perf_events, x86: AMD event scheduling
[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 void __weak hw_perf_event_setup_offline(int cpu)        { barrier(); }
102
103 int __weak
104 hw_perf_group_sched_in(struct perf_event *group_leader,
105                struct perf_cpu_context *cpuctx,
106                struct perf_event_context *ctx, int cpu)
107 {
108         return 0;
109 }
110
111 void __weak perf_event_print_debug(void)        { }
112
113 static DEFINE_PER_CPU(int, perf_disable_count);
114
115 void __perf_disable(void)
116 {
117         __get_cpu_var(perf_disable_count)++;
118 }
119
120 bool __perf_enable(void)
121 {
122         return !--__get_cpu_var(perf_disable_count);
123 }
124
125 void perf_disable(void)
126 {
127         __perf_disable();
128         hw_perf_disable();
129 }
130
131 void perf_enable(void)
132 {
133         if (__perf_enable())
134                 hw_perf_enable();
135 }
136
137 static void get_ctx(struct perf_event_context *ctx)
138 {
139         WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
140 }
141
142 static void free_ctx(struct rcu_head *head)
143 {
144         struct perf_event_context *ctx;
145
146         ctx = container_of(head, struct perf_event_context, rcu_head);
147         kfree(ctx);
148 }
149
150 static void put_ctx(struct perf_event_context *ctx)
151 {
152         if (atomic_dec_and_test(&ctx->refcount)) {
153                 if (ctx->parent_ctx)
154                         put_ctx(ctx->parent_ctx);
155                 if (ctx->task)
156                         put_task_struct(ctx->task);
157                 call_rcu(&ctx->rcu_head, free_ctx);
158         }
159 }
160
161 static void unclone_ctx(struct perf_event_context *ctx)
162 {
163         if (ctx->parent_ctx) {
164                 put_ctx(ctx->parent_ctx);
165                 ctx->parent_ctx = NULL;
166         }
167 }
168
169 /*
170  * If we inherit events we want to return the parent event id
171  * to userspace.
172  */
173 static u64 primary_event_id(struct perf_event *event)
174 {
175         u64 id = event->id;
176
177         if (event->parent)
178                 id = event->parent->id;
179
180         return id;
181 }
182
183 /*
184  * Get the perf_event_context for a task and lock it.
185  * This has to cope with with the fact that until it is locked,
186  * the context could get moved to another task.
187  */
188 static struct perf_event_context *
189 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
190 {
191         struct perf_event_context *ctx;
192
193         rcu_read_lock();
194  retry:
195         ctx = rcu_dereference(task->perf_event_ctxp);
196         if (ctx) {
197                 /*
198                  * If this context is a clone of another, it might
199                  * get swapped for another underneath us by
200                  * perf_event_task_sched_out, though the
201                  * rcu_read_lock() protects us from any context
202                  * getting freed.  Lock the context and check if it
203                  * got swapped before we could get the lock, and retry
204                  * if so.  If we locked the right context, then it
205                  * can't get swapped on us any more.
206                  */
207                 raw_spin_lock_irqsave(&ctx->lock, *flags);
208                 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
209                         raw_spin_unlock_irqrestore(&ctx->lock, *flags);
210                         goto retry;
211                 }
212
213                 if (!atomic_inc_not_zero(&ctx->refcount)) {
214                         raw_spin_unlock_irqrestore(&ctx->lock, *flags);
215                         ctx = NULL;
216                 }
217         }
218         rcu_read_unlock();
219         return ctx;
220 }
221
222 /*
223  * Get the context for a task and increment its pin_count so it
224  * can't get swapped to another task.  This also increments its
225  * reference count so that the context can't get freed.
226  */
227 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
228 {
229         struct perf_event_context *ctx;
230         unsigned long flags;
231
232         ctx = perf_lock_task_context(task, &flags);
233         if (ctx) {
234                 ++ctx->pin_count;
235                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
236         }
237         return ctx;
238 }
239
240 static void perf_unpin_context(struct perf_event_context *ctx)
241 {
242         unsigned long flags;
243
244         raw_spin_lock_irqsave(&ctx->lock, flags);
245         --ctx->pin_count;
246         raw_spin_unlock_irqrestore(&ctx->lock, flags);
247         put_ctx(ctx);
248 }
249
250 static inline u64 perf_clock(void)
251 {
252         return cpu_clock(smp_processor_id());
253 }
254
255 /*
256  * Update the record of the current time in a context.
257  */
258 static void update_context_time(struct perf_event_context *ctx)
259 {
260         u64 now = perf_clock();
261
262         ctx->time += now - ctx->timestamp;
263         ctx->timestamp = now;
264 }
265
266 /*
267  * Update the total_time_enabled and total_time_running fields for a event.
268  */
269 static void update_event_times(struct perf_event *event)
270 {
271         struct perf_event_context *ctx = event->ctx;
272         u64 run_end;
273
274         if (event->state < PERF_EVENT_STATE_INACTIVE ||
275             event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
276                 return;
277
278         if (ctx->is_active)
279                 run_end = ctx->time;
280         else
281                 run_end = event->tstamp_stopped;
282
283         event->total_time_enabled = run_end - event->tstamp_enabled;
284
285         if (event->state == PERF_EVENT_STATE_INACTIVE)
286                 run_end = event->tstamp_stopped;
287         else
288                 run_end = ctx->time;
289
290         event->total_time_running = run_end - event->tstamp_running;
291 }
292
293 static struct list_head *
294 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
295 {
296         if (event->attr.pinned)
297                 return &ctx->pinned_groups;
298         else
299                 return &ctx->flexible_groups;
300 }
301
302 /*
303  * Add a event from the lists for its context.
304  * Must be called with ctx->mutex and ctx->lock held.
305  */
306 static void
307 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
308 {
309         struct perf_event *group_leader = event->group_leader;
310
311         /*
312          * Depending on whether it is a standalone or sibling event,
313          * add it straight to the context's event list, or to the group
314          * leader's sibling list:
315          */
316         if (group_leader == event) {
317                 struct list_head *list;
318
319                 if (is_software_event(event))
320                         event->group_flags |= PERF_GROUP_SOFTWARE;
321
322                 list = ctx_group_list(event, ctx);
323                 list_add_tail(&event->group_entry, list);
324         } else {
325                 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
326                     !is_software_event(event))
327                         group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
328
329                 list_add_tail(&event->group_entry, &group_leader->sibling_list);
330                 group_leader->nr_siblings++;
331         }
332
333         list_add_rcu(&event->event_entry, &ctx->event_list);
334         ctx->nr_events++;
335         if (event->attr.inherit_stat)
336                 ctx->nr_stat++;
337 }
338
339 /*
340  * Remove a event from the lists for its context.
341  * Must be called with ctx->mutex and ctx->lock held.
342  */
343 static void
344 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
345 {
346         struct perf_event *sibling, *tmp;
347
348         if (list_empty(&event->group_entry))
349                 return;
350         ctx->nr_events--;
351         if (event->attr.inherit_stat)
352                 ctx->nr_stat--;
353
354         list_del_init(&event->group_entry);
355         list_del_rcu(&event->event_entry);
356
357         if (event->group_leader != event)
358                 event->group_leader->nr_siblings--;
359
360         update_event_times(event);
361
362         /*
363          * If event was in error state, then keep it
364          * that way, otherwise bogus counts will be
365          * returned on read(). The only way to get out
366          * of error state is by explicit re-enabling
367          * of the event
368          */
369         if (event->state > PERF_EVENT_STATE_OFF)
370                 event->state = PERF_EVENT_STATE_OFF;
371
372         /*
373          * If this was a group event with sibling events then
374          * upgrade the siblings to singleton events by adding them
375          * to the context list directly:
376          */
377         list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
378                 struct list_head *list;
379
380                 list = ctx_group_list(event, ctx);
381                 list_move_tail(&sibling->group_entry, list);
382                 sibling->group_leader = sibling;
383
384                 /* Inherit group flags from the previous leader */
385                 sibling->group_flags = event->group_flags;
386         }
387 }
388
389 static void
390 event_sched_out(struct perf_event *event,
391                   struct perf_cpu_context *cpuctx,
392                   struct perf_event_context *ctx)
393 {
394         if (event->state != PERF_EVENT_STATE_ACTIVE)
395                 return;
396
397         event->state = PERF_EVENT_STATE_INACTIVE;
398         if (event->pending_disable) {
399                 event->pending_disable = 0;
400                 event->state = PERF_EVENT_STATE_OFF;
401         }
402         event->tstamp_stopped = ctx->time;
403         event->pmu->disable(event);
404         event->oncpu = -1;
405
406         if (!is_software_event(event))
407                 cpuctx->active_oncpu--;
408         ctx->nr_active--;
409         if (event->attr.exclusive || !cpuctx->active_oncpu)
410                 cpuctx->exclusive = 0;
411 }
412
413 static void
414 group_sched_out(struct perf_event *group_event,
415                 struct perf_cpu_context *cpuctx,
416                 struct perf_event_context *ctx)
417 {
418         struct perf_event *event;
419
420         if (group_event->state != PERF_EVENT_STATE_ACTIVE)
421                 return;
422
423         event_sched_out(group_event, cpuctx, ctx);
424
425         /*
426          * Schedule out siblings (if any):
427          */
428         list_for_each_entry(event, &group_event->sibling_list, group_entry)
429                 event_sched_out(event, cpuctx, ctx);
430
431         if (group_event->attr.exclusive)
432                 cpuctx->exclusive = 0;
433 }
434
435 /*
436  * Cross CPU call to remove a performance event
437  *
438  * We disable the event on the hardware level first. After that we
439  * remove it from the context list.
440  */
441 static void __perf_event_remove_from_context(void *info)
442 {
443         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
444         struct perf_event *event = info;
445         struct perf_event_context *ctx = event->ctx;
446
447         /*
448          * If this is a task context, we need to check whether it is
449          * the current task context of this cpu. If not it has been
450          * scheduled out before the smp call arrived.
451          */
452         if (ctx->task && cpuctx->task_ctx != ctx)
453                 return;
454
455         raw_spin_lock(&ctx->lock);
456         /*
457          * Protect the list operation against NMI by disabling the
458          * events on a global level.
459          */
460         perf_disable();
461
462         event_sched_out(event, cpuctx, ctx);
463
464         list_del_event(event, ctx);
465
466         if (!ctx->task) {
467                 /*
468                  * Allow more per task events with respect to the
469                  * reservation:
470                  */
471                 cpuctx->max_pertask =
472                         min(perf_max_events - ctx->nr_events,
473                             perf_max_events - perf_reserved_percpu);
474         }
475
476         perf_enable();
477         raw_spin_unlock(&ctx->lock);
478 }
479
480
481 /*
482  * Remove the event from a task's (or a CPU's) list of events.
483  *
484  * Must be called with ctx->mutex held.
485  *
486  * CPU events are removed with a smp call. For task events we only
487  * call when the task is on a CPU.
488  *
489  * If event->ctx is a cloned context, callers must make sure that
490  * every task struct that event->ctx->task could possibly point to
491  * remains valid.  This is OK when called from perf_release since
492  * that only calls us on the top-level context, which can't be a clone.
493  * When called from perf_event_exit_task, it's OK because the
494  * context has been detached from its task.
495  */
496 static void perf_event_remove_from_context(struct perf_event *event)
497 {
498         struct perf_event_context *ctx = event->ctx;
499         struct task_struct *task = ctx->task;
500
501         if (!task) {
502                 /*
503                  * Per cpu events are removed via an smp call and
504                  * the removal is always successful.
505                  */
506                 smp_call_function_single(event->cpu,
507                                          __perf_event_remove_from_context,
508                                          event, 1);
509                 return;
510         }
511
512 retry:
513         task_oncpu_function_call(task, __perf_event_remove_from_context,
514                                  event);
515
516         raw_spin_lock_irq(&ctx->lock);
517         /*
518          * If the context is active we need to retry the smp call.
519          */
520         if (ctx->nr_active && !list_empty(&event->group_entry)) {
521                 raw_spin_unlock_irq(&ctx->lock);
522                 goto retry;
523         }
524
525         /*
526          * The lock prevents that this context is scheduled in so we
527          * can remove the event safely, if the call above did not
528          * succeed.
529          */
530         if (!list_empty(&event->group_entry))
531                 list_del_event(event, ctx);
532         raw_spin_unlock_irq(&ctx->lock);
533 }
534
535 /*
536  * Update total_time_enabled and total_time_running for all events in a group.
537  */
538 static void update_group_times(struct perf_event *leader)
539 {
540         struct perf_event *event;
541
542         update_event_times(leader);
543         list_for_each_entry(event, &leader->sibling_list, group_entry)
544                 update_event_times(event);
545 }
546
547 /*
548  * Cross CPU call to disable a performance event
549  */
550 static void __perf_event_disable(void *info)
551 {
552         struct perf_event *event = info;
553         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
554         struct perf_event_context *ctx = event->ctx;
555
556         /*
557          * If this is a per-task event, need to check whether this
558          * event's task is the current task on this cpu.
559          */
560         if (ctx->task && cpuctx->task_ctx != ctx)
561                 return;
562
563         raw_spin_lock(&ctx->lock);
564
565         /*
566          * If the event is on, turn it off.
567          * If it is in error state, leave it in error state.
568          */
569         if (event->state >= PERF_EVENT_STATE_INACTIVE) {
570                 update_context_time(ctx);
571                 update_group_times(event);
572                 if (event == event->group_leader)
573                         group_sched_out(event, cpuctx, ctx);
574                 else
575                         event_sched_out(event, cpuctx, ctx);
576                 event->state = PERF_EVENT_STATE_OFF;
577         }
578
579         raw_spin_unlock(&ctx->lock);
580 }
581
582 /*
583  * Disable a event.
584  *
585  * If event->ctx is a cloned context, callers must make sure that
586  * every task struct that event->ctx->task could possibly point to
587  * remains valid.  This condition is satisifed when called through
588  * perf_event_for_each_child or perf_event_for_each because they
589  * hold the top-level event's child_mutex, so any descendant that
590  * goes to exit will block in sync_child_event.
591  * When called from perf_pending_event it's OK because event->ctx
592  * is the current context on this CPU and preemption is disabled,
593  * hence we can't get into perf_event_task_sched_out for this context.
594  */
595 void perf_event_disable(struct perf_event *event)
596 {
597         struct perf_event_context *ctx = event->ctx;
598         struct task_struct *task = ctx->task;
599
600         if (!task) {
601                 /*
602                  * Disable the event on the cpu that it's on
603                  */
604                 smp_call_function_single(event->cpu, __perf_event_disable,
605                                          event, 1);
606                 return;
607         }
608
609  retry:
610         task_oncpu_function_call(task, __perf_event_disable, event);
611
612         raw_spin_lock_irq(&ctx->lock);
613         /*
614          * If the event is still active, we need to retry the cross-call.
615          */
616         if (event->state == PERF_EVENT_STATE_ACTIVE) {
617                 raw_spin_unlock_irq(&ctx->lock);
618                 goto retry;
619         }
620
621         /*
622          * Since we have the lock this context can't be scheduled
623          * in, so we can change the state safely.
624          */
625         if (event->state == PERF_EVENT_STATE_INACTIVE) {
626                 update_group_times(event);
627                 event->state = PERF_EVENT_STATE_OFF;
628         }
629
630         raw_spin_unlock_irq(&ctx->lock);
631 }
632
633 static int
634 event_sched_in(struct perf_event *event,
635                  struct perf_cpu_context *cpuctx,
636                  struct perf_event_context *ctx,
637                  int cpu)
638 {
639         if (event->state <= PERF_EVENT_STATE_OFF)
640                 return 0;
641
642         event->state = PERF_EVENT_STATE_ACTIVE;
643         event->oncpu = cpu;     /* TODO: put 'cpu' into cpuctx->cpu */
644         /*
645          * The new state must be visible before we turn it on in the hardware:
646          */
647         smp_wmb();
648
649         if (event->pmu->enable(event)) {
650                 event->state = PERF_EVENT_STATE_INACTIVE;
651                 event->oncpu = -1;
652                 return -EAGAIN;
653         }
654
655         event->tstamp_running += ctx->time - event->tstamp_stopped;
656
657         if (!is_software_event(event))
658                 cpuctx->active_oncpu++;
659         ctx->nr_active++;
660
661         if (event->attr.exclusive)
662                 cpuctx->exclusive = 1;
663
664         return 0;
665 }
666
667 static int
668 group_sched_in(struct perf_event *group_event,
669                struct perf_cpu_context *cpuctx,
670                struct perf_event_context *ctx,
671                int cpu)
672 {
673         struct perf_event *event, *partial_group;
674         int ret;
675
676         if (group_event->state == PERF_EVENT_STATE_OFF)
677                 return 0;
678
679         ret = hw_perf_group_sched_in(group_event, cpuctx, ctx, cpu);
680         if (ret)
681                 return ret < 0 ? ret : 0;
682
683         if (event_sched_in(group_event, cpuctx, ctx, cpu))
684                 return -EAGAIN;
685
686         /*
687          * Schedule in siblings as one group (if any):
688          */
689         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
690                 if (event_sched_in(event, cpuctx, ctx, cpu)) {
691                         partial_group = event;
692                         goto group_error;
693                 }
694         }
695
696         return 0;
697
698 group_error:
699         /*
700          * Groups can be scheduled in as one unit only, so undo any
701          * partial group before returning:
702          */
703         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
704                 if (event == partial_group)
705                         break;
706                 event_sched_out(event, cpuctx, ctx);
707         }
708         event_sched_out(group_event, cpuctx, ctx);
709
710         return -EAGAIN;
711 }
712
713 /*
714  * Work out whether we can put this event group on the CPU now.
715  */
716 static int group_can_go_on(struct perf_event *event,
717                            struct perf_cpu_context *cpuctx,
718                            int can_add_hw)
719 {
720         /*
721          * Groups consisting entirely of software events can always go on.
722          */
723         if (event->group_flags & PERF_GROUP_SOFTWARE)
724                 return 1;
725         /*
726          * If an exclusive group is already on, no other hardware
727          * events can go on.
728          */
729         if (cpuctx->exclusive)
730                 return 0;
731         /*
732          * If this group is exclusive and there are already
733          * events on the CPU, it can't go on.
734          */
735         if (event->attr.exclusive && cpuctx->active_oncpu)
736                 return 0;
737         /*
738          * Otherwise, try to add it if all previous groups were able
739          * to go on.
740          */
741         return can_add_hw;
742 }
743
744 static void add_event_to_ctx(struct perf_event *event,
745                                struct perf_event_context *ctx)
746 {
747         list_add_event(event, ctx);
748         event->tstamp_enabled = ctx->time;
749         event->tstamp_running = ctx->time;
750         event->tstamp_stopped = ctx->time;
751 }
752
753 /*
754  * Cross CPU call to install and enable a performance event
755  *
756  * Must be called with ctx->mutex held
757  */
758 static void __perf_install_in_context(void *info)
759 {
760         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
761         struct perf_event *event = info;
762         struct perf_event_context *ctx = event->ctx;
763         struct perf_event *leader = event->group_leader;
764         int cpu = smp_processor_id();
765         int err;
766
767         /*
768          * If this is a task context, we need to check whether it is
769          * the current task context of this cpu. If not it has been
770          * scheduled out before the smp call arrived.
771          * Or possibly this is the right context but it isn't
772          * on this cpu because it had no events.
773          */
774         if (ctx->task && cpuctx->task_ctx != ctx) {
775                 if (cpuctx->task_ctx || ctx->task != current)
776                         return;
777                 cpuctx->task_ctx = ctx;
778         }
779
780         raw_spin_lock(&ctx->lock);
781         ctx->is_active = 1;
782         update_context_time(ctx);
783
784         /*
785          * Protect the list operation against NMI by disabling the
786          * events on a global level. NOP for non NMI based events.
787          */
788         perf_disable();
789
790         add_event_to_ctx(event, ctx);
791
792         if (event->cpu != -1 && event->cpu != smp_processor_id())
793                 goto unlock;
794
795         /*
796          * Don't put the event on if it is disabled or if
797          * it is in a group and the group isn't on.
798          */
799         if (event->state != PERF_EVENT_STATE_INACTIVE ||
800             (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
801                 goto unlock;
802
803         /*
804          * An exclusive event can't go on if there are already active
805          * hardware events, and no hardware event can go on if there
806          * is already an exclusive event on.
807          */
808         if (!group_can_go_on(event, cpuctx, 1))
809                 err = -EEXIST;
810         else
811                 err = event_sched_in(event, cpuctx, ctx, cpu);
812
813         if (err) {
814                 /*
815                  * This event couldn't go on.  If it is in a group
816                  * then we have to pull the whole group off.
817                  * If the event group is pinned then put it in error state.
818                  */
819                 if (leader != event)
820                         group_sched_out(leader, cpuctx, ctx);
821                 if (leader->attr.pinned) {
822                         update_group_times(leader);
823                         leader->state = PERF_EVENT_STATE_ERROR;
824                 }
825         }
826
827         if (!err && !ctx->task && cpuctx->max_pertask)
828                 cpuctx->max_pertask--;
829
830  unlock:
831         perf_enable();
832
833         raw_spin_unlock(&ctx->lock);
834 }
835
836 /*
837  * Attach a performance event to a context
838  *
839  * First we add the event to the list with the hardware enable bit
840  * in event->hw_config cleared.
841  *
842  * If the event is attached to a task which is on a CPU we use a smp
843  * call to enable it in the task context. The task might have been
844  * scheduled away, but we check this in the smp call again.
845  *
846  * Must be called with ctx->mutex held.
847  */
848 static void
849 perf_install_in_context(struct perf_event_context *ctx,
850                         struct perf_event *event,
851                         int cpu)
852 {
853         struct task_struct *task = ctx->task;
854
855         if (!task) {
856                 /*
857                  * Per cpu events are installed via an smp call and
858                  * the install is always successful.
859                  */
860                 smp_call_function_single(cpu, __perf_install_in_context,
861                                          event, 1);
862                 return;
863         }
864
865 retry:
866         task_oncpu_function_call(task, __perf_install_in_context,
867                                  event);
868
869         raw_spin_lock_irq(&ctx->lock);
870         /*
871          * we need to retry the smp call.
872          */
873         if (ctx->is_active && list_empty(&event->group_entry)) {
874                 raw_spin_unlock_irq(&ctx->lock);
875                 goto retry;
876         }
877
878         /*
879          * The lock prevents that this context is scheduled in so we
880          * can add the event safely, if it the call above did not
881          * succeed.
882          */
883         if (list_empty(&event->group_entry))
884                 add_event_to_ctx(event, ctx);
885         raw_spin_unlock_irq(&ctx->lock);
886 }
887
888 /*
889  * Put a event into inactive state and update time fields.
890  * Enabling the leader of a group effectively enables all
891  * the group members that aren't explicitly disabled, so we
892  * have to update their ->tstamp_enabled also.
893  * Note: this works for group members as well as group leaders
894  * since the non-leader members' sibling_lists will be empty.
895  */
896 static void __perf_event_mark_enabled(struct perf_event *event,
897                                         struct perf_event_context *ctx)
898 {
899         struct perf_event *sub;
900
901         event->state = PERF_EVENT_STATE_INACTIVE;
902         event->tstamp_enabled = ctx->time - event->total_time_enabled;
903         list_for_each_entry(sub, &event->sibling_list, group_entry)
904                 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
905                         sub->tstamp_enabled =
906                                 ctx->time - sub->total_time_enabled;
907 }
908
909 /*
910  * Cross CPU call to enable a performance event
911  */
912 static void __perf_event_enable(void *info)
913 {
914         struct perf_event *event = info;
915         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
916         struct perf_event_context *ctx = event->ctx;
917         struct perf_event *leader = event->group_leader;
918         int err;
919
920         /*
921          * If this is a per-task event, need to check whether this
922          * event's task is the current task on this cpu.
923          */
924         if (ctx->task && cpuctx->task_ctx != ctx) {
925                 if (cpuctx->task_ctx || ctx->task != current)
926                         return;
927                 cpuctx->task_ctx = ctx;
928         }
929
930         raw_spin_lock(&ctx->lock);
931         ctx->is_active = 1;
932         update_context_time(ctx);
933
934         if (event->state >= PERF_EVENT_STATE_INACTIVE)
935                 goto unlock;
936         __perf_event_mark_enabled(event, ctx);
937
938         if (event->cpu != -1 && event->cpu != smp_processor_id())
939                 goto unlock;
940
941         /*
942          * If the event is in a group and isn't the group leader,
943          * then don't put it on unless the group is on.
944          */
945         if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
946                 goto unlock;
947
948         if (!group_can_go_on(event, cpuctx, 1)) {
949                 err = -EEXIST;
950         } else {
951                 perf_disable();
952                 if (event == leader)
953                         err = group_sched_in(event, cpuctx, ctx,
954                                              smp_processor_id());
955                 else
956                         err = event_sched_in(event, cpuctx, ctx,
957                                                smp_processor_id());
958                 perf_enable();
959         }
960
961         if (err) {
962                 /*
963                  * If this event can't go on and it's part of a
964                  * group, then the whole group has to come off.
965                  */
966                 if (leader != event)
967                         group_sched_out(leader, cpuctx, ctx);
968                 if (leader->attr.pinned) {
969                         update_group_times(leader);
970                         leader->state = PERF_EVENT_STATE_ERROR;
971                 }
972         }
973
974  unlock:
975         raw_spin_unlock(&ctx->lock);
976 }
977
978 /*
979  * Enable a event.
980  *
981  * If event->ctx is a cloned context, callers must make sure that
982  * every task struct that event->ctx->task could possibly point to
983  * remains valid.  This condition is satisfied when called through
984  * perf_event_for_each_child or perf_event_for_each as described
985  * for perf_event_disable.
986  */
987 void perf_event_enable(struct perf_event *event)
988 {
989         struct perf_event_context *ctx = event->ctx;
990         struct task_struct *task = ctx->task;
991
992         if (!task) {
993                 /*
994                  * Enable the event on the cpu that it's on
995                  */
996                 smp_call_function_single(event->cpu, __perf_event_enable,
997                                          event, 1);
998                 return;
999         }
1000
1001         raw_spin_lock_irq(&ctx->lock);
1002         if (event->state >= PERF_EVENT_STATE_INACTIVE)
1003                 goto out;
1004
1005         /*
1006          * If the event is in error state, clear that first.
1007          * That way, if we see the event in error state below, we
1008          * know that it has gone back into error state, as distinct
1009          * from the task having been scheduled away before the
1010          * cross-call arrived.
1011          */
1012         if (event->state == PERF_EVENT_STATE_ERROR)
1013                 event->state = PERF_EVENT_STATE_OFF;
1014
1015  retry:
1016         raw_spin_unlock_irq(&ctx->lock);
1017         task_oncpu_function_call(task, __perf_event_enable, event);
1018
1019         raw_spin_lock_irq(&ctx->lock);
1020
1021         /*
1022          * If the context is active and the event is still off,
1023          * we need to retry the cross-call.
1024          */
1025         if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1026                 goto retry;
1027
1028         /*
1029          * Since we have the lock this context can't be scheduled
1030          * in, so we can change the state safely.
1031          */
1032         if (event->state == PERF_EVENT_STATE_OFF)
1033                 __perf_event_mark_enabled(event, ctx);
1034
1035  out:
1036         raw_spin_unlock_irq(&ctx->lock);
1037 }
1038
1039 static int perf_event_refresh(struct perf_event *event, int refresh)
1040 {
1041         /*
1042          * not supported on inherited events
1043          */
1044         if (event->attr.inherit)
1045                 return -EINVAL;
1046
1047         atomic_add(refresh, &event->event_limit);
1048         perf_event_enable(event);
1049
1050         return 0;
1051 }
1052
1053 enum event_type_t {
1054         EVENT_FLEXIBLE = 0x1,
1055         EVENT_PINNED = 0x2,
1056         EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1057 };
1058
1059 static void ctx_sched_out(struct perf_event_context *ctx,
1060                           struct perf_cpu_context *cpuctx,
1061                           enum event_type_t event_type)
1062 {
1063         struct perf_event *event;
1064
1065         raw_spin_lock(&ctx->lock);
1066         ctx->is_active = 0;
1067         if (likely(!ctx->nr_events))
1068                 goto out;
1069         update_context_time(ctx);
1070
1071         perf_disable();
1072         if (!ctx->nr_active)
1073                 goto out_enable;
1074
1075         if (event_type & EVENT_PINNED)
1076                 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1077                         group_sched_out(event, cpuctx, ctx);
1078
1079         if (event_type & EVENT_FLEXIBLE)
1080                 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1081                         group_sched_out(event, cpuctx, ctx);
1082
1083  out_enable:
1084         perf_enable();
1085  out:
1086         raw_spin_unlock(&ctx->lock);
1087 }
1088
1089 /*
1090  * Test whether two contexts are equivalent, i.e. whether they
1091  * have both been cloned from the same version of the same context
1092  * and they both have the same number of enabled events.
1093  * If the number of enabled events is the same, then the set
1094  * of enabled events should be the same, because these are both
1095  * inherited contexts, therefore we can't access individual events
1096  * in them directly with an fd; we can only enable/disable all
1097  * events via prctl, or enable/disable all events in a family
1098  * via ioctl, which will have the same effect on both contexts.
1099  */
1100 static int context_equiv(struct perf_event_context *ctx1,
1101                          struct perf_event_context *ctx2)
1102 {
1103         return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1104                 && ctx1->parent_gen == ctx2->parent_gen
1105                 && !ctx1->pin_count && !ctx2->pin_count;
1106 }
1107
1108 static void __perf_event_sync_stat(struct perf_event *event,
1109                                      struct perf_event *next_event)
1110 {
1111         u64 value;
1112
1113         if (!event->attr.inherit_stat)
1114                 return;
1115
1116         /*
1117          * Update the event value, we cannot use perf_event_read()
1118          * because we're in the middle of a context switch and have IRQs
1119          * disabled, which upsets smp_call_function_single(), however
1120          * we know the event must be on the current CPU, therefore we
1121          * don't need to use it.
1122          */
1123         switch (event->state) {
1124         case PERF_EVENT_STATE_ACTIVE:
1125                 event->pmu->read(event);
1126                 /* fall-through */
1127
1128         case PERF_EVENT_STATE_INACTIVE:
1129                 update_event_times(event);
1130                 break;
1131
1132         default:
1133                 break;
1134         }
1135
1136         /*
1137          * In order to keep per-task stats reliable we need to flip the event
1138          * values when we flip the contexts.
1139          */
1140         value = atomic64_read(&next_event->count);
1141         value = atomic64_xchg(&event->count, value);
1142         atomic64_set(&next_event->count, value);
1143
1144         swap(event->total_time_enabled, next_event->total_time_enabled);
1145         swap(event->total_time_running, next_event->total_time_running);
1146
1147         /*
1148          * Since we swizzled the values, update the user visible data too.
1149          */
1150         perf_event_update_userpage(event);
1151         perf_event_update_userpage(next_event);
1152 }
1153
1154 #define list_next_entry(pos, member) \
1155         list_entry(pos->member.next, typeof(*pos), member)
1156
1157 static void perf_event_sync_stat(struct perf_event_context *ctx,
1158                                    struct perf_event_context *next_ctx)
1159 {
1160         struct perf_event *event, *next_event;
1161
1162         if (!ctx->nr_stat)
1163                 return;
1164
1165         update_context_time(ctx);
1166
1167         event = list_first_entry(&ctx->event_list,
1168                                    struct perf_event, event_entry);
1169
1170         next_event = list_first_entry(&next_ctx->event_list,
1171                                         struct perf_event, event_entry);
1172
1173         while (&event->event_entry != &ctx->event_list &&
1174                &next_event->event_entry != &next_ctx->event_list) {
1175
1176                 __perf_event_sync_stat(event, next_event);
1177
1178                 event = list_next_entry(event, event_entry);
1179                 next_event = list_next_entry(next_event, event_entry);
1180         }
1181 }
1182
1183 /*
1184  * Called from scheduler to remove the events of the current task,
1185  * with interrupts disabled.
1186  *
1187  * We stop each event and update the event value in event->count.
1188  *
1189  * This does not protect us against NMI, but disable()
1190  * sets the disabled bit in the control field of event _before_
1191  * accessing the event control register. If a NMI hits, then it will
1192  * not restart the event.
1193  */
1194 void perf_event_task_sched_out(struct task_struct *task,
1195                                  struct task_struct *next)
1196 {
1197         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1198         struct perf_event_context *ctx = task->perf_event_ctxp;
1199         struct perf_event_context *next_ctx;
1200         struct perf_event_context *parent;
1201         struct pt_regs *regs;
1202         int do_switch = 1;
1203
1204         regs = task_pt_regs(task);
1205         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1206
1207         if (likely(!ctx || !cpuctx->task_ctx))
1208                 return;
1209
1210         rcu_read_lock();
1211         parent = rcu_dereference(ctx->parent_ctx);
1212         next_ctx = next->perf_event_ctxp;
1213         if (parent && next_ctx &&
1214             rcu_dereference(next_ctx->parent_ctx) == parent) {
1215                 /*
1216                  * Looks like the two contexts are clones, so we might be
1217                  * able to optimize the context switch.  We lock both
1218                  * contexts and check that they are clones under the
1219                  * lock (including re-checking that neither has been
1220                  * uncloned in the meantime).  It doesn't matter which
1221                  * order we take the locks because no other cpu could
1222                  * be trying to lock both of these tasks.
1223                  */
1224                 raw_spin_lock(&ctx->lock);
1225                 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1226                 if (context_equiv(ctx, next_ctx)) {
1227                         /*
1228                          * XXX do we need a memory barrier of sorts
1229                          * wrt to rcu_dereference() of perf_event_ctxp
1230                          */
1231                         task->perf_event_ctxp = next_ctx;
1232                         next->perf_event_ctxp = ctx;
1233                         ctx->task = next;
1234                         next_ctx->task = task;
1235                         do_switch = 0;
1236
1237                         perf_event_sync_stat(ctx, next_ctx);
1238                 }
1239                 raw_spin_unlock(&next_ctx->lock);
1240                 raw_spin_unlock(&ctx->lock);
1241         }
1242         rcu_read_unlock();
1243
1244         if (do_switch) {
1245                 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1246                 cpuctx->task_ctx = NULL;
1247         }
1248 }
1249
1250 static void task_ctx_sched_out(struct perf_event_context *ctx,
1251                                enum event_type_t event_type)
1252 {
1253         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1254
1255         if (!cpuctx->task_ctx)
1256                 return;
1257
1258         if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1259                 return;
1260
1261         ctx_sched_out(ctx, cpuctx, event_type);
1262         cpuctx->task_ctx = NULL;
1263 }
1264
1265 /*
1266  * Called with IRQs disabled
1267  */
1268 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1269 {
1270         task_ctx_sched_out(ctx, EVENT_ALL);
1271 }
1272
1273 /*
1274  * Called with IRQs disabled
1275  */
1276 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1277                               enum event_type_t event_type)
1278 {
1279         ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1280 }
1281
1282 static void
1283 ctx_pinned_sched_in(struct perf_event_context *ctx,
1284                     struct perf_cpu_context *cpuctx,
1285                     int cpu)
1286 {
1287         struct perf_event *event;
1288
1289         list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1290                 if (event->state <= PERF_EVENT_STATE_OFF)
1291                         continue;
1292                 if (event->cpu != -1 && event->cpu != cpu)
1293                         continue;
1294
1295                 if (group_can_go_on(event, cpuctx, 1))
1296                         group_sched_in(event, cpuctx, ctx, cpu);
1297
1298                 /*
1299                  * If this pinned group hasn't been scheduled,
1300                  * put it in error state.
1301                  */
1302                 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1303                         update_group_times(event);
1304                         event->state = PERF_EVENT_STATE_ERROR;
1305                 }
1306         }
1307 }
1308
1309 static void
1310 ctx_flexible_sched_in(struct perf_event_context *ctx,
1311                       struct perf_cpu_context *cpuctx,
1312                       int cpu)
1313 {
1314         struct perf_event *event;
1315         int can_add_hw = 1;
1316
1317         list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1318                 /* Ignore events in OFF or ERROR state */
1319                 if (event->state <= PERF_EVENT_STATE_OFF)
1320                         continue;
1321                 /*
1322                  * Listen to the 'cpu' scheduling filter constraint
1323                  * of events:
1324                  */
1325                 if (event->cpu != -1 && event->cpu != cpu)
1326                         continue;
1327
1328                 if (group_can_go_on(event, cpuctx, can_add_hw))
1329                         if (group_sched_in(event, cpuctx, ctx, cpu))
1330                                 can_add_hw = 0;
1331         }
1332 }
1333
1334 static void
1335 ctx_sched_in(struct perf_event_context *ctx,
1336              struct perf_cpu_context *cpuctx,
1337              enum event_type_t event_type)
1338 {
1339         int cpu = smp_processor_id();
1340
1341         raw_spin_lock(&ctx->lock);
1342         ctx->is_active = 1;
1343         if (likely(!ctx->nr_events))
1344                 goto out;
1345
1346         ctx->timestamp = perf_clock();
1347
1348         perf_disable();
1349
1350         /*
1351          * First go through the list and put on any pinned groups
1352          * in order to give them the best chance of going on.
1353          */
1354         if (event_type & EVENT_PINNED)
1355                 ctx_pinned_sched_in(ctx, cpuctx, cpu);
1356
1357         /* Then walk through the lower prio flexible groups */
1358         if (event_type & EVENT_FLEXIBLE)
1359                 ctx_flexible_sched_in(ctx, cpuctx, cpu);
1360
1361         perf_enable();
1362  out:
1363         raw_spin_unlock(&ctx->lock);
1364 }
1365
1366 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1367                              enum event_type_t event_type)
1368 {
1369         struct perf_event_context *ctx = &cpuctx->ctx;
1370
1371         ctx_sched_in(ctx, cpuctx, event_type);
1372 }
1373
1374 static void task_ctx_sched_in(struct task_struct *task,
1375                               enum event_type_t event_type)
1376 {
1377         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1378         struct perf_event_context *ctx = task->perf_event_ctxp;
1379
1380         if (likely(!ctx))
1381                 return;
1382         if (cpuctx->task_ctx == ctx)
1383                 return;
1384         ctx_sched_in(ctx, cpuctx, event_type);
1385         cpuctx->task_ctx = ctx;
1386 }
1387 /*
1388  * Called from scheduler to add the events of the current task
1389  * with interrupts disabled.
1390  *
1391  * We restore the event value and then enable it.
1392  *
1393  * This does not protect us against NMI, but enable()
1394  * sets the enabled bit in the control field of event _before_
1395  * accessing the event control register. If a NMI hits, then it will
1396  * keep the event running.
1397  */
1398 void perf_event_task_sched_in(struct task_struct *task)
1399 {
1400         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1401         struct perf_event_context *ctx = task->perf_event_ctxp;
1402
1403         if (likely(!ctx))
1404                 return;
1405
1406         if (cpuctx->task_ctx == ctx)
1407                 return;
1408
1409         /*
1410          * We want to keep the following priority order:
1411          * cpu pinned (that don't need to move), task pinned,
1412          * cpu flexible, task flexible.
1413          */
1414         cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1415
1416         ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1417         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1418         ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1419
1420         cpuctx->task_ctx = ctx;
1421 }
1422
1423 #define MAX_INTERRUPTS (~0ULL)
1424
1425 static void perf_log_throttle(struct perf_event *event, int enable);
1426
1427 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1428 {
1429         u64 frequency = event->attr.sample_freq;
1430         u64 sec = NSEC_PER_SEC;
1431         u64 divisor, dividend;
1432
1433         int count_fls, nsec_fls, frequency_fls, sec_fls;
1434
1435         count_fls = fls64(count);
1436         nsec_fls = fls64(nsec);
1437         frequency_fls = fls64(frequency);
1438         sec_fls = 30;
1439
1440         /*
1441          * We got @count in @nsec, with a target of sample_freq HZ
1442          * the target period becomes:
1443          *
1444          *             @count * 10^9
1445          * period = -------------------
1446          *          @nsec * sample_freq
1447          *
1448          */
1449
1450         /*
1451          * Reduce accuracy by one bit such that @a and @b converge
1452          * to a similar magnitude.
1453          */
1454 #define REDUCE_FLS(a, b)                \
1455 do {                                    \
1456         if (a##_fls > b##_fls) {        \
1457                 a >>= 1;                \
1458                 a##_fls--;              \
1459         } else {                        \
1460                 b >>= 1;                \
1461                 b##_fls--;              \
1462         }                               \
1463 } while (0)
1464
1465         /*
1466          * Reduce accuracy until either term fits in a u64, then proceed with
1467          * the other, so that finally we can do a u64/u64 division.
1468          */
1469         while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1470                 REDUCE_FLS(nsec, frequency);
1471                 REDUCE_FLS(sec, count);
1472         }
1473
1474         if (count_fls + sec_fls > 64) {
1475                 divisor = nsec * frequency;
1476
1477                 while (count_fls + sec_fls > 64) {
1478                         REDUCE_FLS(count, sec);
1479                         divisor >>= 1;
1480                 }
1481
1482                 dividend = count * sec;
1483         } else {
1484                 dividend = count * sec;
1485
1486                 while (nsec_fls + frequency_fls > 64) {
1487                         REDUCE_FLS(nsec, frequency);
1488                         dividend >>= 1;
1489                 }
1490
1491                 divisor = nsec * frequency;
1492         }
1493
1494         return div64_u64(dividend, divisor);
1495 }
1496
1497 static void perf_event_stop(struct perf_event *event)
1498 {
1499         if (!event->pmu->stop)
1500                 return event->pmu->disable(event);
1501
1502         return event->pmu->stop(event);
1503 }
1504
1505 static int perf_event_start(struct perf_event *event)
1506 {
1507         if (!event->pmu->start)
1508                 return event->pmu->enable(event);
1509
1510         return event->pmu->start(event);
1511 }
1512
1513 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1514 {
1515         struct hw_perf_event *hwc = &event->hw;
1516         u64 period, sample_period;
1517         s64 delta;
1518
1519         period = perf_calculate_period(event, nsec, count);
1520
1521         delta = (s64)(period - hwc->sample_period);
1522         delta = (delta + 7) / 8; /* low pass filter */
1523
1524         sample_period = hwc->sample_period + delta;
1525
1526         if (!sample_period)
1527                 sample_period = 1;
1528
1529         hwc->sample_period = sample_period;
1530
1531         if (atomic64_read(&hwc->period_left) > 8*sample_period) {
1532                 perf_disable();
1533                 perf_event_stop(event);
1534                 atomic64_set(&hwc->period_left, 0);
1535                 perf_event_start(event);
1536                 perf_enable();
1537         }
1538 }
1539
1540 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1541 {
1542         struct perf_event *event;
1543         struct hw_perf_event *hwc;
1544         u64 interrupts, now;
1545         s64 delta;
1546
1547         raw_spin_lock(&ctx->lock);
1548         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1549                 if (event->state != PERF_EVENT_STATE_ACTIVE)
1550                         continue;
1551
1552                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1553                         continue;
1554
1555                 hwc = &event->hw;
1556
1557                 interrupts = hwc->interrupts;
1558                 hwc->interrupts = 0;
1559
1560                 /*
1561                  * unthrottle events on the tick
1562                  */
1563                 if (interrupts == MAX_INTERRUPTS) {
1564                         perf_log_throttle(event, 1);
1565                         event->pmu->unthrottle(event);
1566                 }
1567
1568                 if (!event->attr.freq || !event->attr.sample_freq)
1569                         continue;
1570
1571                 event->pmu->read(event);
1572                 now = atomic64_read(&event->count);
1573                 delta = now - hwc->freq_count_stamp;
1574                 hwc->freq_count_stamp = now;
1575
1576                 if (delta > 0)
1577                         perf_adjust_period(event, TICK_NSEC, delta);
1578         }
1579         raw_spin_unlock(&ctx->lock);
1580 }
1581
1582 /*
1583  * Round-robin a context's events:
1584  */
1585 static void rotate_ctx(struct perf_event_context *ctx)
1586 {
1587         if (!ctx->nr_events)
1588                 return;
1589
1590         raw_spin_lock(&ctx->lock);
1591
1592         /* Rotate the first entry last of non-pinned groups */
1593         list_rotate_left(&ctx->flexible_groups);
1594
1595         raw_spin_unlock(&ctx->lock);
1596 }
1597
1598 void perf_event_task_tick(struct task_struct *curr)
1599 {
1600         struct perf_cpu_context *cpuctx;
1601         struct perf_event_context *ctx;
1602
1603         if (!atomic_read(&nr_events))
1604                 return;
1605
1606         cpuctx = &__get_cpu_var(perf_cpu_context);
1607         ctx = curr->perf_event_ctxp;
1608
1609         perf_disable();
1610
1611         perf_ctx_adjust_freq(&cpuctx->ctx);
1612         if (ctx)
1613                 perf_ctx_adjust_freq(ctx);
1614
1615         cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1616         if (ctx)
1617                 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1618
1619         rotate_ctx(&cpuctx->ctx);
1620         if (ctx)
1621                 rotate_ctx(ctx);
1622
1623         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1624         if (ctx)
1625                 task_ctx_sched_in(curr, EVENT_FLEXIBLE);
1626
1627         perf_enable();
1628 }
1629
1630 static int event_enable_on_exec(struct perf_event *event,
1631                                 struct perf_event_context *ctx)
1632 {
1633         if (!event->attr.enable_on_exec)
1634                 return 0;
1635
1636         event->attr.enable_on_exec = 0;
1637         if (event->state >= PERF_EVENT_STATE_INACTIVE)
1638                 return 0;
1639
1640         __perf_event_mark_enabled(event, ctx);
1641
1642         return 1;
1643 }
1644
1645 /*
1646  * Enable all of a task's events that have been marked enable-on-exec.
1647  * This expects task == current.
1648  */
1649 static void perf_event_enable_on_exec(struct task_struct *task)
1650 {
1651         struct perf_event_context *ctx;
1652         struct perf_event *event;
1653         unsigned long flags;
1654         int enabled = 0;
1655         int ret;
1656
1657         local_irq_save(flags);
1658         ctx = task->perf_event_ctxp;
1659         if (!ctx || !ctx->nr_events)
1660                 goto out;
1661
1662         __perf_event_task_sched_out(ctx);
1663
1664         raw_spin_lock(&ctx->lock);
1665
1666         list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1667                 ret = event_enable_on_exec(event, ctx);
1668                 if (ret)
1669                         enabled = 1;
1670         }
1671
1672         list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1673                 ret = event_enable_on_exec(event, ctx);
1674                 if (ret)
1675                         enabled = 1;
1676         }
1677
1678         /*
1679          * Unclone this context if we enabled any event.
1680          */
1681         if (enabled)
1682                 unclone_ctx(ctx);
1683
1684         raw_spin_unlock(&ctx->lock);
1685
1686         perf_event_task_sched_in(task);
1687  out:
1688         local_irq_restore(flags);
1689 }
1690
1691 /*
1692  * Cross CPU call to read the hardware event
1693  */
1694 static void __perf_event_read(void *info)
1695 {
1696         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1697         struct perf_event *event = info;
1698         struct perf_event_context *ctx = event->ctx;
1699
1700         /*
1701          * If this is a task context, we need to check whether it is
1702          * the current task context of this cpu.  If not it has been
1703          * scheduled out before the smp call arrived.  In that case
1704          * event->count would have been updated to a recent sample
1705          * when the event was scheduled out.
1706          */
1707         if (ctx->task && cpuctx->task_ctx != ctx)
1708                 return;
1709
1710         raw_spin_lock(&ctx->lock);
1711         update_context_time(ctx);
1712         update_event_times(event);
1713         raw_spin_unlock(&ctx->lock);
1714
1715         event->pmu->read(event);
1716 }
1717
1718 static u64 perf_event_read(struct perf_event *event)
1719 {
1720         /*
1721          * If event is enabled and currently active on a CPU, update the
1722          * value in the event structure:
1723          */
1724         if (event->state == PERF_EVENT_STATE_ACTIVE) {
1725                 smp_call_function_single(event->oncpu,
1726                                          __perf_event_read, event, 1);
1727         } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1728                 struct perf_event_context *ctx = event->ctx;
1729                 unsigned long flags;
1730
1731                 raw_spin_lock_irqsave(&ctx->lock, flags);
1732                 update_context_time(ctx);
1733                 update_event_times(event);
1734                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1735         }
1736
1737         return atomic64_read(&event->count);
1738 }
1739
1740 /*
1741  * Initialize the perf_event context in a task_struct:
1742  */
1743 static void
1744 __perf_event_init_context(struct perf_event_context *ctx,
1745                             struct task_struct *task)
1746 {
1747         raw_spin_lock_init(&ctx->lock);
1748         mutex_init(&ctx->mutex);
1749         INIT_LIST_HEAD(&ctx->pinned_groups);
1750         INIT_LIST_HEAD(&ctx->flexible_groups);
1751         INIT_LIST_HEAD(&ctx->event_list);
1752         atomic_set(&ctx->refcount, 1);
1753         ctx->task = task;
1754 }
1755
1756 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1757 {
1758         struct perf_event_context *ctx;
1759         struct perf_cpu_context *cpuctx;
1760         struct task_struct *task;
1761         unsigned long flags;
1762         int err;
1763
1764         if (pid == -1 && cpu != -1) {
1765                 /* Must be root to operate on a CPU event: */
1766                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1767                         return ERR_PTR(-EACCES);
1768
1769                 if (cpu < 0 || cpu >= nr_cpumask_bits)
1770                         return ERR_PTR(-EINVAL);
1771
1772                 /*
1773                  * We could be clever and allow to attach a event to an
1774                  * offline CPU and activate it when the CPU comes up, but
1775                  * that's for later.
1776                  */
1777                 if (!cpu_online(cpu))
1778                         return ERR_PTR(-ENODEV);
1779
1780                 cpuctx = &per_cpu(perf_cpu_context, cpu);
1781                 ctx = &cpuctx->ctx;
1782                 get_ctx(ctx);
1783
1784                 return ctx;
1785         }
1786
1787         rcu_read_lock();
1788         if (!pid)
1789                 task = current;
1790         else
1791                 task = find_task_by_vpid(pid);
1792         if (task)
1793                 get_task_struct(task);
1794         rcu_read_unlock();
1795
1796         if (!task)
1797                 return ERR_PTR(-ESRCH);
1798
1799         /*
1800          * Can't attach events to a dying task.
1801          */
1802         err = -ESRCH;
1803         if (task->flags & PF_EXITING)
1804                 goto errout;
1805
1806         /* Reuse ptrace permission checks for now. */
1807         err = -EACCES;
1808         if (!ptrace_may_access(task, PTRACE_MODE_READ))
1809                 goto errout;
1810
1811  retry:
1812         ctx = perf_lock_task_context(task, &flags);
1813         if (ctx) {
1814                 unclone_ctx(ctx);
1815                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1816         }
1817
1818         if (!ctx) {
1819                 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1820                 err = -ENOMEM;
1821                 if (!ctx)
1822                         goto errout;
1823                 __perf_event_init_context(ctx, task);
1824                 get_ctx(ctx);
1825                 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1826                         /*
1827                          * We raced with some other task; use
1828                          * the context they set.
1829                          */
1830                         kfree(ctx);
1831                         goto retry;
1832                 }
1833                 get_task_struct(task);
1834         }
1835
1836         put_task_struct(task);
1837         return ctx;
1838
1839  errout:
1840         put_task_struct(task);
1841         return ERR_PTR(err);
1842 }
1843
1844 static void perf_event_free_filter(struct perf_event *event);
1845
1846 static void free_event_rcu(struct rcu_head *head)
1847 {
1848         struct perf_event *event;
1849
1850         event = container_of(head, struct perf_event, rcu_head);
1851         if (event->ns)
1852                 put_pid_ns(event->ns);
1853         perf_event_free_filter(event);
1854         kfree(event);
1855 }
1856
1857 static void perf_pending_sync(struct perf_event *event);
1858
1859 static void free_event(struct perf_event *event)
1860 {
1861         perf_pending_sync(event);
1862
1863         if (!event->parent) {
1864                 atomic_dec(&nr_events);
1865                 if (event->attr.mmap)
1866                         atomic_dec(&nr_mmap_events);
1867                 if (event->attr.comm)
1868                         atomic_dec(&nr_comm_events);
1869                 if (event->attr.task)
1870                         atomic_dec(&nr_task_events);
1871         }
1872
1873         if (event->output) {
1874                 fput(event->output->filp);
1875                 event->output = NULL;
1876         }
1877
1878         if (event->destroy)
1879                 event->destroy(event);
1880
1881         put_ctx(event->ctx);
1882         call_rcu(&event->rcu_head, free_event_rcu);
1883 }
1884
1885 int perf_event_release_kernel(struct perf_event *event)
1886 {
1887         struct perf_event_context *ctx = event->ctx;
1888
1889         WARN_ON_ONCE(ctx->parent_ctx);
1890         mutex_lock(&ctx->mutex);
1891         perf_event_remove_from_context(event);
1892         mutex_unlock(&ctx->mutex);
1893
1894         mutex_lock(&event->owner->perf_event_mutex);
1895         list_del_init(&event->owner_entry);
1896         mutex_unlock(&event->owner->perf_event_mutex);
1897         put_task_struct(event->owner);
1898
1899         free_event(event);
1900
1901         return 0;
1902 }
1903 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1904
1905 /*
1906  * Called when the last reference to the file is gone.
1907  */
1908 static int perf_release(struct inode *inode, struct file *file)
1909 {
1910         struct perf_event *event = file->private_data;
1911
1912         file->private_data = NULL;
1913
1914         return perf_event_release_kernel(event);
1915 }
1916
1917 static int perf_event_read_size(struct perf_event *event)
1918 {
1919         int entry = sizeof(u64); /* value */
1920         int size = 0;
1921         int nr = 1;
1922
1923         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1924                 size += sizeof(u64);
1925
1926         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1927                 size += sizeof(u64);
1928
1929         if (event->attr.read_format & PERF_FORMAT_ID)
1930                 entry += sizeof(u64);
1931
1932         if (event->attr.read_format & PERF_FORMAT_GROUP) {
1933                 nr += event->group_leader->nr_siblings;
1934                 size += sizeof(u64);
1935         }
1936
1937         size += entry * nr;
1938
1939         return size;
1940 }
1941
1942 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1943 {
1944         struct perf_event *child;
1945         u64 total = 0;
1946
1947         *enabled = 0;
1948         *running = 0;
1949
1950         mutex_lock(&event->child_mutex);
1951         total += perf_event_read(event);
1952         *enabled += event->total_time_enabled +
1953                         atomic64_read(&event->child_total_time_enabled);
1954         *running += event->total_time_running +
1955                         atomic64_read(&event->child_total_time_running);
1956
1957         list_for_each_entry(child, &event->child_list, child_list) {
1958                 total += perf_event_read(child);
1959                 *enabled += child->total_time_enabled;
1960                 *running += child->total_time_running;
1961         }
1962         mutex_unlock(&event->child_mutex);
1963
1964         return total;
1965 }
1966 EXPORT_SYMBOL_GPL(perf_event_read_value);
1967
1968 static int perf_event_read_group(struct perf_event *event,
1969                                    u64 read_format, char __user *buf)
1970 {
1971         struct perf_event *leader = event->group_leader, *sub;
1972         int n = 0, size = 0, ret = -EFAULT;
1973         struct perf_event_context *ctx = leader->ctx;
1974         u64 values[5];
1975         u64 count, enabled, running;
1976
1977         mutex_lock(&ctx->mutex);
1978         count = perf_event_read_value(leader, &enabled, &running);
1979
1980         values[n++] = 1 + leader->nr_siblings;
1981         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1982                 values[n++] = enabled;
1983         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1984                 values[n++] = running;
1985         values[n++] = count;
1986         if (read_format & PERF_FORMAT_ID)
1987                 values[n++] = primary_event_id(leader);
1988
1989         size = n * sizeof(u64);
1990
1991         if (copy_to_user(buf, values, size))
1992                 goto unlock;
1993
1994         ret = size;
1995
1996         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1997                 n = 0;
1998
1999                 values[n++] = perf_event_read_value(sub, &enabled, &running);
2000                 if (read_format & PERF_FORMAT_ID)
2001                         values[n++] = primary_event_id(sub);
2002
2003                 size = n * sizeof(u64);
2004
2005                 if (copy_to_user(buf + ret, values, size)) {
2006                         ret = -EFAULT;
2007                         goto unlock;
2008                 }
2009
2010                 ret += size;
2011         }
2012 unlock:
2013         mutex_unlock(&ctx->mutex);
2014
2015         return ret;
2016 }
2017
2018 static int perf_event_read_one(struct perf_event *event,
2019                                  u64 read_format, char __user *buf)
2020 {
2021         u64 enabled, running;
2022         u64 values[4];
2023         int n = 0;
2024
2025         values[n++] = perf_event_read_value(event, &enabled, &running);
2026         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2027                 values[n++] = enabled;
2028         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2029                 values[n++] = running;
2030         if (read_format & PERF_FORMAT_ID)
2031                 values[n++] = primary_event_id(event);
2032
2033         if (copy_to_user(buf, values, n * sizeof(u64)))
2034                 return -EFAULT;
2035
2036         return n * sizeof(u64);
2037 }
2038
2039 /*
2040  * Read the performance event - simple non blocking version for now
2041  */
2042 static ssize_t
2043 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2044 {
2045         u64 read_format = event->attr.read_format;
2046         int ret;
2047
2048         /*
2049          * Return end-of-file for a read on a event that is in
2050          * error state (i.e. because it was pinned but it couldn't be
2051          * scheduled on to the CPU at some point).
2052          */
2053         if (event->state == PERF_EVENT_STATE_ERROR)
2054                 return 0;
2055
2056         if (count < perf_event_read_size(event))
2057                 return -ENOSPC;
2058
2059         WARN_ON_ONCE(event->ctx->parent_ctx);
2060         if (read_format & PERF_FORMAT_GROUP)
2061                 ret = perf_event_read_group(event, read_format, buf);
2062         else
2063                 ret = perf_event_read_one(event, read_format, buf);
2064
2065         return ret;
2066 }
2067
2068 static ssize_t
2069 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2070 {
2071         struct perf_event *event = file->private_data;
2072
2073         return perf_read_hw(event, buf, count);
2074 }
2075
2076 static unsigned int perf_poll(struct file *file, poll_table *wait)
2077 {
2078         struct perf_event *event = file->private_data;
2079         struct perf_mmap_data *data;
2080         unsigned int events = POLL_HUP;
2081
2082         rcu_read_lock();
2083         data = rcu_dereference(event->data);
2084         if (data)
2085                 events = atomic_xchg(&data->poll, 0);
2086         rcu_read_unlock();
2087
2088         poll_wait(file, &event->waitq, wait);
2089
2090         return events;
2091 }
2092
2093 static void perf_event_reset(struct perf_event *event)
2094 {
2095         (void)perf_event_read(event);
2096         atomic64_set(&event->count, 0);
2097         perf_event_update_userpage(event);
2098 }
2099
2100 /*
2101  * Holding the top-level event's child_mutex means that any
2102  * descendant process that has inherited this event will block
2103  * in sync_child_event if it goes to exit, thus satisfying the
2104  * task existence requirements of perf_event_enable/disable.
2105  */
2106 static void perf_event_for_each_child(struct perf_event *event,
2107                                         void (*func)(struct perf_event *))
2108 {
2109         struct perf_event *child;
2110
2111         WARN_ON_ONCE(event->ctx->parent_ctx);
2112         mutex_lock(&event->child_mutex);
2113         func(event);
2114         list_for_each_entry(child, &event->child_list, child_list)
2115                 func(child);
2116         mutex_unlock(&event->child_mutex);
2117 }
2118
2119 static void perf_event_for_each(struct perf_event *event,
2120                                   void (*func)(struct perf_event *))
2121 {
2122         struct perf_event_context *ctx = event->ctx;
2123         struct perf_event *sibling;
2124
2125         WARN_ON_ONCE(ctx->parent_ctx);
2126         mutex_lock(&ctx->mutex);
2127         event = event->group_leader;
2128
2129         perf_event_for_each_child(event, func);
2130         func(event);
2131         list_for_each_entry(sibling, &event->sibling_list, group_entry)
2132                 perf_event_for_each_child(event, func);
2133         mutex_unlock(&ctx->mutex);
2134 }
2135
2136 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2137 {
2138         struct perf_event_context *ctx = event->ctx;
2139         unsigned long size;
2140         int ret = 0;
2141         u64 value;
2142
2143         if (!event->attr.sample_period)
2144                 return -EINVAL;
2145
2146         size = copy_from_user(&value, arg, sizeof(value));
2147         if (size != sizeof(value))
2148                 return -EFAULT;
2149
2150         if (!value)
2151                 return -EINVAL;
2152
2153         raw_spin_lock_irq(&ctx->lock);
2154         if (event->attr.freq) {
2155                 if (value > sysctl_perf_event_sample_rate) {
2156                         ret = -EINVAL;
2157                         goto unlock;
2158                 }
2159
2160                 event->attr.sample_freq = value;
2161         } else {
2162                 event->attr.sample_period = value;
2163                 event->hw.sample_period = value;
2164         }
2165 unlock:
2166         raw_spin_unlock_irq(&ctx->lock);
2167
2168         return ret;
2169 }
2170
2171 static int perf_event_set_output(struct perf_event *event, int output_fd);
2172 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2173
2174 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2175 {
2176         struct perf_event *event = file->private_data;
2177         void (*func)(struct perf_event *);
2178         u32 flags = arg;
2179
2180         switch (cmd) {
2181         case PERF_EVENT_IOC_ENABLE:
2182                 func = perf_event_enable;
2183                 break;
2184         case PERF_EVENT_IOC_DISABLE:
2185                 func = perf_event_disable;
2186                 break;
2187         case PERF_EVENT_IOC_RESET:
2188                 func = perf_event_reset;
2189                 break;
2190
2191         case PERF_EVENT_IOC_REFRESH:
2192                 return perf_event_refresh(event, arg);
2193
2194         case PERF_EVENT_IOC_PERIOD:
2195                 return perf_event_period(event, (u64 __user *)arg);
2196
2197         case PERF_EVENT_IOC_SET_OUTPUT:
2198                 return perf_event_set_output(event, arg);
2199
2200         case PERF_EVENT_IOC_SET_FILTER:
2201                 return perf_event_set_filter(event, (void __user *)arg);
2202
2203         default:
2204                 return -ENOTTY;
2205         }
2206
2207         if (flags & PERF_IOC_FLAG_GROUP)
2208                 perf_event_for_each(event, func);
2209         else
2210                 perf_event_for_each_child(event, func);
2211
2212         return 0;
2213 }
2214
2215 int perf_event_task_enable(void)
2216 {
2217         struct perf_event *event;
2218
2219         mutex_lock(&current->perf_event_mutex);
2220         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2221                 perf_event_for_each_child(event, perf_event_enable);
2222         mutex_unlock(&current->perf_event_mutex);
2223
2224         return 0;
2225 }
2226
2227 int perf_event_task_disable(void)
2228 {
2229         struct perf_event *event;
2230
2231         mutex_lock(&current->perf_event_mutex);
2232         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2233                 perf_event_for_each_child(event, perf_event_disable);
2234         mutex_unlock(&current->perf_event_mutex);
2235
2236         return 0;
2237 }
2238
2239 #ifndef PERF_EVENT_INDEX_OFFSET
2240 # define PERF_EVENT_INDEX_OFFSET 0
2241 #endif
2242
2243 static int perf_event_index(struct perf_event *event)
2244 {
2245         if (event->state != PERF_EVENT_STATE_ACTIVE)
2246                 return 0;
2247
2248         return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2249 }
2250
2251 /*
2252  * Callers need to ensure there can be no nesting of this function, otherwise
2253  * the seqlock logic goes bad. We can not serialize this because the arch
2254  * code calls this from NMI context.
2255  */
2256 void perf_event_update_userpage(struct perf_event *event)
2257 {
2258         struct perf_event_mmap_page *userpg;
2259         struct perf_mmap_data *data;
2260
2261         rcu_read_lock();
2262         data = rcu_dereference(event->data);
2263         if (!data)
2264                 goto unlock;
2265
2266         userpg = data->user_page;
2267
2268         /*
2269          * Disable preemption so as to not let the corresponding user-space
2270          * spin too long if we get preempted.
2271          */
2272         preempt_disable();
2273         ++userpg->lock;
2274         barrier();
2275         userpg->index = perf_event_index(event);
2276         userpg->offset = atomic64_read(&event->count);
2277         if (event->state == PERF_EVENT_STATE_ACTIVE)
2278                 userpg->offset -= atomic64_read(&event->hw.prev_count);
2279
2280         userpg->time_enabled = event->total_time_enabled +
2281                         atomic64_read(&event->child_total_time_enabled);
2282
2283         userpg->time_running = event->total_time_running +
2284                         atomic64_read(&event->child_total_time_running);
2285
2286         barrier();
2287         ++userpg->lock;
2288         preempt_enable();
2289 unlock:
2290         rcu_read_unlock();
2291 }
2292
2293 static unsigned long perf_data_size(struct perf_mmap_data *data)
2294 {
2295         return data->nr_pages << (PAGE_SHIFT + data->data_order);
2296 }
2297
2298 #ifndef CONFIG_PERF_USE_VMALLOC
2299
2300 /*
2301  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2302  */
2303
2304 static struct page *
2305 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2306 {
2307         if (pgoff > data->nr_pages)
2308                 return NULL;
2309
2310         if (pgoff == 0)
2311                 return virt_to_page(data->user_page);
2312
2313         return virt_to_page(data->data_pages[pgoff - 1]);
2314 }
2315
2316 static struct perf_mmap_data *
2317 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2318 {
2319         struct perf_mmap_data *data;
2320         unsigned long size;
2321         int i;
2322
2323         WARN_ON(atomic_read(&event->mmap_count));
2324
2325         size = sizeof(struct perf_mmap_data);
2326         size += nr_pages * sizeof(void *);
2327
2328         data = kzalloc(size, GFP_KERNEL);
2329         if (!data)
2330                 goto fail;
2331
2332         data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2333         if (!data->user_page)
2334                 goto fail_user_page;
2335
2336         for (i = 0; i < nr_pages; i++) {
2337                 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2338                 if (!data->data_pages[i])
2339                         goto fail_data_pages;
2340         }
2341
2342         data->data_order = 0;
2343         data->nr_pages = nr_pages;
2344
2345         return data;
2346
2347 fail_data_pages:
2348         for (i--; i >= 0; i--)
2349                 free_page((unsigned long)data->data_pages[i]);
2350
2351         free_page((unsigned long)data->user_page);
2352
2353 fail_user_page:
2354         kfree(data);
2355
2356 fail:
2357         return NULL;
2358 }
2359
2360 static void perf_mmap_free_page(unsigned long addr)
2361 {
2362         struct page *page = virt_to_page((void *)addr);
2363
2364         page->mapping = NULL;
2365         __free_page(page);
2366 }
2367
2368 static void perf_mmap_data_free(struct perf_mmap_data *data)
2369 {
2370         int i;
2371
2372         perf_mmap_free_page((unsigned long)data->user_page);
2373         for (i = 0; i < data->nr_pages; i++)
2374                 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2375         kfree(data);
2376 }
2377
2378 #else
2379
2380 /*
2381  * Back perf_mmap() with vmalloc memory.
2382  *
2383  * Required for architectures that have d-cache aliasing issues.
2384  */
2385
2386 static struct page *
2387 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2388 {
2389         if (pgoff > (1UL << data->data_order))
2390                 return NULL;
2391
2392         return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2393 }
2394
2395 static void perf_mmap_unmark_page(void *addr)
2396 {
2397         struct page *page = vmalloc_to_page(addr);
2398
2399         page->mapping = NULL;
2400 }
2401
2402 static void perf_mmap_data_free_work(struct work_struct *work)
2403 {
2404         struct perf_mmap_data *data;
2405         void *base;
2406         int i, nr;
2407
2408         data = container_of(work, struct perf_mmap_data, work);
2409         nr = 1 << data->data_order;
2410
2411         base = data->user_page;
2412         for (i = 0; i < nr + 1; i++)
2413                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2414
2415         vfree(base);
2416         kfree(data);
2417 }
2418
2419 static void perf_mmap_data_free(struct perf_mmap_data *data)
2420 {
2421         schedule_work(&data->work);
2422 }
2423
2424 static struct perf_mmap_data *
2425 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2426 {
2427         struct perf_mmap_data *data;
2428         unsigned long size;
2429         void *all_buf;
2430
2431         WARN_ON(atomic_read(&event->mmap_count));
2432
2433         size = sizeof(struct perf_mmap_data);
2434         size += sizeof(void *);
2435
2436         data = kzalloc(size, GFP_KERNEL);
2437         if (!data)
2438                 goto fail;
2439
2440         INIT_WORK(&data->work, perf_mmap_data_free_work);
2441
2442         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2443         if (!all_buf)
2444                 goto fail_all_buf;
2445
2446         data->user_page = all_buf;
2447         data->data_pages[0] = all_buf + PAGE_SIZE;
2448         data->data_order = ilog2(nr_pages);
2449         data->nr_pages = 1;
2450
2451         return data;
2452
2453 fail_all_buf:
2454         kfree(data);
2455
2456 fail:
2457         return NULL;
2458 }
2459
2460 #endif
2461
2462 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2463 {
2464         struct perf_event *event = vma->vm_file->private_data;
2465         struct perf_mmap_data *data;
2466         int ret = VM_FAULT_SIGBUS;
2467
2468         if (vmf->flags & FAULT_FLAG_MKWRITE) {
2469                 if (vmf->pgoff == 0)
2470                         ret = 0;
2471                 return ret;
2472         }
2473
2474         rcu_read_lock();
2475         data = rcu_dereference(event->data);
2476         if (!data)
2477                 goto unlock;
2478
2479         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2480                 goto unlock;
2481
2482         vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2483         if (!vmf->page)
2484                 goto unlock;
2485
2486         get_page(vmf->page);
2487         vmf->page->mapping = vma->vm_file->f_mapping;
2488         vmf->page->index   = vmf->pgoff;
2489
2490         ret = 0;
2491 unlock:
2492         rcu_read_unlock();
2493
2494         return ret;
2495 }
2496
2497 static void
2498 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2499 {
2500         long max_size = perf_data_size(data);
2501
2502         atomic_set(&data->lock, -1);
2503
2504         if (event->attr.watermark) {
2505                 data->watermark = min_t(long, max_size,
2506                                         event->attr.wakeup_watermark);
2507         }
2508
2509         if (!data->watermark)
2510                 data->watermark = max_size / 2;
2511
2512
2513         rcu_assign_pointer(event->data, data);
2514 }
2515
2516 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2517 {
2518         struct perf_mmap_data *data;
2519
2520         data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2521         perf_mmap_data_free(data);
2522 }
2523
2524 static void perf_mmap_data_release(struct perf_event *event)
2525 {
2526         struct perf_mmap_data *data = event->data;
2527
2528         WARN_ON(atomic_read(&event->mmap_count));
2529
2530         rcu_assign_pointer(event->data, NULL);
2531         call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2532 }
2533
2534 static void perf_mmap_open(struct vm_area_struct *vma)
2535 {
2536         struct perf_event *event = vma->vm_file->private_data;
2537
2538         atomic_inc(&event->mmap_count);
2539 }
2540
2541 static void perf_mmap_close(struct vm_area_struct *vma)
2542 {
2543         struct perf_event *event = vma->vm_file->private_data;
2544
2545         WARN_ON_ONCE(event->ctx->parent_ctx);
2546         if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2547                 unsigned long size = perf_data_size(event->data);
2548                 struct user_struct *user = current_user();
2549
2550                 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2551                 vma->vm_mm->locked_vm -= event->data->nr_locked;
2552                 perf_mmap_data_release(event);
2553                 mutex_unlock(&event->mmap_mutex);
2554         }
2555 }
2556
2557 static const struct vm_operations_struct perf_mmap_vmops = {
2558         .open           = perf_mmap_open,
2559         .close          = perf_mmap_close,
2560         .fault          = perf_mmap_fault,
2561         .page_mkwrite   = perf_mmap_fault,
2562 };
2563
2564 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2565 {
2566         struct perf_event *event = file->private_data;
2567         unsigned long user_locked, user_lock_limit;
2568         struct user_struct *user = current_user();
2569         unsigned long locked, lock_limit;
2570         struct perf_mmap_data *data;
2571         unsigned long vma_size;
2572         unsigned long nr_pages;
2573         long user_extra, extra;
2574         int ret = 0;
2575
2576         if (!(vma->vm_flags & VM_SHARED))
2577                 return -EINVAL;
2578
2579         vma_size = vma->vm_end - vma->vm_start;
2580         nr_pages = (vma_size / PAGE_SIZE) - 1;
2581
2582         /*
2583          * If we have data pages ensure they're a power-of-two number, so we
2584          * can do bitmasks instead of modulo.
2585          */
2586         if (nr_pages != 0 && !is_power_of_2(nr_pages))
2587                 return -EINVAL;
2588
2589         if (vma_size != PAGE_SIZE * (1 + nr_pages))
2590                 return -EINVAL;
2591
2592         if (vma->vm_pgoff != 0)
2593                 return -EINVAL;
2594
2595         WARN_ON_ONCE(event->ctx->parent_ctx);
2596         mutex_lock(&event->mmap_mutex);
2597         if (event->output) {
2598                 ret = -EINVAL;
2599                 goto unlock;
2600         }
2601
2602         if (atomic_inc_not_zero(&event->mmap_count)) {
2603                 if (nr_pages != event->data->nr_pages)
2604                         ret = -EINVAL;
2605                 goto unlock;
2606         }
2607
2608         user_extra = nr_pages + 1;
2609         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2610
2611         /*
2612          * Increase the limit linearly with more CPUs:
2613          */
2614         user_lock_limit *= num_online_cpus();
2615
2616         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2617
2618         extra = 0;
2619         if (user_locked > user_lock_limit)
2620                 extra = user_locked - user_lock_limit;
2621
2622         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2623         lock_limit >>= PAGE_SHIFT;
2624         locked = vma->vm_mm->locked_vm + extra;
2625
2626         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2627                 !capable(CAP_IPC_LOCK)) {
2628                 ret = -EPERM;
2629                 goto unlock;
2630         }
2631
2632         WARN_ON(event->data);
2633
2634         data = perf_mmap_data_alloc(event, nr_pages);
2635         ret = -ENOMEM;
2636         if (!data)
2637                 goto unlock;
2638
2639         ret = 0;
2640         perf_mmap_data_init(event, data);
2641
2642         atomic_set(&event->mmap_count, 1);
2643         atomic_long_add(user_extra, &user->locked_vm);
2644         vma->vm_mm->locked_vm += extra;
2645         event->data->nr_locked = extra;
2646         if (vma->vm_flags & VM_WRITE)
2647                 event->data->writable = 1;
2648
2649 unlock:
2650         mutex_unlock(&event->mmap_mutex);
2651
2652         vma->vm_flags |= VM_RESERVED;
2653         vma->vm_ops = &perf_mmap_vmops;
2654
2655         return ret;
2656 }
2657
2658 static int perf_fasync(int fd, struct file *filp, int on)
2659 {
2660         struct inode *inode = filp->f_path.dentry->d_inode;
2661         struct perf_event *event = filp->private_data;
2662         int retval;
2663
2664         mutex_lock(&inode->i_mutex);
2665         retval = fasync_helper(fd, filp, on, &event->fasync);
2666         mutex_unlock(&inode->i_mutex);
2667
2668         if (retval < 0)
2669                 return retval;
2670
2671         return 0;
2672 }
2673
2674 static const struct file_operations perf_fops = {
2675         .release                = perf_release,
2676         .read                   = perf_read,
2677         .poll                   = perf_poll,
2678         .unlocked_ioctl         = perf_ioctl,
2679         .compat_ioctl           = perf_ioctl,
2680         .mmap                   = perf_mmap,
2681         .fasync                 = perf_fasync,
2682 };
2683
2684 /*
2685  * Perf event wakeup
2686  *
2687  * If there's data, ensure we set the poll() state and publish everything
2688  * to user-space before waking everybody up.
2689  */
2690
2691 void perf_event_wakeup(struct perf_event *event)
2692 {
2693         wake_up_all(&event->waitq);
2694
2695         if (event->pending_kill) {
2696                 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2697                 event->pending_kill = 0;
2698         }
2699 }
2700
2701 /*
2702  * Pending wakeups
2703  *
2704  * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2705  *
2706  * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2707  * single linked list and use cmpxchg() to add entries lockless.
2708  */
2709
2710 static void perf_pending_event(struct perf_pending_entry *entry)
2711 {
2712         struct perf_event *event = container_of(entry,
2713                         struct perf_event, pending);
2714
2715         if (event->pending_disable) {
2716                 event->pending_disable = 0;
2717                 __perf_event_disable(event);
2718         }
2719
2720         if (event->pending_wakeup) {
2721                 event->pending_wakeup = 0;
2722                 perf_event_wakeup(event);
2723         }
2724 }
2725
2726 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2727
2728 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2729         PENDING_TAIL,
2730 };
2731
2732 static void perf_pending_queue(struct perf_pending_entry *entry,
2733                                void (*func)(struct perf_pending_entry *))
2734 {
2735         struct perf_pending_entry **head;
2736
2737         if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2738                 return;
2739
2740         entry->func = func;
2741
2742         head = &get_cpu_var(perf_pending_head);
2743
2744         do {
2745                 entry->next = *head;
2746         } while (cmpxchg(head, entry->next, entry) != entry->next);
2747
2748         set_perf_event_pending();
2749
2750         put_cpu_var(perf_pending_head);
2751 }
2752
2753 static int __perf_pending_run(void)
2754 {
2755         struct perf_pending_entry *list;
2756         int nr = 0;
2757
2758         list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2759         while (list != PENDING_TAIL) {
2760                 void (*func)(struct perf_pending_entry *);
2761                 struct perf_pending_entry *entry = list;
2762
2763                 list = list->next;
2764
2765                 func = entry->func;
2766                 entry->next = NULL;
2767                 /*
2768                  * Ensure we observe the unqueue before we issue the wakeup,
2769                  * so that we won't be waiting forever.
2770                  * -- see perf_not_pending().
2771                  */
2772                 smp_wmb();
2773
2774                 func(entry);
2775                 nr++;
2776         }
2777
2778         return nr;
2779 }
2780
2781 static inline int perf_not_pending(struct perf_event *event)
2782 {
2783         /*
2784          * If we flush on whatever cpu we run, there is a chance we don't
2785          * need to wait.
2786          */
2787         get_cpu();
2788         __perf_pending_run();
2789         put_cpu();
2790
2791         /*
2792          * Ensure we see the proper queue state before going to sleep
2793          * so that we do not miss the wakeup. -- see perf_pending_handle()
2794          */
2795         smp_rmb();
2796         return event->pending.next == NULL;
2797 }
2798
2799 static void perf_pending_sync(struct perf_event *event)
2800 {
2801         wait_event(event->waitq, perf_not_pending(event));
2802 }
2803
2804 void perf_event_do_pending(void)
2805 {
2806         __perf_pending_run();
2807 }
2808
2809 /*
2810  * Callchain support -- arch specific
2811  */
2812
2813 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2814 {
2815         return NULL;
2816 }
2817
2818 /*
2819  * Output
2820  */
2821 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2822                               unsigned long offset, unsigned long head)
2823 {
2824         unsigned long mask;
2825
2826         if (!data->writable)
2827                 return true;
2828
2829         mask = perf_data_size(data) - 1;
2830
2831         offset = (offset - tail) & mask;
2832         head   = (head   - tail) & mask;
2833
2834         if ((int)(head - offset) < 0)
2835                 return false;
2836
2837         return true;
2838 }
2839
2840 static void perf_output_wakeup(struct perf_output_handle *handle)
2841 {
2842         atomic_set(&handle->data->poll, POLL_IN);
2843
2844         if (handle->nmi) {
2845                 handle->event->pending_wakeup = 1;
2846                 perf_pending_queue(&handle->event->pending,
2847                                    perf_pending_event);
2848         } else
2849                 perf_event_wakeup(handle->event);
2850 }
2851
2852 /*
2853  * Curious locking construct.
2854  *
2855  * We need to ensure a later event_id doesn't publish a head when a former
2856  * event_id isn't done writing. However since we need to deal with NMIs we
2857  * cannot fully serialize things.
2858  *
2859  * What we do is serialize between CPUs so we only have to deal with NMI
2860  * nesting on a single CPU.
2861  *
2862  * We only publish the head (and generate a wakeup) when the outer-most
2863  * event_id completes.
2864  */
2865 static void perf_output_lock(struct perf_output_handle *handle)
2866 {
2867         struct perf_mmap_data *data = handle->data;
2868         int cur, cpu = get_cpu();
2869
2870         handle->locked = 0;
2871
2872         for (;;) {
2873                 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2874                 if (cur == -1) {
2875                         handle->locked = 1;
2876                         break;
2877                 }
2878                 if (cur == cpu)
2879                         break;
2880
2881                 cpu_relax();
2882         }
2883 }
2884
2885 static void perf_output_unlock(struct perf_output_handle *handle)
2886 {
2887         struct perf_mmap_data *data = handle->data;
2888         unsigned long head;
2889         int cpu;
2890
2891         data->done_head = data->head;
2892
2893         if (!handle->locked)
2894                 goto out;
2895
2896 again:
2897         /*
2898          * The xchg implies a full barrier that ensures all writes are done
2899          * before we publish the new head, matched by a rmb() in userspace when
2900          * reading this position.
2901          */
2902         while ((head = atomic_long_xchg(&data->done_head, 0)))
2903                 data->user_page->data_head = head;
2904
2905         /*
2906          * NMI can happen here, which means we can miss a done_head update.
2907          */
2908
2909         cpu = atomic_xchg(&data->lock, -1);
2910         WARN_ON_ONCE(cpu != smp_processor_id());
2911
2912         /*
2913          * Therefore we have to validate we did not indeed do so.
2914          */
2915         if (unlikely(atomic_long_read(&data->done_head))) {
2916                 /*
2917                  * Since we had it locked, we can lock it again.
2918                  */
2919                 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2920                         cpu_relax();
2921
2922                 goto again;
2923         }
2924
2925         if (atomic_xchg(&data->wakeup, 0))
2926                 perf_output_wakeup(handle);
2927 out:
2928         put_cpu();
2929 }
2930
2931 void perf_output_copy(struct perf_output_handle *handle,
2932                       const void *buf, unsigned int len)
2933 {
2934         unsigned int pages_mask;
2935         unsigned long offset;
2936         unsigned int size;
2937         void **pages;
2938
2939         offset          = handle->offset;
2940         pages_mask      = handle->data->nr_pages - 1;
2941         pages           = handle->data->data_pages;
2942
2943         do {
2944                 unsigned long page_offset;
2945                 unsigned long page_size;
2946                 int nr;
2947
2948                 nr          = (offset >> PAGE_SHIFT) & pages_mask;
2949                 page_size   = 1UL << (handle->data->data_order + PAGE_SHIFT);
2950                 page_offset = offset & (page_size - 1);
2951                 size        = min_t(unsigned int, page_size - page_offset, len);
2952
2953                 memcpy(pages[nr] + page_offset, buf, size);
2954
2955                 len         -= size;
2956                 buf         += size;
2957                 offset      += size;
2958         } while (len);
2959
2960         handle->offset = offset;
2961
2962         /*
2963          * Check we didn't copy past our reservation window, taking the
2964          * possible unsigned int wrap into account.
2965          */
2966         WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2967 }
2968
2969 int perf_output_begin(struct perf_output_handle *handle,
2970                       struct perf_event *event, unsigned int size,
2971                       int nmi, int sample)
2972 {
2973         struct perf_event *output_event;
2974         struct perf_mmap_data *data;
2975         unsigned long tail, offset, head;
2976         int have_lost;
2977         struct {
2978                 struct perf_event_header header;
2979                 u64                      id;
2980                 u64                      lost;
2981         } lost_event;
2982
2983         rcu_read_lock();
2984         /*
2985          * For inherited events we send all the output towards the parent.
2986          */
2987         if (event->parent)
2988                 event = event->parent;
2989
2990         output_event = rcu_dereference(event->output);
2991         if (output_event)
2992                 event = output_event;
2993
2994         data = rcu_dereference(event->data);
2995         if (!data)
2996                 goto out;
2997
2998         handle->data    = data;
2999         handle->event   = event;
3000         handle->nmi     = nmi;
3001         handle->sample  = sample;
3002
3003         if (!data->nr_pages)
3004                 goto fail;
3005
3006         have_lost = atomic_read(&data->lost);
3007         if (have_lost)
3008                 size += sizeof(lost_event);
3009
3010         perf_output_lock(handle);
3011
3012         do {
3013                 /*
3014                  * Userspace could choose to issue a mb() before updating the
3015                  * tail pointer. So that all reads will be completed before the
3016                  * write is issued.
3017                  */
3018                 tail = ACCESS_ONCE(data->user_page->data_tail);
3019                 smp_rmb();
3020                 offset = head = atomic_long_read(&data->head);
3021                 head += size;
3022                 if (unlikely(!perf_output_space(data, tail, offset, head)))
3023                         goto fail;
3024         } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
3025
3026         handle->offset  = offset;
3027         handle->head    = head;
3028
3029         if (head - tail > data->watermark)
3030                 atomic_set(&data->wakeup, 1);
3031
3032         if (have_lost) {
3033                 lost_event.header.type = PERF_RECORD_LOST;
3034                 lost_event.header.misc = 0;
3035                 lost_event.header.size = sizeof(lost_event);
3036                 lost_event.id          = event->id;
3037                 lost_event.lost        = atomic_xchg(&data->lost, 0);
3038
3039                 perf_output_put(handle, lost_event);
3040         }
3041
3042         return 0;
3043
3044 fail:
3045         atomic_inc(&data->lost);
3046         perf_output_unlock(handle);
3047 out:
3048         rcu_read_unlock();
3049
3050         return -ENOSPC;
3051 }
3052
3053 void perf_output_end(struct perf_output_handle *handle)
3054 {
3055         struct perf_event *event = handle->event;
3056         struct perf_mmap_data *data = handle->data;
3057
3058         int wakeup_events = event->attr.wakeup_events;
3059
3060         if (handle->sample && wakeup_events) {
3061                 int events = atomic_inc_return(&data->events);
3062                 if (events >= wakeup_events) {
3063                         atomic_sub(wakeup_events, &data->events);
3064                         atomic_set(&data->wakeup, 1);
3065                 }
3066         }
3067
3068         perf_output_unlock(handle);
3069         rcu_read_unlock();
3070 }
3071
3072 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3073 {
3074         /*
3075          * only top level events have the pid namespace they were created in
3076          */
3077         if (event->parent)
3078                 event = event->parent;
3079
3080         return task_tgid_nr_ns(p, event->ns);
3081 }
3082
3083 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3084 {
3085         /*
3086          * only top level events have the pid namespace they were created in
3087          */
3088         if (event->parent)
3089                 event = event->parent;
3090
3091         return task_pid_nr_ns(p, event->ns);
3092 }
3093
3094 static void perf_output_read_one(struct perf_output_handle *handle,
3095                                  struct perf_event *event)
3096 {
3097         u64 read_format = event->attr.read_format;
3098         u64 values[4];
3099         int n = 0;
3100
3101         values[n++] = atomic64_read(&event->count);
3102         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3103                 values[n++] = event->total_time_enabled +
3104                         atomic64_read(&event->child_total_time_enabled);
3105         }
3106         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3107                 values[n++] = event->total_time_running +
3108                         atomic64_read(&event->child_total_time_running);
3109         }
3110         if (read_format & PERF_FORMAT_ID)
3111                 values[n++] = primary_event_id(event);
3112
3113         perf_output_copy(handle, values, n * sizeof(u64));
3114 }
3115
3116 /*
3117  * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3118  */
3119 static void perf_output_read_group(struct perf_output_handle *handle,
3120                             struct perf_event *event)
3121 {
3122         struct perf_event *leader = event->group_leader, *sub;
3123         u64 read_format = event->attr.read_format;
3124         u64 values[5];
3125         int n = 0;
3126
3127         values[n++] = 1 + leader->nr_siblings;
3128
3129         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3130                 values[n++] = leader->total_time_enabled;
3131
3132         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3133                 values[n++] = leader->total_time_running;
3134
3135         if (leader != event)
3136                 leader->pmu->read(leader);
3137
3138         values[n++] = atomic64_read(&leader->count);
3139         if (read_format & PERF_FORMAT_ID)
3140                 values[n++] = primary_event_id(leader);
3141
3142         perf_output_copy(handle, values, n * sizeof(u64));
3143
3144         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3145                 n = 0;
3146
3147                 if (sub != event)
3148                         sub->pmu->read(sub);
3149
3150                 values[n++] = atomic64_read(&sub->count);
3151                 if (read_format & PERF_FORMAT_ID)
3152                         values[n++] = primary_event_id(sub);
3153
3154                 perf_output_copy(handle, values, n * sizeof(u64));
3155         }
3156 }
3157
3158 static void perf_output_read(struct perf_output_handle *handle,
3159                              struct perf_event *event)
3160 {
3161         if (event->attr.read_format & PERF_FORMAT_GROUP)
3162                 perf_output_read_group(handle, event);
3163         else
3164                 perf_output_read_one(handle, event);
3165 }
3166
3167 void perf_output_sample(struct perf_output_handle *handle,
3168                         struct perf_event_header *header,
3169                         struct perf_sample_data *data,
3170                         struct perf_event *event)
3171 {
3172         u64 sample_type = data->type;
3173
3174         perf_output_put(handle, *header);
3175
3176         if (sample_type & PERF_SAMPLE_IP)
3177                 perf_output_put(handle, data->ip);
3178
3179         if (sample_type & PERF_SAMPLE_TID)
3180                 perf_output_put(handle, data->tid_entry);
3181
3182         if (sample_type & PERF_SAMPLE_TIME)
3183                 perf_output_put(handle, data->time);
3184
3185         if (sample_type & PERF_SAMPLE_ADDR)
3186                 perf_output_put(handle, data->addr);
3187
3188         if (sample_type & PERF_SAMPLE_ID)
3189                 perf_output_put(handle, data->id);
3190
3191         if (sample_type & PERF_SAMPLE_STREAM_ID)
3192                 perf_output_put(handle, data->stream_id);
3193
3194         if (sample_type & PERF_SAMPLE_CPU)
3195                 perf_output_put(handle, data->cpu_entry);
3196
3197         if (sample_type & PERF_SAMPLE_PERIOD)
3198                 perf_output_put(handle, data->period);
3199
3200         if (sample_type & PERF_SAMPLE_READ)
3201                 perf_output_read(handle, event);
3202
3203         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3204                 if (data->callchain) {
3205                         int size = 1;
3206
3207                         if (data->callchain)
3208                                 size += data->callchain->nr;
3209
3210                         size *= sizeof(u64);
3211
3212                         perf_output_copy(handle, data->callchain, size);
3213                 } else {
3214                         u64 nr = 0;
3215                         perf_output_put(handle, nr);
3216                 }
3217         }
3218
3219         if (sample_type & PERF_SAMPLE_RAW) {
3220                 if (data->raw) {
3221                         perf_output_put(handle, data->raw->size);
3222                         perf_output_copy(handle, data->raw->data,
3223                                          data->raw->size);
3224                 } else {
3225                         struct {
3226                                 u32     size;
3227                                 u32     data;
3228                         } raw = {
3229                                 .size = sizeof(u32),
3230                                 .data = 0,
3231                         };
3232                         perf_output_put(handle, raw);
3233                 }
3234         }
3235 }
3236
3237 void perf_prepare_sample(struct perf_event_header *header,
3238                          struct perf_sample_data *data,
3239                          struct perf_event *event,
3240                          struct pt_regs *regs)
3241 {
3242         u64 sample_type = event->attr.sample_type;
3243
3244         data->type = sample_type;
3245
3246         header->type = PERF_RECORD_SAMPLE;
3247         header->size = sizeof(*header);
3248
3249         header->misc = 0;
3250         header->misc |= perf_misc_flags(regs);
3251
3252         if (sample_type & PERF_SAMPLE_IP) {
3253                 data->ip = perf_instruction_pointer(regs);
3254
3255                 header->size += sizeof(data->ip);
3256         }
3257
3258         if (sample_type & PERF_SAMPLE_TID) {
3259                 /* namespace issues */
3260                 data->tid_entry.pid = perf_event_pid(event, current);
3261                 data->tid_entry.tid = perf_event_tid(event, current);
3262
3263                 header->size += sizeof(data->tid_entry);
3264         }
3265
3266         if (sample_type & PERF_SAMPLE_TIME) {
3267                 data->time = perf_clock();
3268
3269                 header->size += sizeof(data->time);
3270         }
3271
3272         if (sample_type & PERF_SAMPLE_ADDR)
3273                 header->size += sizeof(data->addr);
3274
3275         if (sample_type & PERF_SAMPLE_ID) {
3276                 data->id = primary_event_id(event);
3277
3278                 header->size += sizeof(data->id);
3279         }
3280
3281         if (sample_type & PERF_SAMPLE_STREAM_ID) {
3282                 data->stream_id = event->id;
3283
3284                 header->size += sizeof(data->stream_id);
3285         }
3286
3287         if (sample_type & PERF_SAMPLE_CPU) {
3288                 data->cpu_entry.cpu             = raw_smp_processor_id();
3289                 data->cpu_entry.reserved        = 0;
3290
3291                 header->size += sizeof(data->cpu_entry);
3292         }
3293
3294         if (sample_type & PERF_SAMPLE_PERIOD)
3295                 header->size += sizeof(data->period);
3296
3297         if (sample_type & PERF_SAMPLE_READ)
3298                 header->size += perf_event_read_size(event);
3299
3300         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3301                 int size = 1;
3302
3303                 data->callchain = perf_callchain(regs);
3304
3305                 if (data->callchain)
3306                         size += data->callchain->nr;
3307
3308                 header->size += size * sizeof(u64);
3309         }
3310
3311         if (sample_type & PERF_SAMPLE_RAW) {
3312                 int size = sizeof(u32);
3313
3314                 if (data->raw)
3315                         size += data->raw->size;
3316                 else
3317                         size += sizeof(u32);
3318
3319                 WARN_ON_ONCE(size & (sizeof(u64)-1));
3320                 header->size += size;
3321         }
3322 }
3323
3324 static void perf_event_output(struct perf_event *event, int nmi,
3325                                 struct perf_sample_data *data,
3326                                 struct pt_regs *regs)
3327 {
3328         struct perf_output_handle handle;
3329         struct perf_event_header header;
3330
3331         perf_prepare_sample(&header, data, event, regs);
3332
3333         if (perf_output_begin(&handle, event, header.size, nmi, 1))
3334                 return;
3335
3336         perf_output_sample(&handle, &header, data, event);
3337
3338         perf_output_end(&handle);
3339 }
3340
3341 /*
3342  * read event_id
3343  */
3344
3345 struct perf_read_event {
3346         struct perf_event_header        header;
3347
3348         u32                             pid;
3349         u32                             tid;
3350 };
3351
3352 static void
3353 perf_event_read_event(struct perf_event *event,
3354                         struct task_struct *task)
3355 {
3356         struct perf_output_handle handle;
3357         struct perf_read_event read_event = {
3358                 .header = {
3359                         .type = PERF_RECORD_READ,
3360                         .misc = 0,
3361                         .size = sizeof(read_event) + perf_event_read_size(event),
3362                 },
3363                 .pid = perf_event_pid(event, task),
3364                 .tid = perf_event_tid(event, task),
3365         };
3366         int ret;
3367
3368         ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3369         if (ret)
3370                 return;
3371
3372         perf_output_put(&handle, read_event);
3373         perf_output_read(&handle, event);
3374
3375         perf_output_end(&handle);
3376 }
3377
3378 /*
3379  * task tracking -- fork/exit
3380  *
3381  * enabled by: attr.comm | attr.mmap | attr.task
3382  */
3383
3384 struct perf_task_event {
3385         struct task_struct              *task;
3386         struct perf_event_context       *task_ctx;
3387
3388         struct {
3389                 struct perf_event_header        header;
3390
3391                 u32                             pid;
3392                 u32                             ppid;
3393                 u32                             tid;
3394                 u32                             ptid;
3395                 u64                             time;
3396         } event_id;
3397 };
3398
3399 static void perf_event_task_output(struct perf_event *event,
3400                                      struct perf_task_event *task_event)
3401 {
3402         struct perf_output_handle handle;
3403         int size;
3404         struct task_struct *task = task_event->task;
3405         int ret;
3406
3407         size  = task_event->event_id.header.size;
3408         ret = perf_output_begin(&handle, event, size, 0, 0);
3409
3410         if (ret)
3411                 return;
3412
3413         task_event->event_id.pid = perf_event_pid(event, task);
3414         task_event->event_id.ppid = perf_event_pid(event, current);
3415
3416         task_event->event_id.tid = perf_event_tid(event, task);
3417         task_event->event_id.ptid = perf_event_tid(event, current);
3418
3419         task_event->event_id.time = perf_clock();
3420
3421         perf_output_put(&handle, task_event->event_id);
3422
3423         perf_output_end(&handle);
3424 }
3425
3426 static int perf_event_task_match(struct perf_event *event)
3427 {
3428         if (event->state != PERF_EVENT_STATE_ACTIVE)
3429                 return 0;
3430
3431         if (event->cpu != -1 && event->cpu != smp_processor_id())
3432                 return 0;
3433
3434         if (event->attr.comm || event->attr.mmap || event->attr.task)
3435                 return 1;
3436
3437         return 0;
3438 }
3439
3440 static void perf_event_task_ctx(struct perf_event_context *ctx,
3441                                   struct perf_task_event *task_event)
3442 {
3443         struct perf_event *event;
3444
3445         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3446                 if (perf_event_task_match(event))
3447                         perf_event_task_output(event, task_event);
3448         }
3449 }
3450
3451 static void perf_event_task_event(struct perf_task_event *task_event)
3452 {
3453         struct perf_cpu_context *cpuctx;
3454         struct perf_event_context *ctx = task_event->task_ctx;
3455
3456         rcu_read_lock();
3457         cpuctx = &get_cpu_var(perf_cpu_context);
3458         perf_event_task_ctx(&cpuctx->ctx, task_event);
3459         if (!ctx)
3460                 ctx = rcu_dereference(task_event->task->perf_event_ctxp);
3461         if (ctx)
3462                 perf_event_task_ctx(ctx, task_event);
3463         put_cpu_var(perf_cpu_context);
3464         rcu_read_unlock();
3465 }
3466
3467 static void perf_event_task(struct task_struct *task,
3468                               struct perf_event_context *task_ctx,
3469                               int new)
3470 {
3471         struct perf_task_event task_event;
3472
3473         if (!atomic_read(&nr_comm_events) &&
3474             !atomic_read(&nr_mmap_events) &&
3475             !atomic_read(&nr_task_events))
3476                 return;
3477
3478         task_event = (struct perf_task_event){
3479                 .task     = task,
3480                 .task_ctx = task_ctx,
3481                 .event_id    = {
3482                         .header = {
3483                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3484                                 .misc = 0,
3485                                 .size = sizeof(task_event.event_id),
3486                         },
3487                         /* .pid  */
3488                         /* .ppid */
3489                         /* .tid  */
3490                         /* .ptid */
3491                 },
3492         };
3493
3494         perf_event_task_event(&task_event);
3495 }
3496
3497 void perf_event_fork(struct task_struct *task)
3498 {
3499         perf_event_task(task, NULL, 1);
3500 }
3501
3502 /*
3503  * comm tracking
3504  */
3505
3506 struct perf_comm_event {
3507         struct task_struct      *task;
3508         char                    *comm;
3509         int                     comm_size;
3510
3511         struct {
3512                 struct perf_event_header        header;
3513
3514                 u32                             pid;
3515                 u32                             tid;
3516         } event_id;
3517 };
3518
3519 static void perf_event_comm_output(struct perf_event *event,
3520                                      struct perf_comm_event *comm_event)
3521 {
3522         struct perf_output_handle handle;
3523         int size = comm_event->event_id.header.size;
3524         int ret = perf_output_begin(&handle, event, size, 0, 0);
3525
3526         if (ret)
3527                 return;
3528
3529         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3530         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3531
3532         perf_output_put(&handle, comm_event->event_id);
3533         perf_output_copy(&handle, comm_event->comm,
3534                                    comm_event->comm_size);
3535         perf_output_end(&handle);
3536 }
3537
3538 static int perf_event_comm_match(struct perf_event *event)
3539 {
3540         if (event->state != PERF_EVENT_STATE_ACTIVE)
3541                 return 0;
3542
3543         if (event->cpu != -1 && event->cpu != smp_processor_id())
3544                 return 0;
3545
3546         if (event->attr.comm)
3547                 return 1;
3548
3549         return 0;
3550 }
3551
3552 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3553                                   struct perf_comm_event *comm_event)
3554 {
3555         struct perf_event *event;
3556
3557         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3558                 if (perf_event_comm_match(event))
3559                         perf_event_comm_output(event, comm_event);
3560         }
3561 }
3562
3563 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3564 {
3565         struct perf_cpu_context *cpuctx;
3566         struct perf_event_context *ctx;
3567         unsigned int size;
3568         char comm[TASK_COMM_LEN];
3569
3570         memset(comm, 0, sizeof(comm));
3571         strlcpy(comm, comm_event->task->comm, sizeof(comm));
3572         size = ALIGN(strlen(comm)+1, sizeof(u64));
3573
3574         comm_event->comm = comm;
3575         comm_event->comm_size = size;
3576
3577         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3578
3579         rcu_read_lock();
3580         cpuctx = &get_cpu_var(perf_cpu_context);
3581         perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3582         ctx = rcu_dereference(current->perf_event_ctxp);
3583         if (ctx)
3584                 perf_event_comm_ctx(ctx, comm_event);
3585         put_cpu_var(perf_cpu_context);
3586         rcu_read_unlock();
3587 }
3588
3589 void perf_event_comm(struct task_struct *task)
3590 {
3591         struct perf_comm_event comm_event;
3592
3593         if (task->perf_event_ctxp)
3594                 perf_event_enable_on_exec(task);
3595
3596         if (!atomic_read(&nr_comm_events))
3597                 return;
3598
3599         comm_event = (struct perf_comm_event){
3600                 .task   = task,
3601                 /* .comm      */
3602                 /* .comm_size */
3603                 .event_id  = {
3604                         .header = {
3605                                 .type = PERF_RECORD_COMM,
3606                                 .misc = 0,
3607                                 /* .size */
3608                         },
3609                         /* .pid */
3610                         /* .tid */
3611                 },
3612         };
3613
3614         perf_event_comm_event(&comm_event);
3615 }
3616
3617 /*
3618  * mmap tracking
3619  */
3620
3621 struct perf_mmap_event {
3622         struct vm_area_struct   *vma;
3623
3624         const char              *file_name;
3625         int                     file_size;
3626
3627         struct {
3628                 struct perf_event_header        header;
3629
3630                 u32                             pid;
3631                 u32                             tid;
3632                 u64                             start;
3633                 u64                             len;
3634                 u64                             pgoff;
3635         } event_id;
3636 };
3637
3638 static void perf_event_mmap_output(struct perf_event *event,
3639                                      struct perf_mmap_event *mmap_event)
3640 {
3641         struct perf_output_handle handle;
3642         int size = mmap_event->event_id.header.size;
3643         int ret = perf_output_begin(&handle, event, size, 0, 0);
3644
3645         if (ret)
3646                 return;
3647
3648         mmap_event->event_id.pid = perf_event_pid(event, current);
3649         mmap_event->event_id.tid = perf_event_tid(event, current);
3650
3651         perf_output_put(&handle, mmap_event->event_id);
3652         perf_output_copy(&handle, mmap_event->file_name,
3653                                    mmap_event->file_size);
3654         perf_output_end(&handle);
3655 }
3656
3657 static int perf_event_mmap_match(struct perf_event *event,
3658                                    struct perf_mmap_event *mmap_event)
3659 {
3660         if (event->state != PERF_EVENT_STATE_ACTIVE)
3661                 return 0;
3662
3663         if (event->cpu != -1 && event->cpu != smp_processor_id())
3664                 return 0;
3665
3666         if (event->attr.mmap)
3667                 return 1;
3668
3669         return 0;
3670 }
3671
3672 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3673                                   struct perf_mmap_event *mmap_event)
3674 {
3675         struct perf_event *event;
3676
3677         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3678                 if (perf_event_mmap_match(event, mmap_event))
3679                         perf_event_mmap_output(event, mmap_event);
3680         }
3681 }
3682
3683 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3684 {
3685         struct perf_cpu_context *cpuctx;
3686         struct perf_event_context *ctx;
3687         struct vm_area_struct *vma = mmap_event->vma;
3688         struct file *file = vma->vm_file;
3689         unsigned int size;
3690         char tmp[16];
3691         char *buf = NULL;
3692         const char *name;
3693
3694         memset(tmp, 0, sizeof(tmp));
3695
3696         if (file) {
3697                 /*
3698                  * d_path works from the end of the buffer backwards, so we
3699                  * need to add enough zero bytes after the string to handle
3700                  * the 64bit alignment we do later.
3701                  */
3702                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3703                 if (!buf) {
3704                         name = strncpy(tmp, "//enomem", sizeof(tmp));
3705                         goto got_name;
3706                 }
3707                 name = d_path(&file->f_path, buf, PATH_MAX);
3708                 if (IS_ERR(name)) {
3709                         name = strncpy(tmp, "//toolong", sizeof(tmp));
3710                         goto got_name;
3711                 }
3712         } else {
3713                 if (arch_vma_name(mmap_event->vma)) {
3714                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3715                                        sizeof(tmp));
3716                         goto got_name;
3717                 }
3718
3719                 if (!vma->vm_mm) {
3720                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
3721                         goto got_name;
3722                 }
3723
3724                 name = strncpy(tmp, "//anon", sizeof(tmp));
3725                 goto got_name;
3726         }
3727
3728 got_name:
3729         size = ALIGN(strlen(name)+1, sizeof(u64));
3730
3731         mmap_event->file_name = name;
3732         mmap_event->file_size = size;
3733
3734         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3735
3736         rcu_read_lock();
3737         cpuctx = &get_cpu_var(perf_cpu_context);
3738         perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3739         ctx = rcu_dereference(current->perf_event_ctxp);
3740         if (ctx)
3741                 perf_event_mmap_ctx(ctx, mmap_event);
3742         put_cpu_var(perf_cpu_context);
3743         rcu_read_unlock();
3744
3745         kfree(buf);
3746 }
3747
3748 void __perf_event_mmap(struct vm_area_struct *vma)
3749 {
3750         struct perf_mmap_event mmap_event;
3751
3752         if (!atomic_read(&nr_mmap_events))
3753                 return;
3754
3755         mmap_event = (struct perf_mmap_event){
3756                 .vma    = vma,
3757                 /* .file_name */
3758                 /* .file_size */
3759                 .event_id  = {
3760                         .header = {
3761                                 .type = PERF_RECORD_MMAP,
3762                                 .misc = 0,
3763                                 /* .size */
3764                         },
3765                         /* .pid */
3766                         /* .tid */
3767                         .start  = vma->vm_start,
3768                         .len    = vma->vm_end - vma->vm_start,
3769                         .pgoff  = (u64)vma->vm_pgoff << PAGE_SHIFT,
3770                 },
3771         };
3772
3773         perf_event_mmap_event(&mmap_event);
3774 }
3775
3776 /*
3777  * IRQ throttle logging
3778  */
3779
3780 static void perf_log_throttle(struct perf_event *event, int enable)
3781 {
3782         struct perf_output_handle handle;
3783         int ret;
3784
3785         struct {
3786                 struct perf_event_header        header;
3787                 u64                             time;
3788                 u64                             id;
3789                 u64                             stream_id;
3790         } throttle_event = {
3791                 .header = {
3792                         .type = PERF_RECORD_THROTTLE,
3793                         .misc = 0,
3794                         .size = sizeof(throttle_event),
3795                 },
3796                 .time           = perf_clock(),
3797                 .id             = primary_event_id(event),
3798                 .stream_id      = event->id,
3799         };
3800
3801         if (enable)
3802                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3803
3804         ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3805         if (ret)
3806                 return;
3807
3808         perf_output_put(&handle, throttle_event);
3809         perf_output_end(&handle);
3810 }
3811
3812 /*
3813  * Generic event overflow handling, sampling.
3814  */
3815
3816 static int __perf_event_overflow(struct perf_event *event, int nmi,
3817                                    int throttle, struct perf_sample_data *data,
3818                                    struct pt_regs *regs)
3819 {
3820         int events = atomic_read(&event->event_limit);
3821         struct hw_perf_event *hwc = &event->hw;
3822         int ret = 0;
3823
3824         throttle = (throttle && event->pmu->unthrottle != NULL);
3825
3826         if (!throttle) {
3827                 hwc->interrupts++;
3828         } else {
3829                 if (hwc->interrupts != MAX_INTERRUPTS) {
3830                         hwc->interrupts++;
3831                         if (HZ * hwc->interrupts >
3832                                         (u64)sysctl_perf_event_sample_rate) {
3833                                 hwc->interrupts = MAX_INTERRUPTS;
3834                                 perf_log_throttle(event, 0);
3835                                 ret = 1;
3836                         }
3837                 } else {
3838                         /*
3839                          * Keep re-disabling events even though on the previous
3840                          * pass we disabled it - just in case we raced with a
3841                          * sched-in and the event got enabled again:
3842                          */
3843                         ret = 1;
3844                 }
3845         }
3846
3847         if (event->attr.freq) {
3848                 u64 now = perf_clock();
3849                 s64 delta = now - hwc->freq_time_stamp;
3850
3851                 hwc->freq_time_stamp = now;
3852
3853                 if (delta > 0 && delta < 2*TICK_NSEC)
3854                         perf_adjust_period(event, delta, hwc->last_period);
3855         }
3856
3857         /*
3858          * XXX event_limit might not quite work as expected on inherited
3859          * events
3860          */
3861
3862         event->pending_kill = POLL_IN;
3863         if (events && atomic_dec_and_test(&event->event_limit)) {
3864                 ret = 1;
3865                 event->pending_kill = POLL_HUP;
3866                 if (nmi) {
3867                         event->pending_disable = 1;
3868                         perf_pending_queue(&event->pending,
3869                                            perf_pending_event);
3870                 } else
3871                         perf_event_disable(event);
3872         }
3873
3874         if (event->overflow_handler)
3875                 event->overflow_handler(event, nmi, data, regs);
3876         else
3877                 perf_event_output(event, nmi, data, regs);
3878
3879         return ret;
3880 }
3881
3882 int perf_event_overflow(struct perf_event *event, int nmi,
3883                           struct perf_sample_data *data,
3884                           struct pt_regs *regs)
3885 {
3886         return __perf_event_overflow(event, nmi, 1, data, regs);
3887 }
3888
3889 /*
3890  * Generic software event infrastructure
3891  */
3892
3893 /*
3894  * We directly increment event->count and keep a second value in
3895  * event->hw.period_left to count intervals. This period event
3896  * is kept in the range [-sample_period, 0] so that we can use the
3897  * sign as trigger.
3898  */
3899
3900 static u64 perf_swevent_set_period(struct perf_event *event)
3901 {
3902         struct hw_perf_event *hwc = &event->hw;
3903         u64 period = hwc->last_period;
3904         u64 nr, offset;
3905         s64 old, val;
3906
3907         hwc->last_period = hwc->sample_period;
3908
3909 again:
3910         old = val = atomic64_read(&hwc->period_left);
3911         if (val < 0)
3912                 return 0;
3913
3914         nr = div64_u64(period + val, period);
3915         offset = nr * period;
3916         val -= offset;
3917         if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3918                 goto again;
3919
3920         return nr;
3921 }
3922
3923 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3924                                     int nmi, struct perf_sample_data *data,
3925                                     struct pt_regs *regs)
3926 {
3927         struct hw_perf_event *hwc = &event->hw;
3928         int throttle = 0;
3929
3930         data->period = event->hw.last_period;
3931         if (!overflow)
3932                 overflow = perf_swevent_set_period(event);
3933
3934         if (hwc->interrupts == MAX_INTERRUPTS)
3935                 return;
3936
3937         for (; overflow; overflow--) {
3938                 if (__perf_event_overflow(event, nmi, throttle,
3939                                             data, regs)) {
3940                         /*
3941                          * We inhibit the overflow from happening when
3942                          * hwc->interrupts == MAX_INTERRUPTS.
3943                          */
3944                         break;
3945                 }
3946                 throttle = 1;
3947         }
3948 }
3949
3950 static void perf_swevent_unthrottle(struct perf_event *event)
3951 {
3952         /*
3953          * Nothing to do, we already reset hwc->interrupts.
3954          */
3955 }
3956
3957 static void perf_swevent_add(struct perf_event *event, u64 nr,
3958                                int nmi, struct perf_sample_data *data,
3959                                struct pt_regs *regs)
3960 {
3961         struct hw_perf_event *hwc = &event->hw;
3962
3963         atomic64_add(nr, &event->count);
3964
3965         if (!regs)
3966                 return;
3967
3968         if (!hwc->sample_period)
3969                 return;
3970
3971         if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3972                 return perf_swevent_overflow(event, 1, nmi, data, regs);
3973
3974         if (atomic64_add_negative(nr, &hwc->period_left))
3975                 return;
3976
3977         perf_swevent_overflow(event, 0, nmi, data, regs);
3978 }
3979
3980 static int perf_swevent_is_counting(struct perf_event *event)
3981 {
3982         /*
3983          * The event is active, we're good!
3984          */
3985         if (event->state == PERF_EVENT_STATE_ACTIVE)
3986                 return 1;
3987
3988         /*
3989          * The event is off/error, not counting.
3990          */
3991         if (event->state != PERF_EVENT_STATE_INACTIVE)
3992                 return 0;
3993
3994         /*
3995          * The event is inactive, if the context is active
3996          * we're part of a group that didn't make it on the 'pmu',
3997          * not counting.
3998          */
3999         if (event->ctx->is_active)
4000                 return 0;
4001
4002         /*
4003          * We're inactive and the context is too, this means the
4004          * task is scheduled out, we're counting events that happen
4005          * to us, like migration events.
4006          */
4007         return 1;
4008 }
4009
4010 static int perf_tp_event_match(struct perf_event *event,
4011                                 struct perf_sample_data *data);
4012
4013 static int perf_exclude_event(struct perf_event *event,
4014                               struct pt_regs *regs)
4015 {
4016         if (regs) {
4017                 if (event->attr.exclude_user && user_mode(regs))
4018                         return 1;
4019
4020                 if (event->attr.exclude_kernel && !user_mode(regs))
4021                         return 1;
4022         }
4023
4024         return 0;
4025 }
4026
4027 static int perf_swevent_match(struct perf_event *event,
4028                                 enum perf_type_id type,
4029                                 u32 event_id,
4030                                 struct perf_sample_data *data,
4031                                 struct pt_regs *regs)
4032 {
4033         if (event->cpu != -1 && event->cpu != smp_processor_id())
4034                 return 0;
4035
4036         if (!perf_swevent_is_counting(event))
4037                 return 0;
4038
4039         if (event->attr.type != type)
4040                 return 0;
4041
4042         if (event->attr.config != event_id)
4043                 return 0;
4044
4045         if (perf_exclude_event(event, regs))
4046                 return 0;
4047
4048         if (event->attr.type == PERF_TYPE_TRACEPOINT &&
4049             !perf_tp_event_match(event, data))
4050                 return 0;
4051
4052         return 1;
4053 }
4054
4055 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
4056                                      enum perf_type_id type,
4057                                      u32 event_id, u64 nr, int nmi,
4058                                      struct perf_sample_data *data,
4059                                      struct pt_regs *regs)
4060 {
4061         struct perf_event *event;
4062
4063         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4064                 if (perf_swevent_match(event, type, event_id, data, regs))
4065                         perf_swevent_add(event, nr, nmi, data, regs);
4066         }
4067 }
4068
4069 int perf_swevent_get_recursion_context(void)
4070 {
4071         struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
4072         int rctx;
4073
4074         if (in_nmi())
4075                 rctx = 3;
4076         else if (in_irq())
4077                 rctx = 2;
4078         else if (in_softirq())
4079                 rctx = 1;
4080         else
4081                 rctx = 0;
4082
4083         if (cpuctx->recursion[rctx]) {
4084                 put_cpu_var(perf_cpu_context);
4085                 return -1;
4086         }
4087
4088         cpuctx->recursion[rctx]++;
4089         barrier();
4090
4091         return rctx;
4092 }
4093 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4094
4095 void perf_swevent_put_recursion_context(int rctx)
4096 {
4097         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4098         barrier();
4099         cpuctx->recursion[rctx]--;
4100         put_cpu_var(perf_cpu_context);
4101 }
4102 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
4103
4104 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4105                                     u64 nr, int nmi,
4106                                     struct perf_sample_data *data,
4107                                     struct pt_regs *regs)
4108 {
4109         struct perf_cpu_context *cpuctx;
4110         struct perf_event_context *ctx;
4111
4112         cpuctx = &__get_cpu_var(perf_cpu_context);
4113         rcu_read_lock();
4114         perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
4115                                  nr, nmi, data, regs);
4116         /*
4117          * doesn't really matter which of the child contexts the
4118          * events ends up in.
4119          */
4120         ctx = rcu_dereference(current->perf_event_ctxp);
4121         if (ctx)
4122                 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
4123         rcu_read_unlock();
4124 }
4125
4126 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4127                             struct pt_regs *regs, u64 addr)
4128 {
4129         struct perf_sample_data data;
4130         int rctx;
4131
4132         rctx = perf_swevent_get_recursion_context();
4133         if (rctx < 0)
4134                 return;
4135
4136         data.addr = addr;
4137         data.raw  = NULL;
4138
4139         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4140
4141         perf_swevent_put_recursion_context(rctx);
4142 }
4143
4144 static void perf_swevent_read(struct perf_event *event)
4145 {
4146 }
4147
4148 static int perf_swevent_enable(struct perf_event *event)
4149 {
4150         struct hw_perf_event *hwc = &event->hw;
4151
4152         if (hwc->sample_period) {
4153                 hwc->last_period = hwc->sample_period;
4154                 perf_swevent_set_period(event);
4155         }
4156         return 0;
4157 }
4158
4159 static void perf_swevent_disable(struct perf_event *event)
4160 {
4161 }
4162
4163 static const struct pmu perf_ops_generic = {
4164         .enable         = perf_swevent_enable,
4165         .disable        = perf_swevent_disable,
4166         .read           = perf_swevent_read,
4167         .unthrottle     = perf_swevent_unthrottle,
4168 };
4169
4170 /*
4171  * hrtimer based swevent callback
4172  */
4173
4174 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4175 {
4176         enum hrtimer_restart ret = HRTIMER_RESTART;
4177         struct perf_sample_data data;
4178         struct pt_regs *regs;
4179         struct perf_event *event;
4180         u64 period;
4181
4182         event   = container_of(hrtimer, struct perf_event, hw.hrtimer);
4183         event->pmu->read(event);
4184
4185         data.addr = 0;
4186         data.raw = NULL;
4187         data.period = event->hw.last_period;
4188         regs = get_irq_regs();
4189         /*
4190          * In case we exclude kernel IPs or are somehow not in interrupt
4191          * context, provide the next best thing, the user IP.
4192          */
4193         if ((event->attr.exclude_kernel || !regs) &&
4194                         !event->attr.exclude_user)
4195                 regs = task_pt_regs(current);
4196
4197         if (regs) {
4198                 if (!(event->attr.exclude_idle && current->pid == 0))
4199                         if (perf_event_overflow(event, 0, &data, regs))
4200                                 ret = HRTIMER_NORESTART;
4201         }
4202
4203         period = max_t(u64, 10000, event->hw.sample_period);
4204         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4205
4206         return ret;
4207 }
4208
4209 static void perf_swevent_start_hrtimer(struct perf_event *event)
4210 {
4211         struct hw_perf_event *hwc = &event->hw;
4212
4213         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4214         hwc->hrtimer.function = perf_swevent_hrtimer;
4215         if (hwc->sample_period) {
4216                 u64 period;
4217
4218                 if (hwc->remaining) {
4219                         if (hwc->remaining < 0)
4220                                 period = 10000;
4221                         else
4222                                 period = hwc->remaining;
4223                         hwc->remaining = 0;
4224                 } else {
4225                         period = max_t(u64, 10000, hwc->sample_period);
4226                 }
4227                 __hrtimer_start_range_ns(&hwc->hrtimer,
4228                                 ns_to_ktime(period), 0,
4229                                 HRTIMER_MODE_REL, 0);
4230         }
4231 }
4232
4233 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4234 {
4235         struct hw_perf_event *hwc = &event->hw;
4236
4237         if (hwc->sample_period) {
4238                 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4239                 hwc->remaining = ktime_to_ns(remaining);
4240
4241                 hrtimer_cancel(&hwc->hrtimer);
4242         }
4243 }
4244
4245 /*
4246  * Software event: cpu wall time clock
4247  */
4248
4249 static void cpu_clock_perf_event_update(struct perf_event *event)
4250 {
4251         int cpu = raw_smp_processor_id();
4252         s64 prev;
4253         u64 now;
4254
4255         now = cpu_clock(cpu);
4256         prev = atomic64_xchg(&event->hw.prev_count, now);
4257         atomic64_add(now - prev, &event->count);
4258 }
4259
4260 static int cpu_clock_perf_event_enable(struct perf_event *event)
4261 {
4262         struct hw_perf_event *hwc = &event->hw;
4263         int cpu = raw_smp_processor_id();
4264
4265         atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4266         perf_swevent_start_hrtimer(event);
4267
4268         return 0;
4269 }
4270
4271 static void cpu_clock_perf_event_disable(struct perf_event *event)
4272 {
4273         perf_swevent_cancel_hrtimer(event);
4274         cpu_clock_perf_event_update(event);
4275 }
4276
4277 static void cpu_clock_perf_event_read(struct perf_event *event)
4278 {
4279         cpu_clock_perf_event_update(event);
4280 }
4281
4282 static const struct pmu perf_ops_cpu_clock = {
4283         .enable         = cpu_clock_perf_event_enable,
4284         .disable        = cpu_clock_perf_event_disable,
4285         .read           = cpu_clock_perf_event_read,
4286 };
4287
4288 /*
4289  * Software event: task time clock
4290  */
4291
4292 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4293 {
4294         u64 prev;
4295         s64 delta;
4296
4297         prev = atomic64_xchg(&event->hw.prev_count, now);
4298         delta = now - prev;
4299         atomic64_add(delta, &event->count);
4300 }
4301
4302 static int task_clock_perf_event_enable(struct perf_event *event)
4303 {
4304         struct hw_perf_event *hwc = &event->hw;
4305         u64 now;
4306
4307         now = event->ctx->time;
4308
4309         atomic64_set(&hwc->prev_count, now);
4310
4311         perf_swevent_start_hrtimer(event);
4312
4313         return 0;
4314 }
4315
4316 static void task_clock_perf_event_disable(struct perf_event *event)
4317 {
4318         perf_swevent_cancel_hrtimer(event);
4319         task_clock_perf_event_update(event, event->ctx->time);
4320
4321 }
4322
4323 static void task_clock_perf_event_read(struct perf_event *event)
4324 {
4325         u64 time;
4326
4327         if (!in_nmi()) {
4328                 update_context_time(event->ctx);
4329                 time = event->ctx->time;
4330         } else {
4331                 u64 now = perf_clock();
4332                 u64 delta = now - event->ctx->timestamp;
4333                 time = event->ctx->time + delta;
4334         }
4335
4336         task_clock_perf_event_update(event, time);
4337 }
4338
4339 static const struct pmu perf_ops_task_clock = {
4340         .enable         = task_clock_perf_event_enable,
4341         .disable        = task_clock_perf_event_disable,
4342         .read           = task_clock_perf_event_read,
4343 };
4344
4345 #ifdef CONFIG_EVENT_TRACING
4346
4347 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4348                           int entry_size)
4349 {
4350         struct perf_raw_record raw = {
4351                 .size = entry_size,
4352                 .data = record,
4353         };
4354
4355         struct perf_sample_data data = {
4356                 .addr = addr,
4357                 .raw = &raw,
4358         };
4359
4360         struct pt_regs *regs = get_irq_regs();
4361
4362         if (!regs)
4363                 regs = task_pt_regs(current);
4364
4365         /* Trace events already protected against recursion */
4366         do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4367                                 &data, regs);
4368 }
4369 EXPORT_SYMBOL_GPL(perf_tp_event);
4370
4371 static int perf_tp_event_match(struct perf_event *event,
4372                                 struct perf_sample_data *data)
4373 {
4374         void *record = data->raw->data;
4375
4376         if (likely(!event->filter) || filter_match_preds(event->filter, record))
4377                 return 1;
4378         return 0;
4379 }
4380
4381 static void tp_perf_event_destroy(struct perf_event *event)
4382 {
4383         ftrace_profile_disable(event->attr.config);
4384 }
4385
4386 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4387 {
4388         /*
4389          * Raw tracepoint data is a severe data leak, only allow root to
4390          * have these.
4391          */
4392         if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4393                         perf_paranoid_tracepoint_raw() &&
4394                         !capable(CAP_SYS_ADMIN))
4395                 return ERR_PTR(-EPERM);
4396
4397         if (ftrace_profile_enable(event->attr.config))
4398                 return NULL;
4399
4400         event->destroy = tp_perf_event_destroy;
4401
4402         return &perf_ops_generic;
4403 }
4404
4405 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4406 {
4407         char *filter_str;
4408         int ret;
4409
4410         if (event->attr.type != PERF_TYPE_TRACEPOINT)
4411                 return -EINVAL;
4412
4413         filter_str = strndup_user(arg, PAGE_SIZE);
4414         if (IS_ERR(filter_str))
4415                 return PTR_ERR(filter_str);
4416
4417         ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4418
4419         kfree(filter_str);
4420         return ret;
4421 }
4422
4423 static void perf_event_free_filter(struct perf_event *event)
4424 {
4425         ftrace_profile_free_filter(event);
4426 }
4427
4428 #else
4429
4430 static int perf_tp_event_match(struct perf_event *event,
4431                                 struct perf_sample_data *data)
4432 {
4433         return 1;
4434 }
4435
4436 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4437 {
4438         return NULL;
4439 }
4440
4441 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4442 {
4443         return -ENOENT;
4444 }
4445
4446 static void perf_event_free_filter(struct perf_event *event)
4447 {
4448 }
4449
4450 #endif /* CONFIG_EVENT_TRACING */
4451
4452 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4453 static void bp_perf_event_destroy(struct perf_event *event)
4454 {
4455         release_bp_slot(event);
4456 }
4457
4458 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4459 {
4460         int err;
4461
4462         err = register_perf_hw_breakpoint(bp);
4463         if (err)
4464                 return ERR_PTR(err);
4465
4466         bp->destroy = bp_perf_event_destroy;
4467
4468         return &perf_ops_bp;
4469 }
4470
4471 void perf_bp_event(struct perf_event *bp, void *data)
4472 {
4473         struct perf_sample_data sample;
4474         struct pt_regs *regs = data;
4475
4476         sample.raw = NULL;
4477         sample.addr = bp->attr.bp_addr;
4478
4479         if (!perf_exclude_event(bp, regs))
4480                 perf_swevent_add(bp, 1, 1, &sample, regs);
4481 }
4482 #else
4483 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4484 {
4485         return NULL;
4486 }
4487
4488 void perf_bp_event(struct perf_event *bp, void *regs)
4489 {
4490 }
4491 #endif
4492
4493 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4494
4495 static void sw_perf_event_destroy(struct perf_event *event)
4496 {
4497         u64 event_id = event->attr.config;
4498
4499         WARN_ON(event->parent);
4500
4501         atomic_dec(&perf_swevent_enabled[event_id]);
4502 }
4503
4504 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4505 {
4506         const struct pmu *pmu = NULL;
4507         u64 event_id = event->attr.config;
4508
4509         /*
4510          * Software events (currently) can't in general distinguish
4511          * between user, kernel and hypervisor events.
4512          * However, context switches and cpu migrations are considered
4513          * to be kernel events, and page faults are never hypervisor
4514          * events.
4515          */
4516         switch (event_id) {
4517         case PERF_COUNT_SW_CPU_CLOCK:
4518                 pmu = &perf_ops_cpu_clock;
4519
4520                 break;
4521         case PERF_COUNT_SW_TASK_CLOCK:
4522                 /*
4523                  * If the user instantiates this as a per-cpu event,
4524                  * use the cpu_clock event instead.
4525                  */
4526                 if (event->ctx->task)
4527                         pmu = &perf_ops_task_clock;
4528                 else
4529                         pmu = &perf_ops_cpu_clock;
4530
4531                 break;
4532         case PERF_COUNT_SW_PAGE_FAULTS:
4533         case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4534         case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4535         case PERF_COUNT_SW_CONTEXT_SWITCHES:
4536         case PERF_COUNT_SW_CPU_MIGRATIONS:
4537         case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4538         case PERF_COUNT_SW_EMULATION_FAULTS:
4539                 if (!event->parent) {
4540                         atomic_inc(&perf_swevent_enabled[event_id]);
4541                         event->destroy = sw_perf_event_destroy;
4542                 }
4543                 pmu = &perf_ops_generic;
4544                 break;
4545         }
4546
4547         return pmu;
4548 }
4549
4550 /*
4551  * Allocate and initialize a event structure
4552  */
4553 static struct perf_event *
4554 perf_event_alloc(struct perf_event_attr *attr,
4555                    int cpu,
4556                    struct perf_event_context *ctx,
4557                    struct perf_event *group_leader,
4558                    struct perf_event *parent_event,
4559                    perf_overflow_handler_t overflow_handler,
4560                    gfp_t gfpflags)
4561 {
4562         const struct pmu *pmu;
4563         struct perf_event *event;
4564         struct hw_perf_event *hwc;
4565         long err;
4566
4567         event = kzalloc(sizeof(*event), gfpflags);
4568         if (!event)
4569                 return ERR_PTR(-ENOMEM);
4570
4571         /*
4572          * Single events are their own group leaders, with an
4573          * empty sibling list:
4574          */
4575         if (!group_leader)
4576                 group_leader = event;
4577
4578         mutex_init(&event->child_mutex);
4579         INIT_LIST_HEAD(&event->child_list);
4580
4581         INIT_LIST_HEAD(&event->group_entry);
4582         INIT_LIST_HEAD(&event->event_entry);
4583         INIT_LIST_HEAD(&event->sibling_list);
4584         init_waitqueue_head(&event->waitq);
4585
4586         mutex_init(&event->mmap_mutex);
4587
4588         event->cpu              = cpu;
4589         event->attr             = *attr;
4590         event->group_leader     = group_leader;
4591         event->pmu              = NULL;
4592         event->ctx              = ctx;
4593         event->oncpu            = -1;
4594
4595         event->parent           = parent_event;
4596
4597         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
4598         event->id               = atomic64_inc_return(&perf_event_id);
4599
4600         event->state            = PERF_EVENT_STATE_INACTIVE;
4601
4602         if (!overflow_handler && parent_event)
4603                 overflow_handler = parent_event->overflow_handler;
4604         
4605         event->overflow_handler = overflow_handler;
4606
4607         if (attr->disabled)
4608                 event->state = PERF_EVENT_STATE_OFF;
4609
4610         pmu = NULL;
4611
4612         hwc = &event->hw;
4613         hwc->sample_period = attr->sample_period;
4614         if (attr->freq && attr->sample_freq)
4615                 hwc->sample_period = 1;
4616         hwc->last_period = hwc->sample_period;
4617
4618         atomic64_set(&hwc->period_left, hwc->sample_period);
4619
4620         /*
4621          * we currently do not support PERF_FORMAT_GROUP on inherited events
4622          */
4623         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4624                 goto done;
4625
4626         switch (attr->type) {
4627         case PERF_TYPE_RAW:
4628         case PERF_TYPE_HARDWARE:
4629         case PERF_TYPE_HW_CACHE:
4630                 pmu = hw_perf_event_init(event);
4631                 break;
4632
4633         case PERF_TYPE_SOFTWARE:
4634                 pmu = sw_perf_event_init(event);
4635                 break;
4636
4637         case PERF_TYPE_TRACEPOINT:
4638                 pmu = tp_perf_event_init(event);
4639                 break;
4640
4641         case PERF_TYPE_BREAKPOINT:
4642                 pmu = bp_perf_event_init(event);
4643                 break;
4644
4645
4646         default:
4647                 break;
4648         }
4649 done:
4650         err = 0;
4651         if (!pmu)
4652                 err = -EINVAL;
4653         else if (IS_ERR(pmu))
4654                 err = PTR_ERR(pmu);
4655
4656         if (err) {
4657                 if (event->ns)
4658                         put_pid_ns(event->ns);
4659                 kfree(event);
4660                 return ERR_PTR(err);
4661         }
4662
4663         event->pmu = pmu;
4664
4665         if (!event->parent) {
4666                 atomic_inc(&nr_events);
4667                 if (event->attr.mmap)
4668                         atomic_inc(&nr_mmap_events);
4669                 if (event->attr.comm)
4670                         atomic_inc(&nr_comm_events);
4671                 if (event->attr.task)
4672                         atomic_inc(&nr_task_events);
4673         }
4674
4675         return event;
4676 }
4677
4678 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4679                           struct perf_event_attr *attr)
4680 {
4681         u32 size;
4682         int ret;
4683
4684         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4685                 return -EFAULT;
4686
4687         /*
4688          * zero the full structure, so that a short copy will be nice.
4689          */
4690         memset(attr, 0, sizeof(*attr));
4691
4692         ret = get_user(size, &uattr->size);
4693         if (ret)
4694                 return ret;
4695
4696         if (size > PAGE_SIZE)   /* silly large */
4697                 goto err_size;
4698
4699         if (!size)              /* abi compat */
4700                 size = PERF_ATTR_SIZE_VER0;
4701
4702         if (size < PERF_ATTR_SIZE_VER0)
4703                 goto err_size;
4704
4705         /*
4706          * If we're handed a bigger struct than we know of,
4707          * ensure all the unknown bits are 0 - i.e. new
4708          * user-space does not rely on any kernel feature
4709          * extensions we dont know about yet.
4710          */
4711         if (size > sizeof(*attr)) {
4712                 unsigned char __user *addr;
4713                 unsigned char __user *end;
4714                 unsigned char val;
4715
4716                 addr = (void __user *)uattr + sizeof(*attr);
4717                 end  = (void __user *)uattr + size;
4718
4719                 for (; addr < end; addr++) {
4720                         ret = get_user(val, addr);
4721                         if (ret)
4722                                 return ret;
4723                         if (val)
4724                                 goto err_size;
4725                 }
4726                 size = sizeof(*attr);
4727         }
4728
4729         ret = copy_from_user(attr, uattr, size);
4730         if (ret)
4731                 return -EFAULT;
4732
4733         /*
4734          * If the type exists, the corresponding creation will verify
4735          * the attr->config.
4736          */
4737         if (attr->type >= PERF_TYPE_MAX)
4738                 return -EINVAL;
4739
4740         if (attr->__reserved_1 || attr->__reserved_2)
4741                 return -EINVAL;
4742
4743         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4744                 return -EINVAL;
4745
4746         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4747                 return -EINVAL;
4748
4749 out:
4750         return ret;
4751
4752 err_size:
4753         put_user(sizeof(*attr), &uattr->size);
4754         ret = -E2BIG;
4755         goto out;
4756 }
4757
4758 static int perf_event_set_output(struct perf_event *event, int output_fd)
4759 {
4760         struct perf_event *output_event = NULL;
4761         struct file *output_file = NULL;
4762         struct perf_event *old_output;
4763         int fput_needed = 0;
4764         int ret = -EINVAL;
4765
4766         if (!output_fd)
4767                 goto set;
4768
4769         output_file = fget_light(output_fd, &fput_needed);
4770         if (!output_file)
4771                 return -EBADF;
4772
4773         if (output_file->f_op != &perf_fops)
4774                 goto out;
4775
4776         output_event = output_file->private_data;
4777
4778         /* Don't chain output fds */
4779         if (output_event->output)
4780                 goto out;
4781
4782         /* Don't set an output fd when we already have an output channel */
4783         if (event->data)
4784                 goto out;
4785
4786         atomic_long_inc(&output_file->f_count);
4787
4788 set:
4789         mutex_lock(&event->mmap_mutex);
4790         old_output = event->output;
4791         rcu_assign_pointer(event->output, output_event);
4792         mutex_unlock(&event->mmap_mutex);
4793
4794         if (old_output) {
4795                 /*
4796                  * we need to make sure no existing perf_output_*()
4797                  * is still referencing this event.
4798                  */
4799                 synchronize_rcu();
4800                 fput(old_output->filp);
4801         }
4802
4803         ret = 0;
4804 out:
4805         fput_light(output_file, fput_needed);
4806         return ret;
4807 }
4808
4809 /**
4810  * sys_perf_event_open - open a performance event, associate it to a task/cpu
4811  *
4812  * @attr_uptr:  event_id type attributes for monitoring/sampling
4813  * @pid:                target pid
4814  * @cpu:                target cpu
4815  * @group_fd:           group leader event fd
4816  */
4817 SYSCALL_DEFINE5(perf_event_open,
4818                 struct perf_event_attr __user *, attr_uptr,
4819                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4820 {
4821         struct perf_event *event, *group_leader;
4822         struct perf_event_attr attr;
4823         struct perf_event_context *ctx;
4824         struct file *event_file = NULL;
4825         struct file *group_file = NULL;
4826         int fput_needed = 0;
4827         int fput_needed2 = 0;
4828         int err;
4829
4830         /* for future expandability... */
4831         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4832                 return -EINVAL;
4833
4834         err = perf_copy_attr(attr_uptr, &attr);
4835         if (err)
4836                 return err;
4837
4838         if (!attr.exclude_kernel) {
4839                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4840                         return -EACCES;
4841         }
4842
4843         if (attr.freq) {
4844                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4845                         return -EINVAL;
4846         }
4847
4848         /*
4849          * Get the target context (task or percpu):
4850          */
4851         ctx = find_get_context(pid, cpu);
4852         if (IS_ERR(ctx))
4853                 return PTR_ERR(ctx);
4854
4855         /*
4856          * Look up the group leader (we will attach this event to it):
4857          */
4858         group_leader = NULL;
4859         if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4860                 err = -EINVAL;
4861                 group_file = fget_light(group_fd, &fput_needed);
4862                 if (!group_file)
4863                         goto err_put_context;
4864                 if (group_file->f_op != &perf_fops)
4865                         goto err_put_context;
4866
4867                 group_leader = group_file->private_data;
4868                 /*
4869                  * Do not allow a recursive hierarchy (this new sibling
4870                  * becoming part of another group-sibling):
4871                  */
4872                 if (group_leader->group_leader != group_leader)
4873                         goto err_put_context;
4874                 /*
4875                  * Do not allow to attach to a group in a different
4876                  * task or CPU context:
4877                  */
4878                 if (group_leader->ctx != ctx)
4879                         goto err_put_context;
4880                 /*
4881                  * Only a group leader can be exclusive or pinned
4882                  */
4883                 if (attr.exclusive || attr.pinned)
4884                         goto err_put_context;
4885         }
4886
4887         event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4888                                      NULL, NULL, GFP_KERNEL);
4889         err = PTR_ERR(event);
4890         if (IS_ERR(event))
4891                 goto err_put_context;
4892
4893         err = anon_inode_getfd("[perf_event]", &perf_fops, event, O_RDWR);
4894         if (err < 0)
4895                 goto err_free_put_context;
4896
4897         event_file = fget_light(err, &fput_needed2);
4898         if (!event_file)
4899                 goto err_free_put_context;
4900
4901         if (flags & PERF_FLAG_FD_OUTPUT) {
4902                 err = perf_event_set_output(event, group_fd);
4903                 if (err)
4904                         goto err_fput_free_put_context;
4905         }
4906
4907         event->filp = event_file;
4908         WARN_ON_ONCE(ctx->parent_ctx);
4909         mutex_lock(&ctx->mutex);
4910         perf_install_in_context(ctx, event, cpu);
4911         ++ctx->generation;
4912         mutex_unlock(&ctx->mutex);
4913
4914         event->owner = current;
4915         get_task_struct(current);
4916         mutex_lock(&current->perf_event_mutex);
4917         list_add_tail(&event->owner_entry, &current->perf_event_list);
4918         mutex_unlock(&current->perf_event_mutex);
4919
4920 err_fput_free_put_context:
4921         fput_light(event_file, fput_needed2);
4922
4923 err_free_put_context:
4924         if (err < 0)
4925                 kfree(event);
4926
4927 err_put_context:
4928         if (err < 0)
4929                 put_ctx(ctx);
4930
4931         fput_light(group_file, fput_needed);
4932
4933         return err;
4934 }
4935
4936 /**
4937  * perf_event_create_kernel_counter
4938  *
4939  * @attr: attributes of the counter to create
4940  * @cpu: cpu in which the counter is bound
4941  * @pid: task to profile
4942  */
4943 struct perf_event *
4944 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4945                                  pid_t pid,
4946                                  perf_overflow_handler_t overflow_handler)
4947 {
4948         struct perf_event *event;
4949         struct perf_event_context *ctx;
4950         int err;
4951
4952         /*
4953          * Get the target context (task or percpu):
4954          */
4955
4956         ctx = find_get_context(pid, cpu);
4957         if (IS_ERR(ctx)) {
4958                 err = PTR_ERR(ctx);
4959                 goto err_exit;
4960         }
4961
4962         event = perf_event_alloc(attr, cpu, ctx, NULL,
4963                                  NULL, overflow_handler, GFP_KERNEL);
4964         if (IS_ERR(event)) {
4965                 err = PTR_ERR(event);
4966                 goto err_put_context;
4967         }
4968
4969         event->filp = NULL;
4970         WARN_ON_ONCE(ctx->parent_ctx);
4971         mutex_lock(&ctx->mutex);
4972         perf_install_in_context(ctx, event, cpu);
4973         ++ctx->generation;
4974         mutex_unlock(&ctx->mutex);
4975
4976         event->owner = current;
4977         get_task_struct(current);
4978         mutex_lock(&current->perf_event_mutex);
4979         list_add_tail(&event->owner_entry, &current->perf_event_list);
4980         mutex_unlock(&current->perf_event_mutex);
4981
4982         return event;
4983
4984  err_put_context:
4985         put_ctx(ctx);
4986  err_exit:
4987         return ERR_PTR(err);
4988 }
4989 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4990
4991 /*
4992  * inherit a event from parent task to child task:
4993  */
4994 static struct perf_event *
4995 inherit_event(struct perf_event *parent_event,
4996               struct task_struct *parent,
4997               struct perf_event_context *parent_ctx,
4998               struct task_struct *child,
4999               struct perf_event *group_leader,
5000               struct perf_event_context *child_ctx)
5001 {
5002         struct perf_event *child_event;
5003
5004         /*
5005          * Instead of creating recursive hierarchies of events,
5006          * we link inherited events back to the original parent,
5007          * which has a filp for sure, which we use as the reference
5008          * count:
5009          */
5010         if (parent_event->parent)
5011                 parent_event = parent_event->parent;
5012
5013         child_event = perf_event_alloc(&parent_event->attr,
5014                                            parent_event->cpu, child_ctx,
5015                                            group_leader, parent_event,
5016                                            NULL, GFP_KERNEL);
5017         if (IS_ERR(child_event))
5018                 return child_event;
5019         get_ctx(child_ctx);
5020
5021         /*
5022          * Make the child state follow the state of the parent event,
5023          * not its attr.disabled bit.  We hold the parent's mutex,
5024          * so we won't race with perf_event_{en, dis}able_family.
5025          */
5026         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
5027                 child_event->state = PERF_EVENT_STATE_INACTIVE;
5028         else
5029                 child_event->state = PERF_EVENT_STATE_OFF;
5030
5031         if (parent_event->attr.freq) {
5032                 u64 sample_period = parent_event->hw.sample_period;
5033                 struct hw_perf_event *hwc = &child_event->hw;
5034
5035                 hwc->sample_period = sample_period;
5036                 hwc->last_period   = sample_period;
5037
5038                 atomic64_set(&hwc->period_left, sample_period);
5039         }
5040
5041         child_event->overflow_handler = parent_event->overflow_handler;
5042
5043         /*
5044          * Link it up in the child's context:
5045          */
5046         add_event_to_ctx(child_event, child_ctx);
5047
5048         /*
5049          * Get a reference to the parent filp - we will fput it
5050          * when the child event exits. This is safe to do because
5051          * we are in the parent and we know that the filp still
5052          * exists and has a nonzero count:
5053          */
5054         atomic_long_inc(&parent_event->filp->f_count);
5055
5056         /*
5057          * Link this into the parent event's child list
5058          */
5059         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5060         mutex_lock(&parent_event->child_mutex);
5061         list_add_tail(&child_event->child_list, &parent_event->child_list);
5062         mutex_unlock(&parent_event->child_mutex);
5063
5064         return child_event;
5065 }
5066
5067 static int inherit_group(struct perf_event *parent_event,
5068               struct task_struct *parent,
5069               struct perf_event_context *parent_ctx,
5070               struct task_struct *child,
5071               struct perf_event_context *child_ctx)
5072 {
5073         struct perf_event *leader;
5074         struct perf_event *sub;
5075         struct perf_event *child_ctr;
5076
5077         leader = inherit_event(parent_event, parent, parent_ctx,
5078                                  child, NULL, child_ctx);
5079         if (IS_ERR(leader))
5080                 return PTR_ERR(leader);
5081         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5082                 child_ctr = inherit_event(sub, parent, parent_ctx,
5083                                             child, leader, child_ctx);
5084                 if (IS_ERR(child_ctr))
5085                         return PTR_ERR(child_ctr);
5086         }
5087         return 0;
5088 }
5089
5090 static void sync_child_event(struct perf_event *child_event,
5091                                struct task_struct *child)
5092 {
5093         struct perf_event *parent_event = child_event->parent;
5094         u64 child_val;
5095
5096         if (child_event->attr.inherit_stat)
5097                 perf_event_read_event(child_event, child);
5098
5099         child_val = atomic64_read(&child_event->count);
5100
5101         /*
5102          * Add back the child's count to the parent's count:
5103          */
5104         atomic64_add(child_val, &parent_event->count);
5105         atomic64_add(child_event->total_time_enabled,
5106                      &parent_event->child_total_time_enabled);
5107         atomic64_add(child_event->total_time_running,
5108                      &parent_event->child_total_time_running);
5109
5110         /*
5111          * Remove this event from the parent's list
5112          */
5113         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5114         mutex_lock(&parent_event->child_mutex);
5115         list_del_init(&child_event->child_list);
5116         mutex_unlock(&parent_event->child_mutex);
5117
5118         /*
5119          * Release the parent event, if this was the last
5120          * reference to it.
5121          */
5122         fput(parent_event->filp);
5123 }
5124
5125 static void
5126 __perf_event_exit_task(struct perf_event *child_event,
5127                          struct perf_event_context *child_ctx,
5128                          struct task_struct *child)
5129 {
5130         struct perf_event *parent_event;
5131
5132         perf_event_remove_from_context(child_event);
5133
5134         parent_event = child_event->parent;
5135         /*
5136          * It can happen that parent exits first, and has events
5137          * that are still around due to the child reference. These
5138          * events need to be zapped - but otherwise linger.
5139          */
5140         if (parent_event) {
5141                 sync_child_event(child_event, child);
5142                 free_event(child_event);
5143         }
5144 }
5145
5146 /*
5147  * When a child task exits, feed back event values to parent events.
5148  */
5149 void perf_event_exit_task(struct task_struct *child)
5150 {
5151         struct perf_event *child_event, *tmp;
5152         struct perf_event_context *child_ctx;
5153         unsigned long flags;
5154
5155         if (likely(!child->perf_event_ctxp)) {
5156                 perf_event_task(child, NULL, 0);
5157                 return;
5158         }
5159
5160         local_irq_save(flags);
5161         /*
5162          * We can't reschedule here because interrupts are disabled,
5163          * and either child is current or it is a task that can't be
5164          * scheduled, so we are now safe from rescheduling changing
5165          * our context.
5166          */
5167         child_ctx = child->perf_event_ctxp;
5168         __perf_event_task_sched_out(child_ctx);
5169
5170         /*
5171          * Take the context lock here so that if find_get_context is
5172          * reading child->perf_event_ctxp, we wait until it has
5173          * incremented the context's refcount before we do put_ctx below.
5174          */
5175         raw_spin_lock(&child_ctx->lock);
5176         child->perf_event_ctxp = NULL;
5177         /*
5178          * If this context is a clone; unclone it so it can't get
5179          * swapped to another process while we're removing all
5180          * the events from it.
5181          */
5182         unclone_ctx(child_ctx);
5183         update_context_time(child_ctx);
5184         raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5185
5186         /*
5187          * Report the task dead after unscheduling the events so that we
5188          * won't get any samples after PERF_RECORD_EXIT. We can however still
5189          * get a few PERF_RECORD_READ events.
5190          */
5191         perf_event_task(child, child_ctx, 0);
5192
5193         /*
5194          * We can recurse on the same lock type through:
5195          *
5196          *   __perf_event_exit_task()
5197          *     sync_child_event()
5198          *       fput(parent_event->filp)
5199          *         perf_release()
5200          *           mutex_lock(&ctx->mutex)
5201          *
5202          * But since its the parent context it won't be the same instance.
5203          */
5204         mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
5205
5206 again:
5207         list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5208                                  group_entry)
5209                 __perf_event_exit_task(child_event, child_ctx, child);
5210
5211         list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5212                                  group_entry)
5213                 __perf_event_exit_task(child_event, child_ctx, child);
5214
5215         /*
5216          * If the last event was a group event, it will have appended all
5217          * its siblings to the list, but we obtained 'tmp' before that which
5218          * will still point to the list head terminating the iteration.
5219          */
5220         if (!list_empty(&child_ctx->pinned_groups) ||
5221             !list_empty(&child_ctx->flexible_groups))
5222                 goto again;
5223
5224         mutex_unlock(&child_ctx->mutex);
5225
5226         put_ctx(child_ctx);
5227 }
5228
5229 static void perf_free_event(struct perf_event *event,
5230                             struct perf_event_context *ctx)
5231 {
5232         struct perf_event *parent = event->parent;
5233
5234         if (WARN_ON_ONCE(!parent))
5235                 return;
5236
5237         mutex_lock(&parent->child_mutex);
5238         list_del_init(&event->child_list);
5239         mutex_unlock(&parent->child_mutex);
5240
5241         fput(parent->filp);
5242
5243         list_del_event(event, ctx);
5244         free_event(event);
5245 }
5246
5247 /*
5248  * free an unexposed, unused context as created by inheritance by
5249  * init_task below, used by fork() in case of fail.
5250  */
5251 void perf_event_free_task(struct task_struct *task)
5252 {
5253         struct perf_event_context *ctx = task->perf_event_ctxp;
5254         struct perf_event *event, *tmp;
5255
5256         if (!ctx)
5257                 return;
5258
5259         mutex_lock(&ctx->mutex);
5260 again:
5261         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5262                 perf_free_event(event, ctx);
5263
5264         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5265                                  group_entry)
5266                 perf_free_event(event, ctx);
5267
5268         if (!list_empty(&ctx->pinned_groups) ||
5269             !list_empty(&ctx->flexible_groups))
5270                 goto again;
5271
5272         mutex_unlock(&ctx->mutex);
5273
5274         put_ctx(ctx);
5275 }
5276
5277 static int
5278 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5279                    struct perf_event_context *parent_ctx,
5280                    struct task_struct *child,
5281                    int *inherited_all)
5282 {
5283         int ret;
5284         struct perf_event_context *child_ctx = child->perf_event_ctxp;
5285
5286         if (!event->attr.inherit) {
5287                 *inherited_all = 0;
5288                 return 0;
5289         }
5290
5291         if (!child_ctx) {
5292                 /*
5293                  * This is executed from the parent task context, so
5294                  * inherit events that have been marked for cloning.
5295                  * First allocate and initialize a context for the
5296                  * child.
5297                  */
5298
5299                 child_ctx = kzalloc(sizeof(struct perf_event_context),
5300                                     GFP_KERNEL);
5301                 if (!child_ctx)
5302                         return -ENOMEM;
5303
5304                 __perf_event_init_context(child_ctx, child);
5305                 child->perf_event_ctxp = child_ctx;
5306                 get_task_struct(child);
5307         }
5308
5309         ret = inherit_group(event, parent, parent_ctx,
5310                             child, child_ctx);
5311
5312         if (ret)
5313                 *inherited_all = 0;
5314
5315         return ret;
5316 }
5317
5318
5319 /*
5320  * Initialize the perf_event context in task_struct
5321  */
5322 int perf_event_init_task(struct task_struct *child)
5323 {
5324         struct perf_event_context *child_ctx, *parent_ctx;
5325         struct perf_event_context *cloned_ctx;
5326         struct perf_event *event;
5327         struct task_struct *parent = current;
5328         int inherited_all = 1;
5329         int ret = 0;
5330
5331         child->perf_event_ctxp = NULL;
5332
5333         mutex_init(&child->perf_event_mutex);
5334         INIT_LIST_HEAD(&child->perf_event_list);
5335
5336         if (likely(!parent->perf_event_ctxp))
5337                 return 0;
5338
5339         /*
5340          * If the parent's context is a clone, pin it so it won't get
5341          * swapped under us.
5342          */
5343         parent_ctx = perf_pin_task_context(parent);
5344
5345         /*
5346          * No need to check if parent_ctx != NULL here; since we saw
5347          * it non-NULL earlier, the only reason for it to become NULL
5348          * is if we exit, and since we're currently in the middle of
5349          * a fork we can't be exiting at the same time.
5350          */
5351
5352         /*
5353          * Lock the parent list. No need to lock the child - not PID
5354          * hashed yet and not running, so nobody can access it.
5355          */
5356         mutex_lock(&parent_ctx->mutex);
5357
5358         /*
5359          * We dont have to disable NMIs - we are only looking at
5360          * the list, not manipulating it:
5361          */
5362         list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5363                 ret = inherit_task_group(event, parent, parent_ctx, child,
5364                                          &inherited_all);
5365                 if (ret)
5366                         break;
5367         }
5368
5369         list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5370                 ret = inherit_task_group(event, parent, parent_ctx, child,
5371                                          &inherited_all);
5372                 if (ret)
5373                         break;
5374         }
5375
5376         child_ctx = child->perf_event_ctxp;
5377
5378         if (child_ctx && inherited_all) {
5379                 /*
5380                  * Mark the child context as a clone of the parent
5381                  * context, or of whatever the parent is a clone of.
5382                  * Note that if the parent is a clone, it could get
5383                  * uncloned at any point, but that doesn't matter
5384                  * because the list of events and the generation
5385                  * count can't have changed since we took the mutex.
5386                  */
5387                 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5388                 if (cloned_ctx) {
5389                         child_ctx->parent_ctx = cloned_ctx;
5390                         child_ctx->parent_gen = parent_ctx->parent_gen;
5391                 } else {
5392                         child_ctx->parent_ctx = parent_ctx;
5393                         child_ctx->parent_gen = parent_ctx->generation;
5394                 }
5395                 get_ctx(child_ctx->parent_ctx);
5396         }
5397
5398         mutex_unlock(&parent_ctx->mutex);
5399
5400         perf_unpin_context(parent_ctx);
5401
5402         return ret;
5403 }
5404
5405 static void __cpuinit perf_event_init_cpu(int cpu)
5406 {
5407         struct perf_cpu_context *cpuctx;
5408
5409         cpuctx = &per_cpu(perf_cpu_context, cpu);
5410         __perf_event_init_context(&cpuctx->ctx, NULL);
5411
5412         spin_lock(&perf_resource_lock);
5413         cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5414         spin_unlock(&perf_resource_lock);
5415
5416         hw_perf_event_setup(cpu);
5417 }
5418
5419 #ifdef CONFIG_HOTPLUG_CPU
5420 static void __perf_event_exit_cpu(void *info)
5421 {
5422         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5423         struct perf_event_context *ctx = &cpuctx->ctx;
5424         struct perf_event *event, *tmp;
5425
5426         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5427                 __perf_event_remove_from_context(event);
5428         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5429                 __perf_event_remove_from_context(event);
5430 }
5431 static void perf_event_exit_cpu(int cpu)
5432 {
5433         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5434         struct perf_event_context *ctx = &cpuctx->ctx;
5435
5436         mutex_lock(&ctx->mutex);
5437         smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5438         mutex_unlock(&ctx->mutex);
5439 }
5440 #else
5441 static inline void perf_event_exit_cpu(int cpu) { }
5442 #endif
5443
5444 static int __cpuinit
5445 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5446 {
5447         unsigned int cpu = (long)hcpu;
5448
5449         switch (action) {
5450
5451         case CPU_UP_PREPARE:
5452         case CPU_UP_PREPARE_FROZEN:
5453                 perf_event_init_cpu(cpu);
5454                 break;
5455
5456         case CPU_ONLINE:
5457         case CPU_ONLINE_FROZEN:
5458                 hw_perf_event_setup_online(cpu);
5459                 break;
5460
5461         case CPU_DOWN_PREPARE:
5462         case CPU_DOWN_PREPARE_FROZEN:
5463                 perf_event_exit_cpu(cpu);
5464                 break;
5465
5466         case CPU_DEAD:
5467                 hw_perf_event_setup_offline(cpu);
5468                 break;
5469
5470         default:
5471                 break;
5472         }
5473
5474         return NOTIFY_OK;
5475 }
5476
5477 /*
5478  * This has to have a higher priority than migration_notifier in sched.c.
5479  */
5480 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5481         .notifier_call          = perf_cpu_notify,
5482         .priority               = 20,
5483 };
5484
5485 void __init perf_event_init(void)
5486 {
5487         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5488                         (void *)(long)smp_processor_id());
5489         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5490                         (void *)(long)smp_processor_id());
5491         register_cpu_notifier(&perf_cpu_nb);
5492 }
5493
5494 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
5495 {
5496         return sprintf(buf, "%d\n", perf_reserved_percpu);
5497 }
5498
5499 static ssize_t
5500 perf_set_reserve_percpu(struct sysdev_class *class,
5501                         const char *buf,
5502                         size_t count)
5503 {
5504         struct perf_cpu_context *cpuctx;
5505         unsigned long val;
5506         int err, cpu, mpt;
5507
5508         err = strict_strtoul(buf, 10, &val);
5509         if (err)
5510                 return err;
5511         if (val > perf_max_events)
5512                 return -EINVAL;
5513
5514         spin_lock(&perf_resource_lock);
5515         perf_reserved_percpu = val;
5516         for_each_online_cpu(cpu) {
5517                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5518                 raw_spin_lock_irq(&cpuctx->ctx.lock);
5519                 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5520                           perf_max_events - perf_reserved_percpu);
5521                 cpuctx->max_pertask = mpt;
5522                 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5523         }
5524         spin_unlock(&perf_resource_lock);
5525
5526         return count;
5527 }
5528
5529 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
5530 {
5531         return sprintf(buf, "%d\n", perf_overcommit);
5532 }
5533
5534 static ssize_t
5535 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
5536 {
5537         unsigned long val;
5538         int err;
5539
5540         err = strict_strtoul(buf, 10, &val);
5541         if (err)
5542                 return err;
5543         if (val > 1)
5544                 return -EINVAL;
5545
5546         spin_lock(&perf_resource_lock);
5547         perf_overcommit = val;
5548         spin_unlock(&perf_resource_lock);
5549
5550         return count;
5551 }
5552
5553 static SYSDEV_CLASS_ATTR(
5554                                 reserve_percpu,
5555                                 0644,
5556                                 perf_show_reserve_percpu,
5557                                 perf_set_reserve_percpu
5558                         );
5559
5560 static SYSDEV_CLASS_ATTR(
5561                                 overcommit,
5562                                 0644,
5563                                 perf_show_overcommit,
5564                                 perf_set_overcommit
5565                         );
5566
5567 static struct attribute *perfclass_attrs[] = {
5568         &attr_reserve_percpu.attr,
5569         &attr_overcommit.attr,
5570         NULL
5571 };
5572
5573 static struct attribute_group perfclass_attr_group = {
5574         .attrs                  = perfclass_attrs,
5575         .name                   = "perf_events",
5576 };
5577
5578 static int __init perf_event_sysfs_init(void)
5579 {
5580         return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5581                                   &perfclass_attr_group);
5582 }
5583 device_initcall(perf_event_sysfs_init);