perf_events: Report the MMAP pgoff value in bytes
[linux-2.6.git] / kernel / perf_event.c
1 /*
2  * Performance events core code:
3  *
4  *  Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5  *  Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6  *  Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7  *  Copyright  ©  2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
8  *
9  * For licensing details see kernel-base/COPYING
10  */
11
12 #include <linux/fs.h>
13 #include <linux/mm.h>
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/file.h>
17 #include <linux/poll.h>
18 #include <linux/sysfs.h>
19 #include <linux/dcache.h>
20 #include <linux/percpu.h>
21 #include <linux/ptrace.h>
22 #include <linux/vmstat.h>
23 #include <linux/vmalloc.h>
24 #include <linux/hardirq.h>
25 #include <linux/rculist.h>
26 #include <linux/uaccess.h>
27 #include <linux/syscalls.h>
28 #include <linux/anon_inodes.h>
29 #include <linux/kernel_stat.h>
30 #include <linux/perf_event.h>
31 #include <linux/ftrace_event.h>
32 #include <linux/hw_breakpoint.h>
33
34 #include <asm/irq_regs.h>
35
36 /*
37  * Each CPU has a list of per CPU events:
38  */
39 static DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
40
41 int perf_max_events __read_mostly = 1;
42 static int perf_reserved_percpu __read_mostly;
43 static int perf_overcommit __read_mostly = 1;
44
45 static atomic_t nr_events __read_mostly;
46 static atomic_t nr_mmap_events __read_mostly;
47 static atomic_t nr_comm_events __read_mostly;
48 static atomic_t nr_task_events __read_mostly;
49
50 /*
51  * perf event paranoia level:
52  *  -1 - not paranoid at all
53  *   0 - disallow raw tracepoint access for unpriv
54  *   1 - disallow cpu events for unpriv
55  *   2 - disallow kernel profiling for unpriv
56  */
57 int sysctl_perf_event_paranoid __read_mostly = 1;
58
59 static inline bool perf_paranoid_tracepoint_raw(void)
60 {
61         return sysctl_perf_event_paranoid > -1;
62 }
63
64 static inline bool perf_paranoid_cpu(void)
65 {
66         return sysctl_perf_event_paranoid > 0;
67 }
68
69 static inline bool perf_paranoid_kernel(void)
70 {
71         return sysctl_perf_event_paranoid > 1;
72 }
73
74 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
75
76 /*
77  * max perf event sample rate
78  */
79 int sysctl_perf_event_sample_rate __read_mostly = 100000;
80
81 static atomic64_t perf_event_id;
82
83 /*
84  * Lock for (sysadmin-configurable) event reservations:
85  */
86 static DEFINE_SPINLOCK(perf_resource_lock);
87
88 /*
89  * Architecture provided APIs - weak aliases:
90  */
91 extern __weak const struct pmu *hw_perf_event_init(struct perf_event *event)
92 {
93         return NULL;
94 }
95
96 void __weak hw_perf_disable(void)               { barrier(); }
97 void __weak hw_perf_enable(void)                { barrier(); }
98
99 void __weak hw_perf_event_setup(int cpu)        { barrier(); }
100 void __weak hw_perf_event_setup_online(int cpu) { barrier(); }
101
102 int __weak
103 hw_perf_group_sched_in(struct perf_event *group_leader,
104                struct perf_cpu_context *cpuctx,
105                struct perf_event_context *ctx, int cpu)
106 {
107         return 0;
108 }
109
110 void __weak perf_event_print_debug(void)        { }
111
112 static DEFINE_PER_CPU(int, perf_disable_count);
113
114 void __perf_disable(void)
115 {
116         __get_cpu_var(perf_disable_count)++;
117 }
118
119 bool __perf_enable(void)
120 {
121         return !--__get_cpu_var(perf_disable_count);
122 }
123
124 void perf_disable(void)
125 {
126         __perf_disable();
127         hw_perf_disable();
128 }
129
130 void perf_enable(void)
131 {
132         if (__perf_enable())
133                 hw_perf_enable();
134 }
135
136 static void get_ctx(struct perf_event_context *ctx)
137 {
138         WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
139 }
140
141 static void free_ctx(struct rcu_head *head)
142 {
143         struct perf_event_context *ctx;
144
145         ctx = container_of(head, struct perf_event_context, rcu_head);
146         kfree(ctx);
147 }
148
149 static void put_ctx(struct perf_event_context *ctx)
150 {
151         if (atomic_dec_and_test(&ctx->refcount)) {
152                 if (ctx->parent_ctx)
153                         put_ctx(ctx->parent_ctx);
154                 if (ctx->task)
155                         put_task_struct(ctx->task);
156                 call_rcu(&ctx->rcu_head, free_ctx);
157         }
158 }
159
160 static void unclone_ctx(struct perf_event_context *ctx)
161 {
162         if (ctx->parent_ctx) {
163                 put_ctx(ctx->parent_ctx);
164                 ctx->parent_ctx = NULL;
165         }
166 }
167
168 /*
169  * If we inherit events we want to return the parent event id
170  * to userspace.
171  */
172 static u64 primary_event_id(struct perf_event *event)
173 {
174         u64 id = event->id;
175
176         if (event->parent)
177                 id = event->parent->id;
178
179         return id;
180 }
181
182 /*
183  * Get the perf_event_context for a task and lock it.
184  * This has to cope with with the fact that until it is locked,
185  * the context could get moved to another task.
186  */
187 static struct perf_event_context *
188 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
189 {
190         struct perf_event_context *ctx;
191
192         rcu_read_lock();
193  retry:
194         ctx = rcu_dereference(task->perf_event_ctxp);
195         if (ctx) {
196                 /*
197                  * If this context is a clone of another, it might
198                  * get swapped for another underneath us by
199                  * perf_event_task_sched_out, though the
200                  * rcu_read_lock() protects us from any context
201                  * getting freed.  Lock the context and check if it
202                  * got swapped before we could get the lock, and retry
203                  * if so.  If we locked the right context, then it
204                  * can't get swapped on us any more.
205                  */
206                 raw_spin_lock_irqsave(&ctx->lock, *flags);
207                 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
208                         raw_spin_unlock_irqrestore(&ctx->lock, *flags);
209                         goto retry;
210                 }
211
212                 if (!atomic_inc_not_zero(&ctx->refcount)) {
213                         raw_spin_unlock_irqrestore(&ctx->lock, *flags);
214                         ctx = NULL;
215                 }
216         }
217         rcu_read_unlock();
218         return ctx;
219 }
220
221 /*
222  * Get the context for a task and increment its pin_count so it
223  * can't get swapped to another task.  This also increments its
224  * reference count so that the context can't get freed.
225  */
226 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
227 {
228         struct perf_event_context *ctx;
229         unsigned long flags;
230
231         ctx = perf_lock_task_context(task, &flags);
232         if (ctx) {
233                 ++ctx->pin_count;
234                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
235         }
236         return ctx;
237 }
238
239 static void perf_unpin_context(struct perf_event_context *ctx)
240 {
241         unsigned long flags;
242
243         raw_spin_lock_irqsave(&ctx->lock, flags);
244         --ctx->pin_count;
245         raw_spin_unlock_irqrestore(&ctx->lock, flags);
246         put_ctx(ctx);
247 }
248
249 static inline u64 perf_clock(void)
250 {
251         return cpu_clock(smp_processor_id());
252 }
253
254 /*
255  * Update the record of the current time in a context.
256  */
257 static void update_context_time(struct perf_event_context *ctx)
258 {
259         u64 now = perf_clock();
260
261         ctx->time += now - ctx->timestamp;
262         ctx->timestamp = now;
263 }
264
265 /*
266  * Update the total_time_enabled and total_time_running fields for a event.
267  */
268 static void update_event_times(struct perf_event *event)
269 {
270         struct perf_event_context *ctx = event->ctx;
271         u64 run_end;
272
273         if (event->state < PERF_EVENT_STATE_INACTIVE ||
274             event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
275                 return;
276
277         if (ctx->is_active)
278                 run_end = ctx->time;
279         else
280                 run_end = event->tstamp_stopped;
281
282         event->total_time_enabled = run_end - event->tstamp_enabled;
283
284         if (event->state == PERF_EVENT_STATE_INACTIVE)
285                 run_end = event->tstamp_stopped;
286         else
287                 run_end = ctx->time;
288
289         event->total_time_running = run_end - event->tstamp_running;
290 }
291
292 static struct list_head *
293 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
294 {
295         if (event->attr.pinned)
296                 return &ctx->pinned_groups;
297         else
298                 return &ctx->flexible_groups;
299 }
300
301 /*
302  * Add a event from the lists for its context.
303  * Must be called with ctx->mutex and ctx->lock held.
304  */
305 static void
306 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
307 {
308         struct perf_event *group_leader = event->group_leader;
309
310         /*
311          * Depending on whether it is a standalone or sibling event,
312          * add it straight to the context's event list, or to the group
313          * leader's sibling list:
314          */
315         if (group_leader == event) {
316                 struct list_head *list;
317
318                 if (is_software_event(event))
319                         event->group_flags |= PERF_GROUP_SOFTWARE;
320
321                 list = ctx_group_list(event, ctx);
322                 list_add_tail(&event->group_entry, list);
323         } else {
324                 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
325                     !is_software_event(event))
326                         group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
327
328                 list_add_tail(&event->group_entry, &group_leader->sibling_list);
329                 group_leader->nr_siblings++;
330         }
331
332         list_add_rcu(&event->event_entry, &ctx->event_list);
333         ctx->nr_events++;
334         if (event->attr.inherit_stat)
335                 ctx->nr_stat++;
336 }
337
338 /*
339  * Remove a event from the lists for its context.
340  * Must be called with ctx->mutex and ctx->lock held.
341  */
342 static void
343 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
344 {
345         struct perf_event *sibling, *tmp;
346
347         if (list_empty(&event->group_entry))
348                 return;
349         ctx->nr_events--;
350         if (event->attr.inherit_stat)
351                 ctx->nr_stat--;
352
353         list_del_init(&event->group_entry);
354         list_del_rcu(&event->event_entry);
355
356         if (event->group_leader != event)
357                 event->group_leader->nr_siblings--;
358
359         update_event_times(event);
360
361         /*
362          * If event was in error state, then keep it
363          * that way, otherwise bogus counts will be
364          * returned on read(). The only way to get out
365          * of error state is by explicit re-enabling
366          * of the event
367          */
368         if (event->state > PERF_EVENT_STATE_OFF)
369                 event->state = PERF_EVENT_STATE_OFF;
370
371         /*
372          * If this was a group event with sibling events then
373          * upgrade the siblings to singleton events by adding them
374          * to the context list directly:
375          */
376         list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
377                 struct list_head *list;
378
379                 list = ctx_group_list(event, ctx);
380                 list_move_tail(&sibling->group_entry, list);
381                 sibling->group_leader = sibling;
382
383                 /* Inherit group flags from the previous leader */
384                 sibling->group_flags = event->group_flags;
385         }
386 }
387
388 static void
389 event_sched_out(struct perf_event *event,
390                   struct perf_cpu_context *cpuctx,
391                   struct perf_event_context *ctx)
392 {
393         if (event->state != PERF_EVENT_STATE_ACTIVE)
394                 return;
395
396         event->state = PERF_EVENT_STATE_INACTIVE;
397         if (event->pending_disable) {
398                 event->pending_disable = 0;
399                 event->state = PERF_EVENT_STATE_OFF;
400         }
401         event->tstamp_stopped = ctx->time;
402         event->pmu->disable(event);
403         event->oncpu = -1;
404
405         if (!is_software_event(event))
406                 cpuctx->active_oncpu--;
407         ctx->nr_active--;
408         if (event->attr.exclusive || !cpuctx->active_oncpu)
409                 cpuctx->exclusive = 0;
410 }
411
412 static void
413 group_sched_out(struct perf_event *group_event,
414                 struct perf_cpu_context *cpuctx,
415                 struct perf_event_context *ctx)
416 {
417         struct perf_event *event;
418
419         if (group_event->state != PERF_EVENT_STATE_ACTIVE)
420                 return;
421
422         event_sched_out(group_event, cpuctx, ctx);
423
424         /*
425          * Schedule out siblings (if any):
426          */
427         list_for_each_entry(event, &group_event->sibling_list, group_entry)
428                 event_sched_out(event, cpuctx, ctx);
429
430         if (group_event->attr.exclusive)
431                 cpuctx->exclusive = 0;
432 }
433
434 /*
435  * Cross CPU call to remove a performance event
436  *
437  * We disable the event on the hardware level first. After that we
438  * remove it from the context list.
439  */
440 static void __perf_event_remove_from_context(void *info)
441 {
442         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
443         struct perf_event *event = info;
444         struct perf_event_context *ctx = event->ctx;
445
446         /*
447          * If this is a task context, we need to check whether it is
448          * the current task context of this cpu. If not it has been
449          * scheduled out before the smp call arrived.
450          */
451         if (ctx->task && cpuctx->task_ctx != ctx)
452                 return;
453
454         raw_spin_lock(&ctx->lock);
455         /*
456          * Protect the list operation against NMI by disabling the
457          * events on a global level.
458          */
459         perf_disable();
460
461         event_sched_out(event, cpuctx, ctx);
462
463         list_del_event(event, ctx);
464
465         if (!ctx->task) {
466                 /*
467                  * Allow more per task events with respect to the
468                  * reservation:
469                  */
470                 cpuctx->max_pertask =
471                         min(perf_max_events - ctx->nr_events,
472                             perf_max_events - perf_reserved_percpu);
473         }
474
475         perf_enable();
476         raw_spin_unlock(&ctx->lock);
477 }
478
479
480 /*
481  * Remove the event from a task's (or a CPU's) list of events.
482  *
483  * Must be called with ctx->mutex held.
484  *
485  * CPU events are removed with a smp call. For task events we only
486  * call when the task is on a CPU.
487  *
488  * If event->ctx is a cloned context, callers must make sure that
489  * every task struct that event->ctx->task could possibly point to
490  * remains valid.  This is OK when called from perf_release since
491  * that only calls us on the top-level context, which can't be a clone.
492  * When called from perf_event_exit_task, it's OK because the
493  * context has been detached from its task.
494  */
495 static void perf_event_remove_from_context(struct perf_event *event)
496 {
497         struct perf_event_context *ctx = event->ctx;
498         struct task_struct *task = ctx->task;
499
500         if (!task) {
501                 /*
502                  * Per cpu events are removed via an smp call and
503                  * the removal is always successful.
504                  */
505                 smp_call_function_single(event->cpu,
506                                          __perf_event_remove_from_context,
507                                          event, 1);
508                 return;
509         }
510
511 retry:
512         task_oncpu_function_call(task, __perf_event_remove_from_context,
513                                  event);
514
515         raw_spin_lock_irq(&ctx->lock);
516         /*
517          * If the context is active we need to retry the smp call.
518          */
519         if (ctx->nr_active && !list_empty(&event->group_entry)) {
520                 raw_spin_unlock_irq(&ctx->lock);
521                 goto retry;
522         }
523
524         /*
525          * The lock prevents that this context is scheduled in so we
526          * can remove the event safely, if the call above did not
527          * succeed.
528          */
529         if (!list_empty(&event->group_entry))
530                 list_del_event(event, ctx);
531         raw_spin_unlock_irq(&ctx->lock);
532 }
533
534 /*
535  * Update total_time_enabled and total_time_running for all events in a group.
536  */
537 static void update_group_times(struct perf_event *leader)
538 {
539         struct perf_event *event;
540
541         update_event_times(leader);
542         list_for_each_entry(event, &leader->sibling_list, group_entry)
543                 update_event_times(event);
544 }
545
546 /*
547  * Cross CPU call to disable a performance event
548  */
549 static void __perf_event_disable(void *info)
550 {
551         struct perf_event *event = info;
552         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
553         struct perf_event_context *ctx = event->ctx;
554
555         /*
556          * If this is a per-task event, need to check whether this
557          * event's task is the current task on this cpu.
558          */
559         if (ctx->task && cpuctx->task_ctx != ctx)
560                 return;
561
562         raw_spin_lock(&ctx->lock);
563
564         /*
565          * If the event is on, turn it off.
566          * If it is in error state, leave it in error state.
567          */
568         if (event->state >= PERF_EVENT_STATE_INACTIVE) {
569                 update_context_time(ctx);
570                 update_group_times(event);
571                 if (event == event->group_leader)
572                         group_sched_out(event, cpuctx, ctx);
573                 else
574                         event_sched_out(event, cpuctx, ctx);
575                 event->state = PERF_EVENT_STATE_OFF;
576         }
577
578         raw_spin_unlock(&ctx->lock);
579 }
580
581 /*
582  * Disable a event.
583  *
584  * If event->ctx is a cloned context, callers must make sure that
585  * every task struct that event->ctx->task could possibly point to
586  * remains valid.  This condition is satisifed when called through
587  * perf_event_for_each_child or perf_event_for_each because they
588  * hold the top-level event's child_mutex, so any descendant that
589  * goes to exit will block in sync_child_event.
590  * When called from perf_pending_event it's OK because event->ctx
591  * is the current context on this CPU and preemption is disabled,
592  * hence we can't get into perf_event_task_sched_out for this context.
593  */
594 void perf_event_disable(struct perf_event *event)
595 {
596         struct perf_event_context *ctx = event->ctx;
597         struct task_struct *task = ctx->task;
598
599         if (!task) {
600                 /*
601                  * Disable the event on the cpu that it's on
602                  */
603                 smp_call_function_single(event->cpu, __perf_event_disable,
604                                          event, 1);
605                 return;
606         }
607
608  retry:
609         task_oncpu_function_call(task, __perf_event_disable, event);
610
611         raw_spin_lock_irq(&ctx->lock);
612         /*
613          * If the event is still active, we need to retry the cross-call.
614          */
615         if (event->state == PERF_EVENT_STATE_ACTIVE) {
616                 raw_spin_unlock_irq(&ctx->lock);
617                 goto retry;
618         }
619
620         /*
621          * Since we have the lock this context can't be scheduled
622          * in, so we can change the state safely.
623          */
624         if (event->state == PERF_EVENT_STATE_INACTIVE) {
625                 update_group_times(event);
626                 event->state = PERF_EVENT_STATE_OFF;
627         }
628
629         raw_spin_unlock_irq(&ctx->lock);
630 }
631
632 static int
633 event_sched_in(struct perf_event *event,
634                  struct perf_cpu_context *cpuctx,
635                  struct perf_event_context *ctx,
636                  int cpu)
637 {
638         if (event->state <= PERF_EVENT_STATE_OFF)
639                 return 0;
640
641         event->state = PERF_EVENT_STATE_ACTIVE;
642         event->oncpu = cpu;     /* TODO: put 'cpu' into cpuctx->cpu */
643         /*
644          * The new state must be visible before we turn it on in the hardware:
645          */
646         smp_wmb();
647
648         if (event->pmu->enable(event)) {
649                 event->state = PERF_EVENT_STATE_INACTIVE;
650                 event->oncpu = -1;
651                 return -EAGAIN;
652         }
653
654         event->tstamp_running += ctx->time - event->tstamp_stopped;
655
656         if (!is_software_event(event))
657                 cpuctx->active_oncpu++;
658         ctx->nr_active++;
659
660         if (event->attr.exclusive)
661                 cpuctx->exclusive = 1;
662
663         return 0;
664 }
665
666 static int
667 group_sched_in(struct perf_event *group_event,
668                struct perf_cpu_context *cpuctx,
669                struct perf_event_context *ctx,
670                int cpu)
671 {
672         struct perf_event *event, *partial_group;
673         int ret;
674
675         if (group_event->state == PERF_EVENT_STATE_OFF)
676                 return 0;
677
678         ret = hw_perf_group_sched_in(group_event, cpuctx, ctx, cpu);
679         if (ret)
680                 return ret < 0 ? ret : 0;
681
682         if (event_sched_in(group_event, cpuctx, ctx, cpu))
683                 return -EAGAIN;
684
685         /*
686          * Schedule in siblings as one group (if any):
687          */
688         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
689                 if (event_sched_in(event, cpuctx, ctx, cpu)) {
690                         partial_group = event;
691                         goto group_error;
692                 }
693         }
694
695         return 0;
696
697 group_error:
698         /*
699          * Groups can be scheduled in as one unit only, so undo any
700          * partial group before returning:
701          */
702         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
703                 if (event == partial_group)
704                         break;
705                 event_sched_out(event, cpuctx, ctx);
706         }
707         event_sched_out(group_event, cpuctx, ctx);
708
709         return -EAGAIN;
710 }
711
712 /*
713  * Work out whether we can put this event group on the CPU now.
714  */
715 static int group_can_go_on(struct perf_event *event,
716                            struct perf_cpu_context *cpuctx,
717                            int can_add_hw)
718 {
719         /*
720          * Groups consisting entirely of software events can always go on.
721          */
722         if (event->group_flags & PERF_GROUP_SOFTWARE)
723                 return 1;
724         /*
725          * If an exclusive group is already on, no other hardware
726          * events can go on.
727          */
728         if (cpuctx->exclusive)
729                 return 0;
730         /*
731          * If this group is exclusive and there are already
732          * events on the CPU, it can't go on.
733          */
734         if (event->attr.exclusive && cpuctx->active_oncpu)
735                 return 0;
736         /*
737          * Otherwise, try to add it if all previous groups were able
738          * to go on.
739          */
740         return can_add_hw;
741 }
742
743 static void add_event_to_ctx(struct perf_event *event,
744                                struct perf_event_context *ctx)
745 {
746         list_add_event(event, ctx);
747         event->tstamp_enabled = ctx->time;
748         event->tstamp_running = ctx->time;
749         event->tstamp_stopped = ctx->time;
750 }
751
752 /*
753  * Cross CPU call to install and enable a performance event
754  *
755  * Must be called with ctx->mutex held
756  */
757 static void __perf_install_in_context(void *info)
758 {
759         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
760         struct perf_event *event = info;
761         struct perf_event_context *ctx = event->ctx;
762         struct perf_event *leader = event->group_leader;
763         int cpu = smp_processor_id();
764         int err;
765
766         /*
767          * If this is a task context, we need to check whether it is
768          * the current task context of this cpu. If not it has been
769          * scheduled out before the smp call arrived.
770          * Or possibly this is the right context but it isn't
771          * on this cpu because it had no events.
772          */
773         if (ctx->task && cpuctx->task_ctx != ctx) {
774                 if (cpuctx->task_ctx || ctx->task != current)
775                         return;
776                 cpuctx->task_ctx = ctx;
777         }
778
779         raw_spin_lock(&ctx->lock);
780         ctx->is_active = 1;
781         update_context_time(ctx);
782
783         /*
784          * Protect the list operation against NMI by disabling the
785          * events on a global level. NOP for non NMI based events.
786          */
787         perf_disable();
788
789         add_event_to_ctx(event, ctx);
790
791         if (event->cpu != -1 && event->cpu != smp_processor_id())
792                 goto unlock;
793
794         /*
795          * Don't put the event on if it is disabled or if
796          * it is in a group and the group isn't on.
797          */
798         if (event->state != PERF_EVENT_STATE_INACTIVE ||
799             (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
800                 goto unlock;
801
802         /*
803          * An exclusive event can't go on if there are already active
804          * hardware events, and no hardware event can go on if there
805          * is already an exclusive event on.
806          */
807         if (!group_can_go_on(event, cpuctx, 1))
808                 err = -EEXIST;
809         else
810                 err = event_sched_in(event, cpuctx, ctx, cpu);
811
812         if (err) {
813                 /*
814                  * This event couldn't go on.  If it is in a group
815                  * then we have to pull the whole group off.
816                  * If the event group is pinned then put it in error state.
817                  */
818                 if (leader != event)
819                         group_sched_out(leader, cpuctx, ctx);
820                 if (leader->attr.pinned) {
821                         update_group_times(leader);
822                         leader->state = PERF_EVENT_STATE_ERROR;
823                 }
824         }
825
826         if (!err && !ctx->task && cpuctx->max_pertask)
827                 cpuctx->max_pertask--;
828
829  unlock:
830         perf_enable();
831
832         raw_spin_unlock(&ctx->lock);
833 }
834
835 /*
836  * Attach a performance event to a context
837  *
838  * First we add the event to the list with the hardware enable bit
839  * in event->hw_config cleared.
840  *
841  * If the event is attached to a task which is on a CPU we use a smp
842  * call to enable it in the task context. The task might have been
843  * scheduled away, but we check this in the smp call again.
844  *
845  * Must be called with ctx->mutex held.
846  */
847 static void
848 perf_install_in_context(struct perf_event_context *ctx,
849                         struct perf_event *event,
850                         int cpu)
851 {
852         struct task_struct *task = ctx->task;
853
854         if (!task) {
855                 /*
856                  * Per cpu events are installed via an smp call and
857                  * the install is always successful.
858                  */
859                 smp_call_function_single(cpu, __perf_install_in_context,
860                                          event, 1);
861                 return;
862         }
863
864 retry:
865         task_oncpu_function_call(task, __perf_install_in_context,
866                                  event);
867
868         raw_spin_lock_irq(&ctx->lock);
869         /*
870          * we need to retry the smp call.
871          */
872         if (ctx->is_active && list_empty(&event->group_entry)) {
873                 raw_spin_unlock_irq(&ctx->lock);
874                 goto retry;
875         }
876
877         /*
878          * The lock prevents that this context is scheduled in so we
879          * can add the event safely, if it the call above did not
880          * succeed.
881          */
882         if (list_empty(&event->group_entry))
883                 add_event_to_ctx(event, ctx);
884         raw_spin_unlock_irq(&ctx->lock);
885 }
886
887 /*
888  * Put a event into inactive state and update time fields.
889  * Enabling the leader of a group effectively enables all
890  * the group members that aren't explicitly disabled, so we
891  * have to update their ->tstamp_enabled also.
892  * Note: this works for group members as well as group leaders
893  * since the non-leader members' sibling_lists will be empty.
894  */
895 static void __perf_event_mark_enabled(struct perf_event *event,
896                                         struct perf_event_context *ctx)
897 {
898         struct perf_event *sub;
899
900         event->state = PERF_EVENT_STATE_INACTIVE;
901         event->tstamp_enabled = ctx->time - event->total_time_enabled;
902         list_for_each_entry(sub, &event->sibling_list, group_entry)
903                 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
904                         sub->tstamp_enabled =
905                                 ctx->time - sub->total_time_enabled;
906 }
907
908 /*
909  * Cross CPU call to enable a performance event
910  */
911 static void __perf_event_enable(void *info)
912 {
913         struct perf_event *event = info;
914         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
915         struct perf_event_context *ctx = event->ctx;
916         struct perf_event *leader = event->group_leader;
917         int err;
918
919         /*
920          * If this is a per-task event, need to check whether this
921          * event's task is the current task on this cpu.
922          */
923         if (ctx->task && cpuctx->task_ctx != ctx) {
924                 if (cpuctx->task_ctx || ctx->task != current)
925                         return;
926                 cpuctx->task_ctx = ctx;
927         }
928
929         raw_spin_lock(&ctx->lock);
930         ctx->is_active = 1;
931         update_context_time(ctx);
932
933         if (event->state >= PERF_EVENT_STATE_INACTIVE)
934                 goto unlock;
935         __perf_event_mark_enabled(event, ctx);
936
937         if (event->cpu != -1 && event->cpu != smp_processor_id())
938                 goto unlock;
939
940         /*
941          * If the event is in a group and isn't the group leader,
942          * then don't put it on unless the group is on.
943          */
944         if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
945                 goto unlock;
946
947         if (!group_can_go_on(event, cpuctx, 1)) {
948                 err = -EEXIST;
949         } else {
950                 perf_disable();
951                 if (event == leader)
952                         err = group_sched_in(event, cpuctx, ctx,
953                                              smp_processor_id());
954                 else
955                         err = event_sched_in(event, cpuctx, ctx,
956                                                smp_processor_id());
957                 perf_enable();
958         }
959
960         if (err) {
961                 /*
962                  * If this event can't go on and it's part of a
963                  * group, then the whole group has to come off.
964                  */
965                 if (leader != event)
966                         group_sched_out(leader, cpuctx, ctx);
967                 if (leader->attr.pinned) {
968                         update_group_times(leader);
969                         leader->state = PERF_EVENT_STATE_ERROR;
970                 }
971         }
972
973  unlock:
974         raw_spin_unlock(&ctx->lock);
975 }
976
977 /*
978  * Enable a event.
979  *
980  * If event->ctx is a cloned context, callers must make sure that
981  * every task struct that event->ctx->task could possibly point to
982  * remains valid.  This condition is satisfied when called through
983  * perf_event_for_each_child or perf_event_for_each as described
984  * for perf_event_disable.
985  */
986 void perf_event_enable(struct perf_event *event)
987 {
988         struct perf_event_context *ctx = event->ctx;
989         struct task_struct *task = ctx->task;
990
991         if (!task) {
992                 /*
993                  * Enable the event on the cpu that it's on
994                  */
995                 smp_call_function_single(event->cpu, __perf_event_enable,
996                                          event, 1);
997                 return;
998         }
999
1000         raw_spin_lock_irq(&ctx->lock);
1001         if (event->state >= PERF_EVENT_STATE_INACTIVE)
1002                 goto out;
1003
1004         /*
1005          * If the event is in error state, clear that first.
1006          * That way, if we see the event in error state below, we
1007          * know that it has gone back into error state, as distinct
1008          * from the task having been scheduled away before the
1009          * cross-call arrived.
1010          */
1011         if (event->state == PERF_EVENT_STATE_ERROR)
1012                 event->state = PERF_EVENT_STATE_OFF;
1013
1014  retry:
1015         raw_spin_unlock_irq(&ctx->lock);
1016         task_oncpu_function_call(task, __perf_event_enable, event);
1017
1018         raw_spin_lock_irq(&ctx->lock);
1019
1020         /*
1021          * If the context is active and the event is still off,
1022          * we need to retry the cross-call.
1023          */
1024         if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1025                 goto retry;
1026
1027         /*
1028          * Since we have the lock this context can't be scheduled
1029          * in, so we can change the state safely.
1030          */
1031         if (event->state == PERF_EVENT_STATE_OFF)
1032                 __perf_event_mark_enabled(event, ctx);
1033
1034  out:
1035         raw_spin_unlock_irq(&ctx->lock);
1036 }
1037
1038 static int perf_event_refresh(struct perf_event *event, int refresh)
1039 {
1040         /*
1041          * not supported on inherited events
1042          */
1043         if (event->attr.inherit)
1044                 return -EINVAL;
1045
1046         atomic_add(refresh, &event->event_limit);
1047         perf_event_enable(event);
1048
1049         return 0;
1050 }
1051
1052 enum event_type_t {
1053         EVENT_FLEXIBLE = 0x1,
1054         EVENT_PINNED = 0x2,
1055         EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1056 };
1057
1058 static void ctx_sched_out(struct perf_event_context *ctx,
1059                           struct perf_cpu_context *cpuctx,
1060                           enum event_type_t event_type)
1061 {
1062         struct perf_event *event;
1063
1064         raw_spin_lock(&ctx->lock);
1065         ctx->is_active = 0;
1066         if (likely(!ctx->nr_events))
1067                 goto out;
1068         update_context_time(ctx);
1069
1070         perf_disable();
1071         if (!ctx->nr_active)
1072                 goto out_enable;
1073
1074         if (event_type & EVENT_PINNED)
1075                 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1076                         group_sched_out(event, cpuctx, ctx);
1077
1078         if (event_type & EVENT_FLEXIBLE)
1079                 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1080                         group_sched_out(event, cpuctx, ctx);
1081
1082  out_enable:
1083         perf_enable();
1084  out:
1085         raw_spin_unlock(&ctx->lock);
1086 }
1087
1088 /*
1089  * Test whether two contexts are equivalent, i.e. whether they
1090  * have both been cloned from the same version of the same context
1091  * and they both have the same number of enabled events.
1092  * If the number of enabled events is the same, then the set
1093  * of enabled events should be the same, because these are both
1094  * inherited contexts, therefore we can't access individual events
1095  * in them directly with an fd; we can only enable/disable all
1096  * events via prctl, or enable/disable all events in a family
1097  * via ioctl, which will have the same effect on both contexts.
1098  */
1099 static int context_equiv(struct perf_event_context *ctx1,
1100                          struct perf_event_context *ctx2)
1101 {
1102         return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1103                 && ctx1->parent_gen == ctx2->parent_gen
1104                 && !ctx1->pin_count && !ctx2->pin_count;
1105 }
1106
1107 static void __perf_event_sync_stat(struct perf_event *event,
1108                                      struct perf_event *next_event)
1109 {
1110         u64 value;
1111
1112         if (!event->attr.inherit_stat)
1113                 return;
1114
1115         /*
1116          * Update the event value, we cannot use perf_event_read()
1117          * because we're in the middle of a context switch and have IRQs
1118          * disabled, which upsets smp_call_function_single(), however
1119          * we know the event must be on the current CPU, therefore we
1120          * don't need to use it.
1121          */
1122         switch (event->state) {
1123         case PERF_EVENT_STATE_ACTIVE:
1124                 event->pmu->read(event);
1125                 /* fall-through */
1126
1127         case PERF_EVENT_STATE_INACTIVE:
1128                 update_event_times(event);
1129                 break;
1130
1131         default:
1132                 break;
1133         }
1134
1135         /*
1136          * In order to keep per-task stats reliable we need to flip the event
1137          * values when we flip the contexts.
1138          */
1139         value = atomic64_read(&next_event->count);
1140         value = atomic64_xchg(&event->count, value);
1141         atomic64_set(&next_event->count, value);
1142
1143         swap(event->total_time_enabled, next_event->total_time_enabled);
1144         swap(event->total_time_running, next_event->total_time_running);
1145
1146         /*
1147          * Since we swizzled the values, update the user visible data too.
1148          */
1149         perf_event_update_userpage(event);
1150         perf_event_update_userpage(next_event);
1151 }
1152
1153 #define list_next_entry(pos, member) \
1154         list_entry(pos->member.next, typeof(*pos), member)
1155
1156 static void perf_event_sync_stat(struct perf_event_context *ctx,
1157                                    struct perf_event_context *next_ctx)
1158 {
1159         struct perf_event *event, *next_event;
1160
1161         if (!ctx->nr_stat)
1162                 return;
1163
1164         update_context_time(ctx);
1165
1166         event = list_first_entry(&ctx->event_list,
1167                                    struct perf_event, event_entry);
1168
1169         next_event = list_first_entry(&next_ctx->event_list,
1170                                         struct perf_event, event_entry);
1171
1172         while (&event->event_entry != &ctx->event_list &&
1173                &next_event->event_entry != &next_ctx->event_list) {
1174
1175                 __perf_event_sync_stat(event, next_event);
1176
1177                 event = list_next_entry(event, event_entry);
1178                 next_event = list_next_entry(next_event, event_entry);
1179         }
1180 }
1181
1182 /*
1183  * Called from scheduler to remove the events of the current task,
1184  * with interrupts disabled.
1185  *
1186  * We stop each event and update the event value in event->count.
1187  *
1188  * This does not protect us against NMI, but disable()
1189  * sets the disabled bit in the control field of event _before_
1190  * accessing the event control register. If a NMI hits, then it will
1191  * not restart the event.
1192  */
1193 void perf_event_task_sched_out(struct task_struct *task,
1194                                  struct task_struct *next)
1195 {
1196         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1197         struct perf_event_context *ctx = task->perf_event_ctxp;
1198         struct perf_event_context *next_ctx;
1199         struct perf_event_context *parent;
1200         struct pt_regs *regs;
1201         int do_switch = 1;
1202
1203         regs = task_pt_regs(task);
1204         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1205
1206         if (likely(!ctx || !cpuctx->task_ctx))
1207                 return;
1208
1209         rcu_read_lock();
1210         parent = rcu_dereference(ctx->parent_ctx);
1211         next_ctx = next->perf_event_ctxp;
1212         if (parent && next_ctx &&
1213             rcu_dereference(next_ctx->parent_ctx) == parent) {
1214                 /*
1215                  * Looks like the two contexts are clones, so we might be
1216                  * able to optimize the context switch.  We lock both
1217                  * contexts and check that they are clones under the
1218                  * lock (including re-checking that neither has been
1219                  * uncloned in the meantime).  It doesn't matter which
1220                  * order we take the locks because no other cpu could
1221                  * be trying to lock both of these tasks.
1222                  */
1223                 raw_spin_lock(&ctx->lock);
1224                 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1225                 if (context_equiv(ctx, next_ctx)) {
1226                         /*
1227                          * XXX do we need a memory barrier of sorts
1228                          * wrt to rcu_dereference() of perf_event_ctxp
1229                          */
1230                         task->perf_event_ctxp = next_ctx;
1231                         next->perf_event_ctxp = ctx;
1232                         ctx->task = next;
1233                         next_ctx->task = task;
1234                         do_switch = 0;
1235
1236                         perf_event_sync_stat(ctx, next_ctx);
1237                 }
1238                 raw_spin_unlock(&next_ctx->lock);
1239                 raw_spin_unlock(&ctx->lock);
1240         }
1241         rcu_read_unlock();
1242
1243         if (do_switch) {
1244                 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1245                 cpuctx->task_ctx = NULL;
1246         }
1247 }
1248
1249 static void task_ctx_sched_out(struct perf_event_context *ctx,
1250                                enum event_type_t event_type)
1251 {
1252         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1253
1254         if (!cpuctx->task_ctx)
1255                 return;
1256
1257         if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1258                 return;
1259
1260         ctx_sched_out(ctx, cpuctx, event_type);
1261         cpuctx->task_ctx = NULL;
1262 }
1263
1264 /*
1265  * Called with IRQs disabled
1266  */
1267 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1268 {
1269         task_ctx_sched_out(ctx, EVENT_ALL);
1270 }
1271
1272 /*
1273  * Called with IRQs disabled
1274  */
1275 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1276                               enum event_type_t event_type)
1277 {
1278         ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1279 }
1280
1281 static void
1282 ctx_pinned_sched_in(struct perf_event_context *ctx,
1283                     struct perf_cpu_context *cpuctx,
1284                     int cpu)
1285 {
1286         struct perf_event *event;
1287
1288         list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1289                 if (event->state <= PERF_EVENT_STATE_OFF)
1290                         continue;
1291                 if (event->cpu != -1 && event->cpu != cpu)
1292                         continue;
1293
1294                 if (group_can_go_on(event, cpuctx, 1))
1295                         group_sched_in(event, cpuctx, ctx, cpu);
1296
1297                 /*
1298                  * If this pinned group hasn't been scheduled,
1299                  * put it in error state.
1300                  */
1301                 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1302                         update_group_times(event);
1303                         event->state = PERF_EVENT_STATE_ERROR;
1304                 }
1305         }
1306 }
1307
1308 static void
1309 ctx_flexible_sched_in(struct perf_event_context *ctx,
1310                       struct perf_cpu_context *cpuctx,
1311                       int cpu)
1312 {
1313         struct perf_event *event;
1314         int can_add_hw = 1;
1315
1316         list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1317                 /* Ignore events in OFF or ERROR state */
1318                 if (event->state <= PERF_EVENT_STATE_OFF)
1319                         continue;
1320                 /*
1321                  * Listen to the 'cpu' scheduling filter constraint
1322                  * of events:
1323                  */
1324                 if (event->cpu != -1 && event->cpu != cpu)
1325                         continue;
1326
1327                 if (group_can_go_on(event, cpuctx, can_add_hw))
1328                         if (group_sched_in(event, cpuctx, ctx, cpu))
1329                                 can_add_hw = 0;
1330         }
1331 }
1332
1333 static void
1334 ctx_sched_in(struct perf_event_context *ctx,
1335              struct perf_cpu_context *cpuctx,
1336              enum event_type_t event_type)
1337 {
1338         int cpu = smp_processor_id();
1339
1340         raw_spin_lock(&ctx->lock);
1341         ctx->is_active = 1;
1342         if (likely(!ctx->nr_events))
1343                 goto out;
1344
1345         ctx->timestamp = perf_clock();
1346
1347         perf_disable();
1348
1349         /*
1350          * First go through the list and put on any pinned groups
1351          * in order to give them the best chance of going on.
1352          */
1353         if (event_type & EVENT_PINNED)
1354                 ctx_pinned_sched_in(ctx, cpuctx, cpu);
1355
1356         /* Then walk through the lower prio flexible groups */
1357         if (event_type & EVENT_FLEXIBLE)
1358                 ctx_flexible_sched_in(ctx, cpuctx, cpu);
1359
1360         perf_enable();
1361  out:
1362         raw_spin_unlock(&ctx->lock);
1363 }
1364
1365 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1366                              enum event_type_t event_type)
1367 {
1368         struct perf_event_context *ctx = &cpuctx->ctx;
1369
1370         ctx_sched_in(ctx, cpuctx, event_type);
1371 }
1372
1373 static void task_ctx_sched_in(struct task_struct *task,
1374                               enum event_type_t event_type)
1375 {
1376         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1377         struct perf_event_context *ctx = task->perf_event_ctxp;
1378
1379         if (likely(!ctx))
1380                 return;
1381         if (cpuctx->task_ctx == ctx)
1382                 return;
1383         ctx_sched_in(ctx, cpuctx, event_type);
1384         cpuctx->task_ctx = ctx;
1385 }
1386 /*
1387  * Called from scheduler to add the events of the current task
1388  * with interrupts disabled.
1389  *
1390  * We restore the event value and then enable it.
1391  *
1392  * This does not protect us against NMI, but enable()
1393  * sets the enabled bit in the control field of event _before_
1394  * accessing the event control register. If a NMI hits, then it will
1395  * keep the event running.
1396  */
1397 void perf_event_task_sched_in(struct task_struct *task)
1398 {
1399         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1400         struct perf_event_context *ctx = task->perf_event_ctxp;
1401
1402         if (likely(!ctx))
1403                 return;
1404
1405         if (cpuctx->task_ctx == ctx)
1406                 return;
1407
1408         /*
1409          * We want to keep the following priority order:
1410          * cpu pinned (that don't need to move), task pinned,
1411          * cpu flexible, task flexible.
1412          */
1413         cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1414
1415         ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1416         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1417         ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1418
1419         cpuctx->task_ctx = ctx;
1420 }
1421
1422 #define MAX_INTERRUPTS (~0ULL)
1423
1424 static void perf_log_throttle(struct perf_event *event, int enable);
1425
1426 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1427 {
1428         u64 frequency = event->attr.sample_freq;
1429         u64 sec = NSEC_PER_SEC;
1430         u64 divisor, dividend;
1431
1432         int count_fls, nsec_fls, frequency_fls, sec_fls;
1433
1434         count_fls = fls64(count);
1435         nsec_fls = fls64(nsec);
1436         frequency_fls = fls64(frequency);
1437         sec_fls = 30;
1438
1439         /*
1440          * We got @count in @nsec, with a target of sample_freq HZ
1441          * the target period becomes:
1442          *
1443          *             @count * 10^9
1444          * period = -------------------
1445          *          @nsec * sample_freq
1446          *
1447          */
1448
1449         /*
1450          * Reduce accuracy by one bit such that @a and @b converge
1451          * to a similar magnitude.
1452          */
1453 #define REDUCE_FLS(a, b)                \
1454 do {                                    \
1455         if (a##_fls > b##_fls) {        \
1456                 a >>= 1;                \
1457                 a##_fls--;              \
1458         } else {                        \
1459                 b >>= 1;                \
1460                 b##_fls--;              \
1461         }                               \
1462 } while (0)
1463
1464         /*
1465          * Reduce accuracy until either term fits in a u64, then proceed with
1466          * the other, so that finally we can do a u64/u64 division.
1467          */
1468         while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1469                 REDUCE_FLS(nsec, frequency);
1470                 REDUCE_FLS(sec, count);
1471         }
1472
1473         if (count_fls + sec_fls > 64) {
1474                 divisor = nsec * frequency;
1475
1476                 while (count_fls + sec_fls > 64) {
1477                         REDUCE_FLS(count, sec);
1478                         divisor >>= 1;
1479                 }
1480
1481                 dividend = count * sec;
1482         } else {
1483                 dividend = count * sec;
1484
1485                 while (nsec_fls + frequency_fls > 64) {
1486                         REDUCE_FLS(nsec, frequency);
1487                         dividend >>= 1;
1488                 }
1489
1490                 divisor = nsec * frequency;
1491         }
1492
1493         return div64_u64(dividend, divisor);
1494 }
1495
1496 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1497 {
1498         struct hw_perf_event *hwc = &event->hw;
1499         u64 period, sample_period;
1500         s64 delta;
1501
1502         period = perf_calculate_period(event, nsec, count);
1503
1504         delta = (s64)(period - hwc->sample_period);
1505         delta = (delta + 7) / 8; /* low pass filter */
1506
1507         sample_period = hwc->sample_period + delta;
1508
1509         if (!sample_period)
1510                 sample_period = 1;
1511
1512         hwc->sample_period = sample_period;
1513
1514         if (atomic64_read(&hwc->period_left) > 8*sample_period) {
1515                 perf_disable();
1516                 event->pmu->disable(event);
1517                 atomic64_set(&hwc->period_left, 0);
1518                 event->pmu->enable(event);
1519                 perf_enable();
1520         }
1521 }
1522
1523 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1524 {
1525         struct perf_event *event;
1526         struct hw_perf_event *hwc;
1527         u64 interrupts, now;
1528         s64 delta;
1529
1530         raw_spin_lock(&ctx->lock);
1531         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1532                 if (event->state != PERF_EVENT_STATE_ACTIVE)
1533                         continue;
1534
1535                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1536                         continue;
1537
1538                 hwc = &event->hw;
1539
1540                 interrupts = hwc->interrupts;
1541                 hwc->interrupts = 0;
1542
1543                 /*
1544                  * unthrottle events on the tick
1545                  */
1546                 if (interrupts == MAX_INTERRUPTS) {
1547                         perf_log_throttle(event, 1);
1548                         event->pmu->unthrottle(event);
1549                 }
1550
1551                 if (!event->attr.freq || !event->attr.sample_freq)
1552                         continue;
1553
1554                 event->pmu->read(event);
1555                 now = atomic64_read(&event->count);
1556                 delta = now - hwc->freq_count_stamp;
1557                 hwc->freq_count_stamp = now;
1558
1559                 if (delta > 0)
1560                         perf_adjust_period(event, TICK_NSEC, delta);
1561         }
1562         raw_spin_unlock(&ctx->lock);
1563 }
1564
1565 /*
1566  * Round-robin a context's events:
1567  */
1568 static void rotate_ctx(struct perf_event_context *ctx)
1569 {
1570         if (!ctx->nr_events)
1571                 return;
1572
1573         raw_spin_lock(&ctx->lock);
1574
1575         /* Rotate the first entry last of non-pinned groups */
1576         list_rotate_left(&ctx->flexible_groups);
1577
1578         raw_spin_unlock(&ctx->lock);
1579 }
1580
1581 void perf_event_task_tick(struct task_struct *curr)
1582 {
1583         struct perf_cpu_context *cpuctx;
1584         struct perf_event_context *ctx;
1585
1586         if (!atomic_read(&nr_events))
1587                 return;
1588
1589         cpuctx = &__get_cpu_var(perf_cpu_context);
1590         ctx = curr->perf_event_ctxp;
1591
1592         perf_disable();
1593
1594         perf_ctx_adjust_freq(&cpuctx->ctx);
1595         if (ctx)
1596                 perf_ctx_adjust_freq(ctx);
1597
1598         cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1599         if (ctx)
1600                 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1601
1602         rotate_ctx(&cpuctx->ctx);
1603         if (ctx)
1604                 rotate_ctx(ctx);
1605
1606         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1607         if (ctx)
1608                 task_ctx_sched_in(curr, EVENT_FLEXIBLE);
1609
1610         perf_enable();
1611 }
1612
1613 static int event_enable_on_exec(struct perf_event *event,
1614                                 struct perf_event_context *ctx)
1615 {
1616         if (!event->attr.enable_on_exec)
1617                 return 0;
1618
1619         event->attr.enable_on_exec = 0;
1620         if (event->state >= PERF_EVENT_STATE_INACTIVE)
1621                 return 0;
1622
1623         __perf_event_mark_enabled(event, ctx);
1624
1625         return 1;
1626 }
1627
1628 /*
1629  * Enable all of a task's events that have been marked enable-on-exec.
1630  * This expects task == current.
1631  */
1632 static void perf_event_enable_on_exec(struct task_struct *task)
1633 {
1634         struct perf_event_context *ctx;
1635         struct perf_event *event;
1636         unsigned long flags;
1637         int enabled = 0;
1638         int ret;
1639
1640         local_irq_save(flags);
1641         ctx = task->perf_event_ctxp;
1642         if (!ctx || !ctx->nr_events)
1643                 goto out;
1644
1645         __perf_event_task_sched_out(ctx);
1646
1647         raw_spin_lock(&ctx->lock);
1648
1649         list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1650                 ret = event_enable_on_exec(event, ctx);
1651                 if (ret)
1652                         enabled = 1;
1653         }
1654
1655         list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1656                 ret = event_enable_on_exec(event, ctx);
1657                 if (ret)
1658                         enabled = 1;
1659         }
1660
1661         /*
1662          * Unclone this context if we enabled any event.
1663          */
1664         if (enabled)
1665                 unclone_ctx(ctx);
1666
1667         raw_spin_unlock(&ctx->lock);
1668
1669         perf_event_task_sched_in(task);
1670  out:
1671         local_irq_restore(flags);
1672 }
1673
1674 /*
1675  * Cross CPU call to read the hardware event
1676  */
1677 static void __perf_event_read(void *info)
1678 {
1679         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1680         struct perf_event *event = info;
1681         struct perf_event_context *ctx = event->ctx;
1682
1683         /*
1684          * If this is a task context, we need to check whether it is
1685          * the current task context of this cpu.  If not it has been
1686          * scheduled out before the smp call arrived.  In that case
1687          * event->count would have been updated to a recent sample
1688          * when the event was scheduled out.
1689          */
1690         if (ctx->task && cpuctx->task_ctx != ctx)
1691                 return;
1692
1693         raw_spin_lock(&ctx->lock);
1694         update_context_time(ctx);
1695         update_event_times(event);
1696         raw_spin_unlock(&ctx->lock);
1697
1698         event->pmu->read(event);
1699 }
1700
1701 static u64 perf_event_read(struct perf_event *event)
1702 {
1703         /*
1704          * If event is enabled and currently active on a CPU, update the
1705          * value in the event structure:
1706          */
1707         if (event->state == PERF_EVENT_STATE_ACTIVE) {
1708                 smp_call_function_single(event->oncpu,
1709                                          __perf_event_read, event, 1);
1710         } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1711                 struct perf_event_context *ctx = event->ctx;
1712                 unsigned long flags;
1713
1714                 raw_spin_lock_irqsave(&ctx->lock, flags);
1715                 update_context_time(ctx);
1716                 update_event_times(event);
1717                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1718         }
1719
1720         return atomic64_read(&event->count);
1721 }
1722
1723 /*
1724  * Initialize the perf_event context in a task_struct:
1725  */
1726 static void
1727 __perf_event_init_context(struct perf_event_context *ctx,
1728                             struct task_struct *task)
1729 {
1730         raw_spin_lock_init(&ctx->lock);
1731         mutex_init(&ctx->mutex);
1732         INIT_LIST_HEAD(&ctx->pinned_groups);
1733         INIT_LIST_HEAD(&ctx->flexible_groups);
1734         INIT_LIST_HEAD(&ctx->event_list);
1735         atomic_set(&ctx->refcount, 1);
1736         ctx->task = task;
1737 }
1738
1739 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1740 {
1741         struct perf_event_context *ctx;
1742         struct perf_cpu_context *cpuctx;
1743         struct task_struct *task;
1744         unsigned long flags;
1745         int err;
1746
1747         if (pid == -1 && cpu != -1) {
1748                 /* Must be root to operate on a CPU event: */
1749                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1750                         return ERR_PTR(-EACCES);
1751
1752                 if (cpu < 0 || cpu >= nr_cpumask_bits)
1753                         return ERR_PTR(-EINVAL);
1754
1755                 /*
1756                  * We could be clever and allow to attach a event to an
1757                  * offline CPU and activate it when the CPU comes up, but
1758                  * that's for later.
1759                  */
1760                 if (!cpu_online(cpu))
1761                         return ERR_PTR(-ENODEV);
1762
1763                 cpuctx = &per_cpu(perf_cpu_context, cpu);
1764                 ctx = &cpuctx->ctx;
1765                 get_ctx(ctx);
1766
1767                 return ctx;
1768         }
1769
1770         rcu_read_lock();
1771         if (!pid)
1772                 task = current;
1773         else
1774                 task = find_task_by_vpid(pid);
1775         if (task)
1776                 get_task_struct(task);
1777         rcu_read_unlock();
1778
1779         if (!task)
1780                 return ERR_PTR(-ESRCH);
1781
1782         /*
1783          * Can't attach events to a dying task.
1784          */
1785         err = -ESRCH;
1786         if (task->flags & PF_EXITING)
1787                 goto errout;
1788
1789         /* Reuse ptrace permission checks for now. */
1790         err = -EACCES;
1791         if (!ptrace_may_access(task, PTRACE_MODE_READ))
1792                 goto errout;
1793
1794  retry:
1795         ctx = perf_lock_task_context(task, &flags);
1796         if (ctx) {
1797                 unclone_ctx(ctx);
1798                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1799         }
1800
1801         if (!ctx) {
1802                 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1803                 err = -ENOMEM;
1804                 if (!ctx)
1805                         goto errout;
1806                 __perf_event_init_context(ctx, task);
1807                 get_ctx(ctx);
1808                 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1809                         /*
1810                          * We raced with some other task; use
1811                          * the context they set.
1812                          */
1813                         kfree(ctx);
1814                         goto retry;
1815                 }
1816                 get_task_struct(task);
1817         }
1818
1819         put_task_struct(task);
1820         return ctx;
1821
1822  errout:
1823         put_task_struct(task);
1824         return ERR_PTR(err);
1825 }
1826
1827 static void perf_event_free_filter(struct perf_event *event);
1828
1829 static void free_event_rcu(struct rcu_head *head)
1830 {
1831         struct perf_event *event;
1832
1833         event = container_of(head, struct perf_event, rcu_head);
1834         if (event->ns)
1835                 put_pid_ns(event->ns);
1836         perf_event_free_filter(event);
1837         kfree(event);
1838 }
1839
1840 static void perf_pending_sync(struct perf_event *event);
1841
1842 static void free_event(struct perf_event *event)
1843 {
1844         perf_pending_sync(event);
1845
1846         if (!event->parent) {
1847                 atomic_dec(&nr_events);
1848                 if (event->attr.mmap)
1849                         atomic_dec(&nr_mmap_events);
1850                 if (event->attr.comm)
1851                         atomic_dec(&nr_comm_events);
1852                 if (event->attr.task)
1853                         atomic_dec(&nr_task_events);
1854         }
1855
1856         if (event->output) {
1857                 fput(event->output->filp);
1858                 event->output = NULL;
1859         }
1860
1861         if (event->destroy)
1862                 event->destroy(event);
1863
1864         put_ctx(event->ctx);
1865         call_rcu(&event->rcu_head, free_event_rcu);
1866 }
1867
1868 int perf_event_release_kernel(struct perf_event *event)
1869 {
1870         struct perf_event_context *ctx = event->ctx;
1871
1872         WARN_ON_ONCE(ctx->parent_ctx);
1873         mutex_lock(&ctx->mutex);
1874         perf_event_remove_from_context(event);
1875         mutex_unlock(&ctx->mutex);
1876
1877         mutex_lock(&event->owner->perf_event_mutex);
1878         list_del_init(&event->owner_entry);
1879         mutex_unlock(&event->owner->perf_event_mutex);
1880         put_task_struct(event->owner);
1881
1882         free_event(event);
1883
1884         return 0;
1885 }
1886 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1887
1888 /*
1889  * Called when the last reference to the file is gone.
1890  */
1891 static int perf_release(struct inode *inode, struct file *file)
1892 {
1893         struct perf_event *event = file->private_data;
1894
1895         file->private_data = NULL;
1896
1897         return perf_event_release_kernel(event);
1898 }
1899
1900 static int perf_event_read_size(struct perf_event *event)
1901 {
1902         int entry = sizeof(u64); /* value */
1903         int size = 0;
1904         int nr = 1;
1905
1906         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1907                 size += sizeof(u64);
1908
1909         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1910                 size += sizeof(u64);
1911
1912         if (event->attr.read_format & PERF_FORMAT_ID)
1913                 entry += sizeof(u64);
1914
1915         if (event->attr.read_format & PERF_FORMAT_GROUP) {
1916                 nr += event->group_leader->nr_siblings;
1917                 size += sizeof(u64);
1918         }
1919
1920         size += entry * nr;
1921
1922         return size;
1923 }
1924
1925 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1926 {
1927         struct perf_event *child;
1928         u64 total = 0;
1929
1930         *enabled = 0;
1931         *running = 0;
1932
1933         mutex_lock(&event->child_mutex);
1934         total += perf_event_read(event);
1935         *enabled += event->total_time_enabled +
1936                         atomic64_read(&event->child_total_time_enabled);
1937         *running += event->total_time_running +
1938                         atomic64_read(&event->child_total_time_running);
1939
1940         list_for_each_entry(child, &event->child_list, child_list) {
1941                 total += perf_event_read(child);
1942                 *enabled += child->total_time_enabled;
1943                 *running += child->total_time_running;
1944         }
1945         mutex_unlock(&event->child_mutex);
1946
1947         return total;
1948 }
1949 EXPORT_SYMBOL_GPL(perf_event_read_value);
1950
1951 static int perf_event_read_group(struct perf_event *event,
1952                                    u64 read_format, char __user *buf)
1953 {
1954         struct perf_event *leader = event->group_leader, *sub;
1955         int n = 0, size = 0, ret = -EFAULT;
1956         struct perf_event_context *ctx = leader->ctx;
1957         u64 values[5];
1958         u64 count, enabled, running;
1959
1960         mutex_lock(&ctx->mutex);
1961         count = perf_event_read_value(leader, &enabled, &running);
1962
1963         values[n++] = 1 + leader->nr_siblings;
1964         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1965                 values[n++] = enabled;
1966         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1967                 values[n++] = running;
1968         values[n++] = count;
1969         if (read_format & PERF_FORMAT_ID)
1970                 values[n++] = primary_event_id(leader);
1971
1972         size = n * sizeof(u64);
1973
1974         if (copy_to_user(buf, values, size))
1975                 goto unlock;
1976
1977         ret = size;
1978
1979         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1980                 n = 0;
1981
1982                 values[n++] = perf_event_read_value(sub, &enabled, &running);
1983                 if (read_format & PERF_FORMAT_ID)
1984                         values[n++] = primary_event_id(sub);
1985
1986                 size = n * sizeof(u64);
1987
1988                 if (copy_to_user(buf + ret, values, size)) {
1989                         ret = -EFAULT;
1990                         goto unlock;
1991                 }
1992
1993                 ret += size;
1994         }
1995 unlock:
1996         mutex_unlock(&ctx->mutex);
1997
1998         return ret;
1999 }
2000
2001 static int perf_event_read_one(struct perf_event *event,
2002                                  u64 read_format, char __user *buf)
2003 {
2004         u64 enabled, running;
2005         u64 values[4];
2006         int n = 0;
2007
2008         values[n++] = perf_event_read_value(event, &enabled, &running);
2009         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2010                 values[n++] = enabled;
2011         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2012                 values[n++] = running;
2013         if (read_format & PERF_FORMAT_ID)
2014                 values[n++] = primary_event_id(event);
2015
2016         if (copy_to_user(buf, values, n * sizeof(u64)))
2017                 return -EFAULT;
2018
2019         return n * sizeof(u64);
2020 }
2021
2022 /*
2023  * Read the performance event - simple non blocking version for now
2024  */
2025 static ssize_t
2026 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2027 {
2028         u64 read_format = event->attr.read_format;
2029         int ret;
2030
2031         /*
2032          * Return end-of-file for a read on a event that is in
2033          * error state (i.e. because it was pinned but it couldn't be
2034          * scheduled on to the CPU at some point).
2035          */
2036         if (event->state == PERF_EVENT_STATE_ERROR)
2037                 return 0;
2038
2039         if (count < perf_event_read_size(event))
2040                 return -ENOSPC;
2041
2042         WARN_ON_ONCE(event->ctx->parent_ctx);
2043         if (read_format & PERF_FORMAT_GROUP)
2044                 ret = perf_event_read_group(event, read_format, buf);
2045         else
2046                 ret = perf_event_read_one(event, read_format, buf);
2047
2048         return ret;
2049 }
2050
2051 static ssize_t
2052 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2053 {
2054         struct perf_event *event = file->private_data;
2055
2056         return perf_read_hw(event, buf, count);
2057 }
2058
2059 static unsigned int perf_poll(struct file *file, poll_table *wait)
2060 {
2061         struct perf_event *event = file->private_data;
2062         struct perf_mmap_data *data;
2063         unsigned int events = POLL_HUP;
2064
2065         rcu_read_lock();
2066         data = rcu_dereference(event->data);
2067         if (data)
2068                 events = atomic_xchg(&data->poll, 0);
2069         rcu_read_unlock();
2070
2071         poll_wait(file, &event->waitq, wait);
2072
2073         return events;
2074 }
2075
2076 static void perf_event_reset(struct perf_event *event)
2077 {
2078         (void)perf_event_read(event);
2079         atomic64_set(&event->count, 0);
2080         perf_event_update_userpage(event);
2081 }
2082
2083 /*
2084  * Holding the top-level event's child_mutex means that any
2085  * descendant process that has inherited this event will block
2086  * in sync_child_event if it goes to exit, thus satisfying the
2087  * task existence requirements of perf_event_enable/disable.
2088  */
2089 static void perf_event_for_each_child(struct perf_event *event,
2090                                         void (*func)(struct perf_event *))
2091 {
2092         struct perf_event *child;
2093
2094         WARN_ON_ONCE(event->ctx->parent_ctx);
2095         mutex_lock(&event->child_mutex);
2096         func(event);
2097         list_for_each_entry(child, &event->child_list, child_list)
2098                 func(child);
2099         mutex_unlock(&event->child_mutex);
2100 }
2101
2102 static void perf_event_for_each(struct perf_event *event,
2103                                   void (*func)(struct perf_event *))
2104 {
2105         struct perf_event_context *ctx = event->ctx;
2106         struct perf_event *sibling;
2107
2108         WARN_ON_ONCE(ctx->parent_ctx);
2109         mutex_lock(&ctx->mutex);
2110         event = event->group_leader;
2111
2112         perf_event_for_each_child(event, func);
2113         func(event);
2114         list_for_each_entry(sibling, &event->sibling_list, group_entry)
2115                 perf_event_for_each_child(event, func);
2116         mutex_unlock(&ctx->mutex);
2117 }
2118
2119 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2120 {
2121         struct perf_event_context *ctx = event->ctx;
2122         unsigned long size;
2123         int ret = 0;
2124         u64 value;
2125
2126         if (!event->attr.sample_period)
2127                 return -EINVAL;
2128
2129         size = copy_from_user(&value, arg, sizeof(value));
2130         if (size != sizeof(value))
2131                 return -EFAULT;
2132
2133         if (!value)
2134                 return -EINVAL;
2135
2136         raw_spin_lock_irq(&ctx->lock);
2137         if (event->attr.freq) {
2138                 if (value > sysctl_perf_event_sample_rate) {
2139                         ret = -EINVAL;
2140                         goto unlock;
2141                 }
2142
2143                 event->attr.sample_freq = value;
2144         } else {
2145                 event->attr.sample_period = value;
2146                 event->hw.sample_period = value;
2147         }
2148 unlock:
2149         raw_spin_unlock_irq(&ctx->lock);
2150
2151         return ret;
2152 }
2153
2154 static int perf_event_set_output(struct perf_event *event, int output_fd);
2155 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2156
2157 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2158 {
2159         struct perf_event *event = file->private_data;
2160         void (*func)(struct perf_event *);
2161         u32 flags = arg;
2162
2163         switch (cmd) {
2164         case PERF_EVENT_IOC_ENABLE:
2165                 func = perf_event_enable;
2166                 break;
2167         case PERF_EVENT_IOC_DISABLE:
2168                 func = perf_event_disable;
2169                 break;
2170         case PERF_EVENT_IOC_RESET:
2171                 func = perf_event_reset;
2172                 break;
2173
2174         case PERF_EVENT_IOC_REFRESH:
2175                 return perf_event_refresh(event, arg);
2176
2177         case PERF_EVENT_IOC_PERIOD:
2178                 return perf_event_period(event, (u64 __user *)arg);
2179
2180         case PERF_EVENT_IOC_SET_OUTPUT:
2181                 return perf_event_set_output(event, arg);
2182
2183         case PERF_EVENT_IOC_SET_FILTER:
2184                 return perf_event_set_filter(event, (void __user *)arg);
2185
2186         default:
2187                 return -ENOTTY;
2188         }
2189
2190         if (flags & PERF_IOC_FLAG_GROUP)
2191                 perf_event_for_each(event, func);
2192         else
2193                 perf_event_for_each_child(event, func);
2194
2195         return 0;
2196 }
2197
2198 int perf_event_task_enable(void)
2199 {
2200         struct perf_event *event;
2201
2202         mutex_lock(&current->perf_event_mutex);
2203         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2204                 perf_event_for_each_child(event, perf_event_enable);
2205         mutex_unlock(&current->perf_event_mutex);
2206
2207         return 0;
2208 }
2209
2210 int perf_event_task_disable(void)
2211 {
2212         struct perf_event *event;
2213
2214         mutex_lock(&current->perf_event_mutex);
2215         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2216                 perf_event_for_each_child(event, perf_event_disable);
2217         mutex_unlock(&current->perf_event_mutex);
2218
2219         return 0;
2220 }
2221
2222 #ifndef PERF_EVENT_INDEX_OFFSET
2223 # define PERF_EVENT_INDEX_OFFSET 0
2224 #endif
2225
2226 static int perf_event_index(struct perf_event *event)
2227 {
2228         if (event->state != PERF_EVENT_STATE_ACTIVE)
2229                 return 0;
2230
2231         return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2232 }
2233
2234 /*
2235  * Callers need to ensure there can be no nesting of this function, otherwise
2236  * the seqlock logic goes bad. We can not serialize this because the arch
2237  * code calls this from NMI context.
2238  */
2239 void perf_event_update_userpage(struct perf_event *event)
2240 {
2241         struct perf_event_mmap_page *userpg;
2242         struct perf_mmap_data *data;
2243
2244         rcu_read_lock();
2245         data = rcu_dereference(event->data);
2246         if (!data)
2247                 goto unlock;
2248
2249         userpg = data->user_page;
2250
2251         /*
2252          * Disable preemption so as to not let the corresponding user-space
2253          * spin too long if we get preempted.
2254          */
2255         preempt_disable();
2256         ++userpg->lock;
2257         barrier();
2258         userpg->index = perf_event_index(event);
2259         userpg->offset = atomic64_read(&event->count);
2260         if (event->state == PERF_EVENT_STATE_ACTIVE)
2261                 userpg->offset -= atomic64_read(&event->hw.prev_count);
2262
2263         userpg->time_enabled = event->total_time_enabled +
2264                         atomic64_read(&event->child_total_time_enabled);
2265
2266         userpg->time_running = event->total_time_running +
2267                         atomic64_read(&event->child_total_time_running);
2268
2269         barrier();
2270         ++userpg->lock;
2271         preempt_enable();
2272 unlock:
2273         rcu_read_unlock();
2274 }
2275
2276 static unsigned long perf_data_size(struct perf_mmap_data *data)
2277 {
2278         return data->nr_pages << (PAGE_SHIFT + data->data_order);
2279 }
2280
2281 #ifndef CONFIG_PERF_USE_VMALLOC
2282
2283 /*
2284  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2285  */
2286
2287 static struct page *
2288 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2289 {
2290         if (pgoff > data->nr_pages)
2291                 return NULL;
2292
2293         if (pgoff == 0)
2294                 return virt_to_page(data->user_page);
2295
2296         return virt_to_page(data->data_pages[pgoff - 1]);
2297 }
2298
2299 static struct perf_mmap_data *
2300 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2301 {
2302         struct perf_mmap_data *data;
2303         unsigned long size;
2304         int i;
2305
2306         WARN_ON(atomic_read(&event->mmap_count));
2307
2308         size = sizeof(struct perf_mmap_data);
2309         size += nr_pages * sizeof(void *);
2310
2311         data = kzalloc(size, GFP_KERNEL);
2312         if (!data)
2313                 goto fail;
2314
2315         data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2316         if (!data->user_page)
2317                 goto fail_user_page;
2318
2319         for (i = 0; i < nr_pages; i++) {
2320                 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2321                 if (!data->data_pages[i])
2322                         goto fail_data_pages;
2323         }
2324
2325         data->data_order = 0;
2326         data->nr_pages = nr_pages;
2327
2328         return data;
2329
2330 fail_data_pages:
2331         for (i--; i >= 0; i--)
2332                 free_page((unsigned long)data->data_pages[i]);
2333
2334         free_page((unsigned long)data->user_page);
2335
2336 fail_user_page:
2337         kfree(data);
2338
2339 fail:
2340         return NULL;
2341 }
2342
2343 static void perf_mmap_free_page(unsigned long addr)
2344 {
2345         struct page *page = virt_to_page((void *)addr);
2346
2347         page->mapping = NULL;
2348         __free_page(page);
2349 }
2350
2351 static void perf_mmap_data_free(struct perf_mmap_data *data)
2352 {
2353         int i;
2354
2355         perf_mmap_free_page((unsigned long)data->user_page);
2356         for (i = 0; i < data->nr_pages; i++)
2357                 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2358         kfree(data);
2359 }
2360
2361 #else
2362
2363 /*
2364  * Back perf_mmap() with vmalloc memory.
2365  *
2366  * Required for architectures that have d-cache aliasing issues.
2367  */
2368
2369 static struct page *
2370 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2371 {
2372         if (pgoff > (1UL << data->data_order))
2373                 return NULL;
2374
2375         return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2376 }
2377
2378 static void perf_mmap_unmark_page(void *addr)
2379 {
2380         struct page *page = vmalloc_to_page(addr);
2381
2382         page->mapping = NULL;
2383 }
2384
2385 static void perf_mmap_data_free_work(struct work_struct *work)
2386 {
2387         struct perf_mmap_data *data;
2388         void *base;
2389         int i, nr;
2390
2391         data = container_of(work, struct perf_mmap_data, work);
2392         nr = 1 << data->data_order;
2393
2394         base = data->user_page;
2395         for (i = 0; i < nr + 1; i++)
2396                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2397
2398         vfree(base);
2399         kfree(data);
2400 }
2401
2402 static void perf_mmap_data_free(struct perf_mmap_data *data)
2403 {
2404         schedule_work(&data->work);
2405 }
2406
2407 static struct perf_mmap_data *
2408 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2409 {
2410         struct perf_mmap_data *data;
2411         unsigned long size;
2412         void *all_buf;
2413
2414         WARN_ON(atomic_read(&event->mmap_count));
2415
2416         size = sizeof(struct perf_mmap_data);
2417         size += sizeof(void *);
2418
2419         data = kzalloc(size, GFP_KERNEL);
2420         if (!data)
2421                 goto fail;
2422
2423         INIT_WORK(&data->work, perf_mmap_data_free_work);
2424
2425         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2426         if (!all_buf)
2427                 goto fail_all_buf;
2428
2429         data->user_page = all_buf;
2430         data->data_pages[0] = all_buf + PAGE_SIZE;
2431         data->data_order = ilog2(nr_pages);
2432         data->nr_pages = 1;
2433
2434         return data;
2435
2436 fail_all_buf:
2437         kfree(data);
2438
2439 fail:
2440         return NULL;
2441 }
2442
2443 #endif
2444
2445 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2446 {
2447         struct perf_event *event = vma->vm_file->private_data;
2448         struct perf_mmap_data *data;
2449         int ret = VM_FAULT_SIGBUS;
2450
2451         if (vmf->flags & FAULT_FLAG_MKWRITE) {
2452                 if (vmf->pgoff == 0)
2453                         ret = 0;
2454                 return ret;
2455         }
2456
2457         rcu_read_lock();
2458         data = rcu_dereference(event->data);
2459         if (!data)
2460                 goto unlock;
2461
2462         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2463                 goto unlock;
2464
2465         vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2466         if (!vmf->page)
2467                 goto unlock;
2468
2469         get_page(vmf->page);
2470         vmf->page->mapping = vma->vm_file->f_mapping;
2471         vmf->page->index   = vmf->pgoff;
2472
2473         ret = 0;
2474 unlock:
2475         rcu_read_unlock();
2476
2477         return ret;
2478 }
2479
2480 static void
2481 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2482 {
2483         long max_size = perf_data_size(data);
2484
2485         atomic_set(&data->lock, -1);
2486
2487         if (event->attr.watermark) {
2488                 data->watermark = min_t(long, max_size,
2489                                         event->attr.wakeup_watermark);
2490         }
2491
2492         if (!data->watermark)
2493                 data->watermark = max_size / 2;
2494
2495
2496         rcu_assign_pointer(event->data, data);
2497 }
2498
2499 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2500 {
2501         struct perf_mmap_data *data;
2502
2503         data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2504         perf_mmap_data_free(data);
2505 }
2506
2507 static void perf_mmap_data_release(struct perf_event *event)
2508 {
2509         struct perf_mmap_data *data = event->data;
2510
2511         WARN_ON(atomic_read(&event->mmap_count));
2512
2513         rcu_assign_pointer(event->data, NULL);
2514         call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2515 }
2516
2517 static void perf_mmap_open(struct vm_area_struct *vma)
2518 {
2519         struct perf_event *event = vma->vm_file->private_data;
2520
2521         atomic_inc(&event->mmap_count);
2522 }
2523
2524 static void perf_mmap_close(struct vm_area_struct *vma)
2525 {
2526         struct perf_event *event = vma->vm_file->private_data;
2527
2528         WARN_ON_ONCE(event->ctx->parent_ctx);
2529         if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2530                 unsigned long size = perf_data_size(event->data);
2531                 struct user_struct *user = current_user();
2532
2533                 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2534                 vma->vm_mm->locked_vm -= event->data->nr_locked;
2535                 perf_mmap_data_release(event);
2536                 mutex_unlock(&event->mmap_mutex);
2537         }
2538 }
2539
2540 static const struct vm_operations_struct perf_mmap_vmops = {
2541         .open           = perf_mmap_open,
2542         .close          = perf_mmap_close,
2543         .fault          = perf_mmap_fault,
2544         .page_mkwrite   = perf_mmap_fault,
2545 };
2546
2547 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2548 {
2549         struct perf_event *event = file->private_data;
2550         unsigned long user_locked, user_lock_limit;
2551         struct user_struct *user = current_user();
2552         unsigned long locked, lock_limit;
2553         struct perf_mmap_data *data;
2554         unsigned long vma_size;
2555         unsigned long nr_pages;
2556         long user_extra, extra;
2557         int ret = 0;
2558
2559         if (!(vma->vm_flags & VM_SHARED))
2560                 return -EINVAL;
2561
2562         vma_size = vma->vm_end - vma->vm_start;
2563         nr_pages = (vma_size / PAGE_SIZE) - 1;
2564
2565         /*
2566          * If we have data pages ensure they're a power-of-two number, so we
2567          * can do bitmasks instead of modulo.
2568          */
2569         if (nr_pages != 0 && !is_power_of_2(nr_pages))
2570                 return -EINVAL;
2571
2572         if (vma_size != PAGE_SIZE * (1 + nr_pages))
2573                 return -EINVAL;
2574
2575         if (vma->vm_pgoff != 0)
2576                 return -EINVAL;
2577
2578         WARN_ON_ONCE(event->ctx->parent_ctx);
2579         mutex_lock(&event->mmap_mutex);
2580         if (event->output) {
2581                 ret = -EINVAL;
2582                 goto unlock;
2583         }
2584
2585         if (atomic_inc_not_zero(&event->mmap_count)) {
2586                 if (nr_pages != event->data->nr_pages)
2587                         ret = -EINVAL;
2588                 goto unlock;
2589         }
2590
2591         user_extra = nr_pages + 1;
2592         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2593
2594         /*
2595          * Increase the limit linearly with more CPUs:
2596          */
2597         user_lock_limit *= num_online_cpus();
2598
2599         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2600
2601         extra = 0;
2602         if (user_locked > user_lock_limit)
2603                 extra = user_locked - user_lock_limit;
2604
2605         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2606         lock_limit >>= PAGE_SHIFT;
2607         locked = vma->vm_mm->locked_vm + extra;
2608
2609         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2610                 !capable(CAP_IPC_LOCK)) {
2611                 ret = -EPERM;
2612                 goto unlock;
2613         }
2614
2615         WARN_ON(event->data);
2616
2617         data = perf_mmap_data_alloc(event, nr_pages);
2618         ret = -ENOMEM;
2619         if (!data)
2620                 goto unlock;
2621
2622         ret = 0;
2623         perf_mmap_data_init(event, data);
2624
2625         atomic_set(&event->mmap_count, 1);
2626         atomic_long_add(user_extra, &user->locked_vm);
2627         vma->vm_mm->locked_vm += extra;
2628         event->data->nr_locked = extra;
2629         if (vma->vm_flags & VM_WRITE)
2630                 event->data->writable = 1;
2631
2632 unlock:
2633         mutex_unlock(&event->mmap_mutex);
2634
2635         vma->vm_flags |= VM_RESERVED;
2636         vma->vm_ops = &perf_mmap_vmops;
2637
2638         return ret;
2639 }
2640
2641 static int perf_fasync(int fd, struct file *filp, int on)
2642 {
2643         struct inode *inode = filp->f_path.dentry->d_inode;
2644         struct perf_event *event = filp->private_data;
2645         int retval;
2646
2647         mutex_lock(&inode->i_mutex);
2648         retval = fasync_helper(fd, filp, on, &event->fasync);
2649         mutex_unlock(&inode->i_mutex);
2650
2651         if (retval < 0)
2652                 return retval;
2653
2654         return 0;
2655 }
2656
2657 static const struct file_operations perf_fops = {
2658         .release                = perf_release,
2659         .read                   = perf_read,
2660         .poll                   = perf_poll,
2661         .unlocked_ioctl         = perf_ioctl,
2662         .compat_ioctl           = perf_ioctl,
2663         .mmap                   = perf_mmap,
2664         .fasync                 = perf_fasync,
2665 };
2666
2667 /*
2668  * Perf event wakeup
2669  *
2670  * If there's data, ensure we set the poll() state and publish everything
2671  * to user-space before waking everybody up.
2672  */
2673
2674 void perf_event_wakeup(struct perf_event *event)
2675 {
2676         wake_up_all(&event->waitq);
2677
2678         if (event->pending_kill) {
2679                 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2680                 event->pending_kill = 0;
2681         }
2682 }
2683
2684 /*
2685  * Pending wakeups
2686  *
2687  * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2688  *
2689  * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2690  * single linked list and use cmpxchg() to add entries lockless.
2691  */
2692
2693 static void perf_pending_event(struct perf_pending_entry *entry)
2694 {
2695         struct perf_event *event = container_of(entry,
2696                         struct perf_event, pending);
2697
2698         if (event->pending_disable) {
2699                 event->pending_disable = 0;
2700                 __perf_event_disable(event);
2701         }
2702
2703         if (event->pending_wakeup) {
2704                 event->pending_wakeup = 0;
2705                 perf_event_wakeup(event);
2706         }
2707 }
2708
2709 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2710
2711 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2712         PENDING_TAIL,
2713 };
2714
2715 static void perf_pending_queue(struct perf_pending_entry *entry,
2716                                void (*func)(struct perf_pending_entry *))
2717 {
2718         struct perf_pending_entry **head;
2719
2720         if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2721                 return;
2722
2723         entry->func = func;
2724
2725         head = &get_cpu_var(perf_pending_head);
2726
2727         do {
2728                 entry->next = *head;
2729         } while (cmpxchg(head, entry->next, entry) != entry->next);
2730
2731         set_perf_event_pending();
2732
2733         put_cpu_var(perf_pending_head);
2734 }
2735
2736 static int __perf_pending_run(void)
2737 {
2738         struct perf_pending_entry *list;
2739         int nr = 0;
2740
2741         list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2742         while (list != PENDING_TAIL) {
2743                 void (*func)(struct perf_pending_entry *);
2744                 struct perf_pending_entry *entry = list;
2745
2746                 list = list->next;
2747
2748                 func = entry->func;
2749                 entry->next = NULL;
2750                 /*
2751                  * Ensure we observe the unqueue before we issue the wakeup,
2752                  * so that we won't be waiting forever.
2753                  * -- see perf_not_pending().
2754                  */
2755                 smp_wmb();
2756
2757                 func(entry);
2758                 nr++;
2759         }
2760
2761         return nr;
2762 }
2763
2764 static inline int perf_not_pending(struct perf_event *event)
2765 {
2766         /*
2767          * If we flush on whatever cpu we run, there is a chance we don't
2768          * need to wait.
2769          */
2770         get_cpu();
2771         __perf_pending_run();
2772         put_cpu();
2773
2774         /*
2775          * Ensure we see the proper queue state before going to sleep
2776          * so that we do not miss the wakeup. -- see perf_pending_handle()
2777          */
2778         smp_rmb();
2779         return event->pending.next == NULL;
2780 }
2781
2782 static void perf_pending_sync(struct perf_event *event)
2783 {
2784         wait_event(event->waitq, perf_not_pending(event));
2785 }
2786
2787 void perf_event_do_pending(void)
2788 {
2789         __perf_pending_run();
2790 }
2791
2792 /*
2793  * Callchain support -- arch specific
2794  */
2795
2796 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2797 {
2798         return NULL;
2799 }
2800
2801 /*
2802  * Output
2803  */
2804 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2805                               unsigned long offset, unsigned long head)
2806 {
2807         unsigned long mask;
2808
2809         if (!data->writable)
2810                 return true;
2811
2812         mask = perf_data_size(data) - 1;
2813
2814         offset = (offset - tail) & mask;
2815         head   = (head   - tail) & mask;
2816
2817         if ((int)(head - offset) < 0)
2818                 return false;
2819
2820         return true;
2821 }
2822
2823 static void perf_output_wakeup(struct perf_output_handle *handle)
2824 {
2825         atomic_set(&handle->data->poll, POLL_IN);
2826
2827         if (handle->nmi) {
2828                 handle->event->pending_wakeup = 1;
2829                 perf_pending_queue(&handle->event->pending,
2830                                    perf_pending_event);
2831         } else
2832                 perf_event_wakeup(handle->event);
2833 }
2834
2835 /*
2836  * Curious locking construct.
2837  *
2838  * We need to ensure a later event_id doesn't publish a head when a former
2839  * event_id isn't done writing. However since we need to deal with NMIs we
2840  * cannot fully serialize things.
2841  *
2842  * What we do is serialize between CPUs so we only have to deal with NMI
2843  * nesting on a single CPU.
2844  *
2845  * We only publish the head (and generate a wakeup) when the outer-most
2846  * event_id completes.
2847  */
2848 static void perf_output_lock(struct perf_output_handle *handle)
2849 {
2850         struct perf_mmap_data *data = handle->data;
2851         int cur, cpu = get_cpu();
2852
2853         handle->locked = 0;
2854
2855         for (;;) {
2856                 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2857                 if (cur == -1) {
2858                         handle->locked = 1;
2859                         break;
2860                 }
2861                 if (cur == cpu)
2862                         break;
2863
2864                 cpu_relax();
2865         }
2866 }
2867
2868 static void perf_output_unlock(struct perf_output_handle *handle)
2869 {
2870         struct perf_mmap_data *data = handle->data;
2871         unsigned long head;
2872         int cpu;
2873
2874         data->done_head = data->head;
2875
2876         if (!handle->locked)
2877                 goto out;
2878
2879 again:
2880         /*
2881          * The xchg implies a full barrier that ensures all writes are done
2882          * before we publish the new head, matched by a rmb() in userspace when
2883          * reading this position.
2884          */
2885         while ((head = atomic_long_xchg(&data->done_head, 0)))
2886                 data->user_page->data_head = head;
2887
2888         /*
2889          * NMI can happen here, which means we can miss a done_head update.
2890          */
2891
2892         cpu = atomic_xchg(&data->lock, -1);
2893         WARN_ON_ONCE(cpu != smp_processor_id());
2894
2895         /*
2896          * Therefore we have to validate we did not indeed do so.
2897          */
2898         if (unlikely(atomic_long_read(&data->done_head))) {
2899                 /*
2900                  * Since we had it locked, we can lock it again.
2901                  */
2902                 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2903                         cpu_relax();
2904
2905                 goto again;
2906         }
2907
2908         if (atomic_xchg(&data->wakeup, 0))
2909                 perf_output_wakeup(handle);
2910 out:
2911         put_cpu();
2912 }
2913
2914 void perf_output_copy(struct perf_output_handle *handle,
2915                       const void *buf, unsigned int len)
2916 {
2917         unsigned int pages_mask;
2918         unsigned long offset;
2919         unsigned int size;
2920         void **pages;
2921
2922         offset          = handle->offset;
2923         pages_mask      = handle->data->nr_pages - 1;
2924         pages           = handle->data->data_pages;
2925
2926         do {
2927                 unsigned long page_offset;
2928                 unsigned long page_size;
2929                 int nr;
2930
2931                 nr          = (offset >> PAGE_SHIFT) & pages_mask;
2932                 page_size   = 1UL << (handle->data->data_order + PAGE_SHIFT);
2933                 page_offset = offset & (page_size - 1);
2934                 size        = min_t(unsigned int, page_size - page_offset, len);
2935
2936                 memcpy(pages[nr] + page_offset, buf, size);
2937
2938                 len         -= size;
2939                 buf         += size;
2940                 offset      += size;
2941         } while (len);
2942
2943         handle->offset = offset;
2944
2945         /*
2946          * Check we didn't copy past our reservation window, taking the
2947          * possible unsigned int wrap into account.
2948          */
2949         WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2950 }
2951
2952 int perf_output_begin(struct perf_output_handle *handle,
2953                       struct perf_event *event, unsigned int size,
2954                       int nmi, int sample)
2955 {
2956         struct perf_event *output_event;
2957         struct perf_mmap_data *data;
2958         unsigned long tail, offset, head;
2959         int have_lost;
2960         struct {
2961                 struct perf_event_header header;
2962                 u64                      id;
2963                 u64                      lost;
2964         } lost_event;
2965
2966         rcu_read_lock();
2967         /*
2968          * For inherited events we send all the output towards the parent.
2969          */
2970         if (event->parent)
2971                 event = event->parent;
2972
2973         output_event = rcu_dereference(event->output);
2974         if (output_event)
2975                 event = output_event;
2976
2977         data = rcu_dereference(event->data);
2978         if (!data)
2979                 goto out;
2980
2981         handle->data    = data;
2982         handle->event   = event;
2983         handle->nmi     = nmi;
2984         handle->sample  = sample;
2985
2986         if (!data->nr_pages)
2987                 goto fail;
2988
2989         have_lost = atomic_read(&data->lost);
2990         if (have_lost)
2991                 size += sizeof(lost_event);
2992
2993         perf_output_lock(handle);
2994
2995         do {
2996                 /*
2997                  * Userspace could choose to issue a mb() before updating the
2998                  * tail pointer. So that all reads will be completed before the
2999                  * write is issued.
3000                  */
3001                 tail = ACCESS_ONCE(data->user_page->data_tail);
3002                 smp_rmb();
3003                 offset = head = atomic_long_read(&data->head);
3004                 head += size;
3005                 if (unlikely(!perf_output_space(data, tail, offset, head)))
3006                         goto fail;
3007         } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
3008
3009         handle->offset  = offset;
3010         handle->head    = head;
3011
3012         if (head - tail > data->watermark)
3013                 atomic_set(&data->wakeup, 1);
3014
3015         if (have_lost) {
3016                 lost_event.header.type = PERF_RECORD_LOST;
3017                 lost_event.header.misc = 0;
3018                 lost_event.header.size = sizeof(lost_event);
3019                 lost_event.id          = event->id;
3020                 lost_event.lost        = atomic_xchg(&data->lost, 0);
3021
3022                 perf_output_put(handle, lost_event);
3023         }
3024
3025         return 0;
3026
3027 fail:
3028         atomic_inc(&data->lost);
3029         perf_output_unlock(handle);
3030 out:
3031         rcu_read_unlock();
3032
3033         return -ENOSPC;
3034 }
3035
3036 void perf_output_end(struct perf_output_handle *handle)
3037 {
3038         struct perf_event *event = handle->event;
3039         struct perf_mmap_data *data = handle->data;
3040
3041         int wakeup_events = event->attr.wakeup_events;
3042
3043         if (handle->sample && wakeup_events) {
3044                 int events = atomic_inc_return(&data->events);
3045                 if (events >= wakeup_events) {
3046                         atomic_sub(wakeup_events, &data->events);
3047                         atomic_set(&data->wakeup, 1);
3048                 }
3049         }
3050
3051         perf_output_unlock(handle);
3052         rcu_read_unlock();
3053 }
3054
3055 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3056 {
3057         /*
3058          * only top level events have the pid namespace they were created in
3059          */
3060         if (event->parent)
3061                 event = event->parent;
3062
3063         return task_tgid_nr_ns(p, event->ns);
3064 }
3065
3066 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3067 {
3068         /*
3069          * only top level events have the pid namespace they were created in
3070          */
3071         if (event->parent)
3072                 event = event->parent;
3073
3074         return task_pid_nr_ns(p, event->ns);
3075 }
3076
3077 static void perf_output_read_one(struct perf_output_handle *handle,
3078                                  struct perf_event *event)
3079 {
3080         u64 read_format = event->attr.read_format;
3081         u64 values[4];
3082         int n = 0;
3083
3084         values[n++] = atomic64_read(&event->count);
3085         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3086                 values[n++] = event->total_time_enabled +
3087                         atomic64_read(&event->child_total_time_enabled);
3088         }
3089         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3090                 values[n++] = event->total_time_running +
3091                         atomic64_read(&event->child_total_time_running);
3092         }
3093         if (read_format & PERF_FORMAT_ID)
3094                 values[n++] = primary_event_id(event);
3095
3096         perf_output_copy(handle, values, n * sizeof(u64));
3097 }
3098
3099 /*
3100  * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3101  */
3102 static void perf_output_read_group(struct perf_output_handle *handle,
3103                             struct perf_event *event)
3104 {
3105         struct perf_event *leader = event->group_leader, *sub;
3106         u64 read_format = event->attr.read_format;
3107         u64 values[5];
3108         int n = 0;
3109
3110         values[n++] = 1 + leader->nr_siblings;
3111
3112         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3113                 values[n++] = leader->total_time_enabled;
3114
3115         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3116                 values[n++] = leader->total_time_running;
3117
3118         if (leader != event)
3119                 leader->pmu->read(leader);
3120
3121         values[n++] = atomic64_read(&leader->count);
3122         if (read_format & PERF_FORMAT_ID)
3123                 values[n++] = primary_event_id(leader);
3124
3125         perf_output_copy(handle, values, n * sizeof(u64));
3126
3127         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3128                 n = 0;
3129
3130                 if (sub != event)
3131                         sub->pmu->read(sub);
3132
3133                 values[n++] = atomic64_read(&sub->count);
3134                 if (read_format & PERF_FORMAT_ID)
3135                         values[n++] = primary_event_id(sub);
3136
3137                 perf_output_copy(handle, values, n * sizeof(u64));
3138         }
3139 }
3140
3141 static void perf_output_read(struct perf_output_handle *handle,
3142                              struct perf_event *event)
3143 {
3144         if (event->attr.read_format & PERF_FORMAT_GROUP)
3145                 perf_output_read_group(handle, event);
3146         else
3147                 perf_output_read_one(handle, event);
3148 }
3149
3150 void perf_output_sample(struct perf_output_handle *handle,
3151                         struct perf_event_header *header,
3152                         struct perf_sample_data *data,
3153                         struct perf_event *event)
3154 {
3155         u64 sample_type = data->type;
3156
3157         perf_output_put(handle, *header);
3158
3159         if (sample_type & PERF_SAMPLE_IP)
3160                 perf_output_put(handle, data->ip);
3161
3162         if (sample_type & PERF_SAMPLE_TID)
3163                 perf_output_put(handle, data->tid_entry);
3164
3165         if (sample_type & PERF_SAMPLE_TIME)
3166                 perf_output_put(handle, data->time);
3167
3168         if (sample_type & PERF_SAMPLE_ADDR)
3169                 perf_output_put(handle, data->addr);
3170
3171         if (sample_type & PERF_SAMPLE_ID)
3172                 perf_output_put(handle, data->id);
3173
3174         if (sample_type & PERF_SAMPLE_STREAM_ID)
3175                 perf_output_put(handle, data->stream_id);
3176
3177         if (sample_type & PERF_SAMPLE_CPU)
3178                 perf_output_put(handle, data->cpu_entry);
3179
3180         if (sample_type & PERF_SAMPLE_PERIOD)
3181                 perf_output_put(handle, data->period);
3182
3183         if (sample_type & PERF_SAMPLE_READ)
3184                 perf_output_read(handle, event);
3185
3186         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3187                 if (data->callchain) {
3188                         int size = 1;
3189
3190                         if (data->callchain)
3191                                 size += data->callchain->nr;
3192
3193                         size *= sizeof(u64);
3194
3195                         perf_output_copy(handle, data->callchain, size);
3196                 } else {
3197                         u64 nr = 0;
3198                         perf_output_put(handle, nr);
3199                 }
3200         }
3201
3202         if (sample_type & PERF_SAMPLE_RAW) {
3203                 if (data->raw) {
3204                         perf_output_put(handle, data->raw->size);
3205                         perf_output_copy(handle, data->raw->data,
3206                                          data->raw->size);
3207                 } else {
3208                         struct {
3209                                 u32     size;
3210                                 u32     data;
3211                         } raw = {
3212                                 .size = sizeof(u32),
3213                                 .data = 0,
3214                         };
3215                         perf_output_put(handle, raw);
3216                 }
3217         }
3218 }
3219
3220 void perf_prepare_sample(struct perf_event_header *header,
3221                          struct perf_sample_data *data,
3222                          struct perf_event *event,
3223                          struct pt_regs *regs)
3224 {
3225         u64 sample_type = event->attr.sample_type;
3226
3227         data->type = sample_type;
3228
3229         header->type = PERF_RECORD_SAMPLE;
3230         header->size = sizeof(*header);
3231
3232         header->misc = 0;
3233         header->misc |= perf_misc_flags(regs);
3234
3235         if (sample_type & PERF_SAMPLE_IP) {
3236                 data->ip = perf_instruction_pointer(regs);
3237
3238                 header->size += sizeof(data->ip);
3239         }
3240
3241         if (sample_type & PERF_SAMPLE_TID) {
3242                 /* namespace issues */
3243                 data->tid_entry.pid = perf_event_pid(event, current);
3244                 data->tid_entry.tid = perf_event_tid(event, current);
3245
3246                 header->size += sizeof(data->tid_entry);
3247         }
3248
3249         if (sample_type & PERF_SAMPLE_TIME) {
3250                 data->time = perf_clock();
3251
3252                 header->size += sizeof(data->time);
3253         }
3254
3255         if (sample_type & PERF_SAMPLE_ADDR)
3256                 header->size += sizeof(data->addr);
3257
3258         if (sample_type & PERF_SAMPLE_ID) {
3259                 data->id = primary_event_id(event);
3260
3261                 header->size += sizeof(data->id);
3262         }
3263
3264         if (sample_type & PERF_SAMPLE_STREAM_ID) {
3265                 data->stream_id = event->id;
3266
3267                 header->size += sizeof(data->stream_id);
3268         }
3269
3270         if (sample_type & PERF_SAMPLE_CPU) {
3271                 data->cpu_entry.cpu             = raw_smp_processor_id();
3272                 data->cpu_entry.reserved        = 0;
3273
3274                 header->size += sizeof(data->cpu_entry);
3275         }
3276
3277         if (sample_type & PERF_SAMPLE_PERIOD)
3278                 header->size += sizeof(data->period);
3279
3280         if (sample_type & PERF_SAMPLE_READ)
3281                 header->size += perf_event_read_size(event);
3282
3283         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3284                 int size = 1;
3285
3286                 data->callchain = perf_callchain(regs);
3287
3288                 if (data->callchain)
3289                         size += data->callchain->nr;
3290
3291                 header->size += size * sizeof(u64);
3292         }
3293
3294         if (sample_type & PERF_SAMPLE_RAW) {
3295                 int size = sizeof(u32);
3296
3297                 if (data->raw)
3298                         size += data->raw->size;
3299                 else
3300                         size += sizeof(u32);
3301
3302                 WARN_ON_ONCE(size & (sizeof(u64)-1));
3303                 header->size += size;
3304         }
3305 }
3306
3307 static void perf_event_output(struct perf_event *event, int nmi,
3308                                 struct perf_sample_data *data,
3309                                 struct pt_regs *regs)
3310 {
3311         struct perf_output_handle handle;
3312         struct perf_event_header header;
3313
3314         perf_prepare_sample(&header, data, event, regs);
3315
3316         if (perf_output_begin(&handle, event, header.size, nmi, 1))
3317                 return;
3318
3319         perf_output_sample(&handle, &header, data, event);
3320
3321         perf_output_end(&handle);
3322 }
3323
3324 /*
3325  * read event_id
3326  */
3327
3328 struct perf_read_event {
3329         struct perf_event_header        header;
3330
3331         u32                             pid;
3332         u32                             tid;
3333 };
3334
3335 static void
3336 perf_event_read_event(struct perf_event *event,
3337                         struct task_struct *task)
3338 {
3339         struct perf_output_handle handle;
3340         struct perf_read_event read_event = {
3341                 .header = {
3342                         .type = PERF_RECORD_READ,
3343                         .misc = 0,
3344                         .size = sizeof(read_event) + perf_event_read_size(event),
3345                 },
3346                 .pid = perf_event_pid(event, task),
3347                 .tid = perf_event_tid(event, task),
3348         };
3349         int ret;
3350
3351         ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3352         if (ret)
3353                 return;
3354
3355         perf_output_put(&handle, read_event);
3356         perf_output_read(&handle, event);
3357
3358         perf_output_end(&handle);
3359 }
3360
3361 /*
3362  * task tracking -- fork/exit
3363  *
3364  * enabled by: attr.comm | attr.mmap | attr.task
3365  */
3366
3367 struct perf_task_event {
3368         struct task_struct              *task;
3369         struct perf_event_context       *task_ctx;
3370
3371         struct {
3372                 struct perf_event_header        header;
3373
3374                 u32                             pid;
3375                 u32                             ppid;
3376                 u32                             tid;
3377                 u32                             ptid;
3378                 u64                             time;
3379         } event_id;
3380 };
3381
3382 static void perf_event_task_output(struct perf_event *event,
3383                                      struct perf_task_event *task_event)
3384 {
3385         struct perf_output_handle handle;
3386         int size;
3387         struct task_struct *task = task_event->task;
3388         int ret;
3389
3390         size  = task_event->event_id.header.size;
3391         ret = perf_output_begin(&handle, event, size, 0, 0);
3392
3393         if (ret)
3394                 return;
3395
3396         task_event->event_id.pid = perf_event_pid(event, task);
3397         task_event->event_id.ppid = perf_event_pid(event, current);
3398
3399         task_event->event_id.tid = perf_event_tid(event, task);
3400         task_event->event_id.ptid = perf_event_tid(event, current);
3401
3402         task_event->event_id.time = perf_clock();
3403
3404         perf_output_put(&handle, task_event->event_id);
3405
3406         perf_output_end(&handle);
3407 }
3408
3409 static int perf_event_task_match(struct perf_event *event)
3410 {
3411         if (event->state != PERF_EVENT_STATE_ACTIVE)
3412                 return 0;
3413
3414         if (event->cpu != -1 && event->cpu != smp_processor_id())
3415                 return 0;
3416
3417         if (event->attr.comm || event->attr.mmap || event->attr.task)
3418                 return 1;
3419
3420         return 0;
3421 }
3422
3423 static void perf_event_task_ctx(struct perf_event_context *ctx,
3424                                   struct perf_task_event *task_event)
3425 {
3426         struct perf_event *event;
3427
3428         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3429                 if (perf_event_task_match(event))
3430                         perf_event_task_output(event, task_event);
3431         }
3432 }
3433
3434 static void perf_event_task_event(struct perf_task_event *task_event)
3435 {
3436         struct perf_cpu_context *cpuctx;
3437         struct perf_event_context *ctx = task_event->task_ctx;
3438
3439         rcu_read_lock();
3440         cpuctx = &get_cpu_var(perf_cpu_context);
3441         perf_event_task_ctx(&cpuctx->ctx, task_event);
3442         if (!ctx)
3443                 ctx = rcu_dereference(task_event->task->perf_event_ctxp);
3444         if (ctx)
3445                 perf_event_task_ctx(ctx, task_event);
3446         put_cpu_var(perf_cpu_context);
3447         rcu_read_unlock();
3448 }
3449
3450 static void perf_event_task(struct task_struct *task,
3451                               struct perf_event_context *task_ctx,
3452                               int new)
3453 {
3454         struct perf_task_event task_event;
3455
3456         if (!atomic_read(&nr_comm_events) &&
3457             !atomic_read(&nr_mmap_events) &&
3458             !atomic_read(&nr_task_events))
3459                 return;
3460
3461         task_event = (struct perf_task_event){
3462                 .task     = task,
3463                 .task_ctx = task_ctx,
3464                 .event_id    = {
3465                         .header = {
3466                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3467                                 .misc = 0,
3468                                 .size = sizeof(task_event.event_id),
3469                         },
3470                         /* .pid  */
3471                         /* .ppid */
3472                         /* .tid  */
3473                         /* .ptid */
3474                 },
3475         };
3476
3477         perf_event_task_event(&task_event);
3478 }
3479
3480 void perf_event_fork(struct task_struct *task)
3481 {
3482         perf_event_task(task, NULL, 1);
3483 }
3484
3485 /*
3486  * comm tracking
3487  */
3488
3489 struct perf_comm_event {
3490         struct task_struct      *task;
3491         char                    *comm;
3492         int                     comm_size;
3493
3494         struct {
3495                 struct perf_event_header        header;
3496
3497                 u32                             pid;
3498                 u32                             tid;
3499         } event_id;
3500 };
3501
3502 static void perf_event_comm_output(struct perf_event *event,
3503                                      struct perf_comm_event *comm_event)
3504 {
3505         struct perf_output_handle handle;
3506         int size = comm_event->event_id.header.size;
3507         int ret = perf_output_begin(&handle, event, size, 0, 0);
3508
3509         if (ret)
3510                 return;
3511
3512         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3513         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3514
3515         perf_output_put(&handle, comm_event->event_id);
3516         perf_output_copy(&handle, comm_event->comm,
3517                                    comm_event->comm_size);
3518         perf_output_end(&handle);
3519 }
3520
3521 static int perf_event_comm_match(struct perf_event *event)
3522 {
3523         if (event->state != PERF_EVENT_STATE_ACTIVE)
3524                 return 0;
3525
3526         if (event->cpu != -1 && event->cpu != smp_processor_id())
3527                 return 0;
3528
3529         if (event->attr.comm)
3530                 return 1;
3531
3532         return 0;
3533 }
3534
3535 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3536                                   struct perf_comm_event *comm_event)
3537 {
3538         struct perf_event *event;
3539
3540         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3541                 if (perf_event_comm_match(event))
3542                         perf_event_comm_output(event, comm_event);
3543         }
3544 }
3545
3546 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3547 {
3548         struct perf_cpu_context *cpuctx;
3549         struct perf_event_context *ctx;
3550         unsigned int size;
3551         char comm[TASK_COMM_LEN];
3552
3553         memset(comm, 0, sizeof(comm));
3554         strlcpy(comm, comm_event->task->comm, sizeof(comm));
3555         size = ALIGN(strlen(comm)+1, sizeof(u64));
3556
3557         comm_event->comm = comm;
3558         comm_event->comm_size = size;
3559
3560         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3561
3562         rcu_read_lock();
3563         cpuctx = &get_cpu_var(perf_cpu_context);
3564         perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3565         ctx = rcu_dereference(current->perf_event_ctxp);
3566         if (ctx)
3567                 perf_event_comm_ctx(ctx, comm_event);
3568         put_cpu_var(perf_cpu_context);
3569         rcu_read_unlock();
3570 }
3571
3572 void perf_event_comm(struct task_struct *task)
3573 {
3574         struct perf_comm_event comm_event;
3575
3576         if (task->perf_event_ctxp)
3577                 perf_event_enable_on_exec(task);
3578
3579         if (!atomic_read(&nr_comm_events))
3580                 return;
3581
3582         comm_event = (struct perf_comm_event){
3583                 .task   = task,
3584                 /* .comm      */
3585                 /* .comm_size */
3586                 .event_id  = {
3587                         .header = {
3588                                 .type = PERF_RECORD_COMM,
3589                                 .misc = 0,
3590                                 /* .size */
3591                         },
3592                         /* .pid */
3593                         /* .tid */
3594                 },
3595         };
3596
3597         perf_event_comm_event(&comm_event);
3598 }
3599
3600 /*
3601  * mmap tracking
3602  */
3603
3604 struct perf_mmap_event {
3605         struct vm_area_struct   *vma;
3606
3607         const char              *file_name;
3608         int                     file_size;
3609
3610         struct {
3611                 struct perf_event_header        header;
3612
3613                 u32                             pid;
3614                 u32                             tid;
3615                 u64                             start;
3616                 u64                             len;
3617                 u64                             pgoff;
3618         } event_id;
3619 };
3620
3621 static void perf_event_mmap_output(struct perf_event *event,
3622                                      struct perf_mmap_event *mmap_event)
3623 {
3624         struct perf_output_handle handle;
3625         int size = mmap_event->event_id.header.size;
3626         int ret = perf_output_begin(&handle, event, size, 0, 0);
3627
3628         if (ret)
3629                 return;
3630
3631         mmap_event->event_id.pid = perf_event_pid(event, current);
3632         mmap_event->event_id.tid = perf_event_tid(event, current);
3633
3634         perf_output_put(&handle, mmap_event->event_id);
3635         perf_output_copy(&handle, mmap_event->file_name,
3636                                    mmap_event->file_size);
3637         perf_output_end(&handle);
3638 }
3639
3640 static int perf_event_mmap_match(struct perf_event *event,
3641                                    struct perf_mmap_event *mmap_event)
3642 {
3643         if (event->state != PERF_EVENT_STATE_ACTIVE)
3644                 return 0;
3645
3646         if (event->cpu != -1 && event->cpu != smp_processor_id())
3647                 return 0;
3648
3649         if (event->attr.mmap)
3650                 return 1;
3651
3652         return 0;
3653 }
3654
3655 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3656                                   struct perf_mmap_event *mmap_event)
3657 {
3658         struct perf_event *event;
3659
3660         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3661                 if (perf_event_mmap_match(event, mmap_event))
3662                         perf_event_mmap_output(event, mmap_event);
3663         }
3664 }
3665
3666 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3667 {
3668         struct perf_cpu_context *cpuctx;
3669         struct perf_event_context *ctx;
3670         struct vm_area_struct *vma = mmap_event->vma;
3671         struct file *file = vma->vm_file;
3672         unsigned int size;
3673         char tmp[16];
3674         char *buf = NULL;
3675         const char *name;
3676
3677         memset(tmp, 0, sizeof(tmp));
3678
3679         if (file) {
3680                 /*
3681                  * d_path works from the end of the buffer backwards, so we
3682                  * need to add enough zero bytes after the string to handle
3683                  * the 64bit alignment we do later.
3684                  */
3685                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3686                 if (!buf) {
3687                         name = strncpy(tmp, "//enomem", sizeof(tmp));
3688                         goto got_name;
3689                 }
3690                 name = d_path(&file->f_path, buf, PATH_MAX);
3691                 if (IS_ERR(name)) {
3692                         name = strncpy(tmp, "//toolong", sizeof(tmp));
3693                         goto got_name;
3694                 }
3695         } else {
3696                 if (arch_vma_name(mmap_event->vma)) {
3697                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3698                                        sizeof(tmp));
3699                         goto got_name;
3700                 }
3701
3702                 if (!vma->vm_mm) {
3703                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
3704                         goto got_name;
3705                 }
3706
3707                 name = strncpy(tmp, "//anon", sizeof(tmp));
3708                 goto got_name;
3709         }
3710
3711 got_name:
3712         size = ALIGN(strlen(name)+1, sizeof(u64));
3713
3714         mmap_event->file_name = name;
3715         mmap_event->file_size = size;
3716
3717         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3718
3719         rcu_read_lock();
3720         cpuctx = &get_cpu_var(perf_cpu_context);
3721         perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3722         ctx = rcu_dereference(current->perf_event_ctxp);
3723         if (ctx)
3724                 perf_event_mmap_ctx(ctx, mmap_event);
3725         put_cpu_var(perf_cpu_context);
3726         rcu_read_unlock();
3727
3728         kfree(buf);
3729 }
3730
3731 void __perf_event_mmap(struct vm_area_struct *vma)
3732 {
3733         struct perf_mmap_event mmap_event;
3734
3735         if (!atomic_read(&nr_mmap_events))
3736                 return;
3737
3738         mmap_event = (struct perf_mmap_event){
3739                 .vma    = vma,
3740                 /* .file_name */
3741                 /* .file_size */
3742                 .event_id  = {
3743                         .header = {
3744                                 .type = PERF_RECORD_MMAP,
3745                                 .misc = 0,
3746                                 /* .size */
3747                         },
3748                         /* .pid */
3749                         /* .tid */
3750                         .start  = vma->vm_start,
3751                         .len    = vma->vm_end - vma->vm_start,
3752                         .pgoff  = (u64)vma->vm_pgoff << PAGE_SHIFT,
3753                 },
3754         };
3755
3756         perf_event_mmap_event(&mmap_event);
3757 }
3758
3759 /*
3760  * IRQ throttle logging
3761  */
3762
3763 static void perf_log_throttle(struct perf_event *event, int enable)
3764 {
3765         struct perf_output_handle handle;
3766         int ret;
3767
3768         struct {
3769                 struct perf_event_header        header;
3770                 u64                             time;
3771                 u64                             id;
3772                 u64                             stream_id;
3773         } throttle_event = {
3774                 .header = {
3775                         .type = PERF_RECORD_THROTTLE,
3776                         .misc = 0,
3777                         .size = sizeof(throttle_event),
3778                 },
3779                 .time           = perf_clock(),
3780                 .id             = primary_event_id(event),
3781                 .stream_id      = event->id,
3782         };
3783
3784         if (enable)
3785                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3786
3787         ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3788         if (ret)
3789                 return;
3790
3791         perf_output_put(&handle, throttle_event);
3792         perf_output_end(&handle);
3793 }
3794
3795 /*
3796  * Generic event overflow handling, sampling.
3797  */
3798
3799 static int __perf_event_overflow(struct perf_event *event, int nmi,
3800                                    int throttle, struct perf_sample_data *data,
3801                                    struct pt_regs *regs)
3802 {
3803         int events = atomic_read(&event->event_limit);
3804         struct hw_perf_event *hwc = &event->hw;
3805         int ret = 0;
3806
3807         throttle = (throttle && event->pmu->unthrottle != NULL);
3808
3809         if (!throttle) {
3810                 hwc->interrupts++;
3811         } else {
3812                 if (hwc->interrupts != MAX_INTERRUPTS) {
3813                         hwc->interrupts++;
3814                         if (HZ * hwc->interrupts >
3815                                         (u64)sysctl_perf_event_sample_rate) {
3816                                 hwc->interrupts = MAX_INTERRUPTS;
3817                                 perf_log_throttle(event, 0);
3818                                 ret = 1;
3819                         }
3820                 } else {
3821                         /*
3822                          * Keep re-disabling events even though on the previous
3823                          * pass we disabled it - just in case we raced with a
3824                          * sched-in and the event got enabled again:
3825                          */
3826                         ret = 1;
3827                 }
3828         }
3829
3830         if (event->attr.freq) {
3831                 u64 now = perf_clock();
3832                 s64 delta = now - hwc->freq_time_stamp;
3833
3834                 hwc->freq_time_stamp = now;
3835
3836                 if (delta > 0 && delta < 2*TICK_NSEC)
3837                         perf_adjust_period(event, delta, hwc->last_period);
3838         }
3839
3840         /*
3841          * XXX event_limit might not quite work as expected on inherited
3842          * events
3843          */
3844
3845         event->pending_kill = POLL_IN;
3846         if (events && atomic_dec_and_test(&event->event_limit)) {
3847                 ret = 1;
3848                 event->pending_kill = POLL_HUP;
3849                 if (nmi) {
3850                         event->pending_disable = 1;
3851                         perf_pending_queue(&event->pending,
3852                                            perf_pending_event);
3853                 } else
3854                         perf_event_disable(event);
3855         }
3856
3857         if (event->overflow_handler)
3858                 event->overflow_handler(event, nmi, data, regs);
3859         else
3860                 perf_event_output(event, nmi, data, regs);
3861
3862         return ret;
3863 }
3864
3865 int perf_event_overflow(struct perf_event *event, int nmi,
3866                           struct perf_sample_data *data,
3867                           struct pt_regs *regs)
3868 {
3869         return __perf_event_overflow(event, nmi, 1, data, regs);
3870 }
3871
3872 /*
3873  * Generic software event infrastructure
3874  */
3875
3876 /*
3877  * We directly increment event->count and keep a second value in
3878  * event->hw.period_left to count intervals. This period event
3879  * is kept in the range [-sample_period, 0] so that we can use the
3880  * sign as trigger.
3881  */
3882
3883 static u64 perf_swevent_set_period(struct perf_event *event)
3884 {
3885         struct hw_perf_event *hwc = &event->hw;
3886         u64 period = hwc->last_period;
3887         u64 nr, offset;
3888         s64 old, val;
3889
3890         hwc->last_period = hwc->sample_period;
3891
3892 again:
3893         old = val = atomic64_read(&hwc->period_left);
3894         if (val < 0)
3895                 return 0;
3896
3897         nr = div64_u64(period + val, period);
3898         offset = nr * period;
3899         val -= offset;
3900         if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3901                 goto again;
3902
3903         return nr;
3904 }
3905
3906 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3907                                     int nmi, struct perf_sample_data *data,
3908                                     struct pt_regs *regs)
3909 {
3910         struct hw_perf_event *hwc = &event->hw;
3911         int throttle = 0;
3912
3913         data->period = event->hw.last_period;
3914         if (!overflow)
3915                 overflow = perf_swevent_set_period(event);
3916
3917         if (hwc->interrupts == MAX_INTERRUPTS)
3918                 return;
3919
3920         for (; overflow; overflow--) {
3921                 if (__perf_event_overflow(event, nmi, throttle,
3922                                             data, regs)) {
3923                         /*
3924                          * We inhibit the overflow from happening when
3925                          * hwc->interrupts == MAX_INTERRUPTS.
3926                          */
3927                         break;
3928                 }
3929                 throttle = 1;
3930         }
3931 }
3932
3933 static void perf_swevent_unthrottle(struct perf_event *event)
3934 {
3935         /*
3936          * Nothing to do, we already reset hwc->interrupts.
3937          */
3938 }
3939
3940 static void perf_swevent_add(struct perf_event *event, u64 nr,
3941                                int nmi, struct perf_sample_data *data,
3942                                struct pt_regs *regs)
3943 {
3944         struct hw_perf_event *hwc = &event->hw;
3945
3946         atomic64_add(nr, &event->count);
3947
3948         if (!regs)
3949                 return;
3950
3951         if (!hwc->sample_period)
3952                 return;
3953
3954         if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3955                 return perf_swevent_overflow(event, 1, nmi, data, regs);
3956
3957         if (atomic64_add_negative(nr, &hwc->period_left))
3958                 return;
3959
3960         perf_swevent_overflow(event, 0, nmi, data, regs);
3961 }
3962
3963 static int perf_swevent_is_counting(struct perf_event *event)
3964 {
3965         /*
3966          * The event is active, we're good!
3967          */
3968         if (event->state == PERF_EVENT_STATE_ACTIVE)
3969                 return 1;
3970
3971         /*
3972          * The event is off/error, not counting.
3973          */
3974         if (event->state != PERF_EVENT_STATE_INACTIVE)
3975                 return 0;
3976
3977         /*
3978          * The event is inactive, if the context is active
3979          * we're part of a group that didn't make it on the 'pmu',
3980          * not counting.
3981          */
3982         if (event->ctx->is_active)
3983                 return 0;
3984
3985         /*
3986          * We're inactive and the context is too, this means the
3987          * task is scheduled out, we're counting events that happen
3988          * to us, like migration events.
3989          */
3990         return 1;
3991 }
3992
3993 static int perf_tp_event_match(struct perf_event *event,
3994                                 struct perf_sample_data *data);
3995
3996 static int perf_exclude_event(struct perf_event *event,
3997                               struct pt_regs *regs)
3998 {
3999         if (regs) {
4000                 if (event->attr.exclude_user && user_mode(regs))
4001                         return 1;
4002
4003                 if (event->attr.exclude_kernel && !user_mode(regs))
4004                         return 1;
4005         }
4006
4007         return 0;
4008 }
4009
4010 static int perf_swevent_match(struct perf_event *event,
4011                                 enum perf_type_id type,
4012                                 u32 event_id,
4013                                 struct perf_sample_data *data,
4014                                 struct pt_regs *regs)
4015 {
4016         if (event->cpu != -1 && event->cpu != smp_processor_id())
4017                 return 0;
4018
4019         if (!perf_swevent_is_counting(event))
4020                 return 0;
4021
4022         if (event->attr.type != type)
4023                 return 0;
4024
4025         if (event->attr.config != event_id)
4026                 return 0;
4027
4028         if (perf_exclude_event(event, regs))
4029                 return 0;
4030
4031         if (event->attr.type == PERF_TYPE_TRACEPOINT &&
4032             !perf_tp_event_match(event, data))
4033                 return 0;
4034
4035         return 1;
4036 }
4037
4038 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
4039                                      enum perf_type_id type,
4040                                      u32 event_id, u64 nr, int nmi,
4041                                      struct perf_sample_data *data,
4042                                      struct pt_regs *regs)
4043 {
4044         struct perf_event *event;
4045
4046         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4047                 if (perf_swevent_match(event, type, event_id, data, regs))
4048                         perf_swevent_add(event, nr, nmi, data, regs);
4049         }
4050 }
4051
4052 int perf_swevent_get_recursion_context(void)
4053 {
4054         struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
4055         int rctx;
4056
4057         if (in_nmi())
4058                 rctx = 3;
4059         else if (in_irq())
4060                 rctx = 2;
4061         else if (in_softirq())
4062                 rctx = 1;
4063         else
4064                 rctx = 0;
4065
4066         if (cpuctx->recursion[rctx]) {
4067                 put_cpu_var(perf_cpu_context);
4068                 return -1;
4069         }
4070
4071         cpuctx->recursion[rctx]++;
4072         barrier();
4073
4074         return rctx;
4075 }
4076 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4077
4078 void perf_swevent_put_recursion_context(int rctx)
4079 {
4080         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4081         barrier();
4082         cpuctx->recursion[rctx]--;
4083         put_cpu_var(perf_cpu_context);
4084 }
4085 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
4086
4087 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4088                                     u64 nr, int nmi,
4089                                     struct perf_sample_data *data,
4090                                     struct pt_regs *regs)
4091 {
4092         struct perf_cpu_context *cpuctx;
4093         struct perf_event_context *ctx;
4094
4095         cpuctx = &__get_cpu_var(perf_cpu_context);
4096         rcu_read_lock();
4097         perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
4098                                  nr, nmi, data, regs);
4099         /*
4100          * doesn't really matter which of the child contexts the
4101          * events ends up in.
4102          */
4103         ctx = rcu_dereference(current->perf_event_ctxp);
4104         if (ctx)
4105                 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
4106         rcu_read_unlock();
4107 }
4108
4109 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4110                             struct pt_regs *regs, u64 addr)
4111 {
4112         struct perf_sample_data data;
4113         int rctx;
4114
4115         rctx = perf_swevent_get_recursion_context();
4116         if (rctx < 0)
4117                 return;
4118
4119         data.addr = addr;
4120         data.raw  = NULL;
4121
4122         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4123
4124         perf_swevent_put_recursion_context(rctx);
4125 }
4126
4127 static void perf_swevent_read(struct perf_event *event)
4128 {
4129 }
4130
4131 static int perf_swevent_enable(struct perf_event *event)
4132 {
4133         struct hw_perf_event *hwc = &event->hw;
4134
4135         if (hwc->sample_period) {
4136                 hwc->last_period = hwc->sample_period;
4137                 perf_swevent_set_period(event);
4138         }
4139         return 0;
4140 }
4141
4142 static void perf_swevent_disable(struct perf_event *event)
4143 {
4144 }
4145
4146 static const struct pmu perf_ops_generic = {
4147         .enable         = perf_swevent_enable,
4148         .disable        = perf_swevent_disable,
4149         .read           = perf_swevent_read,
4150         .unthrottle     = perf_swevent_unthrottle,
4151 };
4152
4153 /*
4154  * hrtimer based swevent callback
4155  */
4156
4157 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4158 {
4159         enum hrtimer_restart ret = HRTIMER_RESTART;
4160         struct perf_sample_data data;
4161         struct pt_regs *regs;
4162         struct perf_event *event;
4163         u64 period;
4164
4165         event   = container_of(hrtimer, struct perf_event, hw.hrtimer);
4166         event->pmu->read(event);
4167
4168         data.addr = 0;
4169         data.raw = NULL;
4170         data.period = event->hw.last_period;
4171         regs = get_irq_regs();
4172         /*
4173          * In case we exclude kernel IPs or are somehow not in interrupt
4174          * context, provide the next best thing, the user IP.
4175          */
4176         if ((event->attr.exclude_kernel || !regs) &&
4177                         !event->attr.exclude_user)
4178                 regs = task_pt_regs(current);
4179
4180         if (regs) {
4181                 if (!(event->attr.exclude_idle && current->pid == 0))
4182                         if (perf_event_overflow(event, 0, &data, regs))
4183                                 ret = HRTIMER_NORESTART;
4184         }
4185
4186         period = max_t(u64, 10000, event->hw.sample_period);
4187         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4188
4189         return ret;
4190 }
4191
4192 static void perf_swevent_start_hrtimer(struct perf_event *event)
4193 {
4194         struct hw_perf_event *hwc = &event->hw;
4195
4196         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4197         hwc->hrtimer.function = perf_swevent_hrtimer;
4198         if (hwc->sample_period) {
4199                 u64 period;
4200
4201                 if (hwc->remaining) {
4202                         if (hwc->remaining < 0)
4203                                 period = 10000;
4204                         else
4205                                 period = hwc->remaining;
4206                         hwc->remaining = 0;
4207                 } else {
4208                         period = max_t(u64, 10000, hwc->sample_period);
4209                 }
4210                 __hrtimer_start_range_ns(&hwc->hrtimer,
4211                                 ns_to_ktime(period), 0,
4212                                 HRTIMER_MODE_REL, 0);
4213         }
4214 }
4215
4216 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4217 {
4218         struct hw_perf_event *hwc = &event->hw;
4219
4220         if (hwc->sample_period) {
4221                 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4222                 hwc->remaining = ktime_to_ns(remaining);
4223
4224                 hrtimer_cancel(&hwc->hrtimer);
4225         }
4226 }
4227
4228 /*
4229  * Software event: cpu wall time clock
4230  */
4231
4232 static void cpu_clock_perf_event_update(struct perf_event *event)
4233 {
4234         int cpu = raw_smp_processor_id();
4235         s64 prev;
4236         u64 now;
4237
4238         now = cpu_clock(cpu);
4239         prev = atomic64_xchg(&event->hw.prev_count, now);
4240         atomic64_add(now - prev, &event->count);
4241 }
4242
4243 static int cpu_clock_perf_event_enable(struct perf_event *event)
4244 {
4245         struct hw_perf_event *hwc = &event->hw;
4246         int cpu = raw_smp_processor_id();
4247
4248         atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4249         perf_swevent_start_hrtimer(event);
4250
4251         return 0;
4252 }
4253
4254 static void cpu_clock_perf_event_disable(struct perf_event *event)
4255 {
4256         perf_swevent_cancel_hrtimer(event);
4257         cpu_clock_perf_event_update(event);
4258 }
4259
4260 static void cpu_clock_perf_event_read(struct perf_event *event)
4261 {
4262         cpu_clock_perf_event_update(event);
4263 }
4264
4265 static const struct pmu perf_ops_cpu_clock = {
4266         .enable         = cpu_clock_perf_event_enable,
4267         .disable        = cpu_clock_perf_event_disable,
4268         .read           = cpu_clock_perf_event_read,
4269 };
4270
4271 /*
4272  * Software event: task time clock
4273  */
4274
4275 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4276 {
4277         u64 prev;
4278         s64 delta;
4279
4280         prev = atomic64_xchg(&event->hw.prev_count, now);
4281         delta = now - prev;
4282         atomic64_add(delta, &event->count);
4283 }
4284
4285 static int task_clock_perf_event_enable(struct perf_event *event)
4286 {
4287         struct hw_perf_event *hwc = &event->hw;
4288         u64 now;
4289
4290         now = event->ctx->time;
4291
4292         atomic64_set(&hwc->prev_count, now);
4293
4294         perf_swevent_start_hrtimer(event);
4295
4296         return 0;
4297 }
4298
4299 static void task_clock_perf_event_disable(struct perf_event *event)
4300 {
4301         perf_swevent_cancel_hrtimer(event);
4302         task_clock_perf_event_update(event, event->ctx->time);
4303
4304 }
4305
4306 static void task_clock_perf_event_read(struct perf_event *event)
4307 {
4308         u64 time;
4309
4310         if (!in_nmi()) {
4311                 update_context_time(event->ctx);
4312                 time = event->ctx->time;
4313         } else {
4314                 u64 now = perf_clock();
4315                 u64 delta = now - event->ctx->timestamp;
4316                 time = event->ctx->time + delta;
4317         }
4318
4319         task_clock_perf_event_update(event, time);
4320 }
4321
4322 static const struct pmu perf_ops_task_clock = {
4323         .enable         = task_clock_perf_event_enable,
4324         .disable        = task_clock_perf_event_disable,
4325         .read           = task_clock_perf_event_read,
4326 };
4327
4328 #ifdef CONFIG_EVENT_TRACING
4329
4330 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4331                           int entry_size)
4332 {
4333         struct perf_raw_record raw = {
4334                 .size = entry_size,
4335                 .data = record,
4336         };
4337
4338         struct perf_sample_data data = {
4339                 .addr = addr,
4340                 .raw = &raw,
4341         };
4342
4343         struct pt_regs *regs = get_irq_regs();
4344
4345         if (!regs)
4346                 regs = task_pt_regs(current);
4347
4348         /* Trace events already protected against recursion */
4349         do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4350                                 &data, regs);
4351 }
4352 EXPORT_SYMBOL_GPL(perf_tp_event);
4353
4354 static int perf_tp_event_match(struct perf_event *event,
4355                                 struct perf_sample_data *data)
4356 {
4357         void *record = data->raw->data;
4358
4359         if (likely(!event->filter) || filter_match_preds(event->filter, record))
4360                 return 1;
4361         return 0;
4362 }
4363
4364 static void tp_perf_event_destroy(struct perf_event *event)
4365 {
4366         ftrace_profile_disable(event->attr.config);
4367 }
4368
4369 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4370 {
4371         /*
4372          * Raw tracepoint data is a severe data leak, only allow root to
4373          * have these.
4374          */
4375         if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4376                         perf_paranoid_tracepoint_raw() &&
4377                         !capable(CAP_SYS_ADMIN))
4378                 return ERR_PTR(-EPERM);
4379
4380         if (ftrace_profile_enable(event->attr.config))
4381                 return NULL;
4382
4383         event->destroy = tp_perf_event_destroy;
4384
4385         return &perf_ops_generic;
4386 }
4387
4388 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4389 {
4390         char *filter_str;
4391         int ret;
4392
4393         if (event->attr.type != PERF_TYPE_TRACEPOINT)
4394                 return -EINVAL;
4395
4396         filter_str = strndup_user(arg, PAGE_SIZE);
4397         if (IS_ERR(filter_str))
4398                 return PTR_ERR(filter_str);
4399
4400         ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4401
4402         kfree(filter_str);
4403         return ret;
4404 }
4405
4406 static void perf_event_free_filter(struct perf_event *event)
4407 {
4408         ftrace_profile_free_filter(event);
4409 }
4410
4411 #else
4412
4413 static int perf_tp_event_match(struct perf_event *event,
4414                                 struct perf_sample_data *data)
4415 {
4416         return 1;
4417 }
4418
4419 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4420 {
4421         return NULL;
4422 }
4423
4424 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4425 {
4426         return -ENOENT;
4427 }
4428
4429 static void perf_event_free_filter(struct perf_event *event)
4430 {
4431 }
4432
4433 #endif /* CONFIG_EVENT_TRACING */
4434
4435 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4436 static void bp_perf_event_destroy(struct perf_event *event)
4437 {
4438         release_bp_slot(event);
4439 }
4440
4441 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4442 {
4443         int err;
4444
4445         err = register_perf_hw_breakpoint(bp);
4446         if (err)
4447                 return ERR_PTR(err);
4448
4449         bp->destroy = bp_perf_event_destroy;
4450
4451         return &perf_ops_bp;
4452 }
4453
4454 void perf_bp_event(struct perf_event *bp, void *data)
4455 {
4456         struct perf_sample_data sample;
4457         struct pt_regs *regs = data;
4458
4459         sample.raw = NULL;
4460         sample.addr = bp->attr.bp_addr;
4461
4462         if (!perf_exclude_event(bp, regs))
4463                 perf_swevent_add(bp, 1, 1, &sample, regs);
4464 }
4465 #else
4466 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4467 {
4468         return NULL;
4469 }
4470
4471 void perf_bp_event(struct perf_event *bp, void *regs)
4472 {
4473 }
4474 #endif
4475
4476 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4477
4478 static void sw_perf_event_destroy(struct perf_event *event)
4479 {
4480         u64 event_id = event->attr.config;
4481
4482         WARN_ON(event->parent);
4483
4484         atomic_dec(&perf_swevent_enabled[event_id]);
4485 }
4486
4487 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4488 {
4489         const struct pmu *pmu = NULL;
4490         u64 event_id = event->attr.config;
4491
4492         /*
4493          * Software events (currently) can't in general distinguish
4494          * between user, kernel and hypervisor events.
4495          * However, context switches and cpu migrations are considered
4496          * to be kernel events, and page faults are never hypervisor
4497          * events.
4498          */
4499         switch (event_id) {
4500         case PERF_COUNT_SW_CPU_CLOCK:
4501                 pmu = &perf_ops_cpu_clock;
4502
4503                 break;
4504         case PERF_COUNT_SW_TASK_CLOCK:
4505                 /*
4506                  * If the user instantiates this as a per-cpu event,
4507                  * use the cpu_clock event instead.
4508                  */
4509                 if (event->ctx->task)
4510                         pmu = &perf_ops_task_clock;
4511                 else
4512                         pmu = &perf_ops_cpu_clock;
4513
4514                 break;
4515         case PERF_COUNT_SW_PAGE_FAULTS:
4516         case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4517         case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4518         case PERF_COUNT_SW_CONTEXT_SWITCHES:
4519         case PERF_COUNT_SW_CPU_MIGRATIONS:
4520         case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4521         case PERF_COUNT_SW_EMULATION_FAULTS:
4522                 if (!event->parent) {
4523                         atomic_inc(&perf_swevent_enabled[event_id]);
4524                         event->destroy = sw_perf_event_destroy;
4525                 }
4526                 pmu = &perf_ops_generic;
4527                 break;
4528         }
4529
4530         return pmu;
4531 }
4532
4533 /*
4534  * Allocate and initialize a event structure
4535  */
4536 static struct perf_event *
4537 perf_event_alloc(struct perf_event_attr *attr,
4538                    int cpu,
4539                    struct perf_event_context *ctx,
4540                    struct perf_event *group_leader,
4541                    struct perf_event *parent_event,
4542                    perf_overflow_handler_t overflow_handler,
4543                    gfp_t gfpflags)
4544 {
4545         const struct pmu *pmu;
4546         struct perf_event *event;
4547         struct hw_perf_event *hwc;
4548         long err;
4549
4550         event = kzalloc(sizeof(*event), gfpflags);
4551         if (!event)
4552                 return ERR_PTR(-ENOMEM);
4553
4554         /*
4555          * Single events are their own group leaders, with an
4556          * empty sibling list:
4557          */
4558         if (!group_leader)
4559                 group_leader = event;
4560
4561         mutex_init(&event->child_mutex);
4562         INIT_LIST_HEAD(&event->child_list);
4563
4564         INIT_LIST_HEAD(&event->group_entry);
4565         INIT_LIST_HEAD(&event->event_entry);
4566         INIT_LIST_HEAD(&event->sibling_list);
4567         init_waitqueue_head(&event->waitq);
4568
4569         mutex_init(&event->mmap_mutex);
4570
4571         event->cpu              = cpu;
4572         event->attr             = *attr;
4573         event->group_leader     = group_leader;
4574         event->pmu              = NULL;
4575         event->ctx              = ctx;
4576         event->oncpu            = -1;
4577
4578         event->parent           = parent_event;
4579
4580         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
4581         event->id               = atomic64_inc_return(&perf_event_id);
4582
4583         event->state            = PERF_EVENT_STATE_INACTIVE;
4584
4585         if (!overflow_handler && parent_event)
4586                 overflow_handler = parent_event->overflow_handler;
4587         
4588         event->overflow_handler = overflow_handler;
4589
4590         if (attr->disabled)
4591                 event->state = PERF_EVENT_STATE_OFF;
4592
4593         pmu = NULL;
4594
4595         hwc = &event->hw;
4596         hwc->sample_period = attr->sample_period;
4597         if (attr->freq && attr->sample_freq)
4598                 hwc->sample_period = 1;
4599         hwc->last_period = hwc->sample_period;
4600
4601         atomic64_set(&hwc->period_left, hwc->sample_period);
4602
4603         /*
4604          * we currently do not support PERF_FORMAT_GROUP on inherited events
4605          */
4606         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4607                 goto done;
4608
4609         switch (attr->type) {
4610         case PERF_TYPE_RAW:
4611         case PERF_TYPE_HARDWARE:
4612         case PERF_TYPE_HW_CACHE:
4613                 pmu = hw_perf_event_init(event);
4614                 break;
4615
4616         case PERF_TYPE_SOFTWARE:
4617                 pmu = sw_perf_event_init(event);
4618                 break;
4619
4620         case PERF_TYPE_TRACEPOINT:
4621                 pmu = tp_perf_event_init(event);
4622                 break;
4623
4624         case PERF_TYPE_BREAKPOINT:
4625                 pmu = bp_perf_event_init(event);
4626                 break;
4627
4628
4629         default:
4630                 break;
4631         }
4632 done:
4633         err = 0;
4634         if (!pmu)
4635                 err = -EINVAL;
4636         else if (IS_ERR(pmu))
4637                 err = PTR_ERR(pmu);
4638
4639         if (err) {
4640                 if (event->ns)
4641                         put_pid_ns(event->ns);
4642                 kfree(event);
4643                 return ERR_PTR(err);
4644         }
4645
4646         event->pmu = pmu;
4647
4648         if (!event->parent) {
4649                 atomic_inc(&nr_events);
4650                 if (event->attr.mmap)
4651                         atomic_inc(&nr_mmap_events);
4652                 if (event->attr.comm)
4653                         atomic_inc(&nr_comm_events);
4654                 if (event->attr.task)
4655                         atomic_inc(&nr_task_events);
4656         }
4657
4658         return event;
4659 }
4660
4661 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4662                           struct perf_event_attr *attr)
4663 {
4664         u32 size;
4665         int ret;
4666
4667         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4668                 return -EFAULT;
4669
4670         /*
4671          * zero the full structure, so that a short copy will be nice.
4672          */
4673         memset(attr, 0, sizeof(*attr));
4674
4675         ret = get_user(size, &uattr->size);
4676         if (ret)
4677                 return ret;
4678
4679         if (size > PAGE_SIZE)   /* silly large */
4680                 goto err_size;
4681
4682         if (!size)              /* abi compat */
4683                 size = PERF_ATTR_SIZE_VER0;
4684
4685         if (size < PERF_ATTR_SIZE_VER0)
4686                 goto err_size;
4687
4688         /*
4689          * If we're handed a bigger struct than we know of,
4690          * ensure all the unknown bits are 0 - i.e. new
4691          * user-space does not rely on any kernel feature
4692          * extensions we dont know about yet.
4693          */
4694         if (size > sizeof(*attr)) {
4695                 unsigned char __user *addr;
4696                 unsigned char __user *end;
4697                 unsigned char val;
4698
4699                 addr = (void __user *)uattr + sizeof(*attr);
4700                 end  = (void __user *)uattr + size;
4701
4702                 for (; addr < end; addr++) {
4703                         ret = get_user(val, addr);
4704                         if (ret)
4705                                 return ret;
4706                         if (val)
4707                                 goto err_size;
4708                 }
4709                 size = sizeof(*attr);
4710         }
4711
4712         ret = copy_from_user(attr, uattr, size);
4713         if (ret)
4714                 return -EFAULT;
4715
4716         /*
4717          * If the type exists, the corresponding creation will verify
4718          * the attr->config.
4719          */
4720         if (attr->type >= PERF_TYPE_MAX)
4721                 return -EINVAL;
4722
4723         if (attr->__reserved_1 || attr->__reserved_2)
4724                 return -EINVAL;
4725
4726         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4727                 return -EINVAL;
4728
4729         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4730                 return -EINVAL;
4731
4732 out:
4733         return ret;
4734
4735 err_size:
4736         put_user(sizeof(*attr), &uattr->size);
4737         ret = -E2BIG;
4738         goto out;
4739 }
4740
4741 static int perf_event_set_output(struct perf_event *event, int output_fd)
4742 {
4743         struct perf_event *output_event = NULL;
4744         struct file *output_file = NULL;
4745         struct perf_event *old_output;
4746         int fput_needed = 0;
4747         int ret = -EINVAL;
4748
4749         if (!output_fd)
4750                 goto set;
4751
4752         output_file = fget_light(output_fd, &fput_needed);
4753         if (!output_file)
4754                 return -EBADF;
4755
4756         if (output_file->f_op != &perf_fops)
4757                 goto out;
4758
4759         output_event = output_file->private_data;
4760
4761         /* Don't chain output fds */
4762         if (output_event->output)
4763                 goto out;
4764
4765         /* Don't set an output fd when we already have an output channel */
4766         if (event->data)
4767                 goto out;
4768
4769         atomic_long_inc(&output_file->f_count);
4770
4771 set:
4772         mutex_lock(&event->mmap_mutex);
4773         old_output = event->output;
4774         rcu_assign_pointer(event->output, output_event);
4775         mutex_unlock(&event->mmap_mutex);
4776
4777         if (old_output) {
4778                 /*
4779                  * we need to make sure no existing perf_output_*()
4780                  * is still referencing this event.
4781                  */
4782                 synchronize_rcu();
4783                 fput(old_output->filp);
4784         }
4785
4786         ret = 0;
4787 out:
4788         fput_light(output_file, fput_needed);
4789         return ret;
4790 }
4791
4792 /**
4793  * sys_perf_event_open - open a performance event, associate it to a task/cpu
4794  *
4795  * @attr_uptr:  event_id type attributes for monitoring/sampling
4796  * @pid:                target pid
4797  * @cpu:                target cpu
4798  * @group_fd:           group leader event fd
4799  */
4800 SYSCALL_DEFINE5(perf_event_open,
4801                 struct perf_event_attr __user *, attr_uptr,
4802                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4803 {
4804         struct perf_event *event, *group_leader;
4805         struct perf_event_attr attr;
4806         struct perf_event_context *ctx;
4807         struct file *event_file = NULL;
4808         struct file *group_file = NULL;
4809         int fput_needed = 0;
4810         int fput_needed2 = 0;
4811         int err;
4812
4813         /* for future expandability... */
4814         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4815                 return -EINVAL;
4816
4817         err = perf_copy_attr(attr_uptr, &attr);
4818         if (err)
4819                 return err;
4820
4821         if (!attr.exclude_kernel) {
4822                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4823                         return -EACCES;
4824         }
4825
4826         if (attr.freq) {
4827                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4828                         return -EINVAL;
4829         }
4830
4831         /*
4832          * Get the target context (task or percpu):
4833          */
4834         ctx = find_get_context(pid, cpu);
4835         if (IS_ERR(ctx))
4836                 return PTR_ERR(ctx);
4837
4838         /*
4839          * Look up the group leader (we will attach this event to it):
4840          */
4841         group_leader = NULL;
4842         if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4843                 err = -EINVAL;
4844                 group_file = fget_light(group_fd, &fput_needed);
4845                 if (!group_file)
4846                         goto err_put_context;
4847                 if (group_file->f_op != &perf_fops)
4848                         goto err_put_context;
4849
4850                 group_leader = group_file->private_data;
4851                 /*
4852                  * Do not allow a recursive hierarchy (this new sibling
4853                  * becoming part of another group-sibling):
4854                  */
4855                 if (group_leader->group_leader != group_leader)
4856                         goto err_put_context;
4857                 /*
4858                  * Do not allow to attach to a group in a different
4859                  * task or CPU context:
4860                  */
4861                 if (group_leader->ctx != ctx)
4862                         goto err_put_context;
4863                 /*
4864                  * Only a group leader can be exclusive or pinned
4865                  */
4866                 if (attr.exclusive || attr.pinned)
4867                         goto err_put_context;
4868         }
4869
4870         event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4871                                      NULL, NULL, GFP_KERNEL);
4872         err = PTR_ERR(event);
4873         if (IS_ERR(event))
4874                 goto err_put_context;
4875
4876         err = anon_inode_getfd("[perf_event]", &perf_fops, event, O_RDWR);
4877         if (err < 0)
4878                 goto err_free_put_context;
4879
4880         event_file = fget_light(err, &fput_needed2);
4881         if (!event_file)
4882                 goto err_free_put_context;
4883
4884         if (flags & PERF_FLAG_FD_OUTPUT) {
4885                 err = perf_event_set_output(event, group_fd);
4886                 if (err)
4887                         goto err_fput_free_put_context;
4888         }
4889
4890         event->filp = event_file;
4891         WARN_ON_ONCE(ctx->parent_ctx);
4892         mutex_lock(&ctx->mutex);
4893         perf_install_in_context(ctx, event, cpu);
4894         ++ctx->generation;
4895         mutex_unlock(&ctx->mutex);
4896
4897         event->owner = current;
4898         get_task_struct(current);
4899         mutex_lock(&current->perf_event_mutex);
4900         list_add_tail(&event->owner_entry, &current->perf_event_list);
4901         mutex_unlock(&current->perf_event_mutex);
4902
4903 err_fput_free_put_context:
4904         fput_light(event_file, fput_needed2);
4905
4906 err_free_put_context:
4907         if (err < 0)
4908                 kfree(event);
4909
4910 err_put_context:
4911         if (err < 0)
4912                 put_ctx(ctx);
4913
4914         fput_light(group_file, fput_needed);
4915
4916         return err;
4917 }
4918
4919 /**
4920  * perf_event_create_kernel_counter
4921  *
4922  * @attr: attributes of the counter to create
4923  * @cpu: cpu in which the counter is bound
4924  * @pid: task to profile
4925  */
4926 struct perf_event *
4927 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4928                                  pid_t pid,
4929                                  perf_overflow_handler_t overflow_handler)
4930 {
4931         struct perf_event *event;
4932         struct perf_event_context *ctx;
4933         int err;
4934
4935         /*
4936          * Get the target context (task or percpu):
4937          */
4938
4939         ctx = find_get_context(pid, cpu);
4940         if (IS_ERR(ctx)) {
4941                 err = PTR_ERR(ctx);
4942                 goto err_exit;
4943         }
4944
4945         event = perf_event_alloc(attr, cpu, ctx, NULL,
4946                                  NULL, overflow_handler, GFP_KERNEL);
4947         if (IS_ERR(event)) {
4948                 err = PTR_ERR(event);
4949                 goto err_put_context;
4950         }
4951
4952         event->filp = NULL;
4953         WARN_ON_ONCE(ctx->parent_ctx);
4954         mutex_lock(&ctx->mutex);
4955         perf_install_in_context(ctx, event, cpu);
4956         ++ctx->generation;
4957         mutex_unlock(&ctx->mutex);
4958
4959         event->owner = current;
4960         get_task_struct(current);
4961         mutex_lock(&current->perf_event_mutex);
4962         list_add_tail(&event->owner_entry, &current->perf_event_list);
4963         mutex_unlock(&current->perf_event_mutex);
4964
4965         return event;
4966
4967  err_put_context:
4968         put_ctx(ctx);
4969  err_exit:
4970         return ERR_PTR(err);
4971 }
4972 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4973
4974 /*
4975  * inherit a event from parent task to child task:
4976  */
4977 static struct perf_event *
4978 inherit_event(struct perf_event *parent_event,
4979               struct task_struct *parent,
4980               struct perf_event_context *parent_ctx,
4981               struct task_struct *child,
4982               struct perf_event *group_leader,
4983               struct perf_event_context *child_ctx)
4984 {
4985         struct perf_event *child_event;
4986
4987         /*
4988          * Instead of creating recursive hierarchies of events,
4989          * we link inherited events back to the original parent,
4990          * which has a filp for sure, which we use as the reference
4991          * count:
4992          */
4993         if (parent_event->parent)
4994                 parent_event = parent_event->parent;
4995
4996         child_event = perf_event_alloc(&parent_event->attr,
4997                                            parent_event->cpu, child_ctx,
4998                                            group_leader, parent_event,
4999                                            NULL, GFP_KERNEL);
5000         if (IS_ERR(child_event))
5001                 return child_event;
5002         get_ctx(child_ctx);
5003
5004         /*
5005          * Make the child state follow the state of the parent event,
5006          * not its attr.disabled bit.  We hold the parent's mutex,
5007          * so we won't race with perf_event_{en, dis}able_family.
5008          */
5009         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
5010                 child_event->state = PERF_EVENT_STATE_INACTIVE;
5011         else
5012                 child_event->state = PERF_EVENT_STATE_OFF;
5013
5014         if (parent_event->attr.freq) {
5015                 u64 sample_period = parent_event->hw.sample_period;
5016                 struct hw_perf_event *hwc = &child_event->hw;
5017
5018                 hwc->sample_period = sample_period;
5019                 hwc->last_period   = sample_period;
5020
5021                 atomic64_set(&hwc->period_left, sample_period);
5022         }
5023
5024         child_event->overflow_handler = parent_event->overflow_handler;
5025
5026         /*
5027          * Link it up in the child's context:
5028          */
5029         add_event_to_ctx(child_event, child_ctx);
5030
5031         /*
5032          * Get a reference to the parent filp - we will fput it
5033          * when the child event exits. This is safe to do because
5034          * we are in the parent and we know that the filp still
5035          * exists and has a nonzero count:
5036          */
5037         atomic_long_inc(&parent_event->filp->f_count);
5038
5039         /*
5040          * Link this into the parent event's child list
5041          */
5042         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5043         mutex_lock(&parent_event->child_mutex);
5044         list_add_tail(&child_event->child_list, &parent_event->child_list);
5045         mutex_unlock(&parent_event->child_mutex);
5046
5047         return child_event;
5048 }
5049
5050 static int inherit_group(struct perf_event *parent_event,
5051               struct task_struct *parent,
5052               struct perf_event_context *parent_ctx,
5053               struct task_struct *child,
5054               struct perf_event_context *child_ctx)
5055 {
5056         struct perf_event *leader;
5057         struct perf_event *sub;
5058         struct perf_event *child_ctr;
5059
5060         leader = inherit_event(parent_event, parent, parent_ctx,
5061                                  child, NULL, child_ctx);
5062         if (IS_ERR(leader))
5063                 return PTR_ERR(leader);
5064         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5065                 child_ctr = inherit_event(sub, parent, parent_ctx,
5066                                             child, leader, child_ctx);
5067                 if (IS_ERR(child_ctr))
5068                         return PTR_ERR(child_ctr);
5069         }
5070         return 0;
5071 }
5072
5073 static void sync_child_event(struct perf_event *child_event,
5074                                struct task_struct *child)
5075 {
5076         struct perf_event *parent_event = child_event->parent;
5077         u64 child_val;
5078
5079         if (child_event->attr.inherit_stat)
5080                 perf_event_read_event(child_event, child);
5081
5082         child_val = atomic64_read(&child_event->count);
5083
5084         /*
5085          * Add back the child's count to the parent's count:
5086          */
5087         atomic64_add(child_val, &parent_event->count);
5088         atomic64_add(child_event->total_time_enabled,
5089                      &parent_event->child_total_time_enabled);
5090         atomic64_add(child_event->total_time_running,
5091                      &parent_event->child_total_time_running);
5092
5093         /*
5094          * Remove this event from the parent's list
5095          */
5096         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5097         mutex_lock(&parent_event->child_mutex);
5098         list_del_init(&child_event->child_list);
5099         mutex_unlock(&parent_event->child_mutex);
5100
5101         /*
5102          * Release the parent event, if this was the last
5103          * reference to it.
5104          */
5105         fput(parent_event->filp);
5106 }
5107
5108 static void
5109 __perf_event_exit_task(struct perf_event *child_event,
5110                          struct perf_event_context *child_ctx,
5111                          struct task_struct *child)
5112 {
5113         struct perf_event *parent_event;
5114
5115         perf_event_remove_from_context(child_event);
5116
5117         parent_event = child_event->parent;
5118         /*
5119          * It can happen that parent exits first, and has events
5120          * that are still around due to the child reference. These
5121          * events need to be zapped - but otherwise linger.
5122          */
5123         if (parent_event) {
5124                 sync_child_event(child_event, child);
5125                 free_event(child_event);
5126         }
5127 }
5128
5129 /*
5130  * When a child task exits, feed back event values to parent events.
5131  */
5132 void perf_event_exit_task(struct task_struct *child)
5133 {
5134         struct perf_event *child_event, *tmp;
5135         struct perf_event_context *child_ctx;
5136         unsigned long flags;
5137
5138         if (likely(!child->perf_event_ctxp)) {
5139                 perf_event_task(child, NULL, 0);
5140                 return;
5141         }
5142
5143         local_irq_save(flags);
5144         /*
5145          * We can't reschedule here because interrupts are disabled,
5146          * and either child is current or it is a task that can't be
5147          * scheduled, so we are now safe from rescheduling changing
5148          * our context.
5149          */
5150         child_ctx = child->perf_event_ctxp;
5151         __perf_event_task_sched_out(child_ctx);
5152
5153         /*
5154          * Take the context lock here so that if find_get_context is
5155          * reading child->perf_event_ctxp, we wait until it has
5156          * incremented the context's refcount before we do put_ctx below.
5157          */
5158         raw_spin_lock(&child_ctx->lock);
5159         child->perf_event_ctxp = NULL;
5160         /*
5161          * If this context is a clone; unclone it so it can't get
5162          * swapped to another process while we're removing all
5163          * the events from it.
5164          */
5165         unclone_ctx(child_ctx);
5166         update_context_time(child_ctx);
5167         raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5168
5169         /*
5170          * Report the task dead after unscheduling the events so that we
5171          * won't get any samples after PERF_RECORD_EXIT. We can however still
5172          * get a few PERF_RECORD_READ events.
5173          */
5174         perf_event_task(child, child_ctx, 0);
5175
5176         /*
5177          * We can recurse on the same lock type through:
5178          *
5179          *   __perf_event_exit_task()
5180          *     sync_child_event()
5181          *       fput(parent_event->filp)
5182          *         perf_release()
5183          *           mutex_lock(&ctx->mutex)
5184          *
5185          * But since its the parent context it won't be the same instance.
5186          */
5187         mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
5188
5189 again:
5190         list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5191                                  group_entry)
5192                 __perf_event_exit_task(child_event, child_ctx, child);
5193
5194         list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5195                                  group_entry)
5196                 __perf_event_exit_task(child_event, child_ctx, child);
5197
5198         /*
5199          * If the last event was a group event, it will have appended all
5200          * its siblings to the list, but we obtained 'tmp' before that which
5201          * will still point to the list head terminating the iteration.
5202          */
5203         if (!list_empty(&child_ctx->pinned_groups) ||
5204             !list_empty(&child_ctx->flexible_groups))
5205                 goto again;
5206
5207         mutex_unlock(&child_ctx->mutex);
5208
5209         put_ctx(child_ctx);
5210 }
5211
5212 static void perf_free_event(struct perf_event *event,
5213                             struct perf_event_context *ctx)
5214 {
5215         struct perf_event *parent = event->parent;
5216
5217         if (WARN_ON_ONCE(!parent))
5218                 return;
5219
5220         mutex_lock(&parent->child_mutex);
5221         list_del_init(&event->child_list);
5222         mutex_unlock(&parent->child_mutex);
5223
5224         fput(parent->filp);
5225
5226         list_del_event(event, ctx);
5227         free_event(event);
5228 }
5229
5230 /*
5231  * free an unexposed, unused context as created by inheritance by
5232  * init_task below, used by fork() in case of fail.
5233  */
5234 void perf_event_free_task(struct task_struct *task)
5235 {
5236         struct perf_event_context *ctx = task->perf_event_ctxp;
5237         struct perf_event *event, *tmp;
5238
5239         if (!ctx)
5240                 return;
5241
5242         mutex_lock(&ctx->mutex);
5243 again:
5244         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5245                 perf_free_event(event, ctx);
5246
5247         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5248                                  group_entry)
5249                 perf_free_event(event, ctx);
5250
5251         if (!list_empty(&ctx->pinned_groups) ||
5252             !list_empty(&ctx->flexible_groups))
5253                 goto again;
5254
5255         mutex_unlock(&ctx->mutex);
5256
5257         put_ctx(ctx);
5258 }
5259
5260 static int
5261 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5262                    struct perf_event_context *parent_ctx,
5263                    struct task_struct *child,
5264                    int *inherited_all)
5265 {
5266         int ret;
5267         struct perf_event_context *child_ctx = child->perf_event_ctxp;
5268
5269         if (!event->attr.inherit) {
5270                 *inherited_all = 0;
5271                 return 0;
5272         }
5273
5274         if (!child_ctx) {
5275                 /*
5276                  * This is executed from the parent task context, so
5277                  * inherit events that have been marked for cloning.
5278                  * First allocate and initialize a context for the
5279                  * child.
5280                  */
5281
5282                 child_ctx = kzalloc(sizeof(struct perf_event_context),
5283                                     GFP_KERNEL);
5284                 if (!child_ctx)
5285                         return -ENOMEM;
5286
5287                 __perf_event_init_context(child_ctx, child);
5288                 child->perf_event_ctxp = child_ctx;
5289                 get_task_struct(child);
5290         }
5291
5292         ret = inherit_group(event, parent, parent_ctx,
5293                             child, child_ctx);
5294
5295         if (ret)
5296                 *inherited_all = 0;
5297
5298         return ret;
5299 }
5300
5301
5302 /*
5303  * Initialize the perf_event context in task_struct
5304  */
5305 int perf_event_init_task(struct task_struct *child)
5306 {
5307         struct perf_event_context *child_ctx, *parent_ctx;
5308         struct perf_event_context *cloned_ctx;
5309         struct perf_event *event;
5310         struct task_struct *parent = current;
5311         int inherited_all = 1;
5312         int ret = 0;
5313
5314         child->perf_event_ctxp = NULL;
5315
5316         mutex_init(&child->perf_event_mutex);
5317         INIT_LIST_HEAD(&child->perf_event_list);
5318
5319         if (likely(!parent->perf_event_ctxp))
5320                 return 0;
5321
5322         /*
5323          * If the parent's context is a clone, pin it so it won't get
5324          * swapped under us.
5325          */
5326         parent_ctx = perf_pin_task_context(parent);
5327
5328         /*
5329          * No need to check if parent_ctx != NULL here; since we saw
5330          * it non-NULL earlier, the only reason for it to become NULL
5331          * is if we exit, and since we're currently in the middle of
5332          * a fork we can't be exiting at the same time.
5333          */
5334
5335         /*
5336          * Lock the parent list. No need to lock the child - not PID
5337          * hashed yet and not running, so nobody can access it.
5338          */
5339         mutex_lock(&parent_ctx->mutex);
5340
5341         /*
5342          * We dont have to disable NMIs - we are only looking at
5343          * the list, not manipulating it:
5344          */
5345         list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5346                 ret = inherit_task_group(event, parent, parent_ctx, child,
5347                                          &inherited_all);
5348                 if (ret)
5349                         break;
5350         }
5351
5352         list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5353                 ret = inherit_task_group(event, parent, parent_ctx, child,
5354                                          &inherited_all);
5355                 if (ret)
5356                         break;
5357         }
5358
5359         child_ctx = child->perf_event_ctxp;
5360
5361         if (child_ctx && inherited_all) {
5362                 /*
5363                  * Mark the child context as a clone of the parent
5364                  * context, or of whatever the parent is a clone of.
5365                  * Note that if the parent is a clone, it could get
5366                  * uncloned at any point, but that doesn't matter
5367                  * because the list of events and the generation
5368                  * count can't have changed since we took the mutex.
5369                  */
5370                 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5371                 if (cloned_ctx) {
5372                         child_ctx->parent_ctx = cloned_ctx;
5373                         child_ctx->parent_gen = parent_ctx->parent_gen;
5374                 } else {
5375                         child_ctx->parent_ctx = parent_ctx;
5376                         child_ctx->parent_gen = parent_ctx->generation;
5377                 }
5378                 get_ctx(child_ctx->parent_ctx);
5379         }
5380
5381         mutex_unlock(&parent_ctx->mutex);
5382
5383         perf_unpin_context(parent_ctx);
5384
5385         return ret;
5386 }
5387
5388 static void __cpuinit perf_event_init_cpu(int cpu)
5389 {
5390         struct perf_cpu_context *cpuctx;
5391
5392         cpuctx = &per_cpu(perf_cpu_context, cpu);
5393         __perf_event_init_context(&cpuctx->ctx, NULL);
5394
5395         spin_lock(&perf_resource_lock);
5396         cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5397         spin_unlock(&perf_resource_lock);
5398
5399         hw_perf_event_setup(cpu);
5400 }
5401
5402 #ifdef CONFIG_HOTPLUG_CPU
5403 static void __perf_event_exit_cpu(void *info)
5404 {
5405         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5406         struct perf_event_context *ctx = &cpuctx->ctx;
5407         struct perf_event *event, *tmp;
5408
5409         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5410                 __perf_event_remove_from_context(event);
5411         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5412                 __perf_event_remove_from_context(event);
5413 }
5414 static void perf_event_exit_cpu(int cpu)
5415 {
5416         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5417         struct perf_event_context *ctx = &cpuctx->ctx;
5418
5419         mutex_lock(&ctx->mutex);
5420         smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5421         mutex_unlock(&ctx->mutex);
5422 }
5423 #else
5424 static inline void perf_event_exit_cpu(int cpu) { }
5425 #endif
5426
5427 static int __cpuinit
5428 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5429 {
5430         unsigned int cpu = (long)hcpu;
5431
5432         switch (action) {
5433
5434         case CPU_UP_PREPARE:
5435         case CPU_UP_PREPARE_FROZEN:
5436                 perf_event_init_cpu(cpu);
5437                 break;
5438
5439         case CPU_ONLINE:
5440         case CPU_ONLINE_FROZEN:
5441                 hw_perf_event_setup_online(cpu);
5442                 break;
5443
5444         case CPU_DOWN_PREPARE:
5445         case CPU_DOWN_PREPARE_FROZEN:
5446                 perf_event_exit_cpu(cpu);
5447                 break;
5448
5449         default:
5450                 break;
5451         }
5452
5453         return NOTIFY_OK;
5454 }
5455
5456 /*
5457  * This has to have a higher priority than migration_notifier in sched.c.
5458  */
5459 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5460         .notifier_call          = perf_cpu_notify,
5461         .priority               = 20,
5462 };
5463
5464 void __init perf_event_init(void)
5465 {
5466         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5467                         (void *)(long)smp_processor_id());
5468         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5469                         (void *)(long)smp_processor_id());
5470         register_cpu_notifier(&perf_cpu_nb);
5471 }
5472
5473 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
5474 {
5475         return sprintf(buf, "%d\n", perf_reserved_percpu);
5476 }
5477
5478 static ssize_t
5479 perf_set_reserve_percpu(struct sysdev_class *class,
5480                         const char *buf,
5481                         size_t count)
5482 {
5483         struct perf_cpu_context *cpuctx;
5484         unsigned long val;
5485         int err, cpu, mpt;
5486
5487         err = strict_strtoul(buf, 10, &val);
5488         if (err)
5489                 return err;
5490         if (val > perf_max_events)
5491                 return -EINVAL;
5492
5493         spin_lock(&perf_resource_lock);
5494         perf_reserved_percpu = val;
5495         for_each_online_cpu(cpu) {
5496                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5497                 raw_spin_lock_irq(&cpuctx->ctx.lock);
5498                 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5499                           perf_max_events - perf_reserved_percpu);
5500                 cpuctx->max_pertask = mpt;
5501                 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5502         }
5503         spin_unlock(&perf_resource_lock);
5504
5505         return count;
5506 }
5507
5508 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
5509 {
5510         return sprintf(buf, "%d\n", perf_overcommit);
5511 }
5512
5513 static ssize_t
5514 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
5515 {
5516         unsigned long val;
5517         int err;
5518
5519         err = strict_strtoul(buf, 10, &val);
5520         if (err)
5521                 return err;
5522         if (val > 1)
5523                 return -EINVAL;
5524
5525         spin_lock(&perf_resource_lock);
5526         perf_overcommit = val;
5527         spin_unlock(&perf_resource_lock);
5528
5529         return count;
5530 }
5531
5532 static SYSDEV_CLASS_ATTR(
5533                                 reserve_percpu,
5534                                 0644,
5535                                 perf_show_reserve_percpu,
5536                                 perf_set_reserve_percpu
5537                         );
5538
5539 static SYSDEV_CLASS_ATTR(
5540                                 overcommit,
5541                                 0644,
5542                                 perf_show_overcommit,
5543                                 perf_set_overcommit
5544                         );
5545
5546 static struct attribute *perfclass_attrs[] = {
5547         &attr_reserve_percpu.attr,
5548         &attr_overcommit.attr,
5549         NULL
5550 };
5551
5552 static struct attribute_group perfclass_attr_group = {
5553         .attrs                  = perfclass_attrs,
5554         .name                   = "perf_events",
5555 };
5556
5557 static int __init perf_event_sysfs_init(void)
5558 {
5559         return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5560                                   &perfclass_attr_group);
5561 }
5562 device_initcall(perf_event_sysfs_init);