perf: Fix 'perf sched record' deadlock
[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         int do_switch = 1;
1168
1169         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0);
1170
1171         if (likely(!ctx || !cpuctx->task_ctx))
1172                 return;
1173
1174         rcu_read_lock();
1175         parent = rcu_dereference(ctx->parent_ctx);
1176         next_ctx = next->perf_event_ctxp;
1177         if (parent && next_ctx &&
1178             rcu_dereference(next_ctx->parent_ctx) == parent) {
1179                 /*
1180                  * Looks like the two contexts are clones, so we might be
1181                  * able to optimize the context switch.  We lock both
1182                  * contexts and check that they are clones under the
1183                  * lock (including re-checking that neither has been
1184                  * uncloned in the meantime).  It doesn't matter which
1185                  * order we take the locks because no other cpu could
1186                  * be trying to lock both of these tasks.
1187                  */
1188                 raw_spin_lock(&ctx->lock);
1189                 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1190                 if (context_equiv(ctx, next_ctx)) {
1191                         /*
1192                          * XXX do we need a memory barrier of sorts
1193                          * wrt to rcu_dereference() of perf_event_ctxp
1194                          */
1195                         task->perf_event_ctxp = next_ctx;
1196                         next->perf_event_ctxp = ctx;
1197                         ctx->task = next;
1198                         next_ctx->task = task;
1199                         do_switch = 0;
1200
1201                         perf_event_sync_stat(ctx, next_ctx);
1202                 }
1203                 raw_spin_unlock(&next_ctx->lock);
1204                 raw_spin_unlock(&ctx->lock);
1205         }
1206         rcu_read_unlock();
1207
1208         if (do_switch) {
1209                 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1210                 cpuctx->task_ctx = NULL;
1211         }
1212 }
1213
1214 static void task_ctx_sched_out(struct perf_event_context *ctx,
1215                                enum event_type_t event_type)
1216 {
1217         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1218
1219         if (!cpuctx->task_ctx)
1220                 return;
1221
1222         if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1223                 return;
1224
1225         ctx_sched_out(ctx, cpuctx, event_type);
1226         cpuctx->task_ctx = NULL;
1227 }
1228
1229 /*
1230  * Called with IRQs disabled
1231  */
1232 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1233 {
1234         task_ctx_sched_out(ctx, EVENT_ALL);
1235 }
1236
1237 /*
1238  * Called with IRQs disabled
1239  */
1240 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1241                               enum event_type_t event_type)
1242 {
1243         ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1244 }
1245
1246 static void
1247 ctx_pinned_sched_in(struct perf_event_context *ctx,
1248                     struct perf_cpu_context *cpuctx)
1249 {
1250         struct perf_event *event;
1251
1252         list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1253                 if (event->state <= PERF_EVENT_STATE_OFF)
1254                         continue;
1255                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1256                         continue;
1257
1258                 if (group_can_go_on(event, cpuctx, 1))
1259                         group_sched_in(event, cpuctx, ctx);
1260
1261                 /*
1262                  * If this pinned group hasn't been scheduled,
1263                  * put it in error state.
1264                  */
1265                 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1266                         update_group_times(event);
1267                         event->state = PERF_EVENT_STATE_ERROR;
1268                 }
1269         }
1270 }
1271
1272 static void
1273 ctx_flexible_sched_in(struct perf_event_context *ctx,
1274                       struct perf_cpu_context *cpuctx)
1275 {
1276         struct perf_event *event;
1277         int can_add_hw = 1;
1278
1279         list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1280                 /* Ignore events in OFF or ERROR state */
1281                 if (event->state <= PERF_EVENT_STATE_OFF)
1282                         continue;
1283                 /*
1284                  * Listen to the 'cpu' scheduling filter constraint
1285                  * of events:
1286                  */
1287                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1288                         continue;
1289
1290                 if (group_can_go_on(event, cpuctx, can_add_hw))
1291                         if (group_sched_in(event, cpuctx, ctx))
1292                                 can_add_hw = 0;
1293         }
1294 }
1295
1296 static void
1297 ctx_sched_in(struct perf_event_context *ctx,
1298              struct perf_cpu_context *cpuctx,
1299              enum event_type_t event_type)
1300 {
1301         raw_spin_lock(&ctx->lock);
1302         ctx->is_active = 1;
1303         if (likely(!ctx->nr_events))
1304                 goto out;
1305
1306         ctx->timestamp = perf_clock();
1307
1308         perf_disable();
1309
1310         /*
1311          * First go through the list and put on any pinned groups
1312          * in order to give them the best chance of going on.
1313          */
1314         if (event_type & EVENT_PINNED)
1315                 ctx_pinned_sched_in(ctx, cpuctx);
1316
1317         /* Then walk through the lower prio flexible groups */
1318         if (event_type & EVENT_FLEXIBLE)
1319                 ctx_flexible_sched_in(ctx, cpuctx);
1320
1321         perf_enable();
1322  out:
1323         raw_spin_unlock(&ctx->lock);
1324 }
1325
1326 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1327                              enum event_type_t event_type)
1328 {
1329         struct perf_event_context *ctx = &cpuctx->ctx;
1330
1331         ctx_sched_in(ctx, cpuctx, event_type);
1332 }
1333
1334 static void task_ctx_sched_in(struct task_struct *task,
1335                               enum event_type_t event_type)
1336 {
1337         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1338         struct perf_event_context *ctx = task->perf_event_ctxp;
1339
1340         if (likely(!ctx))
1341                 return;
1342         if (cpuctx->task_ctx == ctx)
1343                 return;
1344         ctx_sched_in(ctx, cpuctx, event_type);
1345         cpuctx->task_ctx = ctx;
1346 }
1347 /*
1348  * Called from scheduler to add the events of the current task
1349  * with interrupts disabled.
1350  *
1351  * We restore the event value and then enable it.
1352  *
1353  * This does not protect us against NMI, but enable()
1354  * sets the enabled bit in the control field of event _before_
1355  * accessing the event control register. If a NMI hits, then it will
1356  * keep the event running.
1357  */
1358 void perf_event_task_sched_in(struct task_struct *task)
1359 {
1360         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1361         struct perf_event_context *ctx = task->perf_event_ctxp;
1362
1363         if (likely(!ctx))
1364                 return;
1365
1366         if (cpuctx->task_ctx == ctx)
1367                 return;
1368
1369         /*
1370          * We want to keep the following priority order:
1371          * cpu pinned (that don't need to move), task pinned,
1372          * cpu flexible, task flexible.
1373          */
1374         cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1375
1376         ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1377         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1378         ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1379
1380         cpuctx->task_ctx = ctx;
1381 }
1382
1383 #define MAX_INTERRUPTS (~0ULL)
1384
1385 static void perf_log_throttle(struct perf_event *event, int enable);
1386
1387 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1388 {
1389         u64 frequency = event->attr.sample_freq;
1390         u64 sec = NSEC_PER_SEC;
1391         u64 divisor, dividend;
1392
1393         int count_fls, nsec_fls, frequency_fls, sec_fls;
1394
1395         count_fls = fls64(count);
1396         nsec_fls = fls64(nsec);
1397         frequency_fls = fls64(frequency);
1398         sec_fls = 30;
1399
1400         /*
1401          * We got @count in @nsec, with a target of sample_freq HZ
1402          * the target period becomes:
1403          *
1404          *             @count * 10^9
1405          * period = -------------------
1406          *          @nsec * sample_freq
1407          *
1408          */
1409
1410         /*
1411          * Reduce accuracy by one bit such that @a and @b converge
1412          * to a similar magnitude.
1413          */
1414 #define REDUCE_FLS(a, b)                \
1415 do {                                    \
1416         if (a##_fls > b##_fls) {        \
1417                 a >>= 1;                \
1418                 a##_fls--;              \
1419         } else {                        \
1420                 b >>= 1;                \
1421                 b##_fls--;              \
1422         }                               \
1423 } while (0)
1424
1425         /*
1426          * Reduce accuracy until either term fits in a u64, then proceed with
1427          * the other, so that finally we can do a u64/u64 division.
1428          */
1429         while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1430                 REDUCE_FLS(nsec, frequency);
1431                 REDUCE_FLS(sec, count);
1432         }
1433
1434         if (count_fls + sec_fls > 64) {
1435                 divisor = nsec * frequency;
1436
1437                 while (count_fls + sec_fls > 64) {
1438                         REDUCE_FLS(count, sec);
1439                         divisor >>= 1;
1440                 }
1441
1442                 dividend = count * sec;
1443         } else {
1444                 dividend = count * sec;
1445
1446                 while (nsec_fls + frequency_fls > 64) {
1447                         REDUCE_FLS(nsec, frequency);
1448                         dividend >>= 1;
1449                 }
1450
1451                 divisor = nsec * frequency;
1452         }
1453
1454         return div64_u64(dividend, divisor);
1455 }
1456
1457 static void perf_event_stop(struct perf_event *event)
1458 {
1459         if (!event->pmu->stop)
1460                 return event->pmu->disable(event);
1461
1462         return event->pmu->stop(event);
1463 }
1464
1465 static int perf_event_start(struct perf_event *event)
1466 {
1467         if (!event->pmu->start)
1468                 return event->pmu->enable(event);
1469
1470         return event->pmu->start(event);
1471 }
1472
1473 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1474 {
1475         struct hw_perf_event *hwc = &event->hw;
1476         u64 period, sample_period;
1477         s64 delta;
1478
1479         period = perf_calculate_period(event, nsec, count);
1480
1481         delta = (s64)(period - hwc->sample_period);
1482         delta = (delta + 7) / 8; /* low pass filter */
1483
1484         sample_period = hwc->sample_period + delta;
1485
1486         if (!sample_period)
1487                 sample_period = 1;
1488
1489         hwc->sample_period = sample_period;
1490
1491         if (atomic64_read(&hwc->period_left) > 8*sample_period) {
1492                 perf_disable();
1493                 perf_event_stop(event);
1494                 atomic64_set(&hwc->period_left, 0);
1495                 perf_event_start(event);
1496                 perf_enable();
1497         }
1498 }
1499
1500 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1501 {
1502         struct perf_event *event;
1503         struct hw_perf_event *hwc;
1504         u64 interrupts, now;
1505         s64 delta;
1506
1507         raw_spin_lock(&ctx->lock);
1508         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1509                 if (event->state != PERF_EVENT_STATE_ACTIVE)
1510                         continue;
1511
1512                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1513                         continue;
1514
1515                 hwc = &event->hw;
1516
1517                 interrupts = hwc->interrupts;
1518                 hwc->interrupts = 0;
1519
1520                 /*
1521                  * unthrottle events on the tick
1522                  */
1523                 if (interrupts == MAX_INTERRUPTS) {
1524                         perf_log_throttle(event, 1);
1525                         perf_disable();
1526                         event->pmu->unthrottle(event);
1527                         perf_enable();
1528                 }
1529
1530                 if (!event->attr.freq || !event->attr.sample_freq)
1531                         continue;
1532
1533                 perf_disable();
1534                 event->pmu->read(event);
1535                 now = atomic64_read(&event->count);
1536                 delta = now - hwc->freq_count_stamp;
1537                 hwc->freq_count_stamp = now;
1538
1539                 if (delta > 0)
1540                         perf_adjust_period(event, TICK_NSEC, delta);
1541                 perf_enable();
1542         }
1543         raw_spin_unlock(&ctx->lock);
1544 }
1545
1546 /*
1547  * Round-robin a context's events:
1548  */
1549 static void rotate_ctx(struct perf_event_context *ctx)
1550 {
1551         raw_spin_lock(&ctx->lock);
1552
1553         /* Rotate the first entry last of non-pinned groups */
1554         list_rotate_left(&ctx->flexible_groups);
1555
1556         raw_spin_unlock(&ctx->lock);
1557 }
1558
1559 void perf_event_task_tick(struct task_struct *curr)
1560 {
1561         struct perf_cpu_context *cpuctx;
1562         struct perf_event_context *ctx;
1563         int rotate = 0;
1564
1565         if (!atomic_read(&nr_events))
1566                 return;
1567
1568         cpuctx = &__get_cpu_var(perf_cpu_context);
1569         if (cpuctx->ctx.nr_events &&
1570             cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1571                 rotate = 1;
1572
1573         ctx = curr->perf_event_ctxp;
1574         if (ctx && ctx->nr_events && ctx->nr_events != ctx->nr_active)
1575                 rotate = 1;
1576
1577         perf_ctx_adjust_freq(&cpuctx->ctx);
1578         if (ctx)
1579                 perf_ctx_adjust_freq(ctx);
1580
1581         if (!rotate)
1582                 return;
1583
1584         perf_disable();
1585         cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1586         if (ctx)
1587                 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1588
1589         rotate_ctx(&cpuctx->ctx);
1590         if (ctx)
1591                 rotate_ctx(ctx);
1592
1593         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1594         if (ctx)
1595                 task_ctx_sched_in(curr, EVENT_FLEXIBLE);
1596         perf_enable();
1597 }
1598
1599 static int event_enable_on_exec(struct perf_event *event,
1600                                 struct perf_event_context *ctx)
1601 {
1602         if (!event->attr.enable_on_exec)
1603                 return 0;
1604
1605         event->attr.enable_on_exec = 0;
1606         if (event->state >= PERF_EVENT_STATE_INACTIVE)
1607                 return 0;
1608
1609         __perf_event_mark_enabled(event, ctx);
1610
1611         return 1;
1612 }
1613
1614 /*
1615  * Enable all of a task's events that have been marked enable-on-exec.
1616  * This expects task == current.
1617  */
1618 static void perf_event_enable_on_exec(struct task_struct *task)
1619 {
1620         struct perf_event_context *ctx;
1621         struct perf_event *event;
1622         unsigned long flags;
1623         int enabled = 0;
1624         int ret;
1625
1626         local_irq_save(flags);
1627         ctx = task->perf_event_ctxp;
1628         if (!ctx || !ctx->nr_events)
1629                 goto out;
1630
1631         __perf_event_task_sched_out(ctx);
1632
1633         raw_spin_lock(&ctx->lock);
1634
1635         list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1636                 ret = event_enable_on_exec(event, ctx);
1637                 if (ret)
1638                         enabled = 1;
1639         }
1640
1641         list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1642                 ret = event_enable_on_exec(event, ctx);
1643                 if (ret)
1644                         enabled = 1;
1645         }
1646
1647         /*
1648          * Unclone this context if we enabled any event.
1649          */
1650         if (enabled)
1651                 unclone_ctx(ctx);
1652
1653         raw_spin_unlock(&ctx->lock);
1654
1655         perf_event_task_sched_in(task);
1656  out:
1657         local_irq_restore(flags);
1658 }
1659
1660 /*
1661  * Cross CPU call to read the hardware event
1662  */
1663 static void __perf_event_read(void *info)
1664 {
1665         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1666         struct perf_event *event = info;
1667         struct perf_event_context *ctx = event->ctx;
1668
1669         /*
1670          * If this is a task context, we need to check whether it is
1671          * the current task context of this cpu.  If not it has been
1672          * scheduled out before the smp call arrived.  In that case
1673          * event->count would have been updated to a recent sample
1674          * when the event was scheduled out.
1675          */
1676         if (ctx->task && cpuctx->task_ctx != ctx)
1677                 return;
1678
1679         raw_spin_lock(&ctx->lock);
1680         update_context_time(ctx);
1681         update_event_times(event);
1682         raw_spin_unlock(&ctx->lock);
1683
1684         event->pmu->read(event);
1685 }
1686
1687 static u64 perf_event_read(struct perf_event *event)
1688 {
1689         /*
1690          * If event is enabled and currently active on a CPU, update the
1691          * value in the event structure:
1692          */
1693         if (event->state == PERF_EVENT_STATE_ACTIVE) {
1694                 smp_call_function_single(event->oncpu,
1695                                          __perf_event_read, event, 1);
1696         } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1697                 struct perf_event_context *ctx = event->ctx;
1698                 unsigned long flags;
1699
1700                 raw_spin_lock_irqsave(&ctx->lock, flags);
1701                 update_context_time(ctx);
1702                 update_event_times(event);
1703                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1704         }
1705
1706         return atomic64_read(&event->count);
1707 }
1708
1709 /*
1710  * Initialize the perf_event context in a task_struct:
1711  */
1712 static void
1713 __perf_event_init_context(struct perf_event_context *ctx,
1714                             struct task_struct *task)
1715 {
1716         raw_spin_lock_init(&ctx->lock);
1717         mutex_init(&ctx->mutex);
1718         INIT_LIST_HEAD(&ctx->pinned_groups);
1719         INIT_LIST_HEAD(&ctx->flexible_groups);
1720         INIT_LIST_HEAD(&ctx->event_list);
1721         atomic_set(&ctx->refcount, 1);
1722         ctx->task = task;
1723 }
1724
1725 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1726 {
1727         struct perf_event_context *ctx;
1728         struct perf_cpu_context *cpuctx;
1729         struct task_struct *task;
1730         unsigned long flags;
1731         int err;
1732
1733         if (pid == -1 && cpu != -1) {
1734                 /* Must be root to operate on a CPU event: */
1735                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1736                         return ERR_PTR(-EACCES);
1737
1738                 if (cpu < 0 || cpu >= nr_cpumask_bits)
1739                         return ERR_PTR(-EINVAL);
1740
1741                 /*
1742                  * We could be clever and allow to attach a event to an
1743                  * offline CPU and activate it when the CPU comes up, but
1744                  * that's for later.
1745                  */
1746                 if (!cpu_online(cpu))
1747                         return ERR_PTR(-ENODEV);
1748
1749                 cpuctx = &per_cpu(perf_cpu_context, cpu);
1750                 ctx = &cpuctx->ctx;
1751                 get_ctx(ctx);
1752
1753                 return ctx;
1754         }
1755
1756         rcu_read_lock();
1757         if (!pid)
1758                 task = current;
1759         else
1760                 task = find_task_by_vpid(pid);
1761         if (task)
1762                 get_task_struct(task);
1763         rcu_read_unlock();
1764
1765         if (!task)
1766                 return ERR_PTR(-ESRCH);
1767
1768         /*
1769          * Can't attach events to a dying task.
1770          */
1771         err = -ESRCH;
1772         if (task->flags & PF_EXITING)
1773                 goto errout;
1774
1775         /* Reuse ptrace permission checks for now. */
1776         err = -EACCES;
1777         if (!ptrace_may_access(task, PTRACE_MODE_READ))
1778                 goto errout;
1779
1780  retry:
1781         ctx = perf_lock_task_context(task, &flags);
1782         if (ctx) {
1783                 unclone_ctx(ctx);
1784                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1785         }
1786
1787         if (!ctx) {
1788                 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1789                 err = -ENOMEM;
1790                 if (!ctx)
1791                         goto errout;
1792                 __perf_event_init_context(ctx, task);
1793                 get_ctx(ctx);
1794                 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1795                         /*
1796                          * We raced with some other task; use
1797                          * the context they set.
1798                          */
1799                         kfree(ctx);
1800                         goto retry;
1801                 }
1802                 get_task_struct(task);
1803         }
1804
1805         put_task_struct(task);
1806         return ctx;
1807
1808  errout:
1809         put_task_struct(task);
1810         return ERR_PTR(err);
1811 }
1812
1813 static void perf_event_free_filter(struct perf_event *event);
1814
1815 static void free_event_rcu(struct rcu_head *head)
1816 {
1817         struct perf_event *event;
1818
1819         event = container_of(head, struct perf_event, rcu_head);
1820         if (event->ns)
1821                 put_pid_ns(event->ns);
1822         perf_event_free_filter(event);
1823         kfree(event);
1824 }
1825
1826 static void perf_pending_sync(struct perf_event *event);
1827
1828 static void free_event(struct perf_event *event)
1829 {
1830         perf_pending_sync(event);
1831
1832         if (!event->parent) {
1833                 atomic_dec(&nr_events);
1834                 if (event->attr.mmap)
1835                         atomic_dec(&nr_mmap_events);
1836                 if (event->attr.comm)
1837                         atomic_dec(&nr_comm_events);
1838                 if (event->attr.task)
1839                         atomic_dec(&nr_task_events);
1840         }
1841
1842         if (event->output) {
1843                 fput(event->output->filp);
1844                 event->output = NULL;
1845         }
1846
1847         if (event->destroy)
1848                 event->destroy(event);
1849
1850         put_ctx(event->ctx);
1851         call_rcu(&event->rcu_head, free_event_rcu);
1852 }
1853
1854 int perf_event_release_kernel(struct perf_event *event)
1855 {
1856         struct perf_event_context *ctx = event->ctx;
1857
1858         WARN_ON_ONCE(ctx->parent_ctx);
1859         mutex_lock(&ctx->mutex);
1860         perf_event_remove_from_context(event);
1861         mutex_unlock(&ctx->mutex);
1862
1863         mutex_lock(&event->owner->perf_event_mutex);
1864         list_del_init(&event->owner_entry);
1865         mutex_unlock(&event->owner->perf_event_mutex);
1866         put_task_struct(event->owner);
1867
1868         free_event(event);
1869
1870         return 0;
1871 }
1872 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1873
1874 /*
1875  * Called when the last reference to the file is gone.
1876  */
1877 static int perf_release(struct inode *inode, struct file *file)
1878 {
1879         struct perf_event *event = file->private_data;
1880
1881         file->private_data = NULL;
1882
1883         return perf_event_release_kernel(event);
1884 }
1885
1886 static int perf_event_read_size(struct perf_event *event)
1887 {
1888         int entry = sizeof(u64); /* value */
1889         int size = 0;
1890         int nr = 1;
1891
1892         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1893                 size += sizeof(u64);
1894
1895         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1896                 size += sizeof(u64);
1897
1898         if (event->attr.read_format & PERF_FORMAT_ID)
1899                 entry += sizeof(u64);
1900
1901         if (event->attr.read_format & PERF_FORMAT_GROUP) {
1902                 nr += event->group_leader->nr_siblings;
1903                 size += sizeof(u64);
1904         }
1905
1906         size += entry * nr;
1907
1908         return size;
1909 }
1910
1911 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1912 {
1913         struct perf_event *child;
1914         u64 total = 0;
1915
1916         *enabled = 0;
1917         *running = 0;
1918
1919         mutex_lock(&event->child_mutex);
1920         total += perf_event_read(event);
1921         *enabled += event->total_time_enabled +
1922                         atomic64_read(&event->child_total_time_enabled);
1923         *running += event->total_time_running +
1924                         atomic64_read(&event->child_total_time_running);
1925
1926         list_for_each_entry(child, &event->child_list, child_list) {
1927                 total += perf_event_read(child);
1928                 *enabled += child->total_time_enabled;
1929                 *running += child->total_time_running;
1930         }
1931         mutex_unlock(&event->child_mutex);
1932
1933         return total;
1934 }
1935 EXPORT_SYMBOL_GPL(perf_event_read_value);
1936
1937 static int perf_event_read_group(struct perf_event *event,
1938                                    u64 read_format, char __user *buf)
1939 {
1940         struct perf_event *leader = event->group_leader, *sub;
1941         int n = 0, size = 0, ret = -EFAULT;
1942         struct perf_event_context *ctx = leader->ctx;
1943         u64 values[5];
1944         u64 count, enabled, running;
1945
1946         mutex_lock(&ctx->mutex);
1947         count = perf_event_read_value(leader, &enabled, &running);
1948
1949         values[n++] = 1 + leader->nr_siblings;
1950         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1951                 values[n++] = enabled;
1952         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1953                 values[n++] = running;
1954         values[n++] = count;
1955         if (read_format & PERF_FORMAT_ID)
1956                 values[n++] = primary_event_id(leader);
1957
1958         size = n * sizeof(u64);
1959
1960         if (copy_to_user(buf, values, size))
1961                 goto unlock;
1962
1963         ret = size;
1964
1965         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1966                 n = 0;
1967
1968                 values[n++] = perf_event_read_value(sub, &enabled, &running);
1969                 if (read_format & PERF_FORMAT_ID)
1970                         values[n++] = primary_event_id(sub);
1971
1972                 size = n * sizeof(u64);
1973
1974                 if (copy_to_user(buf + ret, values, size)) {
1975                         ret = -EFAULT;
1976                         goto unlock;
1977                 }
1978
1979                 ret += size;
1980         }
1981 unlock:
1982         mutex_unlock(&ctx->mutex);
1983
1984         return ret;
1985 }
1986
1987 static int perf_event_read_one(struct perf_event *event,
1988                                  u64 read_format, char __user *buf)
1989 {
1990         u64 enabled, running;
1991         u64 values[4];
1992         int n = 0;
1993
1994         values[n++] = perf_event_read_value(event, &enabled, &running);
1995         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1996                 values[n++] = enabled;
1997         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1998                 values[n++] = running;
1999         if (read_format & PERF_FORMAT_ID)
2000                 values[n++] = primary_event_id(event);
2001
2002         if (copy_to_user(buf, values, n * sizeof(u64)))
2003                 return -EFAULT;
2004
2005         return n * sizeof(u64);
2006 }
2007
2008 /*
2009  * Read the performance event - simple non blocking version for now
2010  */
2011 static ssize_t
2012 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2013 {
2014         u64 read_format = event->attr.read_format;
2015         int ret;
2016
2017         /*
2018          * Return end-of-file for a read on a event that is in
2019          * error state (i.e. because it was pinned but it couldn't be
2020          * scheduled on to the CPU at some point).
2021          */
2022         if (event->state == PERF_EVENT_STATE_ERROR)
2023                 return 0;
2024
2025         if (count < perf_event_read_size(event))
2026                 return -ENOSPC;
2027
2028         WARN_ON_ONCE(event->ctx->parent_ctx);
2029         if (read_format & PERF_FORMAT_GROUP)
2030                 ret = perf_event_read_group(event, read_format, buf);
2031         else
2032                 ret = perf_event_read_one(event, read_format, buf);
2033
2034         return ret;
2035 }
2036
2037 static ssize_t
2038 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2039 {
2040         struct perf_event *event = file->private_data;
2041
2042         return perf_read_hw(event, buf, count);
2043 }
2044
2045 static unsigned int perf_poll(struct file *file, poll_table *wait)
2046 {
2047         struct perf_event *event = file->private_data;
2048         struct perf_mmap_data *data;
2049         unsigned int events = POLL_HUP;
2050
2051         rcu_read_lock();
2052         data = rcu_dereference(event->data);
2053         if (data)
2054                 events = atomic_xchg(&data->poll, 0);
2055         rcu_read_unlock();
2056
2057         poll_wait(file, &event->waitq, wait);
2058
2059         return events;
2060 }
2061
2062 static void perf_event_reset(struct perf_event *event)
2063 {
2064         (void)perf_event_read(event);
2065         atomic64_set(&event->count, 0);
2066         perf_event_update_userpage(event);
2067 }
2068
2069 /*
2070  * Holding the top-level event's child_mutex means that any
2071  * descendant process that has inherited this event will block
2072  * in sync_child_event if it goes to exit, thus satisfying the
2073  * task existence requirements of perf_event_enable/disable.
2074  */
2075 static void perf_event_for_each_child(struct perf_event *event,
2076                                         void (*func)(struct perf_event *))
2077 {
2078         struct perf_event *child;
2079
2080         WARN_ON_ONCE(event->ctx->parent_ctx);
2081         mutex_lock(&event->child_mutex);
2082         func(event);
2083         list_for_each_entry(child, &event->child_list, child_list)
2084                 func(child);
2085         mutex_unlock(&event->child_mutex);
2086 }
2087
2088 static void perf_event_for_each(struct perf_event *event,
2089                                   void (*func)(struct perf_event *))
2090 {
2091         struct perf_event_context *ctx = event->ctx;
2092         struct perf_event *sibling;
2093
2094         WARN_ON_ONCE(ctx->parent_ctx);
2095         mutex_lock(&ctx->mutex);
2096         event = event->group_leader;
2097
2098         perf_event_for_each_child(event, func);
2099         func(event);
2100         list_for_each_entry(sibling, &event->sibling_list, group_entry)
2101                 perf_event_for_each_child(event, func);
2102         mutex_unlock(&ctx->mutex);
2103 }
2104
2105 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2106 {
2107         struct perf_event_context *ctx = event->ctx;
2108         unsigned long size;
2109         int ret = 0;
2110         u64 value;
2111
2112         if (!event->attr.sample_period)
2113                 return -EINVAL;
2114
2115         size = copy_from_user(&value, arg, sizeof(value));
2116         if (size != sizeof(value))
2117                 return -EFAULT;
2118
2119         if (!value)
2120                 return -EINVAL;
2121
2122         raw_spin_lock_irq(&ctx->lock);
2123         if (event->attr.freq) {
2124                 if (value > sysctl_perf_event_sample_rate) {
2125                         ret = -EINVAL;
2126                         goto unlock;
2127                 }
2128
2129                 event->attr.sample_freq = value;
2130         } else {
2131                 event->attr.sample_period = value;
2132                 event->hw.sample_period = value;
2133         }
2134 unlock:
2135         raw_spin_unlock_irq(&ctx->lock);
2136
2137         return ret;
2138 }
2139
2140 static int perf_event_set_output(struct perf_event *event, int output_fd);
2141 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2142
2143 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2144 {
2145         struct perf_event *event = file->private_data;
2146         void (*func)(struct perf_event *);
2147         u32 flags = arg;
2148
2149         switch (cmd) {
2150         case PERF_EVENT_IOC_ENABLE:
2151                 func = perf_event_enable;
2152                 break;
2153         case PERF_EVENT_IOC_DISABLE:
2154                 func = perf_event_disable;
2155                 break;
2156         case PERF_EVENT_IOC_RESET:
2157                 func = perf_event_reset;
2158                 break;
2159
2160         case PERF_EVENT_IOC_REFRESH:
2161                 return perf_event_refresh(event, arg);
2162
2163         case PERF_EVENT_IOC_PERIOD:
2164                 return perf_event_period(event, (u64 __user *)arg);
2165
2166         case PERF_EVENT_IOC_SET_OUTPUT:
2167                 return perf_event_set_output(event, arg);
2168
2169         case PERF_EVENT_IOC_SET_FILTER:
2170                 return perf_event_set_filter(event, (void __user *)arg);
2171
2172         default:
2173                 return -ENOTTY;
2174         }
2175
2176         if (flags & PERF_IOC_FLAG_GROUP)
2177                 perf_event_for_each(event, func);
2178         else
2179                 perf_event_for_each_child(event, func);
2180
2181         return 0;
2182 }
2183
2184 int perf_event_task_enable(void)
2185 {
2186         struct perf_event *event;
2187
2188         mutex_lock(&current->perf_event_mutex);
2189         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2190                 perf_event_for_each_child(event, perf_event_enable);
2191         mutex_unlock(&current->perf_event_mutex);
2192
2193         return 0;
2194 }
2195
2196 int perf_event_task_disable(void)
2197 {
2198         struct perf_event *event;
2199
2200         mutex_lock(&current->perf_event_mutex);
2201         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2202                 perf_event_for_each_child(event, perf_event_disable);
2203         mutex_unlock(&current->perf_event_mutex);
2204
2205         return 0;
2206 }
2207
2208 #ifndef PERF_EVENT_INDEX_OFFSET
2209 # define PERF_EVENT_INDEX_OFFSET 0
2210 #endif
2211
2212 static int perf_event_index(struct perf_event *event)
2213 {
2214         if (event->state != PERF_EVENT_STATE_ACTIVE)
2215                 return 0;
2216
2217         return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2218 }
2219
2220 /*
2221  * Callers need to ensure there can be no nesting of this function, otherwise
2222  * the seqlock logic goes bad. We can not serialize this because the arch
2223  * code calls this from NMI context.
2224  */
2225 void perf_event_update_userpage(struct perf_event *event)
2226 {
2227         struct perf_event_mmap_page *userpg;
2228         struct perf_mmap_data *data;
2229
2230         rcu_read_lock();
2231         data = rcu_dereference(event->data);
2232         if (!data)
2233                 goto unlock;
2234
2235         userpg = data->user_page;
2236
2237         /*
2238          * Disable preemption so as to not let the corresponding user-space
2239          * spin too long if we get preempted.
2240          */
2241         preempt_disable();
2242         ++userpg->lock;
2243         barrier();
2244         userpg->index = perf_event_index(event);
2245         userpg->offset = atomic64_read(&event->count);
2246         if (event->state == PERF_EVENT_STATE_ACTIVE)
2247                 userpg->offset -= atomic64_read(&event->hw.prev_count);
2248
2249         userpg->time_enabled = event->total_time_enabled +
2250                         atomic64_read(&event->child_total_time_enabled);
2251
2252         userpg->time_running = event->total_time_running +
2253                         atomic64_read(&event->child_total_time_running);
2254
2255         barrier();
2256         ++userpg->lock;
2257         preempt_enable();
2258 unlock:
2259         rcu_read_unlock();
2260 }
2261
2262 static unsigned long perf_data_size(struct perf_mmap_data *data)
2263 {
2264         return data->nr_pages << (PAGE_SHIFT + data->data_order);
2265 }
2266
2267 #ifndef CONFIG_PERF_USE_VMALLOC
2268
2269 /*
2270  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2271  */
2272
2273 static struct page *
2274 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2275 {
2276         if (pgoff > data->nr_pages)
2277                 return NULL;
2278
2279         if (pgoff == 0)
2280                 return virt_to_page(data->user_page);
2281
2282         return virt_to_page(data->data_pages[pgoff - 1]);
2283 }
2284
2285 static struct perf_mmap_data *
2286 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2287 {
2288         struct perf_mmap_data *data;
2289         unsigned long size;
2290         int i;
2291
2292         WARN_ON(atomic_read(&event->mmap_count));
2293
2294         size = sizeof(struct perf_mmap_data);
2295         size += nr_pages * sizeof(void *);
2296
2297         data = kzalloc(size, GFP_KERNEL);
2298         if (!data)
2299                 goto fail;
2300
2301         data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2302         if (!data->user_page)
2303                 goto fail_user_page;
2304
2305         for (i = 0; i < nr_pages; i++) {
2306                 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2307                 if (!data->data_pages[i])
2308                         goto fail_data_pages;
2309         }
2310
2311         data->data_order = 0;
2312         data->nr_pages = nr_pages;
2313
2314         return data;
2315
2316 fail_data_pages:
2317         for (i--; i >= 0; i--)
2318                 free_page((unsigned long)data->data_pages[i]);
2319
2320         free_page((unsigned long)data->user_page);
2321
2322 fail_user_page:
2323         kfree(data);
2324
2325 fail:
2326         return NULL;
2327 }
2328
2329 static void perf_mmap_free_page(unsigned long addr)
2330 {
2331         struct page *page = virt_to_page((void *)addr);
2332
2333         page->mapping = NULL;
2334         __free_page(page);
2335 }
2336
2337 static void perf_mmap_data_free(struct perf_mmap_data *data)
2338 {
2339         int i;
2340
2341         perf_mmap_free_page((unsigned long)data->user_page);
2342         for (i = 0; i < data->nr_pages; i++)
2343                 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2344         kfree(data);
2345 }
2346
2347 #else
2348
2349 /*
2350  * Back perf_mmap() with vmalloc memory.
2351  *
2352  * Required for architectures that have d-cache aliasing issues.
2353  */
2354
2355 static struct page *
2356 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2357 {
2358         if (pgoff > (1UL << data->data_order))
2359                 return NULL;
2360
2361         return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2362 }
2363
2364 static void perf_mmap_unmark_page(void *addr)
2365 {
2366         struct page *page = vmalloc_to_page(addr);
2367
2368         page->mapping = NULL;
2369 }
2370
2371 static void perf_mmap_data_free_work(struct work_struct *work)
2372 {
2373         struct perf_mmap_data *data;
2374         void *base;
2375         int i, nr;
2376
2377         data = container_of(work, struct perf_mmap_data, work);
2378         nr = 1 << data->data_order;
2379
2380         base = data->user_page;
2381         for (i = 0; i < nr + 1; i++)
2382                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2383
2384         vfree(base);
2385         kfree(data);
2386 }
2387
2388 static void perf_mmap_data_free(struct perf_mmap_data *data)
2389 {
2390         schedule_work(&data->work);
2391 }
2392
2393 static struct perf_mmap_data *
2394 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2395 {
2396         struct perf_mmap_data *data;
2397         unsigned long size;
2398         void *all_buf;
2399
2400         WARN_ON(atomic_read(&event->mmap_count));
2401
2402         size = sizeof(struct perf_mmap_data);
2403         size += sizeof(void *);
2404
2405         data = kzalloc(size, GFP_KERNEL);
2406         if (!data)
2407                 goto fail;
2408
2409         INIT_WORK(&data->work, perf_mmap_data_free_work);
2410
2411         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2412         if (!all_buf)
2413                 goto fail_all_buf;
2414
2415         data->user_page = all_buf;
2416         data->data_pages[0] = all_buf + PAGE_SIZE;
2417         data->data_order = ilog2(nr_pages);
2418         data->nr_pages = 1;
2419
2420         return data;
2421
2422 fail_all_buf:
2423         kfree(data);
2424
2425 fail:
2426         return NULL;
2427 }
2428
2429 #endif
2430
2431 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2432 {
2433         struct perf_event *event = vma->vm_file->private_data;
2434         struct perf_mmap_data *data;
2435         int ret = VM_FAULT_SIGBUS;
2436
2437         if (vmf->flags & FAULT_FLAG_MKWRITE) {
2438                 if (vmf->pgoff == 0)
2439                         ret = 0;
2440                 return ret;
2441         }
2442
2443         rcu_read_lock();
2444         data = rcu_dereference(event->data);
2445         if (!data)
2446                 goto unlock;
2447
2448         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2449                 goto unlock;
2450
2451         vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2452         if (!vmf->page)
2453                 goto unlock;
2454
2455         get_page(vmf->page);
2456         vmf->page->mapping = vma->vm_file->f_mapping;
2457         vmf->page->index   = vmf->pgoff;
2458
2459         ret = 0;
2460 unlock:
2461         rcu_read_unlock();
2462
2463         return ret;
2464 }
2465
2466 static void
2467 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2468 {
2469         long max_size = perf_data_size(data);
2470
2471         atomic_set(&data->lock, -1);
2472
2473         if (event->attr.watermark) {
2474                 data->watermark = min_t(long, max_size,
2475                                         event->attr.wakeup_watermark);
2476         }
2477
2478         if (!data->watermark)
2479                 data->watermark = max_size / 2;
2480
2481
2482         rcu_assign_pointer(event->data, data);
2483 }
2484
2485 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2486 {
2487         struct perf_mmap_data *data;
2488
2489         data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2490         perf_mmap_data_free(data);
2491 }
2492
2493 static void perf_mmap_data_release(struct perf_event *event)
2494 {
2495         struct perf_mmap_data *data = event->data;
2496
2497         WARN_ON(atomic_read(&event->mmap_count));
2498
2499         rcu_assign_pointer(event->data, NULL);
2500         call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2501 }
2502
2503 static void perf_mmap_open(struct vm_area_struct *vma)
2504 {
2505         struct perf_event *event = vma->vm_file->private_data;
2506
2507         atomic_inc(&event->mmap_count);
2508 }
2509
2510 static void perf_mmap_close(struct vm_area_struct *vma)
2511 {
2512         struct perf_event *event = vma->vm_file->private_data;
2513
2514         WARN_ON_ONCE(event->ctx->parent_ctx);
2515         if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2516                 unsigned long size = perf_data_size(event->data);
2517                 struct user_struct *user = current_user();
2518
2519                 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2520                 vma->vm_mm->locked_vm -= event->data->nr_locked;
2521                 perf_mmap_data_release(event);
2522                 mutex_unlock(&event->mmap_mutex);
2523         }
2524 }
2525
2526 static const struct vm_operations_struct perf_mmap_vmops = {
2527         .open           = perf_mmap_open,
2528         .close          = perf_mmap_close,
2529         .fault          = perf_mmap_fault,
2530         .page_mkwrite   = perf_mmap_fault,
2531 };
2532
2533 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2534 {
2535         struct perf_event *event = file->private_data;
2536         unsigned long user_locked, user_lock_limit;
2537         struct user_struct *user = current_user();
2538         unsigned long locked, lock_limit;
2539         struct perf_mmap_data *data;
2540         unsigned long vma_size;
2541         unsigned long nr_pages;
2542         long user_extra, extra;
2543         int ret = 0;
2544
2545         if (!(vma->vm_flags & VM_SHARED))
2546                 return -EINVAL;
2547
2548         vma_size = vma->vm_end - vma->vm_start;
2549         nr_pages = (vma_size / PAGE_SIZE) - 1;
2550
2551         /*
2552          * If we have data pages ensure they're a power-of-two number, so we
2553          * can do bitmasks instead of modulo.
2554          */
2555         if (nr_pages != 0 && !is_power_of_2(nr_pages))
2556                 return -EINVAL;
2557
2558         if (vma_size != PAGE_SIZE * (1 + nr_pages))
2559                 return -EINVAL;
2560
2561         if (vma->vm_pgoff != 0)
2562                 return -EINVAL;
2563
2564         WARN_ON_ONCE(event->ctx->parent_ctx);
2565         mutex_lock(&event->mmap_mutex);
2566         if (event->output) {
2567                 ret = -EINVAL;
2568                 goto unlock;
2569         }
2570
2571         if (atomic_inc_not_zero(&event->mmap_count)) {
2572                 if (nr_pages != event->data->nr_pages)
2573                         ret = -EINVAL;
2574                 goto unlock;
2575         }
2576
2577         user_extra = nr_pages + 1;
2578         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2579
2580         /*
2581          * Increase the limit linearly with more CPUs:
2582          */
2583         user_lock_limit *= num_online_cpus();
2584
2585         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2586
2587         extra = 0;
2588         if (user_locked > user_lock_limit)
2589                 extra = user_locked - user_lock_limit;
2590
2591         lock_limit = rlimit(RLIMIT_MEMLOCK);
2592         lock_limit >>= PAGE_SHIFT;
2593         locked = vma->vm_mm->locked_vm + extra;
2594
2595         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2596                 !capable(CAP_IPC_LOCK)) {
2597                 ret = -EPERM;
2598                 goto unlock;
2599         }
2600
2601         WARN_ON(event->data);
2602
2603         data = perf_mmap_data_alloc(event, nr_pages);
2604         ret = -ENOMEM;
2605         if (!data)
2606                 goto unlock;
2607
2608         ret = 0;
2609         perf_mmap_data_init(event, data);
2610
2611         atomic_set(&event->mmap_count, 1);
2612         atomic_long_add(user_extra, &user->locked_vm);
2613         vma->vm_mm->locked_vm += extra;
2614         event->data->nr_locked = extra;
2615         if (vma->vm_flags & VM_WRITE)
2616                 event->data->writable = 1;
2617
2618 unlock:
2619         mutex_unlock(&event->mmap_mutex);
2620
2621         vma->vm_flags |= VM_RESERVED;
2622         vma->vm_ops = &perf_mmap_vmops;
2623
2624         return ret;
2625 }
2626
2627 static int perf_fasync(int fd, struct file *filp, int on)
2628 {
2629         struct inode *inode = filp->f_path.dentry->d_inode;
2630         struct perf_event *event = filp->private_data;
2631         int retval;
2632
2633         mutex_lock(&inode->i_mutex);
2634         retval = fasync_helper(fd, filp, on, &event->fasync);
2635         mutex_unlock(&inode->i_mutex);
2636
2637         if (retval < 0)
2638                 return retval;
2639
2640         return 0;
2641 }
2642
2643 static const struct file_operations perf_fops = {
2644         .release                = perf_release,
2645         .read                   = perf_read,
2646         .poll                   = perf_poll,
2647         .unlocked_ioctl         = perf_ioctl,
2648         .compat_ioctl           = perf_ioctl,
2649         .mmap                   = perf_mmap,
2650         .fasync                 = perf_fasync,
2651 };
2652
2653 /*
2654  * Perf event wakeup
2655  *
2656  * If there's data, ensure we set the poll() state and publish everything
2657  * to user-space before waking everybody up.
2658  */
2659
2660 void perf_event_wakeup(struct perf_event *event)
2661 {
2662         wake_up_all(&event->waitq);
2663
2664         if (event->pending_kill) {
2665                 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2666                 event->pending_kill = 0;
2667         }
2668 }
2669
2670 /*
2671  * Pending wakeups
2672  *
2673  * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2674  *
2675  * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2676  * single linked list and use cmpxchg() to add entries lockless.
2677  */
2678
2679 static void perf_pending_event(struct perf_pending_entry *entry)
2680 {
2681         struct perf_event *event = container_of(entry,
2682                         struct perf_event, pending);
2683
2684         if (event->pending_disable) {
2685                 event->pending_disable = 0;
2686                 __perf_event_disable(event);
2687         }
2688
2689         if (event->pending_wakeup) {
2690                 event->pending_wakeup = 0;
2691                 perf_event_wakeup(event);
2692         }
2693 }
2694
2695 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2696
2697 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2698         PENDING_TAIL,
2699 };
2700
2701 static void perf_pending_queue(struct perf_pending_entry *entry,
2702                                void (*func)(struct perf_pending_entry *))
2703 {
2704         struct perf_pending_entry **head;
2705
2706         if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2707                 return;
2708
2709         entry->func = func;
2710
2711         head = &get_cpu_var(perf_pending_head);
2712
2713         do {
2714                 entry->next = *head;
2715         } while (cmpxchg(head, entry->next, entry) != entry->next);
2716
2717         set_perf_event_pending();
2718
2719         put_cpu_var(perf_pending_head);
2720 }
2721
2722 static int __perf_pending_run(void)
2723 {
2724         struct perf_pending_entry *list;
2725         int nr = 0;
2726
2727         list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2728         while (list != PENDING_TAIL) {
2729                 void (*func)(struct perf_pending_entry *);
2730                 struct perf_pending_entry *entry = list;
2731
2732                 list = list->next;
2733
2734                 func = entry->func;
2735                 entry->next = NULL;
2736                 /*
2737                  * Ensure we observe the unqueue before we issue the wakeup,
2738                  * so that we won't be waiting forever.
2739                  * -- see perf_not_pending().
2740                  */
2741                 smp_wmb();
2742
2743                 func(entry);
2744                 nr++;
2745         }
2746
2747         return nr;
2748 }
2749
2750 static inline int perf_not_pending(struct perf_event *event)
2751 {
2752         /*
2753          * If we flush on whatever cpu we run, there is a chance we don't
2754          * need to wait.
2755          */
2756         get_cpu();
2757         __perf_pending_run();
2758         put_cpu();
2759
2760         /*
2761          * Ensure we see the proper queue state before going to sleep
2762          * so that we do not miss the wakeup. -- see perf_pending_handle()
2763          */
2764         smp_rmb();
2765         return event->pending.next == NULL;
2766 }
2767
2768 static void perf_pending_sync(struct perf_event *event)
2769 {
2770         wait_event(event->waitq, perf_not_pending(event));
2771 }
2772
2773 void perf_event_do_pending(void)
2774 {
2775         __perf_pending_run();
2776 }
2777
2778 /*
2779  * Callchain support -- arch specific
2780  */
2781
2782 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2783 {
2784         return NULL;
2785 }
2786
2787 #ifdef CONFIG_EVENT_TRACING
2788 __weak
2789 void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip, int skip)
2790 {
2791 }
2792 #endif
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         struct task_struct *task = task_event->task;
3380         unsigned long flags;
3381         int size, ret;
3382
3383         /*
3384          * If this CPU attempts to acquire an rq lock held by a CPU spinning
3385          * in perf_output_lock() from interrupt context, it's game over.
3386          */
3387         local_irq_save(flags);
3388
3389         size  = task_event->event_id.header.size;
3390         ret = perf_output_begin(&handle, event, size, 0, 0);
3391
3392         if (ret) {
3393                 local_irq_restore(flags);
3394                 return;
3395         }
3396
3397         task_event->event_id.pid = perf_event_pid(event, task);
3398         task_event->event_id.ppid = perf_event_pid(event, current);
3399
3400         task_event->event_id.tid = perf_event_tid(event, task);
3401         task_event->event_id.ptid = perf_event_tid(event, current);
3402
3403         perf_output_put(&handle, task_event->event_id);
3404
3405         perf_output_end(&handle);
3406         local_irq_restore(flags);
3407 }
3408
3409 static int perf_event_task_match(struct perf_event *event)
3410 {
3411         if (event->state < PERF_EVENT_STATE_INACTIVE)
3412                 return 0;
3413
3414         if (event->cpu != -1 && event->cpu != smp_processor_id())
3415                 return 0;
3416
3417         if (event->attr.comm || event->attr.mmap || event->attr.task)
3418                 return 1;
3419
3420         return 0;
3421 }
3422
3423 static void perf_event_task_ctx(struct perf_event_context *ctx,
3424                                   struct perf_task_event *task_event)
3425 {
3426         struct perf_event *event;
3427
3428         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3429                 if (perf_event_task_match(event))
3430                         perf_event_task_output(event, task_event);
3431         }
3432 }
3433
3434 static void perf_event_task_event(struct perf_task_event *task_event)
3435 {
3436         struct perf_cpu_context *cpuctx;
3437         struct perf_event_context *ctx = task_event->task_ctx;
3438
3439         rcu_read_lock();
3440         cpuctx = &get_cpu_var(perf_cpu_context);
3441         perf_event_task_ctx(&cpuctx->ctx, task_event);
3442         if (!ctx)
3443                 ctx = rcu_dereference(current->perf_event_ctxp);
3444         if (ctx)
3445                 perf_event_task_ctx(ctx, task_event);
3446         put_cpu_var(perf_cpu_context);
3447         rcu_read_unlock();
3448 }
3449
3450 static void perf_event_task(struct task_struct *task,
3451                               struct perf_event_context *task_ctx,
3452                               int new)
3453 {
3454         struct perf_task_event task_event;
3455
3456         if (!atomic_read(&nr_comm_events) &&
3457             !atomic_read(&nr_mmap_events) &&
3458             !atomic_read(&nr_task_events))
3459                 return;
3460
3461         task_event = (struct perf_task_event){
3462                 .task     = task,
3463                 .task_ctx = task_ctx,
3464                 .event_id    = {
3465                         .header = {
3466                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3467                                 .misc = 0,
3468                                 .size = sizeof(task_event.event_id),
3469                         },
3470                         /* .pid  */
3471                         /* .ppid */
3472                         /* .tid  */
3473                         /* .ptid */
3474                         .time = perf_clock(),
3475                 },
3476         };
3477
3478         perf_event_task_event(&task_event);
3479 }
3480
3481 void perf_event_fork(struct task_struct *task)
3482 {
3483         perf_event_task(task, NULL, 1);
3484 }
3485
3486 /*
3487  * comm tracking
3488  */
3489
3490 struct perf_comm_event {
3491         struct task_struct      *task;
3492         char                    *comm;
3493         int                     comm_size;
3494
3495         struct {
3496                 struct perf_event_header        header;
3497
3498                 u32                             pid;
3499                 u32                             tid;
3500         } event_id;
3501 };
3502
3503 static void perf_event_comm_output(struct perf_event *event,
3504                                      struct perf_comm_event *comm_event)
3505 {
3506         struct perf_output_handle handle;
3507         int size = comm_event->event_id.header.size;
3508         int ret = perf_output_begin(&handle, event, size, 0, 0);
3509
3510         if (ret)
3511                 return;
3512
3513         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3514         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3515
3516         perf_output_put(&handle, comm_event->event_id);
3517         perf_output_copy(&handle, comm_event->comm,
3518                                    comm_event->comm_size);
3519         perf_output_end(&handle);
3520 }
3521
3522 static int perf_event_comm_match(struct perf_event *event)
3523 {
3524         if (event->state < PERF_EVENT_STATE_INACTIVE)
3525                 return 0;
3526
3527         if (event->cpu != -1 && event->cpu != smp_processor_id())
3528                 return 0;
3529
3530         if (event->attr.comm)
3531                 return 1;
3532
3533         return 0;
3534 }
3535
3536 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3537                                   struct perf_comm_event *comm_event)
3538 {
3539         struct perf_event *event;
3540
3541         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3542                 if (perf_event_comm_match(event))
3543                         perf_event_comm_output(event, comm_event);
3544         }
3545 }
3546
3547 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3548 {
3549         struct perf_cpu_context *cpuctx;
3550         struct perf_event_context *ctx;
3551         unsigned int size;
3552         char comm[TASK_COMM_LEN];
3553
3554         memset(comm, 0, sizeof(comm));
3555         strlcpy(comm, comm_event->task->comm, sizeof(comm));
3556         size = ALIGN(strlen(comm)+1, sizeof(u64));
3557
3558         comm_event->comm = comm;
3559         comm_event->comm_size = size;
3560
3561         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3562
3563         rcu_read_lock();
3564         cpuctx = &get_cpu_var(perf_cpu_context);
3565         perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3566         ctx = rcu_dereference(current->perf_event_ctxp);
3567         if (ctx)
3568                 perf_event_comm_ctx(ctx, comm_event);
3569         put_cpu_var(perf_cpu_context);
3570         rcu_read_unlock();
3571 }
3572
3573 void perf_event_comm(struct task_struct *task)
3574 {
3575         struct perf_comm_event comm_event;
3576
3577         if (task->perf_event_ctxp)
3578                 perf_event_enable_on_exec(task);
3579
3580         if (!atomic_read(&nr_comm_events))
3581                 return;
3582
3583         comm_event = (struct perf_comm_event){
3584                 .task   = task,
3585                 /* .comm      */
3586                 /* .comm_size */
3587                 .event_id  = {
3588                         .header = {
3589                                 .type = PERF_RECORD_COMM,
3590                                 .misc = 0,
3591                                 /* .size */
3592                         },
3593                         /* .pid */
3594                         /* .tid */
3595                 },
3596         };
3597
3598         perf_event_comm_event(&comm_event);
3599 }
3600
3601 /*
3602  * mmap tracking
3603  */
3604
3605 struct perf_mmap_event {
3606         struct vm_area_struct   *vma;
3607
3608         const char              *file_name;
3609         int                     file_size;
3610
3611         struct {
3612                 struct perf_event_header        header;
3613
3614                 u32                             pid;
3615                 u32                             tid;
3616                 u64                             start;
3617                 u64                             len;
3618                 u64                             pgoff;
3619         } event_id;
3620 };
3621
3622 static void perf_event_mmap_output(struct perf_event *event,
3623                                      struct perf_mmap_event *mmap_event)
3624 {
3625         struct perf_output_handle handle;
3626         int size = mmap_event->event_id.header.size;
3627         int ret = perf_output_begin(&handle, event, size, 0, 0);
3628
3629         if (ret)
3630                 return;
3631
3632         mmap_event->event_id.pid = perf_event_pid(event, current);
3633         mmap_event->event_id.tid = perf_event_tid(event, current);
3634
3635         perf_output_put(&handle, mmap_event->event_id);
3636         perf_output_copy(&handle, mmap_event->file_name,
3637                                    mmap_event->file_size);
3638         perf_output_end(&handle);
3639 }
3640
3641 static int perf_event_mmap_match(struct perf_event *event,
3642                                    struct perf_mmap_event *mmap_event)
3643 {
3644         if (event->state < PERF_EVENT_STATE_INACTIVE)
3645                 return 0;
3646
3647         if (event->cpu != -1 && event->cpu != smp_processor_id())
3648                 return 0;
3649
3650         if (event->attr.mmap)
3651                 return 1;
3652
3653         return 0;
3654 }
3655
3656 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3657                                   struct perf_mmap_event *mmap_event)
3658 {
3659         struct perf_event *event;
3660
3661         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3662                 if (perf_event_mmap_match(event, mmap_event))
3663                         perf_event_mmap_output(event, mmap_event);
3664         }
3665 }
3666
3667 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3668 {
3669         struct perf_cpu_context *cpuctx;
3670         struct perf_event_context *ctx;
3671         struct vm_area_struct *vma = mmap_event->vma;
3672         struct file *file = vma->vm_file;
3673         unsigned int size;
3674         char tmp[16];
3675         char *buf = NULL;
3676         const char *name;
3677
3678         memset(tmp, 0, sizeof(tmp));
3679
3680         if (file) {
3681                 /*
3682                  * d_path works from the end of the buffer backwards, so we
3683                  * need to add enough zero bytes after the string to handle
3684                  * the 64bit alignment we do later.
3685                  */
3686                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3687                 if (!buf) {
3688                         name = strncpy(tmp, "//enomem", sizeof(tmp));
3689                         goto got_name;
3690                 }
3691                 name = d_path(&file->f_path, buf, PATH_MAX);
3692                 if (IS_ERR(name)) {
3693                         name = strncpy(tmp, "//toolong", sizeof(tmp));
3694                         goto got_name;
3695                 }
3696         } else {
3697                 if (arch_vma_name(mmap_event->vma)) {
3698                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3699                                        sizeof(tmp));
3700                         goto got_name;
3701                 }
3702
3703                 if (!vma->vm_mm) {
3704                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
3705                         goto got_name;
3706                 }
3707
3708                 name = strncpy(tmp, "//anon", sizeof(tmp));
3709                 goto got_name;
3710         }
3711
3712 got_name:
3713         size = ALIGN(strlen(name)+1, sizeof(u64));
3714
3715         mmap_event->file_name = name;
3716         mmap_event->file_size = size;
3717
3718         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3719
3720         rcu_read_lock();
3721         cpuctx = &get_cpu_var(perf_cpu_context);
3722         perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3723         ctx = rcu_dereference(current->perf_event_ctxp);
3724         if (ctx)
3725                 perf_event_mmap_ctx(ctx, mmap_event);
3726         put_cpu_var(perf_cpu_context);
3727         rcu_read_unlock();
3728
3729         kfree(buf);
3730 }
3731
3732 void __perf_event_mmap(struct vm_area_struct *vma)
3733 {
3734         struct perf_mmap_event mmap_event;
3735
3736         if (!atomic_read(&nr_mmap_events))
3737                 return;
3738
3739         mmap_event = (struct perf_mmap_event){
3740                 .vma    = vma,
3741                 /* .file_name */
3742                 /* .file_size */
3743                 .event_id  = {
3744                         .header = {
3745                                 .type = PERF_RECORD_MMAP,
3746                                 .misc = 0,
3747                                 /* .size */
3748                         },
3749                         /* .pid */
3750                         /* .tid */
3751                         .start  = vma->vm_start,
3752                         .len    = vma->vm_end - vma->vm_start,
3753                         .pgoff  = (u64)vma->vm_pgoff << PAGE_SHIFT,
3754                 },
3755         };
3756
3757         perf_event_mmap_event(&mmap_event);
3758 }
3759
3760 /*
3761  * IRQ throttle logging
3762  */
3763
3764 static void perf_log_throttle(struct perf_event *event, int enable)
3765 {
3766         struct perf_output_handle handle;
3767         int ret;
3768
3769         struct {
3770                 struct perf_event_header        header;
3771                 u64                             time;
3772                 u64                             id;
3773                 u64                             stream_id;
3774         } throttle_event = {
3775                 .header = {
3776                         .type = PERF_RECORD_THROTTLE,
3777                         .misc = 0,
3778                         .size = sizeof(throttle_event),
3779                 },
3780                 .time           = perf_clock(),
3781                 .id             = primary_event_id(event),
3782                 .stream_id      = event->id,
3783         };
3784
3785         if (enable)
3786                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3787
3788         ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3789         if (ret)
3790                 return;
3791
3792         perf_output_put(&handle, throttle_event);
3793         perf_output_end(&handle);
3794 }
3795
3796 /*
3797  * Generic event overflow handling, sampling.
3798  */
3799
3800 static int __perf_event_overflow(struct perf_event *event, int nmi,
3801                                    int throttle, struct perf_sample_data *data,
3802                                    struct pt_regs *regs)
3803 {
3804         int events = atomic_read(&event->event_limit);
3805         struct hw_perf_event *hwc = &event->hw;
3806         int ret = 0;
3807
3808         throttle = (throttle && event->pmu->unthrottle != NULL);
3809
3810         if (!throttle) {
3811                 hwc->interrupts++;
3812         } else {
3813                 if (hwc->interrupts != MAX_INTERRUPTS) {
3814                         hwc->interrupts++;
3815                         if (HZ * hwc->interrupts >
3816                                         (u64)sysctl_perf_event_sample_rate) {
3817                                 hwc->interrupts = MAX_INTERRUPTS;
3818                                 perf_log_throttle(event, 0);
3819                                 ret = 1;
3820                         }
3821                 } else {
3822                         /*
3823                          * Keep re-disabling events even though on the previous
3824                          * pass we disabled it - just in case we raced with a
3825                          * sched-in and the event got enabled again:
3826                          */
3827                         ret = 1;
3828                 }
3829         }
3830
3831         if (event->attr.freq) {
3832                 u64 now = perf_clock();
3833                 s64 delta = now - hwc->freq_time_stamp;
3834
3835                 hwc->freq_time_stamp = now;
3836
3837                 if (delta > 0 && delta < 2*TICK_NSEC)
3838                         perf_adjust_period(event, delta, hwc->last_period);
3839         }
3840
3841         /*
3842          * XXX event_limit might not quite work as expected on inherited
3843          * events
3844          */
3845
3846         event->pending_kill = POLL_IN;
3847         if (events && atomic_dec_and_test(&event->event_limit)) {
3848                 ret = 1;
3849                 event->pending_kill = POLL_HUP;
3850                 if (nmi) {
3851                         event->pending_disable = 1;
3852                         perf_pending_queue(&event->pending,
3853                                            perf_pending_event);
3854                 } else
3855                         perf_event_disable(event);
3856         }
3857
3858         if (event->overflow_handler)
3859                 event->overflow_handler(event, nmi, data, regs);
3860         else
3861                 perf_event_output(event, nmi, data, regs);
3862
3863         return ret;
3864 }
3865
3866 int perf_event_overflow(struct perf_event *event, int nmi,
3867                           struct perf_sample_data *data,
3868                           struct pt_regs *regs)
3869 {
3870         return __perf_event_overflow(event, nmi, 1, data, regs);
3871 }
3872
3873 /*
3874  * Generic software event infrastructure
3875  */
3876
3877 /*
3878  * We directly increment event->count and keep a second value in
3879  * event->hw.period_left to count intervals. This period event
3880  * is kept in the range [-sample_period, 0] so that we can use the
3881  * sign as trigger.
3882  */
3883
3884 static u64 perf_swevent_set_period(struct perf_event *event)
3885 {
3886         struct hw_perf_event *hwc = &event->hw;
3887         u64 period = hwc->last_period;
3888         u64 nr, offset;
3889         s64 old, val;
3890
3891         hwc->last_period = hwc->sample_period;
3892
3893 again:
3894         old = val = atomic64_read(&hwc->period_left);
3895         if (val < 0)
3896                 return 0;
3897
3898         nr = div64_u64(period + val, period);
3899         offset = nr * period;
3900         val -= offset;
3901         if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3902                 goto again;
3903
3904         return nr;
3905 }
3906
3907 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3908                                     int nmi, struct perf_sample_data *data,
3909                                     struct pt_regs *regs)
3910 {
3911         struct hw_perf_event *hwc = &event->hw;
3912         int throttle = 0;
3913
3914         data->period = event->hw.last_period;
3915         if (!overflow)
3916                 overflow = perf_swevent_set_period(event);
3917
3918         if (hwc->interrupts == MAX_INTERRUPTS)
3919                 return;
3920
3921         for (; overflow; overflow--) {
3922                 if (__perf_event_overflow(event, nmi, throttle,
3923                                             data, regs)) {
3924                         /*
3925                          * We inhibit the overflow from happening when
3926                          * hwc->interrupts == MAX_INTERRUPTS.
3927                          */
3928                         break;
3929                 }
3930                 throttle = 1;
3931         }
3932 }
3933
3934 static void perf_swevent_unthrottle(struct perf_event *event)
3935 {
3936         /*
3937          * Nothing to do, we already reset hwc->interrupts.
3938          */
3939 }
3940
3941 static void perf_swevent_add(struct perf_event *event, u64 nr,
3942                                int nmi, struct perf_sample_data *data,
3943                                struct pt_regs *regs)
3944 {
3945         struct hw_perf_event *hwc = &event->hw;
3946
3947         atomic64_add(nr, &event->count);
3948
3949         if (!regs)
3950                 return;
3951
3952         if (!hwc->sample_period)
3953                 return;
3954
3955         if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3956                 return perf_swevent_overflow(event, 1, nmi, data, regs);
3957
3958         if (atomic64_add_negative(nr, &hwc->period_left))
3959                 return;
3960
3961         perf_swevent_overflow(event, 0, nmi, data, regs);
3962 }
3963
3964 static int perf_swevent_is_counting(struct perf_event *event)
3965 {
3966         /*
3967          * The event is active, we're good!
3968          */
3969         if (event->state == PERF_EVENT_STATE_ACTIVE)
3970                 return 1;
3971
3972         /*
3973          * The event is off/error, not counting.
3974          */
3975         if (event->state != PERF_EVENT_STATE_INACTIVE)
3976                 return 0;
3977
3978         /*
3979          * The event is inactive, if the context is active
3980          * we're part of a group that didn't make it on the 'pmu',
3981          * not counting.
3982          */
3983         if (event->ctx->is_active)
3984                 return 0;
3985
3986         /*
3987          * We're inactive and the context is too, this means the
3988          * task is scheduled out, we're counting events that happen
3989          * to us, like migration events.
3990          */
3991         return 1;
3992 }
3993
3994 static int perf_tp_event_match(struct perf_event *event,
3995                                 struct perf_sample_data *data);
3996
3997 static int perf_exclude_event(struct perf_event *event,
3998                               struct pt_regs *regs)
3999 {
4000         if (regs) {
4001                 if (event->attr.exclude_user && user_mode(regs))
4002                         return 1;
4003
4004                 if (event->attr.exclude_kernel && !user_mode(regs))
4005                         return 1;
4006         }
4007
4008         return 0;
4009 }
4010
4011 static int perf_swevent_match(struct perf_event *event,
4012                                 enum perf_type_id type,
4013                                 u32 event_id,
4014                                 struct perf_sample_data *data,
4015                                 struct pt_regs *regs)
4016 {
4017         if (event->cpu != -1 && event->cpu != smp_processor_id())
4018                 return 0;
4019
4020         if (!perf_swevent_is_counting(event))
4021                 return 0;
4022
4023         if (event->attr.type != type)
4024                 return 0;
4025
4026         if (event->attr.config != event_id)
4027                 return 0;
4028
4029         if (perf_exclude_event(event, regs))
4030                 return 0;
4031
4032         if (event->attr.type == PERF_TYPE_TRACEPOINT &&
4033             !perf_tp_event_match(event, data))
4034                 return 0;
4035
4036         return 1;
4037 }
4038
4039 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
4040                                      enum perf_type_id type,
4041                                      u32 event_id, u64 nr, int nmi,
4042                                      struct perf_sample_data *data,
4043                                      struct pt_regs *regs)
4044 {
4045         struct perf_event *event;
4046
4047         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4048                 if (perf_swevent_match(event, type, event_id, data, regs))
4049                         perf_swevent_add(event, nr, nmi, data, regs);
4050         }
4051 }
4052
4053 int perf_swevent_get_recursion_context(void)
4054 {
4055         struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
4056         int rctx;
4057
4058         if (in_nmi())
4059                 rctx = 3;
4060         else if (in_irq())
4061                 rctx = 2;
4062         else if (in_softirq())
4063                 rctx = 1;
4064         else
4065                 rctx = 0;
4066
4067         if (cpuctx->recursion[rctx]) {
4068                 put_cpu_var(perf_cpu_context);
4069                 return -1;
4070         }
4071
4072         cpuctx->recursion[rctx]++;
4073         barrier();
4074
4075         return rctx;
4076 }
4077 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4078
4079 void perf_swevent_put_recursion_context(int rctx)
4080 {
4081         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4082         barrier();
4083         cpuctx->recursion[rctx]--;
4084         put_cpu_var(perf_cpu_context);
4085 }
4086 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
4087
4088 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4089                                     u64 nr, int nmi,
4090                                     struct perf_sample_data *data,
4091                                     struct pt_regs *regs)
4092 {
4093         struct perf_cpu_context *cpuctx;
4094         struct perf_event_context *ctx;
4095
4096         cpuctx = &__get_cpu_var(perf_cpu_context);
4097         rcu_read_lock();
4098         perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
4099                                  nr, nmi, data, regs);
4100         /*
4101          * doesn't really matter which of the child contexts the
4102          * events ends up in.
4103          */
4104         ctx = rcu_dereference(current->perf_event_ctxp);
4105         if (ctx)
4106                 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
4107         rcu_read_unlock();
4108 }
4109
4110 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4111                             struct pt_regs *regs, u64 addr)
4112 {
4113         struct perf_sample_data data;
4114         int rctx;
4115
4116         rctx = perf_swevent_get_recursion_context();
4117         if (rctx < 0)
4118                 return;
4119
4120         perf_sample_data_init(&data, addr);
4121
4122         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4123
4124         perf_swevent_put_recursion_context(rctx);
4125 }
4126
4127 static void perf_swevent_read(struct perf_event *event)
4128 {
4129 }
4130
4131 static int perf_swevent_enable(struct perf_event *event)
4132 {
4133         struct hw_perf_event *hwc = &event->hw;
4134
4135         if (hwc->sample_period) {
4136                 hwc->last_period = hwc->sample_period;
4137                 perf_swevent_set_period(event);
4138         }
4139         return 0;
4140 }
4141
4142 static void perf_swevent_disable(struct perf_event *event)
4143 {
4144 }
4145
4146 static const struct pmu perf_ops_generic = {
4147         .enable         = perf_swevent_enable,
4148         .disable        = perf_swevent_disable,
4149         .read           = perf_swevent_read,
4150         .unthrottle     = perf_swevent_unthrottle,
4151 };
4152
4153 /*
4154  * hrtimer based swevent callback
4155  */
4156
4157 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4158 {
4159         enum hrtimer_restart ret = HRTIMER_RESTART;
4160         struct perf_sample_data data;
4161         struct pt_regs *regs;
4162         struct perf_event *event;
4163         u64 period;
4164
4165         event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4166         event->pmu->read(event);
4167
4168         perf_sample_data_init(&data, 0);
4169         data.period = event->hw.last_period;
4170         regs = get_irq_regs();
4171         /*
4172          * In case we exclude kernel IPs or are somehow not in interrupt
4173          * context, provide the next best thing, the user IP.
4174          */
4175         if ((event->attr.exclude_kernel || !regs) &&
4176                         !event->attr.exclude_user)
4177                 regs = task_pt_regs(current);
4178
4179         if (regs) {
4180                 if (!(event->attr.exclude_idle && current->pid == 0))
4181                         if (perf_event_overflow(event, 0, &data, regs))
4182                                 ret = HRTIMER_NORESTART;
4183         }
4184
4185         period = max_t(u64, 10000, event->hw.sample_period);
4186         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4187
4188         return ret;
4189 }
4190
4191 static void perf_swevent_start_hrtimer(struct perf_event *event)
4192 {
4193         struct hw_perf_event *hwc = &event->hw;
4194
4195         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4196         hwc->hrtimer.function = perf_swevent_hrtimer;
4197         if (hwc->sample_period) {
4198                 u64 period;
4199
4200                 if (hwc->remaining) {
4201                         if (hwc->remaining < 0)
4202                                 period = 10000;
4203                         else
4204                                 period = hwc->remaining;
4205                         hwc->remaining = 0;
4206                 } else {
4207                         period = max_t(u64, 10000, hwc->sample_period);
4208                 }
4209                 __hrtimer_start_range_ns(&hwc->hrtimer,
4210                                 ns_to_ktime(period), 0,
4211                                 HRTIMER_MODE_REL, 0);
4212         }
4213 }
4214
4215 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4216 {
4217         struct hw_perf_event *hwc = &event->hw;
4218
4219         if (hwc->sample_period) {
4220                 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4221                 hwc->remaining = ktime_to_ns(remaining);
4222
4223                 hrtimer_cancel(&hwc->hrtimer);
4224         }
4225 }
4226
4227 /*
4228  * Software event: cpu wall time clock
4229  */
4230
4231 static void cpu_clock_perf_event_update(struct perf_event *event)
4232 {
4233         int cpu = raw_smp_processor_id();
4234         s64 prev;
4235         u64 now;
4236
4237         now = cpu_clock(cpu);
4238         prev = atomic64_xchg(&event->hw.prev_count, now);
4239         atomic64_add(now - prev, &event->count);
4240 }
4241
4242 static int cpu_clock_perf_event_enable(struct perf_event *event)
4243 {
4244         struct hw_perf_event *hwc = &event->hw;
4245         int cpu = raw_smp_processor_id();
4246
4247         atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4248         perf_swevent_start_hrtimer(event);
4249
4250         return 0;
4251 }
4252
4253 static void cpu_clock_perf_event_disable(struct perf_event *event)
4254 {
4255         perf_swevent_cancel_hrtimer(event);
4256         cpu_clock_perf_event_update(event);
4257 }
4258
4259 static void cpu_clock_perf_event_read(struct perf_event *event)
4260 {
4261         cpu_clock_perf_event_update(event);
4262 }
4263
4264 static const struct pmu perf_ops_cpu_clock = {
4265         .enable         = cpu_clock_perf_event_enable,
4266         .disable        = cpu_clock_perf_event_disable,
4267         .read           = cpu_clock_perf_event_read,
4268 };
4269
4270 /*
4271  * Software event: task time clock
4272  */
4273
4274 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4275 {
4276         u64 prev;
4277         s64 delta;
4278
4279         prev = atomic64_xchg(&event->hw.prev_count, now);
4280         delta = now - prev;
4281         atomic64_add(delta, &event->count);
4282 }
4283
4284 static int task_clock_perf_event_enable(struct perf_event *event)
4285 {
4286         struct hw_perf_event *hwc = &event->hw;
4287         u64 now;
4288
4289         now = event->ctx->time;
4290
4291         atomic64_set(&hwc->prev_count, now);
4292
4293         perf_swevent_start_hrtimer(event);
4294
4295         return 0;
4296 }
4297
4298 static void task_clock_perf_event_disable(struct perf_event *event)
4299 {
4300         perf_swevent_cancel_hrtimer(event);
4301         task_clock_perf_event_update(event, event->ctx->time);
4302
4303 }
4304
4305 static void task_clock_perf_event_read(struct perf_event *event)
4306 {
4307         u64 time;
4308
4309         if (!in_nmi()) {
4310                 update_context_time(event->ctx);
4311                 time = event->ctx->time;
4312         } else {
4313                 u64 now = perf_clock();
4314                 u64 delta = now - event->ctx->timestamp;
4315                 time = event->ctx->time + delta;
4316         }
4317
4318         task_clock_perf_event_update(event, time);
4319 }
4320
4321 static const struct pmu perf_ops_task_clock = {
4322         .enable         = task_clock_perf_event_enable,
4323         .disable        = task_clock_perf_event_disable,
4324         .read           = task_clock_perf_event_read,
4325 };
4326
4327 #ifdef CONFIG_EVENT_TRACING
4328
4329 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4330                    int entry_size, struct pt_regs *regs)
4331 {
4332         struct perf_sample_data data;
4333         struct perf_raw_record raw = {
4334                 .size = entry_size,
4335                 .data = record,
4336         };
4337
4338         perf_sample_data_init(&data, addr);
4339         data.raw = &raw;
4340
4341         /* Trace events already protected against recursion */
4342         do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4343                          &data, regs);
4344 }
4345 EXPORT_SYMBOL_GPL(perf_tp_event);
4346
4347 static int perf_tp_event_match(struct perf_event *event,
4348                                 struct perf_sample_data *data)
4349 {
4350         void *record = data->raw->data;
4351
4352         if (likely(!event->filter) || filter_match_preds(event->filter, record))
4353                 return 1;
4354         return 0;
4355 }
4356
4357 static void tp_perf_event_destroy(struct perf_event *event)
4358 {
4359         perf_trace_disable(event->attr.config);
4360 }
4361
4362 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4363 {
4364         /*
4365          * Raw tracepoint data is a severe data leak, only allow root to
4366          * have these.
4367          */
4368         if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4369                         perf_paranoid_tracepoint_raw() &&
4370                         !capable(CAP_SYS_ADMIN))
4371                 return ERR_PTR(-EPERM);
4372
4373         if (perf_trace_enable(event->attr.config))
4374                 return NULL;
4375
4376         event->destroy = tp_perf_event_destroy;
4377
4378         return &perf_ops_generic;
4379 }
4380
4381 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4382 {
4383         char *filter_str;
4384         int ret;
4385
4386         if (event->attr.type != PERF_TYPE_TRACEPOINT)
4387                 return -EINVAL;
4388
4389         filter_str = strndup_user(arg, PAGE_SIZE);
4390         if (IS_ERR(filter_str))
4391                 return PTR_ERR(filter_str);
4392
4393         ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4394
4395         kfree(filter_str);
4396         return ret;
4397 }
4398
4399 static void perf_event_free_filter(struct perf_event *event)
4400 {
4401         ftrace_profile_free_filter(event);
4402 }
4403
4404 #else
4405
4406 static int perf_tp_event_match(struct perf_event *event,
4407                                 struct perf_sample_data *data)
4408 {
4409         return 1;
4410 }
4411
4412 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4413 {
4414         return NULL;
4415 }
4416
4417 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4418 {
4419         return -ENOENT;
4420 }
4421
4422 static void perf_event_free_filter(struct perf_event *event)
4423 {
4424 }
4425
4426 #endif /* CONFIG_EVENT_TRACING */
4427
4428 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4429 static void bp_perf_event_destroy(struct perf_event *event)
4430 {
4431         release_bp_slot(event);
4432 }
4433
4434 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4435 {
4436         int err;
4437
4438         err = register_perf_hw_breakpoint(bp);
4439         if (err)
4440                 return ERR_PTR(err);
4441
4442         bp->destroy = bp_perf_event_destroy;
4443
4444         return &perf_ops_bp;
4445 }
4446
4447 void perf_bp_event(struct perf_event *bp, void *data)
4448 {
4449         struct perf_sample_data sample;
4450         struct pt_regs *regs = data;
4451
4452         perf_sample_data_init(&sample, bp->attr.bp_addr);
4453
4454         if (!perf_exclude_event(bp, regs))
4455                 perf_swevent_add(bp, 1, 1, &sample, regs);
4456 }
4457 #else
4458 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4459 {
4460         return NULL;
4461 }
4462
4463 void perf_bp_event(struct perf_event *bp, void *regs)
4464 {
4465 }
4466 #endif
4467
4468 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4469
4470 static void sw_perf_event_destroy(struct perf_event *event)
4471 {
4472         u64 event_id = event->attr.config;
4473
4474         WARN_ON(event->parent);
4475
4476         atomic_dec(&perf_swevent_enabled[event_id]);
4477 }
4478
4479 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4480 {
4481         const struct pmu *pmu = NULL;
4482         u64 event_id = event->attr.config;
4483
4484         /*
4485          * Software events (currently) can't in general distinguish
4486          * between user, kernel and hypervisor events.
4487          * However, context switches and cpu migrations are considered
4488          * to be kernel events, and page faults are never hypervisor
4489          * events.
4490          */
4491         switch (event_id) {
4492         case PERF_COUNT_SW_CPU_CLOCK:
4493                 pmu = &perf_ops_cpu_clock;
4494
4495                 break;
4496         case PERF_COUNT_SW_TASK_CLOCK:
4497                 /*
4498                  * If the user instantiates this as a per-cpu event,
4499                  * use the cpu_clock event instead.
4500                  */
4501                 if (event->ctx->task)
4502                         pmu = &perf_ops_task_clock;
4503                 else
4504                         pmu = &perf_ops_cpu_clock;
4505
4506                 break;
4507         case PERF_COUNT_SW_PAGE_FAULTS:
4508         case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4509         case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4510         case PERF_COUNT_SW_CONTEXT_SWITCHES:
4511         case PERF_COUNT_SW_CPU_MIGRATIONS:
4512         case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4513         case PERF_COUNT_SW_EMULATION_FAULTS:
4514                 if (!event->parent) {
4515                         atomic_inc(&perf_swevent_enabled[event_id]);
4516                         event->destroy = sw_perf_event_destroy;
4517                 }
4518                 pmu = &perf_ops_generic;
4519                 break;
4520         }
4521
4522         return pmu;
4523 }
4524
4525 /*
4526  * Allocate and initialize a event structure
4527  */
4528 static struct perf_event *
4529 perf_event_alloc(struct perf_event_attr *attr,
4530                    int cpu,
4531                    struct perf_event_context *ctx,
4532                    struct perf_event *group_leader,
4533                    struct perf_event *parent_event,
4534                    perf_overflow_handler_t overflow_handler,
4535                    gfp_t gfpflags)
4536 {
4537         const struct pmu *pmu;
4538         struct perf_event *event;
4539         struct hw_perf_event *hwc;
4540         long err;
4541
4542         event = kzalloc(sizeof(*event), gfpflags);
4543         if (!event)
4544                 return ERR_PTR(-ENOMEM);
4545
4546         /*
4547          * Single events are their own group leaders, with an
4548          * empty sibling list:
4549          */
4550         if (!group_leader)
4551                 group_leader = event;
4552
4553         mutex_init(&event->child_mutex);
4554         INIT_LIST_HEAD(&event->child_list);
4555
4556         INIT_LIST_HEAD(&event->group_entry);
4557         INIT_LIST_HEAD(&event->event_entry);
4558         INIT_LIST_HEAD(&event->sibling_list);
4559         init_waitqueue_head(&event->waitq);
4560
4561         mutex_init(&event->mmap_mutex);
4562
4563         event->cpu              = cpu;
4564         event->attr             = *attr;
4565         event->group_leader     = group_leader;
4566         event->pmu              = NULL;
4567         event->ctx              = ctx;
4568         event->oncpu            = -1;
4569
4570         event->parent           = parent_event;
4571
4572         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
4573         event->id               = atomic64_inc_return(&perf_event_id);
4574
4575         event->state            = PERF_EVENT_STATE_INACTIVE;
4576
4577         if (!overflow_handler && parent_event)
4578                 overflow_handler = parent_event->overflow_handler;
4579         
4580         event->overflow_handler = overflow_handler;
4581
4582         if (attr->disabled)
4583                 event->state = PERF_EVENT_STATE_OFF;
4584
4585         pmu = NULL;
4586
4587         hwc = &event->hw;
4588         hwc->sample_period = attr->sample_period;
4589         if (attr->freq && attr->sample_freq)
4590                 hwc->sample_period = 1;
4591         hwc->last_period = hwc->sample_period;
4592
4593         atomic64_set(&hwc->period_left, hwc->sample_period);
4594
4595         /*
4596          * we currently do not support PERF_FORMAT_GROUP on inherited events
4597          */
4598         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4599                 goto done;
4600
4601         switch (attr->type) {
4602         case PERF_TYPE_RAW:
4603         case PERF_TYPE_HARDWARE:
4604         case PERF_TYPE_HW_CACHE:
4605                 pmu = hw_perf_event_init(event);
4606                 break;
4607
4608         case PERF_TYPE_SOFTWARE:
4609                 pmu = sw_perf_event_init(event);
4610                 break;
4611
4612         case PERF_TYPE_TRACEPOINT:
4613                 pmu = tp_perf_event_init(event);
4614                 break;
4615
4616         case PERF_TYPE_BREAKPOINT:
4617                 pmu = bp_perf_event_init(event);
4618                 break;
4619
4620
4621         default:
4622                 break;
4623         }
4624 done:
4625         err = 0;
4626         if (!pmu)
4627                 err = -EINVAL;
4628         else if (IS_ERR(pmu))
4629                 err = PTR_ERR(pmu);
4630
4631         if (err) {
4632                 if (event->ns)
4633                         put_pid_ns(event->ns);
4634                 kfree(event);
4635                 return ERR_PTR(err);
4636         }
4637
4638         event->pmu = pmu;
4639
4640         if (!event->parent) {
4641                 atomic_inc(&nr_events);
4642                 if (event->attr.mmap)
4643                         atomic_inc(&nr_mmap_events);
4644                 if (event->attr.comm)
4645                         atomic_inc(&nr_comm_events);
4646                 if (event->attr.task)
4647                         atomic_inc(&nr_task_events);
4648         }
4649
4650         return event;
4651 }
4652
4653 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4654                           struct perf_event_attr *attr)
4655 {
4656         u32 size;
4657         int ret;
4658
4659         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4660                 return -EFAULT;
4661
4662         /*
4663          * zero the full structure, so that a short copy will be nice.
4664          */
4665         memset(attr, 0, sizeof(*attr));
4666
4667         ret = get_user(size, &uattr->size);
4668         if (ret)
4669                 return ret;
4670
4671         if (size > PAGE_SIZE)   /* silly large */
4672                 goto err_size;
4673
4674         if (!size)              /* abi compat */
4675                 size = PERF_ATTR_SIZE_VER0;
4676
4677         if (size < PERF_ATTR_SIZE_VER0)
4678                 goto err_size;
4679
4680         /*
4681          * If we're handed a bigger struct than we know of,
4682          * ensure all the unknown bits are 0 - i.e. new
4683          * user-space does not rely on any kernel feature
4684          * extensions we dont know about yet.
4685          */
4686         if (size > sizeof(*attr)) {
4687                 unsigned char __user *addr;
4688                 unsigned char __user *end;
4689                 unsigned char val;
4690
4691                 addr = (void __user *)uattr + sizeof(*attr);
4692                 end  = (void __user *)uattr + size;
4693
4694                 for (; addr < end; addr++) {
4695                         ret = get_user(val, addr);
4696                         if (ret)
4697                                 return ret;
4698                         if (val)
4699                                 goto err_size;
4700                 }
4701                 size = sizeof(*attr);
4702         }
4703
4704         ret = copy_from_user(attr, uattr, size);
4705         if (ret)
4706                 return -EFAULT;
4707
4708         /*
4709          * If the type exists, the corresponding creation will verify
4710          * the attr->config.
4711          */
4712         if (attr->type >= PERF_TYPE_MAX)
4713                 return -EINVAL;
4714
4715         if (attr->__reserved_1)
4716                 return -EINVAL;
4717
4718         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4719                 return -EINVAL;
4720
4721         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4722                 return -EINVAL;
4723
4724 out:
4725         return ret;
4726
4727 err_size:
4728         put_user(sizeof(*attr), &uattr->size);
4729         ret = -E2BIG;
4730         goto out;
4731 }
4732
4733 static int perf_event_set_output(struct perf_event *event, int output_fd)
4734 {
4735         struct perf_event *output_event = NULL;
4736         struct file *output_file = NULL;
4737         struct perf_event *old_output;
4738         int fput_needed = 0;
4739         int ret = -EINVAL;
4740
4741         if (!output_fd)
4742                 goto set;
4743
4744         output_file = fget_light(output_fd, &fput_needed);
4745         if (!output_file)
4746                 return -EBADF;
4747
4748         if (output_file->f_op != &perf_fops)
4749                 goto out;
4750
4751         output_event = output_file->private_data;
4752
4753         /* Don't chain output fds */
4754         if (output_event->output)
4755                 goto out;
4756
4757         /* Don't set an output fd when we already have an output channel */
4758         if (event->data)
4759                 goto out;
4760
4761         atomic_long_inc(&output_file->f_count);
4762
4763 set:
4764         mutex_lock(&event->mmap_mutex);
4765         old_output = event->output;
4766         rcu_assign_pointer(event->output, output_event);
4767         mutex_unlock(&event->mmap_mutex);
4768
4769         if (old_output) {
4770                 /*
4771                  * we need to make sure no existing perf_output_*()
4772                  * is still referencing this event.
4773                  */
4774                 synchronize_rcu();
4775                 fput(old_output->filp);
4776         }
4777
4778         ret = 0;
4779 out:
4780         fput_light(output_file, fput_needed);
4781         return ret;
4782 }
4783
4784 /**
4785  * sys_perf_event_open - open a performance event, associate it to a task/cpu
4786  *
4787  * @attr_uptr:  event_id type attributes for monitoring/sampling
4788  * @pid:                target pid
4789  * @cpu:                target cpu
4790  * @group_fd:           group leader event fd
4791  */
4792 SYSCALL_DEFINE5(perf_event_open,
4793                 struct perf_event_attr __user *, attr_uptr,
4794                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4795 {
4796         struct perf_event *event, *group_leader;
4797         struct perf_event_attr attr;
4798         struct perf_event_context *ctx;
4799         struct file *event_file = NULL;
4800         struct file *group_file = NULL;
4801         int fput_needed = 0;
4802         int fput_needed2 = 0;
4803         int err;
4804
4805         /* for future expandability... */
4806         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4807                 return -EINVAL;
4808
4809         err = perf_copy_attr(attr_uptr, &attr);
4810         if (err)
4811                 return err;
4812
4813         if (!attr.exclude_kernel) {
4814                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4815                         return -EACCES;
4816         }
4817
4818         if (attr.freq) {
4819                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4820                         return -EINVAL;
4821         }
4822
4823         /*
4824          * Get the target context (task or percpu):
4825          */
4826         ctx = find_get_context(pid, cpu);
4827         if (IS_ERR(ctx))
4828                 return PTR_ERR(ctx);
4829
4830         /*
4831          * Look up the group leader (we will attach this event to it):
4832          */
4833         group_leader = NULL;
4834         if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4835                 err = -EINVAL;
4836                 group_file = fget_light(group_fd, &fput_needed);
4837                 if (!group_file)
4838                         goto err_put_context;
4839                 if (group_file->f_op != &perf_fops)
4840                         goto err_put_context;
4841
4842                 group_leader = group_file->private_data;
4843                 /*
4844                  * Do not allow a recursive hierarchy (this new sibling
4845                  * becoming part of another group-sibling):
4846                  */
4847                 if (group_leader->group_leader != group_leader)
4848                         goto err_put_context;
4849                 /*
4850                  * Do not allow to attach to a group in a different
4851                  * task or CPU context:
4852                  */
4853                 if (group_leader->ctx != ctx)
4854                         goto err_put_context;
4855                 /*
4856                  * Only a group leader can be exclusive or pinned
4857                  */
4858                 if (attr.exclusive || attr.pinned)
4859                         goto err_put_context;
4860         }
4861
4862         event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4863                                      NULL, NULL, GFP_KERNEL);
4864         err = PTR_ERR(event);
4865         if (IS_ERR(event))
4866                 goto err_put_context;
4867
4868         err = anon_inode_getfd("[perf_event]", &perf_fops, event, O_RDWR);
4869         if (err < 0)
4870                 goto err_free_put_context;
4871
4872         event_file = fget_light(err, &fput_needed2);
4873         if (!event_file)
4874                 goto err_free_put_context;
4875
4876         if (flags & PERF_FLAG_FD_OUTPUT) {
4877                 err = perf_event_set_output(event, group_fd);
4878                 if (err)
4879                         goto err_fput_free_put_context;
4880         }
4881
4882         event->filp = event_file;
4883         WARN_ON_ONCE(ctx->parent_ctx);
4884         mutex_lock(&ctx->mutex);
4885         perf_install_in_context(ctx, event, cpu);
4886         ++ctx->generation;
4887         mutex_unlock(&ctx->mutex);
4888
4889         event->owner = current;
4890         get_task_struct(current);
4891         mutex_lock(&current->perf_event_mutex);
4892         list_add_tail(&event->owner_entry, &current->perf_event_list);
4893         mutex_unlock(&current->perf_event_mutex);
4894
4895 err_fput_free_put_context:
4896         fput_light(event_file, fput_needed2);
4897
4898 err_free_put_context:
4899         if (err < 0)
4900                 kfree(event);
4901
4902 err_put_context:
4903         if (err < 0)
4904                 put_ctx(ctx);
4905
4906         fput_light(group_file, fput_needed);
4907
4908         return err;
4909 }
4910
4911 /**
4912  * perf_event_create_kernel_counter
4913  *
4914  * @attr: attributes of the counter to create
4915  * @cpu: cpu in which the counter is bound
4916  * @pid: task to profile
4917  */
4918 struct perf_event *
4919 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4920                                  pid_t pid,
4921                                  perf_overflow_handler_t overflow_handler)
4922 {
4923         struct perf_event *event;
4924         struct perf_event_context *ctx;
4925         int err;
4926
4927         /*
4928          * Get the target context (task or percpu):
4929          */
4930
4931         ctx = find_get_context(pid, cpu);
4932         if (IS_ERR(ctx)) {
4933                 err = PTR_ERR(ctx);
4934                 goto err_exit;
4935         }
4936
4937         event = perf_event_alloc(attr, cpu, ctx, NULL,
4938                                  NULL, overflow_handler, GFP_KERNEL);
4939         if (IS_ERR(event)) {
4940                 err = PTR_ERR(event);
4941                 goto err_put_context;
4942         }
4943
4944         event->filp = NULL;
4945         WARN_ON_ONCE(ctx->parent_ctx);
4946         mutex_lock(&ctx->mutex);
4947         perf_install_in_context(ctx, event, cpu);
4948         ++ctx->generation;
4949         mutex_unlock(&ctx->mutex);
4950
4951         event->owner = current;
4952         get_task_struct(current);
4953         mutex_lock(&current->perf_event_mutex);
4954         list_add_tail(&event->owner_entry, &current->perf_event_list);
4955         mutex_unlock(&current->perf_event_mutex);
4956
4957         return event;
4958
4959  err_put_context:
4960         put_ctx(ctx);
4961  err_exit:
4962         return ERR_PTR(err);
4963 }
4964 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4965
4966 /*
4967  * inherit a event from parent task to child task:
4968  */
4969 static struct perf_event *
4970 inherit_event(struct perf_event *parent_event,
4971               struct task_struct *parent,
4972               struct perf_event_context *parent_ctx,
4973               struct task_struct *child,
4974               struct perf_event *group_leader,
4975               struct perf_event_context *child_ctx)
4976 {
4977         struct perf_event *child_event;
4978
4979         /*
4980          * Instead of creating recursive hierarchies of events,
4981          * we link inherited events back to the original parent,
4982          * which has a filp for sure, which we use as the reference
4983          * count:
4984          */
4985         if (parent_event->parent)
4986                 parent_event = parent_event->parent;
4987
4988         child_event = perf_event_alloc(&parent_event->attr,
4989                                            parent_event->cpu, child_ctx,
4990                                            group_leader, parent_event,
4991                                            NULL, GFP_KERNEL);
4992         if (IS_ERR(child_event))
4993                 return child_event;
4994         get_ctx(child_ctx);
4995
4996         /*
4997          * Make the child state follow the state of the parent event,
4998          * not its attr.disabled bit.  We hold the parent's mutex,
4999          * so we won't race with perf_event_{en, dis}able_family.
5000          */
5001         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
5002                 child_event->state = PERF_EVENT_STATE_INACTIVE;
5003         else
5004                 child_event->state = PERF_EVENT_STATE_OFF;
5005
5006         if (parent_event->attr.freq) {
5007                 u64 sample_period = parent_event->hw.sample_period;
5008                 struct hw_perf_event *hwc = &child_event->hw;
5009
5010                 hwc->sample_period = sample_period;
5011                 hwc->last_period   = sample_period;
5012
5013                 atomic64_set(&hwc->period_left, sample_period);
5014         }
5015
5016         child_event->overflow_handler = parent_event->overflow_handler;
5017
5018         /*
5019          * Link it up in the child's context:
5020          */
5021         add_event_to_ctx(child_event, child_ctx);
5022
5023         /*
5024          * Get a reference to the parent filp - we will fput it
5025          * when the child event exits. This is safe to do because
5026          * we are in the parent and we know that the filp still
5027          * exists and has a nonzero count:
5028          */
5029         atomic_long_inc(&parent_event->filp->f_count);
5030
5031         /*
5032          * Link this into the parent event's child list
5033          */
5034         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5035         mutex_lock(&parent_event->child_mutex);
5036         list_add_tail(&child_event->child_list, &parent_event->child_list);
5037         mutex_unlock(&parent_event->child_mutex);
5038
5039         return child_event;
5040 }
5041
5042 static int inherit_group(struct perf_event *parent_event,
5043               struct task_struct *parent,
5044               struct perf_event_context *parent_ctx,
5045               struct task_struct *child,
5046               struct perf_event_context *child_ctx)
5047 {
5048         struct perf_event *leader;
5049         struct perf_event *sub;
5050         struct perf_event *child_ctr;
5051
5052         leader = inherit_event(parent_event, parent, parent_ctx,
5053                                  child, NULL, child_ctx);
5054         if (IS_ERR(leader))
5055                 return PTR_ERR(leader);
5056         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5057                 child_ctr = inherit_event(sub, parent, parent_ctx,
5058                                             child, leader, child_ctx);
5059                 if (IS_ERR(child_ctr))
5060                         return PTR_ERR(child_ctr);
5061         }
5062         return 0;
5063 }
5064
5065 static void sync_child_event(struct perf_event *child_event,
5066                                struct task_struct *child)
5067 {
5068         struct perf_event *parent_event = child_event->parent;
5069         u64 child_val;
5070
5071         if (child_event->attr.inherit_stat)
5072                 perf_event_read_event(child_event, child);
5073
5074         child_val = atomic64_read(&child_event->count);
5075
5076         /*
5077          * Add back the child's count to the parent's count:
5078          */
5079         atomic64_add(child_val, &parent_event->count);
5080         atomic64_add(child_event->total_time_enabled,
5081                      &parent_event->child_total_time_enabled);
5082         atomic64_add(child_event->total_time_running,
5083                      &parent_event->child_total_time_running);
5084
5085         /*
5086          * Remove this event from the parent's list
5087          */
5088         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5089         mutex_lock(&parent_event->child_mutex);
5090         list_del_init(&child_event->child_list);
5091         mutex_unlock(&parent_event->child_mutex);
5092
5093         /*
5094          * Release the parent event, if this was the last
5095          * reference to it.
5096          */
5097         fput(parent_event->filp);
5098 }
5099
5100 static void
5101 __perf_event_exit_task(struct perf_event *child_event,
5102                          struct perf_event_context *child_ctx,
5103                          struct task_struct *child)
5104 {
5105         struct perf_event *parent_event;
5106
5107         perf_event_remove_from_context(child_event);
5108
5109         parent_event = child_event->parent;
5110         /*
5111          * It can happen that parent exits first, and has events
5112          * that are still around due to the child reference. These
5113          * events need to be zapped - but otherwise linger.
5114          */
5115         if (parent_event) {
5116                 sync_child_event(child_event, child);
5117                 free_event(child_event);
5118         }
5119 }
5120
5121 /*
5122  * When a child task exits, feed back event values to parent events.
5123  */
5124 void perf_event_exit_task(struct task_struct *child)
5125 {
5126         struct perf_event *child_event, *tmp;
5127         struct perf_event_context *child_ctx;
5128         unsigned long flags;
5129
5130         if (likely(!child->perf_event_ctxp)) {
5131                 perf_event_task(child, NULL, 0);
5132                 return;
5133         }
5134
5135         local_irq_save(flags);
5136         /*
5137          * We can't reschedule here because interrupts are disabled,
5138          * and either child is current or it is a task that can't be
5139          * scheduled, so we are now safe from rescheduling changing
5140          * our context.
5141          */
5142         child_ctx = child->perf_event_ctxp;
5143         __perf_event_task_sched_out(child_ctx);
5144
5145         /*
5146          * Take the context lock here so that if find_get_context is
5147          * reading child->perf_event_ctxp, we wait until it has
5148          * incremented the context's refcount before we do put_ctx below.
5149          */
5150         raw_spin_lock(&child_ctx->lock);
5151         child->perf_event_ctxp = NULL;
5152         /*
5153          * If this context is a clone; unclone it so it can't get
5154          * swapped to another process while we're removing all
5155          * the events from it.
5156          */
5157         unclone_ctx(child_ctx);
5158         update_context_time(child_ctx);
5159         raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5160
5161         /*
5162          * Report the task dead after unscheduling the events so that we
5163          * won't get any samples after PERF_RECORD_EXIT. We can however still
5164          * get a few PERF_RECORD_READ events.
5165          */
5166         perf_event_task(child, child_ctx, 0);
5167
5168         /*
5169          * We can recurse on the same lock type through:
5170          *
5171          *   __perf_event_exit_task()
5172          *     sync_child_event()
5173          *       fput(parent_event->filp)
5174          *         perf_release()
5175          *           mutex_lock(&ctx->mutex)
5176          *
5177          * But since its the parent context it won't be the same instance.
5178          */
5179         mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
5180
5181 again:
5182         list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5183                                  group_entry)
5184                 __perf_event_exit_task(child_event, child_ctx, child);
5185
5186         list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5187                                  group_entry)
5188                 __perf_event_exit_task(child_event, child_ctx, child);
5189
5190         /*
5191          * If the last event was a group event, it will have appended all
5192          * its siblings to the list, but we obtained 'tmp' before that which
5193          * will still point to the list head terminating the iteration.
5194          */
5195         if (!list_empty(&child_ctx->pinned_groups) ||
5196             !list_empty(&child_ctx->flexible_groups))
5197                 goto again;
5198
5199         mutex_unlock(&child_ctx->mutex);
5200
5201         put_ctx(child_ctx);
5202 }
5203
5204 static void perf_free_event(struct perf_event *event,
5205                             struct perf_event_context *ctx)
5206 {
5207         struct perf_event *parent = event->parent;
5208
5209         if (WARN_ON_ONCE(!parent))
5210                 return;
5211
5212         mutex_lock(&parent->child_mutex);
5213         list_del_init(&event->child_list);
5214         mutex_unlock(&parent->child_mutex);
5215
5216         fput(parent->filp);
5217
5218         list_del_event(event, ctx);
5219         free_event(event);
5220 }
5221
5222 /*
5223  * free an unexposed, unused context as created by inheritance by
5224  * init_task below, used by fork() in case of fail.
5225  */
5226 void perf_event_free_task(struct task_struct *task)
5227 {
5228         struct perf_event_context *ctx = task->perf_event_ctxp;
5229         struct perf_event *event, *tmp;
5230
5231         if (!ctx)
5232                 return;
5233
5234         mutex_lock(&ctx->mutex);
5235 again:
5236         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5237                 perf_free_event(event, ctx);
5238
5239         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5240                                  group_entry)
5241                 perf_free_event(event, ctx);
5242
5243         if (!list_empty(&ctx->pinned_groups) ||
5244             !list_empty(&ctx->flexible_groups))
5245                 goto again;
5246
5247         mutex_unlock(&ctx->mutex);
5248
5249         put_ctx(ctx);
5250 }
5251
5252 static int
5253 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5254                    struct perf_event_context *parent_ctx,
5255                    struct task_struct *child,
5256                    int *inherited_all)
5257 {
5258         int ret;
5259         struct perf_event_context *child_ctx = child->perf_event_ctxp;
5260
5261         if (!event->attr.inherit) {
5262                 *inherited_all = 0;
5263                 return 0;
5264         }
5265
5266         if (!child_ctx) {
5267                 /*
5268                  * This is executed from the parent task context, so
5269                  * inherit events that have been marked for cloning.
5270                  * First allocate and initialize a context for the
5271                  * child.
5272                  */
5273
5274                 child_ctx = kzalloc(sizeof(struct perf_event_context),
5275                                     GFP_KERNEL);
5276                 if (!child_ctx)
5277                         return -ENOMEM;
5278
5279                 __perf_event_init_context(child_ctx, child);
5280                 child->perf_event_ctxp = child_ctx;
5281                 get_task_struct(child);
5282         }
5283
5284         ret = inherit_group(event, parent, parent_ctx,
5285                             child, child_ctx);
5286
5287         if (ret)
5288                 *inherited_all = 0;
5289
5290         return ret;
5291 }
5292
5293
5294 /*
5295  * Initialize the perf_event context in task_struct
5296  */
5297 int perf_event_init_task(struct task_struct *child)
5298 {
5299         struct perf_event_context *child_ctx, *parent_ctx;
5300         struct perf_event_context *cloned_ctx;
5301         struct perf_event *event;
5302         struct task_struct *parent = current;
5303         int inherited_all = 1;
5304         int ret = 0;
5305
5306         child->perf_event_ctxp = NULL;
5307
5308         mutex_init(&child->perf_event_mutex);
5309         INIT_LIST_HEAD(&child->perf_event_list);
5310
5311         if (likely(!parent->perf_event_ctxp))
5312                 return 0;
5313
5314         /*
5315          * If the parent's context is a clone, pin it so it won't get
5316          * swapped under us.
5317          */
5318         parent_ctx = perf_pin_task_context(parent);
5319
5320         /*
5321          * No need to check if parent_ctx != NULL here; since we saw
5322          * it non-NULL earlier, the only reason for it to become NULL
5323          * is if we exit, and since we're currently in the middle of
5324          * a fork we can't be exiting at the same time.
5325          */
5326
5327         /*
5328          * Lock the parent list. No need to lock the child - not PID
5329          * hashed yet and not running, so nobody can access it.
5330          */
5331         mutex_lock(&parent_ctx->mutex);
5332
5333         /*
5334          * We dont have to disable NMIs - we are only looking at
5335          * the list, not manipulating it:
5336          */
5337         list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5338                 ret = inherit_task_group(event, parent, parent_ctx, child,
5339                                          &inherited_all);
5340                 if (ret)
5341                         break;
5342         }
5343
5344         list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5345                 ret = inherit_task_group(event, parent, parent_ctx, child,
5346                                          &inherited_all);
5347                 if (ret)
5348                         break;
5349         }
5350
5351         child_ctx = child->perf_event_ctxp;
5352
5353         if (child_ctx && inherited_all) {
5354                 /*
5355                  * Mark the child context as a clone of the parent
5356                  * context, or of whatever the parent is a clone of.
5357                  * Note that if the parent is a clone, it could get
5358                  * uncloned at any point, but that doesn't matter
5359                  * because the list of events and the generation
5360                  * count can't have changed since we took the mutex.
5361                  */
5362                 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5363                 if (cloned_ctx) {
5364                         child_ctx->parent_ctx = cloned_ctx;
5365                         child_ctx->parent_gen = parent_ctx->parent_gen;
5366                 } else {
5367                         child_ctx->parent_ctx = parent_ctx;
5368                         child_ctx->parent_gen = parent_ctx->generation;
5369                 }
5370                 get_ctx(child_ctx->parent_ctx);
5371         }
5372
5373         mutex_unlock(&parent_ctx->mutex);
5374
5375         perf_unpin_context(parent_ctx);
5376
5377         return ret;
5378 }
5379
5380 static void __init perf_event_init_all_cpus(void)
5381 {
5382         int cpu;
5383         struct perf_cpu_context *cpuctx;
5384
5385         for_each_possible_cpu(cpu) {
5386                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5387                 __perf_event_init_context(&cpuctx->ctx, NULL);
5388         }
5389 }
5390
5391 static void __cpuinit perf_event_init_cpu(int cpu)
5392 {
5393         struct perf_cpu_context *cpuctx;
5394
5395         cpuctx = &per_cpu(perf_cpu_context, cpu);
5396
5397         spin_lock(&perf_resource_lock);
5398         cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5399         spin_unlock(&perf_resource_lock);
5400 }
5401
5402 #ifdef CONFIG_HOTPLUG_CPU
5403 static void __perf_event_exit_cpu(void *info)
5404 {
5405         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5406         struct perf_event_context *ctx = &cpuctx->ctx;
5407         struct perf_event *event, *tmp;
5408
5409         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5410                 __perf_event_remove_from_context(event);
5411         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5412                 __perf_event_remove_from_context(event);
5413 }
5414 static void perf_event_exit_cpu(int cpu)
5415 {
5416         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5417         struct perf_event_context *ctx = &cpuctx->ctx;
5418
5419         mutex_lock(&ctx->mutex);
5420         smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5421         mutex_unlock(&ctx->mutex);
5422 }
5423 #else
5424 static inline void perf_event_exit_cpu(int cpu) { }
5425 #endif
5426
5427 static int __cpuinit
5428 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5429 {
5430         unsigned int cpu = (long)hcpu;
5431
5432         switch (action) {
5433
5434         case CPU_UP_PREPARE:
5435         case CPU_UP_PREPARE_FROZEN:
5436                 perf_event_init_cpu(cpu);
5437                 break;
5438
5439         case CPU_DOWN_PREPARE:
5440         case CPU_DOWN_PREPARE_FROZEN:
5441                 perf_event_exit_cpu(cpu);
5442                 break;
5443
5444         default:
5445                 break;
5446         }
5447
5448         return NOTIFY_OK;
5449 }
5450
5451 /*
5452  * This has to have a higher priority than migration_notifier in sched.c.
5453  */
5454 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5455         .notifier_call          = perf_cpu_notify,
5456         .priority               = 20,
5457 };
5458
5459 void __init perf_event_init(void)
5460 {
5461         perf_event_init_all_cpus();
5462         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5463                         (void *)(long)smp_processor_id());
5464         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5465                         (void *)(long)smp_processor_id());
5466         register_cpu_notifier(&perf_cpu_nb);
5467 }
5468
5469 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class,
5470                                         struct sysdev_class_attribute *attr,
5471                                         char *buf)
5472 {
5473         return sprintf(buf, "%d\n", perf_reserved_percpu);
5474 }
5475
5476 static ssize_t
5477 perf_set_reserve_percpu(struct sysdev_class *class,
5478                         struct sysdev_class_attribute *attr,
5479                         const char *buf,
5480                         size_t count)
5481 {
5482         struct perf_cpu_context *cpuctx;
5483         unsigned long val;
5484         int err, cpu, mpt;
5485
5486         err = strict_strtoul(buf, 10, &val);
5487         if (err)
5488                 return err;
5489         if (val > perf_max_events)
5490                 return -EINVAL;
5491
5492         spin_lock(&perf_resource_lock);
5493         perf_reserved_percpu = val;
5494         for_each_online_cpu(cpu) {
5495                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5496                 raw_spin_lock_irq(&cpuctx->ctx.lock);
5497                 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5498                           perf_max_events - perf_reserved_percpu);
5499                 cpuctx->max_pertask = mpt;
5500                 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5501         }
5502         spin_unlock(&perf_resource_lock);
5503
5504         return count;
5505 }
5506
5507 static ssize_t perf_show_overcommit(struct sysdev_class *class,
5508                                     struct sysdev_class_attribute *attr,
5509                                     char *buf)
5510 {
5511         return sprintf(buf, "%d\n", perf_overcommit);
5512 }
5513
5514 static ssize_t
5515 perf_set_overcommit(struct sysdev_class *class,
5516                     struct sysdev_class_attribute *attr,
5517                     const char *buf, size_t count)
5518 {
5519         unsigned long val;
5520         int err;
5521
5522         err = strict_strtoul(buf, 10, &val);
5523         if (err)
5524                 return err;
5525         if (val > 1)
5526                 return -EINVAL;
5527
5528         spin_lock(&perf_resource_lock);
5529         perf_overcommit = val;
5530         spin_unlock(&perf_resource_lock);
5531
5532         return count;
5533 }
5534
5535 static SYSDEV_CLASS_ATTR(
5536                                 reserve_percpu,
5537                                 0644,
5538                                 perf_show_reserve_percpu,
5539                                 perf_set_reserve_percpu
5540                         );
5541
5542 static SYSDEV_CLASS_ATTR(
5543                                 overcommit,
5544                                 0644,
5545                                 perf_show_overcommit,
5546                                 perf_set_overcommit
5547                         );
5548
5549 static struct attribute *perfclass_attrs[] = {
5550         &attr_reserve_percpu.attr,
5551         &attr_overcommit.attr,
5552         NULL
5553 };
5554
5555 static struct attribute_group perfclass_attr_group = {
5556         .attrs                  = perfclass_attrs,
5557         .name                   = "perf_events",
5558 };
5559
5560 static int __init perf_event_sysfs_init(void)
5561 {
5562         return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5563                                   &perfclass_attr_group);
5564 }
5565 device_initcall(perf_event_sysfs_init);