perf: Drop the obsolete profile naming for trace events
[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 __weak
2790 void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip, int skip)
2791 {
2792 }
2793
2794 /*
2795  * Output
2796  */
2797 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2798                               unsigned long offset, unsigned long head)
2799 {
2800         unsigned long mask;
2801
2802         if (!data->writable)
2803                 return true;
2804
2805         mask = perf_data_size(data) - 1;
2806
2807         offset = (offset - tail) & mask;
2808         head   = (head   - tail) & mask;
2809
2810         if ((int)(head - offset) < 0)
2811                 return false;
2812
2813         return true;
2814 }
2815
2816 static void perf_output_wakeup(struct perf_output_handle *handle)
2817 {
2818         atomic_set(&handle->data->poll, POLL_IN);
2819
2820         if (handle->nmi) {
2821                 handle->event->pending_wakeup = 1;
2822                 perf_pending_queue(&handle->event->pending,
2823                                    perf_pending_event);
2824         } else
2825                 perf_event_wakeup(handle->event);
2826 }
2827
2828 /*
2829  * Curious locking construct.
2830  *
2831  * We need to ensure a later event_id doesn't publish a head when a former
2832  * event_id isn't done writing. However since we need to deal with NMIs we
2833  * cannot fully serialize things.
2834  *
2835  * What we do is serialize between CPUs so we only have to deal with NMI
2836  * nesting on a single CPU.
2837  *
2838  * We only publish the head (and generate a wakeup) when the outer-most
2839  * event_id completes.
2840  */
2841 static void perf_output_lock(struct perf_output_handle *handle)
2842 {
2843         struct perf_mmap_data *data = handle->data;
2844         int cur, cpu = get_cpu();
2845
2846         handle->locked = 0;
2847
2848         for (;;) {
2849                 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2850                 if (cur == -1) {
2851                         handle->locked = 1;
2852                         break;
2853                 }
2854                 if (cur == cpu)
2855                         break;
2856
2857                 cpu_relax();
2858         }
2859 }
2860
2861 static void perf_output_unlock(struct perf_output_handle *handle)
2862 {
2863         struct perf_mmap_data *data = handle->data;
2864         unsigned long head;
2865         int cpu;
2866
2867         data->done_head = data->head;
2868
2869         if (!handle->locked)
2870                 goto out;
2871
2872 again:
2873         /*
2874          * The xchg implies a full barrier that ensures all writes are done
2875          * before we publish the new head, matched by a rmb() in userspace when
2876          * reading this position.
2877          */
2878         while ((head = atomic_long_xchg(&data->done_head, 0)))
2879                 data->user_page->data_head = head;
2880
2881         /*
2882          * NMI can happen here, which means we can miss a done_head update.
2883          */
2884
2885         cpu = atomic_xchg(&data->lock, -1);
2886         WARN_ON_ONCE(cpu != smp_processor_id());
2887
2888         /*
2889          * Therefore we have to validate we did not indeed do so.
2890          */
2891         if (unlikely(atomic_long_read(&data->done_head))) {
2892                 /*
2893                  * Since we had it locked, we can lock it again.
2894                  */
2895                 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2896                         cpu_relax();
2897
2898                 goto again;
2899         }
2900
2901         if (atomic_xchg(&data->wakeup, 0))
2902                 perf_output_wakeup(handle);
2903 out:
2904         put_cpu();
2905 }
2906
2907 void perf_output_copy(struct perf_output_handle *handle,
2908                       const void *buf, unsigned int len)
2909 {
2910         unsigned int pages_mask;
2911         unsigned long offset;
2912         unsigned int size;
2913         void **pages;
2914
2915         offset          = handle->offset;
2916         pages_mask      = handle->data->nr_pages - 1;
2917         pages           = handle->data->data_pages;
2918
2919         do {
2920                 unsigned long page_offset;
2921                 unsigned long page_size;
2922                 int nr;
2923
2924                 nr          = (offset >> PAGE_SHIFT) & pages_mask;
2925                 page_size   = 1UL << (handle->data->data_order + PAGE_SHIFT);
2926                 page_offset = offset & (page_size - 1);
2927                 size        = min_t(unsigned int, page_size - page_offset, len);
2928
2929                 memcpy(pages[nr] + page_offset, buf, size);
2930
2931                 len         -= size;
2932                 buf         += size;
2933                 offset      += size;
2934         } while (len);
2935
2936         handle->offset = offset;
2937
2938         /*
2939          * Check we didn't copy past our reservation window, taking the
2940          * possible unsigned int wrap into account.
2941          */
2942         WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2943 }
2944
2945 int perf_output_begin(struct perf_output_handle *handle,
2946                       struct perf_event *event, unsigned int size,
2947                       int nmi, int sample)
2948 {
2949         struct perf_event *output_event;
2950         struct perf_mmap_data *data;
2951         unsigned long tail, offset, head;
2952         int have_lost;
2953         struct {
2954                 struct perf_event_header header;
2955                 u64                      id;
2956                 u64                      lost;
2957         } lost_event;
2958
2959         rcu_read_lock();
2960         /*
2961          * For inherited events we send all the output towards the parent.
2962          */
2963         if (event->parent)
2964                 event = event->parent;
2965
2966         output_event = rcu_dereference(event->output);
2967         if (output_event)
2968                 event = output_event;
2969
2970         data = rcu_dereference(event->data);
2971         if (!data)
2972                 goto out;
2973
2974         handle->data    = data;
2975         handle->event   = event;
2976         handle->nmi     = nmi;
2977         handle->sample  = sample;
2978
2979         if (!data->nr_pages)
2980                 goto fail;
2981
2982         have_lost = atomic_read(&data->lost);
2983         if (have_lost)
2984                 size += sizeof(lost_event);
2985
2986         perf_output_lock(handle);
2987
2988         do {
2989                 /*
2990                  * Userspace could choose to issue a mb() before updating the
2991                  * tail pointer. So that all reads will be completed before the
2992                  * write is issued.
2993                  */
2994                 tail = ACCESS_ONCE(data->user_page->data_tail);
2995                 smp_rmb();
2996                 offset = head = atomic_long_read(&data->head);
2997                 head += size;
2998                 if (unlikely(!perf_output_space(data, tail, offset, head)))
2999                         goto fail;
3000         } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
3001
3002         handle->offset  = offset;
3003         handle->head    = head;
3004
3005         if (head - tail > data->watermark)
3006                 atomic_set(&data->wakeup, 1);
3007
3008         if (have_lost) {
3009                 lost_event.header.type = PERF_RECORD_LOST;
3010                 lost_event.header.misc = 0;
3011                 lost_event.header.size = sizeof(lost_event);
3012                 lost_event.id          = event->id;
3013                 lost_event.lost        = atomic_xchg(&data->lost, 0);
3014
3015                 perf_output_put(handle, lost_event);
3016         }
3017
3018         return 0;
3019
3020 fail:
3021         atomic_inc(&data->lost);
3022         perf_output_unlock(handle);
3023 out:
3024         rcu_read_unlock();
3025
3026         return -ENOSPC;
3027 }
3028
3029 void perf_output_end(struct perf_output_handle *handle)
3030 {
3031         struct perf_event *event = handle->event;
3032         struct perf_mmap_data *data = handle->data;
3033
3034         int wakeup_events = event->attr.wakeup_events;
3035
3036         if (handle->sample && wakeup_events) {
3037                 int events = atomic_inc_return(&data->events);
3038                 if (events >= wakeup_events) {
3039                         atomic_sub(wakeup_events, &data->events);
3040                         atomic_set(&data->wakeup, 1);
3041                 }
3042         }
3043
3044         perf_output_unlock(handle);
3045         rcu_read_unlock();
3046 }
3047
3048 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3049 {
3050         /*
3051          * only top level events have the pid namespace they were created in
3052          */
3053         if (event->parent)
3054                 event = event->parent;
3055
3056         return task_tgid_nr_ns(p, event->ns);
3057 }
3058
3059 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3060 {
3061         /*
3062          * only top level events have the pid namespace they were created in
3063          */
3064         if (event->parent)
3065                 event = event->parent;
3066
3067         return task_pid_nr_ns(p, event->ns);
3068 }
3069
3070 static void perf_output_read_one(struct perf_output_handle *handle,
3071                                  struct perf_event *event)
3072 {
3073         u64 read_format = event->attr.read_format;
3074         u64 values[4];
3075         int n = 0;
3076
3077         values[n++] = atomic64_read(&event->count);
3078         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3079                 values[n++] = event->total_time_enabled +
3080                         atomic64_read(&event->child_total_time_enabled);
3081         }
3082         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3083                 values[n++] = event->total_time_running +
3084                         atomic64_read(&event->child_total_time_running);
3085         }
3086         if (read_format & PERF_FORMAT_ID)
3087                 values[n++] = primary_event_id(event);
3088
3089         perf_output_copy(handle, values, n * sizeof(u64));
3090 }
3091
3092 /*
3093  * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3094  */
3095 static void perf_output_read_group(struct perf_output_handle *handle,
3096                             struct perf_event *event)
3097 {
3098         struct perf_event *leader = event->group_leader, *sub;
3099         u64 read_format = event->attr.read_format;
3100         u64 values[5];
3101         int n = 0;
3102
3103         values[n++] = 1 + leader->nr_siblings;
3104
3105         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3106                 values[n++] = leader->total_time_enabled;
3107
3108         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3109                 values[n++] = leader->total_time_running;
3110
3111         if (leader != event)
3112                 leader->pmu->read(leader);
3113
3114         values[n++] = atomic64_read(&leader->count);
3115         if (read_format & PERF_FORMAT_ID)
3116                 values[n++] = primary_event_id(leader);
3117
3118         perf_output_copy(handle, values, n * sizeof(u64));
3119
3120         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3121                 n = 0;
3122
3123                 if (sub != event)
3124                         sub->pmu->read(sub);
3125
3126                 values[n++] = atomic64_read(&sub->count);
3127                 if (read_format & PERF_FORMAT_ID)
3128                         values[n++] = primary_event_id(sub);
3129
3130                 perf_output_copy(handle, values, n * sizeof(u64));
3131         }
3132 }
3133
3134 static void perf_output_read(struct perf_output_handle *handle,
3135                              struct perf_event *event)
3136 {
3137         if (event->attr.read_format & PERF_FORMAT_GROUP)
3138                 perf_output_read_group(handle, event);
3139         else
3140                 perf_output_read_one(handle, event);
3141 }
3142
3143 void perf_output_sample(struct perf_output_handle *handle,
3144                         struct perf_event_header *header,
3145                         struct perf_sample_data *data,
3146                         struct perf_event *event)
3147 {
3148         u64 sample_type = data->type;
3149
3150         perf_output_put(handle, *header);
3151
3152         if (sample_type & PERF_SAMPLE_IP)
3153                 perf_output_put(handle, data->ip);
3154
3155         if (sample_type & PERF_SAMPLE_TID)
3156                 perf_output_put(handle, data->tid_entry);
3157
3158         if (sample_type & PERF_SAMPLE_TIME)
3159                 perf_output_put(handle, data->time);
3160
3161         if (sample_type & PERF_SAMPLE_ADDR)
3162                 perf_output_put(handle, data->addr);
3163
3164         if (sample_type & PERF_SAMPLE_ID)
3165                 perf_output_put(handle, data->id);
3166
3167         if (sample_type & PERF_SAMPLE_STREAM_ID)
3168                 perf_output_put(handle, data->stream_id);
3169
3170         if (sample_type & PERF_SAMPLE_CPU)
3171                 perf_output_put(handle, data->cpu_entry);
3172
3173         if (sample_type & PERF_SAMPLE_PERIOD)
3174                 perf_output_put(handle, data->period);
3175
3176         if (sample_type & PERF_SAMPLE_READ)
3177                 perf_output_read(handle, event);
3178
3179         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3180                 if (data->callchain) {
3181                         int size = 1;
3182
3183                         if (data->callchain)
3184                                 size += data->callchain->nr;
3185
3186                         size *= sizeof(u64);
3187
3188                         perf_output_copy(handle, data->callchain, size);
3189                 } else {
3190                         u64 nr = 0;
3191                         perf_output_put(handle, nr);
3192                 }
3193         }
3194
3195         if (sample_type & PERF_SAMPLE_RAW) {
3196                 if (data->raw) {
3197                         perf_output_put(handle, data->raw->size);
3198                         perf_output_copy(handle, data->raw->data,
3199                                          data->raw->size);
3200                 } else {
3201                         struct {
3202                                 u32     size;
3203                                 u32     data;
3204                         } raw = {
3205                                 .size = sizeof(u32),
3206                                 .data = 0,
3207                         };
3208                         perf_output_put(handle, raw);
3209                 }
3210         }
3211 }
3212
3213 void perf_prepare_sample(struct perf_event_header *header,
3214                          struct perf_sample_data *data,
3215                          struct perf_event *event,
3216                          struct pt_regs *regs)
3217 {
3218         u64 sample_type = event->attr.sample_type;
3219
3220         data->type = sample_type;
3221
3222         header->type = PERF_RECORD_SAMPLE;
3223         header->size = sizeof(*header);
3224
3225         header->misc = 0;
3226         header->misc |= perf_misc_flags(regs);
3227
3228         if (sample_type & PERF_SAMPLE_IP) {
3229                 data->ip = perf_instruction_pointer(regs);
3230
3231                 header->size += sizeof(data->ip);
3232         }
3233
3234         if (sample_type & PERF_SAMPLE_TID) {
3235                 /* namespace issues */
3236                 data->tid_entry.pid = perf_event_pid(event, current);
3237                 data->tid_entry.tid = perf_event_tid(event, current);
3238
3239                 header->size += sizeof(data->tid_entry);
3240         }
3241
3242         if (sample_type & PERF_SAMPLE_TIME) {
3243                 data->time = perf_clock();
3244
3245                 header->size += sizeof(data->time);
3246         }
3247
3248         if (sample_type & PERF_SAMPLE_ADDR)
3249                 header->size += sizeof(data->addr);
3250
3251         if (sample_type & PERF_SAMPLE_ID) {
3252                 data->id = primary_event_id(event);
3253
3254                 header->size += sizeof(data->id);
3255         }
3256
3257         if (sample_type & PERF_SAMPLE_STREAM_ID) {
3258                 data->stream_id = event->id;
3259
3260                 header->size += sizeof(data->stream_id);
3261         }
3262
3263         if (sample_type & PERF_SAMPLE_CPU) {
3264                 data->cpu_entry.cpu             = raw_smp_processor_id();
3265                 data->cpu_entry.reserved        = 0;
3266
3267                 header->size += sizeof(data->cpu_entry);
3268         }
3269
3270         if (sample_type & PERF_SAMPLE_PERIOD)
3271                 header->size += sizeof(data->period);
3272
3273         if (sample_type & PERF_SAMPLE_READ)
3274                 header->size += perf_event_read_size(event);
3275
3276         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3277                 int size = 1;
3278
3279                 data->callchain = perf_callchain(regs);
3280
3281                 if (data->callchain)
3282                         size += data->callchain->nr;
3283
3284                 header->size += size * sizeof(u64);
3285         }
3286
3287         if (sample_type & PERF_SAMPLE_RAW) {
3288                 int size = sizeof(u32);
3289
3290                 if (data->raw)
3291                         size += data->raw->size;
3292                 else
3293                         size += sizeof(u32);
3294
3295                 WARN_ON_ONCE(size & (sizeof(u64)-1));
3296                 header->size += size;
3297         }
3298 }
3299
3300 static void perf_event_output(struct perf_event *event, int nmi,
3301                                 struct perf_sample_data *data,
3302                                 struct pt_regs *regs)
3303 {
3304         struct perf_output_handle handle;
3305         struct perf_event_header header;
3306
3307         perf_prepare_sample(&header, data, event, regs);
3308
3309         if (perf_output_begin(&handle, event, header.size, nmi, 1))
3310                 return;
3311
3312         perf_output_sample(&handle, &header, data, event);
3313
3314         perf_output_end(&handle);
3315 }
3316
3317 /*
3318  * read event_id
3319  */
3320
3321 struct perf_read_event {
3322         struct perf_event_header        header;
3323
3324         u32                             pid;
3325         u32                             tid;
3326 };
3327
3328 static void
3329 perf_event_read_event(struct perf_event *event,
3330                         struct task_struct *task)
3331 {
3332         struct perf_output_handle handle;
3333         struct perf_read_event read_event = {
3334                 .header = {
3335                         .type = PERF_RECORD_READ,
3336                         .misc = 0,
3337                         .size = sizeof(read_event) + perf_event_read_size(event),
3338                 },
3339                 .pid = perf_event_pid(event, task),
3340                 .tid = perf_event_tid(event, task),
3341         };
3342         int ret;
3343
3344         ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3345         if (ret)
3346                 return;
3347
3348         perf_output_put(&handle, read_event);
3349         perf_output_read(&handle, event);
3350
3351         perf_output_end(&handle);
3352 }
3353
3354 /*
3355  * task tracking -- fork/exit
3356  *
3357  * enabled by: attr.comm | attr.mmap | attr.task
3358  */
3359
3360 struct perf_task_event {
3361         struct task_struct              *task;
3362         struct perf_event_context       *task_ctx;
3363
3364         struct {
3365                 struct perf_event_header        header;
3366
3367                 u32                             pid;
3368                 u32                             ppid;
3369                 u32                             tid;
3370                 u32                             ptid;
3371                 u64                             time;
3372         } event_id;
3373 };
3374
3375 static void perf_event_task_output(struct perf_event *event,
3376                                      struct perf_task_event *task_event)
3377 {
3378         struct perf_output_handle handle;
3379         int size;
3380         struct task_struct *task = task_event->task;
3381         int ret;
3382
3383         size  = task_event->event_id.header.size;
3384         ret = perf_output_begin(&handle, event, size, 0, 0);
3385
3386         if (ret)
3387                 return;
3388
3389         task_event->event_id.pid = perf_event_pid(event, task);
3390         task_event->event_id.ppid = perf_event_pid(event, current);
3391
3392         task_event->event_id.tid = perf_event_tid(event, task);
3393         task_event->event_id.ptid = perf_event_tid(event, current);
3394
3395         perf_output_put(&handle, task_event->event_id);
3396
3397         perf_output_end(&handle);
3398 }
3399
3400 static int perf_event_task_match(struct perf_event *event)
3401 {
3402         if (event->state < PERF_EVENT_STATE_INACTIVE)
3403                 return 0;
3404
3405         if (event->cpu != -1 && event->cpu != smp_processor_id())
3406                 return 0;
3407
3408         if (event->attr.comm || event->attr.mmap || event->attr.task)
3409                 return 1;
3410
3411         return 0;
3412 }
3413
3414 static void perf_event_task_ctx(struct perf_event_context *ctx,
3415                                   struct perf_task_event *task_event)
3416 {
3417         struct perf_event *event;
3418
3419         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3420                 if (perf_event_task_match(event))
3421                         perf_event_task_output(event, task_event);
3422         }
3423 }
3424
3425 static void perf_event_task_event(struct perf_task_event *task_event)
3426 {
3427         struct perf_cpu_context *cpuctx;
3428         struct perf_event_context *ctx = task_event->task_ctx;
3429
3430         rcu_read_lock();
3431         cpuctx = &get_cpu_var(perf_cpu_context);
3432         perf_event_task_ctx(&cpuctx->ctx, task_event);
3433         if (!ctx)
3434                 ctx = rcu_dereference(current->perf_event_ctxp);
3435         if (ctx)
3436                 perf_event_task_ctx(ctx, task_event);
3437         put_cpu_var(perf_cpu_context);
3438         rcu_read_unlock();
3439 }
3440
3441 static void perf_event_task(struct task_struct *task,
3442                               struct perf_event_context *task_ctx,
3443                               int new)
3444 {
3445         struct perf_task_event task_event;
3446
3447         if (!atomic_read(&nr_comm_events) &&
3448             !atomic_read(&nr_mmap_events) &&
3449             !atomic_read(&nr_task_events))
3450                 return;
3451
3452         task_event = (struct perf_task_event){
3453                 .task     = task,
3454                 .task_ctx = task_ctx,
3455                 .event_id    = {
3456                         .header = {
3457                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3458                                 .misc = 0,
3459                                 .size = sizeof(task_event.event_id),
3460                         },
3461                         /* .pid  */
3462                         /* .ppid */
3463                         /* .tid  */
3464                         /* .ptid */
3465                         .time = perf_clock(),
3466                 },
3467         };
3468
3469         perf_event_task_event(&task_event);
3470 }
3471
3472 void perf_event_fork(struct task_struct *task)
3473 {
3474         perf_event_task(task, NULL, 1);
3475 }
3476
3477 /*
3478  * comm tracking
3479  */
3480
3481 struct perf_comm_event {
3482         struct task_struct      *task;
3483         char                    *comm;
3484         int                     comm_size;
3485
3486         struct {
3487                 struct perf_event_header        header;
3488
3489                 u32                             pid;
3490                 u32                             tid;
3491         } event_id;
3492 };
3493
3494 static void perf_event_comm_output(struct perf_event *event,
3495                                      struct perf_comm_event *comm_event)
3496 {
3497         struct perf_output_handle handle;
3498         int size = comm_event->event_id.header.size;
3499         int ret = perf_output_begin(&handle, event, size, 0, 0);
3500
3501         if (ret)
3502                 return;
3503
3504         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3505         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3506
3507         perf_output_put(&handle, comm_event->event_id);
3508         perf_output_copy(&handle, comm_event->comm,
3509                                    comm_event->comm_size);
3510         perf_output_end(&handle);
3511 }
3512
3513 static int perf_event_comm_match(struct perf_event *event)
3514 {
3515         if (event->state < PERF_EVENT_STATE_INACTIVE)
3516                 return 0;
3517
3518         if (event->cpu != -1 && event->cpu != smp_processor_id())
3519                 return 0;
3520
3521         if (event->attr.comm)
3522                 return 1;
3523
3524         return 0;
3525 }
3526
3527 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3528                                   struct perf_comm_event *comm_event)
3529 {
3530         struct perf_event *event;
3531
3532         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3533                 if (perf_event_comm_match(event))
3534                         perf_event_comm_output(event, comm_event);
3535         }
3536 }
3537
3538 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3539 {
3540         struct perf_cpu_context *cpuctx;
3541         struct perf_event_context *ctx;
3542         unsigned int size;
3543         char comm[TASK_COMM_LEN];
3544
3545         memset(comm, 0, sizeof(comm));
3546         strlcpy(comm, comm_event->task->comm, sizeof(comm));
3547         size = ALIGN(strlen(comm)+1, sizeof(u64));
3548
3549         comm_event->comm = comm;
3550         comm_event->comm_size = size;
3551
3552         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3553
3554         rcu_read_lock();
3555         cpuctx = &get_cpu_var(perf_cpu_context);
3556         perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3557         ctx = rcu_dereference(current->perf_event_ctxp);
3558         if (ctx)
3559                 perf_event_comm_ctx(ctx, comm_event);
3560         put_cpu_var(perf_cpu_context);
3561         rcu_read_unlock();
3562 }
3563
3564 void perf_event_comm(struct task_struct *task)
3565 {
3566         struct perf_comm_event comm_event;
3567
3568         if (task->perf_event_ctxp)
3569                 perf_event_enable_on_exec(task);
3570
3571         if (!atomic_read(&nr_comm_events))
3572                 return;
3573
3574         comm_event = (struct perf_comm_event){
3575                 .task   = task,
3576                 /* .comm      */
3577                 /* .comm_size */
3578                 .event_id  = {
3579                         .header = {
3580                                 .type = PERF_RECORD_COMM,
3581                                 .misc = 0,
3582                                 /* .size */
3583                         },
3584                         /* .pid */
3585                         /* .tid */
3586                 },
3587         };
3588
3589         perf_event_comm_event(&comm_event);
3590 }
3591
3592 /*
3593  * mmap tracking
3594  */
3595
3596 struct perf_mmap_event {
3597         struct vm_area_struct   *vma;
3598
3599         const char              *file_name;
3600         int                     file_size;
3601
3602         struct {
3603                 struct perf_event_header        header;
3604
3605                 u32                             pid;
3606                 u32                             tid;
3607                 u64                             start;
3608                 u64                             len;
3609                 u64                             pgoff;
3610         } event_id;
3611 };
3612
3613 static void perf_event_mmap_output(struct perf_event *event,
3614                                      struct perf_mmap_event *mmap_event)
3615 {
3616         struct perf_output_handle handle;
3617         int size = mmap_event->event_id.header.size;
3618         int ret = perf_output_begin(&handle, event, size, 0, 0);
3619
3620         if (ret)
3621                 return;
3622
3623         mmap_event->event_id.pid = perf_event_pid(event, current);
3624         mmap_event->event_id.tid = perf_event_tid(event, current);
3625
3626         perf_output_put(&handle, mmap_event->event_id);
3627         perf_output_copy(&handle, mmap_event->file_name,
3628                                    mmap_event->file_size);
3629         perf_output_end(&handle);
3630 }
3631
3632 static int perf_event_mmap_match(struct perf_event *event,
3633                                    struct perf_mmap_event *mmap_event)
3634 {
3635         if (event->state < PERF_EVENT_STATE_INACTIVE)
3636                 return 0;
3637
3638         if (event->cpu != -1 && event->cpu != smp_processor_id())
3639                 return 0;
3640
3641         if (event->attr.mmap)
3642                 return 1;
3643
3644         return 0;
3645 }
3646
3647 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3648                                   struct perf_mmap_event *mmap_event)
3649 {
3650         struct perf_event *event;
3651
3652         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3653                 if (perf_event_mmap_match(event, mmap_event))
3654                         perf_event_mmap_output(event, mmap_event);
3655         }
3656 }
3657
3658 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3659 {
3660         struct perf_cpu_context *cpuctx;
3661         struct perf_event_context *ctx;
3662         struct vm_area_struct *vma = mmap_event->vma;
3663         struct file *file = vma->vm_file;
3664         unsigned int size;
3665         char tmp[16];
3666         char *buf = NULL;
3667         const char *name;
3668
3669         memset(tmp, 0, sizeof(tmp));
3670
3671         if (file) {
3672                 /*
3673                  * d_path works from the end of the buffer backwards, so we
3674                  * need to add enough zero bytes after the string to handle
3675                  * the 64bit alignment we do later.
3676                  */
3677                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3678                 if (!buf) {
3679                         name = strncpy(tmp, "//enomem", sizeof(tmp));
3680                         goto got_name;
3681                 }
3682                 name = d_path(&file->f_path, buf, PATH_MAX);
3683                 if (IS_ERR(name)) {
3684                         name = strncpy(tmp, "//toolong", sizeof(tmp));
3685                         goto got_name;
3686                 }
3687         } else {
3688                 if (arch_vma_name(mmap_event->vma)) {
3689                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3690                                        sizeof(tmp));
3691                         goto got_name;
3692                 }
3693
3694                 if (!vma->vm_mm) {
3695                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
3696                         goto got_name;
3697                 }
3698
3699                 name = strncpy(tmp, "//anon", sizeof(tmp));
3700                 goto got_name;
3701         }
3702
3703 got_name:
3704         size = ALIGN(strlen(name)+1, sizeof(u64));
3705
3706         mmap_event->file_name = name;
3707         mmap_event->file_size = size;
3708
3709         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3710
3711         rcu_read_lock();
3712         cpuctx = &get_cpu_var(perf_cpu_context);
3713         perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3714         ctx = rcu_dereference(current->perf_event_ctxp);
3715         if (ctx)
3716                 perf_event_mmap_ctx(ctx, mmap_event);
3717         put_cpu_var(perf_cpu_context);
3718         rcu_read_unlock();
3719
3720         kfree(buf);
3721 }
3722
3723 void __perf_event_mmap(struct vm_area_struct *vma)
3724 {
3725         struct perf_mmap_event mmap_event;
3726
3727         if (!atomic_read(&nr_mmap_events))
3728                 return;
3729
3730         mmap_event = (struct perf_mmap_event){
3731                 .vma    = vma,
3732                 /* .file_name */
3733                 /* .file_size */
3734                 .event_id  = {
3735                         .header = {
3736                                 .type = PERF_RECORD_MMAP,
3737                                 .misc = 0,
3738                                 /* .size */
3739                         },
3740                         /* .pid */
3741                         /* .tid */
3742                         .start  = vma->vm_start,
3743                         .len    = vma->vm_end - vma->vm_start,
3744                         .pgoff  = (u64)vma->vm_pgoff << PAGE_SHIFT,
3745                 },
3746         };
3747
3748         perf_event_mmap_event(&mmap_event);
3749 }
3750
3751 /*
3752  * IRQ throttle logging
3753  */
3754
3755 static void perf_log_throttle(struct perf_event *event, int enable)
3756 {
3757         struct perf_output_handle handle;
3758         int ret;
3759
3760         struct {
3761                 struct perf_event_header        header;
3762                 u64                             time;
3763                 u64                             id;
3764                 u64                             stream_id;
3765         } throttle_event = {
3766                 .header = {
3767                         .type = PERF_RECORD_THROTTLE,
3768                         .misc = 0,
3769                         .size = sizeof(throttle_event),
3770                 },
3771                 .time           = perf_clock(),
3772                 .id             = primary_event_id(event),
3773                 .stream_id      = event->id,
3774         };
3775
3776         if (enable)
3777                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3778
3779         ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3780         if (ret)
3781                 return;
3782
3783         perf_output_put(&handle, throttle_event);
3784         perf_output_end(&handle);
3785 }
3786
3787 /*
3788  * Generic event overflow handling, sampling.
3789  */
3790
3791 static int __perf_event_overflow(struct perf_event *event, int nmi,
3792                                    int throttle, struct perf_sample_data *data,
3793                                    struct pt_regs *regs)
3794 {
3795         int events = atomic_read(&event->event_limit);
3796         struct hw_perf_event *hwc = &event->hw;
3797         int ret = 0;
3798
3799         throttle = (throttle && event->pmu->unthrottle != NULL);
3800
3801         if (!throttle) {
3802                 hwc->interrupts++;
3803         } else {
3804                 if (hwc->interrupts != MAX_INTERRUPTS) {
3805                         hwc->interrupts++;
3806                         if (HZ * hwc->interrupts >
3807                                         (u64)sysctl_perf_event_sample_rate) {
3808                                 hwc->interrupts = MAX_INTERRUPTS;
3809                                 perf_log_throttle(event, 0);
3810                                 ret = 1;
3811                         }
3812                 } else {
3813                         /*
3814                          * Keep re-disabling events even though on the previous
3815                          * pass we disabled it - just in case we raced with a
3816                          * sched-in and the event got enabled again:
3817                          */
3818                         ret = 1;
3819                 }
3820         }
3821
3822         if (event->attr.freq) {
3823                 u64 now = perf_clock();
3824                 s64 delta = now - hwc->freq_time_stamp;
3825
3826                 hwc->freq_time_stamp = now;
3827
3828                 if (delta > 0 && delta < 2*TICK_NSEC)
3829                         perf_adjust_period(event, delta, hwc->last_period);
3830         }
3831
3832         /*
3833          * XXX event_limit might not quite work as expected on inherited
3834          * events
3835          */
3836
3837         event->pending_kill = POLL_IN;
3838         if (events && atomic_dec_and_test(&event->event_limit)) {
3839                 ret = 1;
3840                 event->pending_kill = POLL_HUP;
3841                 if (nmi) {
3842                         event->pending_disable = 1;
3843                         perf_pending_queue(&event->pending,
3844                                            perf_pending_event);
3845                 } else
3846                         perf_event_disable(event);
3847         }
3848
3849         if (event->overflow_handler)
3850                 event->overflow_handler(event, nmi, data, regs);
3851         else
3852                 perf_event_output(event, nmi, data, regs);
3853
3854         return ret;
3855 }
3856
3857 int perf_event_overflow(struct perf_event *event, int nmi,
3858                           struct perf_sample_data *data,
3859                           struct pt_regs *regs)
3860 {
3861         return __perf_event_overflow(event, nmi, 1, data, regs);
3862 }
3863
3864 /*
3865  * Generic software event infrastructure
3866  */
3867
3868 /*
3869  * We directly increment event->count and keep a second value in
3870  * event->hw.period_left to count intervals. This period event
3871  * is kept in the range [-sample_period, 0] so that we can use the
3872  * sign as trigger.
3873  */
3874
3875 static u64 perf_swevent_set_period(struct perf_event *event)
3876 {
3877         struct hw_perf_event *hwc = &event->hw;
3878         u64 period = hwc->last_period;
3879         u64 nr, offset;
3880         s64 old, val;
3881
3882         hwc->last_period = hwc->sample_period;
3883
3884 again:
3885         old = val = atomic64_read(&hwc->period_left);
3886         if (val < 0)
3887                 return 0;
3888
3889         nr = div64_u64(period + val, period);
3890         offset = nr * period;
3891         val -= offset;
3892         if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3893                 goto again;
3894
3895         return nr;
3896 }
3897
3898 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3899                                     int nmi, struct perf_sample_data *data,
3900                                     struct pt_regs *regs)
3901 {
3902         struct hw_perf_event *hwc = &event->hw;
3903         int throttle = 0;
3904
3905         data->period = event->hw.last_period;
3906         if (!overflow)
3907                 overflow = perf_swevent_set_period(event);
3908
3909         if (hwc->interrupts == MAX_INTERRUPTS)
3910                 return;
3911
3912         for (; overflow; overflow--) {
3913                 if (__perf_event_overflow(event, nmi, throttle,
3914                                             data, regs)) {
3915                         /*
3916                          * We inhibit the overflow from happening when
3917                          * hwc->interrupts == MAX_INTERRUPTS.
3918                          */
3919                         break;
3920                 }
3921                 throttle = 1;
3922         }
3923 }
3924
3925 static void perf_swevent_unthrottle(struct perf_event *event)
3926 {
3927         /*
3928          * Nothing to do, we already reset hwc->interrupts.
3929          */
3930 }
3931
3932 static void perf_swevent_add(struct perf_event *event, u64 nr,
3933                                int nmi, struct perf_sample_data *data,
3934                                struct pt_regs *regs)
3935 {
3936         struct hw_perf_event *hwc = &event->hw;
3937
3938         atomic64_add(nr, &event->count);
3939
3940         if (!regs)
3941                 return;
3942
3943         if (!hwc->sample_period)
3944                 return;
3945
3946         if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3947                 return perf_swevent_overflow(event, 1, nmi, data, regs);
3948
3949         if (atomic64_add_negative(nr, &hwc->period_left))
3950                 return;
3951
3952         perf_swevent_overflow(event, 0, nmi, data, regs);
3953 }
3954
3955 static int perf_swevent_is_counting(struct perf_event *event)
3956 {
3957         /*
3958          * The event is active, we're good!
3959          */
3960         if (event->state == PERF_EVENT_STATE_ACTIVE)
3961                 return 1;
3962
3963         /*
3964          * The event is off/error, not counting.
3965          */
3966         if (event->state != PERF_EVENT_STATE_INACTIVE)
3967                 return 0;
3968
3969         /*
3970          * The event is inactive, if the context is active
3971          * we're part of a group that didn't make it on the 'pmu',
3972          * not counting.
3973          */
3974         if (event->ctx->is_active)
3975                 return 0;
3976
3977         /*
3978          * We're inactive and the context is too, this means the
3979          * task is scheduled out, we're counting events that happen
3980          * to us, like migration events.
3981          */
3982         return 1;
3983 }
3984
3985 static int perf_tp_event_match(struct perf_event *event,
3986                                 struct perf_sample_data *data);
3987
3988 static int perf_exclude_event(struct perf_event *event,
3989                               struct pt_regs *regs)
3990 {
3991         if (regs) {
3992                 if (event->attr.exclude_user && user_mode(regs))
3993                         return 1;
3994
3995                 if (event->attr.exclude_kernel && !user_mode(regs))
3996                         return 1;
3997         }
3998
3999         return 0;
4000 }
4001
4002 static int perf_swevent_match(struct perf_event *event,
4003                                 enum perf_type_id type,
4004                                 u32 event_id,
4005                                 struct perf_sample_data *data,
4006                                 struct pt_regs *regs)
4007 {
4008         if (event->cpu != -1 && event->cpu != smp_processor_id())
4009                 return 0;
4010
4011         if (!perf_swevent_is_counting(event))
4012                 return 0;
4013
4014         if (event->attr.type != type)
4015                 return 0;
4016
4017         if (event->attr.config != event_id)
4018                 return 0;
4019
4020         if (perf_exclude_event(event, regs))
4021                 return 0;
4022
4023         if (event->attr.type == PERF_TYPE_TRACEPOINT &&
4024             !perf_tp_event_match(event, data))
4025                 return 0;
4026
4027         return 1;
4028 }
4029
4030 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
4031                                      enum perf_type_id type,
4032                                      u32 event_id, u64 nr, int nmi,
4033                                      struct perf_sample_data *data,
4034                                      struct pt_regs *regs)
4035 {
4036         struct perf_event *event;
4037
4038         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4039                 if (perf_swevent_match(event, type, event_id, data, regs))
4040                         perf_swevent_add(event, nr, nmi, data, regs);
4041         }
4042 }
4043
4044 int perf_swevent_get_recursion_context(void)
4045 {
4046         struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
4047         int rctx;
4048
4049         if (in_nmi())
4050                 rctx = 3;
4051         else if (in_irq())
4052                 rctx = 2;
4053         else if (in_softirq())
4054                 rctx = 1;
4055         else
4056                 rctx = 0;
4057
4058         if (cpuctx->recursion[rctx]) {
4059                 put_cpu_var(perf_cpu_context);
4060                 return -1;
4061         }
4062
4063         cpuctx->recursion[rctx]++;
4064         barrier();
4065
4066         return rctx;
4067 }
4068 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4069
4070 void perf_swevent_put_recursion_context(int rctx)
4071 {
4072         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4073         barrier();
4074         cpuctx->recursion[rctx]--;
4075         put_cpu_var(perf_cpu_context);
4076 }
4077 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
4078
4079 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4080                                     u64 nr, int nmi,
4081                                     struct perf_sample_data *data,
4082                                     struct pt_regs *regs)
4083 {
4084         struct perf_cpu_context *cpuctx;
4085         struct perf_event_context *ctx;
4086
4087         cpuctx = &__get_cpu_var(perf_cpu_context);
4088         rcu_read_lock();
4089         perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
4090                                  nr, nmi, data, regs);
4091         /*
4092          * doesn't really matter which of the child contexts the
4093          * events ends up in.
4094          */
4095         ctx = rcu_dereference(current->perf_event_ctxp);
4096         if (ctx)
4097                 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
4098         rcu_read_unlock();
4099 }
4100
4101 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4102                             struct pt_regs *regs, u64 addr)
4103 {
4104         struct perf_sample_data data;
4105         int rctx;
4106
4107         rctx = perf_swevent_get_recursion_context();
4108         if (rctx < 0)
4109                 return;
4110
4111         perf_sample_data_init(&data, addr);
4112
4113         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4114
4115         perf_swevent_put_recursion_context(rctx);
4116 }
4117
4118 static void perf_swevent_read(struct perf_event *event)
4119 {
4120 }
4121
4122 static int perf_swevent_enable(struct perf_event *event)
4123 {
4124         struct hw_perf_event *hwc = &event->hw;
4125
4126         if (hwc->sample_period) {
4127                 hwc->last_period = hwc->sample_period;
4128                 perf_swevent_set_period(event);
4129         }
4130         return 0;
4131 }
4132
4133 static void perf_swevent_disable(struct perf_event *event)
4134 {
4135 }
4136
4137 static const struct pmu perf_ops_generic = {
4138         .enable         = perf_swevent_enable,
4139         .disable        = perf_swevent_disable,
4140         .read           = perf_swevent_read,
4141         .unthrottle     = perf_swevent_unthrottle,
4142 };
4143
4144 /*
4145  * hrtimer based swevent callback
4146  */
4147
4148 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4149 {
4150         enum hrtimer_restart ret = HRTIMER_RESTART;
4151         struct perf_sample_data data;
4152         struct pt_regs *regs;
4153         struct perf_event *event;
4154         u64 period;
4155
4156         event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4157         event->pmu->read(event);
4158
4159         perf_sample_data_init(&data, 0);
4160         data.period = event->hw.last_period;
4161         regs = get_irq_regs();
4162         /*
4163          * In case we exclude kernel IPs or are somehow not in interrupt
4164          * context, provide the next best thing, the user IP.
4165          */
4166         if ((event->attr.exclude_kernel || !regs) &&
4167                         !event->attr.exclude_user)
4168                 regs = task_pt_regs(current);
4169
4170         if (regs) {
4171                 if (!(event->attr.exclude_idle && current->pid == 0))
4172                         if (perf_event_overflow(event, 0, &data, regs))
4173                                 ret = HRTIMER_NORESTART;
4174         }
4175
4176         period = max_t(u64, 10000, event->hw.sample_period);
4177         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4178
4179         return ret;
4180 }
4181
4182 static void perf_swevent_start_hrtimer(struct perf_event *event)
4183 {
4184         struct hw_perf_event *hwc = &event->hw;
4185
4186         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4187         hwc->hrtimer.function = perf_swevent_hrtimer;
4188         if (hwc->sample_period) {
4189                 u64 period;
4190
4191                 if (hwc->remaining) {
4192                         if (hwc->remaining < 0)
4193                                 period = 10000;
4194                         else
4195                                 period = hwc->remaining;
4196                         hwc->remaining = 0;
4197                 } else {
4198                         period = max_t(u64, 10000, hwc->sample_period);
4199                 }
4200                 __hrtimer_start_range_ns(&hwc->hrtimer,
4201                                 ns_to_ktime(period), 0,
4202                                 HRTIMER_MODE_REL, 0);
4203         }
4204 }
4205
4206 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4207 {
4208         struct hw_perf_event *hwc = &event->hw;
4209
4210         if (hwc->sample_period) {
4211                 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4212                 hwc->remaining = ktime_to_ns(remaining);
4213
4214                 hrtimer_cancel(&hwc->hrtimer);
4215         }
4216 }
4217
4218 /*
4219  * Software event: cpu wall time clock
4220  */
4221
4222 static void cpu_clock_perf_event_update(struct perf_event *event)
4223 {
4224         int cpu = raw_smp_processor_id();
4225         s64 prev;
4226         u64 now;
4227
4228         now = cpu_clock(cpu);
4229         prev = atomic64_xchg(&event->hw.prev_count, now);
4230         atomic64_add(now - prev, &event->count);
4231 }
4232
4233 static int cpu_clock_perf_event_enable(struct perf_event *event)
4234 {
4235         struct hw_perf_event *hwc = &event->hw;
4236         int cpu = raw_smp_processor_id();
4237
4238         atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4239         perf_swevent_start_hrtimer(event);
4240
4241         return 0;
4242 }
4243
4244 static void cpu_clock_perf_event_disable(struct perf_event *event)
4245 {
4246         perf_swevent_cancel_hrtimer(event);
4247         cpu_clock_perf_event_update(event);
4248 }
4249
4250 static void cpu_clock_perf_event_read(struct perf_event *event)
4251 {
4252         cpu_clock_perf_event_update(event);
4253 }
4254
4255 static const struct pmu perf_ops_cpu_clock = {
4256         .enable         = cpu_clock_perf_event_enable,
4257         .disable        = cpu_clock_perf_event_disable,
4258         .read           = cpu_clock_perf_event_read,
4259 };
4260
4261 /*
4262  * Software event: task time clock
4263  */
4264
4265 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4266 {
4267         u64 prev;
4268         s64 delta;
4269
4270         prev = atomic64_xchg(&event->hw.prev_count, now);
4271         delta = now - prev;
4272         atomic64_add(delta, &event->count);
4273 }
4274
4275 static int task_clock_perf_event_enable(struct perf_event *event)
4276 {
4277         struct hw_perf_event *hwc = &event->hw;
4278         u64 now;
4279
4280         now = event->ctx->time;
4281
4282         atomic64_set(&hwc->prev_count, now);
4283
4284         perf_swevent_start_hrtimer(event);
4285
4286         return 0;
4287 }
4288
4289 static void task_clock_perf_event_disable(struct perf_event *event)
4290 {
4291         perf_swevent_cancel_hrtimer(event);
4292         task_clock_perf_event_update(event, event->ctx->time);
4293
4294 }
4295
4296 static void task_clock_perf_event_read(struct perf_event *event)
4297 {
4298         u64 time;
4299
4300         if (!in_nmi()) {
4301                 update_context_time(event->ctx);
4302                 time = event->ctx->time;
4303         } else {
4304                 u64 now = perf_clock();
4305                 u64 delta = now - event->ctx->timestamp;
4306                 time = event->ctx->time + delta;
4307         }
4308
4309         task_clock_perf_event_update(event, time);
4310 }
4311
4312 static const struct pmu perf_ops_task_clock = {
4313         .enable         = task_clock_perf_event_enable,
4314         .disable        = task_clock_perf_event_disable,
4315         .read           = task_clock_perf_event_read,
4316 };
4317
4318 #ifdef CONFIG_EVENT_TRACING
4319
4320 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4321                    int entry_size, struct pt_regs *regs)
4322 {
4323         struct perf_sample_data data;
4324         struct perf_raw_record raw = {
4325                 .size = entry_size,
4326                 .data = record,
4327         };
4328
4329         perf_sample_data_init(&data, addr);
4330         data.raw = &raw;
4331
4332         /* Trace events already protected against recursion */
4333         do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4334                          &data, regs);
4335 }
4336 EXPORT_SYMBOL_GPL(perf_tp_event);
4337
4338 static int perf_tp_event_match(struct perf_event *event,
4339                                 struct perf_sample_data *data)
4340 {
4341         void *record = data->raw->data;
4342
4343         if (likely(!event->filter) || filter_match_preds(event->filter, record))
4344                 return 1;
4345         return 0;
4346 }
4347
4348 static void tp_perf_event_destroy(struct perf_event *event)
4349 {
4350         perf_trace_disable(event->attr.config);
4351 }
4352
4353 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4354 {
4355         /*
4356          * Raw tracepoint data is a severe data leak, only allow root to
4357          * have these.
4358          */
4359         if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4360                         perf_paranoid_tracepoint_raw() &&
4361                         !capable(CAP_SYS_ADMIN))
4362                 return ERR_PTR(-EPERM);
4363
4364         if (perf_trace_enable(event->attr.config))
4365                 return NULL;
4366
4367         event->destroy = tp_perf_event_destroy;
4368
4369         return &perf_ops_generic;
4370 }
4371
4372 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4373 {
4374         char *filter_str;
4375         int ret;
4376
4377         if (event->attr.type != PERF_TYPE_TRACEPOINT)
4378                 return -EINVAL;
4379
4380         filter_str = strndup_user(arg, PAGE_SIZE);
4381         if (IS_ERR(filter_str))
4382                 return PTR_ERR(filter_str);
4383
4384         ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4385
4386         kfree(filter_str);
4387         return ret;
4388 }
4389
4390 static void perf_event_free_filter(struct perf_event *event)
4391 {
4392         ftrace_profile_free_filter(event);
4393 }
4394
4395 #else
4396
4397 static int perf_tp_event_match(struct perf_event *event,
4398                                 struct perf_sample_data *data)
4399 {
4400         return 1;
4401 }
4402
4403 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4404 {
4405         return NULL;
4406 }
4407
4408 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4409 {
4410         return -ENOENT;
4411 }
4412
4413 static void perf_event_free_filter(struct perf_event *event)
4414 {
4415 }
4416
4417 #endif /* CONFIG_EVENT_TRACING */
4418
4419 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4420 static void bp_perf_event_destroy(struct perf_event *event)
4421 {
4422         release_bp_slot(event);
4423 }
4424
4425 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4426 {
4427         int err;
4428
4429         err = register_perf_hw_breakpoint(bp);
4430         if (err)
4431                 return ERR_PTR(err);
4432
4433         bp->destroy = bp_perf_event_destroy;
4434
4435         return &perf_ops_bp;
4436 }
4437
4438 void perf_bp_event(struct perf_event *bp, void *data)
4439 {
4440         struct perf_sample_data sample;
4441         struct pt_regs *regs = data;
4442
4443         perf_sample_data_init(&sample, bp->attr.bp_addr);
4444
4445         if (!perf_exclude_event(bp, regs))
4446                 perf_swevent_add(bp, 1, 1, &sample, regs);
4447 }
4448 #else
4449 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4450 {
4451         return NULL;
4452 }
4453
4454 void perf_bp_event(struct perf_event *bp, void *regs)
4455 {
4456 }
4457 #endif
4458
4459 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4460
4461 static void sw_perf_event_destroy(struct perf_event *event)
4462 {
4463         u64 event_id = event->attr.config;
4464
4465         WARN_ON(event->parent);
4466
4467         atomic_dec(&perf_swevent_enabled[event_id]);
4468 }
4469
4470 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4471 {
4472         const struct pmu *pmu = NULL;
4473         u64 event_id = event->attr.config;
4474
4475         /*
4476          * Software events (currently) can't in general distinguish
4477          * between user, kernel and hypervisor events.
4478          * However, context switches and cpu migrations are considered
4479          * to be kernel events, and page faults are never hypervisor
4480          * events.
4481          */
4482         switch (event_id) {
4483         case PERF_COUNT_SW_CPU_CLOCK:
4484                 pmu = &perf_ops_cpu_clock;
4485
4486                 break;
4487         case PERF_COUNT_SW_TASK_CLOCK:
4488                 /*
4489                  * If the user instantiates this as a per-cpu event,
4490                  * use the cpu_clock event instead.
4491                  */
4492                 if (event->ctx->task)
4493                         pmu = &perf_ops_task_clock;
4494                 else
4495                         pmu = &perf_ops_cpu_clock;
4496
4497                 break;
4498         case PERF_COUNT_SW_PAGE_FAULTS:
4499         case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4500         case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4501         case PERF_COUNT_SW_CONTEXT_SWITCHES:
4502         case PERF_COUNT_SW_CPU_MIGRATIONS:
4503         case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4504         case PERF_COUNT_SW_EMULATION_FAULTS:
4505                 if (!event->parent) {
4506                         atomic_inc(&perf_swevent_enabled[event_id]);
4507                         event->destroy = sw_perf_event_destroy;
4508                 }
4509                 pmu = &perf_ops_generic;
4510                 break;
4511         }
4512
4513         return pmu;
4514 }
4515
4516 /*
4517  * Allocate and initialize a event structure
4518  */
4519 static struct perf_event *
4520 perf_event_alloc(struct perf_event_attr *attr,
4521                    int cpu,
4522                    struct perf_event_context *ctx,
4523                    struct perf_event *group_leader,
4524                    struct perf_event *parent_event,
4525                    perf_overflow_handler_t overflow_handler,
4526                    gfp_t gfpflags)
4527 {
4528         const struct pmu *pmu;
4529         struct perf_event *event;
4530         struct hw_perf_event *hwc;
4531         long err;
4532
4533         event = kzalloc(sizeof(*event), gfpflags);
4534         if (!event)
4535                 return ERR_PTR(-ENOMEM);
4536
4537         /*
4538          * Single events are their own group leaders, with an
4539          * empty sibling list:
4540          */
4541         if (!group_leader)
4542                 group_leader = event;
4543
4544         mutex_init(&event->child_mutex);
4545         INIT_LIST_HEAD(&event->child_list);
4546
4547         INIT_LIST_HEAD(&event->group_entry);
4548         INIT_LIST_HEAD(&event->event_entry);
4549         INIT_LIST_HEAD(&event->sibling_list);
4550         init_waitqueue_head(&event->waitq);
4551
4552         mutex_init(&event->mmap_mutex);
4553
4554         event->cpu              = cpu;
4555         event->attr             = *attr;
4556         event->group_leader     = group_leader;
4557         event->pmu              = NULL;
4558         event->ctx              = ctx;
4559         event->oncpu            = -1;
4560
4561         event->parent           = parent_event;
4562
4563         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
4564         event->id               = atomic64_inc_return(&perf_event_id);
4565
4566         event->state            = PERF_EVENT_STATE_INACTIVE;
4567
4568         if (!overflow_handler && parent_event)
4569                 overflow_handler = parent_event->overflow_handler;
4570         
4571         event->overflow_handler = overflow_handler;
4572
4573         if (attr->disabled)
4574                 event->state = PERF_EVENT_STATE_OFF;
4575
4576         pmu = NULL;
4577
4578         hwc = &event->hw;
4579         hwc->sample_period = attr->sample_period;
4580         if (attr->freq && attr->sample_freq)
4581                 hwc->sample_period = 1;
4582         hwc->last_period = hwc->sample_period;
4583
4584         atomic64_set(&hwc->period_left, hwc->sample_period);
4585
4586         /*
4587          * we currently do not support PERF_FORMAT_GROUP on inherited events
4588          */
4589         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4590                 goto done;
4591
4592         switch (attr->type) {
4593         case PERF_TYPE_RAW:
4594         case PERF_TYPE_HARDWARE:
4595         case PERF_TYPE_HW_CACHE:
4596                 pmu = hw_perf_event_init(event);
4597                 break;
4598
4599         case PERF_TYPE_SOFTWARE:
4600                 pmu = sw_perf_event_init(event);
4601                 break;
4602
4603         case PERF_TYPE_TRACEPOINT:
4604                 pmu = tp_perf_event_init(event);
4605                 break;
4606
4607         case PERF_TYPE_BREAKPOINT:
4608                 pmu = bp_perf_event_init(event);
4609                 break;
4610
4611
4612         default:
4613                 break;
4614         }
4615 done:
4616         err = 0;
4617         if (!pmu)
4618                 err = -EINVAL;
4619         else if (IS_ERR(pmu))
4620                 err = PTR_ERR(pmu);
4621
4622         if (err) {
4623                 if (event->ns)
4624                         put_pid_ns(event->ns);
4625                 kfree(event);
4626                 return ERR_PTR(err);
4627         }
4628
4629         event->pmu = pmu;
4630
4631         if (!event->parent) {
4632                 atomic_inc(&nr_events);
4633                 if (event->attr.mmap)
4634                         atomic_inc(&nr_mmap_events);
4635                 if (event->attr.comm)
4636                         atomic_inc(&nr_comm_events);
4637                 if (event->attr.task)
4638                         atomic_inc(&nr_task_events);
4639         }
4640
4641         return event;
4642 }
4643
4644 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4645                           struct perf_event_attr *attr)
4646 {
4647         u32 size;
4648         int ret;
4649
4650         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4651                 return -EFAULT;
4652
4653         /*
4654          * zero the full structure, so that a short copy will be nice.
4655          */
4656         memset(attr, 0, sizeof(*attr));
4657
4658         ret = get_user(size, &uattr->size);
4659         if (ret)
4660                 return ret;
4661
4662         if (size > PAGE_SIZE)   /* silly large */
4663                 goto err_size;
4664
4665         if (!size)              /* abi compat */
4666                 size = PERF_ATTR_SIZE_VER0;
4667
4668         if (size < PERF_ATTR_SIZE_VER0)
4669                 goto err_size;
4670
4671         /*
4672          * If we're handed a bigger struct than we know of,
4673          * ensure all the unknown bits are 0 - i.e. new
4674          * user-space does not rely on any kernel feature
4675          * extensions we dont know about yet.
4676          */
4677         if (size > sizeof(*attr)) {
4678                 unsigned char __user *addr;
4679                 unsigned char __user *end;
4680                 unsigned char val;
4681
4682                 addr = (void __user *)uattr + sizeof(*attr);
4683                 end  = (void __user *)uattr + size;
4684
4685                 for (; addr < end; addr++) {
4686                         ret = get_user(val, addr);
4687                         if (ret)
4688                                 return ret;
4689                         if (val)
4690                                 goto err_size;
4691                 }
4692                 size = sizeof(*attr);
4693         }
4694
4695         ret = copy_from_user(attr, uattr, size);
4696         if (ret)
4697                 return -EFAULT;
4698
4699         /*
4700          * If the type exists, the corresponding creation will verify
4701          * the attr->config.
4702          */
4703         if (attr->type >= PERF_TYPE_MAX)
4704                 return -EINVAL;
4705
4706         if (attr->__reserved_1)
4707                 return -EINVAL;
4708
4709         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4710                 return -EINVAL;
4711
4712         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4713                 return -EINVAL;
4714
4715 out:
4716         return ret;
4717
4718 err_size:
4719         put_user(sizeof(*attr), &uattr->size);
4720         ret = -E2BIG;
4721         goto out;
4722 }
4723
4724 static int perf_event_set_output(struct perf_event *event, int output_fd)
4725 {
4726         struct perf_event *output_event = NULL;
4727         struct file *output_file = NULL;
4728         struct perf_event *old_output;
4729         int fput_needed = 0;
4730         int ret = -EINVAL;
4731
4732         if (!output_fd)
4733                 goto set;
4734
4735         output_file = fget_light(output_fd, &fput_needed);
4736         if (!output_file)
4737                 return -EBADF;
4738
4739         if (output_file->f_op != &perf_fops)
4740                 goto out;
4741
4742         output_event = output_file->private_data;
4743
4744         /* Don't chain output fds */
4745         if (output_event->output)
4746                 goto out;
4747
4748         /* Don't set an output fd when we already have an output channel */
4749         if (event->data)
4750                 goto out;
4751
4752         atomic_long_inc(&output_file->f_count);
4753
4754 set:
4755         mutex_lock(&event->mmap_mutex);
4756         old_output = event->output;
4757         rcu_assign_pointer(event->output, output_event);
4758         mutex_unlock(&event->mmap_mutex);
4759
4760         if (old_output) {
4761                 /*
4762                  * we need to make sure no existing perf_output_*()
4763                  * is still referencing this event.
4764                  */
4765                 synchronize_rcu();
4766                 fput(old_output->filp);
4767         }
4768
4769         ret = 0;
4770 out:
4771         fput_light(output_file, fput_needed);
4772         return ret;
4773 }
4774
4775 /**
4776  * sys_perf_event_open - open a performance event, associate it to a task/cpu
4777  *
4778  * @attr_uptr:  event_id type attributes for monitoring/sampling
4779  * @pid:                target pid
4780  * @cpu:                target cpu
4781  * @group_fd:           group leader event fd
4782  */
4783 SYSCALL_DEFINE5(perf_event_open,
4784                 struct perf_event_attr __user *, attr_uptr,
4785                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4786 {
4787         struct perf_event *event, *group_leader;
4788         struct perf_event_attr attr;
4789         struct perf_event_context *ctx;
4790         struct file *event_file = NULL;
4791         struct file *group_file = NULL;
4792         int fput_needed = 0;
4793         int fput_needed2 = 0;
4794         int err;
4795
4796         /* for future expandability... */
4797         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4798                 return -EINVAL;
4799
4800         err = perf_copy_attr(attr_uptr, &attr);
4801         if (err)
4802                 return err;
4803
4804         if (!attr.exclude_kernel) {
4805                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4806                         return -EACCES;
4807         }
4808
4809         if (attr.freq) {
4810                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4811                         return -EINVAL;
4812         }
4813
4814         /*
4815          * Get the target context (task or percpu):
4816          */
4817         ctx = find_get_context(pid, cpu);
4818         if (IS_ERR(ctx))
4819                 return PTR_ERR(ctx);
4820
4821         /*
4822          * Look up the group leader (we will attach this event to it):
4823          */
4824         group_leader = NULL;
4825         if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4826                 err = -EINVAL;
4827                 group_file = fget_light(group_fd, &fput_needed);
4828                 if (!group_file)
4829                         goto err_put_context;
4830                 if (group_file->f_op != &perf_fops)
4831                         goto err_put_context;
4832
4833                 group_leader = group_file->private_data;
4834                 /*
4835                  * Do not allow a recursive hierarchy (this new sibling
4836                  * becoming part of another group-sibling):
4837                  */
4838                 if (group_leader->group_leader != group_leader)
4839                         goto err_put_context;
4840                 /*
4841                  * Do not allow to attach to a group in a different
4842                  * task or CPU context:
4843                  */
4844                 if (group_leader->ctx != ctx)
4845                         goto err_put_context;
4846                 /*
4847                  * Only a group leader can be exclusive or pinned
4848                  */
4849                 if (attr.exclusive || attr.pinned)
4850                         goto err_put_context;
4851         }
4852
4853         event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4854                                      NULL, NULL, GFP_KERNEL);
4855         err = PTR_ERR(event);
4856         if (IS_ERR(event))
4857                 goto err_put_context;
4858
4859         err = anon_inode_getfd("[perf_event]", &perf_fops, event, O_RDWR);
4860         if (err < 0)
4861                 goto err_free_put_context;
4862
4863         event_file = fget_light(err, &fput_needed2);
4864         if (!event_file)
4865                 goto err_free_put_context;
4866
4867         if (flags & PERF_FLAG_FD_OUTPUT) {
4868                 err = perf_event_set_output(event, group_fd);
4869                 if (err)
4870                         goto err_fput_free_put_context;
4871         }
4872
4873         event->filp = event_file;
4874         WARN_ON_ONCE(ctx->parent_ctx);
4875         mutex_lock(&ctx->mutex);
4876         perf_install_in_context(ctx, event, cpu);
4877         ++ctx->generation;
4878         mutex_unlock(&ctx->mutex);
4879
4880         event->owner = current;
4881         get_task_struct(current);
4882         mutex_lock(&current->perf_event_mutex);
4883         list_add_tail(&event->owner_entry, &current->perf_event_list);
4884         mutex_unlock(&current->perf_event_mutex);
4885
4886 err_fput_free_put_context:
4887         fput_light(event_file, fput_needed2);
4888
4889 err_free_put_context:
4890         if (err < 0)
4891                 kfree(event);
4892
4893 err_put_context:
4894         if (err < 0)
4895                 put_ctx(ctx);
4896
4897         fput_light(group_file, fput_needed);
4898
4899         return err;
4900 }
4901
4902 /**
4903  * perf_event_create_kernel_counter
4904  *
4905  * @attr: attributes of the counter to create
4906  * @cpu: cpu in which the counter is bound
4907  * @pid: task to profile
4908  */
4909 struct perf_event *
4910 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4911                                  pid_t pid,
4912                                  perf_overflow_handler_t overflow_handler)
4913 {
4914         struct perf_event *event;
4915         struct perf_event_context *ctx;
4916         int err;
4917
4918         /*
4919          * Get the target context (task or percpu):
4920          */
4921
4922         ctx = find_get_context(pid, cpu);
4923         if (IS_ERR(ctx)) {
4924                 err = PTR_ERR(ctx);
4925                 goto err_exit;
4926         }
4927
4928         event = perf_event_alloc(attr, cpu, ctx, NULL,
4929                                  NULL, overflow_handler, GFP_KERNEL);
4930         if (IS_ERR(event)) {
4931                 err = PTR_ERR(event);
4932                 goto err_put_context;
4933         }
4934
4935         event->filp = NULL;
4936         WARN_ON_ONCE(ctx->parent_ctx);
4937         mutex_lock(&ctx->mutex);
4938         perf_install_in_context(ctx, event, cpu);
4939         ++ctx->generation;
4940         mutex_unlock(&ctx->mutex);
4941
4942         event->owner = current;
4943         get_task_struct(current);
4944         mutex_lock(&current->perf_event_mutex);
4945         list_add_tail(&event->owner_entry, &current->perf_event_list);
4946         mutex_unlock(&current->perf_event_mutex);
4947
4948         return event;
4949
4950  err_put_context:
4951         put_ctx(ctx);
4952  err_exit:
4953         return ERR_PTR(err);
4954 }
4955 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4956
4957 /*
4958  * inherit a event from parent task to child task:
4959  */
4960 static struct perf_event *
4961 inherit_event(struct perf_event *parent_event,
4962               struct task_struct *parent,
4963               struct perf_event_context *parent_ctx,
4964               struct task_struct *child,
4965               struct perf_event *group_leader,
4966               struct perf_event_context *child_ctx)
4967 {
4968         struct perf_event *child_event;
4969
4970         /*
4971          * Instead of creating recursive hierarchies of events,
4972          * we link inherited events back to the original parent,
4973          * which has a filp for sure, which we use as the reference
4974          * count:
4975          */
4976         if (parent_event->parent)
4977                 parent_event = parent_event->parent;
4978
4979         child_event = perf_event_alloc(&parent_event->attr,
4980                                            parent_event->cpu, child_ctx,
4981                                            group_leader, parent_event,
4982                                            NULL, GFP_KERNEL);
4983         if (IS_ERR(child_event))
4984                 return child_event;
4985         get_ctx(child_ctx);
4986
4987         /*
4988          * Make the child state follow the state of the parent event,
4989          * not its attr.disabled bit.  We hold the parent's mutex,
4990          * so we won't race with perf_event_{en, dis}able_family.
4991          */
4992         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
4993                 child_event->state = PERF_EVENT_STATE_INACTIVE;
4994         else
4995                 child_event->state = PERF_EVENT_STATE_OFF;
4996
4997         if (parent_event->attr.freq) {
4998                 u64 sample_period = parent_event->hw.sample_period;
4999                 struct hw_perf_event *hwc = &child_event->hw;
5000
5001                 hwc->sample_period = sample_period;
5002                 hwc->last_period   = sample_period;
5003
5004                 atomic64_set(&hwc->period_left, sample_period);
5005         }
5006
5007         child_event->overflow_handler = parent_event->overflow_handler;
5008
5009         /*
5010          * Link it up in the child's context:
5011          */
5012         add_event_to_ctx(child_event, child_ctx);
5013
5014         /*
5015          * Get a reference to the parent filp - we will fput it
5016          * when the child event exits. This is safe to do because
5017          * we are in the parent and we know that the filp still
5018          * exists and has a nonzero count:
5019          */
5020         atomic_long_inc(&parent_event->filp->f_count);
5021
5022         /*
5023          * Link this into the parent event's child list
5024          */
5025         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5026         mutex_lock(&parent_event->child_mutex);
5027         list_add_tail(&child_event->child_list, &parent_event->child_list);
5028         mutex_unlock(&parent_event->child_mutex);
5029
5030         return child_event;
5031 }
5032
5033 static int inherit_group(struct perf_event *parent_event,
5034               struct task_struct *parent,
5035               struct perf_event_context *parent_ctx,
5036               struct task_struct *child,
5037               struct perf_event_context *child_ctx)
5038 {
5039         struct perf_event *leader;
5040         struct perf_event *sub;
5041         struct perf_event *child_ctr;
5042
5043         leader = inherit_event(parent_event, parent, parent_ctx,
5044                                  child, NULL, child_ctx);
5045         if (IS_ERR(leader))
5046                 return PTR_ERR(leader);
5047         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5048                 child_ctr = inherit_event(sub, parent, parent_ctx,
5049                                             child, leader, child_ctx);
5050                 if (IS_ERR(child_ctr))
5051                         return PTR_ERR(child_ctr);
5052         }
5053         return 0;
5054 }
5055
5056 static void sync_child_event(struct perf_event *child_event,
5057                                struct task_struct *child)
5058 {
5059         struct perf_event *parent_event = child_event->parent;
5060         u64 child_val;
5061
5062         if (child_event->attr.inherit_stat)
5063                 perf_event_read_event(child_event, child);
5064
5065         child_val = atomic64_read(&child_event->count);
5066
5067         /*
5068          * Add back the child's count to the parent's count:
5069          */
5070         atomic64_add(child_val, &parent_event->count);
5071         atomic64_add(child_event->total_time_enabled,
5072                      &parent_event->child_total_time_enabled);
5073         atomic64_add(child_event->total_time_running,
5074                      &parent_event->child_total_time_running);
5075
5076         /*
5077          * Remove this event from the parent's list
5078          */
5079         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5080         mutex_lock(&parent_event->child_mutex);
5081         list_del_init(&child_event->child_list);
5082         mutex_unlock(&parent_event->child_mutex);
5083
5084         /*
5085          * Release the parent event, if this was the last
5086          * reference to it.
5087          */
5088         fput(parent_event->filp);
5089 }
5090
5091 static void
5092 __perf_event_exit_task(struct perf_event *child_event,
5093                          struct perf_event_context *child_ctx,
5094                          struct task_struct *child)
5095 {
5096         struct perf_event *parent_event;
5097
5098         perf_event_remove_from_context(child_event);
5099
5100         parent_event = child_event->parent;
5101         /*
5102          * It can happen that parent exits first, and has events
5103          * that are still around due to the child reference. These
5104          * events need to be zapped - but otherwise linger.
5105          */
5106         if (parent_event) {
5107                 sync_child_event(child_event, child);
5108                 free_event(child_event);
5109         }
5110 }
5111
5112 /*
5113  * When a child task exits, feed back event values to parent events.
5114  */
5115 void perf_event_exit_task(struct task_struct *child)
5116 {
5117         struct perf_event *child_event, *tmp;
5118         struct perf_event_context *child_ctx;
5119         unsigned long flags;
5120
5121         if (likely(!child->perf_event_ctxp)) {
5122                 perf_event_task(child, NULL, 0);
5123                 return;
5124         }
5125
5126         local_irq_save(flags);
5127         /*
5128          * We can't reschedule here because interrupts are disabled,
5129          * and either child is current or it is a task that can't be
5130          * scheduled, so we are now safe from rescheduling changing
5131          * our context.
5132          */
5133         child_ctx = child->perf_event_ctxp;
5134         __perf_event_task_sched_out(child_ctx);
5135
5136         /*
5137          * Take the context lock here so that if find_get_context is
5138          * reading child->perf_event_ctxp, we wait until it has
5139          * incremented the context's refcount before we do put_ctx below.
5140          */
5141         raw_spin_lock(&child_ctx->lock);
5142         child->perf_event_ctxp = NULL;
5143         /*
5144          * If this context is a clone; unclone it so it can't get
5145          * swapped to another process while we're removing all
5146          * the events from it.
5147          */
5148         unclone_ctx(child_ctx);
5149         update_context_time(child_ctx);
5150         raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5151
5152         /*
5153          * Report the task dead after unscheduling the events so that we
5154          * won't get any samples after PERF_RECORD_EXIT. We can however still
5155          * get a few PERF_RECORD_READ events.
5156          */
5157         perf_event_task(child, child_ctx, 0);
5158
5159         /*
5160          * We can recurse on the same lock type through:
5161          *
5162          *   __perf_event_exit_task()
5163          *     sync_child_event()
5164          *       fput(parent_event->filp)
5165          *         perf_release()
5166          *           mutex_lock(&ctx->mutex)
5167          *
5168          * But since its the parent context it won't be the same instance.
5169          */
5170         mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
5171
5172 again:
5173         list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5174                                  group_entry)
5175                 __perf_event_exit_task(child_event, child_ctx, child);
5176
5177         list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5178                                  group_entry)
5179                 __perf_event_exit_task(child_event, child_ctx, child);
5180
5181         /*
5182          * If the last event was a group event, it will have appended all
5183          * its siblings to the list, but we obtained 'tmp' before that which
5184          * will still point to the list head terminating the iteration.
5185          */
5186         if (!list_empty(&child_ctx->pinned_groups) ||
5187             !list_empty(&child_ctx->flexible_groups))
5188                 goto again;
5189
5190         mutex_unlock(&child_ctx->mutex);
5191
5192         put_ctx(child_ctx);
5193 }
5194
5195 static void perf_free_event(struct perf_event *event,
5196                             struct perf_event_context *ctx)
5197 {
5198         struct perf_event *parent = event->parent;
5199
5200         if (WARN_ON_ONCE(!parent))
5201                 return;
5202
5203         mutex_lock(&parent->child_mutex);
5204         list_del_init(&event->child_list);
5205         mutex_unlock(&parent->child_mutex);
5206
5207         fput(parent->filp);
5208
5209         list_del_event(event, ctx);
5210         free_event(event);
5211 }
5212
5213 /*
5214  * free an unexposed, unused context as created by inheritance by
5215  * init_task below, used by fork() in case of fail.
5216  */
5217 void perf_event_free_task(struct task_struct *task)
5218 {
5219         struct perf_event_context *ctx = task->perf_event_ctxp;
5220         struct perf_event *event, *tmp;
5221
5222         if (!ctx)
5223                 return;
5224
5225         mutex_lock(&ctx->mutex);
5226 again:
5227         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5228                 perf_free_event(event, ctx);
5229
5230         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5231                                  group_entry)
5232                 perf_free_event(event, ctx);
5233
5234         if (!list_empty(&ctx->pinned_groups) ||
5235             !list_empty(&ctx->flexible_groups))
5236                 goto again;
5237
5238         mutex_unlock(&ctx->mutex);
5239
5240         put_ctx(ctx);
5241 }
5242
5243 static int
5244 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5245                    struct perf_event_context *parent_ctx,
5246                    struct task_struct *child,
5247                    int *inherited_all)
5248 {
5249         int ret;
5250         struct perf_event_context *child_ctx = child->perf_event_ctxp;
5251
5252         if (!event->attr.inherit) {
5253                 *inherited_all = 0;
5254                 return 0;
5255         }
5256
5257         if (!child_ctx) {
5258                 /*
5259                  * This is executed from the parent task context, so
5260                  * inherit events that have been marked for cloning.
5261                  * First allocate and initialize a context for the
5262                  * child.
5263                  */
5264
5265                 child_ctx = kzalloc(sizeof(struct perf_event_context),
5266                                     GFP_KERNEL);
5267                 if (!child_ctx)
5268                         return -ENOMEM;
5269
5270                 __perf_event_init_context(child_ctx, child);
5271                 child->perf_event_ctxp = child_ctx;
5272                 get_task_struct(child);
5273         }
5274
5275         ret = inherit_group(event, parent, parent_ctx,
5276                             child, child_ctx);
5277
5278         if (ret)
5279                 *inherited_all = 0;
5280
5281         return ret;
5282 }
5283
5284
5285 /*
5286  * Initialize the perf_event context in task_struct
5287  */
5288 int perf_event_init_task(struct task_struct *child)
5289 {
5290         struct perf_event_context *child_ctx, *parent_ctx;
5291         struct perf_event_context *cloned_ctx;
5292         struct perf_event *event;
5293         struct task_struct *parent = current;
5294         int inherited_all = 1;
5295         int ret = 0;
5296
5297         child->perf_event_ctxp = NULL;
5298
5299         mutex_init(&child->perf_event_mutex);
5300         INIT_LIST_HEAD(&child->perf_event_list);
5301
5302         if (likely(!parent->perf_event_ctxp))
5303                 return 0;
5304
5305         /*
5306          * If the parent's context is a clone, pin it so it won't get
5307          * swapped under us.
5308          */
5309         parent_ctx = perf_pin_task_context(parent);
5310
5311         /*
5312          * No need to check if parent_ctx != NULL here; since we saw
5313          * it non-NULL earlier, the only reason for it to become NULL
5314          * is if we exit, and since we're currently in the middle of
5315          * a fork we can't be exiting at the same time.
5316          */
5317
5318         /*
5319          * Lock the parent list. No need to lock the child - not PID
5320          * hashed yet and not running, so nobody can access it.
5321          */
5322         mutex_lock(&parent_ctx->mutex);
5323
5324         /*
5325          * We dont have to disable NMIs - we are only looking at
5326          * the list, not manipulating it:
5327          */
5328         list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5329                 ret = inherit_task_group(event, parent, parent_ctx, child,
5330                                          &inherited_all);
5331                 if (ret)
5332                         break;
5333         }
5334
5335         list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5336                 ret = inherit_task_group(event, parent, parent_ctx, child,
5337                                          &inherited_all);
5338                 if (ret)
5339                         break;
5340         }
5341
5342         child_ctx = child->perf_event_ctxp;
5343
5344         if (child_ctx && inherited_all) {
5345                 /*
5346                  * Mark the child context as a clone of the parent
5347                  * context, or of whatever the parent is a clone of.
5348                  * Note that if the parent is a clone, it could get
5349                  * uncloned at any point, but that doesn't matter
5350                  * because the list of events and the generation
5351                  * count can't have changed since we took the mutex.
5352                  */
5353                 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5354                 if (cloned_ctx) {
5355                         child_ctx->parent_ctx = cloned_ctx;
5356                         child_ctx->parent_gen = parent_ctx->parent_gen;
5357                 } else {
5358                         child_ctx->parent_ctx = parent_ctx;
5359                         child_ctx->parent_gen = parent_ctx->generation;
5360                 }
5361                 get_ctx(child_ctx->parent_ctx);
5362         }
5363
5364         mutex_unlock(&parent_ctx->mutex);
5365
5366         perf_unpin_context(parent_ctx);
5367
5368         return ret;
5369 }
5370
5371 static void __cpuinit perf_event_init_cpu(int cpu)
5372 {
5373         struct perf_cpu_context *cpuctx;
5374
5375         cpuctx = &per_cpu(perf_cpu_context, cpu);
5376         __perf_event_init_context(&cpuctx->ctx, NULL);
5377
5378         spin_lock(&perf_resource_lock);
5379         cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5380         spin_unlock(&perf_resource_lock);
5381 }
5382
5383 #ifdef CONFIG_HOTPLUG_CPU
5384 static void __perf_event_exit_cpu(void *info)
5385 {
5386         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5387         struct perf_event_context *ctx = &cpuctx->ctx;
5388         struct perf_event *event, *tmp;
5389
5390         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5391                 __perf_event_remove_from_context(event);
5392         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5393                 __perf_event_remove_from_context(event);
5394 }
5395 static void perf_event_exit_cpu(int cpu)
5396 {
5397         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5398         struct perf_event_context *ctx = &cpuctx->ctx;
5399
5400         mutex_lock(&ctx->mutex);
5401         smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5402         mutex_unlock(&ctx->mutex);
5403 }
5404 #else
5405 static inline void perf_event_exit_cpu(int cpu) { }
5406 #endif
5407
5408 static int __cpuinit
5409 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5410 {
5411         unsigned int cpu = (long)hcpu;
5412
5413         switch (action) {
5414
5415         case CPU_UP_PREPARE:
5416         case CPU_UP_PREPARE_FROZEN:
5417                 perf_event_init_cpu(cpu);
5418                 break;
5419
5420         case CPU_DOWN_PREPARE:
5421         case CPU_DOWN_PREPARE_FROZEN:
5422                 perf_event_exit_cpu(cpu);
5423                 break;
5424
5425         default:
5426                 break;
5427         }
5428
5429         return NOTIFY_OK;
5430 }
5431
5432 /*
5433  * This has to have a higher priority than migration_notifier in sched.c.
5434  */
5435 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5436         .notifier_call          = perf_cpu_notify,
5437         .priority               = 20,
5438 };
5439
5440 void __init perf_event_init(void)
5441 {
5442         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5443                         (void *)(long)smp_processor_id());
5444         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5445                         (void *)(long)smp_processor_id());
5446         register_cpu_notifier(&perf_cpu_nb);
5447 }
5448
5449 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class,
5450                                         struct sysdev_class_attribute *attr,
5451                                         char *buf)
5452 {
5453         return sprintf(buf, "%d\n", perf_reserved_percpu);
5454 }
5455
5456 static ssize_t
5457 perf_set_reserve_percpu(struct sysdev_class *class,
5458                         struct sysdev_class_attribute *attr,
5459                         const char *buf,
5460                         size_t count)
5461 {
5462         struct perf_cpu_context *cpuctx;
5463         unsigned long val;
5464         int err, cpu, mpt;
5465
5466         err = strict_strtoul(buf, 10, &val);
5467         if (err)
5468                 return err;
5469         if (val > perf_max_events)
5470                 return -EINVAL;
5471
5472         spin_lock(&perf_resource_lock);
5473         perf_reserved_percpu = val;
5474         for_each_online_cpu(cpu) {
5475                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5476                 raw_spin_lock_irq(&cpuctx->ctx.lock);
5477                 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5478                           perf_max_events - perf_reserved_percpu);
5479                 cpuctx->max_pertask = mpt;
5480                 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5481         }
5482         spin_unlock(&perf_resource_lock);
5483
5484         return count;
5485 }
5486
5487 static ssize_t perf_show_overcommit(struct sysdev_class *class,
5488                                     struct sysdev_class_attribute *attr,
5489                                     char *buf)
5490 {
5491         return sprintf(buf, "%d\n", perf_overcommit);
5492 }
5493
5494 static ssize_t
5495 perf_set_overcommit(struct sysdev_class *class,
5496                     struct sysdev_class_attribute *attr,
5497                     const char *buf, size_t count)
5498 {
5499         unsigned long val;
5500         int err;
5501
5502         err = strict_strtoul(buf, 10, &val);
5503         if (err)
5504                 return err;
5505         if (val > 1)
5506                 return -EINVAL;
5507
5508         spin_lock(&perf_resource_lock);
5509         perf_overcommit = val;
5510         spin_unlock(&perf_resource_lock);
5511
5512         return count;
5513 }
5514
5515 static SYSDEV_CLASS_ATTR(
5516                                 reserve_percpu,
5517                                 0644,
5518                                 perf_show_reserve_percpu,
5519                                 perf_set_reserve_percpu
5520                         );
5521
5522 static SYSDEV_CLASS_ATTR(
5523                                 overcommit,
5524                                 0644,
5525                                 perf_show_overcommit,
5526                                 perf_set_overcommit
5527                         );
5528
5529 static struct attribute *perfclass_attrs[] = {
5530         &attr_reserve_percpu.attr,
5531         &attr_overcommit.attr,
5532         NULL
5533 };
5534
5535 static struct attribute_group perfclass_attr_group = {
5536         .attrs                  = perfclass_attrs,
5537         .name                   = "perf_events",
5538 };
5539
5540 static int __init perf_event_sysfs_init(void)
5541 {
5542         return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5543                                   &perfclass_attr_group);
5544 }
5545 device_initcall(perf_event_sysfs_init);