perf: Fix unexported generic perf_arch_fetch_caller_regs
[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 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
60
61 /*
62  * max perf event sample rate
63  */
64 int sysctl_perf_event_sample_rate __read_mostly = 100000;
65
66 static atomic64_t perf_event_id;
67
68 /*
69  * Lock for (sysadmin-configurable) event reservations:
70  */
71 static DEFINE_SPINLOCK(perf_resource_lock);
72
73 /*
74  * Architecture provided APIs - weak aliases:
75  */
76 extern __weak const struct pmu *hw_perf_event_init(struct perf_event *event)
77 {
78         return NULL;
79 }
80
81 void __weak hw_perf_disable(void)               { barrier(); }
82 void __weak hw_perf_enable(void)                { barrier(); }
83
84 int __weak
85 hw_perf_group_sched_in(struct perf_event *group_leader,
86                struct perf_cpu_context *cpuctx,
87                struct perf_event_context *ctx)
88 {
89         return 0;
90 }
91
92 void __weak perf_event_print_debug(void)        { }
93
94 static DEFINE_PER_CPU(int, perf_disable_count);
95
96 void perf_disable(void)
97 {
98         if (!__get_cpu_var(perf_disable_count)++)
99                 hw_perf_disable();
100 }
101
102 void perf_enable(void)
103 {
104         if (!--__get_cpu_var(perf_disable_count))
105                 hw_perf_enable();
106 }
107
108 static void get_ctx(struct perf_event_context *ctx)
109 {
110         WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
111 }
112
113 static void free_ctx(struct rcu_head *head)
114 {
115         struct perf_event_context *ctx;
116
117         ctx = container_of(head, struct perf_event_context, rcu_head);
118         kfree(ctx);
119 }
120
121 static void put_ctx(struct perf_event_context *ctx)
122 {
123         if (atomic_dec_and_test(&ctx->refcount)) {
124                 if (ctx->parent_ctx)
125                         put_ctx(ctx->parent_ctx);
126                 if (ctx->task)
127                         put_task_struct(ctx->task);
128                 call_rcu(&ctx->rcu_head, free_ctx);
129         }
130 }
131
132 static void unclone_ctx(struct perf_event_context *ctx)
133 {
134         if (ctx->parent_ctx) {
135                 put_ctx(ctx->parent_ctx);
136                 ctx->parent_ctx = NULL;
137         }
138 }
139
140 /*
141  * If we inherit events we want to return the parent event id
142  * to userspace.
143  */
144 static u64 primary_event_id(struct perf_event *event)
145 {
146         u64 id = event->id;
147
148         if (event->parent)
149                 id = event->parent->id;
150
151         return id;
152 }
153
154 /*
155  * Get the perf_event_context for a task and lock it.
156  * This has to cope with with the fact that until it is locked,
157  * the context could get moved to another task.
158  */
159 static struct perf_event_context *
160 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
161 {
162         struct perf_event_context *ctx;
163
164         rcu_read_lock();
165  retry:
166         ctx = rcu_dereference(task->perf_event_ctxp);
167         if (ctx) {
168                 /*
169                  * If this context is a clone of another, it might
170                  * get swapped for another underneath us by
171                  * perf_event_task_sched_out, though the
172                  * rcu_read_lock() protects us from any context
173                  * getting freed.  Lock the context and check if it
174                  * got swapped before we could get the lock, and retry
175                  * if so.  If we locked the right context, then it
176                  * can't get swapped on us any more.
177                  */
178                 raw_spin_lock_irqsave(&ctx->lock, *flags);
179                 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
180                         raw_spin_unlock_irqrestore(&ctx->lock, *flags);
181                         goto retry;
182                 }
183
184                 if (!atomic_inc_not_zero(&ctx->refcount)) {
185                         raw_spin_unlock_irqrestore(&ctx->lock, *flags);
186                         ctx = NULL;
187                 }
188         }
189         rcu_read_unlock();
190         return ctx;
191 }
192
193 /*
194  * Get the context for a task and increment its pin_count so it
195  * can't get swapped to another task.  This also increments its
196  * reference count so that the context can't get freed.
197  */
198 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
199 {
200         struct perf_event_context *ctx;
201         unsigned long flags;
202
203         ctx = perf_lock_task_context(task, &flags);
204         if (ctx) {
205                 ++ctx->pin_count;
206                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
207         }
208         return ctx;
209 }
210
211 static void perf_unpin_context(struct perf_event_context *ctx)
212 {
213         unsigned long flags;
214
215         raw_spin_lock_irqsave(&ctx->lock, flags);
216         --ctx->pin_count;
217         raw_spin_unlock_irqrestore(&ctx->lock, flags);
218         put_ctx(ctx);
219 }
220
221 static inline u64 perf_clock(void)
222 {
223         return cpu_clock(raw_smp_processor_id());
224 }
225
226 /*
227  * Update the record of the current time in a context.
228  */
229 static void update_context_time(struct perf_event_context *ctx)
230 {
231         u64 now = perf_clock();
232
233         ctx->time += now - ctx->timestamp;
234         ctx->timestamp = now;
235 }
236
237 /*
238  * Update the total_time_enabled and total_time_running fields for a event.
239  */
240 static void update_event_times(struct perf_event *event)
241 {
242         struct perf_event_context *ctx = event->ctx;
243         u64 run_end;
244
245         if (event->state < PERF_EVENT_STATE_INACTIVE ||
246             event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
247                 return;
248
249         if (ctx->is_active)
250                 run_end = ctx->time;
251         else
252                 run_end = event->tstamp_stopped;
253
254         event->total_time_enabled = run_end - event->tstamp_enabled;
255
256         if (event->state == PERF_EVENT_STATE_INACTIVE)
257                 run_end = event->tstamp_stopped;
258         else
259                 run_end = ctx->time;
260
261         event->total_time_running = run_end - event->tstamp_running;
262 }
263
264 static struct list_head *
265 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
266 {
267         if (event->attr.pinned)
268                 return &ctx->pinned_groups;
269         else
270                 return &ctx->flexible_groups;
271 }
272
273 /*
274  * Add a event from the lists for its context.
275  * Must be called with ctx->mutex and ctx->lock held.
276  */
277 static void
278 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
279 {
280         struct perf_event *group_leader = event->group_leader;
281
282         /*
283          * Depending on whether it is a standalone or sibling event,
284          * add it straight to the context's event list, or to the group
285          * leader's sibling list:
286          */
287         if (group_leader == event) {
288                 struct list_head *list;
289
290                 if (is_software_event(event))
291                         event->group_flags |= PERF_GROUP_SOFTWARE;
292
293                 list = ctx_group_list(event, ctx);
294                 list_add_tail(&event->group_entry, list);
295         } else {
296                 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
297                     !is_software_event(event))
298                         group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
299
300                 list_add_tail(&event->group_entry, &group_leader->sibling_list);
301                 group_leader->nr_siblings++;
302         }
303
304         list_add_rcu(&event->event_entry, &ctx->event_list);
305         ctx->nr_events++;
306         if (event->attr.inherit_stat)
307                 ctx->nr_stat++;
308 }
309
310 /*
311  * Remove a event from the lists for its context.
312  * Must be called with ctx->mutex and ctx->lock held.
313  */
314 static void
315 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
316 {
317         struct perf_event *sibling, *tmp;
318
319         if (list_empty(&event->group_entry))
320                 return;
321         ctx->nr_events--;
322         if (event->attr.inherit_stat)
323                 ctx->nr_stat--;
324
325         list_del_init(&event->group_entry);
326         list_del_rcu(&event->event_entry);
327
328         if (event->group_leader != event)
329                 event->group_leader->nr_siblings--;
330
331         update_event_times(event);
332
333         /*
334          * If event was in error state, then keep it
335          * that way, otherwise bogus counts will be
336          * returned on read(). The only way to get out
337          * of error state is by explicit re-enabling
338          * of the event
339          */
340         if (event->state > PERF_EVENT_STATE_OFF)
341                 event->state = PERF_EVENT_STATE_OFF;
342
343         /*
344          * If this was a group event with sibling events then
345          * upgrade the siblings to singleton events by adding them
346          * to the context list directly:
347          */
348         list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
349                 struct list_head *list;
350
351                 list = ctx_group_list(event, ctx);
352                 list_move_tail(&sibling->group_entry, list);
353                 sibling->group_leader = sibling;
354
355                 /* Inherit group flags from the previous leader */
356                 sibling->group_flags = event->group_flags;
357         }
358 }
359
360 static void
361 event_sched_out(struct perf_event *event,
362                   struct perf_cpu_context *cpuctx,
363                   struct perf_event_context *ctx)
364 {
365         if (event->state != PERF_EVENT_STATE_ACTIVE)
366                 return;
367
368         event->state = PERF_EVENT_STATE_INACTIVE;
369         if (event->pending_disable) {
370                 event->pending_disable = 0;
371                 event->state = PERF_EVENT_STATE_OFF;
372         }
373         event->tstamp_stopped = ctx->time;
374         event->pmu->disable(event);
375         event->oncpu = -1;
376
377         if (!is_software_event(event))
378                 cpuctx->active_oncpu--;
379         ctx->nr_active--;
380         if (event->attr.exclusive || !cpuctx->active_oncpu)
381                 cpuctx->exclusive = 0;
382 }
383
384 static void
385 group_sched_out(struct perf_event *group_event,
386                 struct perf_cpu_context *cpuctx,
387                 struct perf_event_context *ctx)
388 {
389         struct perf_event *event;
390
391         if (group_event->state != PERF_EVENT_STATE_ACTIVE)
392                 return;
393
394         event_sched_out(group_event, cpuctx, ctx);
395
396         /*
397          * Schedule out siblings (if any):
398          */
399         list_for_each_entry(event, &group_event->sibling_list, group_entry)
400                 event_sched_out(event, cpuctx, ctx);
401
402         if (group_event->attr.exclusive)
403                 cpuctx->exclusive = 0;
404 }
405
406 /*
407  * Cross CPU call to remove a performance event
408  *
409  * We disable the event on the hardware level first. After that we
410  * remove it from the context list.
411  */
412 static void __perf_event_remove_from_context(void *info)
413 {
414         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
415         struct perf_event *event = info;
416         struct perf_event_context *ctx = event->ctx;
417
418         /*
419          * If this is a task context, we need to check whether it is
420          * the current task context of this cpu. If not it has been
421          * scheduled out before the smp call arrived.
422          */
423         if (ctx->task && cpuctx->task_ctx != ctx)
424                 return;
425
426         raw_spin_lock(&ctx->lock);
427         /*
428          * Protect the list operation against NMI by disabling the
429          * events on a global level.
430          */
431         perf_disable();
432
433         event_sched_out(event, cpuctx, ctx);
434
435         list_del_event(event, ctx);
436
437         if (!ctx->task) {
438                 /*
439                  * Allow more per task events with respect to the
440                  * reservation:
441                  */
442                 cpuctx->max_pertask =
443                         min(perf_max_events - ctx->nr_events,
444                             perf_max_events - perf_reserved_percpu);
445         }
446
447         perf_enable();
448         raw_spin_unlock(&ctx->lock);
449 }
450
451
452 /*
453  * Remove the event from a task's (or a CPU's) list of events.
454  *
455  * Must be called with ctx->mutex held.
456  *
457  * CPU events are removed with a smp call. For task events we only
458  * call when the task is on a CPU.
459  *
460  * If event->ctx is a cloned context, callers must make sure that
461  * every task struct that event->ctx->task could possibly point to
462  * remains valid.  This is OK when called from perf_release since
463  * that only calls us on the top-level context, which can't be a clone.
464  * When called from perf_event_exit_task, it's OK because the
465  * context has been detached from its task.
466  */
467 static void perf_event_remove_from_context(struct perf_event *event)
468 {
469         struct perf_event_context *ctx = event->ctx;
470         struct task_struct *task = ctx->task;
471
472         if (!task) {
473                 /*
474                  * Per cpu events are removed via an smp call and
475                  * the removal is always successful.
476                  */
477                 smp_call_function_single(event->cpu,
478                                          __perf_event_remove_from_context,
479                                          event, 1);
480                 return;
481         }
482
483 retry:
484         task_oncpu_function_call(task, __perf_event_remove_from_context,
485                                  event);
486
487         raw_spin_lock_irq(&ctx->lock);
488         /*
489          * If the context is active we need to retry the smp call.
490          */
491         if (ctx->nr_active && !list_empty(&event->group_entry)) {
492                 raw_spin_unlock_irq(&ctx->lock);
493                 goto retry;
494         }
495
496         /*
497          * The lock prevents that this context is scheduled in so we
498          * can remove the event safely, if the call above did not
499          * succeed.
500          */
501         if (!list_empty(&event->group_entry))
502                 list_del_event(event, ctx);
503         raw_spin_unlock_irq(&ctx->lock);
504 }
505
506 /*
507  * Update total_time_enabled and total_time_running for all events in a group.
508  */
509 static void update_group_times(struct perf_event *leader)
510 {
511         struct perf_event *event;
512
513         update_event_times(leader);
514         list_for_each_entry(event, &leader->sibling_list, group_entry)
515                 update_event_times(event);
516 }
517
518 /*
519  * Cross CPU call to disable a performance event
520  */
521 static void __perf_event_disable(void *info)
522 {
523         struct perf_event *event = info;
524         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
525         struct perf_event_context *ctx = event->ctx;
526
527         /*
528          * If this is a per-task event, need to check whether this
529          * event's task is the current task on this cpu.
530          */
531         if (ctx->task && cpuctx->task_ctx != ctx)
532                 return;
533
534         raw_spin_lock(&ctx->lock);
535
536         /*
537          * If the event is on, turn it off.
538          * If it is in error state, leave it in error state.
539          */
540         if (event->state >= PERF_EVENT_STATE_INACTIVE) {
541                 update_context_time(ctx);
542                 update_group_times(event);
543                 if (event == event->group_leader)
544                         group_sched_out(event, cpuctx, ctx);
545                 else
546                         event_sched_out(event, cpuctx, ctx);
547                 event->state = PERF_EVENT_STATE_OFF;
548         }
549
550         raw_spin_unlock(&ctx->lock);
551 }
552
553 /*
554  * Disable a event.
555  *
556  * If event->ctx is a cloned context, callers must make sure that
557  * every task struct that event->ctx->task could possibly point to
558  * remains valid.  This condition is satisifed when called through
559  * perf_event_for_each_child or perf_event_for_each because they
560  * hold the top-level event's child_mutex, so any descendant that
561  * goes to exit will block in sync_child_event.
562  * When called from perf_pending_event it's OK because event->ctx
563  * is the current context on this CPU and preemption is disabled,
564  * hence we can't get into perf_event_task_sched_out for this context.
565  */
566 void perf_event_disable(struct perf_event *event)
567 {
568         struct perf_event_context *ctx = event->ctx;
569         struct task_struct *task = ctx->task;
570
571         if (!task) {
572                 /*
573                  * Disable the event on the cpu that it's on
574                  */
575                 smp_call_function_single(event->cpu, __perf_event_disable,
576                                          event, 1);
577                 return;
578         }
579
580  retry:
581         task_oncpu_function_call(task, __perf_event_disable, event);
582
583         raw_spin_lock_irq(&ctx->lock);
584         /*
585          * If the event is still active, we need to retry the cross-call.
586          */
587         if (event->state == PERF_EVENT_STATE_ACTIVE) {
588                 raw_spin_unlock_irq(&ctx->lock);
589                 goto retry;
590         }
591
592         /*
593          * Since we have the lock this context can't be scheduled
594          * in, so we can change the state safely.
595          */
596         if (event->state == PERF_EVENT_STATE_INACTIVE) {
597                 update_group_times(event);
598                 event->state = PERF_EVENT_STATE_OFF;
599         }
600
601         raw_spin_unlock_irq(&ctx->lock);
602 }
603
604 static int
605 event_sched_in(struct perf_event *event,
606                  struct perf_cpu_context *cpuctx,
607                  struct perf_event_context *ctx)
608 {
609         if (event->state <= PERF_EVENT_STATE_OFF)
610                 return 0;
611
612         event->state = PERF_EVENT_STATE_ACTIVE;
613         event->oncpu = smp_processor_id();
614         /*
615          * The new state must be visible before we turn it on in the hardware:
616          */
617         smp_wmb();
618
619         if (event->pmu->enable(event)) {
620                 event->state = PERF_EVENT_STATE_INACTIVE;
621                 event->oncpu = -1;
622                 return -EAGAIN;
623         }
624
625         event->tstamp_running += ctx->time - event->tstamp_stopped;
626
627         if (!is_software_event(event))
628                 cpuctx->active_oncpu++;
629         ctx->nr_active++;
630
631         if (event->attr.exclusive)
632                 cpuctx->exclusive = 1;
633
634         return 0;
635 }
636
637 static int
638 group_sched_in(struct perf_event *group_event,
639                struct perf_cpu_context *cpuctx,
640                struct perf_event_context *ctx)
641 {
642         struct perf_event *event, *partial_group;
643         int ret;
644
645         if (group_event->state == PERF_EVENT_STATE_OFF)
646                 return 0;
647
648         ret = hw_perf_group_sched_in(group_event, cpuctx, ctx);
649         if (ret)
650                 return ret < 0 ? ret : 0;
651
652         if (event_sched_in(group_event, cpuctx, ctx))
653                 return -EAGAIN;
654
655         /*
656          * Schedule in siblings as one group (if any):
657          */
658         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
659                 if (event_sched_in(event, cpuctx, ctx)) {
660                         partial_group = event;
661                         goto group_error;
662                 }
663         }
664
665         return 0;
666
667 group_error:
668         /*
669          * Groups can be scheduled in as one unit only, so undo any
670          * partial group before returning:
671          */
672         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
673                 if (event == partial_group)
674                         break;
675                 event_sched_out(event, cpuctx, ctx);
676         }
677         event_sched_out(group_event, cpuctx, ctx);
678
679         return -EAGAIN;
680 }
681
682 /*
683  * Work out whether we can put this event group on the CPU now.
684  */
685 static int group_can_go_on(struct perf_event *event,
686                            struct perf_cpu_context *cpuctx,
687                            int can_add_hw)
688 {
689         /*
690          * Groups consisting entirely of software events can always go on.
691          */
692         if (event->group_flags & PERF_GROUP_SOFTWARE)
693                 return 1;
694         /*
695          * If an exclusive group is already on, no other hardware
696          * events can go on.
697          */
698         if (cpuctx->exclusive)
699                 return 0;
700         /*
701          * If this group is exclusive and there are already
702          * events on the CPU, it can't go on.
703          */
704         if (event->attr.exclusive && cpuctx->active_oncpu)
705                 return 0;
706         /*
707          * Otherwise, try to add it if all previous groups were able
708          * to go on.
709          */
710         return can_add_hw;
711 }
712
713 static void add_event_to_ctx(struct perf_event *event,
714                                struct perf_event_context *ctx)
715 {
716         list_add_event(event, ctx);
717         event->tstamp_enabled = ctx->time;
718         event->tstamp_running = ctx->time;
719         event->tstamp_stopped = ctx->time;
720 }
721
722 /*
723  * Cross CPU call to install and enable a performance event
724  *
725  * Must be called with ctx->mutex held
726  */
727 static void __perf_install_in_context(void *info)
728 {
729         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
730         struct perf_event *event = info;
731         struct perf_event_context *ctx = event->ctx;
732         struct perf_event *leader = event->group_leader;
733         int err;
734
735         /*
736          * If this is a task context, we need to check whether it is
737          * the current task context of this cpu. If not it has been
738          * scheduled out before the smp call arrived.
739          * Or possibly this is the right context but it isn't
740          * on this cpu because it had no events.
741          */
742         if (ctx->task && cpuctx->task_ctx != ctx) {
743                 if (cpuctx->task_ctx || ctx->task != current)
744                         return;
745                 cpuctx->task_ctx = ctx;
746         }
747
748         raw_spin_lock(&ctx->lock);
749         ctx->is_active = 1;
750         update_context_time(ctx);
751
752         /*
753          * Protect the list operation against NMI by disabling the
754          * events on a global level. NOP for non NMI based events.
755          */
756         perf_disable();
757
758         add_event_to_ctx(event, ctx);
759
760         if (event->cpu != -1 && event->cpu != smp_processor_id())
761                 goto unlock;
762
763         /*
764          * Don't put the event on if it is disabled or if
765          * it is in a group and the group isn't on.
766          */
767         if (event->state != PERF_EVENT_STATE_INACTIVE ||
768             (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
769                 goto unlock;
770
771         /*
772          * An exclusive event can't go on if there are already active
773          * hardware events, and no hardware event can go on if there
774          * is already an exclusive event on.
775          */
776         if (!group_can_go_on(event, cpuctx, 1))
777                 err = -EEXIST;
778         else
779                 err = event_sched_in(event, cpuctx, ctx);
780
781         if (err) {
782                 /*
783                  * This event couldn't go on.  If it is in a group
784                  * then we have to pull the whole group off.
785                  * If the event group is pinned then put it in error state.
786                  */
787                 if (leader != event)
788                         group_sched_out(leader, cpuctx, ctx);
789                 if (leader->attr.pinned) {
790                         update_group_times(leader);
791                         leader->state = PERF_EVENT_STATE_ERROR;
792                 }
793         }
794
795         if (!err && !ctx->task && cpuctx->max_pertask)
796                 cpuctx->max_pertask--;
797
798  unlock:
799         perf_enable();
800
801         raw_spin_unlock(&ctx->lock);
802 }
803
804 /*
805  * Attach a performance event to a context
806  *
807  * First we add the event to the list with the hardware enable bit
808  * in event->hw_config cleared.
809  *
810  * If the event is attached to a task which is on a CPU we use a smp
811  * call to enable it in the task context. The task might have been
812  * scheduled away, but we check this in the smp call again.
813  *
814  * Must be called with ctx->mutex held.
815  */
816 static void
817 perf_install_in_context(struct perf_event_context *ctx,
818                         struct perf_event *event,
819                         int cpu)
820 {
821         struct task_struct *task = ctx->task;
822
823         if (!task) {
824                 /*
825                  * Per cpu events are installed via an smp call and
826                  * the install is always successful.
827                  */
828                 smp_call_function_single(cpu, __perf_install_in_context,
829                                          event, 1);
830                 return;
831         }
832
833 retry:
834         task_oncpu_function_call(task, __perf_install_in_context,
835                                  event);
836
837         raw_spin_lock_irq(&ctx->lock);
838         /*
839          * we need to retry the smp call.
840          */
841         if (ctx->is_active && list_empty(&event->group_entry)) {
842                 raw_spin_unlock_irq(&ctx->lock);
843                 goto retry;
844         }
845
846         /*
847          * The lock prevents that this context is scheduled in so we
848          * can add the event safely, if it the call above did not
849          * succeed.
850          */
851         if (list_empty(&event->group_entry))
852                 add_event_to_ctx(event, ctx);
853         raw_spin_unlock_irq(&ctx->lock);
854 }
855
856 /*
857  * Put a event into inactive state and update time fields.
858  * Enabling the leader of a group effectively enables all
859  * the group members that aren't explicitly disabled, so we
860  * have to update their ->tstamp_enabled also.
861  * Note: this works for group members as well as group leaders
862  * since the non-leader members' sibling_lists will be empty.
863  */
864 static void __perf_event_mark_enabled(struct perf_event *event,
865                                         struct perf_event_context *ctx)
866 {
867         struct perf_event *sub;
868
869         event->state = PERF_EVENT_STATE_INACTIVE;
870         event->tstamp_enabled = ctx->time - event->total_time_enabled;
871         list_for_each_entry(sub, &event->sibling_list, group_entry)
872                 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
873                         sub->tstamp_enabled =
874                                 ctx->time - sub->total_time_enabled;
875 }
876
877 /*
878  * Cross CPU call to enable a performance event
879  */
880 static void __perf_event_enable(void *info)
881 {
882         struct perf_event *event = info;
883         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
884         struct perf_event_context *ctx = event->ctx;
885         struct perf_event *leader = event->group_leader;
886         int err;
887
888         /*
889          * If this is a per-task event, need to check whether this
890          * event's task is the current task on this cpu.
891          */
892         if (ctx->task && cpuctx->task_ctx != ctx) {
893                 if (cpuctx->task_ctx || ctx->task != current)
894                         return;
895                 cpuctx->task_ctx = ctx;
896         }
897
898         raw_spin_lock(&ctx->lock);
899         ctx->is_active = 1;
900         update_context_time(ctx);
901
902         if (event->state >= PERF_EVENT_STATE_INACTIVE)
903                 goto unlock;
904         __perf_event_mark_enabled(event, ctx);
905
906         if (event->cpu != -1 && event->cpu != smp_processor_id())
907                 goto unlock;
908
909         /*
910          * If the event is in a group and isn't the group leader,
911          * then don't put it on unless the group is on.
912          */
913         if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
914                 goto unlock;
915
916         if (!group_can_go_on(event, cpuctx, 1)) {
917                 err = -EEXIST;
918         } else {
919                 perf_disable();
920                 if (event == leader)
921                         err = group_sched_in(event, cpuctx, ctx);
922                 else
923                         err = event_sched_in(event, cpuctx, ctx);
924                 perf_enable();
925         }
926
927         if (err) {
928                 /*
929                  * If this event can't go on and it's part of a
930                  * group, then the whole group has to come off.
931                  */
932                 if (leader != event)
933                         group_sched_out(leader, cpuctx, ctx);
934                 if (leader->attr.pinned) {
935                         update_group_times(leader);
936                         leader->state = PERF_EVENT_STATE_ERROR;
937                 }
938         }
939
940  unlock:
941         raw_spin_unlock(&ctx->lock);
942 }
943
944 /*
945  * Enable a event.
946  *
947  * If event->ctx is a cloned context, callers must make sure that
948  * every task struct that event->ctx->task could possibly point to
949  * remains valid.  This condition is satisfied when called through
950  * perf_event_for_each_child or perf_event_for_each as described
951  * for perf_event_disable.
952  */
953 void perf_event_enable(struct perf_event *event)
954 {
955         struct perf_event_context *ctx = event->ctx;
956         struct task_struct *task = ctx->task;
957
958         if (!task) {
959                 /*
960                  * Enable the event on the cpu that it's on
961                  */
962                 smp_call_function_single(event->cpu, __perf_event_enable,
963                                          event, 1);
964                 return;
965         }
966
967         raw_spin_lock_irq(&ctx->lock);
968         if (event->state >= PERF_EVENT_STATE_INACTIVE)
969                 goto out;
970
971         /*
972          * If the event is in error state, clear that first.
973          * That way, if we see the event in error state below, we
974          * know that it has gone back into error state, as distinct
975          * from the task having been scheduled away before the
976          * cross-call arrived.
977          */
978         if (event->state == PERF_EVENT_STATE_ERROR)
979                 event->state = PERF_EVENT_STATE_OFF;
980
981  retry:
982         raw_spin_unlock_irq(&ctx->lock);
983         task_oncpu_function_call(task, __perf_event_enable, event);
984
985         raw_spin_lock_irq(&ctx->lock);
986
987         /*
988          * If the context is active and the event is still off,
989          * we need to retry the cross-call.
990          */
991         if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
992                 goto retry;
993
994         /*
995          * Since we have the lock this context can't be scheduled
996          * in, so we can change the state safely.
997          */
998         if (event->state == PERF_EVENT_STATE_OFF)
999                 __perf_event_mark_enabled(event, ctx);
1000
1001  out:
1002         raw_spin_unlock_irq(&ctx->lock);
1003 }
1004
1005 static int perf_event_refresh(struct perf_event *event, int refresh)
1006 {
1007         /*
1008          * not supported on inherited events
1009          */
1010         if (event->attr.inherit)
1011                 return -EINVAL;
1012
1013         atomic_add(refresh, &event->event_limit);
1014         perf_event_enable(event);
1015
1016         return 0;
1017 }
1018
1019 enum event_type_t {
1020         EVENT_FLEXIBLE = 0x1,
1021         EVENT_PINNED = 0x2,
1022         EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1023 };
1024
1025 static void ctx_sched_out(struct perf_event_context *ctx,
1026                           struct perf_cpu_context *cpuctx,
1027                           enum event_type_t event_type)
1028 {
1029         struct perf_event *event;
1030
1031         raw_spin_lock(&ctx->lock);
1032         ctx->is_active = 0;
1033         if (likely(!ctx->nr_events))
1034                 goto out;
1035         update_context_time(ctx);
1036
1037         perf_disable();
1038         if (!ctx->nr_active)
1039                 goto out_enable;
1040
1041         if (event_type & EVENT_PINNED)
1042                 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1043                         group_sched_out(event, cpuctx, ctx);
1044
1045         if (event_type & EVENT_FLEXIBLE)
1046                 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1047                         group_sched_out(event, cpuctx, ctx);
1048
1049  out_enable:
1050         perf_enable();
1051  out:
1052         raw_spin_unlock(&ctx->lock);
1053 }
1054
1055 /*
1056  * Test whether two contexts are equivalent, i.e. whether they
1057  * have both been cloned from the same version of the same context
1058  * and they both have the same number of enabled events.
1059  * If the number of enabled events is the same, then the set
1060  * of enabled events should be the same, because these are both
1061  * inherited contexts, therefore we can't access individual events
1062  * in them directly with an fd; we can only enable/disable all
1063  * events via prctl, or enable/disable all events in a family
1064  * via ioctl, which will have the same effect on both contexts.
1065  */
1066 static int context_equiv(struct perf_event_context *ctx1,
1067                          struct perf_event_context *ctx2)
1068 {
1069         return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1070                 && ctx1->parent_gen == ctx2->parent_gen
1071                 && !ctx1->pin_count && !ctx2->pin_count;
1072 }
1073
1074 static void __perf_event_sync_stat(struct perf_event *event,
1075                                      struct perf_event *next_event)
1076 {
1077         u64 value;
1078
1079         if (!event->attr.inherit_stat)
1080                 return;
1081
1082         /*
1083          * Update the event value, we cannot use perf_event_read()
1084          * because we're in the middle of a context switch and have IRQs
1085          * disabled, which upsets smp_call_function_single(), however
1086          * we know the event must be on the current CPU, therefore we
1087          * don't need to use it.
1088          */
1089         switch (event->state) {
1090         case PERF_EVENT_STATE_ACTIVE:
1091                 event->pmu->read(event);
1092                 /* fall-through */
1093
1094         case PERF_EVENT_STATE_INACTIVE:
1095                 update_event_times(event);
1096                 break;
1097
1098         default:
1099                 break;
1100         }
1101
1102         /*
1103          * In order to keep per-task stats reliable we need to flip the event
1104          * values when we flip the contexts.
1105          */
1106         value = atomic64_read(&next_event->count);
1107         value = atomic64_xchg(&event->count, value);
1108         atomic64_set(&next_event->count, value);
1109
1110         swap(event->total_time_enabled, next_event->total_time_enabled);
1111         swap(event->total_time_running, next_event->total_time_running);
1112
1113         /*
1114          * Since we swizzled the values, update the user visible data too.
1115          */
1116         perf_event_update_userpage(event);
1117         perf_event_update_userpage(next_event);
1118 }
1119
1120 #define list_next_entry(pos, member) \
1121         list_entry(pos->member.next, typeof(*pos), member)
1122
1123 static void perf_event_sync_stat(struct perf_event_context *ctx,
1124                                    struct perf_event_context *next_ctx)
1125 {
1126         struct perf_event *event, *next_event;
1127
1128         if (!ctx->nr_stat)
1129                 return;
1130
1131         update_context_time(ctx);
1132
1133         event = list_first_entry(&ctx->event_list,
1134                                    struct perf_event, event_entry);
1135
1136         next_event = list_first_entry(&next_ctx->event_list,
1137                                         struct perf_event, event_entry);
1138
1139         while (&event->event_entry != &ctx->event_list &&
1140                &next_event->event_entry != &next_ctx->event_list) {
1141
1142                 __perf_event_sync_stat(event, next_event);
1143
1144                 event = list_next_entry(event, event_entry);
1145                 next_event = list_next_entry(next_event, event_entry);
1146         }
1147 }
1148
1149 /*
1150  * Called from scheduler to remove the events of the current task,
1151  * with interrupts disabled.
1152  *
1153  * We stop each event and update the event value in event->count.
1154  *
1155  * This does not protect us against NMI, but disable()
1156  * sets the disabled bit in the control field of event _before_
1157  * accessing the event control register. If a NMI hits, then it will
1158  * not restart the event.
1159  */
1160 void perf_event_task_sched_out(struct task_struct *task,
1161                                  struct task_struct *next)
1162 {
1163         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1164         struct perf_event_context *ctx = task->perf_event_ctxp;
1165         struct perf_event_context *next_ctx;
1166         struct perf_event_context *parent;
1167         struct pt_regs *regs;
1168         int do_switch = 1;
1169
1170         regs = task_pt_regs(task);
1171         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1172
1173         if (likely(!ctx || !cpuctx->task_ctx))
1174                 return;
1175
1176         rcu_read_lock();
1177         parent = rcu_dereference(ctx->parent_ctx);
1178         next_ctx = next->perf_event_ctxp;
1179         if (parent && next_ctx &&
1180             rcu_dereference(next_ctx->parent_ctx) == parent) {
1181                 /*
1182                  * Looks like the two contexts are clones, so we might be
1183                  * able to optimize the context switch.  We lock both
1184                  * contexts and check that they are clones under the
1185                  * lock (including re-checking that neither has been
1186                  * uncloned in the meantime).  It doesn't matter which
1187                  * order we take the locks because no other cpu could
1188                  * be trying to lock both of these tasks.
1189                  */
1190                 raw_spin_lock(&ctx->lock);
1191                 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1192                 if (context_equiv(ctx, next_ctx)) {
1193                         /*
1194                          * XXX do we need a memory barrier of sorts
1195                          * wrt to rcu_dereference() of perf_event_ctxp
1196                          */
1197                         task->perf_event_ctxp = next_ctx;
1198                         next->perf_event_ctxp = ctx;
1199                         ctx->task = next;
1200                         next_ctx->task = task;
1201                         do_switch = 0;
1202
1203                         perf_event_sync_stat(ctx, next_ctx);
1204                 }
1205                 raw_spin_unlock(&next_ctx->lock);
1206                 raw_spin_unlock(&ctx->lock);
1207         }
1208         rcu_read_unlock();
1209
1210         if (do_switch) {
1211                 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1212                 cpuctx->task_ctx = NULL;
1213         }
1214 }
1215
1216 static void task_ctx_sched_out(struct perf_event_context *ctx,
1217                                enum event_type_t event_type)
1218 {
1219         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1220
1221         if (!cpuctx->task_ctx)
1222                 return;
1223
1224         if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1225                 return;
1226
1227         ctx_sched_out(ctx, cpuctx, event_type);
1228         cpuctx->task_ctx = NULL;
1229 }
1230
1231 /*
1232  * Called with IRQs disabled
1233  */
1234 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1235 {
1236         task_ctx_sched_out(ctx, EVENT_ALL);
1237 }
1238
1239 /*
1240  * Called with IRQs disabled
1241  */
1242 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1243                               enum event_type_t event_type)
1244 {
1245         ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1246 }
1247
1248 static void
1249 ctx_pinned_sched_in(struct perf_event_context *ctx,
1250                     struct perf_cpu_context *cpuctx)
1251 {
1252         struct perf_event *event;
1253
1254         list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1255                 if (event->state <= PERF_EVENT_STATE_OFF)
1256                         continue;
1257                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1258                         continue;
1259
1260                 if (group_can_go_on(event, cpuctx, 1))
1261                         group_sched_in(event, cpuctx, ctx);
1262
1263                 /*
1264                  * If this pinned group hasn't been scheduled,
1265                  * put it in error state.
1266                  */
1267                 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1268                         update_group_times(event);
1269                         event->state = PERF_EVENT_STATE_ERROR;
1270                 }
1271         }
1272 }
1273
1274 static void
1275 ctx_flexible_sched_in(struct perf_event_context *ctx,
1276                       struct perf_cpu_context *cpuctx)
1277 {
1278         struct perf_event *event;
1279         int can_add_hw = 1;
1280
1281         list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1282                 /* Ignore events in OFF or ERROR state */
1283                 if (event->state <= PERF_EVENT_STATE_OFF)
1284                         continue;
1285                 /*
1286                  * Listen to the 'cpu' scheduling filter constraint
1287                  * of events:
1288                  */
1289                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1290                         continue;
1291
1292                 if (group_can_go_on(event, cpuctx, can_add_hw))
1293                         if (group_sched_in(event, cpuctx, ctx))
1294                                 can_add_hw = 0;
1295         }
1296 }
1297
1298 static void
1299 ctx_sched_in(struct perf_event_context *ctx,
1300              struct perf_cpu_context *cpuctx,
1301              enum event_type_t event_type)
1302 {
1303         raw_spin_lock(&ctx->lock);
1304         ctx->is_active = 1;
1305         if (likely(!ctx->nr_events))
1306                 goto out;
1307
1308         ctx->timestamp = perf_clock();
1309
1310         perf_disable();
1311
1312         /*
1313          * First go through the list and put on any pinned groups
1314          * in order to give them the best chance of going on.
1315          */
1316         if (event_type & EVENT_PINNED)
1317                 ctx_pinned_sched_in(ctx, cpuctx);
1318
1319         /* Then walk through the lower prio flexible groups */
1320         if (event_type & EVENT_FLEXIBLE)
1321                 ctx_flexible_sched_in(ctx, cpuctx);
1322
1323         perf_enable();
1324  out:
1325         raw_spin_unlock(&ctx->lock);
1326 }
1327
1328 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1329                              enum event_type_t event_type)
1330 {
1331         struct perf_event_context *ctx = &cpuctx->ctx;
1332
1333         ctx_sched_in(ctx, cpuctx, event_type);
1334 }
1335
1336 static void task_ctx_sched_in(struct task_struct *task,
1337                               enum event_type_t event_type)
1338 {
1339         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1340         struct perf_event_context *ctx = task->perf_event_ctxp;
1341
1342         if (likely(!ctx))
1343                 return;
1344         if (cpuctx->task_ctx == ctx)
1345                 return;
1346         ctx_sched_in(ctx, cpuctx, event_type);
1347         cpuctx->task_ctx = ctx;
1348 }
1349 /*
1350  * Called from scheduler to add the events of the current task
1351  * with interrupts disabled.
1352  *
1353  * We restore the event value and then enable it.
1354  *
1355  * This does not protect us against NMI, but enable()
1356  * sets the enabled bit in the control field of event _before_
1357  * accessing the event control register. If a NMI hits, then it will
1358  * keep the event running.
1359  */
1360 void perf_event_task_sched_in(struct task_struct *task)
1361 {
1362         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1363         struct perf_event_context *ctx = task->perf_event_ctxp;
1364
1365         if (likely(!ctx))
1366                 return;
1367
1368         if (cpuctx->task_ctx == ctx)
1369                 return;
1370
1371         /*
1372          * We want to keep the following priority order:
1373          * cpu pinned (that don't need to move), task pinned,
1374          * cpu flexible, task flexible.
1375          */
1376         cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1377
1378         ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1379         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1380         ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1381
1382         cpuctx->task_ctx = ctx;
1383 }
1384
1385 #define MAX_INTERRUPTS (~0ULL)
1386
1387 static void perf_log_throttle(struct perf_event *event, int enable);
1388
1389 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1390 {
1391         u64 frequency = event->attr.sample_freq;
1392         u64 sec = NSEC_PER_SEC;
1393         u64 divisor, dividend;
1394
1395         int count_fls, nsec_fls, frequency_fls, sec_fls;
1396
1397         count_fls = fls64(count);
1398         nsec_fls = fls64(nsec);
1399         frequency_fls = fls64(frequency);
1400         sec_fls = 30;
1401
1402         /*
1403          * We got @count in @nsec, with a target of sample_freq HZ
1404          * the target period becomes:
1405          *
1406          *             @count * 10^9
1407          * period = -------------------
1408          *          @nsec * sample_freq
1409          *
1410          */
1411
1412         /*
1413          * Reduce accuracy by one bit such that @a and @b converge
1414          * to a similar magnitude.
1415          */
1416 #define REDUCE_FLS(a, b)                \
1417 do {                                    \
1418         if (a##_fls > b##_fls) {        \
1419                 a >>= 1;                \
1420                 a##_fls--;              \
1421         } else {                        \
1422                 b >>= 1;                \
1423                 b##_fls--;              \
1424         }                               \
1425 } while (0)
1426
1427         /*
1428          * Reduce accuracy until either term fits in a u64, then proceed with
1429          * the other, so that finally we can do a u64/u64 division.
1430          */
1431         while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1432                 REDUCE_FLS(nsec, frequency);
1433                 REDUCE_FLS(sec, count);
1434         }
1435
1436         if (count_fls + sec_fls > 64) {
1437                 divisor = nsec * frequency;
1438
1439                 while (count_fls + sec_fls > 64) {
1440                         REDUCE_FLS(count, sec);
1441                         divisor >>= 1;
1442                 }
1443
1444                 dividend = count * sec;
1445         } else {
1446                 dividend = count * sec;
1447
1448                 while (nsec_fls + frequency_fls > 64) {
1449                         REDUCE_FLS(nsec, frequency);
1450                         dividend >>= 1;
1451                 }
1452
1453                 divisor = nsec * frequency;
1454         }
1455
1456         return div64_u64(dividend, divisor);
1457 }
1458
1459 static void perf_event_stop(struct perf_event *event)
1460 {
1461         if (!event->pmu->stop)
1462                 return event->pmu->disable(event);
1463
1464         return event->pmu->stop(event);
1465 }
1466
1467 static int perf_event_start(struct perf_event *event)
1468 {
1469         if (!event->pmu->start)
1470                 return event->pmu->enable(event);
1471
1472         return event->pmu->start(event);
1473 }
1474
1475 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1476 {
1477         struct hw_perf_event *hwc = &event->hw;
1478         u64 period, sample_period;
1479         s64 delta;
1480
1481         period = perf_calculate_period(event, nsec, count);
1482
1483         delta = (s64)(period - hwc->sample_period);
1484         delta = (delta + 7) / 8; /* low pass filter */
1485
1486         sample_period = hwc->sample_period + delta;
1487
1488         if (!sample_period)
1489                 sample_period = 1;
1490
1491         hwc->sample_period = sample_period;
1492
1493         if (atomic64_read(&hwc->period_left) > 8*sample_period) {
1494                 perf_disable();
1495                 perf_event_stop(event);
1496                 atomic64_set(&hwc->period_left, 0);
1497                 perf_event_start(event);
1498                 perf_enable();
1499         }
1500 }
1501
1502 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1503 {
1504         struct perf_event *event;
1505         struct hw_perf_event *hwc;
1506         u64 interrupts, now;
1507         s64 delta;
1508
1509         raw_spin_lock(&ctx->lock);
1510         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1511                 if (event->state != PERF_EVENT_STATE_ACTIVE)
1512                         continue;
1513
1514                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1515                         continue;
1516
1517                 hwc = &event->hw;
1518
1519                 interrupts = hwc->interrupts;
1520                 hwc->interrupts = 0;
1521
1522                 /*
1523                  * unthrottle events on the tick
1524                  */
1525                 if (interrupts == MAX_INTERRUPTS) {
1526                         perf_log_throttle(event, 1);
1527                         perf_disable();
1528                         event->pmu->unthrottle(event);
1529                         perf_enable();
1530                 }
1531
1532                 if (!event->attr.freq || !event->attr.sample_freq)
1533                         continue;
1534
1535                 perf_disable();
1536                 event->pmu->read(event);
1537                 now = atomic64_read(&event->count);
1538                 delta = now - hwc->freq_count_stamp;
1539                 hwc->freq_count_stamp = now;
1540
1541                 if (delta > 0)
1542                         perf_adjust_period(event, TICK_NSEC, delta);
1543                 perf_enable();
1544         }
1545         raw_spin_unlock(&ctx->lock);
1546 }
1547
1548 /*
1549  * Round-robin a context's events:
1550  */
1551 static void rotate_ctx(struct perf_event_context *ctx)
1552 {
1553         raw_spin_lock(&ctx->lock);
1554
1555         /* Rotate the first entry last of non-pinned groups */
1556         list_rotate_left(&ctx->flexible_groups);
1557
1558         raw_spin_unlock(&ctx->lock);
1559 }
1560
1561 void perf_event_task_tick(struct task_struct *curr)
1562 {
1563         struct perf_cpu_context *cpuctx;
1564         struct perf_event_context *ctx;
1565         int rotate = 0;
1566
1567         if (!atomic_read(&nr_events))
1568                 return;
1569
1570         cpuctx = &__get_cpu_var(perf_cpu_context);
1571         if (cpuctx->ctx.nr_events &&
1572             cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1573                 rotate = 1;
1574
1575         ctx = curr->perf_event_ctxp;
1576         if (ctx && ctx->nr_events && ctx->nr_events != ctx->nr_active)
1577                 rotate = 1;
1578
1579         perf_ctx_adjust_freq(&cpuctx->ctx);
1580         if (ctx)
1581                 perf_ctx_adjust_freq(ctx);
1582
1583         if (!rotate)
1584                 return;
1585
1586         perf_disable();
1587         cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1588         if (ctx)
1589                 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1590
1591         rotate_ctx(&cpuctx->ctx);
1592         if (ctx)
1593                 rotate_ctx(ctx);
1594
1595         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1596         if (ctx)
1597                 task_ctx_sched_in(curr, EVENT_FLEXIBLE);
1598         perf_enable();
1599 }
1600
1601 static int event_enable_on_exec(struct perf_event *event,
1602                                 struct perf_event_context *ctx)
1603 {
1604         if (!event->attr.enable_on_exec)
1605                 return 0;
1606
1607         event->attr.enable_on_exec = 0;
1608         if (event->state >= PERF_EVENT_STATE_INACTIVE)
1609                 return 0;
1610
1611         __perf_event_mark_enabled(event, ctx);
1612
1613         return 1;
1614 }
1615
1616 /*
1617  * Enable all of a task's events that have been marked enable-on-exec.
1618  * This expects task == current.
1619  */
1620 static void perf_event_enable_on_exec(struct task_struct *task)
1621 {
1622         struct perf_event_context *ctx;
1623         struct perf_event *event;
1624         unsigned long flags;
1625         int enabled = 0;
1626         int ret;
1627
1628         local_irq_save(flags);
1629         ctx = task->perf_event_ctxp;
1630         if (!ctx || !ctx->nr_events)
1631                 goto out;
1632
1633         __perf_event_task_sched_out(ctx);
1634
1635         raw_spin_lock(&ctx->lock);
1636
1637         list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1638                 ret = event_enable_on_exec(event, ctx);
1639                 if (ret)
1640                         enabled = 1;
1641         }
1642
1643         list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1644                 ret = event_enable_on_exec(event, ctx);
1645                 if (ret)
1646                         enabled = 1;
1647         }
1648
1649         /*
1650          * Unclone this context if we enabled any event.
1651          */
1652         if (enabled)
1653                 unclone_ctx(ctx);
1654
1655         raw_spin_unlock(&ctx->lock);
1656
1657         perf_event_task_sched_in(task);
1658  out:
1659         local_irq_restore(flags);
1660 }
1661
1662 /*
1663  * Cross CPU call to read the hardware event
1664  */
1665 static void __perf_event_read(void *info)
1666 {
1667         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1668         struct perf_event *event = info;
1669         struct perf_event_context *ctx = event->ctx;
1670
1671         /*
1672          * If this is a task context, we need to check whether it is
1673          * the current task context of this cpu.  If not it has been
1674          * scheduled out before the smp call arrived.  In that case
1675          * event->count would have been updated to a recent sample
1676          * when the event was scheduled out.
1677          */
1678         if (ctx->task && cpuctx->task_ctx != ctx)
1679                 return;
1680
1681         raw_spin_lock(&ctx->lock);
1682         update_context_time(ctx);
1683         update_event_times(event);
1684         raw_spin_unlock(&ctx->lock);
1685
1686         event->pmu->read(event);
1687 }
1688
1689 static u64 perf_event_read(struct perf_event *event)
1690 {
1691         /*
1692          * If event is enabled and currently active on a CPU, update the
1693          * value in the event structure:
1694          */
1695         if (event->state == PERF_EVENT_STATE_ACTIVE) {
1696                 smp_call_function_single(event->oncpu,
1697                                          __perf_event_read, event, 1);
1698         } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1699                 struct perf_event_context *ctx = event->ctx;
1700                 unsigned long flags;
1701
1702                 raw_spin_lock_irqsave(&ctx->lock, flags);
1703                 update_context_time(ctx);
1704                 update_event_times(event);
1705                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1706         }
1707
1708         return atomic64_read(&event->count);
1709 }
1710
1711 /*
1712  * Initialize the perf_event context in a task_struct:
1713  */
1714 static void
1715 __perf_event_init_context(struct perf_event_context *ctx,
1716                             struct task_struct *task)
1717 {
1718         raw_spin_lock_init(&ctx->lock);
1719         mutex_init(&ctx->mutex);
1720         INIT_LIST_HEAD(&ctx->pinned_groups);
1721         INIT_LIST_HEAD(&ctx->flexible_groups);
1722         INIT_LIST_HEAD(&ctx->event_list);
1723         atomic_set(&ctx->refcount, 1);
1724         ctx->task = task;
1725 }
1726
1727 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1728 {
1729         struct perf_event_context *ctx;
1730         struct perf_cpu_context *cpuctx;
1731         struct task_struct *task;
1732         unsigned long flags;
1733         int err;
1734
1735         if (pid == -1 && cpu != -1) {
1736                 /* Must be root to operate on a CPU event: */
1737                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1738                         return ERR_PTR(-EACCES);
1739
1740                 if (cpu < 0 || cpu >= nr_cpumask_bits)
1741                         return ERR_PTR(-EINVAL);
1742
1743                 /*
1744                  * We could be clever and allow to attach a event to an
1745                  * offline CPU and activate it when the CPU comes up, but
1746                  * that's for later.
1747                  */
1748                 if (!cpu_online(cpu))
1749                         return ERR_PTR(-ENODEV);
1750
1751                 cpuctx = &per_cpu(perf_cpu_context, cpu);
1752                 ctx = &cpuctx->ctx;
1753                 get_ctx(ctx);
1754
1755                 return ctx;
1756         }
1757
1758         rcu_read_lock();
1759         if (!pid)
1760                 task = current;
1761         else
1762                 task = find_task_by_vpid(pid);
1763         if (task)
1764                 get_task_struct(task);
1765         rcu_read_unlock();
1766
1767         if (!task)
1768                 return ERR_PTR(-ESRCH);
1769
1770         /*
1771          * Can't attach events to a dying task.
1772          */
1773         err = -ESRCH;
1774         if (task->flags & PF_EXITING)
1775                 goto errout;
1776
1777         /* Reuse ptrace permission checks for now. */
1778         err = -EACCES;
1779         if (!ptrace_may_access(task, PTRACE_MODE_READ))
1780                 goto errout;
1781
1782  retry:
1783         ctx = perf_lock_task_context(task, &flags);
1784         if (ctx) {
1785                 unclone_ctx(ctx);
1786                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1787         }
1788
1789         if (!ctx) {
1790                 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1791                 err = -ENOMEM;
1792                 if (!ctx)
1793                         goto errout;
1794                 __perf_event_init_context(ctx, task);
1795                 get_ctx(ctx);
1796                 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1797                         /*
1798                          * We raced with some other task; use
1799                          * the context they set.
1800                          */
1801                         kfree(ctx);
1802                         goto retry;
1803                 }
1804                 get_task_struct(task);
1805         }
1806
1807         put_task_struct(task);
1808         return ctx;
1809
1810  errout:
1811         put_task_struct(task);
1812         return ERR_PTR(err);
1813 }
1814
1815 static void perf_event_free_filter(struct perf_event *event);
1816
1817 static void free_event_rcu(struct rcu_head *head)
1818 {
1819         struct perf_event *event;
1820
1821         event = container_of(head, struct perf_event, rcu_head);
1822         if (event->ns)
1823                 put_pid_ns(event->ns);
1824         perf_event_free_filter(event);
1825         kfree(event);
1826 }
1827
1828 static void perf_pending_sync(struct perf_event *event);
1829
1830 static void free_event(struct perf_event *event)
1831 {
1832         perf_pending_sync(event);
1833
1834         if (!event->parent) {
1835                 atomic_dec(&nr_events);
1836                 if (event->attr.mmap)
1837                         atomic_dec(&nr_mmap_events);
1838                 if (event->attr.comm)
1839                         atomic_dec(&nr_comm_events);
1840                 if (event->attr.task)
1841                         atomic_dec(&nr_task_events);
1842         }
1843
1844         if (event->output) {
1845                 fput(event->output->filp);
1846                 event->output = NULL;
1847         }
1848
1849         if (event->destroy)
1850                 event->destroy(event);
1851
1852         put_ctx(event->ctx);
1853         call_rcu(&event->rcu_head, free_event_rcu);
1854 }
1855
1856 int perf_event_release_kernel(struct perf_event *event)
1857 {
1858         struct perf_event_context *ctx = event->ctx;
1859
1860         WARN_ON_ONCE(ctx->parent_ctx);
1861         mutex_lock(&ctx->mutex);
1862         perf_event_remove_from_context(event);
1863         mutex_unlock(&ctx->mutex);
1864
1865         mutex_lock(&event->owner->perf_event_mutex);
1866         list_del_init(&event->owner_entry);
1867         mutex_unlock(&event->owner->perf_event_mutex);
1868         put_task_struct(event->owner);
1869
1870         free_event(event);
1871
1872         return 0;
1873 }
1874 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1875
1876 /*
1877  * Called when the last reference to the file is gone.
1878  */
1879 static int perf_release(struct inode *inode, struct file *file)
1880 {
1881         struct perf_event *event = file->private_data;
1882
1883         file->private_data = NULL;
1884
1885         return perf_event_release_kernel(event);
1886 }
1887
1888 static int perf_event_read_size(struct perf_event *event)
1889 {
1890         int entry = sizeof(u64); /* value */
1891         int size = 0;
1892         int nr = 1;
1893
1894         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1895                 size += sizeof(u64);
1896
1897         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1898                 size += sizeof(u64);
1899
1900         if (event->attr.read_format & PERF_FORMAT_ID)
1901                 entry += sizeof(u64);
1902
1903         if (event->attr.read_format & PERF_FORMAT_GROUP) {
1904                 nr += event->group_leader->nr_siblings;
1905                 size += sizeof(u64);
1906         }
1907
1908         size += entry * nr;
1909
1910         return size;
1911 }
1912
1913 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1914 {
1915         struct perf_event *child;
1916         u64 total = 0;
1917
1918         *enabled = 0;
1919         *running = 0;
1920
1921         mutex_lock(&event->child_mutex);
1922         total += perf_event_read(event);
1923         *enabled += event->total_time_enabled +
1924                         atomic64_read(&event->child_total_time_enabled);
1925         *running += event->total_time_running +
1926                         atomic64_read(&event->child_total_time_running);
1927
1928         list_for_each_entry(child, &event->child_list, child_list) {
1929                 total += perf_event_read(child);
1930                 *enabled += child->total_time_enabled;
1931                 *running += child->total_time_running;
1932         }
1933         mutex_unlock(&event->child_mutex);
1934
1935         return total;
1936 }
1937 EXPORT_SYMBOL_GPL(perf_event_read_value);
1938
1939 static int perf_event_read_group(struct perf_event *event,
1940                                    u64 read_format, char __user *buf)
1941 {
1942         struct perf_event *leader = event->group_leader, *sub;
1943         int n = 0, size = 0, ret = -EFAULT;
1944         struct perf_event_context *ctx = leader->ctx;
1945         u64 values[5];
1946         u64 count, enabled, running;
1947
1948         mutex_lock(&ctx->mutex);
1949         count = perf_event_read_value(leader, &enabled, &running);
1950
1951         values[n++] = 1 + leader->nr_siblings;
1952         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1953                 values[n++] = enabled;
1954         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1955                 values[n++] = running;
1956         values[n++] = count;
1957         if (read_format & PERF_FORMAT_ID)
1958                 values[n++] = primary_event_id(leader);
1959
1960         size = n * sizeof(u64);
1961
1962         if (copy_to_user(buf, values, size))
1963                 goto unlock;
1964
1965         ret = size;
1966
1967         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1968                 n = 0;
1969
1970                 values[n++] = perf_event_read_value(sub, &enabled, &running);
1971                 if (read_format & PERF_FORMAT_ID)
1972                         values[n++] = primary_event_id(sub);
1973
1974                 size = n * sizeof(u64);
1975
1976                 if (copy_to_user(buf + ret, values, size)) {
1977                         ret = -EFAULT;
1978                         goto unlock;
1979                 }
1980
1981                 ret += size;
1982         }
1983 unlock:
1984         mutex_unlock(&ctx->mutex);
1985
1986         return ret;
1987 }
1988
1989 static int perf_event_read_one(struct perf_event *event,
1990                                  u64 read_format, char __user *buf)
1991 {
1992         u64 enabled, running;
1993         u64 values[4];
1994         int n = 0;
1995
1996         values[n++] = perf_event_read_value(event, &enabled, &running);
1997         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1998                 values[n++] = enabled;
1999         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2000                 values[n++] = running;
2001         if (read_format & PERF_FORMAT_ID)
2002                 values[n++] = primary_event_id(event);
2003
2004         if (copy_to_user(buf, values, n * sizeof(u64)))
2005                 return -EFAULT;
2006
2007         return n * sizeof(u64);
2008 }
2009
2010 /*
2011  * Read the performance event - simple non blocking version for now
2012  */
2013 static ssize_t
2014 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2015 {
2016         u64 read_format = event->attr.read_format;
2017         int ret;
2018
2019         /*
2020          * Return end-of-file for a read on a event that is in
2021          * error state (i.e. because it was pinned but it couldn't be
2022          * scheduled on to the CPU at some point).
2023          */
2024         if (event->state == PERF_EVENT_STATE_ERROR)
2025                 return 0;
2026
2027         if (count < perf_event_read_size(event))
2028                 return -ENOSPC;
2029
2030         WARN_ON_ONCE(event->ctx->parent_ctx);
2031         if (read_format & PERF_FORMAT_GROUP)
2032                 ret = perf_event_read_group(event, read_format, buf);
2033         else
2034                 ret = perf_event_read_one(event, read_format, buf);
2035
2036         return ret;
2037 }
2038
2039 static ssize_t
2040 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2041 {
2042         struct perf_event *event = file->private_data;
2043
2044         return perf_read_hw(event, buf, count);
2045 }
2046
2047 static unsigned int perf_poll(struct file *file, poll_table *wait)
2048 {
2049         struct perf_event *event = file->private_data;
2050         struct perf_mmap_data *data;
2051         unsigned int events = POLL_HUP;
2052
2053         rcu_read_lock();
2054         data = rcu_dereference(event->data);
2055         if (data)
2056                 events = atomic_xchg(&data->poll, 0);
2057         rcu_read_unlock();
2058
2059         poll_wait(file, &event->waitq, wait);
2060
2061         return events;
2062 }
2063
2064 static void perf_event_reset(struct perf_event *event)
2065 {
2066         (void)perf_event_read(event);
2067         atomic64_set(&event->count, 0);
2068         perf_event_update_userpage(event);
2069 }
2070
2071 /*
2072  * Holding the top-level event's child_mutex means that any
2073  * descendant process that has inherited this event will block
2074  * in sync_child_event if it goes to exit, thus satisfying the
2075  * task existence requirements of perf_event_enable/disable.
2076  */
2077 static void perf_event_for_each_child(struct perf_event *event,
2078                                         void (*func)(struct perf_event *))
2079 {
2080         struct perf_event *child;
2081
2082         WARN_ON_ONCE(event->ctx->parent_ctx);
2083         mutex_lock(&event->child_mutex);
2084         func(event);
2085         list_for_each_entry(child, &event->child_list, child_list)
2086                 func(child);
2087         mutex_unlock(&event->child_mutex);
2088 }
2089
2090 static void perf_event_for_each(struct perf_event *event,
2091                                   void (*func)(struct perf_event *))
2092 {
2093         struct perf_event_context *ctx = event->ctx;
2094         struct perf_event *sibling;
2095
2096         WARN_ON_ONCE(ctx->parent_ctx);
2097         mutex_lock(&ctx->mutex);
2098         event = event->group_leader;
2099
2100         perf_event_for_each_child(event, func);
2101         func(event);
2102         list_for_each_entry(sibling, &event->sibling_list, group_entry)
2103                 perf_event_for_each_child(event, func);
2104         mutex_unlock(&ctx->mutex);
2105 }
2106
2107 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2108 {
2109         struct perf_event_context *ctx = event->ctx;
2110         unsigned long size;
2111         int ret = 0;
2112         u64 value;
2113
2114         if (!event->attr.sample_period)
2115                 return -EINVAL;
2116
2117         size = copy_from_user(&value, arg, sizeof(value));
2118         if (size != sizeof(value))
2119                 return -EFAULT;
2120
2121         if (!value)
2122                 return -EINVAL;
2123
2124         raw_spin_lock_irq(&ctx->lock);
2125         if (event->attr.freq) {
2126                 if (value > sysctl_perf_event_sample_rate) {
2127                         ret = -EINVAL;
2128                         goto unlock;
2129                 }
2130
2131                 event->attr.sample_freq = value;
2132         } else {
2133                 event->attr.sample_period = value;
2134                 event->hw.sample_period = value;
2135         }
2136 unlock:
2137         raw_spin_unlock_irq(&ctx->lock);
2138
2139         return ret;
2140 }
2141
2142 static int perf_event_set_output(struct perf_event *event, int output_fd);
2143 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2144
2145 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2146 {
2147         struct perf_event *event = file->private_data;
2148         void (*func)(struct perf_event *);
2149         u32 flags = arg;
2150
2151         switch (cmd) {
2152         case PERF_EVENT_IOC_ENABLE:
2153                 func = perf_event_enable;
2154                 break;
2155         case PERF_EVENT_IOC_DISABLE:
2156                 func = perf_event_disable;
2157                 break;
2158         case PERF_EVENT_IOC_RESET:
2159                 func = perf_event_reset;
2160                 break;
2161
2162         case PERF_EVENT_IOC_REFRESH:
2163                 return perf_event_refresh(event, arg);
2164
2165         case PERF_EVENT_IOC_PERIOD:
2166                 return perf_event_period(event, (u64 __user *)arg);
2167
2168         case PERF_EVENT_IOC_SET_OUTPUT:
2169                 return perf_event_set_output(event, arg);
2170
2171         case PERF_EVENT_IOC_SET_FILTER:
2172                 return perf_event_set_filter(event, (void __user *)arg);
2173
2174         default:
2175                 return -ENOTTY;
2176         }
2177
2178         if (flags & PERF_IOC_FLAG_GROUP)
2179                 perf_event_for_each(event, func);
2180         else
2181                 perf_event_for_each_child(event, func);
2182
2183         return 0;
2184 }
2185
2186 int perf_event_task_enable(void)
2187 {
2188         struct perf_event *event;
2189
2190         mutex_lock(&current->perf_event_mutex);
2191         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2192                 perf_event_for_each_child(event, perf_event_enable);
2193         mutex_unlock(&current->perf_event_mutex);
2194
2195         return 0;
2196 }
2197
2198 int perf_event_task_disable(void)
2199 {
2200         struct perf_event *event;
2201
2202         mutex_lock(&current->perf_event_mutex);
2203         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2204                 perf_event_for_each_child(event, perf_event_disable);
2205         mutex_unlock(&current->perf_event_mutex);
2206
2207         return 0;
2208 }
2209
2210 #ifndef PERF_EVENT_INDEX_OFFSET
2211 # define PERF_EVENT_INDEX_OFFSET 0
2212 #endif
2213
2214 static int perf_event_index(struct perf_event *event)
2215 {
2216         if (event->state != PERF_EVENT_STATE_ACTIVE)
2217                 return 0;
2218
2219         return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2220 }
2221
2222 /*
2223  * Callers need to ensure there can be no nesting of this function, otherwise
2224  * the seqlock logic goes bad. We can not serialize this because the arch
2225  * code calls this from NMI context.
2226  */
2227 void perf_event_update_userpage(struct perf_event *event)
2228 {
2229         struct perf_event_mmap_page *userpg;
2230         struct perf_mmap_data *data;
2231
2232         rcu_read_lock();
2233         data = rcu_dereference(event->data);
2234         if (!data)
2235                 goto unlock;
2236
2237         userpg = data->user_page;
2238
2239         /*
2240          * Disable preemption so as to not let the corresponding user-space
2241          * spin too long if we get preempted.
2242          */
2243         preempt_disable();
2244         ++userpg->lock;
2245         barrier();
2246         userpg->index = perf_event_index(event);
2247         userpg->offset = atomic64_read(&event->count);
2248         if (event->state == PERF_EVENT_STATE_ACTIVE)
2249                 userpg->offset -= atomic64_read(&event->hw.prev_count);
2250
2251         userpg->time_enabled = event->total_time_enabled +
2252                         atomic64_read(&event->child_total_time_enabled);
2253
2254         userpg->time_running = event->total_time_running +
2255                         atomic64_read(&event->child_total_time_running);
2256
2257         barrier();
2258         ++userpg->lock;
2259         preempt_enable();
2260 unlock:
2261         rcu_read_unlock();
2262 }
2263
2264 static unsigned long perf_data_size(struct perf_mmap_data *data)
2265 {
2266         return data->nr_pages << (PAGE_SHIFT + data->data_order);
2267 }
2268
2269 #ifndef CONFIG_PERF_USE_VMALLOC
2270
2271 /*
2272  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2273  */
2274
2275 static struct page *
2276 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2277 {
2278         if (pgoff > data->nr_pages)
2279                 return NULL;
2280
2281         if (pgoff == 0)
2282                 return virt_to_page(data->user_page);
2283
2284         return virt_to_page(data->data_pages[pgoff - 1]);
2285 }
2286
2287 static struct perf_mmap_data *
2288 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2289 {
2290         struct perf_mmap_data *data;
2291         unsigned long size;
2292         int i;
2293
2294         WARN_ON(atomic_read(&event->mmap_count));
2295
2296         size = sizeof(struct perf_mmap_data);
2297         size += nr_pages * sizeof(void *);
2298
2299         data = kzalloc(size, GFP_KERNEL);
2300         if (!data)
2301                 goto fail;
2302
2303         data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2304         if (!data->user_page)
2305                 goto fail_user_page;
2306
2307         for (i = 0; i < nr_pages; i++) {
2308                 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2309                 if (!data->data_pages[i])
2310                         goto fail_data_pages;
2311         }
2312
2313         data->data_order = 0;
2314         data->nr_pages = nr_pages;
2315
2316         return data;
2317
2318 fail_data_pages:
2319         for (i--; i >= 0; i--)
2320                 free_page((unsigned long)data->data_pages[i]);
2321
2322         free_page((unsigned long)data->user_page);
2323
2324 fail_user_page:
2325         kfree(data);
2326
2327 fail:
2328         return NULL;
2329 }
2330
2331 static void perf_mmap_free_page(unsigned long addr)
2332 {
2333         struct page *page = virt_to_page((void *)addr);
2334
2335         page->mapping = NULL;
2336         __free_page(page);
2337 }
2338
2339 static void perf_mmap_data_free(struct perf_mmap_data *data)
2340 {
2341         int i;
2342
2343         perf_mmap_free_page((unsigned long)data->user_page);
2344         for (i = 0; i < data->nr_pages; i++)
2345                 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2346         kfree(data);
2347 }
2348
2349 #else
2350
2351 /*
2352  * Back perf_mmap() with vmalloc memory.
2353  *
2354  * Required for architectures that have d-cache aliasing issues.
2355  */
2356
2357 static struct page *
2358 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2359 {
2360         if (pgoff > (1UL << data->data_order))
2361                 return NULL;
2362
2363         return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2364 }
2365
2366 static void perf_mmap_unmark_page(void *addr)
2367 {
2368         struct page *page = vmalloc_to_page(addr);
2369
2370         page->mapping = NULL;
2371 }
2372
2373 static void perf_mmap_data_free_work(struct work_struct *work)
2374 {
2375         struct perf_mmap_data *data;
2376         void *base;
2377         int i, nr;
2378
2379         data = container_of(work, struct perf_mmap_data, work);
2380         nr = 1 << data->data_order;
2381
2382         base = data->user_page;
2383         for (i = 0; i < nr + 1; i++)
2384                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2385
2386         vfree(base);
2387         kfree(data);
2388 }
2389
2390 static void perf_mmap_data_free(struct perf_mmap_data *data)
2391 {
2392         schedule_work(&data->work);
2393 }
2394
2395 static struct perf_mmap_data *
2396 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2397 {
2398         struct perf_mmap_data *data;
2399         unsigned long size;
2400         void *all_buf;
2401
2402         WARN_ON(atomic_read(&event->mmap_count));
2403
2404         size = sizeof(struct perf_mmap_data);
2405         size += sizeof(void *);
2406
2407         data = kzalloc(size, GFP_KERNEL);
2408         if (!data)
2409                 goto fail;
2410
2411         INIT_WORK(&data->work, perf_mmap_data_free_work);
2412
2413         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2414         if (!all_buf)
2415                 goto fail_all_buf;
2416
2417         data->user_page = all_buf;
2418         data->data_pages[0] = all_buf + PAGE_SIZE;
2419         data->data_order = ilog2(nr_pages);
2420         data->nr_pages = 1;
2421
2422         return data;
2423
2424 fail_all_buf:
2425         kfree(data);
2426
2427 fail:
2428         return NULL;
2429 }
2430
2431 #endif
2432
2433 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2434 {
2435         struct perf_event *event = vma->vm_file->private_data;
2436         struct perf_mmap_data *data;
2437         int ret = VM_FAULT_SIGBUS;
2438
2439         if (vmf->flags & FAULT_FLAG_MKWRITE) {
2440                 if (vmf->pgoff == 0)
2441                         ret = 0;
2442                 return ret;
2443         }
2444
2445         rcu_read_lock();
2446         data = rcu_dereference(event->data);
2447         if (!data)
2448                 goto unlock;
2449
2450         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2451                 goto unlock;
2452
2453         vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2454         if (!vmf->page)
2455                 goto unlock;
2456
2457         get_page(vmf->page);
2458         vmf->page->mapping = vma->vm_file->f_mapping;
2459         vmf->page->index   = vmf->pgoff;
2460
2461         ret = 0;
2462 unlock:
2463         rcu_read_unlock();
2464
2465         return ret;
2466 }
2467
2468 static void
2469 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2470 {
2471         long max_size = perf_data_size(data);
2472
2473         atomic_set(&data->lock, -1);
2474
2475         if (event->attr.watermark) {
2476                 data->watermark = min_t(long, max_size,
2477                                         event->attr.wakeup_watermark);
2478         }
2479
2480         if (!data->watermark)
2481                 data->watermark = max_size / 2;
2482
2483
2484         rcu_assign_pointer(event->data, data);
2485 }
2486
2487 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2488 {
2489         struct perf_mmap_data *data;
2490
2491         data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2492         perf_mmap_data_free(data);
2493 }
2494
2495 static void perf_mmap_data_release(struct perf_event *event)
2496 {
2497         struct perf_mmap_data *data = event->data;
2498
2499         WARN_ON(atomic_read(&event->mmap_count));
2500
2501         rcu_assign_pointer(event->data, NULL);
2502         call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2503 }
2504
2505 static void perf_mmap_open(struct vm_area_struct *vma)
2506 {
2507         struct perf_event *event = vma->vm_file->private_data;
2508
2509         atomic_inc(&event->mmap_count);
2510 }
2511
2512 static void perf_mmap_close(struct vm_area_struct *vma)
2513 {
2514         struct perf_event *event = vma->vm_file->private_data;
2515
2516         WARN_ON_ONCE(event->ctx->parent_ctx);
2517         if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2518                 unsigned long size = perf_data_size(event->data);
2519                 struct user_struct *user = current_user();
2520
2521                 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2522                 vma->vm_mm->locked_vm -= event->data->nr_locked;
2523                 perf_mmap_data_release(event);
2524                 mutex_unlock(&event->mmap_mutex);
2525         }
2526 }
2527
2528 static const struct vm_operations_struct perf_mmap_vmops = {
2529         .open           = perf_mmap_open,
2530         .close          = perf_mmap_close,
2531         .fault          = perf_mmap_fault,
2532         .page_mkwrite   = perf_mmap_fault,
2533 };
2534
2535 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2536 {
2537         struct perf_event *event = file->private_data;
2538         unsigned long user_locked, user_lock_limit;
2539         struct user_struct *user = current_user();
2540         unsigned long locked, lock_limit;
2541         struct perf_mmap_data *data;
2542         unsigned long vma_size;
2543         unsigned long nr_pages;
2544         long user_extra, extra;
2545         int ret = 0;
2546
2547         if (!(vma->vm_flags & VM_SHARED))
2548                 return -EINVAL;
2549
2550         vma_size = vma->vm_end - vma->vm_start;
2551         nr_pages = (vma_size / PAGE_SIZE) - 1;
2552
2553         /*
2554          * If we have data pages ensure they're a power-of-two number, so we
2555          * can do bitmasks instead of modulo.
2556          */
2557         if (nr_pages != 0 && !is_power_of_2(nr_pages))
2558                 return -EINVAL;
2559
2560         if (vma_size != PAGE_SIZE * (1 + nr_pages))
2561                 return -EINVAL;
2562
2563         if (vma->vm_pgoff != 0)
2564                 return -EINVAL;
2565
2566         WARN_ON_ONCE(event->ctx->parent_ctx);
2567         mutex_lock(&event->mmap_mutex);
2568         if (event->output) {
2569                 ret = -EINVAL;
2570                 goto unlock;
2571         }
2572
2573         if (atomic_inc_not_zero(&event->mmap_count)) {
2574                 if (nr_pages != event->data->nr_pages)
2575                         ret = -EINVAL;
2576                 goto unlock;
2577         }
2578
2579         user_extra = nr_pages + 1;
2580         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2581
2582         /*
2583          * Increase the limit linearly with more CPUs:
2584          */
2585         user_lock_limit *= num_online_cpus();
2586
2587         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2588
2589         extra = 0;
2590         if (user_locked > user_lock_limit)
2591                 extra = user_locked - user_lock_limit;
2592
2593         lock_limit = rlimit(RLIMIT_MEMLOCK);
2594         lock_limit >>= PAGE_SHIFT;
2595         locked = vma->vm_mm->locked_vm + extra;
2596
2597         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2598                 !capable(CAP_IPC_LOCK)) {
2599                 ret = -EPERM;
2600                 goto unlock;
2601         }
2602
2603         WARN_ON(event->data);
2604
2605         data = perf_mmap_data_alloc(event, nr_pages);
2606         ret = -ENOMEM;
2607         if (!data)
2608                 goto unlock;
2609
2610         ret = 0;
2611         perf_mmap_data_init(event, data);
2612
2613         atomic_set(&event->mmap_count, 1);
2614         atomic_long_add(user_extra, &user->locked_vm);
2615         vma->vm_mm->locked_vm += extra;
2616         event->data->nr_locked = extra;
2617         if (vma->vm_flags & VM_WRITE)
2618                 event->data->writable = 1;
2619
2620 unlock:
2621         mutex_unlock(&event->mmap_mutex);
2622
2623         vma->vm_flags |= VM_RESERVED;
2624         vma->vm_ops = &perf_mmap_vmops;
2625
2626         return ret;
2627 }
2628
2629 static int perf_fasync(int fd, struct file *filp, int on)
2630 {
2631         struct inode *inode = filp->f_path.dentry->d_inode;
2632         struct perf_event *event = filp->private_data;
2633         int retval;
2634
2635         mutex_lock(&inode->i_mutex);
2636         retval = fasync_helper(fd, filp, on, &event->fasync);
2637         mutex_unlock(&inode->i_mutex);
2638
2639         if (retval < 0)
2640                 return retval;
2641
2642         return 0;
2643 }
2644
2645 static const struct file_operations perf_fops = {
2646         .release                = perf_release,
2647         .read                   = perf_read,
2648         .poll                   = perf_poll,
2649         .unlocked_ioctl         = perf_ioctl,
2650         .compat_ioctl           = perf_ioctl,
2651         .mmap                   = perf_mmap,
2652         .fasync                 = perf_fasync,
2653 };
2654
2655 /*
2656  * Perf event wakeup
2657  *
2658  * If there's data, ensure we set the poll() state and publish everything
2659  * to user-space before waking everybody up.
2660  */
2661
2662 void perf_event_wakeup(struct perf_event *event)
2663 {
2664         wake_up_all(&event->waitq);
2665
2666         if (event->pending_kill) {
2667                 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2668                 event->pending_kill = 0;
2669         }
2670 }
2671
2672 /*
2673  * Pending wakeups
2674  *
2675  * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2676  *
2677  * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2678  * single linked list and use cmpxchg() to add entries lockless.
2679  */
2680
2681 static void perf_pending_event(struct perf_pending_entry *entry)
2682 {
2683         struct perf_event *event = container_of(entry,
2684                         struct perf_event, pending);
2685
2686         if (event->pending_disable) {
2687                 event->pending_disable = 0;
2688                 __perf_event_disable(event);
2689         }
2690
2691         if (event->pending_wakeup) {
2692                 event->pending_wakeup = 0;
2693                 perf_event_wakeup(event);
2694         }
2695 }
2696
2697 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2698
2699 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2700         PENDING_TAIL,
2701 };
2702
2703 static void perf_pending_queue(struct perf_pending_entry *entry,
2704                                void (*func)(struct perf_pending_entry *))
2705 {
2706         struct perf_pending_entry **head;
2707
2708         if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2709                 return;
2710
2711         entry->func = func;
2712
2713         head = &get_cpu_var(perf_pending_head);
2714
2715         do {
2716                 entry->next = *head;
2717         } while (cmpxchg(head, entry->next, entry) != entry->next);
2718
2719         set_perf_event_pending();
2720
2721         put_cpu_var(perf_pending_head);
2722 }
2723
2724 static int __perf_pending_run(void)
2725 {
2726         struct perf_pending_entry *list;
2727         int nr = 0;
2728
2729         list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2730         while (list != PENDING_TAIL) {
2731                 void (*func)(struct perf_pending_entry *);
2732                 struct perf_pending_entry *entry = list;
2733
2734                 list = list->next;
2735
2736                 func = entry->func;
2737                 entry->next = NULL;
2738                 /*
2739                  * Ensure we observe the unqueue before we issue the wakeup,
2740                  * so that we won't be waiting forever.
2741                  * -- see perf_not_pending().
2742                  */
2743                 smp_wmb();
2744
2745                 func(entry);
2746                 nr++;
2747         }
2748
2749         return nr;
2750 }
2751
2752 static inline int perf_not_pending(struct perf_event *event)
2753 {
2754         /*
2755          * If we flush on whatever cpu we run, there is a chance we don't
2756          * need to wait.
2757          */
2758         get_cpu();
2759         __perf_pending_run();
2760         put_cpu();
2761
2762         /*
2763          * Ensure we see the proper queue state before going to sleep
2764          * so that we do not miss the wakeup. -- see perf_pending_handle()
2765          */
2766         smp_rmb();
2767         return event->pending.next == NULL;
2768 }
2769
2770 static void perf_pending_sync(struct perf_event *event)
2771 {
2772         wait_event(event->waitq, perf_not_pending(event));
2773 }
2774
2775 void perf_event_do_pending(void)
2776 {
2777         __perf_pending_run();
2778 }
2779
2780 /*
2781  * Callchain support -- arch specific
2782  */
2783
2784 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2785 {
2786         return NULL;
2787 }
2788
2789 #ifdef CONFIG_EVENT_TRACING
2790 __weak
2791 void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip, int skip)
2792 {
2793 }
2794 #endif
2795
2796 /*
2797  * Output
2798  */
2799 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2800                               unsigned long offset, unsigned long head)
2801 {
2802         unsigned long mask;
2803
2804         if (!data->writable)
2805                 return true;
2806
2807         mask = perf_data_size(data) - 1;
2808
2809         offset = (offset - tail) & mask;
2810         head   = (head   - tail) & mask;
2811
2812         if ((int)(head - offset) < 0)
2813                 return false;
2814
2815         return true;
2816 }
2817
2818 static void perf_output_wakeup(struct perf_output_handle *handle)
2819 {
2820         atomic_set(&handle->data->poll, POLL_IN);
2821
2822         if (handle->nmi) {
2823                 handle->event->pending_wakeup = 1;
2824                 perf_pending_queue(&handle->event->pending,
2825                                    perf_pending_event);
2826         } else
2827                 perf_event_wakeup(handle->event);
2828 }
2829
2830 /*
2831  * Curious locking construct.
2832  *
2833  * We need to ensure a later event_id doesn't publish a head when a former
2834  * event_id isn't done writing. However since we need to deal with NMIs we
2835  * cannot fully serialize things.
2836  *
2837  * What we do is serialize between CPUs so we only have to deal with NMI
2838  * nesting on a single CPU.
2839  *
2840  * We only publish the head (and generate a wakeup) when the outer-most
2841  * event_id completes.
2842  */
2843 static void perf_output_lock(struct perf_output_handle *handle)
2844 {
2845         struct perf_mmap_data *data = handle->data;
2846         int cur, cpu = get_cpu();
2847
2848         handle->locked = 0;
2849
2850         for (;;) {
2851                 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2852                 if (cur == -1) {
2853                         handle->locked = 1;
2854                         break;
2855                 }
2856                 if (cur == cpu)
2857                         break;
2858
2859                 cpu_relax();
2860         }
2861 }
2862
2863 static void perf_output_unlock(struct perf_output_handle *handle)
2864 {
2865         struct perf_mmap_data *data = handle->data;
2866         unsigned long head;
2867         int cpu;
2868
2869         data->done_head = data->head;
2870
2871         if (!handle->locked)
2872                 goto out;
2873
2874 again:
2875         /*
2876          * The xchg implies a full barrier that ensures all writes are done
2877          * before we publish the new head, matched by a rmb() in userspace when
2878          * reading this position.
2879          */
2880         while ((head = atomic_long_xchg(&data->done_head, 0)))
2881                 data->user_page->data_head = head;
2882
2883         /*
2884          * NMI can happen here, which means we can miss a done_head update.
2885          */
2886
2887         cpu = atomic_xchg(&data->lock, -1);
2888         WARN_ON_ONCE(cpu != smp_processor_id());
2889
2890         /*
2891          * Therefore we have to validate we did not indeed do so.
2892          */
2893         if (unlikely(atomic_long_read(&data->done_head))) {
2894                 /*
2895                  * Since we had it locked, we can lock it again.
2896                  */
2897                 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2898                         cpu_relax();
2899
2900                 goto again;
2901         }
2902
2903         if (atomic_xchg(&data->wakeup, 0))
2904                 perf_output_wakeup(handle);
2905 out:
2906         put_cpu();
2907 }
2908
2909 void perf_output_copy(struct perf_output_handle *handle,
2910                       const void *buf, unsigned int len)
2911 {
2912         unsigned int pages_mask;
2913         unsigned long offset;
2914         unsigned int size;
2915         void **pages;
2916
2917         offset          = handle->offset;
2918         pages_mask      = handle->data->nr_pages - 1;
2919         pages           = handle->data->data_pages;
2920
2921         do {
2922                 unsigned long page_offset;
2923                 unsigned long page_size;
2924                 int nr;
2925
2926                 nr          = (offset >> PAGE_SHIFT) & pages_mask;
2927                 page_size   = 1UL << (handle->data->data_order + PAGE_SHIFT);
2928                 page_offset = offset & (page_size - 1);
2929                 size        = min_t(unsigned int, page_size - page_offset, len);
2930
2931                 memcpy(pages[nr] + page_offset, buf, size);
2932
2933                 len         -= size;
2934                 buf         += size;
2935                 offset      += size;
2936         } while (len);
2937
2938         handle->offset = offset;
2939
2940         /*
2941          * Check we didn't copy past our reservation window, taking the
2942          * possible unsigned int wrap into account.
2943          */
2944         WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2945 }
2946
2947 int perf_output_begin(struct perf_output_handle *handle,
2948                       struct perf_event *event, unsigned int size,
2949                       int nmi, int sample)
2950 {
2951         struct perf_event *output_event;
2952         struct perf_mmap_data *data;
2953         unsigned long tail, offset, head;
2954         int have_lost;
2955         struct {
2956                 struct perf_event_header header;
2957                 u64                      id;
2958                 u64                      lost;
2959         } lost_event;
2960
2961         rcu_read_lock();
2962         /*
2963          * For inherited events we send all the output towards the parent.
2964          */
2965         if (event->parent)
2966                 event = event->parent;
2967
2968         output_event = rcu_dereference(event->output);
2969         if (output_event)
2970                 event = output_event;
2971
2972         data = rcu_dereference(event->data);
2973         if (!data)
2974                 goto out;
2975
2976         handle->data    = data;
2977         handle->event   = event;
2978         handle->nmi     = nmi;
2979         handle->sample  = sample;
2980
2981         if (!data->nr_pages)
2982                 goto fail;
2983
2984         have_lost = atomic_read(&data->lost);
2985         if (have_lost)
2986                 size += sizeof(lost_event);
2987
2988         perf_output_lock(handle);
2989
2990         do {
2991                 /*
2992                  * Userspace could choose to issue a mb() before updating the
2993                  * tail pointer. So that all reads will be completed before the
2994                  * write is issued.
2995                  */
2996                 tail = ACCESS_ONCE(data->user_page->data_tail);
2997                 smp_rmb();
2998                 offset = head = atomic_long_read(&data->head);
2999                 head += size;
3000                 if (unlikely(!perf_output_space(data, tail, offset, head)))
3001                         goto fail;
3002         } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
3003
3004         handle->offset  = offset;
3005         handle->head    = head;
3006
3007         if (head - tail > data->watermark)
3008                 atomic_set(&data->wakeup, 1);
3009
3010         if (have_lost) {
3011                 lost_event.header.type = PERF_RECORD_LOST;
3012                 lost_event.header.misc = 0;
3013                 lost_event.header.size = sizeof(lost_event);
3014                 lost_event.id          = event->id;
3015                 lost_event.lost        = atomic_xchg(&data->lost, 0);
3016
3017                 perf_output_put(handle, lost_event);
3018         }
3019
3020         return 0;
3021
3022 fail:
3023         atomic_inc(&data->lost);
3024         perf_output_unlock(handle);
3025 out:
3026         rcu_read_unlock();
3027
3028         return -ENOSPC;
3029 }
3030
3031 void perf_output_end(struct perf_output_handle *handle)
3032 {
3033         struct perf_event *event = handle->event;
3034         struct perf_mmap_data *data = handle->data;
3035
3036         int wakeup_events = event->attr.wakeup_events;
3037
3038         if (handle->sample && wakeup_events) {
3039                 int events = atomic_inc_return(&data->events);
3040                 if (events >= wakeup_events) {
3041                         atomic_sub(wakeup_events, &data->events);
3042                         atomic_set(&data->wakeup, 1);
3043                 }
3044         }
3045
3046         perf_output_unlock(handle);
3047         rcu_read_unlock();
3048 }
3049
3050 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3051 {
3052         /*
3053          * only top level events have the pid namespace they were created in
3054          */
3055         if (event->parent)
3056                 event = event->parent;
3057
3058         return task_tgid_nr_ns(p, event->ns);
3059 }
3060
3061 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3062 {
3063         /*
3064          * only top level events have the pid namespace they were created in
3065          */
3066         if (event->parent)
3067                 event = event->parent;
3068
3069         return task_pid_nr_ns(p, event->ns);
3070 }
3071
3072 static void perf_output_read_one(struct perf_output_handle *handle,
3073                                  struct perf_event *event)
3074 {
3075         u64 read_format = event->attr.read_format;
3076         u64 values[4];
3077         int n = 0;
3078
3079         values[n++] = atomic64_read(&event->count);
3080         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3081                 values[n++] = event->total_time_enabled +
3082                         atomic64_read(&event->child_total_time_enabled);
3083         }
3084         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3085                 values[n++] = event->total_time_running +
3086                         atomic64_read(&event->child_total_time_running);
3087         }
3088         if (read_format & PERF_FORMAT_ID)
3089                 values[n++] = primary_event_id(event);
3090
3091         perf_output_copy(handle, values, n * sizeof(u64));
3092 }
3093
3094 /*
3095  * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3096  */
3097 static void perf_output_read_group(struct perf_output_handle *handle,
3098                             struct perf_event *event)
3099 {
3100         struct perf_event *leader = event->group_leader, *sub;
3101         u64 read_format = event->attr.read_format;
3102         u64 values[5];
3103         int n = 0;
3104
3105         values[n++] = 1 + leader->nr_siblings;
3106
3107         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3108                 values[n++] = leader->total_time_enabled;
3109
3110         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3111                 values[n++] = leader->total_time_running;
3112
3113         if (leader != event)
3114                 leader->pmu->read(leader);
3115
3116         values[n++] = atomic64_read(&leader->count);
3117         if (read_format & PERF_FORMAT_ID)
3118                 values[n++] = primary_event_id(leader);
3119
3120         perf_output_copy(handle, values, n * sizeof(u64));
3121
3122         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3123                 n = 0;
3124
3125                 if (sub != event)
3126                         sub->pmu->read(sub);
3127
3128                 values[n++] = atomic64_read(&sub->count);
3129                 if (read_format & PERF_FORMAT_ID)
3130                         values[n++] = primary_event_id(sub);
3131
3132                 perf_output_copy(handle, values, n * sizeof(u64));
3133         }
3134 }
3135
3136 static void perf_output_read(struct perf_output_handle *handle,
3137                              struct perf_event *event)
3138 {
3139         if (event->attr.read_format & PERF_FORMAT_GROUP)
3140                 perf_output_read_group(handle, event);
3141         else
3142                 perf_output_read_one(handle, event);
3143 }
3144
3145 void perf_output_sample(struct perf_output_handle *handle,
3146                         struct perf_event_header *header,
3147                         struct perf_sample_data *data,
3148                         struct perf_event *event)
3149 {
3150         u64 sample_type = data->type;
3151
3152         perf_output_put(handle, *header);
3153
3154         if (sample_type & PERF_SAMPLE_IP)
3155                 perf_output_put(handle, data->ip);
3156
3157         if (sample_type & PERF_SAMPLE_TID)
3158                 perf_output_put(handle, data->tid_entry);
3159
3160         if (sample_type & PERF_SAMPLE_TIME)
3161                 perf_output_put(handle, data->time);
3162
3163         if (sample_type & PERF_SAMPLE_ADDR)
3164                 perf_output_put(handle, data->addr);
3165
3166         if (sample_type & PERF_SAMPLE_ID)
3167                 perf_output_put(handle, data->id);
3168
3169         if (sample_type & PERF_SAMPLE_STREAM_ID)
3170                 perf_output_put(handle, data->stream_id);
3171
3172         if (sample_type & PERF_SAMPLE_CPU)
3173                 perf_output_put(handle, data->cpu_entry);
3174
3175         if (sample_type & PERF_SAMPLE_PERIOD)
3176                 perf_output_put(handle, data->period);
3177
3178         if (sample_type & PERF_SAMPLE_READ)
3179                 perf_output_read(handle, event);
3180
3181         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3182                 if (data->callchain) {
3183                         int size = 1;
3184
3185                         if (data->callchain)
3186                                 size += data->callchain->nr;
3187
3188                         size *= sizeof(u64);
3189
3190                         perf_output_copy(handle, data->callchain, size);
3191                 } else {
3192                         u64 nr = 0;
3193                         perf_output_put(handle, nr);
3194                 }
3195         }
3196
3197         if (sample_type & PERF_SAMPLE_RAW) {
3198                 if (data->raw) {
3199                         perf_output_put(handle, data->raw->size);
3200                         perf_output_copy(handle, data->raw->data,
3201                                          data->raw->size);
3202                 } else {
3203                         struct {
3204                                 u32     size;
3205                                 u32     data;
3206                         } raw = {
3207                                 .size = sizeof(u32),
3208                                 .data = 0,
3209                         };
3210                         perf_output_put(handle, raw);
3211                 }
3212         }
3213 }
3214
3215 void perf_prepare_sample(struct perf_event_header *header,
3216                          struct perf_sample_data *data,
3217                          struct perf_event *event,
3218                          struct pt_regs *regs)
3219 {
3220         u64 sample_type = event->attr.sample_type;
3221
3222         data->type = sample_type;
3223
3224         header->type = PERF_RECORD_SAMPLE;
3225         header->size = sizeof(*header);
3226
3227         header->misc = 0;
3228         header->misc |= perf_misc_flags(regs);
3229
3230         if (sample_type & PERF_SAMPLE_IP) {
3231                 data->ip = perf_instruction_pointer(regs);
3232
3233                 header->size += sizeof(data->ip);
3234         }
3235
3236         if (sample_type & PERF_SAMPLE_TID) {
3237                 /* namespace issues */
3238                 data->tid_entry.pid = perf_event_pid(event, current);
3239                 data->tid_entry.tid = perf_event_tid(event, current);
3240
3241                 header->size += sizeof(data->tid_entry);
3242         }
3243
3244         if (sample_type & PERF_SAMPLE_TIME) {
3245                 data->time = perf_clock();
3246
3247                 header->size += sizeof(data->time);
3248         }
3249
3250         if (sample_type & PERF_SAMPLE_ADDR)
3251                 header->size += sizeof(data->addr);
3252
3253         if (sample_type & PERF_SAMPLE_ID) {
3254                 data->id = primary_event_id(event);
3255
3256                 header->size += sizeof(data->id);
3257         }
3258
3259         if (sample_type & PERF_SAMPLE_STREAM_ID) {
3260                 data->stream_id = event->id;
3261
3262                 header->size += sizeof(data->stream_id);
3263         }
3264
3265         if (sample_type & PERF_SAMPLE_CPU) {
3266                 data->cpu_entry.cpu             = raw_smp_processor_id();
3267                 data->cpu_entry.reserved        = 0;
3268
3269                 header->size += sizeof(data->cpu_entry);
3270         }
3271
3272         if (sample_type & PERF_SAMPLE_PERIOD)
3273                 header->size += sizeof(data->period);
3274
3275         if (sample_type & PERF_SAMPLE_READ)
3276                 header->size += perf_event_read_size(event);
3277
3278         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3279                 int size = 1;
3280
3281                 data->callchain = perf_callchain(regs);
3282
3283                 if (data->callchain)
3284                         size += data->callchain->nr;
3285
3286                 header->size += size * sizeof(u64);
3287         }
3288
3289         if (sample_type & PERF_SAMPLE_RAW) {
3290                 int size = sizeof(u32);
3291
3292                 if (data->raw)
3293                         size += data->raw->size;
3294                 else
3295                         size += sizeof(u32);
3296
3297                 WARN_ON_ONCE(size & (sizeof(u64)-1));
3298                 header->size += size;
3299         }
3300 }
3301
3302 static void perf_event_output(struct perf_event *event, int nmi,
3303                                 struct perf_sample_data *data,
3304                                 struct pt_regs *regs)
3305 {
3306         struct perf_output_handle handle;
3307         struct perf_event_header header;
3308
3309         perf_prepare_sample(&header, data, event, regs);
3310
3311         if (perf_output_begin(&handle, event, header.size, nmi, 1))
3312                 return;
3313
3314         perf_output_sample(&handle, &header, data, event);
3315
3316         perf_output_end(&handle);
3317 }
3318
3319 /*
3320  * read event_id
3321  */
3322
3323 struct perf_read_event {
3324         struct perf_event_header        header;
3325
3326         u32                             pid;
3327         u32                             tid;
3328 };
3329
3330 static void
3331 perf_event_read_event(struct perf_event *event,
3332                         struct task_struct *task)
3333 {
3334         struct perf_output_handle handle;
3335         struct perf_read_event read_event = {
3336                 .header = {
3337                         .type = PERF_RECORD_READ,
3338                         .misc = 0,
3339                         .size = sizeof(read_event) + perf_event_read_size(event),
3340                 },
3341                 .pid = perf_event_pid(event, task),
3342                 .tid = perf_event_tid(event, task),
3343         };
3344         int ret;
3345
3346         ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3347         if (ret)
3348                 return;
3349
3350         perf_output_put(&handle, read_event);
3351         perf_output_read(&handle, event);
3352
3353         perf_output_end(&handle);
3354 }
3355
3356 /*
3357  * task tracking -- fork/exit
3358  *
3359  * enabled by: attr.comm | attr.mmap | attr.task
3360  */
3361
3362 struct perf_task_event {
3363         struct task_struct              *task;
3364         struct perf_event_context       *task_ctx;
3365
3366         struct {
3367                 struct perf_event_header        header;
3368
3369                 u32                             pid;
3370                 u32                             ppid;
3371                 u32                             tid;
3372                 u32                             ptid;
3373                 u64                             time;
3374         } event_id;
3375 };
3376
3377 static void perf_event_task_output(struct perf_event *event,
3378                                      struct perf_task_event *task_event)
3379 {
3380         struct perf_output_handle handle;
3381         int size;
3382         struct task_struct *task = task_event->task;
3383         int ret;
3384
3385         size  = task_event->event_id.header.size;
3386         ret = perf_output_begin(&handle, event, size, 0, 0);
3387
3388         if (ret)
3389                 return;
3390
3391         task_event->event_id.pid = perf_event_pid(event, task);
3392         task_event->event_id.ppid = perf_event_pid(event, current);
3393
3394         task_event->event_id.tid = perf_event_tid(event, task);
3395         task_event->event_id.ptid = perf_event_tid(event, current);
3396
3397         perf_output_put(&handle, task_event->event_id);
3398
3399         perf_output_end(&handle);
3400 }
3401
3402 static int perf_event_task_match(struct perf_event *event)
3403 {
3404         if (event->state < PERF_EVENT_STATE_INACTIVE)
3405                 return 0;
3406
3407         if (event->cpu != -1 && event->cpu != smp_processor_id())
3408                 return 0;
3409
3410         if (event->attr.comm || event->attr.mmap || event->attr.task)
3411                 return 1;
3412
3413         return 0;
3414 }
3415
3416 static void perf_event_task_ctx(struct perf_event_context *ctx,
3417                                   struct perf_task_event *task_event)
3418 {
3419         struct perf_event *event;
3420
3421         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3422                 if (perf_event_task_match(event))
3423                         perf_event_task_output(event, task_event);
3424         }
3425 }
3426
3427 static void perf_event_task_event(struct perf_task_event *task_event)
3428 {
3429         struct perf_cpu_context *cpuctx;
3430         struct perf_event_context *ctx = task_event->task_ctx;
3431
3432         rcu_read_lock();
3433         cpuctx = &get_cpu_var(perf_cpu_context);
3434         perf_event_task_ctx(&cpuctx->ctx, task_event);
3435         if (!ctx)
3436                 ctx = rcu_dereference(current->perf_event_ctxp);
3437         if (ctx)
3438                 perf_event_task_ctx(ctx, task_event);
3439         put_cpu_var(perf_cpu_context);
3440         rcu_read_unlock();
3441 }
3442
3443 static void perf_event_task(struct task_struct *task,
3444                               struct perf_event_context *task_ctx,
3445                               int new)
3446 {
3447         struct perf_task_event task_event;
3448
3449         if (!atomic_read(&nr_comm_events) &&
3450             !atomic_read(&nr_mmap_events) &&
3451             !atomic_read(&nr_task_events))
3452                 return;
3453
3454         task_event = (struct perf_task_event){
3455                 .task     = task,
3456                 .task_ctx = task_ctx,
3457                 .event_id    = {
3458                         .header = {
3459                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3460                                 .misc = 0,
3461                                 .size = sizeof(task_event.event_id),
3462                         },
3463                         /* .pid  */
3464                         /* .ppid */
3465                         /* .tid  */
3466                         /* .ptid */
3467                         .time = perf_clock(),
3468                 },
3469         };
3470
3471         perf_event_task_event(&task_event);
3472 }
3473
3474 void perf_event_fork(struct task_struct *task)
3475 {
3476         perf_event_task(task, NULL, 1);
3477 }
3478
3479 /*
3480  * comm tracking
3481  */
3482
3483 struct perf_comm_event {
3484         struct task_struct      *task;
3485         char                    *comm;
3486         int                     comm_size;
3487
3488         struct {
3489                 struct perf_event_header        header;
3490
3491                 u32                             pid;
3492                 u32                             tid;
3493         } event_id;
3494 };
3495
3496 static void perf_event_comm_output(struct perf_event *event,
3497                                      struct perf_comm_event *comm_event)
3498 {
3499         struct perf_output_handle handle;
3500         int size = comm_event->event_id.header.size;
3501         int ret = perf_output_begin(&handle, event, size, 0, 0);
3502
3503         if (ret)
3504                 return;
3505
3506         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3507         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3508
3509         perf_output_put(&handle, comm_event->event_id);
3510         perf_output_copy(&handle, comm_event->comm,
3511                                    comm_event->comm_size);
3512         perf_output_end(&handle);
3513 }
3514
3515 static int perf_event_comm_match(struct perf_event *event)
3516 {
3517         if (event->state < PERF_EVENT_STATE_INACTIVE)
3518                 return 0;
3519
3520         if (event->cpu != -1 && event->cpu != smp_processor_id())
3521                 return 0;
3522
3523         if (event->attr.comm)
3524                 return 1;
3525
3526         return 0;
3527 }
3528
3529 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3530                                   struct perf_comm_event *comm_event)
3531 {
3532         struct perf_event *event;
3533
3534         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3535                 if (perf_event_comm_match(event))
3536                         perf_event_comm_output(event, comm_event);
3537         }
3538 }
3539
3540 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3541 {
3542         struct perf_cpu_context *cpuctx;
3543         struct perf_event_context *ctx;
3544         unsigned int size;
3545         char comm[TASK_COMM_LEN];
3546
3547         memset(comm, 0, sizeof(comm));
3548         strlcpy(comm, comm_event->task->comm, sizeof(comm));
3549         size = ALIGN(strlen(comm)+1, sizeof(u64));
3550
3551         comm_event->comm = comm;
3552         comm_event->comm_size = size;
3553
3554         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3555
3556         rcu_read_lock();
3557         cpuctx = &get_cpu_var(perf_cpu_context);
3558         perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3559         ctx = rcu_dereference(current->perf_event_ctxp);
3560         if (ctx)
3561                 perf_event_comm_ctx(ctx, comm_event);
3562         put_cpu_var(perf_cpu_context);
3563         rcu_read_unlock();
3564 }
3565
3566 void perf_event_comm(struct task_struct *task)
3567 {
3568         struct perf_comm_event comm_event;
3569
3570         if (task->perf_event_ctxp)
3571                 perf_event_enable_on_exec(task);
3572
3573         if (!atomic_read(&nr_comm_events))
3574                 return;
3575
3576         comm_event = (struct perf_comm_event){
3577                 .task   = task,
3578                 /* .comm      */
3579                 /* .comm_size */
3580                 .event_id  = {
3581                         .header = {
3582                                 .type = PERF_RECORD_COMM,
3583                                 .misc = 0,
3584                                 /* .size */
3585                         },
3586                         /* .pid */
3587                         /* .tid */
3588                 },
3589         };
3590
3591         perf_event_comm_event(&comm_event);
3592 }
3593
3594 /*
3595  * mmap tracking
3596  */
3597
3598 struct perf_mmap_event {
3599         struct vm_area_struct   *vma;
3600
3601         const char              *file_name;
3602         int                     file_size;
3603
3604         struct {
3605                 struct perf_event_header        header;
3606
3607                 u32                             pid;
3608                 u32                             tid;
3609                 u64                             start;
3610                 u64                             len;
3611                 u64                             pgoff;
3612         } event_id;
3613 };
3614
3615 static void perf_event_mmap_output(struct perf_event *event,
3616                                      struct perf_mmap_event *mmap_event)
3617 {
3618         struct perf_output_handle handle;
3619         int size = mmap_event->event_id.header.size;
3620         int ret = perf_output_begin(&handle, event, size, 0, 0);
3621
3622         if (ret)
3623                 return;
3624
3625         mmap_event->event_id.pid = perf_event_pid(event, current);
3626         mmap_event->event_id.tid = perf_event_tid(event, current);
3627
3628         perf_output_put(&handle, mmap_event->event_id);
3629         perf_output_copy(&handle, mmap_event->file_name,
3630                                    mmap_event->file_size);
3631         perf_output_end(&handle);
3632 }
3633
3634 static int perf_event_mmap_match(struct perf_event *event,
3635                                    struct perf_mmap_event *mmap_event)
3636 {
3637         if (event->state < PERF_EVENT_STATE_INACTIVE)
3638                 return 0;
3639
3640         if (event->cpu != -1 && event->cpu != smp_processor_id())
3641                 return 0;
3642
3643         if (event->attr.mmap)
3644                 return 1;
3645
3646         return 0;
3647 }
3648
3649 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3650                                   struct perf_mmap_event *mmap_event)
3651 {
3652         struct perf_event *event;
3653
3654         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3655                 if (perf_event_mmap_match(event, mmap_event))
3656                         perf_event_mmap_output(event, mmap_event);
3657         }
3658 }
3659
3660 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3661 {
3662         struct perf_cpu_context *cpuctx;
3663         struct perf_event_context *ctx;
3664         struct vm_area_struct *vma = mmap_event->vma;
3665         struct file *file = vma->vm_file;
3666         unsigned int size;
3667         char tmp[16];
3668         char *buf = NULL;
3669         const char *name;
3670
3671         memset(tmp, 0, sizeof(tmp));
3672
3673         if (file) {
3674                 /*
3675                  * d_path works from the end of the buffer backwards, so we
3676                  * need to add enough zero bytes after the string to handle
3677                  * the 64bit alignment we do later.
3678                  */
3679                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3680                 if (!buf) {
3681                         name = strncpy(tmp, "//enomem", sizeof(tmp));
3682                         goto got_name;
3683                 }
3684                 name = d_path(&file->f_path, buf, PATH_MAX);
3685                 if (IS_ERR(name)) {
3686                         name = strncpy(tmp, "//toolong", sizeof(tmp));
3687                         goto got_name;
3688                 }
3689         } else {
3690                 if (arch_vma_name(mmap_event->vma)) {
3691                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3692                                        sizeof(tmp));
3693                         goto got_name;
3694                 }
3695
3696                 if (!vma->vm_mm) {
3697                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
3698                         goto got_name;
3699                 }
3700
3701                 name = strncpy(tmp, "//anon", sizeof(tmp));
3702                 goto got_name;
3703         }
3704
3705 got_name:
3706         size = ALIGN(strlen(name)+1, sizeof(u64));
3707
3708         mmap_event->file_name = name;
3709         mmap_event->file_size = size;
3710
3711         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3712
3713         rcu_read_lock();
3714         cpuctx = &get_cpu_var(perf_cpu_context);
3715         perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3716         ctx = rcu_dereference(current->perf_event_ctxp);
3717         if (ctx)
3718                 perf_event_mmap_ctx(ctx, mmap_event);
3719         put_cpu_var(perf_cpu_context);
3720         rcu_read_unlock();
3721
3722         kfree(buf);
3723 }
3724
3725 void __perf_event_mmap(struct vm_area_struct *vma)
3726 {
3727         struct perf_mmap_event mmap_event;
3728
3729         if (!atomic_read(&nr_mmap_events))
3730                 return;
3731
3732         mmap_event = (struct perf_mmap_event){
3733                 .vma    = vma,
3734                 /* .file_name */
3735                 /* .file_size */
3736                 .event_id  = {
3737                         .header = {
3738                                 .type = PERF_RECORD_MMAP,
3739                                 .misc = 0,
3740                                 /* .size */
3741                         },
3742                         /* .pid */
3743                         /* .tid */
3744                         .start  = vma->vm_start,
3745                         .len    = vma->vm_end - vma->vm_start,
3746                         .pgoff  = (u64)vma->vm_pgoff << PAGE_SHIFT,
3747                 },
3748         };
3749
3750         perf_event_mmap_event(&mmap_event);
3751 }
3752
3753 /*
3754  * IRQ throttle logging
3755  */
3756
3757 static void perf_log_throttle(struct perf_event *event, int enable)
3758 {
3759         struct perf_output_handle handle;
3760         int ret;
3761
3762         struct {
3763                 struct perf_event_header        header;
3764                 u64                             time;
3765                 u64                             id;
3766                 u64                             stream_id;
3767         } throttle_event = {
3768                 .header = {
3769                         .type = PERF_RECORD_THROTTLE,
3770                         .misc = 0,
3771                         .size = sizeof(throttle_event),
3772                 },
3773                 .time           = perf_clock(),
3774                 .id             = primary_event_id(event),
3775                 .stream_id      = event->id,
3776         };
3777
3778         if (enable)
3779                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3780
3781         ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3782         if (ret)
3783                 return;
3784
3785         perf_output_put(&handle, throttle_event);
3786         perf_output_end(&handle);
3787 }
3788
3789 /*
3790  * Generic event overflow handling, sampling.
3791  */
3792
3793 static int __perf_event_overflow(struct perf_event *event, int nmi,
3794                                    int throttle, struct perf_sample_data *data,
3795                                    struct pt_regs *regs)
3796 {
3797         int events = atomic_read(&event->event_limit);
3798         struct hw_perf_event *hwc = &event->hw;
3799         int ret = 0;
3800
3801         throttle = (throttle && event->pmu->unthrottle != NULL);
3802
3803         if (!throttle) {
3804                 hwc->interrupts++;
3805         } else {
3806                 if (hwc->interrupts != MAX_INTERRUPTS) {
3807                         hwc->interrupts++;
3808                         if (HZ * hwc->interrupts >
3809                                         (u64)sysctl_perf_event_sample_rate) {
3810                                 hwc->interrupts = MAX_INTERRUPTS;
3811                                 perf_log_throttle(event, 0);
3812                                 ret = 1;
3813                         }
3814                 } else {
3815                         /*
3816                          * Keep re-disabling events even though on the previous
3817                          * pass we disabled it - just in case we raced with a
3818                          * sched-in and the event got enabled again:
3819                          */
3820                         ret = 1;
3821                 }
3822         }
3823
3824         if (event->attr.freq) {
3825                 u64 now = perf_clock();
3826                 s64 delta = now - hwc->freq_time_stamp;
3827
3828                 hwc->freq_time_stamp = now;
3829
3830                 if (delta > 0 && delta < 2*TICK_NSEC)
3831                         perf_adjust_period(event, delta, hwc->last_period);
3832         }
3833
3834         /*
3835          * XXX event_limit might not quite work as expected on inherited
3836          * events
3837          */
3838
3839         event->pending_kill = POLL_IN;
3840         if (events && atomic_dec_and_test(&event->event_limit)) {
3841                 ret = 1;
3842                 event->pending_kill = POLL_HUP;
3843                 if (nmi) {
3844                         event->pending_disable = 1;
3845                         perf_pending_queue(&event->pending,
3846                                            perf_pending_event);
3847                 } else
3848                         perf_event_disable(event);
3849         }
3850
3851         if (event->overflow_handler)
3852                 event->overflow_handler(event, nmi, data, regs);
3853         else
3854                 perf_event_output(event, nmi, data, regs);
3855
3856         return ret;
3857 }
3858
3859 int perf_event_overflow(struct perf_event *event, int nmi,
3860                           struct perf_sample_data *data,
3861                           struct pt_regs *regs)
3862 {
3863         return __perf_event_overflow(event, nmi, 1, data, regs);
3864 }
3865
3866 /*
3867  * Generic software event infrastructure
3868  */
3869
3870 /*
3871  * We directly increment event->count and keep a second value in
3872  * event->hw.period_left to count intervals. This period event
3873  * is kept in the range [-sample_period, 0] so that we can use the
3874  * sign as trigger.
3875  */
3876
3877 static u64 perf_swevent_set_period(struct perf_event *event)
3878 {
3879         struct hw_perf_event *hwc = &event->hw;
3880         u64 period = hwc->last_period;
3881         u64 nr, offset;
3882         s64 old, val;
3883
3884         hwc->last_period = hwc->sample_period;
3885
3886 again:
3887         old = val = atomic64_read(&hwc->period_left);
3888         if (val < 0)
3889                 return 0;
3890
3891         nr = div64_u64(period + val, period);
3892         offset = nr * period;
3893         val -= offset;
3894         if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3895                 goto again;
3896
3897         return nr;
3898 }
3899
3900 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3901                                     int nmi, struct perf_sample_data *data,
3902                                     struct pt_regs *regs)
3903 {
3904         struct hw_perf_event *hwc = &event->hw;
3905         int throttle = 0;
3906
3907         data->period = event->hw.last_period;
3908         if (!overflow)
3909                 overflow = perf_swevent_set_period(event);
3910
3911         if (hwc->interrupts == MAX_INTERRUPTS)
3912                 return;
3913
3914         for (; overflow; overflow--) {
3915                 if (__perf_event_overflow(event, nmi, throttle,
3916                                             data, regs)) {
3917                         /*
3918                          * We inhibit the overflow from happening when
3919                          * hwc->interrupts == MAX_INTERRUPTS.
3920                          */
3921                         break;
3922                 }
3923                 throttle = 1;
3924         }
3925 }
3926
3927 static void perf_swevent_unthrottle(struct perf_event *event)
3928 {
3929         /*
3930          * Nothing to do, we already reset hwc->interrupts.
3931          */
3932 }
3933
3934 static void perf_swevent_add(struct perf_event *event, u64 nr,
3935                                int nmi, struct perf_sample_data *data,
3936                                struct pt_regs *regs)
3937 {
3938         struct hw_perf_event *hwc = &event->hw;
3939
3940         atomic64_add(nr, &event->count);
3941
3942         if (!regs)
3943                 return;
3944
3945         if (!hwc->sample_period)
3946                 return;
3947
3948         if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3949                 return perf_swevent_overflow(event, 1, nmi, data, regs);
3950
3951         if (atomic64_add_negative(nr, &hwc->period_left))
3952                 return;
3953
3954         perf_swevent_overflow(event, 0, nmi, data, regs);
3955 }
3956
3957 static int perf_swevent_is_counting(struct perf_event *event)
3958 {
3959         /*
3960          * The event is active, we're good!
3961          */
3962         if (event->state == PERF_EVENT_STATE_ACTIVE)
3963                 return 1;
3964
3965         /*
3966          * The event is off/error, not counting.
3967          */
3968         if (event->state != PERF_EVENT_STATE_INACTIVE)
3969                 return 0;
3970
3971         /*
3972          * The event is inactive, if the context is active
3973          * we're part of a group that didn't make it on the 'pmu',
3974          * not counting.
3975          */
3976         if (event->ctx->is_active)
3977                 return 0;
3978
3979         /*
3980          * We're inactive and the context is too, this means the
3981          * task is scheduled out, we're counting events that happen
3982          * to us, like migration events.
3983          */
3984         return 1;
3985 }
3986
3987 static int perf_tp_event_match(struct perf_event *event,
3988                                 struct perf_sample_data *data);
3989
3990 static int perf_exclude_event(struct perf_event *event,
3991                               struct pt_regs *regs)
3992 {
3993         if (regs) {
3994                 if (event->attr.exclude_user && user_mode(regs))
3995                         return 1;
3996
3997                 if (event->attr.exclude_kernel && !user_mode(regs))
3998                         return 1;
3999         }
4000
4001         return 0;
4002 }
4003
4004 static int perf_swevent_match(struct perf_event *event,
4005                                 enum perf_type_id type,
4006                                 u32 event_id,
4007                                 struct perf_sample_data *data,
4008                                 struct pt_regs *regs)
4009 {
4010         if (event->cpu != -1 && event->cpu != smp_processor_id())
4011                 return 0;
4012
4013         if (!perf_swevent_is_counting(event))
4014                 return 0;
4015
4016         if (event->attr.type != type)
4017                 return 0;
4018
4019         if (event->attr.config != event_id)
4020                 return 0;
4021
4022         if (perf_exclude_event(event, regs))
4023                 return 0;
4024
4025         if (event->attr.type == PERF_TYPE_TRACEPOINT &&
4026             !perf_tp_event_match(event, data))
4027                 return 0;
4028
4029         return 1;
4030 }
4031
4032 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
4033                                      enum perf_type_id type,
4034                                      u32 event_id, u64 nr, int nmi,
4035                                      struct perf_sample_data *data,
4036                                      struct pt_regs *regs)
4037 {
4038         struct perf_event *event;
4039
4040         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4041                 if (perf_swevent_match(event, type, event_id, data, regs))
4042                         perf_swevent_add(event, nr, nmi, data, regs);
4043         }
4044 }
4045
4046 int perf_swevent_get_recursion_context(void)
4047 {
4048         struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
4049         int rctx;
4050
4051         if (in_nmi())
4052                 rctx = 3;
4053         else if (in_irq())
4054                 rctx = 2;
4055         else if (in_softirq())
4056                 rctx = 1;
4057         else
4058                 rctx = 0;
4059
4060         if (cpuctx->recursion[rctx]) {
4061                 put_cpu_var(perf_cpu_context);
4062                 return -1;
4063         }
4064
4065         cpuctx->recursion[rctx]++;
4066         barrier();
4067
4068         return rctx;
4069 }
4070 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4071
4072 void perf_swevent_put_recursion_context(int rctx)
4073 {
4074         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4075         barrier();
4076         cpuctx->recursion[rctx]--;
4077         put_cpu_var(perf_cpu_context);
4078 }
4079 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
4080
4081 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4082                                     u64 nr, int nmi,
4083                                     struct perf_sample_data *data,
4084                                     struct pt_regs *regs)
4085 {
4086         struct perf_cpu_context *cpuctx;
4087         struct perf_event_context *ctx;
4088
4089         cpuctx = &__get_cpu_var(perf_cpu_context);
4090         rcu_read_lock();
4091         perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
4092                                  nr, nmi, data, regs);
4093         /*
4094          * doesn't really matter which of the child contexts the
4095          * events ends up in.
4096          */
4097         ctx = rcu_dereference(current->perf_event_ctxp);
4098         if (ctx)
4099                 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
4100         rcu_read_unlock();
4101 }
4102
4103 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4104                             struct pt_regs *regs, u64 addr)
4105 {
4106         struct perf_sample_data data;
4107         int rctx;
4108
4109         rctx = perf_swevent_get_recursion_context();
4110         if (rctx < 0)
4111                 return;
4112
4113         perf_sample_data_init(&data, addr);
4114
4115         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4116
4117         perf_swevent_put_recursion_context(rctx);
4118 }
4119
4120 static void perf_swevent_read(struct perf_event *event)
4121 {
4122 }
4123
4124 static int perf_swevent_enable(struct perf_event *event)
4125 {
4126         struct hw_perf_event *hwc = &event->hw;
4127
4128         if (hwc->sample_period) {
4129                 hwc->last_period = hwc->sample_period;
4130                 perf_swevent_set_period(event);
4131         }
4132         return 0;
4133 }
4134
4135 static void perf_swevent_disable(struct perf_event *event)
4136 {
4137 }
4138
4139 static const struct pmu perf_ops_generic = {
4140         .enable         = perf_swevent_enable,
4141         .disable        = perf_swevent_disable,
4142         .read           = perf_swevent_read,
4143         .unthrottle     = perf_swevent_unthrottle,
4144 };
4145
4146 /*
4147  * hrtimer based swevent callback
4148  */
4149
4150 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4151 {
4152         enum hrtimer_restart ret = HRTIMER_RESTART;
4153         struct perf_sample_data data;
4154         struct pt_regs *regs;
4155         struct perf_event *event;
4156         u64 period;
4157
4158         event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4159         event->pmu->read(event);
4160
4161         perf_sample_data_init(&data, 0);
4162         data.period = event->hw.last_period;
4163         regs = get_irq_regs();
4164         /*
4165          * In case we exclude kernel IPs or are somehow not in interrupt
4166          * context, provide the next best thing, the user IP.
4167          */
4168         if ((event->attr.exclude_kernel || !regs) &&
4169                         !event->attr.exclude_user)
4170                 regs = task_pt_regs(current);
4171
4172         if (regs) {
4173                 if (!(event->attr.exclude_idle && current->pid == 0))
4174                         if (perf_event_overflow(event, 0, &data, regs))
4175                                 ret = HRTIMER_NORESTART;
4176         }
4177
4178         period = max_t(u64, 10000, event->hw.sample_period);
4179         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4180
4181         return ret;
4182 }
4183
4184 static void perf_swevent_start_hrtimer(struct perf_event *event)
4185 {
4186         struct hw_perf_event *hwc = &event->hw;
4187
4188         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4189         hwc->hrtimer.function = perf_swevent_hrtimer;
4190         if (hwc->sample_period) {
4191                 u64 period;
4192
4193                 if (hwc->remaining) {
4194                         if (hwc->remaining < 0)
4195                                 period = 10000;
4196                         else
4197                                 period = hwc->remaining;
4198                         hwc->remaining = 0;
4199                 } else {
4200                         period = max_t(u64, 10000, hwc->sample_period);
4201                 }
4202                 __hrtimer_start_range_ns(&hwc->hrtimer,
4203                                 ns_to_ktime(period), 0,
4204                                 HRTIMER_MODE_REL, 0);
4205         }
4206 }
4207
4208 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4209 {
4210         struct hw_perf_event *hwc = &event->hw;
4211
4212         if (hwc->sample_period) {
4213                 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4214                 hwc->remaining = ktime_to_ns(remaining);
4215
4216                 hrtimer_cancel(&hwc->hrtimer);
4217         }
4218 }
4219
4220 /*
4221  * Software event: cpu wall time clock
4222  */
4223
4224 static void cpu_clock_perf_event_update(struct perf_event *event)
4225 {
4226         int cpu = raw_smp_processor_id();
4227         s64 prev;
4228         u64 now;
4229
4230         now = cpu_clock(cpu);
4231         prev = atomic64_xchg(&event->hw.prev_count, now);
4232         atomic64_add(now - prev, &event->count);
4233 }
4234
4235 static int cpu_clock_perf_event_enable(struct perf_event *event)
4236 {
4237         struct hw_perf_event *hwc = &event->hw;
4238         int cpu = raw_smp_processor_id();
4239
4240         atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4241         perf_swevent_start_hrtimer(event);
4242
4243         return 0;
4244 }
4245
4246 static void cpu_clock_perf_event_disable(struct perf_event *event)
4247 {
4248         perf_swevent_cancel_hrtimer(event);
4249         cpu_clock_perf_event_update(event);
4250 }
4251
4252 static void cpu_clock_perf_event_read(struct perf_event *event)
4253 {
4254         cpu_clock_perf_event_update(event);
4255 }
4256
4257 static const struct pmu perf_ops_cpu_clock = {
4258         .enable         = cpu_clock_perf_event_enable,
4259         .disable        = cpu_clock_perf_event_disable,
4260         .read           = cpu_clock_perf_event_read,
4261 };
4262
4263 /*
4264  * Software event: task time clock
4265  */
4266
4267 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4268 {
4269         u64 prev;
4270         s64 delta;
4271
4272         prev = atomic64_xchg(&event->hw.prev_count, now);
4273         delta = now - prev;
4274         atomic64_add(delta, &event->count);
4275 }
4276
4277 static int task_clock_perf_event_enable(struct perf_event *event)
4278 {
4279         struct hw_perf_event *hwc = &event->hw;
4280         u64 now;
4281
4282         now = event->ctx->time;
4283
4284         atomic64_set(&hwc->prev_count, now);
4285
4286         perf_swevent_start_hrtimer(event);
4287
4288         return 0;
4289 }
4290
4291 static void task_clock_perf_event_disable(struct perf_event *event)
4292 {
4293         perf_swevent_cancel_hrtimer(event);
4294         task_clock_perf_event_update(event, event->ctx->time);
4295
4296 }
4297
4298 static void task_clock_perf_event_read(struct perf_event *event)
4299 {
4300         u64 time;
4301
4302         if (!in_nmi()) {
4303                 update_context_time(event->ctx);
4304                 time = event->ctx->time;
4305         } else {
4306                 u64 now = perf_clock();
4307                 u64 delta = now - event->ctx->timestamp;
4308                 time = event->ctx->time + delta;
4309         }
4310
4311         task_clock_perf_event_update(event, time);
4312 }
4313
4314 static const struct pmu perf_ops_task_clock = {
4315         .enable         = task_clock_perf_event_enable,
4316         .disable        = task_clock_perf_event_disable,
4317         .read           = task_clock_perf_event_read,
4318 };
4319
4320 #ifdef CONFIG_EVENT_TRACING
4321
4322 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4323                    int entry_size, struct pt_regs *regs)
4324 {
4325         struct perf_sample_data data;
4326         struct perf_raw_record raw = {
4327                 .size = entry_size,
4328                 .data = record,
4329         };
4330
4331         perf_sample_data_init(&data, addr);
4332         data.raw = &raw;
4333
4334         /* Trace events already protected against recursion */
4335         do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4336                          &data, regs);
4337 }
4338 EXPORT_SYMBOL_GPL(perf_tp_event);
4339
4340 static int perf_tp_event_match(struct perf_event *event,
4341                                 struct perf_sample_data *data)
4342 {
4343         void *record = data->raw->data;
4344
4345         if (likely(!event->filter) || filter_match_preds(event->filter, record))
4346                 return 1;
4347         return 0;
4348 }
4349
4350 static void tp_perf_event_destroy(struct perf_event *event)
4351 {
4352         perf_trace_disable(event->attr.config);
4353 }
4354
4355 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4356 {
4357         /*
4358          * Raw tracepoint data is a severe data leak, only allow root to
4359          * have these.
4360          */
4361         if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4362                         perf_paranoid_tracepoint_raw() &&
4363                         !capable(CAP_SYS_ADMIN))
4364                 return ERR_PTR(-EPERM);
4365
4366         if (perf_trace_enable(event->attr.config))
4367                 return NULL;
4368
4369         event->destroy = tp_perf_event_destroy;
4370
4371         return &perf_ops_generic;
4372 }
4373
4374 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4375 {
4376         char *filter_str;
4377         int ret;
4378
4379         if (event->attr.type != PERF_TYPE_TRACEPOINT)
4380                 return -EINVAL;
4381
4382         filter_str = strndup_user(arg, PAGE_SIZE);
4383         if (IS_ERR(filter_str))
4384                 return PTR_ERR(filter_str);
4385
4386         ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4387
4388         kfree(filter_str);
4389         return ret;
4390 }
4391
4392 static void perf_event_free_filter(struct perf_event *event)
4393 {
4394         ftrace_profile_free_filter(event);
4395 }
4396
4397 #else
4398
4399 static int perf_tp_event_match(struct perf_event *event,
4400                                 struct perf_sample_data *data)
4401 {
4402         return 1;
4403 }
4404
4405 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4406 {
4407         return NULL;
4408 }
4409
4410 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4411 {
4412         return -ENOENT;
4413 }
4414
4415 static void perf_event_free_filter(struct perf_event *event)
4416 {
4417 }
4418
4419 #endif /* CONFIG_EVENT_TRACING */
4420
4421 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4422 static void bp_perf_event_destroy(struct perf_event *event)
4423 {
4424         release_bp_slot(event);
4425 }
4426
4427 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4428 {
4429         int err;
4430
4431         err = register_perf_hw_breakpoint(bp);
4432         if (err)
4433                 return ERR_PTR(err);
4434
4435         bp->destroy = bp_perf_event_destroy;
4436
4437         return &perf_ops_bp;
4438 }
4439
4440 void perf_bp_event(struct perf_event *bp, void *data)
4441 {
4442         struct perf_sample_data sample;
4443         struct pt_regs *regs = data;
4444
4445         perf_sample_data_init(&sample, bp->attr.bp_addr);
4446
4447         if (!perf_exclude_event(bp, regs))
4448                 perf_swevent_add(bp, 1, 1, &sample, regs);
4449 }
4450 #else
4451 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4452 {
4453         return NULL;
4454 }
4455
4456 void perf_bp_event(struct perf_event *bp, void *regs)
4457 {
4458 }
4459 #endif
4460
4461 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4462
4463 static void sw_perf_event_destroy(struct perf_event *event)
4464 {
4465         u64 event_id = event->attr.config;
4466
4467         WARN_ON(event->parent);
4468
4469         atomic_dec(&perf_swevent_enabled[event_id]);
4470 }
4471
4472 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4473 {
4474         const struct pmu *pmu = NULL;
4475         u64 event_id = event->attr.config;
4476
4477         /*
4478          * Software events (currently) can't in general distinguish
4479          * between user, kernel and hypervisor events.
4480          * However, context switches and cpu migrations are considered
4481          * to be kernel events, and page faults are never hypervisor
4482          * events.
4483          */
4484         switch (event_id) {
4485         case PERF_COUNT_SW_CPU_CLOCK:
4486                 pmu = &perf_ops_cpu_clock;
4487
4488                 break;
4489         case PERF_COUNT_SW_TASK_CLOCK:
4490                 /*
4491                  * If the user instantiates this as a per-cpu event,
4492                  * use the cpu_clock event instead.
4493                  */
4494                 if (event->ctx->task)
4495                         pmu = &perf_ops_task_clock;
4496                 else
4497                         pmu = &perf_ops_cpu_clock;
4498
4499                 break;
4500         case PERF_COUNT_SW_PAGE_FAULTS:
4501         case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4502         case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4503         case PERF_COUNT_SW_CONTEXT_SWITCHES:
4504         case PERF_COUNT_SW_CPU_MIGRATIONS:
4505         case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4506         case PERF_COUNT_SW_EMULATION_FAULTS:
4507                 if (!event->parent) {
4508                         atomic_inc(&perf_swevent_enabled[event_id]);
4509                         event->destroy = sw_perf_event_destroy;
4510                 }
4511                 pmu = &perf_ops_generic;
4512                 break;
4513         }
4514
4515         return pmu;
4516 }
4517
4518 /*
4519  * Allocate and initialize a event structure
4520  */
4521 static struct perf_event *
4522 perf_event_alloc(struct perf_event_attr *attr,
4523                    int cpu,
4524                    struct perf_event_context *ctx,
4525                    struct perf_event *group_leader,
4526                    struct perf_event *parent_event,
4527                    perf_overflow_handler_t overflow_handler,
4528                    gfp_t gfpflags)
4529 {
4530         const struct pmu *pmu;
4531         struct perf_event *event;
4532         struct hw_perf_event *hwc;
4533         long err;
4534
4535         event = kzalloc(sizeof(*event), gfpflags);
4536         if (!event)
4537                 return ERR_PTR(-ENOMEM);
4538
4539         /*
4540          * Single events are their own group leaders, with an
4541          * empty sibling list:
4542          */
4543         if (!group_leader)
4544                 group_leader = event;
4545
4546         mutex_init(&event->child_mutex);
4547         INIT_LIST_HEAD(&event->child_list);
4548
4549         INIT_LIST_HEAD(&event->group_entry);
4550         INIT_LIST_HEAD(&event->event_entry);
4551         INIT_LIST_HEAD(&event->sibling_list);
4552         init_waitqueue_head(&event->waitq);
4553
4554         mutex_init(&event->mmap_mutex);
4555
4556         event->cpu              = cpu;
4557         event->attr             = *attr;
4558         event->group_leader     = group_leader;
4559         event->pmu              = NULL;
4560         event->ctx              = ctx;
4561         event->oncpu            = -1;
4562
4563         event->parent           = parent_event;
4564
4565         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
4566         event->id               = atomic64_inc_return(&perf_event_id);
4567
4568         event->state            = PERF_EVENT_STATE_INACTIVE;
4569
4570         if (!overflow_handler && parent_event)
4571                 overflow_handler = parent_event->overflow_handler;
4572         
4573         event->overflow_handler = overflow_handler;
4574
4575         if (attr->disabled)
4576                 event->state = PERF_EVENT_STATE_OFF;
4577
4578         pmu = NULL;
4579
4580         hwc = &event->hw;
4581         hwc->sample_period = attr->sample_period;
4582         if (attr->freq && attr->sample_freq)
4583                 hwc->sample_period = 1;
4584         hwc->last_period = hwc->sample_period;
4585
4586         atomic64_set(&hwc->period_left, hwc->sample_period);
4587
4588         /*
4589          * we currently do not support PERF_FORMAT_GROUP on inherited events
4590          */
4591         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4592                 goto done;
4593
4594         switch (attr->type) {
4595         case PERF_TYPE_RAW:
4596         case PERF_TYPE_HARDWARE:
4597         case PERF_TYPE_HW_CACHE:
4598                 pmu = hw_perf_event_init(event);
4599                 break;
4600
4601         case PERF_TYPE_SOFTWARE:
4602                 pmu = sw_perf_event_init(event);
4603                 break;
4604
4605         case PERF_TYPE_TRACEPOINT:
4606                 pmu = tp_perf_event_init(event);
4607                 break;
4608
4609         case PERF_TYPE_BREAKPOINT:
4610                 pmu = bp_perf_event_init(event);
4611                 break;
4612
4613
4614         default:
4615                 break;
4616         }
4617 done:
4618         err = 0;
4619         if (!pmu)
4620                 err = -EINVAL;
4621         else if (IS_ERR(pmu))
4622                 err = PTR_ERR(pmu);
4623
4624         if (err) {
4625                 if (event->ns)
4626                         put_pid_ns(event->ns);
4627                 kfree(event);
4628                 return ERR_PTR(err);
4629         }
4630
4631         event->pmu = pmu;
4632
4633         if (!event->parent) {
4634                 atomic_inc(&nr_events);
4635                 if (event->attr.mmap)
4636                         atomic_inc(&nr_mmap_events);
4637                 if (event->attr.comm)
4638                         atomic_inc(&nr_comm_events);
4639                 if (event->attr.task)
4640                         atomic_inc(&nr_task_events);
4641         }
4642
4643         return event;
4644 }
4645
4646 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4647                           struct perf_event_attr *attr)
4648 {
4649         u32 size;
4650         int ret;
4651
4652         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4653                 return -EFAULT;
4654
4655         /*
4656          * zero the full structure, so that a short copy will be nice.
4657          */
4658         memset(attr, 0, sizeof(*attr));
4659
4660         ret = get_user(size, &uattr->size);
4661         if (ret)
4662                 return ret;
4663
4664         if (size > PAGE_SIZE)   /* silly large */
4665                 goto err_size;
4666
4667         if (!size)              /* abi compat */
4668                 size = PERF_ATTR_SIZE_VER0;
4669
4670         if (size < PERF_ATTR_SIZE_VER0)
4671                 goto err_size;
4672
4673         /*
4674          * If we're handed a bigger struct than we know of,
4675          * ensure all the unknown bits are 0 - i.e. new
4676          * user-space does not rely on any kernel feature
4677          * extensions we dont know about yet.
4678          */
4679         if (size > sizeof(*attr)) {
4680                 unsigned char __user *addr;
4681                 unsigned char __user *end;
4682                 unsigned char val;
4683
4684                 addr = (void __user *)uattr + sizeof(*attr);
4685                 end  = (void __user *)uattr + size;
4686
4687                 for (; addr < end; addr++) {
4688                         ret = get_user(val, addr);
4689                         if (ret)
4690                                 return ret;
4691                         if (val)
4692                                 goto err_size;
4693                 }
4694                 size = sizeof(*attr);
4695         }
4696
4697         ret = copy_from_user(attr, uattr, size);
4698         if (ret)
4699                 return -EFAULT;
4700
4701         /*
4702          * If the type exists, the corresponding creation will verify
4703          * the attr->config.
4704          */
4705         if (attr->type >= PERF_TYPE_MAX)
4706                 return -EINVAL;
4707
4708         if (attr->__reserved_1)
4709                 return -EINVAL;
4710
4711         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4712                 return -EINVAL;
4713
4714         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4715                 return -EINVAL;
4716
4717 out:
4718         return ret;
4719
4720 err_size:
4721         put_user(sizeof(*attr), &uattr->size);
4722         ret = -E2BIG;
4723         goto out;
4724 }
4725
4726 static int perf_event_set_output(struct perf_event *event, int output_fd)
4727 {
4728         struct perf_event *output_event = NULL;
4729         struct file *output_file = NULL;
4730         struct perf_event *old_output;
4731         int fput_needed = 0;
4732         int ret = -EINVAL;
4733
4734         if (!output_fd)
4735                 goto set;
4736
4737         output_file = fget_light(output_fd, &fput_needed);
4738         if (!output_file)
4739                 return -EBADF;
4740
4741         if (output_file->f_op != &perf_fops)
4742                 goto out;
4743
4744         output_event = output_file->private_data;
4745
4746         /* Don't chain output fds */
4747         if (output_event->output)
4748                 goto out;
4749
4750         /* Don't set an output fd when we already have an output channel */
4751         if (event->data)
4752                 goto out;
4753
4754         atomic_long_inc(&output_file->f_count);
4755
4756 set:
4757         mutex_lock(&event->mmap_mutex);
4758         old_output = event->output;
4759         rcu_assign_pointer(event->output, output_event);
4760         mutex_unlock(&event->mmap_mutex);
4761
4762         if (old_output) {
4763                 /*
4764                  * we need to make sure no existing perf_output_*()
4765                  * is still referencing this event.
4766                  */
4767                 synchronize_rcu();
4768                 fput(old_output->filp);
4769         }
4770
4771         ret = 0;
4772 out:
4773         fput_light(output_file, fput_needed);
4774         return ret;
4775 }
4776
4777 /**
4778  * sys_perf_event_open - open a performance event, associate it to a task/cpu
4779  *
4780  * @attr_uptr:  event_id type attributes for monitoring/sampling
4781  * @pid:                target pid
4782  * @cpu:                target cpu
4783  * @group_fd:           group leader event fd
4784  */
4785 SYSCALL_DEFINE5(perf_event_open,
4786                 struct perf_event_attr __user *, attr_uptr,
4787                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4788 {
4789         struct perf_event *event, *group_leader;
4790         struct perf_event_attr attr;
4791         struct perf_event_context *ctx;
4792         struct file *event_file = NULL;
4793         struct file *group_file = NULL;
4794         int fput_needed = 0;
4795         int fput_needed2 = 0;
4796         int err;
4797
4798         /* for future expandability... */
4799         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4800                 return -EINVAL;
4801
4802         err = perf_copy_attr(attr_uptr, &attr);
4803         if (err)
4804                 return err;
4805
4806         if (!attr.exclude_kernel) {
4807                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4808                         return -EACCES;
4809         }
4810
4811         if (attr.freq) {
4812                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4813                         return -EINVAL;
4814         }
4815
4816         /*
4817          * Get the target context (task or percpu):
4818          */
4819         ctx = find_get_context(pid, cpu);
4820         if (IS_ERR(ctx))
4821                 return PTR_ERR(ctx);
4822
4823         /*
4824          * Look up the group leader (we will attach this event to it):
4825          */
4826         group_leader = NULL;
4827         if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4828                 err = -EINVAL;
4829                 group_file = fget_light(group_fd, &fput_needed);
4830                 if (!group_file)
4831                         goto err_put_context;
4832                 if (group_file->f_op != &perf_fops)
4833                         goto err_put_context;
4834
4835                 group_leader = group_file->private_data;
4836                 /*
4837                  * Do not allow a recursive hierarchy (this new sibling
4838                  * becoming part of another group-sibling):
4839                  */
4840                 if (group_leader->group_leader != group_leader)
4841                         goto err_put_context;
4842                 /*
4843                  * Do not allow to attach to a group in a different
4844                  * task or CPU context:
4845                  */
4846                 if (group_leader->ctx != ctx)
4847                         goto err_put_context;
4848                 /*
4849                  * Only a group leader can be exclusive or pinned
4850                  */
4851                 if (attr.exclusive || attr.pinned)
4852                         goto err_put_context;
4853         }
4854
4855         event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4856                                      NULL, NULL, GFP_KERNEL);
4857         err = PTR_ERR(event);
4858         if (IS_ERR(event))
4859                 goto err_put_context;
4860
4861         err = anon_inode_getfd("[perf_event]", &perf_fops, event, O_RDWR);
4862         if (err < 0)
4863                 goto err_free_put_context;
4864
4865         event_file = fget_light(err, &fput_needed2);
4866         if (!event_file)
4867                 goto err_free_put_context;
4868
4869         if (flags & PERF_FLAG_FD_OUTPUT) {
4870                 err = perf_event_set_output(event, group_fd);
4871                 if (err)
4872                         goto err_fput_free_put_context;
4873         }
4874
4875         event->filp = event_file;
4876         WARN_ON_ONCE(ctx->parent_ctx);
4877         mutex_lock(&ctx->mutex);
4878         perf_install_in_context(ctx, event, cpu);
4879         ++ctx->generation;
4880         mutex_unlock(&ctx->mutex);
4881
4882         event->owner = current;
4883         get_task_struct(current);
4884         mutex_lock(&current->perf_event_mutex);
4885         list_add_tail(&event->owner_entry, &current->perf_event_list);
4886         mutex_unlock(&current->perf_event_mutex);
4887
4888 err_fput_free_put_context:
4889         fput_light(event_file, fput_needed2);
4890
4891 err_free_put_context:
4892         if (err < 0)
4893                 kfree(event);
4894
4895 err_put_context:
4896         if (err < 0)
4897                 put_ctx(ctx);
4898
4899         fput_light(group_file, fput_needed);
4900
4901         return err;
4902 }
4903
4904 /**
4905  * perf_event_create_kernel_counter
4906  *
4907  * @attr: attributes of the counter to create
4908  * @cpu: cpu in which the counter is bound
4909  * @pid: task to profile
4910  */
4911 struct perf_event *
4912 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4913                                  pid_t pid,
4914                                  perf_overflow_handler_t overflow_handler)
4915 {
4916         struct perf_event *event;
4917         struct perf_event_context *ctx;
4918         int err;
4919
4920         /*
4921          * Get the target context (task or percpu):
4922          */
4923
4924         ctx = find_get_context(pid, cpu);
4925         if (IS_ERR(ctx)) {
4926                 err = PTR_ERR(ctx);
4927                 goto err_exit;
4928         }
4929
4930         event = perf_event_alloc(attr, cpu, ctx, NULL,
4931                                  NULL, overflow_handler, GFP_KERNEL);
4932         if (IS_ERR(event)) {
4933                 err = PTR_ERR(event);
4934                 goto err_put_context;
4935         }
4936
4937         event->filp = NULL;
4938         WARN_ON_ONCE(ctx->parent_ctx);
4939         mutex_lock(&ctx->mutex);
4940         perf_install_in_context(ctx, event, cpu);
4941         ++ctx->generation;
4942         mutex_unlock(&ctx->mutex);
4943
4944         event->owner = current;
4945         get_task_struct(current);
4946         mutex_lock(&current->perf_event_mutex);
4947         list_add_tail(&event->owner_entry, &current->perf_event_list);
4948         mutex_unlock(&current->perf_event_mutex);
4949
4950         return event;
4951
4952  err_put_context:
4953         put_ctx(ctx);
4954  err_exit:
4955         return ERR_PTR(err);
4956 }
4957 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4958
4959 /*
4960  * inherit a event from parent task to child task:
4961  */
4962 static struct perf_event *
4963 inherit_event(struct perf_event *parent_event,
4964               struct task_struct *parent,
4965               struct perf_event_context *parent_ctx,
4966               struct task_struct *child,
4967               struct perf_event *group_leader,
4968               struct perf_event_context *child_ctx)
4969 {
4970         struct perf_event *child_event;
4971
4972         /*
4973          * Instead of creating recursive hierarchies of events,
4974          * we link inherited events back to the original parent,
4975          * which has a filp for sure, which we use as the reference
4976          * count:
4977          */
4978         if (parent_event->parent)
4979                 parent_event = parent_event->parent;
4980
4981         child_event = perf_event_alloc(&parent_event->attr,
4982                                            parent_event->cpu, child_ctx,
4983                                            group_leader, parent_event,
4984                                            NULL, GFP_KERNEL);
4985         if (IS_ERR(child_event))
4986                 return child_event;
4987         get_ctx(child_ctx);
4988
4989         /*
4990          * Make the child state follow the state of the parent event,
4991          * not its attr.disabled bit.  We hold the parent's mutex,
4992          * so we won't race with perf_event_{en, dis}able_family.
4993          */
4994         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
4995                 child_event->state = PERF_EVENT_STATE_INACTIVE;
4996         else
4997                 child_event->state = PERF_EVENT_STATE_OFF;
4998
4999         if (parent_event->attr.freq) {
5000                 u64 sample_period = parent_event->hw.sample_period;
5001                 struct hw_perf_event *hwc = &child_event->hw;
5002
5003                 hwc->sample_period = sample_period;
5004                 hwc->last_period   = sample_period;
5005
5006                 atomic64_set(&hwc->period_left, sample_period);
5007         }
5008
5009         child_event->overflow_handler = parent_event->overflow_handler;
5010
5011         /*
5012          * Link it up in the child's context:
5013          */
5014         add_event_to_ctx(child_event, child_ctx);
5015
5016         /*
5017          * Get a reference to the parent filp - we will fput it
5018          * when the child event exits. This is safe to do because
5019          * we are in the parent and we know that the filp still
5020          * exists and has a nonzero count:
5021          */
5022         atomic_long_inc(&parent_event->filp->f_count);
5023
5024         /*
5025          * Link this into the parent event's child list
5026          */
5027         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5028         mutex_lock(&parent_event->child_mutex);
5029         list_add_tail(&child_event->child_list, &parent_event->child_list);
5030         mutex_unlock(&parent_event->child_mutex);
5031
5032         return child_event;
5033 }
5034
5035 static int inherit_group(struct perf_event *parent_event,
5036               struct task_struct *parent,
5037               struct perf_event_context *parent_ctx,
5038               struct task_struct *child,
5039               struct perf_event_context *child_ctx)
5040 {
5041         struct perf_event *leader;
5042         struct perf_event *sub;
5043         struct perf_event *child_ctr;
5044
5045         leader = inherit_event(parent_event, parent, parent_ctx,
5046                                  child, NULL, child_ctx);
5047         if (IS_ERR(leader))
5048                 return PTR_ERR(leader);
5049         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5050                 child_ctr = inherit_event(sub, parent, parent_ctx,
5051                                             child, leader, child_ctx);
5052                 if (IS_ERR(child_ctr))
5053                         return PTR_ERR(child_ctr);
5054         }
5055         return 0;
5056 }
5057
5058 static void sync_child_event(struct perf_event *child_event,
5059                                struct task_struct *child)
5060 {
5061         struct perf_event *parent_event = child_event->parent;
5062         u64 child_val;
5063
5064         if (child_event->attr.inherit_stat)
5065                 perf_event_read_event(child_event, child);
5066
5067         child_val = atomic64_read(&child_event->count);
5068
5069         /*
5070          * Add back the child's count to the parent's count:
5071          */
5072         atomic64_add(child_val, &parent_event->count);
5073         atomic64_add(child_event->total_time_enabled,
5074                      &parent_event->child_total_time_enabled);
5075         atomic64_add(child_event->total_time_running,
5076                      &parent_event->child_total_time_running);
5077
5078         /*
5079          * Remove this event from the parent's list
5080          */
5081         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5082         mutex_lock(&parent_event->child_mutex);
5083         list_del_init(&child_event->child_list);
5084         mutex_unlock(&parent_event->child_mutex);
5085
5086         /*
5087          * Release the parent event, if this was the last
5088          * reference to it.
5089          */
5090         fput(parent_event->filp);
5091 }
5092
5093 static void
5094 __perf_event_exit_task(struct perf_event *child_event,
5095                          struct perf_event_context *child_ctx,
5096                          struct task_struct *child)
5097 {
5098         struct perf_event *parent_event;
5099
5100         perf_event_remove_from_context(child_event);
5101
5102         parent_event = child_event->parent;
5103         /*
5104          * It can happen that parent exits first, and has events
5105          * that are still around due to the child reference. These
5106          * events need to be zapped - but otherwise linger.
5107          */
5108         if (parent_event) {
5109                 sync_child_event(child_event, child);
5110                 free_event(child_event);
5111         }
5112 }
5113
5114 /*
5115  * When a child task exits, feed back event values to parent events.
5116  */
5117 void perf_event_exit_task(struct task_struct *child)
5118 {
5119         struct perf_event *child_event, *tmp;
5120         struct perf_event_context *child_ctx;
5121         unsigned long flags;
5122
5123         if (likely(!child->perf_event_ctxp)) {
5124                 perf_event_task(child, NULL, 0);
5125                 return;
5126         }
5127
5128         local_irq_save(flags);
5129         /*
5130          * We can't reschedule here because interrupts are disabled,
5131          * and either child is current or it is a task that can't be
5132          * scheduled, so we are now safe from rescheduling changing
5133          * our context.
5134          */
5135         child_ctx = child->perf_event_ctxp;
5136         __perf_event_task_sched_out(child_ctx);
5137
5138         /*
5139          * Take the context lock here so that if find_get_context is
5140          * reading child->perf_event_ctxp, we wait until it has
5141          * incremented the context's refcount before we do put_ctx below.
5142          */
5143         raw_spin_lock(&child_ctx->lock);
5144         child->perf_event_ctxp = NULL;
5145         /*
5146          * If this context is a clone; unclone it so it can't get
5147          * swapped to another process while we're removing all
5148          * the events from it.
5149          */
5150         unclone_ctx(child_ctx);
5151         update_context_time(child_ctx);
5152         raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5153
5154         /*
5155          * Report the task dead after unscheduling the events so that we
5156          * won't get any samples after PERF_RECORD_EXIT. We can however still
5157          * get a few PERF_RECORD_READ events.
5158          */
5159         perf_event_task(child, child_ctx, 0);
5160
5161         /*
5162          * We can recurse on the same lock type through:
5163          *
5164          *   __perf_event_exit_task()
5165          *     sync_child_event()
5166          *       fput(parent_event->filp)
5167          *         perf_release()
5168          *           mutex_lock(&ctx->mutex)
5169          *
5170          * But since its the parent context it won't be the same instance.
5171          */
5172         mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
5173
5174 again:
5175         list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5176                                  group_entry)
5177                 __perf_event_exit_task(child_event, child_ctx, child);
5178
5179         list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5180                                  group_entry)
5181                 __perf_event_exit_task(child_event, child_ctx, child);
5182
5183         /*
5184          * If the last event was a group event, it will have appended all
5185          * its siblings to the list, but we obtained 'tmp' before that which
5186          * will still point to the list head terminating the iteration.
5187          */
5188         if (!list_empty(&child_ctx->pinned_groups) ||
5189             !list_empty(&child_ctx->flexible_groups))
5190                 goto again;
5191
5192         mutex_unlock(&child_ctx->mutex);
5193
5194         put_ctx(child_ctx);
5195 }
5196
5197 static void perf_free_event(struct perf_event *event,
5198                             struct perf_event_context *ctx)
5199 {
5200         struct perf_event *parent = event->parent;
5201
5202         if (WARN_ON_ONCE(!parent))
5203                 return;
5204
5205         mutex_lock(&parent->child_mutex);
5206         list_del_init(&event->child_list);
5207         mutex_unlock(&parent->child_mutex);
5208
5209         fput(parent->filp);
5210
5211         list_del_event(event, ctx);
5212         free_event(event);
5213 }
5214
5215 /*
5216  * free an unexposed, unused context as created by inheritance by
5217  * init_task below, used by fork() in case of fail.
5218  */
5219 void perf_event_free_task(struct task_struct *task)
5220 {
5221         struct perf_event_context *ctx = task->perf_event_ctxp;
5222         struct perf_event *event, *tmp;
5223
5224         if (!ctx)
5225                 return;
5226
5227         mutex_lock(&ctx->mutex);
5228 again:
5229         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5230                 perf_free_event(event, ctx);
5231
5232         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5233                                  group_entry)
5234                 perf_free_event(event, ctx);
5235
5236         if (!list_empty(&ctx->pinned_groups) ||
5237             !list_empty(&ctx->flexible_groups))
5238                 goto again;
5239
5240         mutex_unlock(&ctx->mutex);
5241
5242         put_ctx(ctx);
5243 }
5244
5245 static int
5246 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5247                    struct perf_event_context *parent_ctx,
5248                    struct task_struct *child,
5249                    int *inherited_all)
5250 {
5251         int ret;
5252         struct perf_event_context *child_ctx = child->perf_event_ctxp;
5253
5254         if (!event->attr.inherit) {
5255                 *inherited_all = 0;
5256                 return 0;
5257         }
5258
5259         if (!child_ctx) {
5260                 /*
5261                  * This is executed from the parent task context, so
5262                  * inherit events that have been marked for cloning.
5263                  * First allocate and initialize a context for the
5264                  * child.
5265                  */
5266
5267                 child_ctx = kzalloc(sizeof(struct perf_event_context),
5268                                     GFP_KERNEL);
5269                 if (!child_ctx)
5270                         return -ENOMEM;
5271
5272                 __perf_event_init_context(child_ctx, child);
5273                 child->perf_event_ctxp = child_ctx;
5274                 get_task_struct(child);
5275         }
5276
5277         ret = inherit_group(event, parent, parent_ctx,
5278                             child, child_ctx);
5279
5280         if (ret)
5281                 *inherited_all = 0;
5282
5283         return ret;
5284 }
5285
5286
5287 /*
5288  * Initialize the perf_event context in task_struct
5289  */
5290 int perf_event_init_task(struct task_struct *child)
5291 {
5292         struct perf_event_context *child_ctx, *parent_ctx;
5293         struct perf_event_context *cloned_ctx;
5294         struct perf_event *event;
5295         struct task_struct *parent = current;
5296         int inherited_all = 1;
5297         int ret = 0;
5298
5299         child->perf_event_ctxp = NULL;
5300
5301         mutex_init(&child->perf_event_mutex);
5302         INIT_LIST_HEAD(&child->perf_event_list);
5303
5304         if (likely(!parent->perf_event_ctxp))
5305                 return 0;
5306
5307         /*
5308          * If the parent's context is a clone, pin it so it won't get
5309          * swapped under us.
5310          */
5311         parent_ctx = perf_pin_task_context(parent);
5312
5313         /*
5314          * No need to check if parent_ctx != NULL here; since we saw
5315          * it non-NULL earlier, the only reason for it to become NULL
5316          * is if we exit, and since we're currently in the middle of
5317          * a fork we can't be exiting at the same time.
5318          */
5319
5320         /*
5321          * Lock the parent list. No need to lock the child - not PID
5322          * hashed yet and not running, so nobody can access it.
5323          */
5324         mutex_lock(&parent_ctx->mutex);
5325
5326         /*
5327          * We dont have to disable NMIs - we are only looking at
5328          * the list, not manipulating it:
5329          */
5330         list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5331                 ret = inherit_task_group(event, parent, parent_ctx, child,
5332                                          &inherited_all);
5333                 if (ret)
5334                         break;
5335         }
5336
5337         list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5338                 ret = inherit_task_group(event, parent, parent_ctx, child,
5339                                          &inherited_all);
5340                 if (ret)
5341                         break;
5342         }
5343
5344         child_ctx = child->perf_event_ctxp;
5345
5346         if (child_ctx && inherited_all) {
5347                 /*
5348                  * Mark the child context as a clone of the parent
5349                  * context, or of whatever the parent is a clone of.
5350                  * Note that if the parent is a clone, it could get
5351                  * uncloned at any point, but that doesn't matter
5352                  * because the list of events and the generation
5353                  * count can't have changed since we took the mutex.
5354                  */
5355                 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5356                 if (cloned_ctx) {
5357                         child_ctx->parent_ctx = cloned_ctx;
5358                         child_ctx->parent_gen = parent_ctx->parent_gen;
5359                 } else {
5360                         child_ctx->parent_ctx = parent_ctx;
5361                         child_ctx->parent_gen = parent_ctx->generation;
5362                 }
5363                 get_ctx(child_ctx->parent_ctx);
5364         }
5365
5366         mutex_unlock(&parent_ctx->mutex);
5367
5368         perf_unpin_context(parent_ctx);
5369
5370         return ret;
5371 }
5372
5373 static void __init perf_event_init_all_cpus(void)
5374 {
5375         int cpu;
5376         struct perf_cpu_context *cpuctx;
5377
5378         for_each_possible_cpu(cpu) {
5379                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5380                 __perf_event_init_context(&cpuctx->ctx, NULL);
5381         }
5382 }
5383
5384 static void __cpuinit perf_event_init_cpu(int cpu)
5385 {
5386         struct perf_cpu_context *cpuctx;
5387
5388         cpuctx = &per_cpu(perf_cpu_context, cpu);
5389
5390         spin_lock(&perf_resource_lock);
5391         cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5392         spin_unlock(&perf_resource_lock);
5393 }
5394
5395 #ifdef CONFIG_HOTPLUG_CPU
5396 static void __perf_event_exit_cpu(void *info)
5397 {
5398         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5399         struct perf_event_context *ctx = &cpuctx->ctx;
5400         struct perf_event *event, *tmp;
5401
5402         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5403                 __perf_event_remove_from_context(event);
5404         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5405                 __perf_event_remove_from_context(event);
5406 }
5407 static void perf_event_exit_cpu(int cpu)
5408 {
5409         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5410         struct perf_event_context *ctx = &cpuctx->ctx;
5411
5412         mutex_lock(&ctx->mutex);
5413         smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5414         mutex_unlock(&ctx->mutex);
5415 }
5416 #else
5417 static inline void perf_event_exit_cpu(int cpu) { }
5418 #endif
5419
5420 static int __cpuinit
5421 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5422 {
5423         unsigned int cpu = (long)hcpu;
5424
5425         switch (action) {
5426
5427         case CPU_UP_PREPARE:
5428         case CPU_UP_PREPARE_FROZEN:
5429                 perf_event_init_cpu(cpu);
5430                 break;
5431
5432         case CPU_DOWN_PREPARE:
5433         case CPU_DOWN_PREPARE_FROZEN:
5434                 perf_event_exit_cpu(cpu);
5435                 break;
5436
5437         default:
5438                 break;
5439         }
5440
5441         return NOTIFY_OK;
5442 }
5443
5444 /*
5445  * This has to have a higher priority than migration_notifier in sched.c.
5446  */
5447 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5448         .notifier_call          = perf_cpu_notify,
5449         .priority               = 20,
5450 };
5451
5452 void __init perf_event_init(void)
5453 {
5454         perf_event_init_all_cpus();
5455         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5456                         (void *)(long)smp_processor_id());
5457         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5458                         (void *)(long)smp_processor_id());
5459         register_cpu_notifier(&perf_cpu_nb);
5460 }
5461
5462 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class,
5463                                         struct sysdev_class_attribute *attr,
5464                                         char *buf)
5465 {
5466         return sprintf(buf, "%d\n", perf_reserved_percpu);
5467 }
5468
5469 static ssize_t
5470 perf_set_reserve_percpu(struct sysdev_class *class,
5471                         struct sysdev_class_attribute *attr,
5472                         const char *buf,
5473                         size_t count)
5474 {
5475         struct perf_cpu_context *cpuctx;
5476         unsigned long val;
5477         int err, cpu, mpt;
5478
5479         err = strict_strtoul(buf, 10, &val);
5480         if (err)
5481                 return err;
5482         if (val > perf_max_events)
5483                 return -EINVAL;
5484
5485         spin_lock(&perf_resource_lock);
5486         perf_reserved_percpu = val;
5487         for_each_online_cpu(cpu) {
5488                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5489                 raw_spin_lock_irq(&cpuctx->ctx.lock);
5490                 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5491                           perf_max_events - perf_reserved_percpu);
5492                 cpuctx->max_pertask = mpt;
5493                 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5494         }
5495         spin_unlock(&perf_resource_lock);
5496
5497         return count;
5498 }
5499
5500 static ssize_t perf_show_overcommit(struct sysdev_class *class,
5501                                     struct sysdev_class_attribute *attr,
5502                                     char *buf)
5503 {
5504         return sprintf(buf, "%d\n", perf_overcommit);
5505 }
5506
5507 static ssize_t
5508 perf_set_overcommit(struct sysdev_class *class,
5509                     struct sysdev_class_attribute *attr,
5510                     const char *buf, size_t count)
5511 {
5512         unsigned long val;
5513         int err;
5514
5515         err = strict_strtoul(buf, 10, &val);
5516         if (err)
5517                 return err;
5518         if (val > 1)
5519                 return -EINVAL;
5520
5521         spin_lock(&perf_resource_lock);
5522         perf_overcommit = val;
5523         spin_unlock(&perf_resource_lock);
5524
5525         return count;
5526 }
5527
5528 static SYSDEV_CLASS_ATTR(
5529                                 reserve_percpu,
5530                                 0644,
5531                                 perf_show_reserve_percpu,
5532                                 perf_set_reserve_percpu
5533                         );
5534
5535 static SYSDEV_CLASS_ATTR(
5536                                 overcommit,
5537                                 0644,
5538                                 perf_show_overcommit,
5539                                 perf_set_overcommit
5540                         );
5541
5542 static struct attribute *perfclass_attrs[] = {
5543         &attr_reserve_percpu.attr,
5544         &attr_overcommit.attr,
5545         NULL
5546 };
5547
5548 static struct attribute_group perfclass_attr_group = {
5549         .attrs                  = perfclass_attrs,
5550         .name                   = "perf_events",
5551 };
5552
5553 static int __init perf_event_sysfs_init(void)
5554 {
5555         return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5556                                   &perfclass_attr_group);
5557 }
5558 device_initcall(perf_event_sysfs_init);