Merge branch 'perf/urgent' into perf/core
[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                 raw_spin_lock_irqsave(&ctx->lock, *flags);
207                 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
208                         raw_spin_unlock_irqrestore(&ctx->lock, *flags);
209                         goto retry;
210                 }
211
212                 if (!atomic_inc_not_zero(&ctx->refcount)) {
213                         raw_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                 raw_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         raw_spin_lock_irqsave(&ctx->lock, flags);
244         --ctx->pin_count;
245         raw_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         raw_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         raw_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 successful.
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         raw_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                 raw_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         raw_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         raw_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         raw_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         raw_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                 raw_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         raw_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         raw_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         raw_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 successful.
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         raw_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                 raw_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         raw_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         raw_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         raw_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         raw_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         raw_spin_unlock_irq(&ctx->lock);
1010         task_oncpu_function_call(task, __perf_event_enable, event);
1011
1012         raw_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         raw_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         raw_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         raw_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)
1174 {
1175         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
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                 raw_spin_lock(&ctx->lock);
1203                 raw_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                 raw_spin_unlock(&next_ctx->lock);
1218                 raw_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)
1256 {
1257         int cpu = smp_processor_id();
1258         struct perf_event *event;
1259         int can_add_hw = 1;
1260
1261         raw_spin_lock(&ctx->lock);
1262         ctx->is_active = 1;
1263         if (likely(!ctx->nr_events))
1264                 goto out;
1265
1266         ctx->timestamp = perf_clock();
1267
1268         perf_disable();
1269
1270         /*
1271          * First go through the list and put on any pinned groups
1272          * in order to give them the best chance of going on.
1273          */
1274         list_for_each_entry(event, &ctx->group_list, group_entry) {
1275                 if (event->state <= PERF_EVENT_STATE_OFF ||
1276                     !event->attr.pinned)
1277                         continue;
1278                 if (event->cpu != -1 && event->cpu != cpu)
1279                         continue;
1280
1281                 if (group_can_go_on(event, cpuctx, 1))
1282                         group_sched_in(event, cpuctx, ctx, cpu);
1283
1284                 /*
1285                  * If this pinned group hasn't been scheduled,
1286                  * put it in error state.
1287                  */
1288                 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1289                         update_group_times(event);
1290                         event->state = PERF_EVENT_STATE_ERROR;
1291                 }
1292         }
1293
1294         list_for_each_entry(event, &ctx->group_list, group_entry) {
1295                 /*
1296                  * Ignore events in OFF or ERROR state, and
1297                  * ignore pinned events since we did them already.
1298                  */
1299                 if (event->state <= PERF_EVENT_STATE_OFF ||
1300                     event->attr.pinned)
1301                         continue;
1302
1303                 /*
1304                  * Listen to the 'cpu' scheduling filter constraint
1305                  * of events:
1306                  */
1307                 if (event->cpu != -1 && event->cpu != cpu)
1308                         continue;
1309
1310                 if (group_can_go_on(event, cpuctx, can_add_hw))
1311                         if (group_sched_in(event, cpuctx, ctx, cpu))
1312                                 can_add_hw = 0;
1313         }
1314         perf_enable();
1315  out:
1316         raw_spin_unlock(&ctx->lock);
1317 }
1318
1319 /*
1320  * Called from scheduler to add the events of the current task
1321  * with interrupts disabled.
1322  *
1323  * We restore the event value and then enable it.
1324  *
1325  * This does not protect us against NMI, but enable()
1326  * sets the enabled bit in the control field of event _before_
1327  * accessing the event control register. If a NMI hits, then it will
1328  * keep the event running.
1329  */
1330 void perf_event_task_sched_in(struct task_struct *task)
1331 {
1332         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1333         struct perf_event_context *ctx = task->perf_event_ctxp;
1334
1335         if (likely(!ctx))
1336                 return;
1337         if (cpuctx->task_ctx == ctx)
1338                 return;
1339         __perf_event_sched_in(ctx, cpuctx);
1340         cpuctx->task_ctx = ctx;
1341 }
1342
1343 static void perf_event_cpu_sched_in(struct perf_cpu_context *cpuctx)
1344 {
1345         struct perf_event_context *ctx = &cpuctx->ctx;
1346
1347         __perf_event_sched_in(ctx, cpuctx);
1348 }
1349
1350 #define MAX_INTERRUPTS (~0ULL)
1351
1352 static void perf_log_throttle(struct perf_event *event, int enable);
1353
1354 static void perf_adjust_period(struct perf_event *event, u64 events)
1355 {
1356         struct hw_perf_event *hwc = &event->hw;
1357         u64 period, sample_period;
1358         s64 delta;
1359
1360         events *= hwc->sample_period;
1361         period = div64_u64(events, event->attr.sample_freq);
1362
1363         delta = (s64)(period - hwc->sample_period);
1364         delta = (delta + 7) / 8; /* low pass filter */
1365
1366         sample_period = hwc->sample_period + delta;
1367
1368         if (!sample_period)
1369                 sample_period = 1;
1370
1371         hwc->sample_period = sample_period;
1372 }
1373
1374 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1375 {
1376         struct perf_event *event;
1377         struct hw_perf_event *hwc;
1378         u64 interrupts, freq;
1379
1380         raw_spin_lock(&ctx->lock);
1381         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1382                 if (event->state != PERF_EVENT_STATE_ACTIVE)
1383                         continue;
1384
1385                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1386                         continue;
1387
1388                 hwc = &event->hw;
1389
1390                 interrupts = hwc->interrupts;
1391                 hwc->interrupts = 0;
1392
1393                 /*
1394                  * unthrottle events on the tick
1395                  */
1396                 if (interrupts == MAX_INTERRUPTS) {
1397                         perf_log_throttle(event, 1);
1398                         event->pmu->unthrottle(event);
1399                         interrupts = 2*sysctl_perf_event_sample_rate/HZ;
1400                 }
1401
1402                 if (!event->attr.freq || !event->attr.sample_freq)
1403                         continue;
1404
1405                 /*
1406                  * if the specified freq < HZ then we need to skip ticks
1407                  */
1408                 if (event->attr.sample_freq < HZ) {
1409                         freq = event->attr.sample_freq;
1410
1411                         hwc->freq_count += freq;
1412                         hwc->freq_interrupts += interrupts;
1413
1414                         if (hwc->freq_count < HZ)
1415                                 continue;
1416
1417                         interrupts = hwc->freq_interrupts;
1418                         hwc->freq_interrupts = 0;
1419                         hwc->freq_count -= HZ;
1420                 } else
1421                         freq = HZ;
1422
1423                 perf_adjust_period(event, freq * interrupts);
1424
1425                 /*
1426                  * In order to avoid being stalled by an (accidental) huge
1427                  * sample period, force reset the sample period if we didn't
1428                  * get any events in this freq period.
1429                  */
1430                 if (!interrupts) {
1431                         perf_disable();
1432                         event->pmu->disable(event);
1433                         atomic64_set(&hwc->period_left, 0);
1434                         event->pmu->enable(event);
1435                         perf_enable();
1436                 }
1437         }
1438         raw_spin_unlock(&ctx->lock);
1439 }
1440
1441 /*
1442  * Round-robin a context's events:
1443  */
1444 static void rotate_ctx(struct perf_event_context *ctx)
1445 {
1446         struct perf_event *event;
1447
1448         if (!ctx->nr_events)
1449                 return;
1450
1451         raw_spin_lock(&ctx->lock);
1452         /*
1453          * Rotate the first entry last (works just fine for group events too):
1454          */
1455         perf_disable();
1456         list_for_each_entry(event, &ctx->group_list, group_entry) {
1457                 list_move_tail(&event->group_entry, &ctx->group_list);
1458                 break;
1459         }
1460         perf_enable();
1461
1462         raw_spin_unlock(&ctx->lock);
1463 }
1464
1465 void perf_event_task_tick(struct task_struct *curr)
1466 {
1467         struct perf_cpu_context *cpuctx;
1468         struct perf_event_context *ctx;
1469
1470         if (!atomic_read(&nr_events))
1471                 return;
1472
1473         cpuctx = &__get_cpu_var(perf_cpu_context);
1474         ctx = curr->perf_event_ctxp;
1475
1476         perf_ctx_adjust_freq(&cpuctx->ctx);
1477         if (ctx)
1478                 perf_ctx_adjust_freq(ctx);
1479
1480         perf_event_cpu_sched_out(cpuctx);
1481         if (ctx)
1482                 __perf_event_task_sched_out(ctx);
1483
1484         rotate_ctx(&cpuctx->ctx);
1485         if (ctx)
1486                 rotate_ctx(ctx);
1487
1488         perf_event_cpu_sched_in(cpuctx);
1489         if (ctx)
1490                 perf_event_task_sched_in(curr);
1491 }
1492
1493 /*
1494  * Enable all of a task's events that have been marked enable-on-exec.
1495  * This expects task == current.
1496  */
1497 static void perf_event_enable_on_exec(struct task_struct *task)
1498 {
1499         struct perf_event_context *ctx;
1500         struct perf_event *event;
1501         unsigned long flags;
1502         int enabled = 0;
1503
1504         local_irq_save(flags);
1505         ctx = task->perf_event_ctxp;
1506         if (!ctx || !ctx->nr_events)
1507                 goto out;
1508
1509         __perf_event_task_sched_out(ctx);
1510
1511         raw_spin_lock(&ctx->lock);
1512
1513         list_for_each_entry(event, &ctx->group_list, group_entry) {
1514                 if (!event->attr.enable_on_exec)
1515                         continue;
1516                 event->attr.enable_on_exec = 0;
1517                 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1518                         continue;
1519                 __perf_event_mark_enabled(event, ctx);
1520                 enabled = 1;
1521         }
1522
1523         /*
1524          * Unclone this context if we enabled any event.
1525          */
1526         if (enabled)
1527                 unclone_ctx(ctx);
1528
1529         raw_spin_unlock(&ctx->lock);
1530
1531         perf_event_task_sched_in(task);
1532  out:
1533         local_irq_restore(flags);
1534 }
1535
1536 /*
1537  * Cross CPU call to read the hardware event
1538  */
1539 static void __perf_event_read(void *info)
1540 {
1541         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1542         struct perf_event *event = info;
1543         struct perf_event_context *ctx = event->ctx;
1544
1545         /*
1546          * If this is a task context, we need to check whether it is
1547          * the current task context of this cpu.  If not it has been
1548          * scheduled out before the smp call arrived.  In that case
1549          * event->count would have been updated to a recent sample
1550          * when the event was scheduled out.
1551          */
1552         if (ctx->task && cpuctx->task_ctx != ctx)
1553                 return;
1554
1555         raw_spin_lock(&ctx->lock);
1556         update_context_time(ctx);
1557         update_event_times(event);
1558         raw_spin_unlock(&ctx->lock);
1559
1560         event->pmu->read(event);
1561 }
1562
1563 static u64 perf_event_read(struct perf_event *event)
1564 {
1565         /*
1566          * If event is enabled and currently active on a CPU, update the
1567          * value in the event structure:
1568          */
1569         if (event->state == PERF_EVENT_STATE_ACTIVE) {
1570                 smp_call_function_single(event->oncpu,
1571                                          __perf_event_read, event, 1);
1572         } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1573                 struct perf_event_context *ctx = event->ctx;
1574                 unsigned long flags;
1575
1576                 raw_spin_lock_irqsave(&ctx->lock, flags);
1577                 update_context_time(ctx);
1578                 update_event_times(event);
1579                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1580         }
1581
1582         return atomic64_read(&event->count);
1583 }
1584
1585 /*
1586  * Initialize the perf_event context in a task_struct:
1587  */
1588 static void
1589 __perf_event_init_context(struct perf_event_context *ctx,
1590                             struct task_struct *task)
1591 {
1592         raw_spin_lock_init(&ctx->lock);
1593         mutex_init(&ctx->mutex);
1594         INIT_LIST_HEAD(&ctx->group_list);
1595         INIT_LIST_HEAD(&ctx->event_list);
1596         atomic_set(&ctx->refcount, 1);
1597         ctx->task = task;
1598 }
1599
1600 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1601 {
1602         struct perf_event_context *ctx;
1603         struct perf_cpu_context *cpuctx;
1604         struct task_struct *task;
1605         unsigned long flags;
1606         int err;
1607
1608         if (pid == -1 && cpu != -1) {
1609                 /* Must be root to operate on a CPU event: */
1610                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1611                         return ERR_PTR(-EACCES);
1612
1613                 if (cpu < 0 || cpu >= nr_cpumask_bits)
1614                         return ERR_PTR(-EINVAL);
1615
1616                 /*
1617                  * We could be clever and allow to attach a event to an
1618                  * offline CPU and activate it when the CPU comes up, but
1619                  * that's for later.
1620                  */
1621                 if (!cpu_online(cpu))
1622                         return ERR_PTR(-ENODEV);
1623
1624                 cpuctx = &per_cpu(perf_cpu_context, cpu);
1625                 ctx = &cpuctx->ctx;
1626                 get_ctx(ctx);
1627
1628                 return ctx;
1629         }
1630
1631         rcu_read_lock();
1632         if (!pid)
1633                 task = current;
1634         else
1635                 task = find_task_by_vpid(pid);
1636         if (task)
1637                 get_task_struct(task);
1638         rcu_read_unlock();
1639
1640         if (!task)
1641                 return ERR_PTR(-ESRCH);
1642
1643         /*
1644          * Can't attach events to a dying task.
1645          */
1646         err = -ESRCH;
1647         if (task->flags & PF_EXITING)
1648                 goto errout;
1649
1650         /* Reuse ptrace permission checks for now. */
1651         err = -EACCES;
1652         if (!ptrace_may_access(task, PTRACE_MODE_READ))
1653                 goto errout;
1654
1655  retry:
1656         ctx = perf_lock_task_context(task, &flags);
1657         if (ctx) {
1658                 unclone_ctx(ctx);
1659                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1660         }
1661
1662         if (!ctx) {
1663                 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1664                 err = -ENOMEM;
1665                 if (!ctx)
1666                         goto errout;
1667                 __perf_event_init_context(ctx, task);
1668                 get_ctx(ctx);
1669                 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1670                         /*
1671                          * We raced with some other task; use
1672                          * the context they set.
1673                          */
1674                         kfree(ctx);
1675                         goto retry;
1676                 }
1677                 get_task_struct(task);
1678         }
1679
1680         put_task_struct(task);
1681         return ctx;
1682
1683  errout:
1684         put_task_struct(task);
1685         return ERR_PTR(err);
1686 }
1687
1688 static void perf_event_free_filter(struct perf_event *event);
1689
1690 static void free_event_rcu(struct rcu_head *head)
1691 {
1692         struct perf_event *event;
1693
1694         event = container_of(head, struct perf_event, rcu_head);
1695         if (event->ns)
1696                 put_pid_ns(event->ns);
1697         perf_event_free_filter(event);
1698         kfree(event);
1699 }
1700
1701 static void perf_pending_sync(struct perf_event *event);
1702
1703 static void free_event(struct perf_event *event)
1704 {
1705         perf_pending_sync(event);
1706
1707         if (!event->parent) {
1708                 atomic_dec(&nr_events);
1709                 if (event->attr.mmap)
1710                         atomic_dec(&nr_mmap_events);
1711                 if (event->attr.comm)
1712                         atomic_dec(&nr_comm_events);
1713                 if (event->attr.task)
1714                         atomic_dec(&nr_task_events);
1715         }
1716
1717         if (event->output) {
1718                 fput(event->output->filp);
1719                 event->output = NULL;
1720         }
1721
1722         if (event->destroy)
1723                 event->destroy(event);
1724
1725         put_ctx(event->ctx);
1726         call_rcu(&event->rcu_head, free_event_rcu);
1727 }
1728
1729 int perf_event_release_kernel(struct perf_event *event)
1730 {
1731         struct perf_event_context *ctx = event->ctx;
1732
1733         WARN_ON_ONCE(ctx->parent_ctx);
1734         mutex_lock(&ctx->mutex);
1735         perf_event_remove_from_context(event);
1736         mutex_unlock(&ctx->mutex);
1737
1738         mutex_lock(&event->owner->perf_event_mutex);
1739         list_del_init(&event->owner_entry);
1740         mutex_unlock(&event->owner->perf_event_mutex);
1741         put_task_struct(event->owner);
1742
1743         free_event(event);
1744
1745         return 0;
1746 }
1747 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1748
1749 /*
1750  * Called when the last reference to the file is gone.
1751  */
1752 static int perf_release(struct inode *inode, struct file *file)
1753 {
1754         struct perf_event *event = file->private_data;
1755
1756         file->private_data = NULL;
1757
1758         return perf_event_release_kernel(event);
1759 }
1760
1761 static int perf_event_read_size(struct perf_event *event)
1762 {
1763         int entry = sizeof(u64); /* value */
1764         int size = 0;
1765         int nr = 1;
1766
1767         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1768                 size += sizeof(u64);
1769
1770         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1771                 size += sizeof(u64);
1772
1773         if (event->attr.read_format & PERF_FORMAT_ID)
1774                 entry += sizeof(u64);
1775
1776         if (event->attr.read_format & PERF_FORMAT_GROUP) {
1777                 nr += event->group_leader->nr_siblings;
1778                 size += sizeof(u64);
1779         }
1780
1781         size += entry * nr;
1782
1783         return size;
1784 }
1785
1786 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1787 {
1788         struct perf_event *child;
1789         u64 total = 0;
1790
1791         *enabled = 0;
1792         *running = 0;
1793
1794         mutex_lock(&event->child_mutex);
1795         total += perf_event_read(event);
1796         *enabled += event->total_time_enabled +
1797                         atomic64_read(&event->child_total_time_enabled);
1798         *running += event->total_time_running +
1799                         atomic64_read(&event->child_total_time_running);
1800
1801         list_for_each_entry(child, &event->child_list, child_list) {
1802                 total += perf_event_read(child);
1803                 *enabled += child->total_time_enabled;
1804                 *running += child->total_time_running;
1805         }
1806         mutex_unlock(&event->child_mutex);
1807
1808         return total;
1809 }
1810 EXPORT_SYMBOL_GPL(perf_event_read_value);
1811
1812 static int perf_event_read_group(struct perf_event *event,
1813                                    u64 read_format, char __user *buf)
1814 {
1815         struct perf_event *leader = event->group_leader, *sub;
1816         int n = 0, size = 0, ret = -EFAULT;
1817         struct perf_event_context *ctx = leader->ctx;
1818         u64 values[5];
1819         u64 count, enabled, running;
1820
1821         mutex_lock(&ctx->mutex);
1822         count = perf_event_read_value(leader, &enabled, &running);
1823
1824         values[n++] = 1 + leader->nr_siblings;
1825         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1826                 values[n++] = enabled;
1827         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1828                 values[n++] = running;
1829         values[n++] = count;
1830         if (read_format & PERF_FORMAT_ID)
1831                 values[n++] = primary_event_id(leader);
1832
1833         size = n * sizeof(u64);
1834
1835         if (copy_to_user(buf, values, size))
1836                 goto unlock;
1837
1838         ret = size;
1839
1840         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1841                 n = 0;
1842
1843                 values[n++] = perf_event_read_value(sub, &enabled, &running);
1844                 if (read_format & PERF_FORMAT_ID)
1845                         values[n++] = primary_event_id(sub);
1846
1847                 size = n * sizeof(u64);
1848
1849                 if (copy_to_user(buf + ret, values, size)) {
1850                         ret = -EFAULT;
1851                         goto unlock;
1852                 }
1853
1854                 ret += size;
1855         }
1856 unlock:
1857         mutex_unlock(&ctx->mutex);
1858
1859         return ret;
1860 }
1861
1862 static int perf_event_read_one(struct perf_event *event,
1863                                  u64 read_format, char __user *buf)
1864 {
1865         u64 enabled, running;
1866         u64 values[4];
1867         int n = 0;
1868
1869         values[n++] = perf_event_read_value(event, &enabled, &running);
1870         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1871                 values[n++] = enabled;
1872         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1873                 values[n++] = running;
1874         if (read_format & PERF_FORMAT_ID)
1875                 values[n++] = primary_event_id(event);
1876
1877         if (copy_to_user(buf, values, n * sizeof(u64)))
1878                 return -EFAULT;
1879
1880         return n * sizeof(u64);
1881 }
1882
1883 /*
1884  * Read the performance event - simple non blocking version for now
1885  */
1886 static ssize_t
1887 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
1888 {
1889         u64 read_format = event->attr.read_format;
1890         int ret;
1891
1892         /*
1893          * Return end-of-file for a read on a event that is in
1894          * error state (i.e. because it was pinned but it couldn't be
1895          * scheduled on to the CPU at some point).
1896          */
1897         if (event->state == PERF_EVENT_STATE_ERROR)
1898                 return 0;
1899
1900         if (count < perf_event_read_size(event))
1901                 return -ENOSPC;
1902
1903         WARN_ON_ONCE(event->ctx->parent_ctx);
1904         if (read_format & PERF_FORMAT_GROUP)
1905                 ret = perf_event_read_group(event, read_format, buf);
1906         else
1907                 ret = perf_event_read_one(event, read_format, buf);
1908
1909         return ret;
1910 }
1911
1912 static ssize_t
1913 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1914 {
1915         struct perf_event *event = file->private_data;
1916
1917         return perf_read_hw(event, buf, count);
1918 }
1919
1920 static unsigned int perf_poll(struct file *file, poll_table *wait)
1921 {
1922         struct perf_event *event = file->private_data;
1923         struct perf_mmap_data *data;
1924         unsigned int events = POLL_HUP;
1925
1926         rcu_read_lock();
1927         data = rcu_dereference(event->data);
1928         if (data)
1929                 events = atomic_xchg(&data->poll, 0);
1930         rcu_read_unlock();
1931
1932         poll_wait(file, &event->waitq, wait);
1933
1934         return events;
1935 }
1936
1937 static void perf_event_reset(struct perf_event *event)
1938 {
1939         (void)perf_event_read(event);
1940         atomic64_set(&event->count, 0);
1941         perf_event_update_userpage(event);
1942 }
1943
1944 /*
1945  * Holding the top-level event's child_mutex means that any
1946  * descendant process that has inherited this event will block
1947  * in sync_child_event if it goes to exit, thus satisfying the
1948  * task existence requirements of perf_event_enable/disable.
1949  */
1950 static void perf_event_for_each_child(struct perf_event *event,
1951                                         void (*func)(struct perf_event *))
1952 {
1953         struct perf_event *child;
1954
1955         WARN_ON_ONCE(event->ctx->parent_ctx);
1956         mutex_lock(&event->child_mutex);
1957         func(event);
1958         list_for_each_entry(child, &event->child_list, child_list)
1959                 func(child);
1960         mutex_unlock(&event->child_mutex);
1961 }
1962
1963 static void perf_event_for_each(struct perf_event *event,
1964                                   void (*func)(struct perf_event *))
1965 {
1966         struct perf_event_context *ctx = event->ctx;
1967         struct perf_event *sibling;
1968
1969         WARN_ON_ONCE(ctx->parent_ctx);
1970         mutex_lock(&ctx->mutex);
1971         event = event->group_leader;
1972
1973         perf_event_for_each_child(event, func);
1974         func(event);
1975         list_for_each_entry(sibling, &event->sibling_list, group_entry)
1976                 perf_event_for_each_child(event, func);
1977         mutex_unlock(&ctx->mutex);
1978 }
1979
1980 static int perf_event_period(struct perf_event *event, u64 __user *arg)
1981 {
1982         struct perf_event_context *ctx = event->ctx;
1983         unsigned long size;
1984         int ret = 0;
1985         u64 value;
1986
1987         if (!event->attr.sample_period)
1988                 return -EINVAL;
1989
1990         size = copy_from_user(&value, arg, sizeof(value));
1991         if (size != sizeof(value))
1992                 return -EFAULT;
1993
1994         if (!value)
1995                 return -EINVAL;
1996
1997         raw_spin_lock_irq(&ctx->lock);
1998         if (event->attr.freq) {
1999                 if (value > sysctl_perf_event_sample_rate) {
2000                         ret = -EINVAL;
2001                         goto unlock;
2002                 }
2003
2004                 event->attr.sample_freq = value;
2005         } else {
2006                 event->attr.sample_period = value;
2007                 event->hw.sample_period = value;
2008         }
2009 unlock:
2010         raw_spin_unlock_irq(&ctx->lock);
2011
2012         return ret;
2013 }
2014
2015 static int perf_event_set_output(struct perf_event *event, int output_fd);
2016 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2017
2018 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2019 {
2020         struct perf_event *event = file->private_data;
2021         void (*func)(struct perf_event *);
2022         u32 flags = arg;
2023
2024         switch (cmd) {
2025         case PERF_EVENT_IOC_ENABLE:
2026                 func = perf_event_enable;
2027                 break;
2028         case PERF_EVENT_IOC_DISABLE:
2029                 func = perf_event_disable;
2030                 break;
2031         case PERF_EVENT_IOC_RESET:
2032                 func = perf_event_reset;
2033                 break;
2034
2035         case PERF_EVENT_IOC_REFRESH:
2036                 return perf_event_refresh(event, arg);
2037
2038         case PERF_EVENT_IOC_PERIOD:
2039                 return perf_event_period(event, (u64 __user *)arg);
2040
2041         case PERF_EVENT_IOC_SET_OUTPUT:
2042                 return perf_event_set_output(event, arg);
2043
2044         case PERF_EVENT_IOC_SET_FILTER:
2045                 return perf_event_set_filter(event, (void __user *)arg);
2046
2047         default:
2048                 return -ENOTTY;
2049         }
2050
2051         if (flags & PERF_IOC_FLAG_GROUP)
2052                 perf_event_for_each(event, func);
2053         else
2054                 perf_event_for_each_child(event, func);
2055
2056         return 0;
2057 }
2058
2059 int perf_event_task_enable(void)
2060 {
2061         struct perf_event *event;
2062
2063         mutex_lock(&current->perf_event_mutex);
2064         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2065                 perf_event_for_each_child(event, perf_event_enable);
2066         mutex_unlock(&current->perf_event_mutex);
2067
2068         return 0;
2069 }
2070
2071 int perf_event_task_disable(void)
2072 {
2073         struct perf_event *event;
2074
2075         mutex_lock(&current->perf_event_mutex);
2076         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2077                 perf_event_for_each_child(event, perf_event_disable);
2078         mutex_unlock(&current->perf_event_mutex);
2079
2080         return 0;
2081 }
2082
2083 #ifndef PERF_EVENT_INDEX_OFFSET
2084 # define PERF_EVENT_INDEX_OFFSET 0
2085 #endif
2086
2087 static int perf_event_index(struct perf_event *event)
2088 {
2089         if (event->state != PERF_EVENT_STATE_ACTIVE)
2090                 return 0;
2091
2092         return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2093 }
2094
2095 /*
2096  * Callers need to ensure there can be no nesting of this function, otherwise
2097  * the seqlock logic goes bad. We can not serialize this because the arch
2098  * code calls this from NMI context.
2099  */
2100 void perf_event_update_userpage(struct perf_event *event)
2101 {
2102         struct perf_event_mmap_page *userpg;
2103         struct perf_mmap_data *data;
2104
2105         rcu_read_lock();
2106         data = rcu_dereference(event->data);
2107         if (!data)
2108                 goto unlock;
2109
2110         userpg = data->user_page;
2111
2112         /*
2113          * Disable preemption so as to not let the corresponding user-space
2114          * spin too long if we get preempted.
2115          */
2116         preempt_disable();
2117         ++userpg->lock;
2118         barrier();
2119         userpg->index = perf_event_index(event);
2120         userpg->offset = atomic64_read(&event->count);
2121         if (event->state == PERF_EVENT_STATE_ACTIVE)
2122                 userpg->offset -= atomic64_read(&event->hw.prev_count);
2123
2124         userpg->time_enabled = event->total_time_enabled +
2125                         atomic64_read(&event->child_total_time_enabled);
2126
2127         userpg->time_running = event->total_time_running +
2128                         atomic64_read(&event->child_total_time_running);
2129
2130         barrier();
2131         ++userpg->lock;
2132         preempt_enable();
2133 unlock:
2134         rcu_read_unlock();
2135 }
2136
2137 static unsigned long perf_data_size(struct perf_mmap_data *data)
2138 {
2139         return data->nr_pages << (PAGE_SHIFT + data->data_order);
2140 }
2141
2142 #ifndef CONFIG_PERF_USE_VMALLOC
2143
2144 /*
2145  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2146  */
2147
2148 static struct page *
2149 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2150 {
2151         if (pgoff > data->nr_pages)
2152                 return NULL;
2153
2154         if (pgoff == 0)
2155                 return virt_to_page(data->user_page);
2156
2157         return virt_to_page(data->data_pages[pgoff - 1]);
2158 }
2159
2160 static struct perf_mmap_data *
2161 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2162 {
2163         struct perf_mmap_data *data;
2164         unsigned long size;
2165         int i;
2166
2167         WARN_ON(atomic_read(&event->mmap_count));
2168
2169         size = sizeof(struct perf_mmap_data);
2170         size += nr_pages * sizeof(void *);
2171
2172         data = kzalloc(size, GFP_KERNEL);
2173         if (!data)
2174                 goto fail;
2175
2176         data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2177         if (!data->user_page)
2178                 goto fail_user_page;
2179
2180         for (i = 0; i < nr_pages; i++) {
2181                 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2182                 if (!data->data_pages[i])
2183                         goto fail_data_pages;
2184         }
2185
2186         data->data_order = 0;
2187         data->nr_pages = nr_pages;
2188
2189         return data;
2190
2191 fail_data_pages:
2192         for (i--; i >= 0; i--)
2193                 free_page((unsigned long)data->data_pages[i]);
2194
2195         free_page((unsigned long)data->user_page);
2196
2197 fail_user_page:
2198         kfree(data);
2199
2200 fail:
2201         return NULL;
2202 }
2203
2204 static void perf_mmap_free_page(unsigned long addr)
2205 {
2206         struct page *page = virt_to_page((void *)addr);
2207
2208         page->mapping = NULL;
2209         __free_page(page);
2210 }
2211
2212 static void perf_mmap_data_free(struct perf_mmap_data *data)
2213 {
2214         int i;
2215
2216         perf_mmap_free_page((unsigned long)data->user_page);
2217         for (i = 0; i < data->nr_pages; i++)
2218                 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2219         kfree(data);
2220 }
2221
2222 #else
2223
2224 /*
2225  * Back perf_mmap() with vmalloc memory.
2226  *
2227  * Required for architectures that have d-cache aliasing issues.
2228  */
2229
2230 static struct page *
2231 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2232 {
2233         if (pgoff > (1UL << data->data_order))
2234                 return NULL;
2235
2236         return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2237 }
2238
2239 static void perf_mmap_unmark_page(void *addr)
2240 {
2241         struct page *page = vmalloc_to_page(addr);
2242
2243         page->mapping = NULL;
2244 }
2245
2246 static void perf_mmap_data_free_work(struct work_struct *work)
2247 {
2248         struct perf_mmap_data *data;
2249         void *base;
2250         int i, nr;
2251
2252         data = container_of(work, struct perf_mmap_data, work);
2253         nr = 1 << data->data_order;
2254
2255         base = data->user_page;
2256         for (i = 0; i < nr + 1; i++)
2257                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2258
2259         vfree(base);
2260         kfree(data);
2261 }
2262
2263 static void perf_mmap_data_free(struct perf_mmap_data *data)
2264 {
2265         schedule_work(&data->work);
2266 }
2267
2268 static struct perf_mmap_data *
2269 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2270 {
2271         struct perf_mmap_data *data;
2272         unsigned long size;
2273         void *all_buf;
2274
2275         WARN_ON(atomic_read(&event->mmap_count));
2276
2277         size = sizeof(struct perf_mmap_data);
2278         size += sizeof(void *);
2279
2280         data = kzalloc(size, GFP_KERNEL);
2281         if (!data)
2282                 goto fail;
2283
2284         INIT_WORK(&data->work, perf_mmap_data_free_work);
2285
2286         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2287         if (!all_buf)
2288                 goto fail_all_buf;
2289
2290         data->user_page = all_buf;
2291         data->data_pages[0] = all_buf + PAGE_SIZE;
2292         data->data_order = ilog2(nr_pages);
2293         data->nr_pages = 1;
2294
2295         return data;
2296
2297 fail_all_buf:
2298         kfree(data);
2299
2300 fail:
2301         return NULL;
2302 }
2303
2304 #endif
2305
2306 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2307 {
2308         struct perf_event *event = vma->vm_file->private_data;
2309         struct perf_mmap_data *data;
2310         int ret = VM_FAULT_SIGBUS;
2311
2312         if (vmf->flags & FAULT_FLAG_MKWRITE) {
2313                 if (vmf->pgoff == 0)
2314                         ret = 0;
2315                 return ret;
2316         }
2317
2318         rcu_read_lock();
2319         data = rcu_dereference(event->data);
2320         if (!data)
2321                 goto unlock;
2322
2323         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2324                 goto unlock;
2325
2326         vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2327         if (!vmf->page)
2328                 goto unlock;
2329
2330         get_page(vmf->page);
2331         vmf->page->mapping = vma->vm_file->f_mapping;
2332         vmf->page->index   = vmf->pgoff;
2333
2334         ret = 0;
2335 unlock:
2336         rcu_read_unlock();
2337
2338         return ret;
2339 }
2340
2341 static void
2342 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2343 {
2344         long max_size = perf_data_size(data);
2345
2346         atomic_set(&data->lock, -1);
2347
2348         if (event->attr.watermark) {
2349                 data->watermark = min_t(long, max_size,
2350                                         event->attr.wakeup_watermark);
2351         }
2352
2353         if (!data->watermark)
2354                 data->watermark = max_size / 2;
2355
2356
2357         rcu_assign_pointer(event->data, data);
2358 }
2359
2360 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2361 {
2362         struct perf_mmap_data *data;
2363
2364         data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2365         perf_mmap_data_free(data);
2366 }
2367
2368 static void perf_mmap_data_release(struct perf_event *event)
2369 {
2370         struct perf_mmap_data *data = event->data;
2371
2372         WARN_ON(atomic_read(&event->mmap_count));
2373
2374         rcu_assign_pointer(event->data, NULL);
2375         call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2376 }
2377
2378 static void perf_mmap_open(struct vm_area_struct *vma)
2379 {
2380         struct perf_event *event = vma->vm_file->private_data;
2381
2382         atomic_inc(&event->mmap_count);
2383 }
2384
2385 static void perf_mmap_close(struct vm_area_struct *vma)
2386 {
2387         struct perf_event *event = vma->vm_file->private_data;
2388
2389         WARN_ON_ONCE(event->ctx->parent_ctx);
2390         if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2391                 unsigned long size = perf_data_size(event->data);
2392                 struct user_struct *user = current_user();
2393
2394                 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2395                 vma->vm_mm->locked_vm -= event->data->nr_locked;
2396                 perf_mmap_data_release(event);
2397                 mutex_unlock(&event->mmap_mutex);
2398         }
2399 }
2400
2401 static const struct vm_operations_struct perf_mmap_vmops = {
2402         .open           = perf_mmap_open,
2403         .close          = perf_mmap_close,
2404         .fault          = perf_mmap_fault,
2405         .page_mkwrite   = perf_mmap_fault,
2406 };
2407
2408 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2409 {
2410         struct perf_event *event = file->private_data;
2411         unsigned long user_locked, user_lock_limit;
2412         struct user_struct *user = current_user();
2413         unsigned long locked, lock_limit;
2414         struct perf_mmap_data *data;
2415         unsigned long vma_size;
2416         unsigned long nr_pages;
2417         long user_extra, extra;
2418         int ret = 0;
2419
2420         if (!(vma->vm_flags & VM_SHARED))
2421                 return -EINVAL;
2422
2423         vma_size = vma->vm_end - vma->vm_start;
2424         nr_pages = (vma_size / PAGE_SIZE) - 1;
2425
2426         /*
2427          * If we have data pages ensure they're a power-of-two number, so we
2428          * can do bitmasks instead of modulo.
2429          */
2430         if (nr_pages != 0 && !is_power_of_2(nr_pages))
2431                 return -EINVAL;
2432
2433         if (vma_size != PAGE_SIZE * (1 + nr_pages))
2434                 return -EINVAL;
2435
2436         if (vma->vm_pgoff != 0)
2437                 return -EINVAL;
2438
2439         WARN_ON_ONCE(event->ctx->parent_ctx);
2440         mutex_lock(&event->mmap_mutex);
2441         if (event->output) {
2442                 ret = -EINVAL;
2443                 goto unlock;
2444         }
2445
2446         if (atomic_inc_not_zero(&event->mmap_count)) {
2447                 if (nr_pages != event->data->nr_pages)
2448                         ret = -EINVAL;
2449                 goto unlock;
2450         }
2451
2452         user_extra = nr_pages + 1;
2453         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2454
2455         /*
2456          * Increase the limit linearly with more CPUs:
2457          */
2458         user_lock_limit *= num_online_cpus();
2459
2460         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2461
2462         extra = 0;
2463         if (user_locked > user_lock_limit)
2464                 extra = user_locked - user_lock_limit;
2465
2466         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2467         lock_limit >>= PAGE_SHIFT;
2468         locked = vma->vm_mm->locked_vm + extra;
2469
2470         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2471                 !capable(CAP_IPC_LOCK)) {
2472                 ret = -EPERM;
2473                 goto unlock;
2474         }
2475
2476         WARN_ON(event->data);
2477
2478         data = perf_mmap_data_alloc(event, nr_pages);
2479         ret = -ENOMEM;
2480         if (!data)
2481                 goto unlock;
2482
2483         ret = 0;
2484         perf_mmap_data_init(event, data);
2485
2486         atomic_set(&event->mmap_count, 1);
2487         atomic_long_add(user_extra, &user->locked_vm);
2488         vma->vm_mm->locked_vm += extra;
2489         event->data->nr_locked = extra;
2490         if (vma->vm_flags & VM_WRITE)
2491                 event->data->writable = 1;
2492
2493 unlock:
2494         mutex_unlock(&event->mmap_mutex);
2495
2496         vma->vm_flags |= VM_RESERVED;
2497         vma->vm_ops = &perf_mmap_vmops;
2498
2499         return ret;
2500 }
2501
2502 static int perf_fasync(int fd, struct file *filp, int on)
2503 {
2504         struct inode *inode = filp->f_path.dentry->d_inode;
2505         struct perf_event *event = filp->private_data;
2506         int retval;
2507
2508         mutex_lock(&inode->i_mutex);
2509         retval = fasync_helper(fd, filp, on, &event->fasync);
2510         mutex_unlock(&inode->i_mutex);
2511
2512         if (retval < 0)
2513                 return retval;
2514
2515         return 0;
2516 }
2517
2518 static const struct file_operations perf_fops = {
2519         .release                = perf_release,
2520         .read                   = perf_read,
2521         .poll                   = perf_poll,
2522         .unlocked_ioctl         = perf_ioctl,
2523         .compat_ioctl           = perf_ioctl,
2524         .mmap                   = perf_mmap,
2525         .fasync                 = perf_fasync,
2526 };
2527
2528 /*
2529  * Perf event wakeup
2530  *
2531  * If there's data, ensure we set the poll() state and publish everything
2532  * to user-space before waking everybody up.
2533  */
2534
2535 void perf_event_wakeup(struct perf_event *event)
2536 {
2537         wake_up_all(&event->waitq);
2538
2539         if (event->pending_kill) {
2540                 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2541                 event->pending_kill = 0;
2542         }
2543 }
2544
2545 /*
2546  * Pending wakeups
2547  *
2548  * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2549  *
2550  * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2551  * single linked list and use cmpxchg() to add entries lockless.
2552  */
2553
2554 static void perf_pending_event(struct perf_pending_entry *entry)
2555 {
2556         struct perf_event *event = container_of(entry,
2557                         struct perf_event, pending);
2558
2559         if (event->pending_disable) {
2560                 event->pending_disable = 0;
2561                 __perf_event_disable(event);
2562         }
2563
2564         if (event->pending_wakeup) {
2565                 event->pending_wakeup = 0;
2566                 perf_event_wakeup(event);
2567         }
2568 }
2569
2570 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2571
2572 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2573         PENDING_TAIL,
2574 };
2575
2576 static void perf_pending_queue(struct perf_pending_entry *entry,
2577                                void (*func)(struct perf_pending_entry *))
2578 {
2579         struct perf_pending_entry **head;
2580
2581         if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2582                 return;
2583
2584         entry->func = func;
2585
2586         head = &get_cpu_var(perf_pending_head);
2587
2588         do {
2589                 entry->next = *head;
2590         } while (cmpxchg(head, entry->next, entry) != entry->next);
2591
2592         set_perf_event_pending();
2593
2594         put_cpu_var(perf_pending_head);
2595 }
2596
2597 static int __perf_pending_run(void)
2598 {
2599         struct perf_pending_entry *list;
2600         int nr = 0;
2601
2602         list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2603         while (list != PENDING_TAIL) {
2604                 void (*func)(struct perf_pending_entry *);
2605                 struct perf_pending_entry *entry = list;
2606
2607                 list = list->next;
2608
2609                 func = entry->func;
2610                 entry->next = NULL;
2611                 /*
2612                  * Ensure we observe the unqueue before we issue the wakeup,
2613                  * so that we won't be waiting forever.
2614                  * -- see perf_not_pending().
2615                  */
2616                 smp_wmb();
2617
2618                 func(entry);
2619                 nr++;
2620         }
2621
2622         return nr;
2623 }
2624
2625 static inline int perf_not_pending(struct perf_event *event)
2626 {
2627         /*
2628          * If we flush on whatever cpu we run, there is a chance we don't
2629          * need to wait.
2630          */
2631         get_cpu();
2632         __perf_pending_run();
2633         put_cpu();
2634
2635         /*
2636          * Ensure we see the proper queue state before going to sleep
2637          * so that we do not miss the wakeup. -- see perf_pending_handle()
2638          */
2639         smp_rmb();
2640         return event->pending.next == NULL;
2641 }
2642
2643 static void perf_pending_sync(struct perf_event *event)
2644 {
2645         wait_event(event->waitq, perf_not_pending(event));
2646 }
2647
2648 void perf_event_do_pending(void)
2649 {
2650         __perf_pending_run();
2651 }
2652
2653 /*
2654  * Callchain support -- arch specific
2655  */
2656
2657 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2658 {
2659         return NULL;
2660 }
2661
2662 /*
2663  * Output
2664  */
2665 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2666                               unsigned long offset, unsigned long head)
2667 {
2668         unsigned long mask;
2669
2670         if (!data->writable)
2671                 return true;
2672
2673         mask = perf_data_size(data) - 1;
2674
2675         offset = (offset - tail) & mask;
2676         head   = (head   - tail) & mask;
2677
2678         if ((int)(head - offset) < 0)
2679                 return false;
2680
2681         return true;
2682 }
2683
2684 static void perf_output_wakeup(struct perf_output_handle *handle)
2685 {
2686         atomic_set(&handle->data->poll, POLL_IN);
2687
2688         if (handle->nmi) {
2689                 handle->event->pending_wakeup = 1;
2690                 perf_pending_queue(&handle->event->pending,
2691                                    perf_pending_event);
2692         } else
2693                 perf_event_wakeup(handle->event);
2694 }
2695
2696 /*
2697  * Curious locking construct.
2698  *
2699  * We need to ensure a later event_id doesn't publish a head when a former
2700  * event_id isn't done writing. However since we need to deal with NMIs we
2701  * cannot fully serialize things.
2702  *
2703  * What we do is serialize between CPUs so we only have to deal with NMI
2704  * nesting on a single CPU.
2705  *
2706  * We only publish the head (and generate a wakeup) when the outer-most
2707  * event_id completes.
2708  */
2709 static void perf_output_lock(struct perf_output_handle *handle)
2710 {
2711         struct perf_mmap_data *data = handle->data;
2712         int cur, cpu = get_cpu();
2713
2714         handle->locked = 0;
2715
2716         for (;;) {
2717                 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2718                 if (cur == -1) {
2719                         handle->locked = 1;
2720                         break;
2721                 }
2722                 if (cur == cpu)
2723                         break;
2724
2725                 cpu_relax();
2726         }
2727 }
2728
2729 static void perf_output_unlock(struct perf_output_handle *handle)
2730 {
2731         struct perf_mmap_data *data = handle->data;
2732         unsigned long head;
2733         int cpu;
2734
2735         data->done_head = data->head;
2736
2737         if (!handle->locked)
2738                 goto out;
2739
2740 again:
2741         /*
2742          * The xchg implies a full barrier that ensures all writes are done
2743          * before we publish the new head, matched by a rmb() in userspace when
2744          * reading this position.
2745          */
2746         while ((head = atomic_long_xchg(&data->done_head, 0)))
2747                 data->user_page->data_head = head;
2748
2749         /*
2750          * NMI can happen here, which means we can miss a done_head update.
2751          */
2752
2753         cpu = atomic_xchg(&data->lock, -1);
2754         WARN_ON_ONCE(cpu != smp_processor_id());
2755
2756         /*
2757          * Therefore we have to validate we did not indeed do so.
2758          */
2759         if (unlikely(atomic_long_read(&data->done_head))) {
2760                 /*
2761                  * Since we had it locked, we can lock it again.
2762                  */
2763                 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2764                         cpu_relax();
2765
2766                 goto again;
2767         }
2768
2769         if (atomic_xchg(&data->wakeup, 0))
2770                 perf_output_wakeup(handle);
2771 out:
2772         put_cpu();
2773 }
2774
2775 void perf_output_copy(struct perf_output_handle *handle,
2776                       const void *buf, unsigned int len)
2777 {
2778         unsigned int pages_mask;
2779         unsigned long offset;
2780         unsigned int size;
2781         void **pages;
2782
2783         offset          = handle->offset;
2784         pages_mask      = handle->data->nr_pages - 1;
2785         pages           = handle->data->data_pages;
2786
2787         do {
2788                 unsigned long page_offset;
2789                 unsigned long page_size;
2790                 int nr;
2791
2792                 nr          = (offset >> PAGE_SHIFT) & pages_mask;
2793                 page_size   = 1UL << (handle->data->data_order + PAGE_SHIFT);
2794                 page_offset = offset & (page_size - 1);
2795                 size        = min_t(unsigned int, page_size - page_offset, len);
2796
2797                 memcpy(pages[nr] + page_offset, buf, size);
2798
2799                 len         -= size;
2800                 buf         += size;
2801                 offset      += size;
2802         } while (len);
2803
2804         handle->offset = offset;
2805
2806         /*
2807          * Check we didn't copy past our reservation window, taking the
2808          * possible unsigned int wrap into account.
2809          */
2810         WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2811 }
2812
2813 int perf_output_begin(struct perf_output_handle *handle,
2814                       struct perf_event *event, unsigned int size,
2815                       int nmi, int sample)
2816 {
2817         struct perf_event *output_event;
2818         struct perf_mmap_data *data;
2819         unsigned long tail, offset, head;
2820         int have_lost;
2821         struct {
2822                 struct perf_event_header header;
2823                 u64                      id;
2824                 u64                      lost;
2825         } lost_event;
2826
2827         rcu_read_lock();
2828         /*
2829          * For inherited events we send all the output towards the parent.
2830          */
2831         if (event->parent)
2832                 event = event->parent;
2833
2834         output_event = rcu_dereference(event->output);
2835         if (output_event)
2836                 event = output_event;
2837
2838         data = rcu_dereference(event->data);
2839         if (!data)
2840                 goto out;
2841
2842         handle->data    = data;
2843         handle->event   = event;
2844         handle->nmi     = nmi;
2845         handle->sample  = sample;
2846
2847         if (!data->nr_pages)
2848                 goto fail;
2849
2850         have_lost = atomic_read(&data->lost);
2851         if (have_lost)
2852                 size += sizeof(lost_event);
2853
2854         perf_output_lock(handle);
2855
2856         do {
2857                 /*
2858                  * Userspace could choose to issue a mb() before updating the
2859                  * tail pointer. So that all reads will be completed before the
2860                  * write is issued.
2861                  */
2862                 tail = ACCESS_ONCE(data->user_page->data_tail);
2863                 smp_rmb();
2864                 offset = head = atomic_long_read(&data->head);
2865                 head += size;
2866                 if (unlikely(!perf_output_space(data, tail, offset, head)))
2867                         goto fail;
2868         } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2869
2870         handle->offset  = offset;
2871         handle->head    = head;
2872
2873         if (head - tail > data->watermark)
2874                 atomic_set(&data->wakeup, 1);
2875
2876         if (have_lost) {
2877                 lost_event.header.type = PERF_RECORD_LOST;
2878                 lost_event.header.misc = 0;
2879                 lost_event.header.size = sizeof(lost_event);
2880                 lost_event.id          = event->id;
2881                 lost_event.lost        = atomic_xchg(&data->lost, 0);
2882
2883                 perf_output_put(handle, lost_event);
2884         }
2885
2886         return 0;
2887
2888 fail:
2889         atomic_inc(&data->lost);
2890         perf_output_unlock(handle);
2891 out:
2892         rcu_read_unlock();
2893
2894         return -ENOSPC;
2895 }
2896
2897 void perf_output_end(struct perf_output_handle *handle)
2898 {
2899         struct perf_event *event = handle->event;
2900         struct perf_mmap_data *data = handle->data;
2901
2902         int wakeup_events = event->attr.wakeup_events;
2903
2904         if (handle->sample && wakeup_events) {
2905                 int events = atomic_inc_return(&data->events);
2906                 if (events >= wakeup_events) {
2907                         atomic_sub(wakeup_events, &data->events);
2908                         atomic_set(&data->wakeup, 1);
2909                 }
2910         }
2911
2912         perf_output_unlock(handle);
2913         rcu_read_unlock();
2914 }
2915
2916 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
2917 {
2918         /*
2919          * only top level events have the pid namespace they were created in
2920          */
2921         if (event->parent)
2922                 event = event->parent;
2923
2924         return task_tgid_nr_ns(p, event->ns);
2925 }
2926
2927 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
2928 {
2929         /*
2930          * only top level events have the pid namespace they were created in
2931          */
2932         if (event->parent)
2933                 event = event->parent;
2934
2935         return task_pid_nr_ns(p, event->ns);
2936 }
2937
2938 static void perf_output_read_one(struct perf_output_handle *handle,
2939                                  struct perf_event *event)
2940 {
2941         u64 read_format = event->attr.read_format;
2942         u64 values[4];
2943         int n = 0;
2944
2945         values[n++] = atomic64_read(&event->count);
2946         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2947                 values[n++] = event->total_time_enabled +
2948                         atomic64_read(&event->child_total_time_enabled);
2949         }
2950         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2951                 values[n++] = event->total_time_running +
2952                         atomic64_read(&event->child_total_time_running);
2953         }
2954         if (read_format & PERF_FORMAT_ID)
2955                 values[n++] = primary_event_id(event);
2956
2957         perf_output_copy(handle, values, n * sizeof(u64));
2958 }
2959
2960 /*
2961  * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
2962  */
2963 static void perf_output_read_group(struct perf_output_handle *handle,
2964                             struct perf_event *event)
2965 {
2966         struct perf_event *leader = event->group_leader, *sub;
2967         u64 read_format = event->attr.read_format;
2968         u64 values[5];
2969         int n = 0;
2970
2971         values[n++] = 1 + leader->nr_siblings;
2972
2973         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2974                 values[n++] = leader->total_time_enabled;
2975
2976         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2977                 values[n++] = leader->total_time_running;
2978
2979         if (leader != event)
2980                 leader->pmu->read(leader);
2981
2982         values[n++] = atomic64_read(&leader->count);
2983         if (read_format & PERF_FORMAT_ID)
2984                 values[n++] = primary_event_id(leader);
2985
2986         perf_output_copy(handle, values, n * sizeof(u64));
2987
2988         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2989                 n = 0;
2990
2991                 if (sub != event)
2992                         sub->pmu->read(sub);
2993
2994                 values[n++] = atomic64_read(&sub->count);
2995                 if (read_format & PERF_FORMAT_ID)
2996                         values[n++] = primary_event_id(sub);
2997
2998                 perf_output_copy(handle, values, n * sizeof(u64));
2999         }
3000 }
3001
3002 static void perf_output_read(struct perf_output_handle *handle,
3003                              struct perf_event *event)
3004 {
3005         if (event->attr.read_format & PERF_FORMAT_GROUP)
3006                 perf_output_read_group(handle, event);
3007         else
3008                 perf_output_read_one(handle, event);
3009 }
3010
3011 void perf_output_sample(struct perf_output_handle *handle,
3012                         struct perf_event_header *header,
3013                         struct perf_sample_data *data,
3014                         struct perf_event *event)
3015 {
3016         u64 sample_type = data->type;
3017
3018         perf_output_put(handle, *header);
3019
3020         if (sample_type & PERF_SAMPLE_IP)
3021                 perf_output_put(handle, data->ip);
3022
3023         if (sample_type & PERF_SAMPLE_TID)
3024                 perf_output_put(handle, data->tid_entry);
3025
3026         if (sample_type & PERF_SAMPLE_TIME)
3027                 perf_output_put(handle, data->time);
3028
3029         if (sample_type & PERF_SAMPLE_ADDR)
3030                 perf_output_put(handle, data->addr);
3031
3032         if (sample_type & PERF_SAMPLE_ID)
3033                 perf_output_put(handle, data->id);
3034
3035         if (sample_type & PERF_SAMPLE_STREAM_ID)
3036                 perf_output_put(handle, data->stream_id);
3037
3038         if (sample_type & PERF_SAMPLE_CPU)
3039                 perf_output_put(handle, data->cpu_entry);
3040
3041         if (sample_type & PERF_SAMPLE_PERIOD)
3042                 perf_output_put(handle, data->period);
3043
3044         if (sample_type & PERF_SAMPLE_READ)
3045                 perf_output_read(handle, event);
3046
3047         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3048                 if (data->callchain) {
3049                         int size = 1;
3050
3051                         if (data->callchain)
3052                                 size += data->callchain->nr;
3053
3054                         size *= sizeof(u64);
3055
3056                         perf_output_copy(handle, data->callchain, size);
3057                 } else {
3058                         u64 nr = 0;
3059                         perf_output_put(handle, nr);
3060                 }
3061         }
3062
3063         if (sample_type & PERF_SAMPLE_RAW) {
3064                 if (data->raw) {
3065                         perf_output_put(handle, data->raw->size);
3066                         perf_output_copy(handle, data->raw->data,
3067                                          data->raw->size);
3068                 } else {
3069                         struct {
3070                                 u32     size;
3071                                 u32     data;
3072                         } raw = {
3073                                 .size = sizeof(u32),
3074                                 .data = 0,
3075                         };
3076                         perf_output_put(handle, raw);
3077                 }
3078         }
3079 }
3080
3081 void perf_prepare_sample(struct perf_event_header *header,
3082                          struct perf_sample_data *data,
3083                          struct perf_event *event,
3084                          struct pt_regs *regs)
3085 {
3086         u64 sample_type = event->attr.sample_type;
3087
3088         data->type = sample_type;
3089
3090         header->type = PERF_RECORD_SAMPLE;
3091         header->size = sizeof(*header);
3092
3093         header->misc = 0;
3094         header->misc |= perf_misc_flags(regs);
3095
3096         if (sample_type & PERF_SAMPLE_IP) {
3097                 data->ip = perf_instruction_pointer(regs);
3098
3099                 header->size += sizeof(data->ip);
3100         }
3101
3102         if (sample_type & PERF_SAMPLE_TID) {
3103                 /* namespace issues */
3104                 data->tid_entry.pid = perf_event_pid(event, current);
3105                 data->tid_entry.tid = perf_event_tid(event, current);
3106
3107                 header->size += sizeof(data->tid_entry);
3108         }
3109
3110         if (sample_type & PERF_SAMPLE_TIME) {
3111                 data->time = perf_clock();
3112
3113                 header->size += sizeof(data->time);
3114         }
3115
3116         if (sample_type & PERF_SAMPLE_ADDR)
3117                 header->size += sizeof(data->addr);
3118
3119         if (sample_type & PERF_SAMPLE_ID) {
3120                 data->id = primary_event_id(event);
3121
3122                 header->size += sizeof(data->id);
3123         }
3124
3125         if (sample_type & PERF_SAMPLE_STREAM_ID) {
3126                 data->stream_id = event->id;
3127
3128                 header->size += sizeof(data->stream_id);
3129         }
3130
3131         if (sample_type & PERF_SAMPLE_CPU) {
3132                 data->cpu_entry.cpu             = raw_smp_processor_id();
3133                 data->cpu_entry.reserved        = 0;
3134
3135                 header->size += sizeof(data->cpu_entry);
3136         }
3137
3138         if (sample_type & PERF_SAMPLE_PERIOD)
3139                 header->size += sizeof(data->period);
3140
3141         if (sample_type & PERF_SAMPLE_READ)
3142                 header->size += perf_event_read_size(event);
3143
3144         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3145                 int size = 1;
3146
3147                 data->callchain = perf_callchain(regs);
3148
3149                 if (data->callchain)
3150                         size += data->callchain->nr;
3151
3152                 header->size += size * sizeof(u64);
3153         }
3154
3155         if (sample_type & PERF_SAMPLE_RAW) {
3156                 int size = sizeof(u32);
3157
3158                 if (data->raw)
3159                         size += data->raw->size;
3160                 else
3161                         size += sizeof(u32);
3162
3163                 WARN_ON_ONCE(size & (sizeof(u64)-1));
3164                 header->size += size;
3165         }
3166 }
3167
3168 static void perf_event_output(struct perf_event *event, int nmi,
3169                                 struct perf_sample_data *data,
3170                                 struct pt_regs *regs)
3171 {
3172         struct perf_output_handle handle;
3173         struct perf_event_header header;
3174
3175         perf_prepare_sample(&header, data, event, regs);
3176
3177         if (perf_output_begin(&handle, event, header.size, nmi, 1))
3178                 return;
3179
3180         perf_output_sample(&handle, &header, data, event);
3181
3182         perf_output_end(&handle);
3183 }
3184
3185 /*
3186  * read event_id
3187  */
3188
3189 struct perf_read_event {
3190         struct perf_event_header        header;
3191
3192         u32                             pid;
3193         u32                             tid;
3194 };
3195
3196 static void
3197 perf_event_read_event(struct perf_event *event,
3198                         struct task_struct *task)
3199 {
3200         struct perf_output_handle handle;
3201         struct perf_read_event read_event = {
3202                 .header = {
3203                         .type = PERF_RECORD_READ,
3204                         .misc = 0,
3205                         .size = sizeof(read_event) + perf_event_read_size(event),
3206                 },
3207                 .pid = perf_event_pid(event, task),
3208                 .tid = perf_event_tid(event, task),
3209         };
3210         int ret;
3211
3212         ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3213         if (ret)
3214                 return;
3215
3216         perf_output_put(&handle, read_event);
3217         perf_output_read(&handle, event);
3218
3219         perf_output_end(&handle);
3220 }
3221
3222 /*
3223  * task tracking -- fork/exit
3224  *
3225  * enabled by: attr.comm | attr.mmap | attr.task
3226  */
3227
3228 struct perf_task_event {
3229         struct task_struct              *task;
3230         struct perf_event_context       *task_ctx;
3231
3232         struct {
3233                 struct perf_event_header        header;
3234
3235                 u32                             pid;
3236                 u32                             ppid;
3237                 u32                             tid;
3238                 u32                             ptid;
3239                 u64                             time;
3240         } event_id;
3241 };
3242
3243 static void perf_event_task_output(struct perf_event *event,
3244                                      struct perf_task_event *task_event)
3245 {
3246         struct perf_output_handle handle;
3247         int size;
3248         struct task_struct *task = task_event->task;
3249         int ret;
3250
3251         size  = task_event->event_id.header.size;
3252         ret = perf_output_begin(&handle, event, size, 0, 0);
3253
3254         if (ret)
3255                 return;
3256
3257         task_event->event_id.pid = perf_event_pid(event, task);
3258         task_event->event_id.ppid = perf_event_pid(event, current);
3259
3260         task_event->event_id.tid = perf_event_tid(event, task);
3261         task_event->event_id.ptid = perf_event_tid(event, current);
3262
3263         task_event->event_id.time = perf_clock();
3264
3265         perf_output_put(&handle, task_event->event_id);
3266
3267         perf_output_end(&handle);
3268 }
3269
3270 static int perf_event_task_match(struct perf_event *event)
3271 {
3272         if (event->cpu != -1 && event->cpu != smp_processor_id())
3273                 return 0;
3274
3275         if (event->attr.comm || event->attr.mmap || event->attr.task)
3276                 return 1;
3277
3278         return 0;
3279 }
3280
3281 static void perf_event_task_ctx(struct perf_event_context *ctx,
3282                                   struct perf_task_event *task_event)
3283 {
3284         struct perf_event *event;
3285
3286         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3287                 if (perf_event_task_match(event))
3288                         perf_event_task_output(event, task_event);
3289         }
3290 }
3291
3292 static void perf_event_task_event(struct perf_task_event *task_event)
3293 {
3294         struct perf_cpu_context *cpuctx;
3295         struct perf_event_context *ctx = task_event->task_ctx;
3296
3297         rcu_read_lock();
3298         cpuctx = &get_cpu_var(perf_cpu_context);
3299         perf_event_task_ctx(&cpuctx->ctx, task_event);
3300         if (!ctx)
3301                 ctx = rcu_dereference(task_event->task->perf_event_ctxp);
3302         if (ctx)
3303                 perf_event_task_ctx(ctx, task_event);
3304         put_cpu_var(perf_cpu_context);
3305         rcu_read_unlock();
3306 }
3307
3308 static void perf_event_task(struct task_struct *task,
3309                               struct perf_event_context *task_ctx,
3310                               int new)
3311 {
3312         struct perf_task_event task_event;
3313
3314         if (!atomic_read(&nr_comm_events) &&
3315             !atomic_read(&nr_mmap_events) &&
3316             !atomic_read(&nr_task_events))
3317                 return;
3318
3319         task_event = (struct perf_task_event){
3320                 .task     = task,
3321                 .task_ctx = task_ctx,
3322                 .event_id    = {
3323                         .header = {
3324                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3325                                 .misc = 0,
3326                                 .size = sizeof(task_event.event_id),
3327                         },
3328                         /* .pid  */
3329                         /* .ppid */
3330                         /* .tid  */
3331                         /* .ptid */
3332                 },
3333         };
3334
3335         perf_event_task_event(&task_event);
3336 }
3337
3338 void perf_event_fork(struct task_struct *task)
3339 {
3340         perf_event_task(task, NULL, 1);
3341 }
3342
3343 /*
3344  * comm tracking
3345  */
3346
3347 struct perf_comm_event {
3348         struct task_struct      *task;
3349         char                    *comm;
3350         int                     comm_size;
3351
3352         struct {
3353                 struct perf_event_header        header;
3354
3355                 u32                             pid;
3356                 u32                             tid;
3357         } event_id;
3358 };
3359
3360 static void perf_event_comm_output(struct perf_event *event,
3361                                      struct perf_comm_event *comm_event)
3362 {
3363         struct perf_output_handle handle;
3364         int size = comm_event->event_id.header.size;
3365         int ret = perf_output_begin(&handle, event, size, 0, 0);
3366
3367         if (ret)
3368                 return;
3369
3370         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3371         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3372
3373         perf_output_put(&handle, comm_event->event_id);
3374         perf_output_copy(&handle, comm_event->comm,
3375                                    comm_event->comm_size);
3376         perf_output_end(&handle);
3377 }
3378
3379 static int perf_event_comm_match(struct perf_event *event)
3380 {
3381         if (event->cpu != -1 && event->cpu != smp_processor_id())
3382                 return 0;
3383
3384         if (event->attr.comm)
3385                 return 1;
3386
3387         return 0;
3388 }
3389
3390 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3391                                   struct perf_comm_event *comm_event)
3392 {
3393         struct perf_event *event;
3394
3395         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3396                 if (perf_event_comm_match(event))
3397                         perf_event_comm_output(event, comm_event);
3398         }
3399 }
3400
3401 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3402 {
3403         struct perf_cpu_context *cpuctx;
3404         struct perf_event_context *ctx;
3405         unsigned int size;
3406         char comm[TASK_COMM_LEN];
3407
3408         memset(comm, 0, sizeof(comm));
3409         strlcpy(comm, comm_event->task->comm, sizeof(comm));
3410         size = ALIGN(strlen(comm)+1, sizeof(u64));
3411
3412         comm_event->comm = comm;
3413         comm_event->comm_size = size;
3414
3415         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3416
3417         rcu_read_lock();
3418         cpuctx = &get_cpu_var(perf_cpu_context);
3419         perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3420         ctx = rcu_dereference(current->perf_event_ctxp);
3421         if (ctx)
3422                 perf_event_comm_ctx(ctx, comm_event);
3423         put_cpu_var(perf_cpu_context);
3424         rcu_read_unlock();
3425 }
3426
3427 void perf_event_comm(struct task_struct *task)
3428 {
3429         struct perf_comm_event comm_event;
3430
3431         if (task->perf_event_ctxp)
3432                 perf_event_enable_on_exec(task);
3433
3434         if (!atomic_read(&nr_comm_events))
3435                 return;
3436
3437         comm_event = (struct perf_comm_event){
3438                 .task   = task,
3439                 /* .comm      */
3440                 /* .comm_size */
3441                 .event_id  = {
3442                         .header = {
3443                                 .type = PERF_RECORD_COMM,
3444                                 .misc = 0,
3445                                 /* .size */
3446                         },
3447                         /* .pid */
3448                         /* .tid */
3449                 },
3450         };
3451
3452         perf_event_comm_event(&comm_event);
3453 }
3454
3455 /*
3456  * mmap tracking
3457  */
3458
3459 struct perf_mmap_event {
3460         struct vm_area_struct   *vma;
3461
3462         const char              *file_name;
3463         int                     file_size;
3464
3465         struct {
3466                 struct perf_event_header        header;
3467
3468                 u32                             pid;
3469                 u32                             tid;
3470                 u64                             start;
3471                 u64                             len;
3472                 u64                             pgoff;
3473         } event_id;
3474 };
3475
3476 static void perf_event_mmap_output(struct perf_event *event,
3477                                      struct perf_mmap_event *mmap_event)
3478 {
3479         struct perf_output_handle handle;
3480         int size = mmap_event->event_id.header.size;
3481         int ret = perf_output_begin(&handle, event, size, 0, 0);
3482
3483         if (ret)
3484                 return;
3485
3486         mmap_event->event_id.pid = perf_event_pid(event, current);
3487         mmap_event->event_id.tid = perf_event_tid(event, current);
3488
3489         perf_output_put(&handle, mmap_event->event_id);
3490         perf_output_copy(&handle, mmap_event->file_name,
3491                                    mmap_event->file_size);
3492         perf_output_end(&handle);
3493 }
3494
3495 static int perf_event_mmap_match(struct perf_event *event,
3496                                    struct perf_mmap_event *mmap_event)
3497 {
3498         if (event->cpu != -1 && event->cpu != smp_processor_id())
3499                 return 0;
3500
3501         if (event->attr.mmap)
3502                 return 1;
3503
3504         return 0;
3505 }
3506
3507 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3508                                   struct perf_mmap_event *mmap_event)
3509 {
3510         struct perf_event *event;
3511
3512         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3513                 if (perf_event_mmap_match(event, mmap_event))
3514                         perf_event_mmap_output(event, mmap_event);
3515         }
3516 }
3517
3518 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3519 {
3520         struct perf_cpu_context *cpuctx;
3521         struct perf_event_context *ctx;
3522         struct vm_area_struct *vma = mmap_event->vma;
3523         struct file *file = vma->vm_file;
3524         unsigned int size;
3525         char tmp[16];
3526         char *buf = NULL;
3527         const char *name;
3528
3529         memset(tmp, 0, sizeof(tmp));
3530
3531         if (file) {
3532                 /*
3533                  * d_path works from the end of the buffer backwards, so we
3534                  * need to add enough zero bytes after the string to handle
3535                  * the 64bit alignment we do later.
3536                  */
3537                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3538                 if (!buf) {
3539                         name = strncpy(tmp, "//enomem", sizeof(tmp));
3540                         goto got_name;
3541                 }
3542                 name = d_path(&file->f_path, buf, PATH_MAX);
3543                 if (IS_ERR(name)) {
3544                         name = strncpy(tmp, "//toolong", sizeof(tmp));
3545                         goto got_name;
3546                 }
3547         } else {
3548                 if (arch_vma_name(mmap_event->vma)) {
3549                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3550                                        sizeof(tmp));
3551                         goto got_name;
3552                 }
3553
3554                 if (!vma->vm_mm) {
3555                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
3556                         goto got_name;
3557                 }
3558
3559                 name = strncpy(tmp, "//anon", sizeof(tmp));
3560                 goto got_name;
3561         }
3562
3563 got_name:
3564         size = ALIGN(strlen(name)+1, sizeof(u64));
3565
3566         mmap_event->file_name = name;
3567         mmap_event->file_size = size;
3568
3569         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3570
3571         rcu_read_lock();
3572         cpuctx = &get_cpu_var(perf_cpu_context);
3573         perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3574         ctx = rcu_dereference(current->perf_event_ctxp);
3575         if (ctx)
3576                 perf_event_mmap_ctx(ctx, mmap_event);
3577         put_cpu_var(perf_cpu_context);
3578         rcu_read_unlock();
3579
3580         kfree(buf);
3581 }
3582
3583 void __perf_event_mmap(struct vm_area_struct *vma)
3584 {
3585         struct perf_mmap_event mmap_event;
3586
3587         if (!atomic_read(&nr_mmap_events))
3588                 return;
3589
3590         mmap_event = (struct perf_mmap_event){
3591                 .vma    = vma,
3592                 /* .file_name */
3593                 /* .file_size */
3594                 .event_id  = {
3595                         .header = {
3596                                 .type = PERF_RECORD_MMAP,
3597                                 .misc = 0,
3598                                 /* .size */
3599                         },
3600                         /* .pid */
3601                         /* .tid */
3602                         .start  = vma->vm_start,
3603                         .len    = vma->vm_end - vma->vm_start,
3604                         .pgoff  = vma->vm_pgoff,
3605                 },
3606         };
3607
3608         perf_event_mmap_event(&mmap_event);
3609 }
3610
3611 /*
3612  * IRQ throttle logging
3613  */
3614
3615 static void perf_log_throttle(struct perf_event *event, int enable)
3616 {
3617         struct perf_output_handle handle;
3618         int ret;
3619
3620         struct {
3621                 struct perf_event_header        header;
3622                 u64                             time;
3623                 u64                             id;
3624                 u64                             stream_id;
3625         } throttle_event = {
3626                 .header = {
3627                         .type = PERF_RECORD_THROTTLE,
3628                         .misc = 0,
3629                         .size = sizeof(throttle_event),
3630                 },
3631                 .time           = perf_clock(),
3632                 .id             = primary_event_id(event),
3633                 .stream_id      = event->id,
3634         };
3635
3636         if (enable)
3637                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3638
3639         ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3640         if (ret)
3641                 return;
3642
3643         perf_output_put(&handle, throttle_event);
3644         perf_output_end(&handle);
3645 }
3646
3647 /*
3648  * Generic event overflow handling, sampling.
3649  */
3650
3651 static int __perf_event_overflow(struct perf_event *event, int nmi,
3652                                    int throttle, struct perf_sample_data *data,
3653                                    struct pt_regs *regs)
3654 {
3655         int events = atomic_read(&event->event_limit);
3656         struct hw_perf_event *hwc = &event->hw;
3657         int ret = 0;
3658
3659         throttle = (throttle && event->pmu->unthrottle != NULL);
3660
3661         if (!throttle) {
3662                 hwc->interrupts++;
3663         } else {
3664                 if (hwc->interrupts != MAX_INTERRUPTS) {
3665                         hwc->interrupts++;
3666                         if (HZ * hwc->interrupts >
3667                                         (u64)sysctl_perf_event_sample_rate) {
3668                                 hwc->interrupts = MAX_INTERRUPTS;
3669                                 perf_log_throttle(event, 0);
3670                                 ret = 1;
3671                         }
3672                 } else {
3673                         /*
3674                          * Keep re-disabling events even though on the previous
3675                          * pass we disabled it - just in case we raced with a
3676                          * sched-in and the event got enabled again:
3677                          */
3678                         ret = 1;
3679                 }
3680         }
3681
3682         if (event->attr.freq) {
3683                 u64 now = perf_clock();
3684                 s64 delta = now - hwc->freq_stamp;
3685
3686                 hwc->freq_stamp = now;
3687
3688                 if (delta > 0 && delta < TICK_NSEC)
3689                         perf_adjust_period(event, NSEC_PER_SEC / (int)delta);
3690         }
3691
3692         /*
3693          * XXX event_limit might not quite work as expected on inherited
3694          * events
3695          */
3696
3697         event->pending_kill = POLL_IN;
3698         if (events && atomic_dec_and_test(&event->event_limit)) {
3699                 ret = 1;
3700                 event->pending_kill = POLL_HUP;
3701                 if (nmi) {
3702                         event->pending_disable = 1;
3703                         perf_pending_queue(&event->pending,
3704                                            perf_pending_event);
3705                 } else
3706                         perf_event_disable(event);
3707         }
3708
3709         if (event->overflow_handler)
3710                 event->overflow_handler(event, nmi, data, regs);
3711         else
3712                 perf_event_output(event, nmi, data, regs);
3713
3714         return ret;
3715 }
3716
3717 int perf_event_overflow(struct perf_event *event, int nmi,
3718                           struct perf_sample_data *data,
3719                           struct pt_regs *regs)
3720 {
3721         return __perf_event_overflow(event, nmi, 1, data, regs);
3722 }
3723
3724 /*
3725  * Generic software event infrastructure
3726  */
3727
3728 /*
3729  * We directly increment event->count and keep a second value in
3730  * event->hw.period_left to count intervals. This period event
3731  * is kept in the range [-sample_period, 0] so that we can use the
3732  * sign as trigger.
3733  */
3734
3735 static u64 perf_swevent_set_period(struct perf_event *event)
3736 {
3737         struct hw_perf_event *hwc = &event->hw;
3738         u64 period = hwc->last_period;
3739         u64 nr, offset;
3740         s64 old, val;
3741
3742         hwc->last_period = hwc->sample_period;
3743
3744 again:
3745         old = val = atomic64_read(&hwc->period_left);
3746         if (val < 0)
3747                 return 0;
3748
3749         nr = div64_u64(period + val, period);
3750         offset = nr * period;
3751         val -= offset;
3752         if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3753                 goto again;
3754
3755         return nr;
3756 }
3757
3758 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3759                                     int nmi, struct perf_sample_data *data,
3760                                     struct pt_regs *regs)
3761 {
3762         struct hw_perf_event *hwc = &event->hw;
3763         int throttle = 0;
3764
3765         data->period = event->hw.last_period;
3766         if (!overflow)
3767                 overflow = perf_swevent_set_period(event);
3768
3769         if (hwc->interrupts == MAX_INTERRUPTS)
3770                 return;
3771
3772         for (; overflow; overflow--) {
3773                 if (__perf_event_overflow(event, nmi, throttle,
3774                                             data, regs)) {
3775                         /*
3776                          * We inhibit the overflow from happening when
3777                          * hwc->interrupts == MAX_INTERRUPTS.
3778                          */
3779                         break;
3780                 }
3781                 throttle = 1;
3782         }
3783 }
3784
3785 static void perf_swevent_unthrottle(struct perf_event *event)
3786 {
3787         /*
3788          * Nothing to do, we already reset hwc->interrupts.
3789          */
3790 }
3791
3792 static void perf_swevent_add(struct perf_event *event, u64 nr,
3793                                int nmi, struct perf_sample_data *data,
3794                                struct pt_regs *regs)
3795 {
3796         struct hw_perf_event *hwc = &event->hw;
3797
3798         atomic64_add(nr, &event->count);
3799
3800         if (!regs)
3801                 return;
3802
3803         if (!hwc->sample_period)
3804                 return;
3805
3806         if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3807                 return perf_swevent_overflow(event, 1, nmi, data, regs);
3808
3809         if (atomic64_add_negative(nr, &hwc->period_left))
3810                 return;
3811
3812         perf_swevent_overflow(event, 0, nmi, data, regs);
3813 }
3814
3815 static int perf_swevent_is_counting(struct perf_event *event)
3816 {
3817         /*
3818          * The event is active, we're good!
3819          */
3820         if (event->state == PERF_EVENT_STATE_ACTIVE)
3821                 return 1;
3822
3823         /*
3824          * The event is off/error, not counting.
3825          */
3826         if (event->state != PERF_EVENT_STATE_INACTIVE)
3827                 return 0;
3828
3829         /*
3830          * The event is inactive, if the context is active
3831          * we're part of a group that didn't make it on the 'pmu',
3832          * not counting.
3833          */
3834         if (event->ctx->is_active)
3835                 return 0;
3836
3837         /*
3838          * We're inactive and the context is too, this means the
3839          * task is scheduled out, we're counting events that happen
3840          * to us, like migration events.
3841          */
3842         return 1;
3843 }
3844
3845 static int perf_tp_event_match(struct perf_event *event,
3846                                 struct perf_sample_data *data);
3847
3848 static int perf_exclude_event(struct perf_event *event,
3849                               struct pt_regs *regs)
3850 {
3851         if (regs) {
3852                 if (event->attr.exclude_user && user_mode(regs))
3853                         return 1;
3854
3855                 if (event->attr.exclude_kernel && !user_mode(regs))
3856                         return 1;
3857         }
3858
3859         return 0;
3860 }
3861
3862 static int perf_swevent_match(struct perf_event *event,
3863                                 enum perf_type_id type,
3864                                 u32 event_id,
3865                                 struct perf_sample_data *data,
3866                                 struct pt_regs *regs)
3867 {
3868         if (event->cpu != -1 && event->cpu != smp_processor_id())
3869                 return 0;
3870
3871         if (!perf_swevent_is_counting(event))
3872                 return 0;
3873
3874         if (event->attr.type != type)
3875                 return 0;
3876
3877         if (event->attr.config != event_id)
3878                 return 0;
3879
3880         if (perf_exclude_event(event, regs))
3881                 return 0;
3882
3883         if (event->attr.type == PERF_TYPE_TRACEPOINT &&
3884             !perf_tp_event_match(event, data))
3885                 return 0;
3886
3887         return 1;
3888 }
3889
3890 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
3891                                      enum perf_type_id type,
3892                                      u32 event_id, u64 nr, int nmi,
3893                                      struct perf_sample_data *data,
3894                                      struct pt_regs *regs)
3895 {
3896         struct perf_event *event;
3897
3898         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3899                 if (perf_swevent_match(event, type, event_id, data, regs))
3900                         perf_swevent_add(event, nr, nmi, data, regs);
3901         }
3902 }
3903
3904 int perf_swevent_get_recursion_context(void)
3905 {
3906         struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3907         int rctx;
3908
3909         if (in_nmi())
3910                 rctx = 3;
3911         else if (in_irq())
3912                 rctx = 2;
3913         else if (in_softirq())
3914                 rctx = 1;
3915         else
3916                 rctx = 0;
3917
3918         if (cpuctx->recursion[rctx]) {
3919                 put_cpu_var(perf_cpu_context);
3920                 return -1;
3921         }
3922
3923         cpuctx->recursion[rctx]++;
3924         barrier();
3925
3926         return rctx;
3927 }
3928 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
3929
3930 void perf_swevent_put_recursion_context(int rctx)
3931 {
3932         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
3933         barrier();
3934         cpuctx->recursion[rctx]--;
3935         put_cpu_var(perf_cpu_context);
3936 }
3937 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
3938
3939 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
3940                                     u64 nr, int nmi,
3941                                     struct perf_sample_data *data,
3942                                     struct pt_regs *regs)
3943 {
3944         struct perf_cpu_context *cpuctx;
3945         struct perf_event_context *ctx;
3946
3947         cpuctx = &__get_cpu_var(perf_cpu_context);
3948         rcu_read_lock();
3949         perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
3950                                  nr, nmi, data, regs);
3951         /*
3952          * doesn't really matter which of the child contexts the
3953          * events ends up in.
3954          */
3955         ctx = rcu_dereference(current->perf_event_ctxp);
3956         if (ctx)
3957                 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
3958         rcu_read_unlock();
3959 }
3960
3961 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
3962                             struct pt_regs *regs, u64 addr)
3963 {
3964         struct perf_sample_data data;
3965         int rctx;
3966
3967         rctx = perf_swevent_get_recursion_context();
3968         if (rctx < 0)
3969                 return;
3970
3971         data.addr = addr;
3972         data.raw  = NULL;
3973
3974         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
3975
3976         perf_swevent_put_recursion_context(rctx);
3977 }
3978
3979 static void perf_swevent_read(struct perf_event *event)
3980 {
3981 }
3982
3983 static int perf_swevent_enable(struct perf_event *event)
3984 {
3985         struct hw_perf_event *hwc = &event->hw;
3986
3987         if (hwc->sample_period) {
3988                 hwc->last_period = hwc->sample_period;
3989                 perf_swevent_set_period(event);
3990         }
3991         return 0;
3992 }
3993
3994 static void perf_swevent_disable(struct perf_event *event)
3995 {
3996 }
3997
3998 static const struct pmu perf_ops_generic = {
3999         .enable         = perf_swevent_enable,
4000         .disable        = perf_swevent_disable,
4001         .read           = perf_swevent_read,
4002         .unthrottle     = perf_swevent_unthrottle,
4003 };
4004
4005 /*
4006  * hrtimer based swevent callback
4007  */
4008
4009 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4010 {
4011         enum hrtimer_restart ret = HRTIMER_RESTART;
4012         struct perf_sample_data data;
4013         struct pt_regs *regs;
4014         struct perf_event *event;
4015         u64 period;
4016
4017         event   = container_of(hrtimer, struct perf_event, hw.hrtimer);
4018         event->pmu->read(event);
4019
4020         data.addr = 0;
4021         data.raw = NULL;
4022         data.period = event->hw.last_period;
4023         regs = get_irq_regs();
4024         /*
4025          * In case we exclude kernel IPs or are somehow not in interrupt
4026          * context, provide the next best thing, the user IP.
4027          */
4028         if ((event->attr.exclude_kernel || !regs) &&
4029                         !event->attr.exclude_user)
4030                 regs = task_pt_regs(current);
4031
4032         if (regs) {
4033                 if (!(event->attr.exclude_idle && current->pid == 0))
4034                         if (perf_event_overflow(event, 0, &data, regs))
4035                                 ret = HRTIMER_NORESTART;
4036         }
4037
4038         period = max_t(u64, 10000, event->hw.sample_period);
4039         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4040
4041         return ret;
4042 }
4043
4044 static void perf_swevent_start_hrtimer(struct perf_event *event)
4045 {
4046         struct hw_perf_event *hwc = &event->hw;
4047
4048         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4049         hwc->hrtimer.function = perf_swevent_hrtimer;
4050         if (hwc->sample_period) {
4051                 u64 period;
4052
4053                 if (hwc->remaining) {
4054                         if (hwc->remaining < 0)
4055                                 period = 10000;
4056                         else
4057                                 period = hwc->remaining;
4058                         hwc->remaining = 0;
4059                 } else {
4060                         period = max_t(u64, 10000, hwc->sample_period);
4061                 }
4062                 __hrtimer_start_range_ns(&hwc->hrtimer,
4063                                 ns_to_ktime(period), 0,
4064                                 HRTIMER_MODE_REL, 0);
4065         }
4066 }
4067
4068 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4069 {
4070         struct hw_perf_event *hwc = &event->hw;
4071
4072         if (hwc->sample_period) {
4073                 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4074                 hwc->remaining = ktime_to_ns(remaining);
4075
4076                 hrtimer_cancel(&hwc->hrtimer);
4077         }
4078 }
4079
4080 /*
4081  * Software event: cpu wall time clock
4082  */
4083
4084 static void cpu_clock_perf_event_update(struct perf_event *event)
4085 {
4086         int cpu = raw_smp_processor_id();
4087         s64 prev;
4088         u64 now;
4089
4090         now = cpu_clock(cpu);
4091         prev = atomic64_xchg(&event->hw.prev_count, now);
4092         atomic64_add(now - prev, &event->count);
4093 }
4094
4095 static int cpu_clock_perf_event_enable(struct perf_event *event)
4096 {
4097         struct hw_perf_event *hwc = &event->hw;
4098         int cpu = raw_smp_processor_id();
4099
4100         atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4101         perf_swevent_start_hrtimer(event);
4102
4103         return 0;
4104 }
4105
4106 static void cpu_clock_perf_event_disable(struct perf_event *event)
4107 {
4108         perf_swevent_cancel_hrtimer(event);
4109         cpu_clock_perf_event_update(event);
4110 }
4111
4112 static void cpu_clock_perf_event_read(struct perf_event *event)
4113 {
4114         cpu_clock_perf_event_update(event);
4115 }
4116
4117 static const struct pmu perf_ops_cpu_clock = {
4118         .enable         = cpu_clock_perf_event_enable,
4119         .disable        = cpu_clock_perf_event_disable,
4120         .read           = cpu_clock_perf_event_read,
4121 };
4122
4123 /*
4124  * Software event: task time clock
4125  */
4126
4127 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4128 {
4129         u64 prev;
4130         s64 delta;
4131
4132         prev = atomic64_xchg(&event->hw.prev_count, now);
4133         delta = now - prev;
4134         atomic64_add(delta, &event->count);
4135 }
4136
4137 static int task_clock_perf_event_enable(struct perf_event *event)
4138 {
4139         struct hw_perf_event *hwc = &event->hw;
4140         u64 now;
4141
4142         now = event->ctx->time;
4143
4144         atomic64_set(&hwc->prev_count, now);
4145
4146         perf_swevent_start_hrtimer(event);
4147
4148         return 0;
4149 }
4150
4151 static void task_clock_perf_event_disable(struct perf_event *event)
4152 {
4153         perf_swevent_cancel_hrtimer(event);
4154         task_clock_perf_event_update(event, event->ctx->time);
4155
4156 }
4157
4158 static void task_clock_perf_event_read(struct perf_event *event)
4159 {
4160         u64 time;
4161
4162         if (!in_nmi()) {
4163                 update_context_time(event->ctx);
4164                 time = event->ctx->time;
4165         } else {
4166                 u64 now = perf_clock();
4167                 u64 delta = now - event->ctx->timestamp;
4168                 time = event->ctx->time + delta;
4169         }
4170
4171         task_clock_perf_event_update(event, time);
4172 }
4173
4174 static const struct pmu perf_ops_task_clock = {
4175         .enable         = task_clock_perf_event_enable,
4176         .disable        = task_clock_perf_event_disable,
4177         .read           = task_clock_perf_event_read,
4178 };
4179
4180 #ifdef CONFIG_EVENT_TRACING
4181
4182 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4183                           int entry_size)
4184 {
4185         struct perf_raw_record raw = {
4186                 .size = entry_size,
4187                 .data = record,
4188         };
4189
4190         struct perf_sample_data data = {
4191                 .addr = addr,
4192                 .raw = &raw,
4193         };
4194
4195         struct pt_regs *regs = get_irq_regs();
4196
4197         if (!regs)
4198                 regs = task_pt_regs(current);
4199
4200         /* Trace events already protected against recursion */
4201         do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4202                                 &data, regs);
4203 }
4204 EXPORT_SYMBOL_GPL(perf_tp_event);
4205
4206 static int perf_tp_event_match(struct perf_event *event,
4207                                 struct perf_sample_data *data)
4208 {
4209         void *record = data->raw->data;
4210
4211         if (likely(!event->filter) || filter_match_preds(event->filter, record))
4212                 return 1;
4213         return 0;
4214 }
4215
4216 static void tp_perf_event_destroy(struct perf_event *event)
4217 {
4218         ftrace_profile_disable(event->attr.config);
4219 }
4220
4221 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4222 {
4223         /*
4224          * Raw tracepoint data is a severe data leak, only allow root to
4225          * have these.
4226          */
4227         if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4228                         perf_paranoid_tracepoint_raw() &&
4229                         !capable(CAP_SYS_ADMIN))
4230                 return ERR_PTR(-EPERM);
4231
4232         if (ftrace_profile_enable(event->attr.config))
4233                 return NULL;
4234
4235         event->destroy = tp_perf_event_destroy;
4236
4237         return &perf_ops_generic;
4238 }
4239
4240 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4241 {
4242         char *filter_str;
4243         int ret;
4244
4245         if (event->attr.type != PERF_TYPE_TRACEPOINT)
4246                 return -EINVAL;
4247
4248         filter_str = strndup_user(arg, PAGE_SIZE);
4249         if (IS_ERR(filter_str))
4250                 return PTR_ERR(filter_str);
4251
4252         ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4253
4254         kfree(filter_str);
4255         return ret;
4256 }
4257
4258 static void perf_event_free_filter(struct perf_event *event)
4259 {
4260         ftrace_profile_free_filter(event);
4261 }
4262
4263 #else
4264
4265 static int perf_tp_event_match(struct perf_event *event,
4266                                 struct perf_sample_data *data)
4267 {
4268         return 1;
4269 }
4270
4271 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4272 {
4273         return NULL;
4274 }
4275
4276 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4277 {
4278         return -ENOENT;
4279 }
4280
4281 static void perf_event_free_filter(struct perf_event *event)
4282 {
4283 }
4284
4285 #endif /* CONFIG_EVENT_TRACING */
4286
4287 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4288 static void bp_perf_event_destroy(struct perf_event *event)
4289 {
4290         release_bp_slot(event);
4291 }
4292
4293 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4294 {
4295         int err;
4296
4297         err = register_perf_hw_breakpoint(bp);
4298         if (err)
4299                 return ERR_PTR(err);
4300
4301         bp->destroy = bp_perf_event_destroy;
4302
4303         return &perf_ops_bp;
4304 }
4305
4306 void perf_bp_event(struct perf_event *bp, void *data)
4307 {
4308         struct perf_sample_data sample;
4309         struct pt_regs *regs = data;
4310
4311         sample.raw = NULL;
4312         sample.addr = bp->attr.bp_addr;
4313
4314         if (!perf_exclude_event(bp, regs))
4315                 perf_swevent_add(bp, 1, 1, &sample, regs);
4316 }
4317 #else
4318 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4319 {
4320         return NULL;
4321 }
4322
4323 void perf_bp_event(struct perf_event *bp, void *regs)
4324 {
4325 }
4326 #endif
4327
4328 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4329
4330 static void sw_perf_event_destroy(struct perf_event *event)
4331 {
4332         u64 event_id = event->attr.config;
4333
4334         WARN_ON(event->parent);
4335
4336         atomic_dec(&perf_swevent_enabled[event_id]);
4337 }
4338
4339 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4340 {
4341         const struct pmu *pmu = NULL;
4342         u64 event_id = event->attr.config;
4343
4344         /*
4345          * Software events (currently) can't in general distinguish
4346          * between user, kernel and hypervisor events.
4347          * However, context switches and cpu migrations are considered
4348          * to be kernel events, and page faults are never hypervisor
4349          * events.
4350          */
4351         switch (event_id) {
4352         case PERF_COUNT_SW_CPU_CLOCK:
4353                 pmu = &perf_ops_cpu_clock;
4354
4355                 break;
4356         case PERF_COUNT_SW_TASK_CLOCK:
4357                 /*
4358                  * If the user instantiates this as a per-cpu event,
4359                  * use the cpu_clock event instead.
4360                  */
4361                 if (event->ctx->task)
4362                         pmu = &perf_ops_task_clock;
4363                 else
4364                         pmu = &perf_ops_cpu_clock;
4365
4366                 break;
4367         case PERF_COUNT_SW_PAGE_FAULTS:
4368         case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4369         case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4370         case PERF_COUNT_SW_CONTEXT_SWITCHES:
4371         case PERF_COUNT_SW_CPU_MIGRATIONS:
4372         case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4373         case PERF_COUNT_SW_EMULATION_FAULTS:
4374                 if (!event->parent) {
4375                         atomic_inc(&perf_swevent_enabled[event_id]);
4376                         event->destroy = sw_perf_event_destroy;
4377                 }
4378                 pmu = &perf_ops_generic;
4379                 break;
4380         }
4381
4382         return pmu;
4383 }
4384
4385 /*
4386  * Allocate and initialize a event structure
4387  */
4388 static struct perf_event *
4389 perf_event_alloc(struct perf_event_attr *attr,
4390                    int cpu,
4391                    struct perf_event_context *ctx,
4392                    struct perf_event *group_leader,
4393                    struct perf_event *parent_event,
4394                    perf_overflow_handler_t overflow_handler,
4395                    gfp_t gfpflags)
4396 {
4397         const struct pmu *pmu;
4398         struct perf_event *event;
4399         struct hw_perf_event *hwc;
4400         long err;
4401
4402         event = kzalloc(sizeof(*event), gfpflags);
4403         if (!event)
4404                 return ERR_PTR(-ENOMEM);
4405
4406         /*
4407          * Single events are their own group leaders, with an
4408          * empty sibling list:
4409          */
4410         if (!group_leader)
4411                 group_leader = event;
4412
4413         mutex_init(&event->child_mutex);
4414         INIT_LIST_HEAD(&event->child_list);
4415
4416         INIT_LIST_HEAD(&event->group_entry);
4417         INIT_LIST_HEAD(&event->event_entry);
4418         INIT_LIST_HEAD(&event->sibling_list);
4419         init_waitqueue_head(&event->waitq);
4420
4421         mutex_init(&event->mmap_mutex);
4422
4423         event->cpu              = cpu;
4424         event->attr             = *attr;
4425         event->group_leader     = group_leader;
4426         event->pmu              = NULL;
4427         event->ctx              = ctx;
4428         event->oncpu            = -1;
4429
4430         event->parent           = parent_event;
4431
4432         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
4433         event->id               = atomic64_inc_return(&perf_event_id);
4434
4435         event->state            = PERF_EVENT_STATE_INACTIVE;
4436
4437         if (!overflow_handler && parent_event)
4438                 overflow_handler = parent_event->overflow_handler;
4439         
4440         event->overflow_handler = overflow_handler;
4441
4442         if (attr->disabled)
4443                 event->state = PERF_EVENT_STATE_OFF;
4444
4445         pmu = NULL;
4446
4447         hwc = &event->hw;
4448         hwc->sample_period = attr->sample_period;
4449         if (attr->freq && attr->sample_freq)
4450                 hwc->sample_period = 1;
4451         hwc->last_period = hwc->sample_period;
4452
4453         atomic64_set(&hwc->period_left, hwc->sample_period);
4454
4455         /*
4456          * we currently do not support PERF_FORMAT_GROUP on inherited events
4457          */
4458         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4459                 goto done;
4460
4461         switch (attr->type) {
4462         case PERF_TYPE_RAW:
4463         case PERF_TYPE_HARDWARE:
4464         case PERF_TYPE_HW_CACHE:
4465                 pmu = hw_perf_event_init(event);
4466                 break;
4467
4468         case PERF_TYPE_SOFTWARE:
4469                 pmu = sw_perf_event_init(event);
4470                 break;
4471
4472         case PERF_TYPE_TRACEPOINT:
4473                 pmu = tp_perf_event_init(event);
4474                 break;
4475
4476         case PERF_TYPE_BREAKPOINT:
4477                 pmu = bp_perf_event_init(event);
4478                 break;
4479
4480
4481         default:
4482                 break;
4483         }
4484 done:
4485         err = 0;
4486         if (!pmu)
4487                 err = -EINVAL;
4488         else if (IS_ERR(pmu))
4489                 err = PTR_ERR(pmu);
4490
4491         if (err) {
4492                 if (event->ns)
4493                         put_pid_ns(event->ns);
4494                 kfree(event);
4495                 return ERR_PTR(err);
4496         }
4497
4498         event->pmu = pmu;
4499
4500         if (!event->parent) {
4501                 atomic_inc(&nr_events);
4502                 if (event->attr.mmap)
4503                         atomic_inc(&nr_mmap_events);
4504                 if (event->attr.comm)
4505                         atomic_inc(&nr_comm_events);
4506                 if (event->attr.task)
4507                         atomic_inc(&nr_task_events);
4508         }
4509
4510         return event;
4511 }
4512
4513 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4514                           struct perf_event_attr *attr)
4515 {
4516         u32 size;
4517         int ret;
4518
4519         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4520                 return -EFAULT;
4521
4522         /*
4523          * zero the full structure, so that a short copy will be nice.
4524          */
4525         memset(attr, 0, sizeof(*attr));
4526
4527         ret = get_user(size, &uattr->size);
4528         if (ret)
4529                 return ret;
4530
4531         if (size > PAGE_SIZE)   /* silly large */
4532                 goto err_size;
4533
4534         if (!size)              /* abi compat */
4535                 size = PERF_ATTR_SIZE_VER0;
4536
4537         if (size < PERF_ATTR_SIZE_VER0)
4538                 goto err_size;
4539
4540         /*
4541          * If we're handed a bigger struct than we know of,
4542          * ensure all the unknown bits are 0 - i.e. new
4543          * user-space does not rely on any kernel feature
4544          * extensions we dont know about yet.
4545          */
4546         if (size > sizeof(*attr)) {
4547                 unsigned char __user *addr;
4548                 unsigned char __user *end;
4549                 unsigned char val;
4550
4551                 addr = (void __user *)uattr + sizeof(*attr);
4552                 end  = (void __user *)uattr + size;
4553
4554                 for (; addr < end; addr++) {
4555                         ret = get_user(val, addr);
4556                         if (ret)
4557                                 return ret;
4558                         if (val)
4559                                 goto err_size;
4560                 }
4561                 size = sizeof(*attr);
4562         }
4563
4564         ret = copy_from_user(attr, uattr, size);
4565         if (ret)
4566                 return -EFAULT;
4567
4568         /*
4569          * If the type exists, the corresponding creation will verify
4570          * the attr->config.
4571          */
4572         if (attr->type >= PERF_TYPE_MAX)
4573                 return -EINVAL;
4574
4575         if (attr->__reserved_1 || attr->__reserved_2)
4576                 return -EINVAL;
4577
4578         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4579                 return -EINVAL;
4580
4581         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4582                 return -EINVAL;
4583
4584 out:
4585         return ret;
4586
4587 err_size:
4588         put_user(sizeof(*attr), &uattr->size);
4589         ret = -E2BIG;
4590         goto out;
4591 }
4592
4593 static int perf_event_set_output(struct perf_event *event, int output_fd)
4594 {
4595         struct perf_event *output_event = NULL;
4596         struct file *output_file = NULL;
4597         struct perf_event *old_output;
4598         int fput_needed = 0;
4599         int ret = -EINVAL;
4600
4601         if (!output_fd)
4602                 goto set;
4603
4604         output_file = fget_light(output_fd, &fput_needed);
4605         if (!output_file)
4606                 return -EBADF;
4607
4608         if (output_file->f_op != &perf_fops)
4609                 goto out;
4610
4611         output_event = output_file->private_data;
4612
4613         /* Don't chain output fds */
4614         if (output_event->output)
4615                 goto out;
4616
4617         /* Don't set an output fd when we already have an output channel */
4618         if (event->data)
4619                 goto out;
4620
4621         atomic_long_inc(&output_file->f_count);
4622
4623 set:
4624         mutex_lock(&event->mmap_mutex);
4625         old_output = event->output;
4626         rcu_assign_pointer(event->output, output_event);
4627         mutex_unlock(&event->mmap_mutex);
4628
4629         if (old_output) {
4630                 /*
4631                  * we need to make sure no existing perf_output_*()
4632                  * is still referencing this event.
4633                  */
4634                 synchronize_rcu();
4635                 fput(old_output->filp);
4636         }
4637
4638         ret = 0;
4639 out:
4640         fput_light(output_file, fput_needed);
4641         return ret;
4642 }
4643
4644 /**
4645  * sys_perf_event_open - open a performance event, associate it to a task/cpu
4646  *
4647  * @attr_uptr:  event_id type attributes for monitoring/sampling
4648  * @pid:                target pid
4649  * @cpu:                target cpu
4650  * @group_fd:           group leader event fd
4651  */
4652 SYSCALL_DEFINE5(perf_event_open,
4653                 struct perf_event_attr __user *, attr_uptr,
4654                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4655 {
4656         struct perf_event *event, *group_leader;
4657         struct perf_event_attr attr;
4658         struct perf_event_context *ctx;
4659         struct file *event_file = NULL;
4660         struct file *group_file = NULL;
4661         int fput_needed = 0;
4662         int fput_needed2 = 0;
4663         int err;
4664
4665         /* for future expandability... */
4666         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4667                 return -EINVAL;
4668
4669         err = perf_copy_attr(attr_uptr, &attr);
4670         if (err)
4671                 return err;
4672
4673         if (!attr.exclude_kernel) {
4674                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4675                         return -EACCES;
4676         }
4677
4678         if (attr.freq) {
4679                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4680                         return -EINVAL;
4681         }
4682
4683         /*
4684          * Get the target context (task or percpu):
4685          */
4686         ctx = find_get_context(pid, cpu);
4687         if (IS_ERR(ctx))
4688                 return PTR_ERR(ctx);
4689
4690         /*
4691          * Look up the group leader (we will attach this event to it):
4692          */
4693         group_leader = NULL;
4694         if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4695                 err = -EINVAL;
4696                 group_file = fget_light(group_fd, &fput_needed);
4697                 if (!group_file)
4698                         goto err_put_context;
4699                 if (group_file->f_op != &perf_fops)
4700                         goto err_put_context;
4701
4702                 group_leader = group_file->private_data;
4703                 /*
4704                  * Do not allow a recursive hierarchy (this new sibling
4705                  * becoming part of another group-sibling):
4706                  */
4707                 if (group_leader->group_leader != group_leader)
4708                         goto err_put_context;
4709                 /*
4710                  * Do not allow to attach to a group in a different
4711                  * task or CPU context:
4712                  */
4713                 if (group_leader->ctx != ctx)
4714                         goto err_put_context;
4715                 /*
4716                  * Only a group leader can be exclusive or pinned
4717                  */
4718                 if (attr.exclusive || attr.pinned)
4719                         goto err_put_context;
4720         }
4721
4722         event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4723                                      NULL, NULL, GFP_KERNEL);
4724         err = PTR_ERR(event);
4725         if (IS_ERR(event))
4726                 goto err_put_context;
4727
4728         err = anon_inode_getfd("[perf_event]", &perf_fops, event, O_RDWR);
4729         if (err < 0)
4730                 goto err_free_put_context;
4731
4732         event_file = fget_light(err, &fput_needed2);
4733         if (!event_file)
4734                 goto err_free_put_context;
4735
4736         if (flags & PERF_FLAG_FD_OUTPUT) {
4737                 err = perf_event_set_output(event, group_fd);
4738                 if (err)
4739                         goto err_fput_free_put_context;
4740         }
4741
4742         event->filp = event_file;
4743         WARN_ON_ONCE(ctx->parent_ctx);
4744         mutex_lock(&ctx->mutex);
4745         perf_install_in_context(ctx, event, cpu);
4746         ++ctx->generation;
4747         mutex_unlock(&ctx->mutex);
4748
4749         event->owner = current;
4750         get_task_struct(current);
4751         mutex_lock(&current->perf_event_mutex);
4752         list_add_tail(&event->owner_entry, &current->perf_event_list);
4753         mutex_unlock(&current->perf_event_mutex);
4754
4755 err_fput_free_put_context:
4756         fput_light(event_file, fput_needed2);
4757
4758 err_free_put_context:
4759         if (err < 0)
4760                 kfree(event);
4761
4762 err_put_context:
4763         if (err < 0)
4764                 put_ctx(ctx);
4765
4766         fput_light(group_file, fput_needed);
4767
4768         return err;
4769 }
4770
4771 /**
4772  * perf_event_create_kernel_counter
4773  *
4774  * @attr: attributes of the counter to create
4775  * @cpu: cpu in which the counter is bound
4776  * @pid: task to profile
4777  */
4778 struct perf_event *
4779 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4780                                  pid_t pid,
4781                                  perf_overflow_handler_t overflow_handler)
4782 {
4783         struct perf_event *event;
4784         struct perf_event_context *ctx;
4785         int err;
4786
4787         /*
4788          * Get the target context (task or percpu):
4789          */
4790
4791         ctx = find_get_context(pid, cpu);
4792         if (IS_ERR(ctx)) {
4793                 err = PTR_ERR(ctx);
4794                 goto err_exit;
4795         }
4796
4797         event = perf_event_alloc(attr, cpu, ctx, NULL,
4798                                  NULL, overflow_handler, GFP_KERNEL);
4799         if (IS_ERR(event)) {
4800                 err = PTR_ERR(event);
4801                 goto err_put_context;
4802         }
4803
4804         event->filp = NULL;
4805         WARN_ON_ONCE(ctx->parent_ctx);
4806         mutex_lock(&ctx->mutex);
4807         perf_install_in_context(ctx, event, cpu);
4808         ++ctx->generation;
4809         mutex_unlock(&ctx->mutex);
4810
4811         event->owner = current;
4812         get_task_struct(current);
4813         mutex_lock(&current->perf_event_mutex);
4814         list_add_tail(&event->owner_entry, &current->perf_event_list);
4815         mutex_unlock(&current->perf_event_mutex);
4816
4817         return event;
4818
4819  err_put_context:
4820         put_ctx(ctx);
4821  err_exit:
4822         return ERR_PTR(err);
4823 }
4824 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4825
4826 /*
4827  * inherit a event from parent task to child task:
4828  */
4829 static struct perf_event *
4830 inherit_event(struct perf_event *parent_event,
4831               struct task_struct *parent,
4832               struct perf_event_context *parent_ctx,
4833               struct task_struct *child,
4834               struct perf_event *group_leader,
4835               struct perf_event_context *child_ctx)
4836 {
4837         struct perf_event *child_event;
4838
4839         /*
4840          * Instead of creating recursive hierarchies of events,
4841          * we link inherited events back to the original parent,
4842          * which has a filp for sure, which we use as the reference
4843          * count:
4844          */
4845         if (parent_event->parent)
4846                 parent_event = parent_event->parent;
4847
4848         child_event = perf_event_alloc(&parent_event->attr,
4849                                            parent_event->cpu, child_ctx,
4850                                            group_leader, parent_event,
4851                                            NULL, GFP_KERNEL);
4852         if (IS_ERR(child_event))
4853                 return child_event;
4854         get_ctx(child_ctx);
4855
4856         /*
4857          * Make the child state follow the state of the parent event,
4858          * not its attr.disabled bit.  We hold the parent's mutex,
4859          * so we won't race with perf_event_{en, dis}able_family.
4860          */
4861         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
4862                 child_event->state = PERF_EVENT_STATE_INACTIVE;
4863         else
4864                 child_event->state = PERF_EVENT_STATE_OFF;
4865
4866         if (parent_event->attr.freq)
4867                 child_event->hw.sample_period = parent_event->hw.sample_period;
4868
4869         child_event->overflow_handler = parent_event->overflow_handler;
4870
4871         /*
4872          * Link it up in the child's context:
4873          */
4874         add_event_to_ctx(child_event, child_ctx);
4875
4876         /*
4877          * Get a reference to the parent filp - we will fput it
4878          * when the child event exits. This is safe to do because
4879          * we are in the parent and we know that the filp still
4880          * exists and has a nonzero count:
4881          */
4882         atomic_long_inc(&parent_event->filp->f_count);
4883
4884         /*
4885          * Link this into the parent event's child list
4886          */
4887         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4888         mutex_lock(&parent_event->child_mutex);
4889         list_add_tail(&child_event->child_list, &parent_event->child_list);
4890         mutex_unlock(&parent_event->child_mutex);
4891
4892         return child_event;
4893 }
4894
4895 static int inherit_group(struct perf_event *parent_event,
4896               struct task_struct *parent,
4897               struct perf_event_context *parent_ctx,
4898               struct task_struct *child,
4899               struct perf_event_context *child_ctx)
4900 {
4901         struct perf_event *leader;
4902         struct perf_event *sub;
4903         struct perf_event *child_ctr;
4904
4905         leader = inherit_event(parent_event, parent, parent_ctx,
4906                                  child, NULL, child_ctx);
4907         if (IS_ERR(leader))
4908                 return PTR_ERR(leader);
4909         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
4910                 child_ctr = inherit_event(sub, parent, parent_ctx,
4911                                             child, leader, child_ctx);
4912                 if (IS_ERR(child_ctr))
4913                         return PTR_ERR(child_ctr);
4914         }
4915         return 0;
4916 }
4917
4918 static void sync_child_event(struct perf_event *child_event,
4919                                struct task_struct *child)
4920 {
4921         struct perf_event *parent_event = child_event->parent;
4922         u64 child_val;
4923
4924         if (child_event->attr.inherit_stat)
4925                 perf_event_read_event(child_event, child);
4926
4927         child_val = atomic64_read(&child_event->count);
4928
4929         /*
4930          * Add back the child's count to the parent's count:
4931          */
4932         atomic64_add(child_val, &parent_event->count);
4933         atomic64_add(child_event->total_time_enabled,
4934                      &parent_event->child_total_time_enabled);
4935         atomic64_add(child_event->total_time_running,
4936                      &parent_event->child_total_time_running);
4937
4938         /*
4939          * Remove this event from the parent's list
4940          */
4941         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4942         mutex_lock(&parent_event->child_mutex);
4943         list_del_init(&child_event->child_list);
4944         mutex_unlock(&parent_event->child_mutex);
4945
4946         /*
4947          * Release the parent event, if this was the last
4948          * reference to it.
4949          */
4950         fput(parent_event->filp);
4951 }
4952
4953 static void
4954 __perf_event_exit_task(struct perf_event *child_event,
4955                          struct perf_event_context *child_ctx,
4956                          struct task_struct *child)
4957 {
4958         struct perf_event *parent_event;
4959
4960         perf_event_remove_from_context(child_event);
4961
4962         parent_event = child_event->parent;
4963         /*
4964          * It can happen that parent exits first, and has events
4965          * that are still around due to the child reference. These
4966          * events need to be zapped - but otherwise linger.
4967          */
4968         if (parent_event) {
4969                 sync_child_event(child_event, child);
4970                 free_event(child_event);
4971         }
4972 }
4973
4974 /*
4975  * When a child task exits, feed back event values to parent events.
4976  */
4977 void perf_event_exit_task(struct task_struct *child)
4978 {
4979         struct perf_event *child_event, *tmp;
4980         struct perf_event_context *child_ctx;
4981         unsigned long flags;
4982
4983         if (likely(!child->perf_event_ctxp)) {
4984                 perf_event_task(child, NULL, 0);
4985                 return;
4986         }
4987
4988         local_irq_save(flags);
4989         /*
4990          * We can't reschedule here because interrupts are disabled,
4991          * and either child is current or it is a task that can't be
4992          * scheduled, so we are now safe from rescheduling changing
4993          * our context.
4994          */
4995         child_ctx = child->perf_event_ctxp;
4996         __perf_event_task_sched_out(child_ctx);
4997
4998         /*
4999          * Take the context lock here so that if find_get_context is
5000          * reading child->perf_event_ctxp, we wait until it has
5001          * incremented the context's refcount before we do put_ctx below.
5002          */
5003         raw_spin_lock(&child_ctx->lock);
5004         child->perf_event_ctxp = NULL;
5005         /*
5006          * If this context is a clone; unclone it so it can't get
5007          * swapped to another process while we're removing all
5008          * the events from it.
5009          */
5010         unclone_ctx(child_ctx);
5011         update_context_time(child_ctx);
5012         raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5013
5014         /*
5015          * Report the task dead after unscheduling the events so that we
5016          * won't get any samples after PERF_RECORD_EXIT. We can however still
5017          * get a few PERF_RECORD_READ events.
5018          */
5019         perf_event_task(child, child_ctx, 0);
5020
5021         /*
5022          * We can recurse on the same lock type through:
5023          *
5024          *   __perf_event_exit_task()
5025          *     sync_child_event()
5026          *       fput(parent_event->filp)
5027          *         perf_release()
5028          *           mutex_lock(&ctx->mutex)
5029          *
5030          * But since its the parent context it won't be the same instance.
5031          */
5032         mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
5033
5034 again:
5035         list_for_each_entry_safe(child_event, tmp, &child_ctx->group_list,
5036                                  group_entry)
5037                 __perf_event_exit_task(child_event, child_ctx, child);
5038
5039         /*
5040          * If the last event was a group event, it will have appended all
5041          * its siblings to the list, but we obtained 'tmp' before that which
5042          * will still point to the list head terminating the iteration.
5043          */
5044         if (!list_empty(&child_ctx->group_list))
5045                 goto again;
5046
5047         mutex_unlock(&child_ctx->mutex);
5048
5049         put_ctx(child_ctx);
5050 }
5051
5052 /*
5053  * free an unexposed, unused context as created by inheritance by
5054  * init_task below, used by fork() in case of fail.
5055  */
5056 void perf_event_free_task(struct task_struct *task)
5057 {
5058         struct perf_event_context *ctx = task->perf_event_ctxp;
5059         struct perf_event *event, *tmp;
5060
5061         if (!ctx)
5062                 return;
5063
5064         mutex_lock(&ctx->mutex);
5065 again:
5066         list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry) {
5067                 struct perf_event *parent = event->parent;
5068
5069                 if (WARN_ON_ONCE(!parent))
5070                         continue;
5071
5072                 mutex_lock(&parent->child_mutex);
5073                 list_del_init(&event->child_list);
5074                 mutex_unlock(&parent->child_mutex);
5075
5076                 fput(parent->filp);
5077
5078                 list_del_event(event, ctx);
5079                 free_event(event);
5080         }
5081
5082         if (!list_empty(&ctx->group_list))
5083                 goto again;
5084
5085         mutex_unlock(&ctx->mutex);
5086
5087         put_ctx(ctx);
5088 }
5089
5090 /*
5091  * Initialize the perf_event context in task_struct
5092  */
5093 int perf_event_init_task(struct task_struct *child)
5094 {
5095         struct perf_event_context *child_ctx = NULL, *parent_ctx;
5096         struct perf_event_context *cloned_ctx;
5097         struct perf_event *event;
5098         struct task_struct *parent = current;
5099         int inherited_all = 1;
5100         int ret = 0;
5101
5102         child->perf_event_ctxp = NULL;
5103
5104         mutex_init(&child->perf_event_mutex);
5105         INIT_LIST_HEAD(&child->perf_event_list);
5106
5107         if (likely(!parent->perf_event_ctxp))
5108                 return 0;
5109
5110         /*
5111          * If the parent's context is a clone, pin it so it won't get
5112          * swapped under us.
5113          */
5114         parent_ctx = perf_pin_task_context(parent);
5115
5116         /*
5117          * No need to check if parent_ctx != NULL here; since we saw
5118          * it non-NULL earlier, the only reason for it to become NULL
5119          * is if we exit, and since we're currently in the middle of
5120          * a fork we can't be exiting at the same time.
5121          */
5122
5123         /*
5124          * Lock the parent list. No need to lock the child - not PID
5125          * hashed yet and not running, so nobody can access it.
5126          */
5127         mutex_lock(&parent_ctx->mutex);
5128
5129         /*
5130          * We dont have to disable NMIs - we are only looking at
5131          * the list, not manipulating it:
5132          */
5133         list_for_each_entry(event, &parent_ctx->group_list, group_entry) {
5134
5135                 if (!event->attr.inherit) {
5136                         inherited_all = 0;
5137                         continue;
5138                 }
5139
5140                 if (!child->perf_event_ctxp) {
5141                         /*
5142                          * This is executed from the parent task context, so
5143                          * inherit events that have been marked for cloning.
5144                          * First allocate and initialize a context for the
5145                          * child.
5146                          */
5147
5148                         child_ctx = kzalloc(sizeof(struct perf_event_context),
5149                                             GFP_KERNEL);
5150                         if (!child_ctx) {
5151                                 ret = -ENOMEM;
5152                                 break;
5153                         }
5154
5155                         __perf_event_init_context(child_ctx, child);
5156                         child->perf_event_ctxp = child_ctx;
5157                         get_task_struct(child);
5158                 }
5159
5160                 ret = inherit_group(event, parent, parent_ctx,
5161                                              child, child_ctx);
5162                 if (ret) {
5163                         inherited_all = 0;
5164                         break;
5165                 }
5166         }
5167
5168         if (child_ctx && inherited_all) {
5169                 /*
5170                  * Mark the child context as a clone of the parent
5171                  * context, or of whatever the parent is a clone of.
5172                  * Note that if the parent is a clone, it could get
5173                  * uncloned at any point, but that doesn't matter
5174                  * because the list of events and the generation
5175                  * count can't have changed since we took the mutex.
5176                  */
5177                 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5178                 if (cloned_ctx) {
5179                         child_ctx->parent_ctx = cloned_ctx;
5180                         child_ctx->parent_gen = parent_ctx->parent_gen;
5181                 } else {
5182                         child_ctx->parent_ctx = parent_ctx;
5183                         child_ctx->parent_gen = parent_ctx->generation;
5184                 }
5185                 get_ctx(child_ctx->parent_ctx);
5186         }
5187
5188         mutex_unlock(&parent_ctx->mutex);
5189
5190         perf_unpin_context(parent_ctx);
5191
5192         return ret;
5193 }
5194
5195 static void __cpuinit perf_event_init_cpu(int cpu)
5196 {
5197         struct perf_cpu_context *cpuctx;
5198
5199         cpuctx = &per_cpu(perf_cpu_context, cpu);
5200         __perf_event_init_context(&cpuctx->ctx, NULL);
5201
5202         spin_lock(&perf_resource_lock);
5203         cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5204         spin_unlock(&perf_resource_lock);
5205
5206         hw_perf_event_setup(cpu);
5207 }
5208
5209 #ifdef CONFIG_HOTPLUG_CPU
5210 static void __perf_event_exit_cpu(void *info)
5211 {
5212         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5213         struct perf_event_context *ctx = &cpuctx->ctx;
5214         struct perf_event *event, *tmp;
5215
5216         list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry)
5217                 __perf_event_remove_from_context(event);
5218 }
5219 static void perf_event_exit_cpu(int cpu)
5220 {
5221         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5222         struct perf_event_context *ctx = &cpuctx->ctx;
5223
5224         mutex_lock(&ctx->mutex);
5225         smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5226         mutex_unlock(&ctx->mutex);
5227 }
5228 #else
5229 static inline void perf_event_exit_cpu(int cpu) { }
5230 #endif
5231
5232 static int __cpuinit
5233 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5234 {
5235         unsigned int cpu = (long)hcpu;
5236
5237         switch (action) {
5238
5239         case CPU_UP_PREPARE:
5240         case CPU_UP_PREPARE_FROZEN:
5241                 perf_event_init_cpu(cpu);
5242                 break;
5243
5244         case CPU_ONLINE:
5245         case CPU_ONLINE_FROZEN:
5246                 hw_perf_event_setup_online(cpu);
5247                 break;
5248
5249         case CPU_DOWN_PREPARE:
5250         case CPU_DOWN_PREPARE_FROZEN:
5251                 perf_event_exit_cpu(cpu);
5252                 break;
5253
5254         default:
5255                 break;
5256         }
5257
5258         return NOTIFY_OK;
5259 }
5260
5261 /*
5262  * This has to have a higher priority than migration_notifier in sched.c.
5263  */
5264 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5265         .notifier_call          = perf_cpu_notify,
5266         .priority               = 20,
5267 };
5268
5269 void __init perf_event_init(void)
5270 {
5271         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5272                         (void *)(long)smp_processor_id());
5273         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5274                         (void *)(long)smp_processor_id());
5275         register_cpu_notifier(&perf_cpu_nb);
5276 }
5277
5278 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
5279 {
5280         return sprintf(buf, "%d\n", perf_reserved_percpu);
5281 }
5282
5283 static ssize_t
5284 perf_set_reserve_percpu(struct sysdev_class *class,
5285                         const char *buf,
5286                         size_t count)
5287 {
5288         struct perf_cpu_context *cpuctx;
5289         unsigned long val;
5290         int err, cpu, mpt;
5291
5292         err = strict_strtoul(buf, 10, &val);
5293         if (err)
5294                 return err;
5295         if (val > perf_max_events)
5296                 return -EINVAL;
5297
5298         spin_lock(&perf_resource_lock);
5299         perf_reserved_percpu = val;
5300         for_each_online_cpu(cpu) {
5301                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5302                 raw_spin_lock_irq(&cpuctx->ctx.lock);
5303                 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5304                           perf_max_events - perf_reserved_percpu);
5305                 cpuctx->max_pertask = mpt;
5306                 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5307         }
5308         spin_unlock(&perf_resource_lock);
5309
5310         return count;
5311 }
5312
5313 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
5314 {
5315         return sprintf(buf, "%d\n", perf_overcommit);
5316 }
5317
5318 static ssize_t
5319 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
5320 {
5321         unsigned long val;
5322         int err;
5323
5324         err = strict_strtoul(buf, 10, &val);
5325         if (err)
5326                 return err;
5327         if (val > 1)
5328                 return -EINVAL;
5329
5330         spin_lock(&perf_resource_lock);
5331         perf_overcommit = val;
5332         spin_unlock(&perf_resource_lock);
5333
5334         return count;
5335 }
5336
5337 static SYSDEV_CLASS_ATTR(
5338                                 reserve_percpu,
5339                                 0644,
5340                                 perf_show_reserve_percpu,
5341                                 perf_set_reserve_percpu
5342                         );
5343
5344 static SYSDEV_CLASS_ATTR(
5345                                 overcommit,
5346                                 0644,
5347                                 perf_show_overcommit,
5348                                 perf_set_overcommit
5349                         );
5350
5351 static struct attribute *perfclass_attrs[] = {
5352         &attr_reserve_percpu.attr,
5353         &attr_overcommit.attr,
5354         NULL
5355 };
5356
5357 static struct attribute_group perfclass_attr_group = {
5358         .attrs                  = perfclass_attrs,
5359         .name                   = "perf_events",
5360 };
5361
5362 static int __init perf_event_sysfs_init(void)
5363 {
5364         return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5365                                   &perfclass_attr_group);
5366 }
5367 device_initcall(perf_event_sysfs_init);