barrier of sorts
1207                          * wrt to rcu_dereference() of perf_event_ctxp
1208                          */
1209                         task->perf_event_ctxp = next_ctx;
1210                         next->perf_event_ctxp = ctx;
1211                         ctx->task = next;
1212                         next_ctx->task = task;
1213                         do_switch = 0;
1214
1215                         perf_event_sync_stat(ctx, next_ctx);
1216                 }
1217                 spin_unlock(&next_ctx->lock);
1218                 spin_unlock(&ctx->lock);
1219         }
1220         rcu_read_unlock();
1221
1222         if (do_switch) {
1223                 __perf_event_sched_out(ctx, cpuctx);
1224                 cpuctx->task_ctx = NULL;
1225         }
1226 }
1227
1228 /*
1229  * Called with IRQs disabled
1230  */
1231 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1232 {
1233         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1234
1235         if (!cpuctx->task_ctx)
1236                 return;
1237
1238         if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1239                 return;
1240
1241         __perf_event_sched_out(ctx, cpuctx);
1242         cpuctx->task_ctx = NULL;
1243 }
1244
1245 /*
1246  * Called with IRQs disabled
1247  */
1248 static void perf_event_cpu_sched_out(struct perf_cpu_context *cpuctx)
1249 {
1250         __perf_event_sched_out(&cpuctx->ctx, cpuctx);
1251 }
1252
1253 static void
1254 __perf_event_sched_in(struct perf_event_context *ctx,
1255                         struct perf_cpu_context *cpuctx, int cpu)
1256 {
1257         struct perf_event *event;
1258         int can_add_hw = 1;
1259
1260         spin_lock(&ctx->lock);
1261         ctx->is_active = 1;
1262         if (likely(!ctx->nr_events))
1263                 goto out;
1264
1265         ctx->timestamp = perf_clock();
1266
1267         perf_disable();
1268
1269         /*
1270          * First go through the list and put on any pinned groups
1271          * in order to give them the best chance of going on.
1272          */
1273         list_for_each_entry(event, &ctx->group_list, group_entry) {
1274                 if (event->state <= PERF_EVENT_STATE_OFF ||
1275                     !event->attr.pinned)
1276                         continue;
1277                 if (event->cpu != -1 && event->cpu != cpu)
1278                         continue;
1279
1280                 if (group_can_go_on(event, cpuctx, 1))
1281                         group_sched_in(event, cpuctx, ctx, cpu);
1282
1283                 /*
1284                  * If this pinned group hasn't been scheduled,
1285                  * put it in error state.
1286                  */
1287                 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1288                         update_group_times(event);
1289                         event->state = PERF_EVENT_STATE_ERROR;
1290                 }
1291         }
1292
1293         list_for_each_entry(event, &ctx->group_list, group_entry) {
1294                 /*
1295                  * Ignore events in OFF or ERROR state, and
1296                  * ignore pinned events since we did them already.
1297                  */
1298                 if (event->state <= PERF_EVENT_STATE_OFF ||
1299                     event->attr.pinned)
1300                         continue;
1301
1302                 /*
1303                  * Listen to the 'cpu' scheduling filter constraint
1304                  * of events:
1305                  */
1306                 if (event->cpu != -1 && event->cpu != cpu)
1307                         continue;
1308
1309                 if (group_can_go_on(event, cpuctx, can_add_hw))
1310                         if (group_sched_in(event, cpuctx, ctx, cpu))
1311                                 can_add_hw = 0;
1312         }
1313         perf_enable();
1314  out:
1315         spin_unlock(&ctx->lock);
1316 }
1317
1318 /*
1319  * Called from scheduler to add the events of the current task
1320  * with interrupts disabled.
1321  *
1322  * We restore the event value and then enable it.
1323  *
1324  * This does not protect us against NMI, but enable()
1325  * sets the enabled bit in the control field of event _before_
1326  * accessing the event control register. If a NMI hits, then it will
1327  * keep the event running.
1328  */
1329 void perf_event_task_sched_in(struct task_struct *task, int cpu)
1330 {
1331         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1332         struct perf_event_context *ctx = task->perf_event_ctxp;
1333
1334         if (likely(!ctx))
1335                 return;
1336         if (cpuctx->task_ctx == ctx)
1337                 return;
1338         __perf_event_sched_in(ctx, cpuctx, cpu);
1339         cpuctx->task_ctx = ctx;
1340 }
1341
1342 static void perf_event_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
1343 {
1344         struct perf_event_context *ctx = &cpuctx->ctx;
1345
1346         __perf_event_sched_in(ctx, cpuctx, cpu);
1347 }
1348
1349 #define MAX_INTERRUPTS (~0ULL)
1350
1351 static void perf_log_throttle(struct perf_event *event, int enable);
1352
1353 static void perf_adjust_period(struct perf_event *event, u64 events)
1354 {
1355         struct hw_perf_event *hwc = &event->hw;
1356         u64 period, sample_period;
1357         s64 delta;
1358
1359         events *= hwc->sample_period;
1360         period = div64_u64(events, event->attr.sample_freq);
1361
1362         delta = (s64)(period - hwc->sample_period);
1363         delta = (delta + 7) / 8; /* low pass filter */
1364
1365         sample_period = hwc->sample_period + delta;
1366
1367         if (!sample_period)
1368                 sample_period = 1;
1369
1370         hwc->sample_period = sample_period;
1371 }
1372
1373 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1374 {
1375         struct perf_event *event;
1376         struct hw_perf_event *hwc;
1377         u64 interrupts, freq;
1378
1379         spin_lock(&ctx->lock);
1380         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1381                 if (event->state != PERF_EVENT_STATE_ACTIVE)
1382                         continue;
1383
1384                 hwc = &event->hw;
1385
1386                 interrupts = hwc->interrupts;
1387                 hwc->interrupts = 0;
1388
1389                 /*
1390                  * unthrottle events on the tick
1391                  */
1392                 if (interrupts == MAX_INTERRUPTS) {
1393                         perf_log_throttle(event, 1);
1394                         event->pmu->unthrottle(event);
1395                         interrupts = 2*sysctl_perf_event_sample_rate/HZ;
1396                 }
1397
1398                 if (!event->attr.freq || !event->attr.sample_freq)
1399                         continue;
1400
1401                 /*
1402                  * if the specified freq < HZ then we need to skip ticks
1403                  */
1404                 if (event->attr.sample_freq < HZ) {
1405                         freq = event->attr.sample_freq;
1406
1407                         hwc->freq_count += freq;
1408                         hwc->freq_interrupts += interrupts;
1409
1410                         if (hwc->freq_count < HZ)
1411                                 continue;
1412
1413                         interrupts = hwc->freq_interrupts;
1414                         hwc->freq_interrupts = 0;
1415                         hwc->freq_count -= HZ;
1416                 } else
1417                         freq = HZ;
1418
1419                 perf_adjust_period(event, freq * interrupts);
1420
1421                 /*
1422                  * In order to avoid being stalled by an (accidental) huge
1423                  * sample period, force reset the sample period if we didn't
1424                  * get any events in this freq period.
1425                  */
1426                 if (!interrupts) {
1427                         perf_disable();
1428                         event->pmu->disable(event);
1429                         atomic64_set(&hwc->period_left, 0);
1430                         event->pmu->enable(event);
1431                         perf_enable();
1432                 }
1433         }
1434         spin_unlock(&ctx->lock);
1435 }
1436
1437 /*
1438  * Round-robin a context's events:
1439  */
1440 static void rotate_ctx(struct perf_event_context *ctx)
1441 {
1442         struct perf_event *event;
1443
1444         if (!ctx->nr_events)
1445                 return;
1446
1447         spin_lock(&ctx->lock);
1448         /*
1449          * Rotate the first entry last (works just fine for group events too):
1450          */
1451         perf_disable();
1452         list_for_each_entry(event, &ctx->group_list, group_entry) {
1453                 list_move_tail(&event->group_entry, &ctx->group_list);
1454                 break;
1455         }
1456         perf_enable();
1457
1458         spin_unlock(&ctx->lock);
1459 }
1460
1461 void perf_event_task_tick(struct task_struct *curr, int cpu)
1462 {
1463         struct perf_cpu_context *cpuctx;
1464         struct perf_event_context *ctx;
1465
1466         if (!atomic_read(&nr_events))
1467                 return;
1468
1469         cpuctx = &per_cpu(perf_cpu_context, cpu);
1470         ctx = curr->perf_event_ctxp;
1471
1472         perf_ctx_adjust_freq(&cpuctx->ctx);
1473         if (ctx)
1474                 perf_ctx_adjust_freq(ctx);
1475
1476         perf_event_cpu_sched_out(cpuctx);
1477         if (ctx)
1478                 __perf_event_task_sched_out(ctx);
1479
1480         rotate_ctx(&cpuctx->ctx);
1481         if (ctx)
1482                 rotate_ctx(ctx);
1483
1484         perf_event_cpu_sched_in(cpuctx, cpu);
1485         if (ctx)
1486                 perf_event_task_sched_in(curr, cpu);
1487 }
1488
1489 /*
1490  * Enable all of a task's events that have been marked enable-on-exec.
1491  * This expects task == current.
1492  */
1493 static void perf_event_enable_on_exec(struct task_struct *task)
1494 {
1495         struct perf_event_context *ctx;
1496         struct perf_event *event;
1497         unsigned long flags;
1498         int enabled = 0;
1499
1500         local_irq_save(flags);
1501         ctx = task->perf_event_ctxp;
1502         if (!ctx || !ctx->nr_events)
1503                 goto out;
1504
1505         __perf_event_task_sched_out(ctx);
1506
1507         spin_lock(&ctx->lock);
1508
1509         list_for_each_entry(event, &ctx->group_list, group_entry) {
1510                 if (!event->attr.enable_on_exec)
1511                         continue;
1512                 event->attr.enable_on_exec = 0;
1513                 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1514                         continue;
1515                 __perf_event_mark_enabled(event, ctx);
1516                 enabled = 1;
1517         }
1518
1519         /*
1520          * Unclone this context if we enabled any event.
1521          */
1522         if (enabled)
1523                 unclone_ctx(ctx);
1524
1525         spin_unlock(&ctx->lock);
1526
1527         perf_event_task_sched_in(task, smp_processor_id());
1528  out:
1529         local_irq_restore(flags);
1530 }
1531
1532 /*
1533  * Cross CPU call to read the hardware event
1534  */
1535 static void __perf_event_read(void *info)
1536 {
1537         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1538         struct perf_event *event = info;
1539         struct perf_event_context *ctx = event->ctx;
1540
1541         /*
1542          * If this is a task context, we need to check whether it is
1543          * the current task context of this cpu.  If not it has been
1544          * scheduled out before the smp call arrived.  In that case
1545          * event->count would have been updated to a recent sample
1546          * when the event was scheduled out.
1547          */
1548         if (ctx->task && cpuctx->task_ctx != ctx)
1549                 return;
1550
1551         spin_lock(&ctx->lock);
1552         update_context_time(ctx);
1553         update_event_times(event);
1554         spin_unlock(&ctx->lock);
1555
1556         event->pmu->read(event);
1557 }
1558
1559 static u64 perf_event_read(struct perf_event *event)
1560 {
1561         /*
1562          * If event is enabled and currently active on a CPU, update the
1563          * value in the event structure:
1564          */
1565         if (event->state == PERF_EVENT_STATE_ACTIVE) {
1566                 smp_call_function_single(event->oncpu,
1567                                          __perf_event_read, event, 1);
1568         } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1569                 struct perf_event_context *ctx = event->ctx;
1570                 unsigned long flags;
1571
1572                 spin_lock_irqsave(&ctx->lock, flags);
1573                 update_context_time(ctx);
1574                 update_event_times(event);
1575                 spin_unlock_irqrestore(&ctx->lock, flags);
1576         }
1577
1578         return atomic64_read(&event->count);
1579 }
1580
1581 /*
1582  * Initialize the perf_event context in a task_struct:
1583  */
1584 static void
1585 __perf_event_init_context(struct perf_event_context *ctx,
1586                             struct task_struct *task)
1587 {
1588         spin_lock_init(&ctx->lock);
1589         mutex_init(&ctx->mutex);
1590         INIT_LIST_HEAD(&ctx->group_list);
1591         INIT_LIST_HEAD(&ctx->event_list);
1592         atomic_set(&ctx->refcount, 1);
1593         ctx->task = task;
1594 }
1595
1596 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1597 {
1598         struct perf_event_context *ctx;
1599         struct perf_cpu_context *cpuctx;
1600         struct task_struct *task;
1601         unsigned long flags;
1602         int err;
1603
1604         if (pid == -1 && cpu != -1) {
1605                 /* Must be root to operate on a CPU event: */
1606                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1607                         return ERR_PTR(-EACCES);
1608
1609                 if (cpu < 0 || cpu >= nr_cpumask_bits)
1610                         return ERR_PTR(-EINVAL);
1611
1612                 /*
1613                  * We could be clever and allow to attach a event to an
1614                  * offline CPU and activate it when the CPU comes up, but
1615                  * that's for later.
1616                  */
1617                 if (!cpu_isset(cpu, cpu_online_map))
1618                         return ERR_PTR(-ENODEV);
1619
1620                 cpuctx = &per_cpu(perf_cpu_context, cpu);
1621                 ctx = &cpuctx->ctx;
1622                 get_ctx(ctx);
1623
1624                 return ctx;
1625         }
1626
1627         rcu_read_lock();
1628         if (!pid)
1629                 task = current;
1630         else
1631                 task = find_task_by_vpid(pid);
1632         if (task)
1633                 get_task_struct(task);
1634         rcu_read_unlock();
1635
1636         if (!task)
1637                 return ERR_PTR(-ESRCH);
1638
1639         /*
1640          * Can't attach events to a dying task.
1641          */
1642         err = -ESRCH;
1643         if (task->flags & PF_EXITING)
1644                 goto errout;
1645
1646         /* Reuse ptrace permission checks for now. */
1647         err = -EACCES;
1648         if (!ptrace_may_access(task, PTRACE_MODE_READ))
1649                 goto errout;
1650
1651  retry:
1652         ctx = perf_lock_task_context(task, &flags);
1653         if (ctx) {
1654                 unclone_ctx(ctx);
1655                 spin_unlock_irqrestore(&ctx->lock, flags);
1656         }
1657
1658         if (!ctx) {
1659                 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1660                 err = -ENOMEM;
1661                 if (!ctx)
1662                         goto errout;
1663                 __perf_event_init_context(ctx, task);
1664                 get_ctx(ctx);
1665                 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1666                         /*
1667                          * We raced with some other task; use
1668                          * the context they set.
1669                          */
1670                         kfree(ctx);
1671                         goto retry;
1672                 }
1673                 get_task_struct(task);
1674         }
1675
1676         put_task_struct(task);
1677         return ctx;
1678
1679  errout:
1680         put_task_struct(task);
1681         return ERR_PTR(err);
1682 }
1683
1684 static void perf_event_free_filter(struct perf_event *event);
1685
1686 static void free_event_rcu(struct rcu_head *head)
1687 {
1688         struct perf_event *event;
1689
1690         event = container_of(head, struct perf_event, rcu_head);
1691         if (event->ns)
1692                 put_pid_ns(event->ns);
1693         perf_event_free_filter(event);
1694         kfree(event);
1695 }
1696
1697 static void perf_pending_sync(struct perf_event *event);
1698
1699 static void free_event(struct perf_event *event)
1700 {
1701         perf_pending_sync(event);
1702
1703         if (!event->parent) {
1704                 atomic_dec(&nr_events);
1705                 if (event->attr.mmap)
1706                         atomic_dec(&nr_mmap_events);
1707                 if (event->attr.comm)
1708                         atomic_dec(&nr_comm_events);
1709                 if (event->attr.task)
1710                         atomic_dec(&nr_task_events);
1711         }
1712
1713         if (event->output) {
1714                 fput(event->output->filp);
1715                 event->output = NULL;
1716         }
1717
1718         if (event->destroy)
1719                 event->destroy(event);
1720
1721         put_ctx(event->ctx);
1722         call_rcu(&event->rcu_head, free_event_rcu);
1723 }
1724
1725 int perf_event_release_kernel(struct perf_event *event)
1726 {
1727         struct perf_event_context *ctx = event->ctx;
1728
1729         WARN_ON_ONCE(ctx->parent_ctx);
1730         mutex_lock(&ctx->mutex);
1731         perf_event_remove_from_context(event);
1732         mutex_unlock(&ctx->mutex);
1733
1734         mutex_lock(&event->owner->perf_event_mutex);
1735         list_del_init(&event->owner_entry);
1736         mutex_unlock(&event->owner->perf_event_mutex);
1737         put_task_struct(event->owner);
1738
1739         free_event(event);
1740
1741         return 0;
1742 }
1743 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1744
1745 /*
1746  * Called when the last reference to the file is gone.
1747  */
1748 static int perf_release(struct inode *inode, struct file *file)
1749 {
1750         struct perf_event *event = file->private_data;
1751
1752         file->private_data = NULL;
1753
1754         return perf_event_release_kernel(event);
1755 }
1756
1757 static int perf_event_read_size(struct perf_event *event)
1758 {
1759         int entry = sizeof(u64); /* value */
1760         int size = 0;
1761         int nr = 1;
1762
1763         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1764                 size += sizeof(u64);
1765
1766         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1767                 size += sizeof(u64);
1768
1769         if (event->attr.read_format & PERF_FORMAT_ID)
1770                 entry += sizeof(u64);
1771
1772         if (event->attr.read_format & PERF_FORMAT_GROUP) {
1773                 nr += event->group_leader->nr_siblings;
1774                 size += sizeof(u64);
1775         }
1776
1777         size += entry * nr;
1778
1779         return size;
1780 }
1781
1782 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1783 {
1784         struct perf_event *child;
1785         u64 total = 0;
1786
1787         *enabled = 0;
1788         *running = 0;
1789
1790         mutex_lock(&event->child_mutex);
1791         total += perf_event_read(event);
1792         *enabled += event->total_time_enabled +
1793                         atomic64_read(&event->child_total_time_enabled);
1794         *running += event->total_time_running +
1795                         atomic64_read(&event->child_total_time_running);
1796
1797         list_for_each_entry(child, &event->child_list, child_list) {
1798                 total += perf_event_read(child);
1799                 *enabled += child->total_time_enabled;
1800                 *running += child->total_time_running;
1801         }
1802         mutex_unlock(&event->child_mutex);
1803
1804         return total;
1805 }
1806 EXPORT_SYMBOL_GPL(perf_event_read_value);
1807
1808 static int perf_event_read_group(struct perf_event *event,
1809                                    u64 read_format, char __user *buf)
1810 {
1811         struct perf_event *leader = event->group_leader, *sub;
1812         int n = 0, size = 0, ret = -EFAULT;
1813         struct perf_event_context *ctx = leader->ctx;
1814         u64 values[5];
1815         u64 count, enabled, running;
1816
1817         mutex_lock(&ctx->mutex);
1818         count = perf_event_read_value(leader, &enabled, &running);
1819
1820         values[n++] = 1 + leader->nr_siblings;
1821         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1822                 values[n++] = enabled;
1823         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1824                 values[n++] = running;
1825         values[n++] = count;
1826         if (read_format & PERF_FORMAT_ID)
1827                 values[n++] = primary_event_id(leader);
1828
1829         size = n * sizeof(u64);
1830
1831         if (copy_to_user(buf, values, size))
1832                 goto unlock;
1833
1834         ret = size;
1835
1836         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1837                 n = 0;
1838
1839                 values[n++] = perf_event_read_value(sub, &enabled, &running);
1840                 if (read_format & PERF_FORMAT_ID)
1841                         values[n++] = primary_event_id(sub);
1842
1843                 size = n * sizeof(u64);
1844
1845                 if (copy_to_user(buf + ret, values, size)) {
1846                         ret = -EFAULT;
1847                         goto unlock;
1848                 }
1849
1850                 ret += size;
1851         }
1852 unlock:
1853         mutex_unlock(&ctx->mutex);
1854
1855         return ret;
1856 }
1857
1858 static int perf_event_read_one(struct perf_event *event,
1859                                  u64 read_format, char __user *buf)
1860 {
1861         u64 enabled, running;
1862         u64 values[4];
1863         int n = 0;
1864
1865         values[n++] = perf_event_read_value(event, &enabled, &running);
1866         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1867                 values[n++] = enabled;
1868         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1869                 values[n++] = running;
1870         if (read_format & PERF_FORMAT_ID)
1871                 values[n++] = primary_event_id(event);
1872
1873         if (copy_to_user(buf, values, n * sizeof(u64)))
1874                 return -EFAULT;
1875
1876         return n * sizeof(u64);
1877 }
1878
1879 /*
1880  * Read the performance event - simple non blocking version for now
1881  */
1882 static ssize_t
1883 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
1884 {
1885         u64 read_format = event->attr.read_format;
1886         int ret;
1887
1888         /*
1889          * Return end-of-file for a read on a event that is in
1890          * error state (i.e. because it was pinned but it couldn't be
1891          * scheduled on to the CPU at some point).
1892          */
1893         if (event->state == PERF_EVENT_STATE_ERROR)
1894                 return 0;
1895
1896         if (count < perf_event_read_size(event))
1897                 return -ENOSPC;
1898
1899         WARN_ON_ONCE(event->ctx->parent_ctx);
1900         if (read_format & PERF_FORMAT_GROUP)
1901                 ret = perf_event_read_group(event, read_format, buf);
1902         else
1903                 ret = perf_event_read_one(event, read_format, buf);
1904
1905         return ret;
1906 }
1907
1908 static ssize_t
1909 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1910 {
1911         struct perf_event *event = file->private_data;
1912
1913         return perf_read_hw(event, buf, count);
1914 }
1915
1916 static unsigned int perf_poll(struct file *file, poll_table *wait)
1917 {
1918         struct perf_event *event = file->private_data;
1919         struct perf_mmap_data *data;
1920         unsigned int events = POLL_HUP;
1921
1922         rcu_read_lock();
1923         data = rcu_dereference(event->data);
1924         if (data)
1925                 events = atomic_xchg(&data->poll, 0);
1926         rcu_read_unlock();
1927
1928         poll_wait(file, &event->waitq, wait);
1929
1930         return events;
1931 }
1932
1933 static void perf_event_reset(struct perf_event *event)
1934 {
1935         (void)perf_event_read(event);
1936         atomic64_set(&event->count, 0);
1937         perf_event_update_userpage(event);
1938 }
1939
1940 /*
1941  * Holding the top-level event's child_mutex means that any
1942  * descendant process that has inherited this event will block
1943  * in sync_child_event if it goes to exit, thus satisfying the
1944  * task existence requirements of perf_event_enable/disable.
1945  */
1946 static void perf_event_for_each_child(struct perf_event *event,
1947                                         void (*func)(struct perf_event *))
1948 {
1949         struct perf_event *child;
1950
1951         WARN_ON_ONCE(event->ctx->parent_ctx);
1952         mutex_lock(&event->child_mutex);
1953         func(event);
1954         list_for_each_entry(child, &event->child_list, child_list)
1955                 func(child);
1956         mutex_unlock(&event->child_mutex);
1957 }
1958
1959 static void perf_event_for_each(struct perf_event *event,
1960                                   void (*func)(struct perf_event *))
1961 {
1962         struct perf_event_context *ctx = event->ctx;
1963         struct perf_event *sibling;
1964
1965         WARN_ON_ONCE(ctx->parent_ctx);
1966         mutex_lock(&ctx->mutex);
1967         event = event->group_leader;
1968
1969         perf_event_for_each_child(event, func);
1970         func(event);
1971         list_for_each_entry(sibling, &event->sibling_list, group_entry)
1972                 perf_event_for_each_child(event, func);
1973         mutex_unlock(&ctx->mutex);
1974 }
1975
1976 static int perf_event_period(struct perf_event *event, u64 __user *arg)
1977 {
1978         struct perf_event_context *ctx = event->ctx;
1979         unsigned long size;
1980         int ret = 0;
1981         u64 value;
1982
1983         if (!event->attr.sample_period)
1984                 return -EINVAL;
1985
1986         size = copy_from_user(&value, arg, sizeof(value));
1987         if (size != sizeof(value))
1988                 return -EFAULT;
1989
1990         if (!value)
1991                 return -EINVAL;
1992
1993         spin_lock_irq(&ctx->lock);
1994         if (event->attr.freq) {
1995                 if (value > sysctl_perf_event_sample_rate) {
1996                         ret = -EINVAL;
1997                         goto unlock;
1998                 }
1999
2000                 event->attr.sample_freq = value;
2001         } else {
2002                 event->attr.sample_period = value;
2003                 event->hw.sample_period = value;
2004         }
2005 unlock:
2006         spin_unlock_irq(&ctx->lock);
2007
2008         return ret;
2009 }
2010
2011 static int perf_event_set_output(struct perf_event *event, int output_fd);
2012 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2013
2014 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2015 {
2016         struct perf_event *event = file->private_data;
2017         void (*func)(struct perf_event *);
2018         u32 flags = arg;
2019
2020         switch (cmd) {
2021         case PERF_EVENT_IOC_ENABLE:
2022                 func = perf_event_enable;
2023                 break;
2024         case PERF_EVENT_IOC_DISABLE:
2025                 func = perf_event_disable;
2026                 break;
2027         case PERF_EVENT_IOC_RESET:
2028                 func = perf_event_reset;
2029                 break;
2030
2031         case PERF_EVENT_IOC_REFRESH:
2032                 return perf_event_refresh(event, arg);
2033
2034         case PERF_EVENT_IOC_PERIOD:
2035                 return perf_event_period(event, (u64 __user *)arg);
2036
2037         case PERF_EVENT_IOC_SET_OUTPUT:
2038                 return perf_event_set_output(event, arg);
2039
2040         case PERF_EVENT_IOC_SET_FILTER:
2041                 return perf_event_set_filter(event, (void __user *)arg);
2042
2043         default:
2044                 return -ENOTTY;
2045         }
2046
2047         if (flags & PERF_IOC_FLAG_GROUP)
2048                 perf_event_for_each(event, func);
2049         else
2050                 perf_event_for_each_child(event, func);
2051
2052         return 0;
2053 }
2054
2055 int perf_event_task_enable(void)
2056 {
2057         struct perf_event *event;
2058
2059         mutex_lock(&current->perf_event_mutex);
2060         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2061                 perf_event_for_each_child(event, perf_event_enable);
2062         mutex_unlock(&current->perf_event_mutex);
2063
2064         return 0;
2065 }
2066
2067 int perf_event_task_disable(void)
2068 {
2069         struct perf_event *event;
2070
2071         mutex_lock(&current->perf_event_mutex);
2072         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2073                 perf_event_for_each_child(event, perf_event_disable);
2074         mutex_unlock(&current->perf_event_mutex);
2075
2076         return 0;
2077 }
2078
2079 #ifndef PERF_EVENT_INDEX_OFFSET
2080 # define PERF_EVENT_INDEX_OFFSET 0
2081 #endif
2082
2083 static int perf_event_index(struct perf_event *event)
2084 {
2085         if (event->state != PERF_EVENT_STATE_ACTIVE)
2086                 return 0;
2087
2088         return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2089 }
2090
2091 /*
2092  * Callers need to ensure there can be no nesting of this function, otherwise
2093  * the seqlock logic goes bad. We can not serialize this because the arch
2094  * code calls this from NMI context.
2095  */
2096 void perf_event_update_userpage(struct perf_event *event)
2097 {
2098         struct perf_event_mmap_page *userpg;
2099         struct perf_mmap_data *data;
2100
2101         rcu_read_lock();
2102         data = rcu_dereference(event->data);
2103         if (!data)
2104                 goto unlock;
2105
2106         userpg = data->user_page;
2107
2108         /*
2109          * Disable preemption so as to not let the corresponding user-space
2110          * spin too long if we get preempted.
2111          */
2112         preempt_disable();
2113         ++userpg->lock;
2114         barrier();
2115         userpg->index = perf_event_index(event);
2116         userpg->offset = atomic64_read(&event->count);
2117         if (event->state == PERF_EVENT_STATE_ACTIVE)
2118                 userpg->offset -= atomic64_read(&event->hw.prev_count);
2119
2120         userpg->time_enabled = event->total_time_enabled +
2121                         atomic64_read(&event->child_total_time_enabled);
2122
2123         userpg->time_running = event->total_time_running +
2124                         atomic64_read(&event->child_total_time_running);
2125
2126         barrier();
2127         ++userpg->lock;
2128         preempt_enable();
2129 unlock:
2130         rcu_read_unlock();
2131 }
2132
2133 static unsigned long perf_data_size(struct perf_mmap_data *data)
2134 {
2135         return data->nr_pages << (PAGE_SHIFT + data->data_order);
2136 }
2137
2138 #ifndef CONFIG_PERF_USE_VMALLOC
2139
2140 /*
2141  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2142  */
2143
2144 static struct page *
2145 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2146 {
2147         if (pgoff > data->nr_pages)
2148                 return NULL;
2149
2150         if (pgoff == 0)
2151                 return virt_to_page(data->user_page);
2152
2153         return virt_to_page(data->data_pages[pgoff - 1]);
2154 }
2155
2156 static struct perf_mmap_data *
2157 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2158 {
2159         struct perf_mmap_data *data;
2160         unsigned long size;
2161         int i;
2162
2163         WARN_ON(atomic_read(&event->mmap_count));
2164
2165         size = sizeof(struct perf_mmap_data);
2166         size += nr_pages * sizeof(void *);
2167
2168         data = kzalloc(size, GFP_KERNEL);
2169         if (!data)
2170                 goto fail;
2171
2172         data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2173         if (!data->user_page)
2174                 goto fail_user_page;
2175
2176         for (i = 0; i < nr_pages; i++) {
2177                 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2178                 if (!data->data_pages[i])
2179                         goto fail_data_pages;
2180         }
2181
2182         data->data_order = 0;
2183         data->nr_pages = nr_pages;
2184
2185         return data;
2186
2187 fail_data_pages:
2188         for (i--; i >= 0; i--)
2189                 free_page((unsigned long)data->data_pages[i]);
2190
2191         free_page((unsigned long)data->user_page);
2192
2193 fail_user_page:
2194         kfree(data);
2195
2196 fail:
2197         return NULL;
2198 }
2199
2200 static void perf_mmap_free_page(unsigned long addr)
2201 {
2202         struct page *page = virt_to_page((void *)addr);
2203
2204         page->mapping = NULL;
2205         __free_page(page);
2206 }
2207
2208 static void perf_mmap_data_free(struct perf_mmap_data *data)
2209 {
2210         int i;
2211
2212         perf_mmap_free_page((unsigned long)data->user_page);
2213         for (i = 0; i < data->nr_pages; i++)
2214                 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2215         kfree(data);
2216 }
2217
2218 #else
2219
2220 /*
2221  * Back perf_mmap() with vmalloc memory.
2222  *
2223  * Required for architectures that have d-cache aliasing issues.
2224  */
2225
2226 static struct page *
2227 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2228 {
2229         if (pgoff > (1UL << data->data_order))
2230                 return NULL;
2231
2232         return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2233 }
2234
2235 static void perf_mmap_unmark_page(void *addr)
2236 {
2237         struct page *page = vmalloc_to_page(addr);
2238
2239         page->mapping = NULL;
2240 }
2241
2242 static void perf_mmap_data_free_work(struct work_struct *work)
2243 {
2244         struct perf_mmap_data *data;
2245         void *base;
2246         int i, nr;
2247
2248         data = container_of(work, struct perf_mmap_data, work);
2249         nr = 1 << data->data_order;
2250
2251         base = data->user_page;
2252         for (i = 0; i < nr + 1; i++)
2253                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2254
2255         vfree(base);
2256         kfree(data);
2257 }
2258
2259 static void perf_mmap_data_free(struct perf_mmap_data *data)
2260 {
2261         schedule_work(&data->work);
2262 }
2263
2264 static struct perf_mmap_data *
2265 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2266 {
2267         struct perf_mmap_data *data;
2268         unsigned long size;
2269         void *all_buf;
2270
2271         WARN_ON(atomic_read(&event->mmap_count));
2272
2273         size = sizeof(struct perf_mmap_data);
2274         size += sizeof(void *);
2275
2276         data = kzalloc(size, GFP_KERNEL);
2277         if (!data)
2278                 goto fail;
2279
2280         INIT_WORK(&data->work, perf_mmap_data_free_work);
2281
2282         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2283         if (!all_buf)
2284                 goto fail_all_buf;
2285
2286         data->user_page = all_buf;
2287         data->data_pages[0] = all_buf + PAGE_SIZE;
2288         data->data_order = ilog2(nr_pages);
2289         data->nr_pages = 1;
2290
2291         return data;
2292
2293 fail_all_buf:
2294         kfree(data);
2295
2296 fail:
2297         return NULL;
2298 }
2299
2300 #endif
2301
2302 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2303 {
2304         struct perf_event *event = vma->vm_file->private_data;
2305         struct perf_mmap_data *data;
2306         int ret = VM_FAULT_SIGBUS;
2307
2308         if (vmf->flags & FAULT_FLAG_MKWRITE) {
2309                 if (vmf->pgoff == 0)
2310                         ret = 0;
2311                 return ret;
2312         }
2313
2314         rcu_read_lock();
2315         data = rcu_dereference(event->data);
2316         if (!data)
2317                 goto unlock;
2318
2319         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2320                 goto unlock;
2321
2322         vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2323         if (!vmf->page)
2324                 goto unlock;
2325
2326         get_page(vmf->page);
2327         vmf->page->mapping = vma->vm_file->f_mapping;
2328         vmf->page->index   = vmf->pgoff;
2329
2330         ret = 0;
2331 unlock:
2332         rcu_read_unlock();
2333
2334         return ret;
2335 }
2336
2337 static void
2338 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2339 {
2340         long max_size = perf_data_size(data);
2341
2342         atomic_set(&data->lock, -1);
2343
2344         if (event->attr.watermark) {
2345                 data->watermark = min_t(long, max_size,
2346                                         event->attr.wakeup_watermark);
2347         }
2348
2349         if (!data->watermark)
2350                 data->watermark = max_size / 2;
2351
2352
2353         rcu_assign_pointer(event->data, data);
2354 }
2355
2356 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2357 {
2358         struct perf_mmap_data *data;
2359
2360         data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2361         perf_mmap_data_free(data);
2362 }
2363
2364 static void perf_mmap_data_release(struct perf_event *event)
2365 {
2366         struct perf_mmap_data *data = event->data;
2367
2368         WARN_ON(atomic_read(&event->mmap_count));
2369
2370         rcu_assign_pointer(event->data, NULL);
2371         call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2372 }
2373
2374 static void perf_mmap_open(struct vm_area_struct *vma)
2375 {
2376         struct perf_event *event = vma->vm_file->private_data;
2377
2378         atomic_inc(&event->mmap_count);
2379 }
2380
2381 static void perf_mmap_close(struct vm_area_struct *vma)
2382 {
2383         struct perf_event *event = vma->vm_file->private_data;
2384
2385         WARN_ON_ONCE(event->ctx->parent_ctx);
2386         if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2387                 unsigned long size = perf_data_size(event->data);
2388                 struct user_struct *user = current_user();
2389
2390                 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2391                 vma->vm_mm->locked_vm -= event->data->nr_locked;
2392                 perf_mmap_data_release(event);
2393                 mutex_unlock(&event->mmap_mutex);
2394         }
2395 }
2396
2397 static const struct vm_operations_struct perf_mmap_vmops = {
2398         .open           = perf_mmap_open,
2399         .close          = perf_mmap_close,
2400         .fault          = perf_mmap_fault,
2401         .page_mkwrite   = perf_mmap_fault,
2402 };
2403
2404 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2405 {
2406         struct perf_event *event = file->private_data;
2407         unsigned long user_locked, user_lock_limit;
2408         struct user_struct *user = current_user();
2409         unsigned long locked, lock_limit;
2410         struct perf_mmap_data *data;
2411         unsigned long vma_size;
2412         unsigned long nr_pages;
2413         long user_extra, extra;
2414         int ret = 0;
2415
2416         if (!(vma->vm_flags & VM_SHARED))
2417                 return -EINVAL;
2418
2419         vma_size = vma->vm_end - vma->vm_start;
2420         nr_pages = (vma_size / PAGE_SIZE) - 1;
2421
2422         /*
2423          * If we have data pages ensure they're a power-of-two number, so we
2424          * can do bitmasks instead of modulo.
2425          */
2426         if (nr_pages != 0 && !is_power_of_2(nr_pages))
2427                 return -EINVAL;
2428
2429         if (vma_size != PAGE_SIZE * (1 + nr_pages))
2430                 return -EINVAL;
2431
2432         if (vma->vm_pgoff != 0)
2433                 return -EINVAL;
2434
2435         WARN_ON_ONCE(event->ctx->parent_ctx);
2436         mutex_lock(&event->mmap_mutex);
2437         if (event->output) {
2438                 ret = -EINVAL;
2439                 goto unlock;
2440         }
2441
2442         if (atomic_inc_not_zero(&event->mmap_count)) {
2443                 if (nr_pages != event->data->nr_pages)
2444                         ret = -EINVAL;
2445                 goto unlock;
2446         }
2447
2448         user_extra = nr_pages + 1;
2449         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2450
2451         /*
2452          * Increase the limit linearly with more CPUs:
2453          */
2454         user_lock_limit *= num_online_cpus();
2455
2456         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2457
2458         extra = 0;
2459         if (user_locked > user_lock_limit)
2460                 extra = user_locked - user_lock_limit;
2461
2462         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2463         lock_limit >>= PAGE_SHIFT;
2464         locked = vma->vm_mm->locked_vm + extra;
2465
2466         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2467                 !capable(CAP_IPC_LOCK)) {
2468                 ret = -EPERM;
2469                 goto unlock;
2470         }
2471
2472         WARN_ON(event->data);
2473
2474         data = perf_mmap_data_alloc(event, nr_pages);
2475         ret = -ENOMEM;
2476         if (!data)
2477                 goto unlock;
2478
2479         ret = 0;
2480         perf_mmap_data_init(event, data);
2481
2482         atomic_set(&event->mmap_count, 1);
2483         atomic_long_add(user_extra, &user->locked_vm);
2484         vma->vm_mm->locked_vm += extra;
2485         event->data->nr_locked = extra;
2486         if (vma->vm_flags & VM_WRITE)
2487                 event->data->writable = 1;
2488
2489 unlock:
2490         mutex_unlock(&event->mmap_mutex);
2491
2492         vma->vm_flags |= VM_RESERVED;
2493         vma->vm_ops = &perf_mmap_vmops;
2494
2495         return ret;
2496 }
2497
2498 static int perf_fasync(int fd, struct file *filp, int on)
2499 {
2500         struct inode *inode = filp->f_path.dentry->d_inode;
2501         struct perf_event *event = filp->private_data;
2502         int retval;
2503
2504         mutex_lock(&inode->i_mutex);
2505         retval = fasync_helper(fd, filp, on, &event->fasync);
2506         mutex_unlock(&inode->i_mutex);
2507
2508         if (retval < 0)
2509                 return retval;
2510
2511         return 0;
2512 }
2513
2514 static const struct file_operations perf_fops = {
2515         .release                = perf_release,
2516         .read                   = perf_read,
2517         .poll                   = perf_poll,
2518         .unlocked_ioctl         = perf_ioctl,
2519         .compat_ioctl           = perf_ioctl,
2520         .mmap                   = perf_mmap,
2521         .fasync                 = perf_fasync,
2522 };
2523
2524 /*
2525  * Perf event wakeup
2526  *
2527  * If there's data, ensure we set the poll() state and publish everything
2528  * to user-space before waking everybody up.
2529  */
2530
2531 void perf_event_wakeup(struct perf_event *event)
2532 {
2533         wake_up_all(&event->waitq);
2534
2535         if (event->pending_kill) {
2536                 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2537                 event->pending_kill = 0;
2538         }
2539 }
2540
2541 /*
2542  * Pending wakeups
2543  *
2544  * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2545  *
2546  * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2547  * single linked list and use cmpxchg() to add entries lockless.
2548  */
2549
2550 static void perf_pending_event(struct perf_pending_entry *entry)
2551 {
2552         struct perf_event *event = container_of(entry,
2553                         struct perf_event, pending);
2554
2555         if (event->pending_disable) {
2556                 event->pending_disable = 0;
2557                 __perf_event_disable(event);
2558         }
2559
2560         if (event->pending_wakeup) {
2561                 event->pending_wakeup = 0;
2562                 perf_event_wakeup(event);
2563         }
2564 }
2565
2566 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2567
2568 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2569         PENDING_TAIL,
2570 };
2571
2572 static void perf_pending_queue(struct perf_pending_entry *entry,
2573                                void (*func)(struct perf_pending_entry *))
2574 {
2575         struct perf_pending_entry **head;
2576
2577         if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2578                 return;
2579
2580         entry->func = func;
2581
2582         head = &get_cpu_var(perf_pending_head);
2583
2584         do {
2585                 entry->next = *head;
2586         } while (cmpxchg(head, entry->next, entry) != entry->next);
2587
2588         set_perf_event_pending();
2589
2590         put_cpu_var(perf_pending_head);
2591 }
2592
2593 static int __perf_pending_run(void)
2594 {
2595         struct perf_pending_entry *list;
2596         int nr = 0;
2597
2598         list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2599         while (list != PENDING_TAIL) {
2600                 void (*func)(struct perf_pending_entry *);
2601                 struct perf_pending_entry *entry = list;
2602
2603                 list = list->next;
2604
2605                 func = entry->func;
2606                 entry->next = NULL;
2607                 /*
2608                  * Ensure we observe the unqueue before we issue the wakeup,
2609                  * so that we won't be waiting forever.
2610                  * -- see perf_not_pending().
2611                  */
2612                 smp_wmb();
2613
2614                 func(entry);
2615                 nr++;
2616         }
2617
2618         return nr;
2619 }
2620
2621 static inline int perf_not_pending(struct perf_event *event)
2622 {
2623         /*
2624          * If we flush on whatever cpu we run, there is a chance we don't
2625          * need to wait.
2626          */
2627         get_cpu();
2628         __perf_pending_run();
2629         put_cpu();
2630
2631         /*
2632          * Ensure we see the proper queue state before going to sleep
2633          * so that we do not miss the wakeup. -- see perf_pending_handle()
2634          */
2635         smp_rmb();
2636         return event->pending.next == NULL;
2637 }
2638
2639 static void perf_pending_sync(struct perf_event *event)
2640 {
2641         wait_event(event->waitq, perf_not_pending(event));
2642 }
2643
2644 void perf_event_do_pending(void)
2645 {
2646         __perf_pending_run();
2647 }
2648
2649 /*
2650  * Callchain support -- arch specific
2651  */
2652
2653 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2654 {
2655         return NULL;
2656 }
2657
2658 /*
2659  * Output
2660  */
2661 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2662                               unsigned long offset, unsigned long head)
2663 {
2664         unsigned long mask;
2665
2666         if (!data->writable)
2667                 return true;
2668
2669         mask = perf_data_size(data) - 1;
2670
2671         offset = (offset - tail) & mask;
2672         head   = (head   - tail) & mask;
2673
2674         if ((int)(head - offset) < 0)
2675                 return false;
2676
2677         return true;
2678 }
2679
2680 static void perf_output_wakeup(struct perf_output_handle *handle)
2681 {
2682         atomic_set(&handle->data->poll, POLL_IN);
2683
2684         if (handle->nmi) {
2685                 handle->event->pending_wakeup = 1;
2686                 perf_pending_queue(&handle->event->pending,
2687                                    perf_pending_event);
2688         } else
2689                 perf_event_wakeup(handle->event);
2690 }
2691
2692 /*
2693  * Curious locking construct.
2694  *
2695  * We need to ensure a later event_id doesn't publish a head when a former
2696  * event_id isn't done writing. However since we need to deal with NMIs we
2697  * cannot fully serialize things.
2698  *
2699  * What we do is serialize between CPUs so we only have to deal with NMI
2700  * nesting on a single CPU.
2701  *
2702  * We only publish the head (and generate a wakeup) when the outer-most
2703  * event_id completes.
2704  */
2705 static void perf_output_lock(struct perf_output_handle *handle)
2706 {
2707         struct perf_mmap_data *data = handle->data;
2708         int cur, cpu = get_cpu();
2709
2710         handle->locked = 0;
2711
2712         for (;;) {
2713                 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2714                 if (cur == -1) {
2715                         handle->locked = 1;
2716                         break;
2717                 }
2718                 if (cur == cpu)
2719                         break;
2720
2721                 cpu_relax();
2722         }
2723 }
2724
2725 static void perf_output_unlock(struct perf_output_handle *handle)
2726 {
2727         struct perf_mmap_data *data = handle->data;
2728         unsigned long head;
2729         int cpu;
2730
2731         data->done_head = data->head;
2732
2733         if (!handle->locked)
2734                 goto out;
2735
2736 again:
2737         /*
2738          * The xchg implies a full barrier that ensures all writes are done
2739          * before we publish the new head, matched by a rmb() in userspace when
2740          * reading this position.
2741          */
2742         while ((head = atomic_long_xchg(&data->done_head, 0)))
2743                 data->user_page->data_head = head;
2744
2745         /*
2746          * NMI can happen here, which means we can miss a done_head update.
2747          */
2748
2749         cpu = atomic_xchg(&data->lock, -1);
2750         WARN_ON_ONCE(cpu != smp_processor_id());
2751
2752         /*
2753          * Therefore we have to validate we did not indeed do so.
2754          */
2755         if (unlikely(atomic_long_read(&data->done_head))) {
2756                 /*
2757                  * Since we had it locked, we can lock it again.
2758                  */
2759                 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2760                         cpu_relax();
2761
2762                 goto again;
2763         }
2764
2765         if (atomic_xchg(&data->wakeup, 0))
2766                 perf_output_wakeup(handle);
2767 out:
2768         put_cpu();
2769 }
2770
2771 void perf_output_copy(struct perf_output_handle *handle,
2772                       const void *buf, unsigned int len)
2773 {
2774         unsigned int pages_mask;
2775         unsigned long offset;
2776         unsigned int size;
2777         void **pages;
2778
2779         offset          = handle->offset;
2780         pages_mask      = handle->data->nr_pages - 1;
2781         pages           = handle->data->data_pages;
2782
2783         do {
2784                 unsigned long page_offset;
2785                 unsigned long page_size;
2786                 int nr;
2787
2788                 nr          = (offset >> PAGE_SHIFT) & pages_mask;
2789                 page_size   = 1UL << (handle->data->data_order + PAGE_SHIFT);
2790                 page_offset = offset & (page_size - 1);
2791                 size        = min_t(unsigned int, page_size - page_offset, len);
2792
2793                 memcpy(pages[nr] + page_offset, buf, size);
2794
2795                 len         -= size;
2796                 buf         += size;
2797                 offset      += size;
2798         } while (len);
2799
2800         handle->offset = offset;
2801
2802         /*
2803          * Check we didn't copy past our reservation window, taking the
2804          * possible unsigned int wrap into account.
2805          */
2806         WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2807 }
2808
2809 int perf_output_begin(struct perf_output_handle *handle,
2810                       struct perf_event *event, unsigned int size,
2811                       int nmi, int sample)
2812 {
2813         struct perf_event *output_event;
2814         struct perf_mmap_data *data;
2815         unsigned long tail, offset, head;
2816         int have_lost;
2817         struct {
2818                 struct perf_event_header header;
2819                 u64                      id;
2820                 u64                      lost;
2821         } lost_event;
2822
2823         rcu_read_lock();
2824         /*
2825          * For inherited events we send all the output towards the parent.
2826          */
2827         if (event->parent)
2828                 event = event->parent;
2829
2830         output_event = rcu_dereference(event->output);
2831         if (output_event)
2832                 event = output_event;
2833
2834         data = rcu_dereference(event->data);
2835         if (!data)
2836                 goto out;
2837
2838         handle->data    = data;
2839         handle->event   = event;
2840         handle->nmi     = nmi;
2841         handle->sample  = sample;
2842
2843         if (!data->nr_pages)
2844                 goto fail;
2845
2846         have_lost = atomic_read(&data->lost);
2847         if (have_lost)
2848                 size += sizeof(lost_event);
2849
2850         perf_output_lock(handle);
2851
2852         do {
2853                 /*
2854                  * Userspace could choose to issue a mb() before updating the
2855                  * tail pointer. So that all reads will be completed before the
2856                  * write is issued.
2857                  */
2858                 tail = ACCESS_ONCE(data->user_page->data_tail);
2859                 smp_rmb();
2860                 offset = head = atomic_long_read(&data->head);
2861                 head += size;
2862                 if (unlikely(!perf_output_space(data, tail, offset, head)))
2863                         goto fail;
2864         } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2865
2866         handle->offset  = offset;
2867         handle->head    = head;
2868
2869         if (head - tail > data->watermark)
2870                 atomic_set(&data->wakeup, 1);
2871
2872         if (have_lost) {
2873                 lost_event.header.type = PERF_RECORD_LOST;
2874                 lost_event.header.misc = 0;
2875                 lost_event.header.size = sizeof(lost_event);
2876                 lost_event.id          = event->id;
2877                 lost_event.lost        = atomic_xchg(&data->lost, 0);
2878
2879                 perf_output_put(handle, lost_event);
2880         }
2881
2882         return 0;
2883
2884 fail:
2885         atomic_inc(&data->lost);
2886         perf_output_unlock(handle);
2887 out:
2888         rcu_read_unlock();
2889
2890         return -ENOSPC;
2891 }
2892
2893 void perf_output_end(struct perf_output_handle *handle)
2894 {
2895         struct perf_event *event = handle->event;
2896         struct perf_mmap_data *data = handle->data;
2897
2898         int wakeup_events = event->attr.wakeup_events;
2899
2900         if (handle->sample && wakeup_events) {
2901                 int events = atomic_inc_return(&data->events);
2902                 if (events >= wakeup_events) {
2903                         atomic_sub(wakeup_events, &data->events);
2904                         atomic_set(&data->wakeup, 1);
2905                 }
2906         }
2907
2908         perf_output_unlock(handle);
2909         rcu_read_unlock();
2910 }
2911
2912 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
2913 {
2914         /*
2915          * only top level events have the pid namespace they were created in
2916          */
2917         if (event->parent)
2918                 event = event->parent;
2919
2920         return task_tgid_nr_ns(p, event->ns);
2921 }
2922
2923 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
2924 {
2925         /*
2926          * only top level events have the pid namespace they were created in
2927          */
2928         if (event->parent)
2929                 event = event->parent;
2930
2931         return task_pid_nr_ns(p, event->ns);
2932 }
2933
2934 static void perf_output_read_one(struct perf_output_handle *handle,
2935                                  struct perf_event *event)
2936 {
2937         u64 read_format = event->attr.read_format;
2938         u64 values[4];
2939         int n = 0;
2940
2941         values[n++] = atomic64_read(&event->count);
2942         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2943                 values[n++] = event->total_time_enabled +
2944                         atomic64_read(&event->child_total_time_enabled);
2945         }
2946         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2947                 values[n++] = event->total_time_running +
2948                         atomic64_read(&event->child_total_time_running);
2949         }
2950         if (read_format & PERF_FORMAT_ID)
2951                 values[n++] = primary_event_id(event);
2952
2953         perf_output_copy(handle, values, n * sizeof(u64));
2954 }
2955
2956 /*
2957  * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
2958  */
2959 static void perf_output_read_group(struct perf_output_handle *handle,
2960                             struct perf_event *event)
2961 {
2962         struct perf_event *leader = event->group_leader, *sub;
2963         u64 read_format = event->attr.read_format;
2964         u64 values[5];
2965         int n = 0;
2966
2967         values[n++] = 1 + leader->nr_siblings;
2968
2969         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2970                 values[n++] = leader->total_time_enabled;
2971
2972         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2973                 values[n++] = leader->total_time_running;
2974
2975         if (leader != event)
2976                 leader->pmu->read(leader);
2977
2978         values[n++] = atomic64_read(&leader->count);
2979         if (read_format & PERF_FORMAT_ID)
2980                 values[n++] = primary_event_id(leader);
2981
2982         perf_output_copy(handle, values, n * sizeof(u64));
2983
2984         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2985                 n = 0;
2986
2987                 if (sub != event)
2988                         sub->pmu->read(sub);
2989
2990                 values[n++] = atomic64_read(&sub->count);
2991                 if (read_format & PERF_FORMAT_ID)
2992                         values[n++] = primary_event_id(sub);
2993
2994                 perf_output_copy(handle, values, n * sizeof(u64));
2995         }
2996 }
2997
2998 static void perf_output_read(struct perf_output_handle *handle,
2999                              struct perf_event *event)
3000 {
3001         if (event->attr.read_format & PERF_FORMAT_GROUP)
3002                 perf_output_read_group(handle, event);
3003         else
3004                 perf_output_read_one(handle, event);
3005 }
3006
3007 void perf_output_sample(struct perf_output_handle *handle,
3008                         struct perf_event_header *header,
3009                         struct perf_sample_data *data,
3010                         struct perf_event *event)
3011 {
3012         u64 sample_type = data->type;
3013
3014         perf_output_put(handle, *header);
3015
3016         if (sample_type & PERF_SAMPLE_IP)
3017                 perf_output_put(handle, data->ip);
3018
3019         if (sample_type & PERF_SAMPLE_TID)
3020                 perf_output_put(handle, data->tid_entry);
3021
3022         if (sample_type & PERF_SAMPLE_TIME)
3023                 perf_output_put(handle, data->time);
3024
3025         if (sample_type & PERF_SAMPLE_ADDR)
3026                 perf_output_put(handle, data->addr);
3027
3028         if (sample_type & PERF_SAMPLE_ID)
3029                 perf_output_put(handle, data->id);
3030
3031         if (sample_type & PERF_SAMPLE_STREAM_ID)
3032                 perf_output_put(handle, data->stream_id);
3033
3034         if (sample_type & PERF_SAMPLE_CPU)
3035                 perf_output_put(handle, data->cpu_entry);
3036
3037         if (sample_type & PERF_SAMPLE_PERIOD)
3038                 perf_output_put(handle, data->period);
3039
3040         if (sample_type & PERF_SAMPLE_READ)
3041                 perf_output_read(handle, event);
3042
3043         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3044                 if (data->callchain) {
3045                         int size = 1;
3046
3047                         if (data->callchain)
3048                                 size += data->callchain->nr;
3049
3050                         size *= sizeof(u64);
3051
3052                         perf_output_copy(handle, data->callchain, size);
3053                 } else {
3054                         u64 nr = 0;
3055                         perf_output_put(handle, nr);
3056                 }
3057         }
3058
3059         if (sample_type & PERF_SAMPLE_RAW) {
3060                 if (data->raw) {
3061                         perf_output_put(handle, data->raw->size);
3062                         perf_output_copy(handle, data->raw->data,
3063                                          data->raw->size);
3064                 } else {
3065                         struct {
3066                                 u32     size;
3067                                 u32     data;
3068                         } raw = {
3069                                 .size = sizeof(u32),
3070                                 .data = 0,
3071                         };
3072                         perf_output_put(handle, raw);
3073                 }
3074         }
3075 }
3076
3077 void perf_prepare_sample(struct perf_event_header *header,
3078                          struct perf_sample_data *data,
3079                          struct perf_event *event,
3080                          struct pt_regs *regs)
3081 {
3082         u64 sample_type = event->attr.sample_type;
3083
3084         data->type = sample_type;
3085
3086         header->type = PERF_RECORD_SAMPLE;
3087         header->size = sizeof(*header);
3088
3089         header->misc = 0;
3090         header->misc |= perf_misc_flags(regs);
3091
3092         if (sample_type & PERF_SAMPLE_IP) {
3093                 data->ip = perf_instruction_pointer(regs);
3094
3095                 header->size += sizeof(data->ip);
3096         }
3097
3098         if (sample_type & PERF_SAMPLE_TID) {
3099                 /* namespace issues */
3100                 data->tid_entry.pid = perf_event_pid(event, current);
3101                 data->tid_entry.tid = perf_event_tid(event, current);
3102
3103                 header->size += sizeof(data->tid_entry);
3104         }
3105
3106         if (sample_type & PERF_SAMPLE_TIME) {
3107                 data->time = perf_clock();
3108
3109                 header->size += sizeof(data->time);
3110         }
3111
3112         if (sample_type & PERF_SAMPLE_ADDR)
3113                 header->size += sizeof(data->addr);
3114
3115         if (sample_type & PERF_SAMPLE_ID) {
3116                 data->id = primary_event_id(event);
3117
3118                 header->size += sizeof(data->id);
3119         }
3120
3121         if (sample_type & PERF_SAMPLE_STREAM_ID) {
3122                 data->stream_id = event->id;
3123
3124                 header->size += sizeof(data->stream_id);
3125         }
3126
3127         if (sample_type & PERF_SAMPLE_CPU) {
3128                 data->cpu_entry.cpu             = raw_smp_processor_id();
3129                 data->cpu_entry.reserved        = 0;
3130
3131                 header->size += sizeof(data->cpu_entry);
3132         }
3133
3134         if (sample_type & PERF_SAMPLE_PERIOD)
3135                 header->size += sizeof(data->period);
3136
3137         if (sample_type & PERF_SAMPLE_READ)
3138                 header->size += perf_event_read_size(event);
3139
3140         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3141                 int size = 1;
3142
3143                 data->callchain = perf_callchain(regs);
3144
3145                 if (data->callchain)
3146                         size += data->callchain->nr;
3147
3148                 header->size += size * sizeof(u64);
3149         }
3150
3151         if (sample_type & PERF_SAMPLE_RAW) {
3152                 int size = sizeof(u32);
3153
3154                 if (data->raw)
3155                         size += data->raw->size;
3156                 else
3157                         size += sizeof(u32);
3158
3159                 WARN_ON_ONCE(size & (sizeof(u64)-1));
3160                 header->size += size;
3161         }
3162 }
3163
3164 static void perf_event_output(struct perf_event *event, int nmi,
3165                                 struct perf_sample_data *data,
3166                                 struct pt_regs *regs)
3167 {
3168         struct perf_output_handle handle;
3169         struct perf_event_header header;
3170
3171         perf_prepare_sample(&header, data, event, regs);
3172
3173         if (perf_output_begin(&handle, event, header.size, nmi, 1))
3174                 return;
3175
3176         perf_output_sample(&handle, &header, data, event);
3177
3178         perf_output_end(&handle);
3179 }
3180
3181 /*
3182  * read event_id
3183  */
3184
3185 struct perf_read_event {
3186         struct perf_event_header        header;
3187
3188         u32                             pid;
3189         u32                             tid;
3190 };
3191
3192 static void
3193 perf_event_read_event(struct perf_event *event,
3194                         struct task_struct *task)
3195 {
3196         struct perf_output_handle handle;
3197         struct perf_read_event read_event = {
3198                 .header = {
3199                         .type = PERF_RECORD_READ,
3200                         .misc = 0,
3201                         .size = sizeof(read_event) + perf_event_read_size(event),
3202                 },
3203                 .pid = perf_event_pid(event, task),
3204                 .tid = perf_event_tid(event, task),
3205         };
3206         int ret;
3207
3208         ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3209         if (ret)
3210                 return;
3211
3212         perf_output_put(&handle, read_event);
3213         perf_output_read(&handle, event);
3214
3215         perf_output_end(&handle);
3216 }
3217
3218 /*
3219  * task tracking -- fork/exit
3220  *
3221  * enabled by: attr.comm | attr.mmap | attr.task
3222  */
3223
3224 struct perf_task_event {
3225         struct task_struct              *task;
3226         struct perf_event_context       *task_ctx;
3227
3228         struct {
3229                 struct perf_event_header        header;
3230
3231                 u32                             pid;
3232                 u32                             ppid;
3233                 u32                             tid;
3234                 u32                             ptid;
3235                 u64                             time;
3236         } event_id;
3237 };
3238
3239 static void perf_event_task_output(struct perf_event *event,
3240                                      struct perf_task_event *task_event)
3241 {
3242         struct perf_output_handle handle;
3243         int size;
3244         struct task_struct *task = task_event->task;
3245         int ret;
3246
3247         size  = task_event->event_id.header.size;
3248         ret = perf_output_begin(&handle, event, size, 0, 0);
3249
3250         if (ret)
3251                 return;
3252
3253         task_event->event_id.pid = perf_event_pid(event, task);
3254         task_event->event_id.ppid = perf_event_pid(event, current);
3255
3256         task_event->event_id.tid = perf_event_tid(event, task);
3257         task_event->event_id.ptid = perf_event_tid(event, current);
3258
3259         task_event->event_id.time = perf_clock();
3260
3261         perf_output_put(&handle, task_event->event_id);
3262
3263         perf_output_end(&handle);
3264 }
3265
3266 static int perf_event_task_match(struct perf_event *event)
3267 {
3268         if (event->attr.comm || event->attr.mmap || event->attr.task)
3269                 return 1;
3270
3271         return 0;
3272 }
3273
3274 static void perf_event_task_ctx(struct perf_event_context *ctx,
3275                                   struct perf_task_event *task_event)
3276 {
3277         struct perf_event *event;
3278
3279         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3280                 if (perf_event_task_match(event))
3281                         perf_event_task_output(event, task_event);
3282         }
3283 }
3284
3285 static void perf_event_task_event(struct perf_task_event *task_event)
3286 {
3287         struct perf_cpu_context *cpuctx;
3288         struct perf_event_context *ctx = task_event->task_ctx;
3289
3290         rcu_read_lock();
3291         cpuctx = &get_cpu_var(perf_cpu_context);
3292         perf_event_task_ctx(&cpuctx->ctx, task_event);
3293         put_cpu_var(perf_cpu_context);
3294
3295         if (!ctx)
3296                 ctx = rcu_dereference(task_event->task->perf_event_ctxp);
3297         if (ctx)
3298                 perf_event_task_ctx(ctx, task_event);
3299         rcu_read_unlock();
3300 }
3301
3302 static void perf_event_task(struct task_struct *task,
3303                               struct perf_event_context *task_ctx,
3304                               int new)
3305 {
3306         struct perf_task_event task_event;
3307
3308         if (!atomic_read(&nr_comm_events) &&
3309             !atomic_read(&nr_mmap_events) &&
3310             !atomic_read(&nr_task_events))
3311                 return;
3312
3313         task_event = (struct perf_task_event){
3314                 .task     = task,
3315                 .task_ctx = task_ctx,
3316                 .event_id    = {
3317                         .header = {
3318                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3319                                 .misc = 0,
3320                                 .size = sizeof(task_event.event_id),
3321                         },
3322                         /* .pid  */
3323                         /* .ppid */
3324                         /* .tid  */
3325                         /* .ptid */
3326                 },
3327         };
3328
3329         perf_event_task_event(&task_event);
3330 }
3331
3332 void perf_event_fork(struct task_struct *task)
3333 {
3334         perf_event_task(task, NULL, 1);
3335 }
3336
3337 /*
3338  * comm tracking
3339  */
3340
3341 struct perf_comm_event {
3342         struct task_struct      *task;
3343         char                    *comm;
3344         int                     comm_size;
3345
3346         struct {
3347                 struct perf_event_header        header;
3348
3349                 u32                             pid;
3350                 u32                             tid;
3351         } event_id;
3352 };
3353
3354 static void perf_event_comm_output(struct perf_event *event,
3355                                      struct perf_comm_event *comm_event)
3356 {
3357         struct perf_output_handle handle;
3358         int size = comm_event->event_id.header.size;
3359         int ret = perf_output_begin(&handle, event, size, 0, 0);
3360
3361         if (ret)
3362                 return;
3363
3364         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3365         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3366
3367         perf_output_put(&handle, comm_event->event_id);
3368         perf_output_copy(&handle, comm_event->comm,
3369                                    comm_event->comm_size);
3370         perf_output_end(&handle);
3371 }
3372
3373 static int perf_event_comm_match(struct perf_event *event)
3374 {
3375         if (event->attr.comm)
3376                 return 1;
3377
3378         return 0;
3379 }
3380
3381 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3382                                   struct perf_comm_event *comm_event)
3383 {
3384         struct perf_event *event;
3385
3386         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3387                 if (perf_event_comm_match(event))
3388                         perf_event_comm_output(event, comm_event);
3389         }
3390 }
3391
3392 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3393 {
3394         struct perf_cpu_context *cpuctx;
3395         struct perf_event_context *ctx;
3396         unsigned int size;
3397         char comm[TASK_COMM_LEN];
3398
3399         memset(comm, 0, sizeof(comm));
3400         strlcpy(comm, comm_event->task->comm, sizeof(comm));
3401         size = ALIGN(strlen(comm)+1, sizeof(u64));
3402
3403         comm_event->comm = comm;
3404         comm_event->comm_size = size;
3405
3406         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3407
3408         rcu_read_lock();
3409         cpuctx = &get_cpu_var(perf_cpu_context);
3410         perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3411         put_cpu_var(perf_cpu_context);
3412
3413         /*
3414          * doesn't really matter which of the child contexts the
3415          * events ends up in.
3416          */
3417         ctx = rcu_dereference(current->perf_event_ctxp);
3418         if (ctx)
3419                 perf_event_comm_ctx(ctx, comm_event);
3420         rcu_read_unlock();
3421 }
3422
3423 void perf_event_comm(struct task_struct *task)
3424 {
3425         struct perf_comm_event comm_event;
3426
3427         if (task->perf_event_ctxp)
3428                 perf_event_enable_on_exec(task);
3429
3430         if (!atomic_read(&nr_comm_events))
3431                 return;
3432
3433         comm_event = (struct perf_comm_event){
3434                 .task   = task,
3435                 /* .comm      */
3436                 /* .comm_size */
3437                 .event_id  = {
3438                         .header = {
3439                                 .type = PERF_RECORD_COMM,
3440                                 .misc = 0,
3441                                 /* .size */
3442                         },
3443                         /* .pid */
3444                         /* .tid */
3445                 },
3446         };
3447
3448         perf_event_comm_event(&comm_event);
3449 }
3450
3451 /*
3452  * mmap tracking
3453  */
3454
3455 struct perf_mmap_event {
3456         struct vm_area_struct   *vma;
3457
3458         const char              *file_name;
3459         int                     file_size;
3460
3461         struct {
3462                 struct perf_event_header        header;
3463
3464                 u32                             pid;
3465                 u32                             tid;
3466                 u64                             start;
3467                 u64                             len;
3468                 u64                             pgoff;
3469         } event_id;
3470 };
3471
3472 static void perf_event_mmap_output(struct perf_event *event,
3473                                      struct perf_mmap_event *mmap_event)
3474 {
3475         struct perf_output_handle handle;
3476         int size = mmap_event->event_id.header.size;
3477         int ret = perf_output_begin(&handle, event, size, 0, 0);
3478
3479         if (ret)
3480                 return;
3481
3482         mmap_event->event_id.pid = perf_event_pid(event, current);
3483         mmap_event->event_id.tid = perf_event_tid(event, current);
3484
3485         perf_output_put(&handle, mmap_event->event_id);
3486         perf_output_copy(&handle, mmap_event->file_name,
3487                                    mmap_event->file_size);
3488         perf_output_end(&handle);
3489 }
3490
3491 static int perf_event_mmap_match(struct perf_event *event,
3492                                    struct perf_mmap_event *mmap_event)
3493 {
3494         if (event->attr.mmap)
3495                 return 1;
3496
3497         return 0;
3498 }
3499
3500 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3501                                   struct perf_mmap_event *mmap_event)
3502 {
3503         struct perf_event *event;
3504
3505         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3506                 if (perf_event_mmap_match(event, mmap_event))
3507                         perf_event_mmap_output(event, mmap_event);
3508         }
3509 }
3510
3511 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3512 {
3513         struct perf_cpu_context *cpuctx;
3514         struct perf_event_context *ctx;
3515         struct vm_area_struct *vma = mmap_event->vma;
3516         struct file *file = vma->vm_file;
3517         unsigned int size;
3518         char tmp[16];
3519         char *buf = NULL;
3520         const char *name;
3521
3522         memset(tmp, 0, sizeof(tmp));
3523
3524         if (file) {
3525                 /*
3526                  * d_path works from the end of the buffer backwards, so we
3527                  * need to add enough zero bytes after the string to handle
3528                  * the 64bit alignment we do later.
3529                  */
3530                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3531                 if (!buf) {
3532                         name = strncpy(tmp, "//enomem", sizeof(tmp));
3533                         goto got_name;
3534                 }
3535                 name = d_path(&file->f_path, buf, PATH_MAX);
3536                 if (IS_ERR(name)) {
3537                         name = strncpy(tmp, "//toolong", sizeof(tmp));
3538                         goto got_name;
3539                 }
3540         } else {
3541                 if (arch_vma_name(mmap_event->vma)) {
3542                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3543                                        sizeof(tmp));
3544                         goto got_name;
3545                 }
3546
3547                 if (!vma->vm_mm) {
3548                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
3549                         goto got_name;
3550                 }
3551
3552                 name = strncpy(tmp, "//anon", sizeof(tmp));
3553                 goto got_name;
3554         }
3555
3556 got_name:
3557         size = ALIGN(strlen(name)+1, sizeof(u64));
3558
3559         mmap_event->file_name = name;
3560         mmap_event->file_size = size;
3561
3562         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3563
3564         rcu_read_lock();
3565         cpuctx = &get_cpu_var(perf_cpu_context);
3566         perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3567         put_cpu_var(perf_cpu_context);
3568
3569         /*
3570          * doesn't really matter which of the child contexts the
3571          * events ends up in.
3572          */
3573         ctx = rcu_dereference(current->perf_event_ctxp);
3574         if (ctx)
3575                 perf_event_mmap_ctx(ctx, mmap_event);
3576         rcu_read_unlock();
3577
3578         kfree(buf);
3579 }
3580
3581 void __perf_event_mmap(struct vm_area_struct *vma)
3582 {
3583         struct perf_mmap_event mmap_event;
3584
3585         if (!atomic_read(&nr_mmap_events))
3586                 return;
3587
3588         mmap_event = (struct perf_mmap_event){
3589                 .vma    = vma,
3590                 /* .file_name */
3591                 /* .file_size */
3592                 .event_id  = {
3593                         .header = {
3594                                 .type = PERF_RECORD_MMAP,
3595                                 .misc = 0,
3596                                 /* .size */
3597                         },
3598                         /* .pid */
3599                         /* .tid */
3600                         .start  = vma->vm_start,
3601                         .len    = vma->vm_end - vma->vm_start,
3602                         .pgoff  = vma->vm_pgoff,
3603                 },
3604         };
3605
3606         perf_event_mmap_event(&mmap_event);
3607 }
3608
3609 /*
3610  * IRQ throttle logging
3611  */
3612
3613 static void perf_log_throttle(struct perf_event *event, int enable)
3614 {
3615         struct perf_output_handle handle;
3616         int ret;
3617
3618         struct {
3619                 struct perf_event_header        header;
3620                 u64                             time;
3621                 u64                             id;
3622                 u64                             stream_id;
3623         } throttle_event = {
3624                 .header = {
3625                         .type = PERF_RECORD_THROTTLE,
3626                         .misc = 0,
3627                         .size = sizeof(throttle_event),
3628                 },
3629                 .time           = perf_clock(),
3630                 .id             = primary_event_id(event),
3631                 .stream_id      = event->id,
3632         };
3633
3634         if (enable)
3635                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3636
3637         ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3638         if (ret)
3639                 return;
3640
3641         perf_output_put(&handle, throttle_event);
3642         perf_output_end(&handle);
3643 }
3644
3645 /*
3646  * Generic event overflow handling, sampling.
3647  */
3648
3649 static int __perf_event_overflow(struct perf_event *event, int nmi,
3650                                    int throttle, struct perf_sample_data *data,
3651                                    struct pt_regs *regs)
3652 {
3653         int events = atomic_read(&event->event_limit);
3654         struct hw_perf_event *hwc = &event->hw;
3655         int ret = 0;
3656
3657         throttle = (throttle && event->pmu->unthrottle != NULL);
3658
3659         if (!throttle) {
3660                 hwc->interrupts++;
3661         } else {
3662                 if (hwc->interrupts != MAX_INTERRUPTS) {
3663                         hwc->interrupts++;
3664                         if (HZ * hwc->interrupts >
3665                                         (u64)sysctl_perf_event_sample_rate) {
3666                                 hwc->interrupts = MAX_INTERRUPTS;
3667                                 perf_log_throttle(event, 0);
3668                                 ret = 1;
3669                         }
3670                 } else {
3671                         /*
3672                          * Keep re-disabling events even though on the previous
3673                          * pass we disabled it - just in case we raced with a
3674                          * sched-in and the event got enabled again:
3675                          */
3676                         ret = 1;
3677                 }
3678         }
3679
3680         if (event->attr.freq) {
3681                 u64 now = perf_clock();
3682                 s64 delta = now - hwc->freq_stamp;
3683
3684                 hwc->freq_stamp = now;
3685
3686                 if (delta > 0 && delta < TICK_NSEC)
3687                         perf_adjust_period(event, NSEC_PER_SEC / (int)delta);
3688         }
3689
3690         /*
3691          * XXX event_limit might not quite work as expected on inherited
3692          * events
3693          */
3694
3695         event->pending_kill = POLL_IN;
3696         if (events && atomic_dec_and_test(&event->event_limit)) {
3697                 ret = 1;
3698                 event->pending_kill = POLL_HUP;
3699                 if (nmi) {
3700                         event->pending_disable = 1;
3701                         perf_pending_queue(&event->pending,
3702                                            perf_pending_event);
3703                 } else
3704                         perf_event_disable(event);
3705         }
3706
3707         if (event->overflow_handler)
3708                 event->overflow_handler(event, nmi, data, regs);
3709         else
3710                 perf_event_output(event, nmi, data, regs);
3711
3712         return ret;
3713 }
3714
3715 int perf_event_overflow(struct perf_event *event, int nmi,
3716                           struct perf_sample_data *data,
3717                           struct pt_regs *regs)
3718 {
3719         return __perf_event_overflow(event, nmi, 1, data, regs);
3720 }
3721
3722 /*
3723  * Generic software event infrastructure
3724  */
3725
3726 /*
3727  * We directly increment event->count and keep a second value in
3728  * event->hw.period_left to count intervals. This period event
3729  * is kept in the range [-sample_period, 0] so that we can use the
3730  * sign as trigger.
3731  */
3732
3733 static u64 perf_swevent_set_period(struct perf_event *event)
3734 {
3735         struct hw_perf_event *hwc = &event->hw;
3736         u64 period = hwc->last_period;
3737         u64 nr, offset;
3738         s64 old, val;
3739
3740         hwc->last_period = hwc->sample_period;
3741
3742 again:
3743         old = val = atomic64_read(&hwc->period_left);
3744         if (val < 0)
3745                 return 0;
3746
3747         nr = div64_u64(period + val, period);
3748         offset = nr * period;
3749         val -= offset;
3750         if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3751                 goto again;
3752
3753         return nr;
3754 }
3755
3756 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3757                                     int nmi, struct perf_sample_data *data,
3758                                     struct pt_regs *regs)
3759 {
3760         struct hw_perf_event *hwc = &event->hw;
3761         int throttle = 0;
3762
3763         data->period = event->hw.last_period;
3764         if (!overflow)
3765                 overflow = perf_swevent_set_period(event);
3766
3767         if (hwc->interrupts == MAX_INTERRUPTS)
3768                 return;
3769
3770         for (; overflow; overflow--) {
3771                 if (__perf_event_overflow(event, nmi, throttle,
3772                                             data, regs)) {
3773                         /*
3774                          * We inhibit the overflow from happening when
3775                          * hwc->interrupts == MAX_INTERRUPTS.
3776                          */
3777                         break;
3778                 }
3779                 throttle = 1;
3780         }
3781 }
3782
3783 static void perf_swevent_unthrottle(struct perf_event *event)
3784 {
3785         /*
3786          * Nothing to do, we already reset hwc->interrupts.
3787          */
3788 }
3789
3790 static void perf_swevent_add(struct perf_event *event, u64 nr,
3791                                int nmi, struct perf_sample_data *data,
3792                                struct pt_regs *regs)
3793 {
3794         struct hw_perf_event *hwc = &event->hw;
3795
3796         atomic64_add(nr, &event->count);
3797
3798         if (!regs)
3799                 return;
3800
3801         if (!hwc->sample_period)
3802                 return;
3803
3804         if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3805                 return perf_swevent_overflow(event, 1, nmi, data, regs);
3806
3807         if (atomic64_add_negative(nr, &hwc->period_left))
3808                 return;
3809
3810         perf_swevent_overflow(event, 0, nmi, data, regs);
3811 }
3812
3813 static int perf_swevent_is_counting(struct perf_event *event)
3814 {
3815         /*
3816          * The event is active, we're good!
3817          */
3818         if (event->state == PERF_EVENT_STATE_ACTIVE)
3819                 return 1;
3820
3821         /*
3822          * The event is off/error, not counting.
3823          */
3824         if (event->state != PERF_EVENT_STATE_INACTIVE)
3825                 return 0;
3826
3827         /*
3828          * The event is inactive, if the context is active
3829          * we're part of a group that didn't make it on the 'pmu',
3830          * not counting.
3831          */
3832         if (event->ctx->is_active)
3833                 return 0;
3834
3835         /*
3836          * We're inactive and the context is too, this means the
3837          * task is scheduled out, we're counting events that happen
3838          * to us, like migration events.
3839          */
3840         return 1;
3841 }
3842
3843 static int perf_tp_event_match(struct perf_event *event,
3844                                 struct perf_sample_data *data);
3845
3846 static int perf_exclude_event(struct perf_event *event,
3847                               struct pt_regs *regs)
3848 {
3849         if (regs) {
3850                 if (event->attr.exclude_user && user_mode(regs))
3851                         return 1;
3852
3853                 if (event->attr.exclude_kernel && !user_mode(regs))
3854                         return 1;
3855         }
3856
3857         return 0;
3858 }
3859
3860 static int perf_swevent_match(struct perf_event *event,
3861                                 enum perf_type_id type,
3862                                 u32 event_id,
3863                                 struct perf_sample_data *data,
3864                                 struct pt_regs *regs)
3865 {
3866         if (!perf_swevent_is_counting(event))
3867                 return 0;
3868
3869         if (event->attr.type != type)
3870                 return 0;
3871
3872         if (event->attr.config != event_id)
3873                 return 0;
3874
3875         if (perf_exclude_event(event, regs))
3876                 return 0;
3877
3878         if (event->attr.type == PERF_TYPE_TRACEPOINT &&
3879             !perf_tp_event_match(event, data))
3880                 return 0;
3881
3882         return 1;
3883 }
3884
3885 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
3886                                      enum perf_type_id type,
3887                                      u32 event_id, u64 nr, int nmi,
3888                                      struct perf_sample_data *data,
3889                                      struct pt_regs *regs)
3890 {
3891         struct perf_event *event;
3892
3893         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3894                 if (perf_swevent_match(event, type, event_id, data, regs))
3895                         perf_swevent_add(event, nr, nmi, data, regs);
3896         }
3897 }
3898
3899 int perf_swevent_get_recursion_context(void)
3900 {
3901         struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3902         int rctx;
3903
3904         if (in_nmi())
3905                 rctx = 3;
3906         else if (in_irq())
3907                 rctx = 2;
3908         else if (in_softirq())
3909                 rctx = 1;
3910         else
3911                 rctx = 0;
3912
3913         if (cpuctx->recursion[rctx]) {
3914                 put_cpu_var(perf_cpu_context);
3915                 return -1;
3916         }
3917
3918         cpuctx->recursion[rctx]++;
3919         barrier();
3920
3921         return rctx;
3922 }
3923 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
3924
3925 void perf_swevent_put_recursion_context(int rctx)
3926 {
3927         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
3928         barrier();
3929         cpuctx->recursion[rctx]--;
3930         put_cpu_var(perf_cpu_context);
3931 }
3932 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
3933
3934 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
3935                                     u64 nr, int nmi,
3936                                     struct perf_sample_data *data,
3937                                     struct pt_regs *regs)
3938 {
3939         struct perf_cpu_context *cpuctx;
3940         struct perf_event_context *ctx;
3941
3942         cpuctx = &__get_cpu_var(perf_cpu_context);
3943         rcu_read_lock();
3944         perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
3945                                  nr, nmi, data, regs);
3946         /*
3947          * doesn't really matter which of the child contexts the
3948          * events ends up in.
3949          */
3950         ctx = rcu_dereference(current->perf_event_ctxp);
3951         if (ctx)
3952                 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
3953         rcu_read_unlock();
3954 }
3955
3956 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
3957                             struct pt_regs *regs, u64 addr)
3958 {
3959         struct perf_sample_data data;
3960         int rctx;
3961
3962         rctx = perf_swevent_get_recursion_context();
3963         if (rctx < 0)
3964                 return;
3965
3966         data.addr = addr;
3967         data.raw  = NULL;
3968
3969         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
3970
3971         perf_swevent_put_recursion_context(rctx);
3972 }
3973
3974 static void perf_swevent_read(struct perf_event *event)
3975 {
3976 }
3977
3978 static int perf_swevent_enable(struct perf_event *event)
3979 {
3980         struct hw_perf_event *hwc = &event->hw;
3981
3982         if (hwc->sample_period) {
3983                 hwc->last_period = hwc->sample_period;
3984                 perf_swevent_set_period(event);
3985         }
3986         return 0;
3987 }
3988
3989 static void perf_swevent_disable(struct perf_event *event)
3990 {
3991 }
3992
3993 static const struct pmu perf_ops_generic = {
3994         .enable         = perf_swevent_enable,
3995         .disable        = perf_swevent_disable,
3996         .read           = perf_swevent_read,
3997         .unthrottle     = perf_swevent_unthrottle,
3998 };
3999
4000 /*
4001  * hrtimer based swevent callback
4002  */
4003
4004 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4005 {
4006         enum hrtimer_restart ret = HRTIMER_RESTART;
4007         struct perf_sample_data data;
4008         struct pt_regs *regs;
4009         struct perf_event *event;
4010         u64 period;
4011
4012         event   = container_of(hrtimer, struct perf_event, hw.hrtimer);
4013         event->pmu->read(event);
4014
4015         data.addr = 0;
4016         data.raw = NULL;
4017         data.period = event->hw.last_period;
4018         regs = get_irq_regs();
4019         /*
4020          * In case we exclude kernel IPs or are somehow not in interrupt
4021          * context, provide the next best thing, the user IP.
4022          */
4023         if ((event->attr.exclude_kernel || !regs) &&
4024                         !event->attr.exclude_user)
4025                 regs = task_pt_regs(current);
4026
4027         if (regs) {
4028                 if (!(event->attr.exclude_idle && current->pid == 0))
4029                         if (perf_event_overflow(event, 0, &data, regs))
4030                                 ret = HRTIMER_NORESTART;
4031         }
4032
4033         period = max_t(u64, 10000, event->hw.sample_period);
4034         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4035
4036         return ret;
4037 }
4038
4039 static void perf_swevent_start_hrtimer(struct perf_event *event)
4040 {
4041         struct hw_perf_event *hwc = &event->hw;
4042
4043         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4044         hwc->hrtimer.function = perf_swevent_hrtimer;
4045         if (hwc->sample_period) {
4046                 u64 period;
4047
4048                 if (hwc->remaining) {
4049                         if (hwc->remaining < 0)
4050                                 period = 10000;
4051                         else
4052                                 period = hwc->remaining;
4053                         hwc->remaining = 0;
4054                 } else {
4055                         period = max_t(u64, 10000, hwc->sample_period);
4056                 }
4057                 __hrtimer_start_range_ns(&hwc->hrtimer,
4058                                 ns_to_ktime(period), 0,
4059                                 HRTIMER_MODE_REL, 0);
4060         }
4061 }
4062
4063 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4064 {
4065         struct hw_perf_event *hwc = &event->hw;
4066
4067         if (hwc->sample_period) {
4068                 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4069                 hwc->remaining = ktime_to_ns(remaining);
4070
4071                 hrtimer_cancel(&hwc->hrtimer);
4072         }
4073 }
4074
4075 /*
4076  * Software event: cpu wall time clock
4077  */
4078
4079 static void cpu_clock_perf_event_update(struct perf_event *event)
4080 {
4081         int cpu = raw_smp_processor_id();
4082         s64 prev;
4083         u64 now;
4084
4085         now = cpu_clock(cpu);
4086         prev = atomic64_xchg(&event->hw.prev_count, now);
4087         atomic64_add(now - prev, &event->count);
4088 }
4089
4090 static int cpu_clock_perf_event_enable(struct perf_event *event)
4091 {
4092         struct hw_perf_event *hwc = &event->hw;
4093         int cpu = raw_smp_processor_id();
4094
4095         atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4096         perf_swevent_start_hrtimer(event);
4097
4098         return 0;
4099 }
4100
4101 static void cpu_clock_perf_event_disable(struct perf_event *event)
4102 {
4103         perf_swevent_cancel_hrtimer(event);
4104         cpu_clock_perf_event_update(event);
4105 }
4106
4107 static void cpu_clock_perf_event_read(struct perf_event *event)
4108 {
4109         cpu_clock_perf_event_update(event);
4110 }
4111
4112 static const struct pmu perf_ops_cpu_clock = {
4113         .enable         = cpu_clock_perf_event_enable,
4114         .disable        = cpu_clock_perf_event_disable,
4115         .read           = cpu_clock_perf_event_read,
4116 };
4117
4118 /*
4119  * Software event: task time clock
4120  */
4121
4122 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4123 {
4124         u64 prev;
4125         s64 delta;
4126
4127         prev = atomic64_xchg(&event->hw.prev_count, now);
4128         delta = now - prev;
4129         atomic64_add(delta, &event->count);
4130 }
4131
4132 static int task_clock_perf_event_enable(struct perf_event *event)
4133 {
4134         struct hw_perf_event *hwc = &event->hw;
4135         u64 now;
4136
4137         now = event->ctx->time;
4138
4139         atomic64_set(&hwc->prev_count, now);
4140
4141         perf_swevent_start_hrtimer(event);
4142
4143         return 0;
4144 }
4145
4146 static void task_clock_perf_event_disable(struct perf_event *event)
4147 {
4148         perf_swevent_cancel_hrtimer(event);
4149         task_clock_perf_event_update(event, event->ctx->time);
4150
4151 }
4152
4153 static void task_clock_perf_event_read(struct perf_event *event)
4154 {
4155         u64 time;
4156
4157         if (!in_nmi()) {
4158                 update_context_time(event->ctx);
4159                 time = event->ctx->time;
4160         } else {
4161                 u64 now = perf_clock();
4162                 u64 delta = now - event->ctx->timestamp;
4163                 time = event->ctx->time + delta;
4164         }
4165
4166         task_clock_perf_event_update(event, time);
4167 }
4168
4169 static const struct pmu perf_ops_task_clock = {
4170         .enable         = task_clock_perf_event_enable,
4171         .disable        = task_clock_perf_event_disable,
4172         .read           = task_clock_perf_event_read,
4173 };
4174
4175 #ifdef CONFIG_EVENT_PROFILE
4176
4177 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4178                           int entry_size)
4179 {
4180         struct perf_raw_record raw = {
4181                 .size = entry_size,
4182                 .data = record,
4183         };
4184
4185         struct perf_sample_data data = {
4186                 .addr = addr,
4187                 .raw = &raw,
4188         };
4189
4190         struct pt_regs *regs = get_irq_regs();
4191
4192         if (!regs)
4193                 regs = task_pt_regs(current);
4194
4195         /* Trace events already protected against recursion */
4196         do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4197                                 &data, regs);
4198 }
4199 EXPORT_SYMBOL_GPL(perf_tp_event);
4200
4201 static int perf_tp_event_match(struct perf_event *event,
4202                                 struct perf_sample_data *data)
4203 {
4204         void *record = data->raw->data;
4205
4206         if (likely(!event->filter) || filter_match_preds(event->filter, record))
4207                 return 1;
4208         return 0;
4209 }
4210
4211 static void tp_perf_event_destroy(struct perf_event *event)
4212 {
4213         ftrace_profile_disable(event->attr.config);
4214 }
4215
4216 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4217 {
4218         /*
4219          * Raw tracepoint data is a severe data leak, only allow root to
4220          * have these.
4221          */
4222         if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4223                         perf_paranoid_tracepoint_raw() &&
4224                         !capable(CAP_SYS_ADMIN))
4225                 return ERR_PTR(-EPERM);
4226
4227         if (ftrace_profile_enable(event->attr.config))
4228                 return NULL;
4229
4230         event->destroy = tp_perf_event_destroy;
4231
4232         return &perf_ops_generic;
4233 }
4234
4235 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4236 {
4237         char *filter_str;
4238         int ret;
4239
4240         if (event->attr.type != PERF_TYPE_TRACEPOINT)
4241                 return -EINVAL;
4242
4243         filter_str = strndup_user(arg, PAGE_SIZE);
4244         if (IS_ERR(filter_str))
4245                 return PTR_ERR(filter_str);
4246
4247         ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4248
4249         kfree(filter_str);
4250         return ret;
4251 }
4252
4253 static void perf_event_free_filter(struct perf_event *event)
4254 {
4255         ftrace_profile_free_filter(event);
4256 }
4257
4258 #else
4259
4260 static int perf_tp_event_match(struct perf_event *event,
4261                                 struct perf_sample_data *data)
4262 {
4263         return 1;
4264 }
4265
4266 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4267 {
4268         return NULL;
4269 }
4270
4271 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4272 {
4273         return -ENOENT;
4274 }
4275
4276 static void perf_event_free_filter(struct perf_event *event)
4277 {
4278 }
4279
4280 #endif /* CONFIG_EVENT_PROFILE */
4281
4282 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4283 static void bp_perf_event_destroy(struct perf_event *event)
4284 {
4285         release_bp_slot(event);
4286 }
4287
4288 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4289 {
4290         int err;
4291
4292         err = register_perf_hw_breakpoint(bp);
4293         if (err)
4294                 return ERR_PTR(err);
4295
4296         bp->destroy = bp_perf_event_destroy;
4297
4298         return &perf_ops_bp;
4299 }
4300
4301 void perf_bp_event(struct perf_event *bp, void *data)
4302 {
4303         struct perf_sample_data sample;
4304         struct pt_regs *regs = data;
4305
4306         sample.raw = NULL;
4307         sample.addr = bp->attr.bp_addr;
4308
4309         if (!perf_exclude_event(bp, regs))
4310                 perf_swevent_add(bp, 1, 1, &sample, regs);
4311 }
4312 #else
4313 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4314 {
4315         return NULL;
4316 }
4317
4318 void perf_bp_event(struct perf_event *bp, void *regs)
4319 {
4320 }
4321 #endif
4322
4323 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4324
4325 static void sw_perf_event_destroy(struct perf_event *event)
4326 {
4327         u64 event_id = event->attr.config;
4328
4329         WARN_ON(event->parent);
4330
4331         atomic_dec(&perf_swevent_enabled[event_id]);
4332 }
4333
4334 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4335 {
4336         const struct pmu *pmu = NULL;
4337         u64 event_id = event->attr.config;
4338
4339         /*
4340          * Software events (currently) can't in general distinguish
4341          * between user, kernel and hypervisor events.
4342          * However, context switches and cpu migrations are considered
4343          * to be kernel events, and page faults are never hypervisor
4344          * events.
4345          */
4346         switch (event_id) {
4347         case PERF_COUNT_SW_CPU_CLOCK:
4348                 pmu = &perf_ops_cpu_clock;
4349
4350                 break;
4351         case PERF_COUNT_SW_TASK_CLOCK:
4352                 /*
4353                  * If the user instantiates this as a per-cpu event,
4354                  * use the cpu_clock event instead.
4355                  */
4356                 if (event->ctx->task)
4357                         pmu = &perf_ops_task_clock;
4358                 else
4359                         pmu = &perf_ops_cpu_clock;
4360
4361                 break;
4362         case PERF_COUNT_SW_PAGE_FAULTS:
4363         case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4364         case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4365         case PERF_COUNT_SW_CONTEXT_SWITCHES:
4366         case PERF_COUNT_SW_CPU_MIGRATIONS:
4367         case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4368         case PERF_COUNT_SW_EMULATION_FAULTS:
4369                 if (!event->parent) {
4370                         atomic_inc(&perf_swevent_enabled[event_id]);
4371                         event->destroy = sw_perf_event_destroy;
4372                 }
4373                 pmu = &perf_ops_generic;
4374                 break;
4375         }
4376
4377         return pmu;
4378 }
4379
4380 /*
4381  * Allocate and initialize a event structure
4382  */
4383 static struct perf_event *
4384 perf_event_alloc(struct perf_event_attr *attr,
4385                    int cpu,
4386                    struct perf_event_context *ctx,
4387                    struct perf_event *group_leader,
4388                    struct perf_event *parent_event,
4389                    perf_overflow_handler_t overflow_handler,
4390                    gfp_t gfpflags)
4391 {
4392         const struct pmu *pmu;
4393         struct perf_event *event;
4394         struct hw_perf_event *hwc;
4395         long err;
4396
4397         event = kzalloc(sizeof(*event), gfpflags);
4398         if (!event)
4399                 return ERR_PTR(-ENOMEM);
4400
4401         /*
4402          * Single events are their own group leaders, with an
4403          * empty sibling list:
4404          */
4405         if (!group_leader)
4406                 group_leader = event;
4407
4408         mutex_init(&event->child_mutex);
4409         INIT_LIST_HEAD(&event->child_list);
4410
4411         INIT_LIST_HEAD(&event->group_entry);
4412         INIT_LIST_HEAD(&event->event_entry);
4413         INIT_LIST_HEAD(&event->sibling_list);
4414         init_waitqueue_head(&event->waitq);
4415
4416         mutex_init(&event->mmap_mutex);
4417
4418         event->cpu              = cpu;
4419         event->attr             = *attr;
4420         event->group_leader     = group_leader;
4421         event->pmu              = NULL;
4422         event->ctx              = ctx;
4423         event->oncpu            = -1;
4424
4425         event->parent           = parent_event;
4426
4427         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
4428         event->id               = atomic64_inc_return(&perf_event_id);
4429
4430         event->state            = PERF_EVENT_STATE_INACTIVE;
4431
4432         if (!overflow_handler && parent_event)
4433                 overflow_handler = parent_event->overflow_handler;
4434         
4435         event->overflow_handler = overflow_handler;
4436
4437         if (attr->disabled)
4438                 event->state = PERF_EVENT_STATE_OFF;
4439
4440         pmu = NULL;
4441
4442         hwc = &event->hw;
4443         hwc->sample_period = attr->sample_period;
4444         if (attr->freq && attr->sample_freq)
4445                 hwc->sample_period = 1;
4446         hwc->last_period = hwc->sample_period;
4447
4448         atomic64_set(&hwc->period_left, hwc->sample_period);
4449
4450         /*
4451          * we currently do not support PERF_FORMAT_GROUP on inherited events
4452          */
4453         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4454                 goto done;
4455
4456         switch (attr->type) {
4457         case PERF_TYPE_RAW:
4458         case PERF_TYPE_HARDWARE:
4459         case PERF_TYPE_HW_CACHE:
4460                 pmu = hw_perf_event_init(event);
4461                 break;
4462
4463         case PERF_TYPE_SOFTWARE:
4464                 pmu = sw_perf_event_init(event);
4465                 break;
4466
4467         case PERF_TYPE_TRACEPOINT:
4468                 pmu = tp_perf_event_init(event);
4469                 break;
4470
4471         case PERF_TYPE_BREAKPOINT:
4472                 pmu = bp_perf_event_init(event);
4473                 break;
4474
4475
4476         default:
4477                 break;
4478         }
4479 done:
4480         err = 0;
4481         if (!pmu)
4482                 err = -EINVAL;
4483         else if (IS_ERR(pmu))
4484                 err = PTR_ERR(pmu);
4485
4486         if (err) {
4487                 if (event->ns)
4488                         put_pid_ns(event->ns);
4489                 kfree(event);
4490                 return ERR_PTR(err);
4491         }
4492
4493         event->pmu = pmu;
4494
4495         if (!event->parent) {
4496                 atomic_inc(&nr_events);
4497                 if (event->attr.mmap)
4498                         atomic_inc(&nr_mmap_events);
4499                 if (event->attr.comm)
4500                         atomic_inc(&nr_comm_events);
4501                 if (event->attr.task)
4502                         atomic_inc(&nr_task_events);
4503         }
4504
4505         return event;
4506 }
4507
4508 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4509                           struct perf_event_attr *attr)
4510 {
4511         u32 size;
4512         int ret;
4513
4514         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4515                 return -EFAULT;
4516
4517         /*
4518          * zero the full structure, so that a short copy will be nice.
4519          */
4520         memset(attr, 0, sizeof(*attr));
4521
4522         ret = get_user(size, &uattr->size);
4523         if (ret)
4524                 return ret;
4525
4526         if (size > PAGE_SIZE)   /* silly large */
4527                 goto err_size;
4528
4529         if (!size)              /* abi compat */
4530                 size = PERF_ATTR_SIZE_VER0;
4531
4532         if (size < PERF_ATTR_SIZE_VER0)
4533                 goto err_size;
4534
4535         /*
4536          * If we're handed a bigger struct than we know of,
4537          * ensure all the unknown bits are 0 - i.e. new
4538          * user-space does not rely on any kernel feature
4539          * extensions we dont know about yet.
4540          */
4541         if (size > sizeof(*attr)) {
4542                 unsigned char __user *addr;
4543                 unsigned char __user *end;
4544                 unsigned char val;
4545
4546                 addr = (void __user *)uattr + sizeof(*attr);
4547                 end  = (void __user *)uattr + size;
4548
4549                 for (; addr < end; addr++) {
4550                         ret = get_user(val, addr);
4551                         if (ret)
4552                                 return ret;
4553                         if (val)
4554                                 goto err_size;
4555                 }
4556                 size = sizeof(*attr);
4557         }
4558
4559         ret = copy_from_user(attr, uattr, size);
4560         if (ret)
4561                 return -EFAULT;
4562
4563         /*
4564          * If the type exists, the corresponding creation will verify
4565          * the attr->config.
4566          */
4567         if (attr->type >= PERF_TYPE_MAX)
4568                 return -EINVAL;
4569
4570         if (attr->__reserved_1 || attr->__reserved_2)
4571                 return -EINVAL;
4572
4573         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4574                 return -EINVAL;
4575
4576         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4577                 return -EINVAL;
4578
4579 out:
4580         return ret;
4581
4582 err_size:
4583         put_user(sizeof(*attr), &uattr->size);
4584         ret = -E2BIG;
4585         goto out;
4586 }
4587
4588 static int perf_event_set_output(struct perf_event *event, int output_fd)
4589 {
4590         struct perf_event *output_event = NULL;
4591         struct file *output_file = NULL;
4592         struct perf_event *old_output;
4593         int fput_needed = 0;
4594         int ret = -EINVAL;
4595
4596         if (!output_fd)
4597                 goto set;
4598
4599         output_file = fget_light(output_fd, &fput_needed);
4600         if (!output_file)
4601                 return -EBADF;
4602
4603         if (output_file->f_op != &perf_fops)
4604                 goto out;
4605
4606         output_event = output_file->private_data;
4607
4608         /* Don't chain output fds */
4609         if (output_event->output)
4610                 goto out;
4611
4612         /* Don't set an output fd when we already have an output channel */
4613         if (event->data)
4614                 goto out;
4615
4616         atomic_long_inc(&output_file->f_count);
4617
4618 set:
4619         mutex_lock(&event->mmap_mutex);
4620         old_output = event->output;
4621         rcu_assign_pointer(event->output, output_event);
4622         mutex_unlock(&event->mmap_mutex);
4623
4624         if (old_output) {
4625                 /*
4626                  * we need to make sure no existing perf_output_*()
4627                  * is still referencing this event.
4628                  */
4629                 synchronize_rcu();
4630                 fput(old_output->filp);
4631         }
4632
4633         ret = 0;
4634 out:
4635         fput_light(output_file, fput_needed);
4636         return ret;
4637 }
4638
4639 /**
4640  * sys_perf_event_open - open a performance event, associate it to a task/cpu
4641  *
4642  * @attr_uptr:  event_id type attributes for monitoring/sampling
4643  * @pid:                target pid
4644  * @cpu:                target cpu
4645  * @group_fd:           group leader event fd
4646  */
4647 SYSCALL_DEFINE5(perf_event_open,
4648                 struct perf_event_attr __user *, attr_uptr,
4649                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4650 {
4651         struct perf_event *event, *group_leader;
4652         struct perf_event_attr attr;
4653         struct perf_event_context *ctx;
4654         struct file *event_file = NULL;
4655         struct file *group_file = NULL;
4656         int fput_needed = 0;
4657         int fput_needed2 = 0;
4658         int err;
4659
4660         /* for future expandability... */
4661         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4662                 return -EINVAL;
4663
4664         err = perf_copy_attr(attr_uptr, &attr);
4665         if (err)
4666                 return err;
4667
4668         if (!attr.exclude_kernel) {
4669                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4670                         return -EACCES;
4671         }
4672
4673         if (attr.freq) {
4674                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4675                         return -EINVAL;
4676         }
4677
4678         /*
4679          * Get the target context (task or percpu):
4680          */
4681         ctx = find_get_context(pid, cpu);
4682         if (IS_ERR(ctx))
4683                 return PTR_ERR(ctx);
4684
4685         /*
4686          * Look up the group leader (we will attach this event to it):
4687          */
4688         group_leader = NULL;
4689         if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4690                 err = -EINVAL;
4691                 group_file = fget_light(group_fd, &fput_needed);
4692                 if (!group_file)
4693                         goto err_put_context;
4694                 if (group_file->f_op != &perf_fops)
4695                         goto err_put_context;
4696
4697                 group_leader = group_file->private_data;
4698                 /*
4699                  * Do not allow a recursive hierarchy (this new sibling
4700                  * becoming part of another group-sibling):
4701                  */
4702                 if (group_leader->group_leader != group_leader)
4703                         goto err_put_context;
4704                 /*
4705                  * Do not allow to attach to a group in a different
4706                  * task or CPU context:
4707                  */
4708                 if (group_leader->ctx != ctx)
4709                         goto err_put_context;
4710                 /*
4711                  * Only a group leader can be exclusive or pinned
4712                  */
4713                 if (attr.exclusive || attr.pinned)
4714                         goto err_put_context;
4715         }
4716
4717         event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4718                                      NULL, NULL, GFP_KERNEL);
4719         err = PTR_ERR(event);
4720         if (IS_ERR(event))
4721                 goto err_put_context;
4722
4723         err = anon_inode_getfd("[perf_event]", &perf_fops, event, 0);
4724         if (err < 0)
4725                 goto err_free_put_context;
4726
4727         event_file = fget_light(err, &fput_needed2);
4728         if (!event_file)
4729                 goto err_free_put_context;
4730
4731         if (flags & PERF_FLAG_FD_OUTPUT) {
4732                 err = perf_event_set_output(event, group_fd);
4733                 if (err)
4734                         goto err_fput_free_put_context;
4735         }
4736
4737         event->filp = event_file;
4738         WARN_ON_ONCE(ctx->parent_ctx);
4739         mutex_lock(&ctx->mutex);
4740         perf_install_in_context(ctx, event, cpu);
4741         ++ctx->generation;
4742         mutex_unlock(&ctx->mutex);
4743
4744         event->owner = current;
4745         get_task_struct(current);
4746         mutex_lock(&current->perf_event_mutex);
4747         list_add_tail(&event->owner_entry, &current->perf_event_list);
4748         mutex_unlock(&current->perf_event_mutex);
4749
4750 err_fput_free_put_context:
4751         fput_light(event_file, fput_needed2);
4752
4753 err_free_put_context:
4754         if (err < 0)
4755                 kfree(event);
4756
4757 err_put_context:
4758         if (err < 0)
4759                 put_ctx(ctx);
4760
4761         fput_light(group_file, fput_needed);
4762
4763         return err;
4764 }
4765
4766 /**
4767  * perf_event_create_kernel_counter
4768  *
4769  * @attr: attributes of the counter to create
4770  * @cpu: cpu in which the counter is bound
4771  * @pid: task to profile
4772  */
4773 struct perf_event *
4774 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4775                                  pid_t pid,
4776                                  perf_overflow_handler_t overflow_handler)
4777 {
4778         struct perf_event *event;
4779         struct perf_event_context *ctx;
4780         int err;
4781
4782         /*
4783          * Get the target context (task or percpu):
4784          */
4785
4786         ctx = find_get_context(pid, cpu);
4787         if (IS_ERR(ctx)) {
4788                 err = PTR_ERR(ctx);
4789                 goto err_exit;
4790         }
4791
4792         event = perf_event_alloc(attr, cpu, ctx, NULL,
4793                                  NULL, overflow_handler, GFP_KERNEL);
4794         if (IS_ERR(event)) {
4795                 err = PTR_ERR(event);
4796                 goto err_put_context;
4797         }
4798
4799         event->filp = NULL;
4800         WARN_ON_ONCE(ctx->parent_ctx);
4801         mutex_lock(&ctx->mutex);
4802         perf_install_in_context(ctx, event, cpu);
4803         ++ctx->generation;
4804         mutex_unlock(&ctx->mutex);
4805
4806         event->owner = current;
4807         get_task_struct(current);
4808         mutex_lock(&current->perf_event_mutex);
4809         list_add_tail(&event->owner_entry, &current->perf_event_list);
4810         mutex_unlock(&current->perf_event_mutex);
4811
4812         return event;
4813
4814  err_put_context:
4815         put_ctx(ctx);
4816  err_exit:
4817         return ERR_PTR(err);
4818 }
4819 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4820
4821 /*
4822  * inherit a event from parent task to child task:
4823  */
4824 static struct perf_event *
4825 inherit_event(struct perf_event *parent_event,
4826               struct task_struct *parent,
4827               struct perf_event_context *parent_ctx,
4828               struct task_struct *child,
4829               struct perf_event *group_leader,
4830               struct perf_event_context *child_ctx)
4831 {
4832         struct perf_event *child_event;
4833
4834         /*
4835          * Instead of creating recursive hierarchies of events,
4836          * we link inherited events back to the original parent,
4837          * which has a filp for sure, which we use as the reference
4838          * count:
4839          */
4840         if (parent_event->parent)
4841                 parent_event = parent_event->parent;
4842
4843         child_event = perf_event_alloc(&parent_event->attr,
4844                                            parent_event->cpu, child_ctx,
4845                                            group_leader, parent_event,
4846                                            NULL, GFP_KERNEL);
4847         if (IS_ERR(child_event))
4848                 return child_event;
4849         get_ctx(child_ctx);
4850
4851         /*
4852          * Make the child state follow the state of the parent event,
4853          * not its attr.disabled bit.  We hold the parent's mutex,
4854          * so we won't race with perf_event_{en, dis}able_family.
4855          */
4856         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
4857                 child_event->state = PERF_EVENT_STATE_INACTIVE;
4858         else
4859                 child_event->state = PERF_EVENT_STATE_OFF;
4860
4861         if (parent_event->attr.freq)
4862                 child_event->hw.sample_period = parent_event->hw.sample_period;
4863
4864         child_event->overflow_handler = parent_event->overflow_handler;
4865
4866         /*
4867          * Link it up in the child's context:
4868          */
4869         add_event_to_ctx(child_event, child_ctx);
4870
4871         /*
4872          * Get a reference to the parent filp - we will fput it
4873          * when the child event exits. This is safe to do because
4874          * we are in the parent and we know that the filp still
4875          * exists and has a nonzero count:
4876          */
4877         atomic_long_inc(&parent_event->filp->f_count);
4878
4879         /*
4880          * Link this into the parent event's child list
4881          */
4882         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4883         mutex_lock(&parent_event->child_mutex);
4884         list_add_tail(&child_event->child_list, &parent_event->child_list);
4885         mutex_unlock(&parent_event->child_mutex);
4886
4887         return child_event;
4888 }
4889
4890 static int inherit_group(struct perf_event *parent_event,
4891               struct task_struct *parent,
4892               struct perf_event_context *parent_ctx,
4893               struct task_struct *child,
4894               struct perf_event_context *child_ctx)
4895 {
4896         struct perf_event *leader;
4897         struct perf_event *sub;
4898         struct perf_event *child_ctr;
4899
4900         leader = inherit_event(parent_event, parent, parent_ctx,
4901                                  child, NULL, child_ctx);
4902         if (IS_ERR(leader))
4903                 return PTR_ERR(leader);
4904         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
4905                 child_ctr = inherit_event(sub, parent, parent_ctx,
4906                                             child, leader, child_ctx);
4907                 if (IS_ERR(child_ctr))
4908                         return PTR_ERR(child_ctr);
4909         }
4910         return 0;
4911 }
4912
4913 static void sync_child_event(struct perf_event *child_event,
4914                                struct task_struct *child)
4915 {
4916         struct perf_event *parent_event = child_event->parent;
4917         u64 child_val;
4918
4919         if (child_event->attr.inherit_stat)
4920                 perf_event_read_event(child_event, child);
4921
4922         child_val = atomic64_read(&child_event->count);
4923
4924         /*
4925          * Add back the child's count to the parent's count:
4926          */
4927         atomic64_add(child_val, &parent_event->count);
4928         atomic64_add(child_event->total_time_enabled,
4929                      &parent_event->child_total_time_enabled);
4930         atomic64_add(child_event->total_time_running,
4931                      &parent_event->child_total_time_running);
4932
4933         /*
4934          * Remove this event from the parent's list
4935          */
4936         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4937         mutex_lock(&parent_event->child_mutex);
4938         list_del_init(&child_event->child_list);
4939         mutex_unlock(&parent_event->child_mutex);
4940
4941         /*
4942          * Release the parent event, if this was the last
4943          * reference to it.
4944          */
4945         fput(parent_event->filp);
4946 }
4947
4948 static void
4949 __perf_event_exit_task(struct perf_event *child_event,
4950                          struct perf_event_context *child_ctx,
4951                          struct task_struct *child)
4952 {
4953         struct perf_event *parent_event;
4954
4955         perf_event_remove_from_context(child_event);
4956
4957         parent_event = child_event->parent;
4958         /*
4959          * It can happen that parent exits first, and has events
4960          * that are still around due to the child reference. These
4961          * events need to be zapped - but otherwise linger.
4962          */
4963         if (parent_event) {
4964                 sync_child_event(child_event, child);
4965                 free_event(child_event);
4966         }
4967 }
4968
4969 /*
4970  * When a child task exits, feed back event values to parent events.
4971  */
4972 void perf_event_exit_task(struct task_struct *child)
4973 {
4974         struct perf_event *child_event, *tmp;
4975         struct perf_event_context *child_ctx;
4976         unsigned long flags;
4977
4978         if (likely(!child->perf_event_ctxp)) {
4979                 perf_event_task(child, NULL, 0);
4980                 return;
4981         }
4982
4983         local_irq_save(flags);
4984         /*
4985          * We can't reschedule here because interrupts are disabled,
4986          * and either child is current or it is a task that can't be
4987          * scheduled, so we are now safe from rescheduling changing
4988          * our context.
4989          */
4990         child_ctx = child->perf_event_ctxp;
4991         __perf_event_task_sched_out(child_ctx);
4992
4993         /*
4994          * Take the context lock here so that if find_get_context is
4995          * reading child->perf_event_ctxp, we wait until it has
4996          * incremented the context's refcount before we do put_ctx below.
4997          */
4998         spin_lock(&child_ctx->lock);
4999         child->perf_event_ctxp = NULL;
5000         /*
5001          * If this context is a clone; unclone it so it can't get
5002          * swapped to another process while we're removing all
5003          * the events from it.
5004          */
5005         unclone_ctx(child_ctx);
5006         update_context_time(child_ctx);
5007         spin_unlock_irqrestore(&child_ctx->lock, flags);
5008
5009         /*
5010          * Report the task dead after unscheduling the events so that we
5011          * won't get any samples after PERF_RECORD_EXIT. We can however still
5012          * get a few PERF_RECORD_READ events.
5013          */
5014         perf_event_task(child, child_ctx, 0);
5015
5016         /*
5017          * We can recurse on the same lock type through:
5018          *
5019          *   __perf_event_exit_task()
5020          *     sync_child_event()
5021          *       fput(parent_event->filp)
5022          *         perf_release()
5023          *           mutex_lock(&ctx->mutex)
5024          *
5025          * But since its the parent context it won't be the same instance.
5026          */
5027         mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
5028
5029 again:
5030         list_for_each_entry_safe(child_event, tmp, &child_ctx->group_list,
5031                                  group_entry)
5032                 __perf_event_exit_task(child_event, child_ctx, child);
5033
5034         /*
5035          * If the last event was a group event, it will have appended all
5036          * its siblings to the list, but we obtained 'tmp' before that which
5037          * will still point to the list head terminating the iteration.
5038          */
5039         if (!list_empty(&child_ctx->group_list))
5040                 goto again;
5041
5042         mutex_unlock(&child_ctx->mutex);
5043
5044         put_ctx(child_ctx);
5045 }
5046
5047 /*
5048  * free an unexposed, unused context as created by inheritance by
5049  * init_task below, used by fork() in case of fail.
5050  */
5051 void perf_event_free_task(struct task_struct *task)
5052 {
5053         struct perf_event_context *ctx = task->perf_event_ctxp;
5054         struct perf_event *event, *tmp;
5055
5056         if (!ctx)
5057                 return;
5058
5059         mutex_lock(&ctx->mutex);
5060 again:
5061         list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry) {
5062                 struct perf_event *parent = event->parent;
5063
5064                 if (WARN_ON_ONCE(!parent))
5065                         continue;
5066
5067                 mutex_lock(&parent->child_mutex);
5068                 list_del_init(&event->child_list);
5069                 mutex_unlock(&parent->child_mutex);
5070
5071                 fput(parent->filp);
5072
5073                 list_del_event(event, ctx);
5074                 free_event(event);
5075         }
5076
5077         if (!list_empty(&ctx->group_list))
5078                 goto again;
5079
5080         mutex_unlock(&ctx->mutex);
5081
5082         put_ctx(ctx);
5083 }
5084
5085 /*
5086  * Initialize the perf_event context in task_struct
5087  */
5088 int perf_event_init_task(struct task_struct *child)
5089 {
5090         struct perf_event_context *child_ctx = NULL, *parent_ctx;
5091         struct perf_event_context *cloned_ctx;
5092         struct perf_event *event;
5093         struct task_struct *parent = current;
5094         int inherited_all = 1;
5095         int ret = 0;
5096
5097         child->perf_event_ctxp = NULL;
5098
5099         mutex_init(&child->perf_event_mutex);
5100         INIT_LIST_HEAD(&child->perf_event_list);
5101
5102         if (likely(!parent->perf_event_ctxp))
5103                 return 0;
5104
5105         /*
5106          * If the parent's context is a clone, pin it so it won't get
5107          * swapped under us.
5108          */
5109         parent_ctx = perf_pin_task_context(parent);
5110
5111         /*
5112          * No need to check if parent_ctx != NULL here; since we saw
5113          * it non-NULL earlier, the only reason for it to become NULL
5114          * is if we exit, and since we're currently in the middle of
5115          * a fork we can't be exiting at the same time.
5116          */
5117
5118         /*
5119          * Lock the parent list. No need to lock the child - not PID
5120          * hashed yet and not running, so nobody can access it.
5121          */
5122         mutex_lock(&parent_ctx->mutex);
5123
5124         /*
5125          * We dont have to disable NMIs - we are only looking at
5126          * the list, not manipulating it:
5127          */
5128         list_for_each_entry(event, &parent_ctx->group_list, group_entry) {
5129
5130                 if (!event->attr.inherit) {
5131                         inherited_all = 0;
5132                         continue;
5133                 }
5134
5135                 if (!child->perf_event_ctxp) {
5136                         /*
5137                          * This is executed from the parent task context, so
5138                          * inherit events that have been marked for cloning.
5139                          * First allocate and initialize a context for the
5140                          * child.
5141                          */
5142
5143                         child_ctx = kzalloc(sizeof(struct perf_event_context),
5144                                             GFP_KERNEL);
5145                         if (!child_ctx) {
5146                                 ret = -ENOMEM;
5147                                 goto exit;
5148                         }
5149
5150                         __perf_event_init_context(child_ctx, child);
5151                         child->perf_event_ctxp = child_ctx;
5152                         get_task_struct(child);
5153                 }
5154
5155                 ret = inherit_group(event, parent, parent_ctx,
5156                                              child, child_ctx);
5157                 if (ret) {
5158                         inherited_all = 0;
5159                         break;
5160                 }
5161         }
5162
5163         if (inherited_all) {
5164                 /*
5165                  * Mark the child context as a clone of the parent
5166                  * context, or of whatever the parent is a clone of.
5167                  * Note that if the parent is a clone, it could get
5168                  * uncloned at any point, but that doesn't matter
5169                  * because the list of events and the generation
5170                  * count can't have changed since we took the mutex.
5171                  */
5172                 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5173                 if (cloned_ctx) {
5174                         child_ctx->parent_ctx = cloned_ctx;
5175                         child_ctx->parent_gen = parent_ctx->parent_gen;
5176                 } else {
5177                         child_ctx->parent_ctx = parent_ctx;
5178                         child_ctx->parent_gen = parent_ctx->generation;
5179                 }
5180                 get_ctx(child_ctx->parent_ctx);
5181         }
5182
5183 exit:
5184         mutex_unlock(&parent_ctx->mutex);
5185
5186         perf_unpin_context(parent_ctx);
5187
5188         return ret;
5189 }
5190
5191 static void __cpuinit perf_event_init_cpu(int cpu)
5192 {
5193         struct perf_cpu_context *cpuctx;
5194
5195         cpuctx = &per_cpu(perf_cpu_context, cpu);
5196         __perf_event_init_context(&cpuctx->ctx, NULL);
5197
5198         spin_lock(&perf_resource_lock);
5199         cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5200         spin_unlock(&perf_resource_lock);
5201
5202         hw_perf_event_setup(cpu);
5203 }
5204
5205 #ifdef CONFIG_HOTPLUG_CPU
5206 static void __perf_event_exit_cpu(void *info)
5207 {
5208         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5209         struct perf_event_context *ctx = &cpuctx->ctx;
5210         struct perf_event *event, *tmp;
5211
5212         list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry)
5213                 __perf_event_remove_from_context(event);
5214 }
5215 static void perf_event_exit_cpu(int cpu)
5216 {
5217         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5218         struct perf_event_context *ctx = &cpuctx->ctx;
5219
5220         mutex_lock(&ctx->mutex);
5221         smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5222         mutex_unlock(&ctx->mutex);
5223 }
5224 #else
5225 static inline void perf_event_exit_cpu(int cpu) { }
5226 #endif
5227
5228 static int __cpuinit
5229 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5230 {
5231         unsigned int cpu = (long)hcpu;
5232
5233         switch (action) {
5234
5235         case CPU_UP_PREPARE:
5236         case CPU_UP_PREPARE_FROZEN:
5237                 perf_event_init_cpu(cpu);
5238                 break;
5239
5240         case CPU_ONLINE:
5241         case CPU_ONLINE_FROZEN:
5242                 hw_perf_event_setup_online(cpu);
5243                 break;
5244
5245         case CPU_DOWN_PREPARE:
5246         case CPU_DOWN_PREPARE_FROZEN:
5247                 perf_event_exit_cpu(cpu);
5248                 break;
5249
5250         default:
5251                 break;
5252         }
5253
5254         return NOTIFY_OK;
5255 }
5256
5257 /*
5258  * This has to have a higher priority than migration_notifier in sched.c.
5259  */
5260 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5261         .notifier_call          = perf_cpu_notify,
5262         .priority               = 20,
5263 };
5264
5265 void __init perf_event_init(void)
5266 {
5267         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5268                         (void *)(long)smp_processor_id());
5269         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5270                         (void *)(long)smp_processor_id());
5271         register_cpu_notifier(&perf_cpu_nb);
5272 }
5273
5274 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
5275 {
5276         return sprintf(buf, "%d\n", perf_reserved_percpu);
5277 }
5278
5279 static ssize_t
5280 perf_set_reserve_percpu(struct sysdev_class *class,
5281                         const char *buf,
5282                         size_t count)
5283 {
5284         struct perf_cpu_context *cpuctx;
5285         unsigned long val;
5286         int err, cpu, mpt;
5287
5288         err = strict_strtoul(buf, 10, &val);
5289         if (err)
5290                 return err;
5291         if (val > perf_max_events)
5292                 return -EINVAL;
5293
5294         spin_lock(&perf_resource_lock);
5295         perf_reserved_percpu = val;
5296         for_each_online_cpu(cpu) {
5297                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5298                 spin_lock_irq(&cpuctx->ctx.lock);
5299                 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5300                           perf_max_events - perf_reserved_percpu);
5301                 cpuctx->max_pertask = mpt;
5302                 spin_unlock_irq(&cpuctx->ctx.lock);
5303         }
5304         spin_unlock(&perf_resource_lock);
5305
5306         return count;
5307 }
5308
5309 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
5310 {
5311         return sprintf(buf, "%d\n", perf_overcommit);
5312 }
5313
5314 static ssize_t
5315 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
5316 {
5317         unsigned long val;
5318         int err;
5319
5320         err = strict_strtoul(buf, 10, &val);
5321         if (err)
5322                 return err;
5323         if (val > 1)
5324                 return -EINVAL;
5325
5326         spin_lock(&perf_resource_lock);
5327         perf_overcommit = val;
5328         spin_unlock(&perf_resource_lock);
5329
5330         return count;
5331 }
5332
5333 static SYSDEV_CLASS_ATTR(
5334                                 reserve_percpu,
5335                                 0644,
5336                                 perf_show_reserve_percpu,
5337                                 perf_set_reserve_percpu
5338                         );
5339
5340 static SYSDEV_CLASS_ATTR(
5341                                 overcommit,
5342                                 0644,
5343                                 perf_show_overcommit,
5344                                 perf_set_overcommit
5345                         );
5346
5347 static struct attribute *perfclass_attrs[] = {
5348         &attr_reserve_percpu.attr,
5349         &attr_overcommit.attr,
5350         NULL
5351 };
5352
5353 static struct attribute_group perfclass_attr_group = {
5354         .attrs                  = perfclass_attrs,
5355         .name                   = "perf_events",
5356 };
5357
5358 static int __init perf_event_sysfs_init(void)
5359 {
5360         return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5361                                   &perfclass_attr_group);
5362 }
5363 device_initcall(perf_event_sysfs_init);