Merge branch 'timers-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git...
[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/hardirq.h>
24 #include <linux/rculist.h>
25 #include <linux/uaccess.h>
26 #include <linux/syscalls.h>
27 #include <linux/anon_inodes.h>
28 #include <linux/kernel_stat.h>
29 #include <linux/perf_event.h>
30
31 #include <asm/irq_regs.h>
32
33 /*
34  * Each CPU has a list of per CPU events:
35  */
36 DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
37
38 int perf_max_events __read_mostly = 1;
39 static int perf_reserved_percpu __read_mostly;
40 static int perf_overcommit __read_mostly = 1;
41
42 static atomic_t nr_events __read_mostly;
43 static atomic_t nr_mmap_events __read_mostly;
44 static atomic_t nr_comm_events __read_mostly;
45 static atomic_t nr_task_events __read_mostly;
46
47 /*
48  * perf event paranoia level:
49  *  -1 - not paranoid at all
50  *   0 - disallow raw tracepoint access for unpriv
51  *   1 - disallow cpu events for unpriv
52  *   2 - disallow kernel profiling for unpriv
53  */
54 int sysctl_perf_event_paranoid __read_mostly = 1;
55
56 static inline bool perf_paranoid_tracepoint_raw(void)
57 {
58         return sysctl_perf_event_paranoid > -1;
59 }
60
61 static inline bool perf_paranoid_cpu(void)
62 {
63         return sysctl_perf_event_paranoid > 0;
64 }
65
66 static inline bool perf_paranoid_kernel(void)
67 {
68         return sysctl_perf_event_paranoid > 1;
69 }
70
71 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
72
73 /*
74  * max perf event sample rate
75  */
76 int sysctl_perf_event_sample_rate __read_mostly = 100000;
77
78 static atomic64_t perf_event_id;
79
80 /*
81  * Lock for (sysadmin-configurable) event reservations:
82  */
83 static DEFINE_SPINLOCK(perf_resource_lock);
84
85 /*
86  * Architecture provided APIs - weak aliases:
87  */
88 extern __weak const struct pmu *hw_perf_event_init(struct perf_event *event)
89 {
90         return NULL;
91 }
92
93 void __weak hw_perf_disable(void)               { barrier(); }
94 void __weak hw_perf_enable(void)                { barrier(); }
95
96 void __weak hw_perf_event_setup(int cpu)        { barrier(); }
97 void __weak hw_perf_event_setup_online(int cpu) { barrier(); }
98
99 int __weak
100 hw_perf_group_sched_in(struct perf_event *group_leader,
101                struct perf_cpu_context *cpuctx,
102                struct perf_event_context *ctx, int cpu)
103 {
104         return 0;
105 }
106
107 void __weak perf_event_print_debug(void)        { }
108
109 static DEFINE_PER_CPU(int, perf_disable_count);
110
111 void __perf_disable(void)
112 {
113         __get_cpu_var(perf_disable_count)++;
114 }
115
116 bool __perf_enable(void)
117 {
118         return !--__get_cpu_var(perf_disable_count);
119 }
120
121 void perf_disable(void)
122 {
123         __perf_disable();
124         hw_perf_disable();
125 }
126
127 void perf_enable(void)
128 {
129         if (__perf_enable())
130                 hw_perf_enable();
131 }
132
133 static void get_ctx(struct perf_event_context *ctx)
134 {
135         WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
136 }
137
138 static void free_ctx(struct rcu_head *head)
139 {
140         struct perf_event_context *ctx;
141
142         ctx = container_of(head, struct perf_event_context, rcu_head);
143         kfree(ctx);
144 }
145
146 static void put_ctx(struct perf_event_context *ctx)
147 {
148         if (atomic_dec_and_test(&ctx->refcount)) {
149                 if (ctx->parent_ctx)
150                         put_ctx(ctx->parent_ctx);
151                 if (ctx->task)
152                         put_task_struct(ctx->task);
153                 call_rcu(&ctx->rcu_head, free_ctx);
154         }
155 }
156
157 static void unclone_ctx(struct perf_event_context *ctx)
158 {
159         if (ctx->parent_ctx) {
160                 put_ctx(ctx->parent_ctx);
161                 ctx->parent_ctx = NULL;
162         }
163 }
164
165 /*
166  * If we inherit events we want to return the parent event id
167  * to userspace.
168  */
169 static u64 primary_event_id(struct perf_event *event)
170 {
171         u64 id = event->id;
172
173         if (event->parent)
174                 id = event->parent->id;
175
176         return id;
177 }
178
179 /*
180  * Get the perf_event_context for a task and lock it.
181  * This has to cope with with the fact that until it is locked,
182  * the context could get moved to another task.
183  */
184 static struct perf_event_context *
185 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
186 {
187         struct perf_event_context *ctx;
188
189         rcu_read_lock();
190  retry:
191         ctx = rcu_dereference(task->perf_event_ctxp);
192         if (ctx) {
193                 /*
194                  * If this context is a clone of another, it might
195                  * get swapped for another underneath us by
196                  * perf_event_task_sched_out, though the
197                  * rcu_read_lock() protects us from any context
198                  * getting freed.  Lock the context and check if it
199                  * got swapped before we could get the lock, and retry
200                  * if so.  If we locked the right context, then it
201                  * can't get swapped on us any more.
202                  */
203                 spin_lock_irqsave(&ctx->lock, *flags);
204                 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
205                         spin_unlock_irqrestore(&ctx->lock, *flags);
206                         goto retry;
207                 }
208
209                 if (!atomic_inc_not_zero(&ctx->refcount)) {
210                         spin_unlock_irqrestore(&ctx->lock, *flags);
211                         ctx = NULL;
212                 }
213         }
214         rcu_read_unlock();
215         return ctx;
216 }
217
218 /*
219  * Get the context for a task and increment its pin_count so it
220  * can't get swapped to another task.  This also increments its
221  * reference count so that the context can't get freed.
222  */
223 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
224 {
225         struct perf_event_context *ctx;
226         unsigned long flags;
227
228         ctx = perf_lock_task_context(task, &flags);
229         if (ctx) {
230                 ++ctx->pin_count;
231                 spin_unlock_irqrestore(&ctx->lock, flags);
232         }
233         return ctx;
234 }
235
236 static void perf_unpin_context(struct perf_event_context *ctx)
237 {
238         unsigned long flags;
239
240         spin_lock_irqsave(&ctx->lock, flags);
241         --ctx->pin_count;
242         spin_unlock_irqrestore(&ctx->lock, flags);
243         put_ctx(ctx);
244 }
245
246 /*
247  * Add a event from the lists for its context.
248  * Must be called with ctx->mutex and ctx->lock held.
249  */
250 static void
251 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
252 {
253         struct perf_event *group_leader = event->group_leader;
254
255         /*
256          * Depending on whether it is a standalone or sibling event,
257          * add it straight to the context's event list, or to the group
258          * leader's sibling list:
259          */
260         if (group_leader == event)
261                 list_add_tail(&event->group_entry, &ctx->group_list);
262         else {
263                 list_add_tail(&event->group_entry, &group_leader->sibling_list);
264                 group_leader->nr_siblings++;
265         }
266
267         list_add_rcu(&event->event_entry, &ctx->event_list);
268         ctx->nr_events++;
269         if (event->attr.inherit_stat)
270                 ctx->nr_stat++;
271 }
272
273 /*
274  * Remove 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_del_event(struct perf_event *event, struct perf_event_context *ctx)
279 {
280         struct perf_event *sibling, *tmp;
281
282         if (list_empty(&event->group_entry))
283                 return;
284         ctx->nr_events--;
285         if (event->attr.inherit_stat)
286                 ctx->nr_stat--;
287
288         list_del_init(&event->group_entry);
289         list_del_rcu(&event->event_entry);
290
291         if (event->group_leader != event)
292                 event->group_leader->nr_siblings--;
293
294         /*
295          * If this was a group event with sibling events then
296          * upgrade the siblings to singleton events by adding them
297          * to the context list directly:
298          */
299         list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
300
301                 list_move_tail(&sibling->group_entry, &ctx->group_list);
302                 sibling->group_leader = sibling;
303         }
304 }
305
306 static void
307 event_sched_out(struct perf_event *event,
308                   struct perf_cpu_context *cpuctx,
309                   struct perf_event_context *ctx)
310 {
311         if (event->state != PERF_EVENT_STATE_ACTIVE)
312                 return;
313
314         event->state = PERF_EVENT_STATE_INACTIVE;
315         if (event->pending_disable) {
316                 event->pending_disable = 0;
317                 event->state = PERF_EVENT_STATE_OFF;
318         }
319         event->tstamp_stopped = ctx->time;
320         event->pmu->disable(event);
321         event->oncpu = -1;
322
323         if (!is_software_event(event))
324                 cpuctx->active_oncpu--;
325         ctx->nr_active--;
326         if (event->attr.exclusive || !cpuctx->active_oncpu)
327                 cpuctx->exclusive = 0;
328 }
329
330 static void
331 group_sched_out(struct perf_event *group_event,
332                 struct perf_cpu_context *cpuctx,
333                 struct perf_event_context *ctx)
334 {
335         struct perf_event *event;
336
337         if (group_event->state != PERF_EVENT_STATE_ACTIVE)
338                 return;
339
340         event_sched_out(group_event, cpuctx, ctx);
341
342         /*
343          * Schedule out siblings (if any):
344          */
345         list_for_each_entry(event, &group_event->sibling_list, group_entry)
346                 event_sched_out(event, cpuctx, ctx);
347
348         if (group_event->attr.exclusive)
349                 cpuctx->exclusive = 0;
350 }
351
352 /*
353  * Cross CPU call to remove a performance event
354  *
355  * We disable the event on the hardware level first. After that we
356  * remove it from the context list.
357  */
358 static void __perf_event_remove_from_context(void *info)
359 {
360         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
361         struct perf_event *event = info;
362         struct perf_event_context *ctx = event->ctx;
363
364         /*
365          * If this is a task context, we need to check whether it is
366          * the current task context of this cpu. If not it has been
367          * scheduled out before the smp call arrived.
368          */
369         if (ctx->task && cpuctx->task_ctx != ctx)
370                 return;
371
372         spin_lock(&ctx->lock);
373         /*
374          * Protect the list operation against NMI by disabling the
375          * events on a global level.
376          */
377         perf_disable();
378
379         event_sched_out(event, cpuctx, ctx);
380
381         list_del_event(event, ctx);
382
383         if (!ctx->task) {
384                 /*
385                  * Allow more per task events with respect to the
386                  * reservation:
387                  */
388                 cpuctx->max_pertask =
389                         min(perf_max_events - ctx->nr_events,
390                             perf_max_events - perf_reserved_percpu);
391         }
392
393         perf_enable();
394         spin_unlock(&ctx->lock);
395 }
396
397
398 /*
399  * Remove the event from a task's (or a CPU's) list of events.
400  *
401  * Must be called with ctx->mutex held.
402  *
403  * CPU events are removed with a smp call. For task events we only
404  * call when the task is on a CPU.
405  *
406  * If event->ctx is a cloned context, callers must make sure that
407  * every task struct that event->ctx->task could possibly point to
408  * remains valid.  This is OK when called from perf_release since
409  * that only calls us on the top-level context, which can't be a clone.
410  * When called from perf_event_exit_task, it's OK because the
411  * context has been detached from its task.
412  */
413 static void perf_event_remove_from_context(struct perf_event *event)
414 {
415         struct perf_event_context *ctx = event->ctx;
416         struct task_struct *task = ctx->task;
417
418         if (!task) {
419                 /*
420                  * Per cpu events are removed via an smp call and
421                  * the removal is always sucessful.
422                  */
423                 smp_call_function_single(event->cpu,
424                                          __perf_event_remove_from_context,
425                                          event, 1);
426                 return;
427         }
428
429 retry:
430         task_oncpu_function_call(task, __perf_event_remove_from_context,
431                                  event);
432
433         spin_lock_irq(&ctx->lock);
434         /*
435          * If the context is active we need to retry the smp call.
436          */
437         if (ctx->nr_active && !list_empty(&event->group_entry)) {
438                 spin_unlock_irq(&ctx->lock);
439                 goto retry;
440         }
441
442         /*
443          * The lock prevents that this context is scheduled in so we
444          * can remove the event safely, if the call above did not
445          * succeed.
446          */
447         if (!list_empty(&event->group_entry)) {
448                 list_del_event(event, ctx);
449         }
450         spin_unlock_irq(&ctx->lock);
451 }
452
453 static inline u64 perf_clock(void)
454 {
455         return cpu_clock(smp_processor_id());
456 }
457
458 /*
459  * Update the record of the current time in a context.
460  */
461 static void update_context_time(struct perf_event_context *ctx)
462 {
463         u64 now = perf_clock();
464
465         ctx->time += now - ctx->timestamp;
466         ctx->timestamp = now;
467 }
468
469 /*
470  * Update the total_time_enabled and total_time_running fields for a event.
471  */
472 static void update_event_times(struct perf_event *event)
473 {
474         struct perf_event_context *ctx = event->ctx;
475         u64 run_end;
476
477         if (event->state < PERF_EVENT_STATE_INACTIVE ||
478             event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
479                 return;
480
481         event->total_time_enabled = ctx->time - event->tstamp_enabled;
482
483         if (event->state == PERF_EVENT_STATE_INACTIVE)
484                 run_end = event->tstamp_stopped;
485         else
486                 run_end = ctx->time;
487
488         event->total_time_running = run_end - event->tstamp_running;
489 }
490
491 /*
492  * Update total_time_enabled and total_time_running for all events in a group.
493  */
494 static void update_group_times(struct perf_event *leader)
495 {
496         struct perf_event *event;
497
498         update_event_times(leader);
499         list_for_each_entry(event, &leader->sibling_list, group_entry)
500                 update_event_times(event);
501 }
502
503 /*
504  * Cross CPU call to disable a performance event
505  */
506 static void __perf_event_disable(void *info)
507 {
508         struct perf_event *event = info;
509         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
510         struct perf_event_context *ctx = event->ctx;
511
512         /*
513          * If this is a per-task event, need to check whether this
514          * event's task is the current task on this cpu.
515          */
516         if (ctx->task && cpuctx->task_ctx != ctx)
517                 return;
518
519         spin_lock(&ctx->lock);
520
521         /*
522          * If the event is on, turn it off.
523          * If it is in error state, leave it in error state.
524          */
525         if (event->state >= PERF_EVENT_STATE_INACTIVE) {
526                 update_context_time(ctx);
527                 update_group_times(event);
528                 if (event == event->group_leader)
529                         group_sched_out(event, cpuctx, ctx);
530                 else
531                         event_sched_out(event, cpuctx, ctx);
532                 event->state = PERF_EVENT_STATE_OFF;
533         }
534
535         spin_unlock(&ctx->lock);
536 }
537
538 /*
539  * Disable a event.
540  *
541  * If event->ctx is a cloned context, callers must make sure that
542  * every task struct that event->ctx->task could possibly point to
543  * remains valid.  This condition is satisifed when called through
544  * perf_event_for_each_child or perf_event_for_each because they
545  * hold the top-level event's child_mutex, so any descendant that
546  * goes to exit will block in sync_child_event.
547  * When called from perf_pending_event it's OK because event->ctx
548  * is the current context on this CPU and preemption is disabled,
549  * hence we can't get into perf_event_task_sched_out for this context.
550  */
551 static void perf_event_disable(struct perf_event *event)
552 {
553         struct perf_event_context *ctx = event->ctx;
554         struct task_struct *task = ctx->task;
555
556         if (!task) {
557                 /*
558                  * Disable the event on the cpu that it's on
559                  */
560                 smp_call_function_single(event->cpu, __perf_event_disable,
561                                          event, 1);
562                 return;
563         }
564
565  retry:
566         task_oncpu_function_call(task, __perf_event_disable, event);
567
568         spin_lock_irq(&ctx->lock);
569         /*
570          * If the event is still active, we need to retry the cross-call.
571          */
572         if (event->state == PERF_EVENT_STATE_ACTIVE) {
573                 spin_unlock_irq(&ctx->lock);
574                 goto retry;
575         }
576
577         /*
578          * Since we have the lock this context can't be scheduled
579          * in, so we can change the state safely.
580          */
581         if (event->state == PERF_EVENT_STATE_INACTIVE) {
582                 update_group_times(event);
583                 event->state = PERF_EVENT_STATE_OFF;
584         }
585
586         spin_unlock_irq(&ctx->lock);
587 }
588
589 static int
590 event_sched_in(struct perf_event *event,
591                  struct perf_cpu_context *cpuctx,
592                  struct perf_event_context *ctx,
593                  int cpu)
594 {
595         if (event->state <= PERF_EVENT_STATE_OFF)
596                 return 0;
597
598         event->state = PERF_EVENT_STATE_ACTIVE;
599         event->oncpu = cpu;     /* TODO: put 'cpu' into cpuctx->cpu */
600         /*
601          * The new state must be visible before we turn it on in the hardware:
602          */
603         smp_wmb();
604
605         if (event->pmu->enable(event)) {
606                 event->state = PERF_EVENT_STATE_INACTIVE;
607                 event->oncpu = -1;
608                 return -EAGAIN;
609         }
610
611         event->tstamp_running += ctx->time - event->tstamp_stopped;
612
613         if (!is_software_event(event))
614                 cpuctx->active_oncpu++;
615         ctx->nr_active++;
616
617         if (event->attr.exclusive)
618                 cpuctx->exclusive = 1;
619
620         return 0;
621 }
622
623 static int
624 group_sched_in(struct perf_event *group_event,
625                struct perf_cpu_context *cpuctx,
626                struct perf_event_context *ctx,
627                int cpu)
628 {
629         struct perf_event *event, *partial_group;
630         int ret;
631
632         if (group_event->state == PERF_EVENT_STATE_OFF)
633                 return 0;
634
635         ret = hw_perf_group_sched_in(group_event, cpuctx, ctx, cpu);
636         if (ret)
637                 return ret < 0 ? ret : 0;
638
639         if (event_sched_in(group_event, cpuctx, ctx, cpu))
640                 return -EAGAIN;
641
642         /*
643          * Schedule in siblings as one group (if any):
644          */
645         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
646                 if (event_sched_in(event, cpuctx, ctx, cpu)) {
647                         partial_group = event;
648                         goto group_error;
649                 }
650         }
651
652         return 0;
653
654 group_error:
655         /*
656          * Groups can be scheduled in as one unit only, so undo any
657          * partial group before returning:
658          */
659         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
660                 if (event == partial_group)
661                         break;
662                 event_sched_out(event, cpuctx, ctx);
663         }
664         event_sched_out(group_event, cpuctx, ctx);
665
666         return -EAGAIN;
667 }
668
669 /*
670  * Return 1 for a group consisting entirely of software events,
671  * 0 if the group contains any hardware events.
672  */
673 static int is_software_only_group(struct perf_event *leader)
674 {
675         struct perf_event *event;
676
677         if (!is_software_event(leader))
678                 return 0;
679
680         list_for_each_entry(event, &leader->sibling_list, group_entry)
681                 if (!is_software_event(event))
682                         return 0;
683
684         return 1;
685 }
686
687 /*
688  * Work out whether we can put this event group on the CPU now.
689  */
690 static int group_can_go_on(struct perf_event *event,
691                            struct perf_cpu_context *cpuctx,
692                            int can_add_hw)
693 {
694         /*
695          * Groups consisting entirely of software events can always go on.
696          */
697         if (is_software_only_group(event))
698                 return 1;
699         /*
700          * If an exclusive group is already on, no other hardware
701          * events can go on.
702          */
703         if (cpuctx->exclusive)
704                 return 0;
705         /*
706          * If this group is exclusive and there are already
707          * events on the CPU, it can't go on.
708          */
709         if (event->attr.exclusive && cpuctx->active_oncpu)
710                 return 0;
711         /*
712          * Otherwise, try to add it if all previous groups were able
713          * to go on.
714          */
715         return can_add_hw;
716 }
717
718 static void add_event_to_ctx(struct perf_event *event,
719                                struct perf_event_context *ctx)
720 {
721         list_add_event(event, ctx);
722         event->tstamp_enabled = ctx->time;
723         event->tstamp_running = ctx->time;
724         event->tstamp_stopped = ctx->time;
725 }
726
727 /*
728  * Cross CPU call to install and enable a performance event
729  *
730  * Must be called with ctx->mutex held
731  */
732 static void __perf_install_in_context(void *info)
733 {
734         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
735         struct perf_event *event = info;
736         struct perf_event_context *ctx = event->ctx;
737         struct perf_event *leader = event->group_leader;
738         int cpu = smp_processor_id();
739         int err;
740
741         /*
742          * If this is a task context, we need to check whether it is
743          * the current task context of this cpu. If not it has been
744          * scheduled out before the smp call arrived.
745          * Or possibly this is the right context but it isn't
746          * on this cpu because it had no events.
747          */
748         if (ctx->task && cpuctx->task_ctx != ctx) {
749                 if (cpuctx->task_ctx || ctx->task != current)
750                         return;
751                 cpuctx->task_ctx = ctx;
752         }
753
754         spin_lock(&ctx->lock);
755         ctx->is_active = 1;
756         update_context_time(ctx);
757
758         /*
759          * Protect the list operation against NMI by disabling the
760          * events on a global level. NOP for non NMI based events.
761          */
762         perf_disable();
763
764         add_event_to_ctx(event, ctx);
765
766         /*
767          * Don't put the event on if it is disabled or if
768          * it is in a group and the group isn't on.
769          */
770         if (event->state != PERF_EVENT_STATE_INACTIVE ||
771             (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
772                 goto unlock;
773
774         /*
775          * An exclusive event can't go on if there are already active
776          * hardware events, and no hardware event can go on if there
777          * is already an exclusive event on.
778          */
779         if (!group_can_go_on(event, cpuctx, 1))
780                 err = -EEXIST;
781         else
782                 err = event_sched_in(event, cpuctx, ctx, cpu);
783
784         if (err) {
785                 /*
786                  * This event couldn't go on.  If it is in a group
787                  * then we have to pull the whole group off.
788                  * If the event group is pinned then put it in error state.
789                  */
790                 if (leader != event)
791                         group_sched_out(leader, cpuctx, ctx);
792                 if (leader->attr.pinned) {
793                         update_group_times(leader);
794                         leader->state = PERF_EVENT_STATE_ERROR;
795                 }
796         }
797
798         if (!err && !ctx->task && cpuctx->max_pertask)
799                 cpuctx->max_pertask--;
800
801  unlock:
802         perf_enable();
803
804         spin_unlock(&ctx->lock);
805 }
806
807 /*
808  * Attach a performance event to a context
809  *
810  * First we add the event to the list with the hardware enable bit
811  * in event->hw_config cleared.
812  *
813  * If the event is attached to a task which is on a CPU we use a smp
814  * call to enable it in the task context. The task might have been
815  * scheduled away, but we check this in the smp call again.
816  *
817  * Must be called with ctx->mutex held.
818  */
819 static void
820 perf_install_in_context(struct perf_event_context *ctx,
821                         struct perf_event *event,
822                         int cpu)
823 {
824         struct task_struct *task = ctx->task;
825
826         if (!task) {
827                 /*
828                  * Per cpu events are installed via an smp call and
829                  * the install is always sucessful.
830                  */
831                 smp_call_function_single(cpu, __perf_install_in_context,
832                                          event, 1);
833                 return;
834         }
835
836 retry:
837         task_oncpu_function_call(task, __perf_install_in_context,
838                                  event);
839
840         spin_lock_irq(&ctx->lock);
841         /*
842          * we need to retry the smp call.
843          */
844         if (ctx->is_active && list_empty(&event->group_entry)) {
845                 spin_unlock_irq(&ctx->lock);
846                 goto retry;
847         }
848
849         /*
850          * The lock prevents that this context is scheduled in so we
851          * can add the event safely, if it the call above did not
852          * succeed.
853          */
854         if (list_empty(&event->group_entry))
855                 add_event_to_ctx(event, ctx);
856         spin_unlock_irq(&ctx->lock);
857 }
858
859 /*
860  * Put a event into inactive state and update time fields.
861  * Enabling the leader of a group effectively enables all
862  * the group members that aren't explicitly disabled, so we
863  * have to update their ->tstamp_enabled also.
864  * Note: this works for group members as well as group leaders
865  * since the non-leader members' sibling_lists will be empty.
866  */
867 static void __perf_event_mark_enabled(struct perf_event *event,
868                                         struct perf_event_context *ctx)
869 {
870         struct perf_event *sub;
871
872         event->state = PERF_EVENT_STATE_INACTIVE;
873         event->tstamp_enabled = ctx->time - event->total_time_enabled;
874         list_for_each_entry(sub, &event->sibling_list, group_entry)
875                 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
876                         sub->tstamp_enabled =
877                                 ctx->time - sub->total_time_enabled;
878 }
879
880 /*
881  * Cross CPU call to enable a performance event
882  */
883 static void __perf_event_enable(void *info)
884 {
885         struct perf_event *event = info;
886         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
887         struct perf_event_context *ctx = event->ctx;
888         struct perf_event *leader = event->group_leader;
889         int err;
890
891         /*
892          * If this is a per-task event, need to check whether this
893          * event's task is the current task on this cpu.
894          */
895         if (ctx->task && cpuctx->task_ctx != ctx) {
896                 if (cpuctx->task_ctx || ctx->task != current)
897                         return;
898                 cpuctx->task_ctx = ctx;
899         }
900
901         spin_lock(&ctx->lock);
902         ctx->is_active = 1;
903         update_context_time(ctx);
904
905         if (event->state >= PERF_EVENT_STATE_INACTIVE)
906                 goto unlock;
907         __perf_event_mark_enabled(event, ctx);
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                                              smp_processor_id());
923                 else
924                         err = event_sched_in(event, cpuctx, ctx,
925                                                smp_processor_id());
926                 perf_enable();
927         }
928
929         if (err) {
930                 /*
931                  * If this event can't go on and it's part of a
932                  * group, then the whole group has to come off.
933                  */
934                 if (leader != event)
935                         group_sched_out(leader, cpuctx, ctx);
936                 if (leader->attr.pinned) {
937                         update_group_times(leader);
938                         leader->state = PERF_EVENT_STATE_ERROR;
939                 }
940         }
941
942  unlock:
943         spin_unlock(&ctx->lock);
944 }
945
946 /*
947  * Enable a event.
948  *
949  * If event->ctx is a cloned context, callers must make sure that
950  * every task struct that event->ctx->task could possibly point to
951  * remains valid.  This condition is satisfied when called through
952  * perf_event_for_each_child or perf_event_for_each as described
953  * for perf_event_disable.
954  */
955 static void perf_event_enable(struct perf_event *event)
956 {
957         struct perf_event_context *ctx = event->ctx;
958         struct task_struct *task = ctx->task;
959
960         if (!task) {
961                 /*
962                  * Enable the event on the cpu that it's on
963                  */
964                 smp_call_function_single(event->cpu, __perf_event_enable,
965                                          event, 1);
966                 return;
967         }
968
969         spin_lock_irq(&ctx->lock);
970         if (event->state >= PERF_EVENT_STATE_INACTIVE)
971                 goto out;
972
973         /*
974          * If the event is in error state, clear that first.
975          * That way, if we see the event in error state below, we
976          * know that it has gone back into error state, as distinct
977          * from the task having been scheduled away before the
978          * cross-call arrived.
979          */
980         if (event->state == PERF_EVENT_STATE_ERROR)
981                 event->state = PERF_EVENT_STATE_OFF;
982
983  retry:
984         spin_unlock_irq(&ctx->lock);
985         task_oncpu_function_call(task, __perf_event_enable, event);
986
987         spin_lock_irq(&ctx->lock);
988
989         /*
990          * If the context is active and the event is still off,
991          * we need to retry the cross-call.
992          */
993         if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
994                 goto retry;
995
996         /*
997          * Since we have the lock this context can't be scheduled
998          * in, so we can change the state safely.
999          */
1000         if (event->state == PERF_EVENT_STATE_OFF)
1001                 __perf_event_mark_enabled(event, ctx);
1002
1003  out:
1004         spin_unlock_irq(&ctx->lock);
1005 }
1006
1007 static int perf_event_refresh(struct perf_event *event, int refresh)
1008 {
1009         /*
1010          * not supported on inherited events
1011          */
1012         if (event->attr.inherit)
1013                 return -EINVAL;
1014
1015         atomic_add(refresh, &event->event_limit);
1016         perf_event_enable(event);
1017
1018         return 0;
1019 }
1020
1021 void __perf_event_sched_out(struct perf_event_context *ctx,
1022                               struct perf_cpu_context *cpuctx)
1023 {
1024         struct perf_event *event;
1025
1026         spin_lock(&ctx->lock);
1027         ctx->is_active = 0;
1028         if (likely(!ctx->nr_events))
1029                 goto out;
1030         update_context_time(ctx);
1031
1032         perf_disable();
1033         if (ctx->nr_active) {
1034                 list_for_each_entry(event, &ctx->group_list, group_entry) {
1035                         if (event != event->group_leader)
1036                                 event_sched_out(event, cpuctx, ctx);
1037                         else
1038                                 group_sched_out(event, cpuctx, ctx);
1039                 }
1040         }
1041         perf_enable();
1042  out:
1043         spin_unlock(&ctx->lock);
1044 }
1045
1046 /*
1047  * Test whether two contexts are equivalent, i.e. whether they
1048  * have both been cloned from the same version of the same context
1049  * and they both have the same number of enabled events.
1050  * If the number of enabled events is the same, then the set
1051  * of enabled events should be the same, because these are both
1052  * inherited contexts, therefore we can't access individual events
1053  * in them directly with an fd; we can only enable/disable all
1054  * events via prctl, or enable/disable all events in a family
1055  * via ioctl, which will have the same effect on both contexts.
1056  */
1057 static int context_equiv(struct perf_event_context *ctx1,
1058                          struct perf_event_context *ctx2)
1059 {
1060         return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1061                 && ctx1->parent_gen == ctx2->parent_gen
1062                 && !ctx1->pin_count && !ctx2->pin_count;
1063 }
1064
1065 static void __perf_event_read(void *event);
1066
1067 static void __perf_event_sync_stat(struct perf_event *event,
1068                                      struct perf_event *next_event)
1069 {
1070         u64 value;
1071
1072         if (!event->attr.inherit_stat)
1073                 return;
1074
1075         /*
1076          * Update the event value, we cannot use perf_event_read()
1077          * because we're in the middle of a context switch and have IRQs
1078          * disabled, which upsets smp_call_function_single(), however
1079          * we know the event must be on the current CPU, therefore we
1080          * don't need to use it.
1081          */
1082         switch (event->state) {
1083         case PERF_EVENT_STATE_ACTIVE:
1084                 __perf_event_read(event);
1085                 break;
1086
1087         case PERF_EVENT_STATE_INACTIVE:
1088                 update_event_times(event);
1089                 break;
1090
1091         default:
1092                 break;
1093         }
1094
1095         /*
1096          * In order to keep per-task stats reliable we need to flip the event
1097          * values when we flip the contexts.
1098          */
1099         value = atomic64_read(&next_event->count);
1100         value = atomic64_xchg(&event->count, value);
1101         atomic64_set(&next_event->count, value);
1102
1103         swap(event->total_time_enabled, next_event->total_time_enabled);
1104         swap(event->total_time_running, next_event->total_time_running);
1105
1106         /*
1107          * Since we swizzled the values, update the user visible data too.
1108          */
1109         perf_event_update_userpage(event);
1110         perf_event_update_userpage(next_event);
1111 }
1112
1113 #define list_next_entry(pos, member) \
1114         list_entry(pos->member.next, typeof(*pos), member)
1115
1116 static void perf_event_sync_stat(struct perf_event_context *ctx,
1117                                    struct perf_event_context *next_ctx)
1118 {
1119         struct perf_event *event, *next_event;
1120
1121         if (!ctx->nr_stat)
1122                 return;
1123
1124         event = list_first_entry(&ctx->event_list,
1125                                    struct perf_event, event_entry);
1126
1127         next_event = list_first_entry(&next_ctx->event_list,
1128                                         struct perf_event, event_entry);
1129
1130         while (&event->event_entry != &ctx->event_list &&
1131                &next_event->event_entry != &next_ctx->event_list) {
1132
1133                 __perf_event_sync_stat(event, next_event);
1134
1135                 event = list_next_entry(event, event_entry);
1136                 next_event = list_next_entry(next_event, event_entry);
1137         }
1138 }
1139
1140 /*
1141  * Called from scheduler to remove the events of the current task,
1142  * with interrupts disabled.
1143  *
1144  * We stop each event and update the event value in event->count.
1145  *
1146  * This does not protect us against NMI, but disable()
1147  * sets the disabled bit in the control field of event _before_
1148  * accessing the event control register. If a NMI hits, then it will
1149  * not restart the event.
1150  */
1151 void perf_event_task_sched_out(struct task_struct *task,
1152                                  struct task_struct *next, int cpu)
1153 {
1154         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1155         struct perf_event_context *ctx = task->perf_event_ctxp;
1156         struct perf_event_context *next_ctx;
1157         struct perf_event_context *parent;
1158         struct pt_regs *regs;
1159         int do_switch = 1;
1160
1161         regs = task_pt_regs(task);
1162         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1163
1164         if (likely(!ctx || !cpuctx->task_ctx))
1165                 return;
1166
1167         update_context_time(ctx);
1168
1169         rcu_read_lock();
1170         parent = rcu_dereference(ctx->parent_ctx);
1171         next_ctx = next->perf_event_ctxp;
1172         if (parent && next_ctx &&
1173             rcu_dereference(next_ctx->parent_ctx) == parent) {
1174                 /*
1175                  * Looks like the two contexts are clones, so we might be
1176                  * able to optimize the context switch.  We lock both
1177                  * contexts and check that they are clones under the
1178                  * lock (including re-checking that neither has been
1179                  * uncloned in the meantime).  It doesn't matter which
1180                  * order we take the locks because no other cpu could
1181                  * be trying to lock both of these tasks.
1182                  */
1183                 spin_lock(&ctx->lock);
1184                 spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1185                 if (context_equiv(ctx, next_ctx)) {
1186                         /*
1187                          * XXX do we need a memory barrier of sorts
1188                          * wrt to rcu_dereference() of perf_event_ctxp
1189                          */
1190                         task->perf_event_ctxp = next_ctx;
1191                         next->perf_event_ctxp = ctx;
1192                         ctx->task = next;
1193                         next_ctx->task = task;
1194                         do_switch = 0;
1195
1196                         perf_event_sync_stat(ctx, next_ctx);
1197                 }
1198                 spin_unlock(&next_ctx->lock);
1199                 spin_unlock(&ctx->lock);
1200         }
1201         rcu_read_unlock();
1202
1203         if (do_switch) {
1204                 __perf_event_sched_out(ctx, cpuctx);
1205                 cpuctx->task_ctx = NULL;
1206         }
1207 }
1208
1209 /*
1210  * Called with IRQs disabled
1211  */
1212 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1213 {
1214         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1215
1216         if (!cpuctx->task_ctx)
1217                 return;
1218
1219         if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1220                 return;
1221
1222         __perf_event_sched_out(ctx, cpuctx);
1223         cpuctx->task_ctx = NULL;
1224 }
1225
1226 /*
1227  * Called with IRQs disabled
1228  */
1229 static void perf_event_cpu_sched_out(struct perf_cpu_context *cpuctx)
1230 {
1231         __perf_event_sched_out(&cpuctx->ctx, cpuctx);
1232 }
1233
1234 static void
1235 __perf_event_sched_in(struct perf_event_context *ctx,
1236                         struct perf_cpu_context *cpuctx, int cpu)
1237 {
1238         struct perf_event *event;
1239         int can_add_hw = 1;
1240
1241         spin_lock(&ctx->lock);
1242         ctx->is_active = 1;
1243         if (likely(!ctx->nr_events))
1244                 goto out;
1245
1246         ctx->timestamp = perf_clock();
1247
1248         perf_disable();
1249
1250         /*
1251          * First go through the list and put on any pinned groups
1252          * in order to give them the best chance of going on.
1253          */
1254         list_for_each_entry(event, &ctx->group_list, group_entry) {
1255                 if (event->state <= PERF_EVENT_STATE_OFF ||
1256                     !event->attr.pinned)
1257                         continue;
1258                 if (event->cpu != -1 && event->cpu != cpu)
1259                         continue;
1260
1261                 if (event != event->group_leader)
1262                         event_sched_in(event, cpuctx, ctx, cpu);
1263                 else {
1264                         if (group_can_go_on(event, cpuctx, 1))
1265                                 group_sched_in(event, cpuctx, ctx, cpu);
1266                 }
1267
1268                 /*
1269                  * If this pinned group hasn't been scheduled,
1270                  * put it in error state.
1271                  */
1272                 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1273                         update_group_times(event);
1274                         event->state = PERF_EVENT_STATE_ERROR;
1275                 }
1276         }
1277
1278         list_for_each_entry(event, &ctx->group_list, group_entry) {
1279                 /*
1280                  * Ignore events in OFF or ERROR state, and
1281                  * ignore pinned events since we did them already.
1282                  */
1283                 if (event->state <= PERF_EVENT_STATE_OFF ||
1284                     event->attr.pinned)
1285                         continue;
1286
1287                 /*
1288                  * Listen to the 'cpu' scheduling filter constraint
1289                  * of events:
1290                  */
1291                 if (event->cpu != -1 && event->cpu != cpu)
1292                         continue;
1293
1294                 if (event != event->group_leader) {
1295                         if (event_sched_in(event, cpuctx, ctx, cpu))
1296                                 can_add_hw = 0;
1297                 } else {
1298                         if (group_can_go_on(event, cpuctx, can_add_hw)) {
1299                                 if (group_sched_in(event, cpuctx, ctx, cpu))
1300                                         can_add_hw = 0;
1301                         }
1302                 }
1303         }
1304         perf_enable();
1305  out:
1306         spin_unlock(&ctx->lock);
1307 }
1308
1309 /*
1310  * Called from scheduler to add the events of the current task
1311  * with interrupts disabled.
1312  *
1313  * We restore the event value and then enable it.
1314  *
1315  * This does not protect us against NMI, but enable()
1316  * sets the enabled bit in the control field of event _before_
1317  * accessing the event control register. If a NMI hits, then it will
1318  * keep the event running.
1319  */
1320 void perf_event_task_sched_in(struct task_struct *task, int cpu)
1321 {
1322         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1323         struct perf_event_context *ctx = task->perf_event_ctxp;
1324
1325         if (likely(!ctx))
1326                 return;
1327         if (cpuctx->task_ctx == ctx)
1328                 return;
1329         __perf_event_sched_in(ctx, cpuctx, cpu);
1330         cpuctx->task_ctx = ctx;
1331 }
1332
1333 static void perf_event_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
1334 {
1335         struct perf_event_context *ctx = &cpuctx->ctx;
1336
1337         __perf_event_sched_in(ctx, cpuctx, cpu);
1338 }
1339
1340 #define MAX_INTERRUPTS (~0ULL)
1341
1342 static void perf_log_throttle(struct perf_event *event, int enable);
1343
1344 static void perf_adjust_period(struct perf_event *event, u64 events)
1345 {
1346         struct hw_perf_event *hwc = &event->hw;
1347         u64 period, sample_period;
1348         s64 delta;
1349
1350         events *= hwc->sample_period;
1351         period = div64_u64(events, event->attr.sample_freq);
1352
1353         delta = (s64)(period - hwc->sample_period);
1354         delta = (delta + 7) / 8; /* low pass filter */
1355
1356         sample_period = hwc->sample_period + delta;
1357
1358         if (!sample_period)
1359                 sample_period = 1;
1360
1361         hwc->sample_period = sample_period;
1362 }
1363
1364 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1365 {
1366         struct perf_event *event;
1367         struct hw_perf_event *hwc;
1368         u64 interrupts, freq;
1369
1370         spin_lock(&ctx->lock);
1371         list_for_each_entry(event, &ctx->group_list, group_entry) {
1372                 if (event->state != PERF_EVENT_STATE_ACTIVE)
1373                         continue;
1374
1375                 hwc = &event->hw;
1376
1377                 interrupts = hwc->interrupts;
1378                 hwc->interrupts = 0;
1379
1380                 /*
1381                  * unthrottle events on the tick
1382                  */
1383                 if (interrupts == MAX_INTERRUPTS) {
1384                         perf_log_throttle(event, 1);
1385                         event->pmu->unthrottle(event);
1386                         interrupts = 2*sysctl_perf_event_sample_rate/HZ;
1387                 }
1388
1389                 if (!event->attr.freq || !event->attr.sample_freq)
1390                         continue;
1391
1392                 /*
1393                  * if the specified freq < HZ then we need to skip ticks
1394                  */
1395                 if (event->attr.sample_freq < HZ) {
1396                         freq = event->attr.sample_freq;
1397
1398                         hwc->freq_count += freq;
1399                         hwc->freq_interrupts += interrupts;
1400
1401                         if (hwc->freq_count < HZ)
1402                                 continue;
1403
1404                         interrupts = hwc->freq_interrupts;
1405                         hwc->freq_interrupts = 0;
1406                         hwc->freq_count -= HZ;
1407                 } else
1408                         freq = HZ;
1409
1410                 perf_adjust_period(event, freq * interrupts);
1411
1412                 /*
1413                  * In order to avoid being stalled by an (accidental) huge
1414                  * sample period, force reset the sample period if we didn't
1415                  * get any events in this freq period.
1416                  */
1417                 if (!interrupts) {
1418                         perf_disable();
1419                         event->pmu->disable(event);
1420                         atomic64_set(&hwc->period_left, 0);
1421                         event->pmu->enable(event);
1422                         perf_enable();
1423                 }
1424         }
1425         spin_unlock(&ctx->lock);
1426 }
1427
1428 /*
1429  * Round-robin a context's events:
1430  */
1431 static void rotate_ctx(struct perf_event_context *ctx)
1432 {
1433         struct perf_event *event;
1434
1435         if (!ctx->nr_events)
1436                 return;
1437
1438         spin_lock(&ctx->lock);
1439         /*
1440          * Rotate the first entry last (works just fine for group events too):
1441          */
1442         perf_disable();
1443         list_for_each_entry(event, &ctx->group_list, group_entry) {
1444                 list_move_tail(&event->group_entry, &ctx->group_list);
1445                 break;
1446         }
1447         perf_enable();
1448
1449         spin_unlock(&ctx->lock);
1450 }
1451
1452 void perf_event_task_tick(struct task_struct *curr, int cpu)
1453 {
1454         struct perf_cpu_context *cpuctx;
1455         struct perf_event_context *ctx;
1456
1457         if (!atomic_read(&nr_events))
1458                 return;
1459
1460         cpuctx = &per_cpu(perf_cpu_context, cpu);
1461         ctx = curr->perf_event_ctxp;
1462
1463         perf_ctx_adjust_freq(&cpuctx->ctx);
1464         if (ctx)
1465                 perf_ctx_adjust_freq(ctx);
1466
1467         perf_event_cpu_sched_out(cpuctx);
1468         if (ctx)
1469                 __perf_event_task_sched_out(ctx);
1470
1471         rotate_ctx(&cpuctx->ctx);
1472         if (ctx)
1473                 rotate_ctx(ctx);
1474
1475         perf_event_cpu_sched_in(cpuctx, cpu);
1476         if (ctx)
1477                 perf_event_task_sched_in(curr, cpu);
1478 }
1479
1480 /*
1481  * Enable all of a task's events that have been marked enable-on-exec.
1482  * This expects task == current.
1483  */
1484 static void perf_event_enable_on_exec(struct task_struct *task)
1485 {
1486         struct perf_event_context *ctx;
1487         struct perf_event *event;
1488         unsigned long flags;
1489         int enabled = 0;
1490
1491         local_irq_save(flags);
1492         ctx = task->perf_event_ctxp;
1493         if (!ctx || !ctx->nr_events)
1494                 goto out;
1495
1496         __perf_event_task_sched_out(ctx);
1497
1498         spin_lock(&ctx->lock);
1499
1500         list_for_each_entry(event, &ctx->group_list, group_entry) {
1501                 if (!event->attr.enable_on_exec)
1502                         continue;
1503                 event->attr.enable_on_exec = 0;
1504                 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1505                         continue;
1506                 __perf_event_mark_enabled(event, ctx);
1507                 enabled = 1;
1508         }
1509
1510         /*
1511          * Unclone this context if we enabled any event.
1512          */
1513         if (enabled)
1514                 unclone_ctx(ctx);
1515
1516         spin_unlock(&ctx->lock);
1517
1518         perf_event_task_sched_in(task, smp_processor_id());
1519  out:
1520         local_irq_restore(flags);
1521 }
1522
1523 /*
1524  * Cross CPU call to read the hardware event
1525  */
1526 static void __perf_event_read(void *info)
1527 {
1528         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1529         struct perf_event *event = info;
1530         struct perf_event_context *ctx = event->ctx;
1531         unsigned long flags;
1532
1533         /*
1534          * If this is a task context, we need to check whether it is
1535          * the current task context of this cpu.  If not it has been
1536          * scheduled out before the smp call arrived.  In that case
1537          * event->count would have been updated to a recent sample
1538          * when the event was scheduled out.
1539          */
1540         if (ctx->task && cpuctx->task_ctx != ctx)
1541                 return;
1542
1543         local_irq_save(flags);
1544         if (ctx->is_active)
1545                 update_context_time(ctx);
1546         event->pmu->read(event);
1547         update_event_times(event);
1548         local_irq_restore(flags);
1549 }
1550
1551 static u64 perf_event_read(struct perf_event *event)
1552 {
1553         /*
1554          * If event is enabled and currently active on a CPU, update the
1555          * value in the event structure:
1556          */
1557         if (event->state == PERF_EVENT_STATE_ACTIVE) {
1558                 smp_call_function_single(event->oncpu,
1559                                          __perf_event_read, event, 1);
1560         } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1561                 update_event_times(event);
1562         }
1563
1564         return atomic64_read(&event->count);
1565 }
1566
1567 /*
1568  * Initialize the perf_event context in a task_struct:
1569  */
1570 static void
1571 __perf_event_init_context(struct perf_event_context *ctx,
1572                             struct task_struct *task)
1573 {
1574         memset(ctx, 0, sizeof(*ctx));
1575         spin_lock_init(&ctx->lock);
1576         mutex_init(&ctx->mutex);
1577         INIT_LIST_HEAD(&ctx->group_list);
1578         INIT_LIST_HEAD(&ctx->event_list);
1579         atomic_set(&ctx->refcount, 1);
1580         ctx->task = task;
1581 }
1582
1583 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1584 {
1585         struct perf_event_context *ctx;
1586         struct perf_cpu_context *cpuctx;
1587         struct task_struct *task;
1588         unsigned long flags;
1589         int err;
1590
1591         /*
1592          * If cpu is not a wildcard then this is a percpu event:
1593          */
1594         if (cpu != -1) {
1595                 /* Must be root to operate on a CPU event: */
1596                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1597                         return ERR_PTR(-EACCES);
1598
1599                 if (cpu < 0 || cpu > num_possible_cpus())
1600                         return ERR_PTR(-EINVAL);
1601
1602                 /*
1603                  * We could be clever and allow to attach a event to an
1604                  * offline CPU and activate it when the CPU comes up, but
1605                  * that's for later.
1606                  */
1607                 if (!cpu_isset(cpu, cpu_online_map))
1608                         return ERR_PTR(-ENODEV);
1609
1610                 cpuctx = &per_cpu(perf_cpu_context, cpu);
1611                 ctx = &cpuctx->ctx;
1612                 get_ctx(ctx);
1613
1614                 return ctx;
1615         }
1616
1617         rcu_read_lock();
1618         if (!pid)
1619                 task = current;
1620         else
1621                 task = find_task_by_vpid(pid);
1622         if (task)
1623                 get_task_struct(task);
1624         rcu_read_unlock();
1625
1626         if (!task)
1627                 return ERR_PTR(-ESRCH);
1628
1629         /*
1630          * Can't attach events to a dying task.
1631          */
1632         err = -ESRCH;
1633         if (task->flags & PF_EXITING)
1634                 goto errout;
1635
1636         /* Reuse ptrace permission checks for now. */
1637         err = -EACCES;
1638         if (!ptrace_may_access(task, PTRACE_MODE_READ))
1639                 goto errout;
1640
1641  retry:
1642         ctx = perf_lock_task_context(task, &flags);
1643         if (ctx) {
1644                 unclone_ctx(ctx);
1645                 spin_unlock_irqrestore(&ctx->lock, flags);
1646         }
1647
1648         if (!ctx) {
1649                 ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1650                 err = -ENOMEM;
1651                 if (!ctx)
1652                         goto errout;
1653                 __perf_event_init_context(ctx, task);
1654                 get_ctx(ctx);
1655                 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1656                         /*
1657                          * We raced with some other task; use
1658                          * the context they set.
1659                          */
1660                         kfree(ctx);
1661                         goto retry;
1662                 }
1663                 get_task_struct(task);
1664         }
1665
1666         put_task_struct(task);
1667         return ctx;
1668
1669  errout:
1670         put_task_struct(task);
1671         return ERR_PTR(err);
1672 }
1673
1674 static void free_event_rcu(struct rcu_head *head)
1675 {
1676         struct perf_event *event;
1677
1678         event = container_of(head, struct perf_event, rcu_head);
1679         if (event->ns)
1680                 put_pid_ns(event->ns);
1681         kfree(event);
1682 }
1683
1684 static void perf_pending_sync(struct perf_event *event);
1685
1686 static void free_event(struct perf_event *event)
1687 {
1688         perf_pending_sync(event);
1689
1690         if (!event->parent) {
1691                 atomic_dec(&nr_events);
1692                 if (event->attr.mmap)
1693                         atomic_dec(&nr_mmap_events);
1694                 if (event->attr.comm)
1695                         atomic_dec(&nr_comm_events);
1696                 if (event->attr.task)
1697                         atomic_dec(&nr_task_events);
1698         }
1699
1700         if (event->output) {
1701                 fput(event->output->filp);
1702                 event->output = NULL;
1703         }
1704
1705         if (event->destroy)
1706                 event->destroy(event);
1707
1708         put_ctx(event->ctx);
1709         call_rcu(&event->rcu_head, free_event_rcu);
1710 }
1711
1712 /*
1713  * Called when the last reference to the file is gone.
1714  */
1715 static int perf_release(struct inode *inode, struct file *file)
1716 {
1717         struct perf_event *event = file->private_data;
1718         struct perf_event_context *ctx = event->ctx;
1719
1720         file->private_data = NULL;
1721
1722         WARN_ON_ONCE(ctx->parent_ctx);
1723         mutex_lock(&ctx->mutex);
1724         perf_event_remove_from_context(event);
1725         mutex_unlock(&ctx->mutex);
1726
1727         mutex_lock(&event->owner->perf_event_mutex);
1728         list_del_init(&event->owner_entry);
1729         mutex_unlock(&event->owner->perf_event_mutex);
1730         put_task_struct(event->owner);
1731
1732         free_event(event);
1733
1734         return 0;
1735 }
1736
1737 static int perf_event_read_size(struct perf_event *event)
1738 {
1739         int entry = sizeof(u64); /* value */
1740         int size = 0;
1741         int nr = 1;
1742
1743         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1744                 size += sizeof(u64);
1745
1746         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1747                 size += sizeof(u64);
1748
1749         if (event->attr.read_format & PERF_FORMAT_ID)
1750                 entry += sizeof(u64);
1751
1752         if (event->attr.read_format & PERF_FORMAT_GROUP) {
1753                 nr += event->group_leader->nr_siblings;
1754                 size += sizeof(u64);
1755         }
1756
1757         size += entry * nr;
1758
1759         return size;
1760 }
1761
1762 static u64 perf_event_read_value(struct perf_event *event)
1763 {
1764         struct perf_event *child;
1765         u64 total = 0;
1766
1767         total += perf_event_read(event);
1768         list_for_each_entry(child, &event->child_list, child_list)
1769                 total += perf_event_read(child);
1770
1771         return total;
1772 }
1773
1774 static int perf_event_read_entry(struct perf_event *event,
1775                                    u64 read_format, char __user *buf)
1776 {
1777         int n = 0, count = 0;
1778         u64 values[2];
1779
1780         values[n++] = perf_event_read_value(event);
1781         if (read_format & PERF_FORMAT_ID)
1782                 values[n++] = primary_event_id(event);
1783
1784         count = n * sizeof(u64);
1785
1786         if (copy_to_user(buf, values, count))
1787                 return -EFAULT;
1788
1789         return count;
1790 }
1791
1792 static int perf_event_read_group(struct perf_event *event,
1793                                    u64 read_format, char __user *buf)
1794 {
1795         struct perf_event *leader = event->group_leader, *sub;
1796         int n = 0, size = 0, err = -EFAULT;
1797         u64 values[3];
1798
1799         values[n++] = 1 + leader->nr_siblings;
1800         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1801                 values[n++] = leader->total_time_enabled +
1802                         atomic64_read(&leader->child_total_time_enabled);
1803         }
1804         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1805                 values[n++] = leader->total_time_running +
1806                         atomic64_read(&leader->child_total_time_running);
1807         }
1808
1809         size = n * sizeof(u64);
1810
1811         if (copy_to_user(buf, values, size))
1812                 return -EFAULT;
1813
1814         err = perf_event_read_entry(leader, read_format, buf + size);
1815         if (err < 0)
1816                 return err;
1817
1818         size += err;
1819
1820         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1821                 err = perf_event_read_entry(sub, read_format,
1822                                 buf + size);
1823                 if (err < 0)
1824                         return err;
1825
1826                 size += err;
1827         }
1828
1829         return size;
1830 }
1831
1832 static int perf_event_read_one(struct perf_event *event,
1833                                  u64 read_format, char __user *buf)
1834 {
1835         u64 values[4];
1836         int n = 0;
1837
1838         values[n++] = perf_event_read_value(event);
1839         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1840                 values[n++] = event->total_time_enabled +
1841                         atomic64_read(&event->child_total_time_enabled);
1842         }
1843         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1844                 values[n++] = event->total_time_running +
1845                         atomic64_read(&event->child_total_time_running);
1846         }
1847         if (read_format & PERF_FORMAT_ID)
1848                 values[n++] = primary_event_id(event);
1849
1850         if (copy_to_user(buf, values, n * sizeof(u64)))
1851                 return -EFAULT;
1852
1853         return n * sizeof(u64);
1854 }
1855
1856 /*
1857  * Read the performance event - simple non blocking version for now
1858  */
1859 static ssize_t
1860 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
1861 {
1862         u64 read_format = event->attr.read_format;
1863         int ret;
1864
1865         /*
1866          * Return end-of-file for a read on a event that is in
1867          * error state (i.e. because it was pinned but it couldn't be
1868          * scheduled on to the CPU at some point).
1869          */
1870         if (event->state == PERF_EVENT_STATE_ERROR)
1871                 return 0;
1872
1873         if (count < perf_event_read_size(event))
1874                 return -ENOSPC;
1875
1876         WARN_ON_ONCE(event->ctx->parent_ctx);
1877         mutex_lock(&event->child_mutex);
1878         if (read_format & PERF_FORMAT_GROUP)
1879                 ret = perf_event_read_group(event, read_format, buf);
1880         else
1881                 ret = perf_event_read_one(event, read_format, buf);
1882         mutex_unlock(&event->child_mutex);
1883
1884         return ret;
1885 }
1886
1887 static ssize_t
1888 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1889 {
1890         struct perf_event *event = file->private_data;
1891
1892         return perf_read_hw(event, buf, count);
1893 }
1894
1895 static unsigned int perf_poll(struct file *file, poll_table *wait)
1896 {
1897         struct perf_event *event = file->private_data;
1898         struct perf_mmap_data *data;
1899         unsigned int events = POLL_HUP;
1900
1901         rcu_read_lock();
1902         data = rcu_dereference(event->data);
1903         if (data)
1904                 events = atomic_xchg(&data->poll, 0);
1905         rcu_read_unlock();
1906
1907         poll_wait(file, &event->waitq, wait);
1908
1909         return events;
1910 }
1911
1912 static void perf_event_reset(struct perf_event *event)
1913 {
1914         (void)perf_event_read(event);
1915         atomic64_set(&event->count, 0);
1916         perf_event_update_userpage(event);
1917 }
1918
1919 /*
1920  * Holding the top-level event's child_mutex means that any
1921  * descendant process that has inherited this event will block
1922  * in sync_child_event if it goes to exit, thus satisfying the
1923  * task existence requirements of perf_event_enable/disable.
1924  */
1925 static void perf_event_for_each_child(struct perf_event *event,
1926                                         void (*func)(struct perf_event *))
1927 {
1928         struct perf_event *child;
1929
1930         WARN_ON_ONCE(event->ctx->parent_ctx);
1931         mutex_lock(&event->child_mutex);
1932         func(event);
1933         list_for_each_entry(child, &event->child_list, child_list)
1934                 func(child);
1935         mutex_unlock(&event->child_mutex);
1936 }
1937
1938 static void perf_event_for_each(struct perf_event *event,
1939                                   void (*func)(struct perf_event *))
1940 {
1941         struct perf_event_context *ctx = event->ctx;
1942         struct perf_event *sibling;
1943
1944         WARN_ON_ONCE(ctx->parent_ctx);
1945         mutex_lock(&ctx->mutex);
1946         event = event->group_leader;
1947
1948         perf_event_for_each_child(event, func);
1949         func(event);
1950         list_for_each_entry(sibling, &event->sibling_list, group_entry)
1951                 perf_event_for_each_child(event, func);
1952         mutex_unlock(&ctx->mutex);
1953 }
1954
1955 static int perf_event_period(struct perf_event *event, u64 __user *arg)
1956 {
1957         struct perf_event_context *ctx = event->ctx;
1958         unsigned long size;
1959         int ret = 0;
1960         u64 value;
1961
1962         if (!event->attr.sample_period)
1963                 return -EINVAL;
1964
1965         size = copy_from_user(&value, arg, sizeof(value));
1966         if (size != sizeof(value))
1967                 return -EFAULT;
1968
1969         if (!value)
1970                 return -EINVAL;
1971
1972         spin_lock_irq(&ctx->lock);
1973         if (event->attr.freq) {
1974                 if (value > sysctl_perf_event_sample_rate) {
1975                         ret = -EINVAL;
1976                         goto unlock;
1977                 }
1978
1979                 event->attr.sample_freq = value;
1980         } else {
1981                 event->attr.sample_period = value;
1982                 event->hw.sample_period = value;
1983         }
1984 unlock:
1985         spin_unlock_irq(&ctx->lock);
1986
1987         return ret;
1988 }
1989
1990 int perf_event_set_output(struct perf_event *event, int output_fd);
1991
1992 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1993 {
1994         struct perf_event *event = file->private_data;
1995         void (*func)(struct perf_event *);
1996         u32 flags = arg;
1997
1998         switch (cmd) {
1999         case PERF_EVENT_IOC_ENABLE:
2000                 func = perf_event_enable;
2001                 break;
2002         case PERF_EVENT_IOC_DISABLE:
2003                 func = perf_event_disable;
2004                 break;
2005         case PERF_EVENT_IOC_RESET:
2006                 func = perf_event_reset;
2007                 break;
2008
2009         case PERF_EVENT_IOC_REFRESH:
2010                 return perf_event_refresh(event, arg);
2011
2012         case PERF_EVENT_IOC_PERIOD:
2013                 return perf_event_period(event, (u64 __user *)arg);
2014
2015         case PERF_EVENT_IOC_SET_OUTPUT:
2016                 return perf_event_set_output(event, arg);
2017
2018         default:
2019                 return -ENOTTY;
2020         }
2021
2022         if (flags & PERF_IOC_FLAG_GROUP)
2023                 perf_event_for_each(event, func);
2024         else
2025                 perf_event_for_each_child(event, func);
2026
2027         return 0;
2028 }
2029
2030 int perf_event_task_enable(void)
2031 {
2032         struct perf_event *event;
2033
2034         mutex_lock(&current->perf_event_mutex);
2035         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2036                 perf_event_for_each_child(event, perf_event_enable);
2037         mutex_unlock(&current->perf_event_mutex);
2038
2039         return 0;
2040 }
2041
2042 int perf_event_task_disable(void)
2043 {
2044         struct perf_event *event;
2045
2046         mutex_lock(&current->perf_event_mutex);
2047         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2048                 perf_event_for_each_child(event, perf_event_disable);
2049         mutex_unlock(&current->perf_event_mutex);
2050
2051         return 0;
2052 }
2053
2054 #ifndef PERF_EVENT_INDEX_OFFSET
2055 # define PERF_EVENT_INDEX_OFFSET 0
2056 #endif
2057
2058 static int perf_event_index(struct perf_event *event)
2059 {
2060         if (event->state != PERF_EVENT_STATE_ACTIVE)
2061                 return 0;
2062
2063         return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2064 }
2065
2066 /*
2067  * Callers need to ensure there can be no nesting of this function, otherwise
2068  * the seqlock logic goes bad. We can not serialize this because the arch
2069  * code calls this from NMI context.
2070  */
2071 void perf_event_update_userpage(struct perf_event *event)
2072 {
2073         struct perf_event_mmap_page *userpg;
2074         struct perf_mmap_data *data;
2075
2076         rcu_read_lock();
2077         data = rcu_dereference(event->data);
2078         if (!data)
2079                 goto unlock;
2080
2081         userpg = data->user_page;
2082
2083         /*
2084          * Disable preemption so as to not let the corresponding user-space
2085          * spin too long if we get preempted.
2086          */
2087         preempt_disable();
2088         ++userpg->lock;
2089         barrier();
2090         userpg->index = perf_event_index(event);
2091         userpg->offset = atomic64_read(&event->count);
2092         if (event->state == PERF_EVENT_STATE_ACTIVE)
2093                 userpg->offset -= atomic64_read(&event->hw.prev_count);
2094
2095         userpg->time_enabled = event->total_time_enabled +
2096                         atomic64_read(&event->child_total_time_enabled);
2097
2098         userpg->time_running = event->total_time_running +
2099                         atomic64_read(&event->child_total_time_running);
2100
2101         barrier();
2102         ++userpg->lock;
2103         preempt_enable();
2104 unlock:
2105         rcu_read_unlock();
2106 }
2107
2108 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2109 {
2110         struct perf_event *event = vma->vm_file->private_data;
2111         struct perf_mmap_data *data;
2112         int ret = VM_FAULT_SIGBUS;
2113
2114         if (vmf->flags & FAULT_FLAG_MKWRITE) {
2115                 if (vmf->pgoff == 0)
2116                         ret = 0;
2117                 return ret;
2118         }
2119
2120         rcu_read_lock();
2121         data = rcu_dereference(event->data);
2122         if (!data)
2123                 goto unlock;
2124
2125         if (vmf->pgoff == 0) {
2126                 vmf->page = virt_to_page(data->user_page);
2127         } else {
2128                 int nr = vmf->pgoff - 1;
2129
2130                 if ((unsigned)nr > data->nr_pages)
2131                         goto unlock;
2132
2133                 if (vmf->flags & FAULT_FLAG_WRITE)
2134                         goto unlock;
2135
2136                 vmf->page = virt_to_page(data->data_pages[nr]);
2137         }
2138
2139         get_page(vmf->page);
2140         vmf->page->mapping = vma->vm_file->f_mapping;
2141         vmf->page->index   = vmf->pgoff;
2142
2143         ret = 0;
2144 unlock:
2145         rcu_read_unlock();
2146
2147         return ret;
2148 }
2149
2150 static int perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2151 {
2152         struct perf_mmap_data *data;
2153         unsigned long size;
2154         int i;
2155
2156         WARN_ON(atomic_read(&event->mmap_count));
2157
2158         size = sizeof(struct perf_mmap_data);
2159         size += nr_pages * sizeof(void *);
2160
2161         data = kzalloc(size, GFP_KERNEL);
2162         if (!data)
2163                 goto fail;
2164
2165         data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2166         if (!data->user_page)
2167                 goto fail_user_page;
2168
2169         for (i = 0; i < nr_pages; i++) {
2170                 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2171                 if (!data->data_pages[i])
2172                         goto fail_data_pages;
2173         }
2174
2175         data->nr_pages = nr_pages;
2176         atomic_set(&data->lock, -1);
2177
2178         if (event->attr.watermark) {
2179                 data->watermark = min_t(long, PAGE_SIZE * nr_pages,
2180                                       event->attr.wakeup_watermark);
2181         }
2182         if (!data->watermark)
2183                 data->watermark = max(PAGE_SIZE, PAGE_SIZE * nr_pages / 4);
2184
2185         rcu_assign_pointer(event->data, data);
2186
2187         return 0;
2188
2189 fail_data_pages:
2190         for (i--; i >= 0; i--)
2191                 free_page((unsigned long)data->data_pages[i]);
2192
2193         free_page((unsigned long)data->user_page);
2194
2195 fail_user_page:
2196         kfree(data);
2197
2198 fail:
2199         return -ENOMEM;
2200 }
2201
2202 static void perf_mmap_free_page(unsigned long addr)
2203 {
2204         struct page *page = virt_to_page((void *)addr);
2205
2206         page->mapping = NULL;
2207         __free_page(page);
2208 }
2209
2210 static void __perf_mmap_data_free(struct rcu_head *rcu_head)
2211 {
2212         struct perf_mmap_data *data;
2213         int i;
2214
2215         data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2216
2217         perf_mmap_free_page((unsigned long)data->user_page);
2218         for (i = 0; i < data->nr_pages; i++)
2219                 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2220
2221         kfree(data);
2222 }
2223
2224 static void perf_mmap_data_free(struct perf_event *event)
2225 {
2226         struct perf_mmap_data *data = event->data;
2227
2228         WARN_ON(atomic_read(&event->mmap_count));
2229
2230         rcu_assign_pointer(event->data, NULL);
2231         call_rcu(&data->rcu_head, __perf_mmap_data_free);
2232 }
2233
2234 static void perf_mmap_open(struct vm_area_struct *vma)
2235 {
2236         struct perf_event *event = vma->vm_file->private_data;
2237
2238         atomic_inc(&event->mmap_count);
2239 }
2240
2241 static void perf_mmap_close(struct vm_area_struct *vma)
2242 {
2243         struct perf_event *event = vma->vm_file->private_data;
2244
2245         WARN_ON_ONCE(event->ctx->parent_ctx);
2246         if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2247                 struct user_struct *user = current_user();
2248
2249                 atomic_long_sub(event->data->nr_pages + 1, &user->locked_vm);
2250                 vma->vm_mm->locked_vm -= event->data->nr_locked;
2251                 perf_mmap_data_free(event);
2252                 mutex_unlock(&event->mmap_mutex);
2253         }
2254 }
2255
2256 static struct vm_operations_struct perf_mmap_vmops = {
2257         .open           = perf_mmap_open,
2258         .close          = perf_mmap_close,
2259         .fault          = perf_mmap_fault,
2260         .page_mkwrite   = perf_mmap_fault,
2261 };
2262
2263 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2264 {
2265         struct perf_event *event = file->private_data;
2266         unsigned long user_locked, user_lock_limit;
2267         struct user_struct *user = current_user();
2268         unsigned long locked, lock_limit;
2269         unsigned long vma_size;
2270         unsigned long nr_pages;
2271         long user_extra, extra;
2272         int ret = 0;
2273
2274         if (!(vma->vm_flags & VM_SHARED))
2275                 return -EINVAL;
2276
2277         vma_size = vma->vm_end - vma->vm_start;
2278         nr_pages = (vma_size / PAGE_SIZE) - 1;
2279
2280         /*
2281          * If we have data pages ensure they're a power-of-two number, so we
2282          * can do bitmasks instead of modulo.
2283          */
2284         if (nr_pages != 0 && !is_power_of_2(nr_pages))
2285                 return -EINVAL;
2286
2287         if (vma_size != PAGE_SIZE * (1 + nr_pages))
2288                 return -EINVAL;
2289
2290         if (vma->vm_pgoff != 0)
2291                 return -EINVAL;
2292
2293         WARN_ON_ONCE(event->ctx->parent_ctx);
2294         mutex_lock(&event->mmap_mutex);
2295         if (event->output) {
2296                 ret = -EINVAL;
2297                 goto unlock;
2298         }
2299
2300         if (atomic_inc_not_zero(&event->mmap_count)) {
2301                 if (nr_pages != event->data->nr_pages)
2302                         ret = -EINVAL;
2303                 goto unlock;
2304         }
2305
2306         user_extra = nr_pages + 1;
2307         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2308
2309         /*
2310          * Increase the limit linearly with more CPUs:
2311          */
2312         user_lock_limit *= num_online_cpus();
2313
2314         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2315
2316         extra = 0;
2317         if (user_locked > user_lock_limit)
2318                 extra = user_locked - user_lock_limit;
2319
2320         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2321         lock_limit >>= PAGE_SHIFT;
2322         locked = vma->vm_mm->locked_vm + extra;
2323
2324         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2325                 !capable(CAP_IPC_LOCK)) {
2326                 ret = -EPERM;
2327                 goto unlock;
2328         }
2329
2330         WARN_ON(event->data);
2331         ret = perf_mmap_data_alloc(event, nr_pages);
2332         if (ret)
2333                 goto unlock;
2334
2335         atomic_set(&event->mmap_count, 1);
2336         atomic_long_add(user_extra, &user->locked_vm);
2337         vma->vm_mm->locked_vm += extra;
2338         event->data->nr_locked = extra;
2339         if (vma->vm_flags & VM_WRITE)
2340                 event->data->writable = 1;
2341
2342 unlock:
2343         mutex_unlock(&event->mmap_mutex);
2344
2345         vma->vm_flags |= VM_RESERVED;
2346         vma->vm_ops = &perf_mmap_vmops;
2347
2348         return ret;
2349 }
2350
2351 static int perf_fasync(int fd, struct file *filp, int on)
2352 {
2353         struct inode *inode = filp->f_path.dentry->d_inode;
2354         struct perf_event *event = filp->private_data;
2355         int retval;
2356
2357         mutex_lock(&inode->i_mutex);
2358         retval = fasync_helper(fd, filp, on, &event->fasync);
2359         mutex_unlock(&inode->i_mutex);
2360
2361         if (retval < 0)
2362                 return retval;
2363
2364         return 0;
2365 }
2366
2367 static const struct file_operations perf_fops = {
2368         .release                = perf_release,
2369         .read                   = perf_read,
2370         .poll                   = perf_poll,
2371         .unlocked_ioctl         = perf_ioctl,
2372         .compat_ioctl           = perf_ioctl,
2373         .mmap                   = perf_mmap,
2374         .fasync                 = perf_fasync,
2375 };
2376
2377 /*
2378  * Perf event wakeup
2379  *
2380  * If there's data, ensure we set the poll() state and publish everything
2381  * to user-space before waking everybody up.
2382  */
2383
2384 void perf_event_wakeup(struct perf_event *event)
2385 {
2386         wake_up_all(&event->waitq);
2387
2388         if (event->pending_kill) {
2389                 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2390                 event->pending_kill = 0;
2391         }
2392 }
2393
2394 /*
2395  * Pending wakeups
2396  *
2397  * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2398  *
2399  * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2400  * single linked list and use cmpxchg() to add entries lockless.
2401  */
2402
2403 static void perf_pending_event(struct perf_pending_entry *entry)
2404 {
2405         struct perf_event *event = container_of(entry,
2406                         struct perf_event, pending);
2407
2408         if (event->pending_disable) {
2409                 event->pending_disable = 0;
2410                 __perf_event_disable(event);
2411         }
2412
2413         if (event->pending_wakeup) {
2414                 event->pending_wakeup = 0;
2415                 perf_event_wakeup(event);
2416         }
2417 }
2418
2419 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2420
2421 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2422         PENDING_TAIL,
2423 };
2424
2425 static void perf_pending_queue(struct perf_pending_entry *entry,
2426                                void (*func)(struct perf_pending_entry *))
2427 {
2428         struct perf_pending_entry **head;
2429
2430         if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2431                 return;
2432
2433         entry->func = func;
2434
2435         head = &get_cpu_var(perf_pending_head);
2436
2437         do {
2438                 entry->next = *head;
2439         } while (cmpxchg(head, entry->next, entry) != entry->next);
2440
2441         set_perf_event_pending();
2442
2443         put_cpu_var(perf_pending_head);
2444 }
2445
2446 static int __perf_pending_run(void)
2447 {
2448         struct perf_pending_entry *list;
2449         int nr = 0;
2450
2451         list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2452         while (list != PENDING_TAIL) {
2453                 void (*func)(struct perf_pending_entry *);
2454                 struct perf_pending_entry *entry = list;
2455
2456                 list = list->next;
2457
2458                 func = entry->func;
2459                 entry->next = NULL;
2460                 /*
2461                  * Ensure we observe the unqueue before we issue the wakeup,
2462                  * so that we won't be waiting forever.
2463                  * -- see perf_not_pending().
2464                  */
2465                 smp_wmb();
2466
2467                 func(entry);
2468                 nr++;
2469         }
2470
2471         return nr;
2472 }
2473
2474 static inline int perf_not_pending(struct perf_event *event)
2475 {
2476         /*
2477          * If we flush on whatever cpu we run, there is a chance we don't
2478          * need to wait.
2479          */
2480         get_cpu();
2481         __perf_pending_run();
2482         put_cpu();
2483
2484         /*
2485          * Ensure we see the proper queue state before going to sleep
2486          * so that we do not miss the wakeup. -- see perf_pending_handle()
2487          */
2488         smp_rmb();
2489         return event->pending.next == NULL;
2490 }
2491
2492 static void perf_pending_sync(struct perf_event *event)
2493 {
2494         wait_event(event->waitq, perf_not_pending(event));
2495 }
2496
2497 void perf_event_do_pending(void)
2498 {
2499         __perf_pending_run();
2500 }
2501
2502 /*
2503  * Callchain support -- arch specific
2504  */
2505
2506 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2507 {
2508         return NULL;
2509 }
2510
2511 /*
2512  * Output
2513  */
2514 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2515                               unsigned long offset, unsigned long head)
2516 {
2517         unsigned long mask;
2518
2519         if (!data->writable)
2520                 return true;
2521
2522         mask = (data->nr_pages << PAGE_SHIFT) - 1;
2523
2524         offset = (offset - tail) & mask;
2525         head   = (head   - tail) & mask;
2526
2527         if ((int)(head - offset) < 0)
2528                 return false;
2529
2530         return true;
2531 }
2532
2533 static void perf_output_wakeup(struct perf_output_handle *handle)
2534 {
2535         atomic_set(&handle->data->poll, POLL_IN);
2536
2537         if (handle->nmi) {
2538                 handle->event->pending_wakeup = 1;
2539                 perf_pending_queue(&handle->event->pending,
2540                                    perf_pending_event);
2541         } else
2542                 perf_event_wakeup(handle->event);
2543 }
2544
2545 /*
2546  * Curious locking construct.
2547  *
2548  * We need to ensure a later event_id doesn't publish a head when a former
2549  * event_id isn't done writing. However since we need to deal with NMIs we
2550  * cannot fully serialize things.
2551  *
2552  * What we do is serialize between CPUs so we only have to deal with NMI
2553  * nesting on a single CPU.
2554  *
2555  * We only publish the head (and generate a wakeup) when the outer-most
2556  * event_id completes.
2557  */
2558 static void perf_output_lock(struct perf_output_handle *handle)
2559 {
2560         struct perf_mmap_data *data = handle->data;
2561         int cpu;
2562
2563         handle->locked = 0;
2564
2565         local_irq_save(handle->flags);
2566         cpu = smp_processor_id();
2567
2568         if (in_nmi() && atomic_read(&data->lock) == cpu)
2569                 return;
2570
2571         while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2572                 cpu_relax();
2573
2574         handle->locked = 1;
2575 }
2576
2577 static void perf_output_unlock(struct perf_output_handle *handle)
2578 {
2579         struct perf_mmap_data *data = handle->data;
2580         unsigned long head;
2581         int cpu;
2582
2583         data->done_head = data->head;
2584
2585         if (!handle->locked)
2586                 goto out;
2587
2588 again:
2589         /*
2590          * The xchg implies a full barrier that ensures all writes are done
2591          * before we publish the new head, matched by a rmb() in userspace when
2592          * reading this position.
2593          */
2594         while ((head = atomic_long_xchg(&data->done_head, 0)))
2595                 data->user_page->data_head = head;
2596
2597         /*
2598          * NMI can happen here, which means we can miss a done_head update.
2599          */
2600
2601         cpu = atomic_xchg(&data->lock, -1);
2602         WARN_ON_ONCE(cpu != smp_processor_id());
2603
2604         /*
2605          * Therefore we have to validate we did not indeed do so.
2606          */
2607         if (unlikely(atomic_long_read(&data->done_head))) {
2608                 /*
2609                  * Since we had it locked, we can lock it again.
2610                  */
2611                 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2612                         cpu_relax();
2613
2614                 goto again;
2615         }
2616
2617         if (atomic_xchg(&data->wakeup, 0))
2618                 perf_output_wakeup(handle);
2619 out:
2620         local_irq_restore(handle->flags);
2621 }
2622
2623 void perf_output_copy(struct perf_output_handle *handle,
2624                       const void *buf, unsigned int len)
2625 {
2626         unsigned int pages_mask;
2627         unsigned int offset;
2628         unsigned int size;
2629         void **pages;
2630
2631         offset          = handle->offset;
2632         pages_mask      = handle->data->nr_pages - 1;
2633         pages           = handle->data->data_pages;
2634
2635         do {
2636                 unsigned int page_offset;
2637                 int nr;
2638
2639                 nr          = (offset >> PAGE_SHIFT) & pages_mask;
2640                 page_offset = offset & (PAGE_SIZE - 1);
2641                 size        = min_t(unsigned int, PAGE_SIZE - page_offset, len);
2642
2643                 memcpy(pages[nr] + page_offset, buf, size);
2644
2645                 len         -= size;
2646                 buf         += size;
2647                 offset      += size;
2648         } while (len);
2649
2650         handle->offset = offset;
2651
2652         /*
2653          * Check we didn't copy past our reservation window, taking the
2654          * possible unsigned int wrap into account.
2655          */
2656         WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2657 }
2658
2659 int perf_output_begin(struct perf_output_handle *handle,
2660                       struct perf_event *event, unsigned int size,
2661                       int nmi, int sample)
2662 {
2663         struct perf_event *output_event;
2664         struct perf_mmap_data *data;
2665         unsigned long tail, offset, head;
2666         int have_lost;
2667         struct {
2668                 struct perf_event_header header;
2669                 u64                      id;
2670                 u64                      lost;
2671         } lost_event;
2672
2673         rcu_read_lock();
2674         /*
2675          * For inherited events we send all the output towards the parent.
2676          */
2677         if (event->parent)
2678                 event = event->parent;
2679
2680         output_event = rcu_dereference(event->output);
2681         if (output_event)
2682                 event = output_event;
2683
2684         data = rcu_dereference(event->data);
2685         if (!data)
2686                 goto out;
2687
2688         handle->data    = data;
2689         handle->event   = event;
2690         handle->nmi     = nmi;
2691         handle->sample  = sample;
2692
2693         if (!data->nr_pages)
2694                 goto fail;
2695
2696         have_lost = atomic_read(&data->lost);
2697         if (have_lost)
2698                 size += sizeof(lost_event);
2699
2700         perf_output_lock(handle);
2701
2702         do {
2703                 /*
2704                  * Userspace could choose to issue a mb() before updating the
2705                  * tail pointer. So that all reads will be completed before the
2706                  * write is issued.
2707                  */
2708                 tail = ACCESS_ONCE(data->user_page->data_tail);
2709                 smp_rmb();
2710                 offset = head = atomic_long_read(&data->head);
2711                 head += size;
2712                 if (unlikely(!perf_output_space(data, tail, offset, head)))
2713                         goto fail;
2714         } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2715
2716         handle->offset  = offset;
2717         handle->head    = head;
2718
2719         if (head - tail > data->watermark)
2720                 atomic_set(&data->wakeup, 1);
2721
2722         if (have_lost) {
2723                 lost_event.header.type = PERF_RECORD_LOST;
2724                 lost_event.header.misc = 0;
2725                 lost_event.header.size = sizeof(lost_event);
2726                 lost_event.id          = event->id;
2727                 lost_event.lost        = atomic_xchg(&data->lost, 0);
2728
2729                 perf_output_put(handle, lost_event);
2730         }
2731
2732         return 0;
2733
2734 fail:
2735         atomic_inc(&data->lost);
2736         perf_output_unlock(handle);
2737 out:
2738         rcu_read_unlock();
2739
2740         return -ENOSPC;
2741 }
2742
2743 void perf_output_end(struct perf_output_handle *handle)
2744 {
2745         struct perf_event *event = handle->event;
2746         struct perf_mmap_data *data = handle->data;
2747
2748         int wakeup_events = event->attr.wakeup_events;
2749
2750         if (handle->sample && wakeup_events) {
2751                 int events = atomic_inc_return(&data->events);
2752                 if (events >= wakeup_events) {
2753                         atomic_sub(wakeup_events, &data->events);
2754                         atomic_set(&data->wakeup, 1);
2755                 }
2756         }
2757
2758         perf_output_unlock(handle);
2759         rcu_read_unlock();
2760 }
2761
2762 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
2763 {
2764         /*
2765          * only top level events have the pid namespace they were created in
2766          */
2767         if (event->parent)
2768                 event = event->parent;
2769
2770         return task_tgid_nr_ns(p, event->ns);
2771 }
2772
2773 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
2774 {
2775         /*
2776          * only top level events have the pid namespace they were created in
2777          */
2778         if (event->parent)
2779                 event = event->parent;
2780
2781         return task_pid_nr_ns(p, event->ns);
2782 }
2783
2784 static void perf_output_read_one(struct perf_output_handle *handle,
2785                                  struct perf_event *event)
2786 {
2787         u64 read_format = event->attr.read_format;
2788         u64 values[4];
2789         int n = 0;
2790
2791         values[n++] = atomic64_read(&event->count);
2792         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2793                 values[n++] = event->total_time_enabled +
2794                         atomic64_read(&event->child_total_time_enabled);
2795         }
2796         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2797                 values[n++] = event->total_time_running +
2798                         atomic64_read(&event->child_total_time_running);
2799         }
2800         if (read_format & PERF_FORMAT_ID)
2801                 values[n++] = primary_event_id(event);
2802
2803         perf_output_copy(handle, values, n * sizeof(u64));
2804 }
2805
2806 /*
2807  * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
2808  */
2809 static void perf_output_read_group(struct perf_output_handle *handle,
2810                             struct perf_event *event)
2811 {
2812         struct perf_event *leader = event->group_leader, *sub;
2813         u64 read_format = event->attr.read_format;
2814         u64 values[5];
2815         int n = 0;
2816
2817         values[n++] = 1 + leader->nr_siblings;
2818
2819         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2820                 values[n++] = leader->total_time_enabled;
2821
2822         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2823                 values[n++] = leader->total_time_running;
2824
2825         if (leader != event)
2826                 leader->pmu->read(leader);
2827
2828         values[n++] = atomic64_read(&leader->count);
2829         if (read_format & PERF_FORMAT_ID)
2830                 values[n++] = primary_event_id(leader);
2831
2832         perf_output_copy(handle, values, n * sizeof(u64));
2833
2834         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2835                 n = 0;
2836
2837                 if (sub != event)
2838                         sub->pmu->read(sub);
2839
2840                 values[n++] = atomic64_read(&sub->count);
2841                 if (read_format & PERF_FORMAT_ID)
2842                         values[n++] = primary_event_id(sub);
2843
2844                 perf_output_copy(handle, values, n * sizeof(u64));
2845         }
2846 }
2847
2848 static void perf_output_read(struct perf_output_handle *handle,
2849                              struct perf_event *event)
2850 {
2851         if (event->attr.read_format & PERF_FORMAT_GROUP)
2852                 perf_output_read_group(handle, event);
2853         else
2854                 perf_output_read_one(handle, event);
2855 }
2856
2857 void perf_output_sample(struct perf_output_handle *handle,
2858                         struct perf_event_header *header,
2859                         struct perf_sample_data *data,
2860                         struct perf_event *event)
2861 {
2862         u64 sample_type = data->type;
2863
2864         perf_output_put(handle, *header);
2865
2866         if (sample_type & PERF_SAMPLE_IP)
2867                 perf_output_put(handle, data->ip);
2868
2869         if (sample_type & PERF_SAMPLE_TID)
2870                 perf_output_put(handle, data->tid_entry);
2871
2872         if (sample_type & PERF_SAMPLE_TIME)
2873                 perf_output_put(handle, data->time);
2874
2875         if (sample_type & PERF_SAMPLE_ADDR)
2876                 perf_output_put(handle, data->addr);
2877
2878         if (sample_type & PERF_SAMPLE_ID)
2879                 perf_output_put(handle, data->id);
2880
2881         if (sample_type & PERF_SAMPLE_STREAM_ID)
2882                 perf_output_put(handle, data->stream_id);
2883
2884         if (sample_type & PERF_SAMPLE_CPU)
2885                 perf_output_put(handle, data->cpu_entry);
2886
2887         if (sample_type & PERF_SAMPLE_PERIOD)
2888                 perf_output_put(handle, data->period);
2889
2890         if (sample_type & PERF_SAMPLE_READ)
2891                 perf_output_read(handle, event);
2892
2893         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
2894                 if (data->callchain) {
2895                         int size = 1;
2896
2897                         if (data->callchain)
2898                                 size += data->callchain->nr;
2899
2900                         size *= sizeof(u64);
2901
2902                         perf_output_copy(handle, data->callchain, size);
2903                 } else {
2904                         u64 nr = 0;
2905                         perf_output_put(handle, nr);
2906                 }
2907         }
2908
2909         if (sample_type & PERF_SAMPLE_RAW) {
2910                 if (data->raw) {
2911                         perf_output_put(handle, data->raw->size);
2912                         perf_output_copy(handle, data->raw->data,
2913                                          data->raw->size);
2914                 } else {
2915                         struct {
2916                                 u32     size;
2917                                 u32     data;
2918                         } raw = {
2919                                 .size = sizeof(u32),
2920                                 .data = 0,
2921                         };
2922                         perf_output_put(handle, raw);
2923                 }
2924         }
2925 }
2926
2927 void perf_prepare_sample(struct perf_event_header *header,
2928                          struct perf_sample_data *data,
2929                          struct perf_event *event,
2930                          struct pt_regs *regs)
2931 {
2932         u64 sample_type = event->attr.sample_type;
2933
2934         data->type = sample_type;
2935
2936         header->type = PERF_RECORD_SAMPLE;
2937         header->size = sizeof(*header);
2938
2939         header->misc = 0;
2940         header->misc |= perf_misc_flags(regs);
2941
2942         if (sample_type & PERF_SAMPLE_IP) {
2943                 data->ip = perf_instruction_pointer(regs);
2944
2945                 header->size += sizeof(data->ip);
2946         }
2947
2948         if (sample_type & PERF_SAMPLE_TID) {
2949                 /* namespace issues */
2950                 data->tid_entry.pid = perf_event_pid(event, current);
2951                 data->tid_entry.tid = perf_event_tid(event, current);
2952
2953                 header->size += sizeof(data->tid_entry);
2954         }
2955
2956         if (sample_type & PERF_SAMPLE_TIME) {
2957                 data->time = perf_clock();
2958
2959                 header->size += sizeof(data->time);
2960         }
2961
2962         if (sample_type & PERF_SAMPLE_ADDR)
2963                 header->size += sizeof(data->addr);
2964
2965         if (sample_type & PERF_SAMPLE_ID) {
2966                 data->id = primary_event_id(event);
2967
2968                 header->size += sizeof(data->id);
2969         }
2970
2971         if (sample_type & PERF_SAMPLE_STREAM_ID) {
2972                 data->stream_id = event->id;
2973
2974                 header->size += sizeof(data->stream_id);
2975         }
2976
2977         if (sample_type & PERF_SAMPLE_CPU) {
2978                 data->cpu_entry.cpu             = raw_smp_processor_id();
2979                 data->cpu_entry.reserved        = 0;
2980
2981                 header->size += sizeof(data->cpu_entry);
2982         }
2983
2984         if (sample_type & PERF_SAMPLE_PERIOD)
2985                 header->size += sizeof(data->period);
2986
2987         if (sample_type & PERF_SAMPLE_READ)
2988                 header->size += perf_event_read_size(event);
2989
2990         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
2991                 int size = 1;
2992
2993                 data->callchain = perf_callchain(regs);
2994
2995                 if (data->callchain)
2996                         size += data->callchain->nr;
2997
2998                 header->size += size * sizeof(u64);
2999         }
3000
3001         if (sample_type & PERF_SAMPLE_RAW) {
3002                 int size = sizeof(u32);
3003
3004                 if (data->raw)
3005                         size += data->raw->size;
3006                 else
3007                         size += sizeof(u32);
3008
3009                 WARN_ON_ONCE(size & (sizeof(u64)-1));
3010                 header->size += size;
3011         }
3012 }
3013
3014 static void perf_event_output(struct perf_event *event, int nmi,
3015                                 struct perf_sample_data *data,
3016                                 struct pt_regs *regs)
3017 {
3018         struct perf_output_handle handle;
3019         struct perf_event_header header;
3020
3021         perf_prepare_sample(&header, data, event, regs);
3022
3023         if (perf_output_begin(&handle, event, header.size, nmi, 1))
3024                 return;
3025
3026         perf_output_sample(&handle, &header, data, event);
3027
3028         perf_output_end(&handle);
3029 }
3030
3031 /*
3032  * read event_id
3033  */
3034
3035 struct perf_read_event {
3036         struct perf_event_header        header;
3037
3038         u32                             pid;
3039         u32                             tid;
3040 };
3041
3042 static void
3043 perf_event_read_event(struct perf_event *event,
3044                         struct task_struct *task)
3045 {
3046         struct perf_output_handle handle;
3047         struct perf_read_event read_event = {
3048                 .header = {
3049                         .type = PERF_RECORD_READ,
3050                         .misc = 0,
3051                         .size = sizeof(read_event) + perf_event_read_size(event),
3052                 },
3053                 .pid = perf_event_pid(event, task),
3054                 .tid = perf_event_tid(event, task),
3055         };
3056         int ret;
3057
3058         ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3059         if (ret)
3060                 return;
3061
3062         perf_output_put(&handle, read_event);
3063         perf_output_read(&handle, event);
3064
3065         perf_output_end(&handle);
3066 }
3067
3068 /*
3069  * task tracking -- fork/exit
3070  *
3071  * enabled by: attr.comm | attr.mmap | attr.task
3072  */
3073
3074 struct perf_task_event {
3075         struct task_struct              *task;
3076         struct perf_event_context       *task_ctx;
3077
3078         struct {
3079                 struct perf_event_header        header;
3080
3081                 u32                             pid;
3082                 u32                             ppid;
3083                 u32                             tid;
3084                 u32                             ptid;
3085                 u64                             time;
3086         } event_id;
3087 };
3088
3089 static void perf_event_task_output(struct perf_event *event,
3090                                      struct perf_task_event *task_event)
3091 {
3092         struct perf_output_handle handle;
3093         int size;
3094         struct task_struct *task = task_event->task;
3095         int ret;
3096
3097         size  = task_event->event_id.header.size;
3098         ret = perf_output_begin(&handle, event, size, 0, 0);
3099
3100         if (ret)
3101                 return;
3102
3103         task_event->event_id.pid = perf_event_pid(event, task);
3104         task_event->event_id.ppid = perf_event_pid(event, current);
3105
3106         task_event->event_id.tid = perf_event_tid(event, task);
3107         task_event->event_id.ptid = perf_event_tid(event, current);
3108
3109         task_event->event_id.time = perf_clock();
3110
3111         perf_output_put(&handle, task_event->event_id);
3112
3113         perf_output_end(&handle);
3114 }
3115
3116 static int perf_event_task_match(struct perf_event *event)
3117 {
3118         if (event->attr.comm || event->attr.mmap || event->attr.task)
3119                 return 1;
3120
3121         return 0;
3122 }
3123
3124 static void perf_event_task_ctx(struct perf_event_context *ctx,
3125                                   struct perf_task_event *task_event)
3126 {
3127         struct perf_event *event;
3128
3129         if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3130                 return;
3131
3132         rcu_read_lock();
3133         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3134                 if (perf_event_task_match(event))
3135                         perf_event_task_output(event, task_event);
3136         }
3137         rcu_read_unlock();
3138 }
3139
3140 static void perf_event_task_event(struct perf_task_event *task_event)
3141 {
3142         struct perf_cpu_context *cpuctx;
3143         struct perf_event_context *ctx = task_event->task_ctx;
3144
3145         cpuctx = &get_cpu_var(perf_cpu_context);
3146         perf_event_task_ctx(&cpuctx->ctx, task_event);
3147         put_cpu_var(perf_cpu_context);
3148
3149         rcu_read_lock();
3150         if (!ctx)
3151                 ctx = rcu_dereference(task_event->task->perf_event_ctxp);
3152         if (ctx)
3153                 perf_event_task_ctx(ctx, task_event);
3154         rcu_read_unlock();
3155 }
3156
3157 static void perf_event_task(struct task_struct *task,
3158                               struct perf_event_context *task_ctx,
3159                               int new)
3160 {
3161         struct perf_task_event task_event;
3162
3163         if (!atomic_read(&nr_comm_events) &&
3164             !atomic_read(&nr_mmap_events) &&
3165             !atomic_read(&nr_task_events))
3166                 return;
3167
3168         task_event = (struct perf_task_event){
3169                 .task     = task,
3170                 .task_ctx = task_ctx,
3171                 .event_id    = {
3172                         .header = {
3173                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3174                                 .misc = 0,
3175                                 .size = sizeof(task_event.event_id),
3176                         },
3177                         /* .pid  */
3178                         /* .ppid */
3179                         /* .tid  */
3180                         /* .ptid */
3181                 },
3182         };
3183
3184         perf_event_task_event(&task_event);
3185 }
3186
3187 void perf_event_fork(struct task_struct *task)
3188 {
3189         perf_event_task(task, NULL, 1);
3190 }
3191
3192 /*
3193  * comm tracking
3194  */
3195
3196 struct perf_comm_event {
3197         struct task_struct      *task;
3198         char                    *comm;
3199         int                     comm_size;
3200
3201         struct {
3202                 struct perf_event_header        header;
3203
3204                 u32                             pid;
3205                 u32                             tid;
3206         } event_id;
3207 };
3208
3209 static void perf_event_comm_output(struct perf_event *event,
3210                                      struct perf_comm_event *comm_event)
3211 {
3212         struct perf_output_handle handle;
3213         int size = comm_event->event_id.header.size;
3214         int ret = perf_output_begin(&handle, event, size, 0, 0);
3215
3216         if (ret)
3217                 return;
3218
3219         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3220         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3221
3222         perf_output_put(&handle, comm_event->event_id);
3223         perf_output_copy(&handle, comm_event->comm,
3224                                    comm_event->comm_size);
3225         perf_output_end(&handle);
3226 }
3227
3228 static int perf_event_comm_match(struct perf_event *event)
3229 {
3230         if (event->attr.comm)
3231                 return 1;
3232
3233         return 0;
3234 }
3235
3236 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3237                                   struct perf_comm_event *comm_event)
3238 {
3239         struct perf_event *event;
3240
3241         if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3242                 return;
3243
3244         rcu_read_lock();
3245         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3246                 if (perf_event_comm_match(event))
3247                         perf_event_comm_output(event, comm_event);
3248         }
3249         rcu_read_unlock();
3250 }
3251
3252 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3253 {
3254         struct perf_cpu_context *cpuctx;
3255         struct perf_event_context *ctx;
3256         unsigned int size;
3257         char comm[TASK_COMM_LEN];
3258
3259         memset(comm, 0, sizeof(comm));
3260         strncpy(comm, comm_event->task->comm, sizeof(comm));
3261         size = ALIGN(strlen(comm)+1, sizeof(u64));
3262
3263         comm_event->comm = comm;
3264         comm_event->comm_size = size;
3265
3266         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3267
3268         cpuctx = &get_cpu_var(perf_cpu_context);
3269         perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3270         put_cpu_var(perf_cpu_context);
3271
3272         rcu_read_lock();
3273         /*
3274          * doesn't really matter which of the child contexts the
3275          * events ends up in.
3276          */
3277         ctx = rcu_dereference(current->perf_event_ctxp);
3278         if (ctx)
3279                 perf_event_comm_ctx(ctx, comm_event);
3280         rcu_read_unlock();
3281 }
3282
3283 void perf_event_comm(struct task_struct *task)
3284 {
3285         struct perf_comm_event comm_event;
3286
3287         if (task->perf_event_ctxp)
3288                 perf_event_enable_on_exec(task);
3289
3290         if (!atomic_read(&nr_comm_events))
3291                 return;
3292
3293         comm_event = (struct perf_comm_event){
3294                 .task   = task,
3295                 /* .comm      */
3296                 /* .comm_size */
3297                 .event_id  = {
3298                         .header = {
3299                                 .type = PERF_RECORD_COMM,
3300                                 .misc = 0,
3301                                 /* .size */
3302                         },
3303                         /* .pid */
3304                         /* .tid */
3305                 },
3306         };
3307
3308         perf_event_comm_event(&comm_event);
3309 }
3310
3311 /*
3312  * mmap tracking
3313  */
3314
3315 struct perf_mmap_event {
3316         struct vm_area_struct   *vma;
3317
3318         const char              *file_name;
3319         int                     file_size;
3320
3321         struct {
3322                 struct perf_event_header        header;
3323
3324                 u32                             pid;
3325                 u32                             tid;
3326                 u64                             start;
3327                 u64                             len;
3328                 u64                             pgoff;
3329         } event_id;
3330 };
3331
3332 static void perf_event_mmap_output(struct perf_event *event,
3333                                      struct perf_mmap_event *mmap_event)
3334 {
3335         struct perf_output_handle handle;
3336         int size = mmap_event->event_id.header.size;
3337         int ret = perf_output_begin(&handle, event, size, 0, 0);
3338
3339         if (ret)
3340                 return;
3341
3342         mmap_event->event_id.pid = perf_event_pid(event, current);
3343         mmap_event->event_id.tid = perf_event_tid(event, current);
3344
3345         perf_output_put(&handle, mmap_event->event_id);
3346         perf_output_copy(&handle, mmap_event->file_name,
3347                                    mmap_event->file_size);
3348         perf_output_end(&handle);
3349 }
3350
3351 static int perf_event_mmap_match(struct perf_event *event,
3352                                    struct perf_mmap_event *mmap_event)
3353 {
3354         if (event->attr.mmap)
3355                 return 1;
3356
3357         return 0;
3358 }
3359
3360 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3361                                   struct perf_mmap_event *mmap_event)
3362 {
3363         struct perf_event *event;
3364
3365         if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3366                 return;
3367
3368         rcu_read_lock();
3369         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3370                 if (perf_event_mmap_match(event, mmap_event))
3371                         perf_event_mmap_output(event, mmap_event);
3372         }
3373         rcu_read_unlock();
3374 }
3375
3376 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3377 {
3378         struct perf_cpu_context *cpuctx;
3379         struct perf_event_context *ctx;
3380         struct vm_area_struct *vma = mmap_event->vma;
3381         struct file *file = vma->vm_file;
3382         unsigned int size;
3383         char tmp[16];
3384         char *buf = NULL;
3385         const char *name;
3386
3387         memset(tmp, 0, sizeof(tmp));
3388
3389         if (file) {
3390                 /*
3391                  * d_path works from the end of the buffer backwards, so we
3392                  * need to add enough zero bytes after the string to handle
3393                  * the 64bit alignment we do later.
3394                  */
3395                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3396                 if (!buf) {
3397                         name = strncpy(tmp, "//enomem", sizeof(tmp));
3398                         goto got_name;
3399                 }
3400                 name = d_path(&file->f_path, buf, PATH_MAX);
3401                 if (IS_ERR(name)) {
3402                         name = strncpy(tmp, "//toolong", sizeof(tmp));
3403                         goto got_name;
3404                 }
3405         } else {
3406                 if (arch_vma_name(mmap_event->vma)) {
3407                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3408                                        sizeof(tmp));
3409                         goto got_name;
3410                 }
3411
3412                 if (!vma->vm_mm) {
3413                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
3414                         goto got_name;
3415                 }
3416
3417                 name = strncpy(tmp, "//anon", sizeof(tmp));
3418                 goto got_name;
3419         }
3420
3421 got_name:
3422         size = ALIGN(strlen(name)+1, sizeof(u64));
3423
3424         mmap_event->file_name = name;
3425         mmap_event->file_size = size;
3426
3427         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3428
3429         cpuctx = &get_cpu_var(perf_cpu_context);
3430         perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3431         put_cpu_var(perf_cpu_context);
3432
3433         rcu_read_lock();
3434         /*
3435          * doesn't really matter which of the child contexts the
3436          * events ends up in.
3437          */
3438         ctx = rcu_dereference(current->perf_event_ctxp);
3439         if (ctx)
3440                 perf_event_mmap_ctx(ctx, mmap_event);
3441         rcu_read_unlock();
3442
3443         kfree(buf);
3444 }
3445
3446 void __perf_event_mmap(struct vm_area_struct *vma)
3447 {
3448         struct perf_mmap_event mmap_event;
3449
3450         if (!atomic_read(&nr_mmap_events))
3451                 return;
3452
3453         mmap_event = (struct perf_mmap_event){
3454                 .vma    = vma,
3455                 /* .file_name */
3456                 /* .file_size */
3457                 .event_id  = {
3458                         .header = {
3459                                 .type = PERF_RECORD_MMAP,
3460                                 .misc = 0,
3461                                 /* .size */
3462                         },
3463                         /* .pid */
3464                         /* .tid */
3465                         .start  = vma->vm_start,
3466                         .len    = vma->vm_end - vma->vm_start,
3467                         .pgoff  = vma->vm_pgoff,
3468                 },
3469         };
3470
3471         perf_event_mmap_event(&mmap_event);
3472 }
3473
3474 /*
3475  * IRQ throttle logging
3476  */
3477
3478 static void perf_log_throttle(struct perf_event *event, int enable)
3479 {
3480         struct perf_output_handle handle;
3481         int ret;
3482
3483         struct {
3484                 struct perf_event_header        header;
3485                 u64                             time;
3486                 u64                             id;
3487                 u64                             stream_id;
3488         } throttle_event = {
3489                 .header = {
3490                         .type = PERF_RECORD_THROTTLE,
3491                         .misc = 0,
3492                         .size = sizeof(throttle_event),
3493                 },
3494                 .time           = perf_clock(),
3495                 .id             = primary_event_id(event),
3496                 .stream_id      = event->id,
3497         };
3498
3499         if (enable)
3500                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3501
3502         ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3503         if (ret)
3504                 return;
3505
3506         perf_output_put(&handle, throttle_event);
3507         perf_output_end(&handle);
3508 }
3509
3510 /*
3511  * Generic event overflow handling, sampling.
3512  */
3513
3514 static int __perf_event_overflow(struct perf_event *event, int nmi,
3515                                    int throttle, struct perf_sample_data *data,
3516                                    struct pt_regs *regs)
3517 {
3518         int events = atomic_read(&event->event_limit);
3519         struct hw_perf_event *hwc = &event->hw;
3520         int ret = 0;
3521
3522         throttle = (throttle && event->pmu->unthrottle != NULL);
3523
3524         if (!throttle) {
3525                 hwc->interrupts++;
3526         } else {
3527                 if (hwc->interrupts != MAX_INTERRUPTS) {
3528                         hwc->interrupts++;
3529                         if (HZ * hwc->interrupts >
3530                                         (u64)sysctl_perf_event_sample_rate) {
3531                                 hwc->interrupts = MAX_INTERRUPTS;
3532                                 perf_log_throttle(event, 0);
3533                                 ret = 1;
3534                         }
3535                 } else {
3536                         /*
3537                          * Keep re-disabling events even though on the previous
3538                          * pass we disabled it - just in case we raced with a
3539                          * sched-in and the event got enabled again:
3540                          */
3541                         ret = 1;
3542                 }
3543         }
3544
3545         if (event->attr.freq) {
3546                 u64 now = perf_clock();
3547                 s64 delta = now - hwc->freq_stamp;
3548
3549                 hwc->freq_stamp = now;
3550
3551                 if (delta > 0 && delta < TICK_NSEC)
3552                         perf_adjust_period(event, NSEC_PER_SEC / (int)delta);
3553         }
3554
3555         /*
3556          * XXX event_limit might not quite work as expected on inherited
3557          * events
3558          */
3559
3560         event->pending_kill = POLL_IN;
3561         if (events && atomic_dec_and_test(&event->event_limit)) {
3562                 ret = 1;
3563                 event->pending_kill = POLL_HUP;
3564                 if (nmi) {
3565                         event->pending_disable = 1;
3566                         perf_pending_queue(&event->pending,
3567                                            perf_pending_event);
3568                 } else
3569                         perf_event_disable(event);
3570         }
3571
3572         perf_event_output(event, nmi, data, regs);
3573         return ret;
3574 }
3575
3576 int perf_event_overflow(struct perf_event *event, int nmi,
3577                           struct perf_sample_data *data,
3578                           struct pt_regs *regs)
3579 {
3580         return __perf_event_overflow(event, nmi, 1, data, regs);
3581 }
3582
3583 /*
3584  * Generic software event infrastructure
3585  */
3586
3587 /*
3588  * We directly increment event->count and keep a second value in
3589  * event->hw.period_left to count intervals. This period event
3590  * is kept in the range [-sample_period, 0] so that we can use the
3591  * sign as trigger.
3592  */
3593
3594 static u64 perf_swevent_set_period(struct perf_event *event)
3595 {
3596         struct hw_perf_event *hwc = &event->hw;
3597         u64 period = hwc->last_period;
3598         u64 nr, offset;
3599         s64 old, val;
3600
3601         hwc->last_period = hwc->sample_period;
3602
3603 again:
3604         old = val = atomic64_read(&hwc->period_left);
3605         if (val < 0)
3606                 return 0;
3607
3608         nr = div64_u64(period + val, period);
3609         offset = nr * period;
3610         val -= offset;
3611         if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3612                 goto again;
3613
3614         return nr;
3615 }
3616
3617 static void perf_swevent_overflow(struct perf_event *event,
3618                                     int nmi, struct perf_sample_data *data,
3619                                     struct pt_regs *regs)
3620 {
3621         struct hw_perf_event *hwc = &event->hw;
3622         int throttle = 0;
3623         u64 overflow;
3624
3625         data->period = event->hw.last_period;
3626         overflow = perf_swevent_set_period(event);
3627
3628         if (hwc->interrupts == MAX_INTERRUPTS)
3629                 return;
3630
3631         for (; overflow; overflow--) {
3632                 if (__perf_event_overflow(event, nmi, throttle,
3633                                             data, regs)) {
3634                         /*
3635                          * We inhibit the overflow from happening when
3636                          * hwc->interrupts == MAX_INTERRUPTS.
3637                          */
3638                         break;
3639                 }
3640                 throttle = 1;
3641         }
3642 }
3643
3644 static void perf_swevent_unthrottle(struct perf_event *event)
3645 {
3646         /*
3647          * Nothing to do, we already reset hwc->interrupts.
3648          */
3649 }
3650
3651 static void perf_swevent_add(struct perf_event *event, u64 nr,
3652                                int nmi, struct perf_sample_data *data,
3653                                struct pt_regs *regs)
3654 {
3655         struct hw_perf_event *hwc = &event->hw;
3656
3657         atomic64_add(nr, &event->count);
3658
3659         if (!hwc->sample_period)
3660                 return;
3661
3662         if (!regs)
3663                 return;
3664
3665         if (!atomic64_add_negative(nr, &hwc->period_left))
3666                 perf_swevent_overflow(event, nmi, data, regs);
3667 }
3668
3669 static int perf_swevent_is_counting(struct perf_event *event)
3670 {
3671         /*
3672          * The event is active, we're good!
3673          */
3674         if (event->state == PERF_EVENT_STATE_ACTIVE)
3675                 return 1;
3676
3677         /*
3678          * The event is off/error, not counting.
3679          */
3680         if (event->state != PERF_EVENT_STATE_INACTIVE)
3681                 return 0;
3682
3683         /*
3684          * The event is inactive, if the context is active
3685          * we're part of a group that didn't make it on the 'pmu',
3686          * not counting.
3687          */
3688         if (event->ctx->is_active)
3689                 return 0;
3690
3691         /*
3692          * We're inactive and the context is too, this means the
3693          * task is scheduled out, we're counting events that happen
3694          * to us, like migration events.
3695          */
3696         return 1;
3697 }
3698
3699 static int perf_swevent_match(struct perf_event *event,
3700                                 enum perf_type_id type,
3701                                 u32 event_id, struct pt_regs *regs)
3702 {
3703         if (!perf_swevent_is_counting(event))
3704                 return 0;
3705
3706         if (event->attr.type != type)
3707                 return 0;
3708         if (event->attr.config != event_id)
3709                 return 0;
3710
3711         if (regs) {
3712                 if (event->attr.exclude_user && user_mode(regs))
3713                         return 0;
3714
3715                 if (event->attr.exclude_kernel && !user_mode(regs))
3716                         return 0;
3717         }
3718
3719         return 1;
3720 }
3721
3722 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
3723                                      enum perf_type_id type,
3724                                      u32 event_id, u64 nr, int nmi,
3725                                      struct perf_sample_data *data,
3726                                      struct pt_regs *regs)
3727 {
3728         struct perf_event *event;
3729
3730         if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3731                 return;
3732
3733         rcu_read_lock();
3734         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3735                 if (perf_swevent_match(event, type, event_id, regs))
3736                         perf_swevent_add(event, nr, nmi, data, regs);
3737         }
3738         rcu_read_unlock();
3739 }
3740
3741 static int *perf_swevent_recursion_context(struct perf_cpu_context *cpuctx)
3742 {
3743         if (in_nmi())
3744                 return &cpuctx->recursion[3];
3745
3746         if (in_irq())
3747                 return &cpuctx->recursion[2];
3748
3749         if (in_softirq())
3750                 return &cpuctx->recursion[1];
3751
3752         return &cpuctx->recursion[0];
3753 }
3754
3755 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
3756                                     u64 nr, int nmi,
3757                                     struct perf_sample_data *data,
3758                                     struct pt_regs *regs)
3759 {
3760         struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3761         int *recursion = perf_swevent_recursion_context(cpuctx);
3762         struct perf_event_context *ctx;
3763
3764         if (*recursion)
3765                 goto out;
3766
3767         (*recursion)++;
3768         barrier();
3769
3770         perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
3771                                  nr, nmi, data, regs);
3772         rcu_read_lock();
3773         /*
3774          * doesn't really matter which of the child contexts the
3775          * events ends up in.
3776          */
3777         ctx = rcu_dereference(current->perf_event_ctxp);
3778         if (ctx)
3779                 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
3780         rcu_read_unlock();
3781
3782         barrier();
3783         (*recursion)--;
3784
3785 out:
3786         put_cpu_var(perf_cpu_context);
3787 }
3788
3789 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
3790                             struct pt_regs *regs, u64 addr)
3791 {
3792         struct perf_sample_data data = {
3793                 .addr = addr,
3794         };
3795
3796         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi,
3797                                 &data, regs);
3798 }
3799
3800 static void perf_swevent_read(struct perf_event *event)
3801 {
3802 }
3803
3804 static int perf_swevent_enable(struct perf_event *event)
3805 {
3806         struct hw_perf_event *hwc = &event->hw;
3807
3808         if (hwc->sample_period) {
3809                 hwc->last_period = hwc->sample_period;
3810                 perf_swevent_set_period(event);
3811         }
3812         return 0;
3813 }
3814
3815 static void perf_swevent_disable(struct perf_event *event)
3816 {
3817 }
3818
3819 static const struct pmu perf_ops_generic = {
3820         .enable         = perf_swevent_enable,
3821         .disable        = perf_swevent_disable,
3822         .read           = perf_swevent_read,
3823         .unthrottle     = perf_swevent_unthrottle,
3824 };
3825
3826 /*
3827  * hrtimer based swevent callback
3828  */
3829
3830 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
3831 {
3832         enum hrtimer_restart ret = HRTIMER_RESTART;
3833         struct perf_sample_data data;
3834         struct pt_regs *regs;
3835         struct perf_event *event;
3836         u64 period;
3837
3838         event   = container_of(hrtimer, struct perf_event, hw.hrtimer);
3839         event->pmu->read(event);
3840
3841         data.addr = 0;
3842         regs = get_irq_regs();
3843         /*
3844          * In case we exclude kernel IPs or are somehow not in interrupt
3845          * context, provide the next best thing, the user IP.
3846          */
3847         if ((event->attr.exclude_kernel || !regs) &&
3848                         !event->attr.exclude_user)
3849                 regs = task_pt_regs(current);
3850
3851         if (regs) {
3852                 if (perf_event_overflow(event, 0, &data, regs))
3853                         ret = HRTIMER_NORESTART;
3854         }
3855
3856         period = max_t(u64, 10000, event->hw.sample_period);
3857         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
3858
3859         return ret;
3860 }
3861
3862 /*
3863  * Software event: cpu wall time clock
3864  */
3865
3866 static void cpu_clock_perf_event_update(struct perf_event *event)
3867 {
3868         int cpu = raw_smp_processor_id();
3869         s64 prev;
3870         u64 now;
3871
3872         now = cpu_clock(cpu);
3873         prev = atomic64_read(&event->hw.prev_count);
3874         atomic64_set(&event->hw.prev_count, now);
3875         atomic64_add(now - prev, &event->count);
3876 }
3877
3878 static int cpu_clock_perf_event_enable(struct perf_event *event)
3879 {
3880         struct hw_perf_event *hwc = &event->hw;
3881         int cpu = raw_smp_processor_id();
3882
3883         atomic64_set(&hwc->prev_count, cpu_clock(cpu));
3884         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3885         hwc->hrtimer.function = perf_swevent_hrtimer;
3886         if (hwc->sample_period) {
3887                 u64 period = max_t(u64, 10000, hwc->sample_period);
3888                 __hrtimer_start_range_ns(&hwc->hrtimer,
3889                                 ns_to_ktime(period), 0,
3890                                 HRTIMER_MODE_REL, 0);
3891         }
3892
3893         return 0;
3894 }
3895
3896 static void cpu_clock_perf_event_disable(struct perf_event *event)
3897 {
3898         if (event->hw.sample_period)
3899                 hrtimer_cancel(&event->hw.hrtimer);
3900         cpu_clock_perf_event_update(event);
3901 }
3902
3903 static void cpu_clock_perf_event_read(struct perf_event *event)
3904 {
3905         cpu_clock_perf_event_update(event);
3906 }
3907
3908 static const struct pmu perf_ops_cpu_clock = {
3909         .enable         = cpu_clock_perf_event_enable,
3910         .disable        = cpu_clock_perf_event_disable,
3911         .read           = cpu_clock_perf_event_read,
3912 };
3913
3914 /*
3915  * Software event: task time clock
3916  */
3917
3918 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
3919 {
3920         u64 prev;
3921         s64 delta;
3922
3923         prev = atomic64_xchg(&event->hw.prev_count, now);
3924         delta = now - prev;
3925         atomic64_add(delta, &event->count);
3926 }
3927
3928 static int task_clock_perf_event_enable(struct perf_event *event)
3929 {
3930         struct hw_perf_event *hwc = &event->hw;
3931         u64 now;
3932
3933         now = event->ctx->time;
3934
3935         atomic64_set(&hwc->prev_count, now);
3936         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3937         hwc->hrtimer.function = perf_swevent_hrtimer;
3938         if (hwc->sample_period) {
3939                 u64 period = max_t(u64, 10000, hwc->sample_period);
3940                 __hrtimer_start_range_ns(&hwc->hrtimer,
3941                                 ns_to_ktime(period), 0,
3942                                 HRTIMER_MODE_REL, 0);
3943         }
3944
3945         return 0;
3946 }
3947
3948 static void task_clock_perf_event_disable(struct perf_event *event)
3949 {
3950         if (event->hw.sample_period)
3951                 hrtimer_cancel(&event->hw.hrtimer);
3952         task_clock_perf_event_update(event, event->ctx->time);
3953
3954 }
3955
3956 static void task_clock_perf_event_read(struct perf_event *event)
3957 {
3958         u64 time;
3959
3960         if (!in_nmi()) {
3961                 update_context_time(event->ctx);
3962                 time = event->ctx->time;
3963         } else {
3964                 u64 now = perf_clock();
3965                 u64 delta = now - event->ctx->timestamp;
3966                 time = event->ctx->time + delta;
3967         }
3968
3969         task_clock_perf_event_update(event, time);
3970 }
3971
3972 static const struct pmu perf_ops_task_clock = {
3973         .enable         = task_clock_perf_event_enable,
3974         .disable        = task_clock_perf_event_disable,
3975         .read           = task_clock_perf_event_read,
3976 };
3977
3978 #ifdef CONFIG_EVENT_PROFILE
3979 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
3980                           int entry_size)
3981 {
3982         struct perf_raw_record raw = {
3983                 .size = entry_size,
3984                 .data = record,
3985         };
3986
3987         struct perf_sample_data data = {
3988                 .addr = addr,
3989                 .raw = &raw,
3990         };
3991
3992         struct pt_regs *regs = get_irq_regs();
3993
3994         if (!regs)
3995                 regs = task_pt_regs(current);
3996
3997         do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
3998                                 &data, regs);
3999 }
4000 EXPORT_SYMBOL_GPL(perf_tp_event);
4001
4002 extern int ftrace_profile_enable(int);
4003 extern void ftrace_profile_disable(int);
4004
4005 static void tp_perf_event_destroy(struct perf_event *event)
4006 {
4007         ftrace_profile_disable(event->attr.config);
4008 }
4009
4010 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4011 {
4012         /*
4013          * Raw tracepoint data is a severe data leak, only allow root to
4014          * have these.
4015          */
4016         if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4017                         perf_paranoid_tracepoint_raw() &&
4018                         !capable(CAP_SYS_ADMIN))
4019                 return ERR_PTR(-EPERM);
4020
4021         if (ftrace_profile_enable(event->attr.config))
4022                 return NULL;
4023
4024         event->destroy = tp_perf_event_destroy;
4025
4026         return &perf_ops_generic;
4027 }
4028 #else
4029 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4030 {
4031         return NULL;
4032 }
4033 #endif
4034
4035 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4036
4037 static void sw_perf_event_destroy(struct perf_event *event)
4038 {
4039         u64 event_id = event->attr.config;
4040
4041         WARN_ON(event->parent);
4042
4043         atomic_dec(&perf_swevent_enabled[event_id]);
4044 }
4045
4046 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4047 {
4048         const struct pmu *pmu = NULL;
4049         u64 event_id = event->attr.config;
4050
4051         /*
4052          * Software events (currently) can't in general distinguish
4053          * between user, kernel and hypervisor events.
4054          * However, context switches and cpu migrations are considered
4055          * to be kernel events, and page faults are never hypervisor
4056          * events.
4057          */
4058         switch (event_id) {
4059         case PERF_COUNT_SW_CPU_CLOCK:
4060                 pmu = &perf_ops_cpu_clock;
4061
4062                 break;
4063         case PERF_COUNT_SW_TASK_CLOCK:
4064                 /*
4065                  * If the user instantiates this as a per-cpu event,
4066                  * use the cpu_clock event instead.
4067                  */
4068                 if (event->ctx->task)
4069                         pmu = &perf_ops_task_clock;
4070                 else
4071                         pmu = &perf_ops_cpu_clock;
4072
4073                 break;
4074         case PERF_COUNT_SW_PAGE_FAULTS:
4075         case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4076         case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4077         case PERF_COUNT_SW_CONTEXT_SWITCHES:
4078         case PERF_COUNT_SW_CPU_MIGRATIONS:
4079                 if (!event->parent) {
4080                         atomic_inc(&perf_swevent_enabled[event_id]);
4081                         event->destroy = sw_perf_event_destroy;
4082                 }
4083                 pmu = &perf_ops_generic;
4084                 break;
4085         }
4086
4087         return pmu;
4088 }
4089
4090 /*
4091  * Allocate and initialize a event structure
4092  */
4093 static struct perf_event *
4094 perf_event_alloc(struct perf_event_attr *attr,
4095                    int cpu,
4096                    struct perf_event_context *ctx,
4097                    struct perf_event *group_leader,
4098                    struct perf_event *parent_event,
4099                    gfp_t gfpflags)
4100 {
4101         const struct pmu *pmu;
4102         struct perf_event *event;
4103         struct hw_perf_event *hwc;
4104         long err;
4105
4106         event = kzalloc(sizeof(*event), gfpflags);
4107         if (!event)
4108                 return ERR_PTR(-ENOMEM);
4109
4110         /*
4111          * Single events are their own group leaders, with an
4112          * empty sibling list:
4113          */
4114         if (!group_leader)
4115                 group_leader = event;
4116
4117         mutex_init(&event->child_mutex);
4118         INIT_LIST_HEAD(&event->child_list);
4119
4120         INIT_LIST_HEAD(&event->group_entry);
4121         INIT_LIST_HEAD(&event->event_entry);
4122         INIT_LIST_HEAD(&event->sibling_list);
4123         init_waitqueue_head(&event->waitq);
4124
4125         mutex_init(&event->mmap_mutex);
4126
4127         event->cpu              = cpu;
4128         event->attr             = *attr;
4129         event->group_leader     = group_leader;
4130         event->pmu              = NULL;
4131         event->ctx              = ctx;
4132         event->oncpu            = -1;
4133
4134         event->parent           = parent_event;
4135
4136         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
4137         event->id               = atomic64_inc_return(&perf_event_id);
4138
4139         event->state            = PERF_EVENT_STATE_INACTIVE;
4140
4141         if (attr->disabled)
4142                 event->state = PERF_EVENT_STATE_OFF;
4143
4144         pmu = NULL;
4145
4146         hwc = &event->hw;
4147         hwc->sample_period = attr->sample_period;
4148         if (attr->freq && attr->sample_freq)
4149                 hwc->sample_period = 1;
4150         hwc->last_period = hwc->sample_period;
4151
4152         atomic64_set(&hwc->period_left, hwc->sample_period);
4153
4154         /*
4155          * we currently do not support PERF_FORMAT_GROUP on inherited events
4156          */
4157         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4158                 goto done;
4159
4160         switch (attr->type) {
4161         case PERF_TYPE_RAW:
4162         case PERF_TYPE_HARDWARE:
4163         case PERF_TYPE_HW_CACHE:
4164                 pmu = hw_perf_event_init(event);
4165                 break;
4166
4167         case PERF_TYPE_SOFTWARE:
4168                 pmu = sw_perf_event_init(event);
4169                 break;
4170
4171         case PERF_TYPE_TRACEPOINT:
4172                 pmu = tp_perf_event_init(event);
4173                 break;
4174
4175         default:
4176                 break;
4177         }
4178 done:
4179         err = 0;
4180         if (!pmu)
4181                 err = -EINVAL;
4182         else if (IS_ERR(pmu))
4183                 err = PTR_ERR(pmu);
4184
4185         if (err) {
4186                 if (event->ns)
4187                         put_pid_ns(event->ns);
4188                 kfree(event);
4189                 return ERR_PTR(err);
4190         }
4191
4192         event->pmu = pmu;
4193
4194         if (!event->parent) {
4195                 atomic_inc(&nr_events);
4196                 if (event->attr.mmap)
4197                         atomic_inc(&nr_mmap_events);
4198                 if (event->attr.comm)
4199                         atomic_inc(&nr_comm_events);
4200                 if (event->attr.task)
4201                         atomic_inc(&nr_task_events);
4202         }
4203
4204         return event;
4205 }
4206
4207 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4208                           struct perf_event_attr *attr)
4209 {
4210         u32 size;
4211         int ret;
4212
4213         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4214                 return -EFAULT;
4215
4216         /*
4217          * zero the full structure, so that a short copy will be nice.
4218          */
4219         memset(attr, 0, sizeof(*attr));
4220
4221         ret = get_user(size, &uattr->size);
4222         if (ret)
4223                 return ret;
4224
4225         if (size > PAGE_SIZE)   /* silly large */
4226                 goto err_size;
4227
4228         if (!size)              /* abi compat */
4229                 size = PERF_ATTR_SIZE_VER0;
4230
4231         if (size < PERF_ATTR_SIZE_VER0)
4232                 goto err_size;
4233
4234         /*
4235          * If we're handed a bigger struct than we know of,
4236          * ensure all the unknown bits are 0 - i.e. new
4237          * user-space does not rely on any kernel feature
4238          * extensions we dont know about yet.
4239          */
4240         if (size > sizeof(*attr)) {
4241                 unsigned char __user *addr;
4242                 unsigned char __user *end;
4243                 unsigned char val;
4244
4245                 addr = (void __user *)uattr + sizeof(*attr);
4246                 end  = (void __user *)uattr + size;
4247
4248                 for (; addr < end; addr++) {
4249                         ret = get_user(val, addr);
4250                         if (ret)
4251                                 return ret;
4252                         if (val)
4253                                 goto err_size;
4254                 }
4255                 size = sizeof(*attr);
4256         }
4257
4258         ret = copy_from_user(attr, uattr, size);
4259         if (ret)
4260                 return -EFAULT;
4261
4262         /*
4263          * If the type exists, the corresponding creation will verify
4264          * the attr->config.
4265          */
4266         if (attr->type >= PERF_TYPE_MAX)
4267                 return -EINVAL;
4268
4269         if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3)
4270                 return -EINVAL;
4271
4272         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4273                 return -EINVAL;
4274
4275         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4276                 return -EINVAL;
4277
4278 out:
4279         return ret;
4280
4281 err_size:
4282         put_user(sizeof(*attr), &uattr->size);
4283         ret = -E2BIG;
4284         goto out;
4285 }
4286
4287 int perf_event_set_output(struct perf_event *event, int output_fd)
4288 {
4289         struct perf_event *output_event = NULL;
4290         struct file *output_file = NULL;
4291         struct perf_event *old_output;
4292         int fput_needed = 0;
4293         int ret = -EINVAL;
4294
4295         if (!output_fd)
4296                 goto set;
4297
4298         output_file = fget_light(output_fd, &fput_needed);
4299         if (!output_file)
4300                 return -EBADF;
4301
4302         if (output_file->f_op != &perf_fops)
4303                 goto out;
4304
4305         output_event = output_file->private_data;
4306
4307         /* Don't chain output fds */
4308         if (output_event->output)
4309                 goto out;
4310
4311         /* Don't set an output fd when we already have an output channel */
4312         if (event->data)
4313                 goto out;
4314
4315         atomic_long_inc(&output_file->f_count);
4316
4317 set:
4318         mutex_lock(&event->mmap_mutex);
4319         old_output = event->output;
4320         rcu_assign_pointer(event->output, output_event);
4321         mutex_unlock(&event->mmap_mutex);
4322
4323         if (old_output) {
4324                 /*
4325                  * we need to make sure no existing perf_output_*()
4326                  * is still referencing this event.
4327                  */
4328                 synchronize_rcu();
4329                 fput(old_output->filp);
4330         }
4331
4332         ret = 0;
4333 out:
4334         fput_light(output_file, fput_needed);
4335         return ret;
4336 }
4337
4338 /**
4339  * sys_perf_event_open - open a performance event, associate it to a task/cpu
4340  *
4341  * @attr_uptr:  event_id type attributes for monitoring/sampling
4342  * @pid:                target pid
4343  * @cpu:                target cpu
4344  * @group_fd:           group leader event fd
4345  */
4346 SYSCALL_DEFINE5(perf_event_open,
4347                 struct perf_event_attr __user *, attr_uptr,
4348                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4349 {
4350         struct perf_event *event, *group_leader;
4351         struct perf_event_attr attr;
4352         struct perf_event_context *ctx;
4353         struct file *event_file = NULL;
4354         struct file *group_file = NULL;
4355         int fput_needed = 0;
4356         int fput_needed2 = 0;
4357         int err;
4358
4359         /* for future expandability... */
4360         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4361                 return -EINVAL;
4362
4363         err = perf_copy_attr(attr_uptr, &attr);
4364         if (err)
4365                 return err;
4366
4367         if (!attr.exclude_kernel) {
4368                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4369                         return -EACCES;
4370         }
4371
4372         if (attr.freq) {
4373                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4374                         return -EINVAL;
4375         }
4376
4377         /*
4378          * Get the target context (task or percpu):
4379          */
4380         ctx = find_get_context(pid, cpu);
4381         if (IS_ERR(ctx))
4382                 return PTR_ERR(ctx);
4383
4384         /*
4385          * Look up the group leader (we will attach this event to it):
4386          */
4387         group_leader = NULL;
4388         if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4389                 err = -EINVAL;
4390                 group_file = fget_light(group_fd, &fput_needed);
4391                 if (!group_file)
4392                         goto err_put_context;
4393                 if (group_file->f_op != &perf_fops)
4394                         goto err_put_context;
4395
4396                 group_leader = group_file->private_data;
4397                 /*
4398                  * Do not allow a recursive hierarchy (this new sibling
4399                  * becoming part of another group-sibling):
4400                  */
4401                 if (group_leader->group_leader != group_leader)
4402                         goto err_put_context;
4403                 /*
4404                  * Do not allow to attach to a group in a different
4405                  * task or CPU context:
4406                  */
4407                 if (group_leader->ctx != ctx)
4408                         goto err_put_context;
4409                 /*
4410                  * Only a group leader can be exclusive or pinned
4411                  */
4412                 if (attr.exclusive || attr.pinned)
4413                         goto err_put_context;
4414         }
4415
4416         event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4417                                      NULL, GFP_KERNEL);
4418         err = PTR_ERR(event);
4419         if (IS_ERR(event))
4420                 goto err_put_context;
4421
4422         err = anon_inode_getfd("[perf_event]", &perf_fops, event, 0);
4423         if (err < 0)
4424                 goto err_free_put_context;
4425
4426         event_file = fget_light(err, &fput_needed2);
4427         if (!event_file)
4428                 goto err_free_put_context;
4429
4430         if (flags & PERF_FLAG_FD_OUTPUT) {
4431                 err = perf_event_set_output(event, group_fd);
4432                 if (err)
4433                         goto err_fput_free_put_context;
4434         }
4435
4436         event->filp = event_file;
4437         WARN_ON_ONCE(ctx->parent_ctx);
4438         mutex_lock(&ctx->mutex);
4439         perf_install_in_context(ctx, event, cpu);
4440         ++ctx->generation;
4441         mutex_unlock(&ctx->mutex);
4442
4443         event->owner = current;
4444         get_task_struct(current);
4445         mutex_lock(&current->perf_event_mutex);
4446         list_add_tail(&event->owner_entry, &current->perf_event_list);
4447         mutex_unlock(&current->perf_event_mutex);
4448
4449 err_fput_free_put_context:
4450         fput_light(event_file, fput_needed2);
4451
4452 err_free_put_context:
4453         if (err < 0)
4454                 kfree(event);
4455
4456 err_put_context:
4457         if (err < 0)
4458                 put_ctx(ctx);
4459
4460         fput_light(group_file, fput_needed);
4461
4462         return err;
4463 }
4464
4465 /*
4466  * inherit a event from parent task to child task:
4467  */
4468 static struct perf_event *
4469 inherit_event(struct perf_event *parent_event,
4470               struct task_struct *parent,
4471               struct perf_event_context *parent_ctx,
4472               struct task_struct *child,
4473               struct perf_event *group_leader,
4474               struct perf_event_context *child_ctx)
4475 {
4476         struct perf_event *child_event;
4477
4478         /*
4479          * Instead of creating recursive hierarchies of events,
4480          * we link inherited events back to the original parent,
4481          * which has a filp for sure, which we use as the reference
4482          * count:
4483          */
4484         if (parent_event->parent)
4485                 parent_event = parent_event->parent;
4486
4487         child_event = perf_event_alloc(&parent_event->attr,
4488                                            parent_event->cpu, child_ctx,
4489                                            group_leader, parent_event,
4490                                            GFP_KERNEL);
4491         if (IS_ERR(child_event))
4492                 return child_event;
4493         get_ctx(child_ctx);
4494
4495         /*
4496          * Make the child state follow the state of the parent event,
4497          * not its attr.disabled bit.  We hold the parent's mutex,
4498          * so we won't race with perf_event_{en, dis}able_family.
4499          */
4500         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
4501                 child_event->state = PERF_EVENT_STATE_INACTIVE;
4502         else
4503                 child_event->state = PERF_EVENT_STATE_OFF;
4504
4505         if (parent_event->attr.freq)
4506                 child_event->hw.sample_period = parent_event->hw.sample_period;
4507
4508         /*
4509          * Link it up in the child's context:
4510          */
4511         add_event_to_ctx(child_event, child_ctx);
4512
4513         /*
4514          * Get a reference to the parent filp - we will fput it
4515          * when the child event exits. This is safe to do because
4516          * we are in the parent and we know that the filp still
4517          * exists and has a nonzero count:
4518          */
4519         atomic_long_inc(&parent_event->filp->f_count);
4520
4521         /*
4522          * Link this into the parent event's child list
4523          */
4524         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4525         mutex_lock(&parent_event->child_mutex);
4526         list_add_tail(&child_event->child_list, &parent_event->child_list);
4527         mutex_unlock(&parent_event->child_mutex);
4528
4529         return child_event;
4530 }
4531
4532 static int inherit_group(struct perf_event *parent_event,
4533               struct task_struct *parent,
4534               struct perf_event_context *parent_ctx,
4535               struct task_struct *child,
4536               struct perf_event_context *child_ctx)
4537 {
4538         struct perf_event *leader;
4539         struct perf_event *sub;
4540         struct perf_event *child_ctr;
4541
4542         leader = inherit_event(parent_event, parent, parent_ctx,
4543                                  child, NULL, child_ctx);
4544         if (IS_ERR(leader))
4545                 return PTR_ERR(leader);
4546         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
4547                 child_ctr = inherit_event(sub, parent, parent_ctx,
4548                                             child, leader, child_ctx);
4549                 if (IS_ERR(child_ctr))
4550                         return PTR_ERR(child_ctr);
4551         }
4552         return 0;
4553 }
4554
4555 static void sync_child_event(struct perf_event *child_event,
4556                                struct task_struct *child)
4557 {
4558         struct perf_event *parent_event = child_event->parent;
4559         u64 child_val;
4560
4561         if (child_event->attr.inherit_stat)
4562                 perf_event_read_event(child_event, child);
4563
4564         child_val = atomic64_read(&child_event->count);
4565
4566         /*
4567          * Add back the child's count to the parent's count:
4568          */
4569         atomic64_add(child_val, &parent_event->count);
4570         atomic64_add(child_event->total_time_enabled,
4571                      &parent_event->child_total_time_enabled);
4572         atomic64_add(child_event->total_time_running,
4573                      &parent_event->child_total_time_running);
4574
4575         /*
4576          * Remove this event from the parent's list
4577          */
4578         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4579         mutex_lock(&parent_event->child_mutex);
4580         list_del_init(&child_event->child_list);
4581         mutex_unlock(&parent_event->child_mutex);
4582
4583         /*
4584          * Release the parent event, if this was the last
4585          * reference to it.
4586          */
4587         fput(parent_event->filp);
4588 }
4589
4590 static void
4591 __perf_event_exit_task(struct perf_event *child_event,
4592                          struct perf_event_context *child_ctx,
4593                          struct task_struct *child)
4594 {
4595         struct perf_event *parent_event;
4596
4597         update_event_times(child_event);
4598         perf_event_remove_from_context(child_event);
4599
4600         parent_event = child_event->parent;
4601         /*
4602          * It can happen that parent exits first, and has events
4603          * that are still around due to the child reference. These
4604          * events need to be zapped - but otherwise linger.
4605          */
4606         if (parent_event) {
4607                 sync_child_event(child_event, child);
4608                 free_event(child_event);
4609         }
4610 }
4611
4612 /*
4613  * When a child task exits, feed back event values to parent events.
4614  */
4615 void perf_event_exit_task(struct task_struct *child)
4616 {
4617         struct perf_event *child_event, *tmp;
4618         struct perf_event_context *child_ctx;
4619         unsigned long flags;
4620
4621         if (likely(!child->perf_event_ctxp)) {
4622                 perf_event_task(child, NULL, 0);
4623                 return;
4624         }
4625
4626         local_irq_save(flags);
4627         /*
4628          * We can't reschedule here because interrupts are disabled,
4629          * and either child is current or it is a task that can't be
4630          * scheduled, so we are now safe from rescheduling changing
4631          * our context.
4632          */
4633         child_ctx = child->perf_event_ctxp;
4634         __perf_event_task_sched_out(child_ctx);
4635
4636         /*
4637          * Take the context lock here so that if find_get_context is
4638          * reading child->perf_event_ctxp, we wait until it has
4639          * incremented the context's refcount before we do put_ctx below.
4640          */
4641         spin_lock(&child_ctx->lock);
4642         child->perf_event_ctxp = NULL;
4643         /*
4644          * If this context is a clone; unclone it so it can't get
4645          * swapped to another process while we're removing all
4646          * the events from it.
4647          */
4648         unclone_ctx(child_ctx);
4649         spin_unlock_irqrestore(&child_ctx->lock, flags);
4650
4651         /*
4652          * Report the task dead after unscheduling the events so that we
4653          * won't get any samples after PERF_RECORD_EXIT. We can however still
4654          * get a few PERF_RECORD_READ events.
4655          */
4656         perf_event_task(child, child_ctx, 0);
4657
4658         /*
4659          * We can recurse on the same lock type through:
4660          *
4661          *   __perf_event_exit_task()
4662          *     sync_child_event()
4663          *       fput(parent_event->filp)
4664          *         perf_release()
4665          *           mutex_lock(&ctx->mutex)
4666          *
4667          * But since its the parent context it won't be the same instance.
4668          */
4669         mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
4670
4671 again:
4672         list_for_each_entry_safe(child_event, tmp, &child_ctx->group_list,
4673                                  group_entry)
4674                 __perf_event_exit_task(child_event, child_ctx, child);
4675
4676         /*
4677          * If the last event was a group event, it will have appended all
4678          * its siblings to the list, but we obtained 'tmp' before that which
4679          * will still point to the list head terminating the iteration.
4680          */
4681         if (!list_empty(&child_ctx->group_list))
4682                 goto again;
4683
4684         mutex_unlock(&child_ctx->mutex);
4685
4686         put_ctx(child_ctx);
4687 }
4688
4689 /*
4690  * free an unexposed, unused context as created by inheritance by
4691  * init_task below, used by fork() in case of fail.
4692  */
4693 void perf_event_free_task(struct task_struct *task)
4694 {
4695         struct perf_event_context *ctx = task->perf_event_ctxp;
4696         struct perf_event *event, *tmp;
4697
4698         if (!ctx)
4699                 return;
4700
4701         mutex_lock(&ctx->mutex);
4702 again:
4703         list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry) {
4704                 struct perf_event *parent = event->parent;
4705
4706                 if (WARN_ON_ONCE(!parent))
4707                         continue;
4708
4709                 mutex_lock(&parent->child_mutex);
4710                 list_del_init(&event->child_list);
4711                 mutex_unlock(&parent->child_mutex);
4712
4713                 fput(parent->filp);
4714
4715                 list_del_event(event, ctx);
4716                 free_event(event);
4717         }
4718
4719         if (!list_empty(&ctx->group_list))
4720                 goto again;
4721
4722         mutex_unlock(&ctx->mutex);
4723
4724         put_ctx(ctx);
4725 }
4726
4727 /*
4728  * Initialize the perf_event context in task_struct
4729  */
4730 int perf_event_init_task(struct task_struct *child)
4731 {
4732         struct perf_event_context *child_ctx, *parent_ctx;
4733         struct perf_event_context *cloned_ctx;
4734         struct perf_event *event;
4735         struct task_struct *parent = current;
4736         int inherited_all = 1;
4737         int ret = 0;
4738
4739         child->perf_event_ctxp = NULL;
4740
4741         mutex_init(&child->perf_event_mutex);
4742         INIT_LIST_HEAD(&child->perf_event_list);
4743
4744         if (likely(!parent->perf_event_ctxp))
4745                 return 0;
4746
4747         /*
4748          * This is executed from the parent task context, so inherit
4749          * events that have been marked for cloning.
4750          * First allocate and initialize a context for the child.
4751          */
4752
4753         child_ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
4754         if (!child_ctx)
4755                 return -ENOMEM;
4756
4757         __perf_event_init_context(child_ctx, child);
4758         child->perf_event_ctxp = child_ctx;
4759         get_task_struct(child);
4760
4761         /*
4762          * If the parent's context is a clone, pin it so it won't get
4763          * swapped under us.
4764          */
4765         parent_ctx = perf_pin_task_context(parent);
4766
4767         /*
4768          * No need to check if parent_ctx != NULL here; since we saw
4769          * it non-NULL earlier, the only reason for it to become NULL
4770          * is if we exit, and since we're currently in the middle of
4771          * a fork we can't be exiting at the same time.
4772          */
4773
4774         /*
4775          * Lock the parent list. No need to lock the child - not PID
4776          * hashed yet and not running, so nobody can access it.
4777          */
4778         mutex_lock(&parent_ctx->mutex);
4779
4780         /*
4781          * We dont have to disable NMIs - we are only looking at
4782          * the list, not manipulating it:
4783          */
4784         list_for_each_entry_rcu(event, &parent_ctx->event_list, event_entry) {
4785                 if (event != event->group_leader)
4786                         continue;
4787
4788                 if (!event->attr.inherit) {
4789                         inherited_all = 0;
4790                         continue;
4791                 }
4792
4793                 ret = inherit_group(event, parent, parent_ctx,
4794                                              child, child_ctx);
4795                 if (ret) {
4796                         inherited_all = 0;
4797                         break;
4798                 }
4799         }
4800
4801         if (inherited_all) {
4802                 /*
4803                  * Mark the child context as a clone of the parent
4804                  * context, or of whatever the parent is a clone of.
4805                  * Note that if the parent is a clone, it could get
4806                  * uncloned at any point, but that doesn't matter
4807                  * because the list of events and the generation
4808                  * count can't have changed since we took the mutex.
4809                  */
4810                 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
4811                 if (cloned_ctx) {
4812                         child_ctx->parent_ctx = cloned_ctx;
4813                         child_ctx->parent_gen = parent_ctx->parent_gen;
4814                 } else {
4815                         child_ctx->parent_ctx = parent_ctx;
4816                         child_ctx->parent_gen = parent_ctx->generation;
4817                 }
4818                 get_ctx(child_ctx->parent_ctx);
4819         }
4820
4821         mutex_unlock(&parent_ctx->mutex);
4822
4823         perf_unpin_context(parent_ctx);
4824
4825         return ret;
4826 }
4827
4828 static void __cpuinit perf_event_init_cpu(int cpu)
4829 {
4830         struct perf_cpu_context *cpuctx;
4831
4832         cpuctx = &per_cpu(perf_cpu_context, cpu);
4833         __perf_event_init_context(&cpuctx->ctx, NULL);
4834
4835         spin_lock(&perf_resource_lock);
4836         cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
4837         spin_unlock(&perf_resource_lock);
4838
4839         hw_perf_event_setup(cpu);
4840 }
4841
4842 #ifdef CONFIG_HOTPLUG_CPU
4843 static void __perf_event_exit_cpu(void *info)
4844 {
4845         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4846         struct perf_event_context *ctx = &cpuctx->ctx;
4847         struct perf_event *event, *tmp;
4848
4849         list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry)
4850                 __perf_event_remove_from_context(event);
4851 }
4852 static void perf_event_exit_cpu(int cpu)
4853 {
4854         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4855         struct perf_event_context *ctx = &cpuctx->ctx;
4856
4857         mutex_lock(&ctx->mutex);
4858         smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
4859         mutex_unlock(&ctx->mutex);
4860 }
4861 #else
4862 static inline void perf_event_exit_cpu(int cpu) { }
4863 #endif
4864
4865 static int __cpuinit
4866 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
4867 {
4868         unsigned int cpu = (long)hcpu;
4869
4870         switch (action) {
4871
4872         case CPU_UP_PREPARE:
4873         case CPU_UP_PREPARE_FROZEN:
4874                 perf_event_init_cpu(cpu);
4875                 break;
4876
4877         case CPU_ONLINE:
4878         case CPU_ONLINE_FROZEN:
4879                 hw_perf_event_setup_online(cpu);
4880                 break;
4881
4882         case CPU_DOWN_PREPARE:
4883         case CPU_DOWN_PREPARE_FROZEN:
4884                 perf_event_exit_cpu(cpu);
4885                 break;
4886
4887         default:
4888                 break;
4889         }
4890
4891         return NOTIFY_OK;
4892 }
4893
4894 /*
4895  * This has to have a higher priority than migration_notifier in sched.c.
4896  */
4897 static struct notifier_block __cpuinitdata perf_cpu_nb = {
4898         .notifier_call          = perf_cpu_notify,
4899         .priority               = 20,
4900 };
4901
4902 void __init perf_event_init(void)
4903 {
4904         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
4905                         (void *)(long)smp_processor_id());
4906         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
4907                         (void *)(long)smp_processor_id());
4908         register_cpu_notifier(&perf_cpu_nb);
4909 }
4910
4911 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
4912 {
4913         return sprintf(buf, "%d\n", perf_reserved_percpu);
4914 }
4915
4916 static ssize_t
4917 perf_set_reserve_percpu(struct sysdev_class *class,
4918                         const char *buf,
4919                         size_t count)
4920 {
4921         struct perf_cpu_context *cpuctx;
4922         unsigned long val;
4923         int err, cpu, mpt;
4924
4925         err = strict_strtoul(buf, 10, &val);
4926         if (err)
4927                 return err;
4928         if (val > perf_max_events)
4929                 return -EINVAL;
4930
4931         spin_lock(&perf_resource_lock);
4932         perf_reserved_percpu = val;
4933         for_each_online_cpu(cpu) {
4934                 cpuctx = &per_cpu(perf_cpu_context, cpu);
4935                 spin_lock_irq(&cpuctx->ctx.lock);
4936                 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
4937                           perf_max_events - perf_reserved_percpu);
4938                 cpuctx->max_pertask = mpt;
4939                 spin_unlock_irq(&cpuctx->ctx.lock);
4940         }
4941         spin_unlock(&perf_resource_lock);
4942
4943         return count;
4944 }
4945
4946 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
4947 {
4948         return sprintf(buf, "%d\n", perf_overcommit);
4949 }
4950
4951 static ssize_t
4952 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
4953 {
4954         unsigned long val;
4955         int err;
4956
4957         err = strict_strtoul(buf, 10, &val);
4958         if (err)
4959                 return err;
4960         if (val > 1)
4961                 return -EINVAL;
4962
4963         spin_lock(&perf_resource_lock);
4964         perf_overcommit = val;
4965         spin_unlock(&perf_resource_lock);
4966
4967         return count;
4968 }
4969
4970 static SYSDEV_CLASS_ATTR(
4971                                 reserve_percpu,
4972                                 0644,
4973                                 perf_show_reserve_percpu,
4974                                 perf_set_reserve_percpu
4975                         );
4976
4977 static SYSDEV_CLASS_ATTR(
4978                                 overcommit,
4979                                 0644,
4980                                 perf_show_overcommit,
4981                                 perf_set_overcommit
4982                         );
4983
4984 static struct attribute *perfclass_attrs[] = {
4985         &attr_reserve_percpu.attr,
4986         &attr_overcommit.attr,
4987         NULL
4988 };
4989
4990 static struct attribute_group perfclass_attr_group = {
4991         .attrs                  = perfclass_attrs,
4992         .name                   = "perf_events",
4993 };
4994
4995 static int __init perf_event_sysfs_init(void)
4996 {
4997         return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
4998                                   &perfclass_attr_group);
4999 }
5000 device_initcall(perf_event_sysfs_init);