perf: Don't schedule out/in pinned events on task tick
[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 task_ctx_sched_in(struct task_struct *task,
1366                               enum event_type_t event_type)
1367 {
1368         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1369         struct perf_event_context *ctx = task->perf_event_ctxp;
1370
1371         if (likely(!ctx))
1372                 return;
1373         if (cpuctx->task_ctx == ctx)
1374                 return;
1375         ctx_sched_in(ctx, cpuctx, event_type);
1376         cpuctx->task_ctx = ctx;
1377 }
1378 /*
1379  * Called from scheduler to add the events of the current task
1380  * with interrupts disabled.
1381  *
1382  * We restore the event value and then enable it.
1383  *
1384  * This does not protect us against NMI, but enable()
1385  * sets the enabled bit in the control field of event _before_
1386  * accessing the event control register. If a NMI hits, then it will
1387  * keep the event running.
1388  */
1389 void perf_event_task_sched_in(struct task_struct *task)
1390 {
1391         task_ctx_sched_in(task, EVENT_ALL);
1392 }
1393
1394 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1395                              enum event_type_t event_type)
1396 {
1397         struct perf_event_context *ctx = &cpuctx->ctx;
1398
1399         ctx_sched_in(ctx, cpuctx, event_type);
1400 }
1401
1402 #define MAX_INTERRUPTS (~0ULL)
1403
1404 static void perf_log_throttle(struct perf_event *event, int enable);
1405
1406 static void perf_adjust_period(struct perf_event *event, u64 events)
1407 {
1408         struct hw_perf_event *hwc = &event->hw;
1409         u64 period, sample_period;
1410         s64 delta;
1411
1412         events *= hwc->sample_period;
1413         period = div64_u64(events, event->attr.sample_freq);
1414
1415         delta = (s64)(period - hwc->sample_period);
1416         delta = (delta + 7) / 8; /* low pass filter */
1417
1418         sample_period = hwc->sample_period + delta;
1419
1420         if (!sample_period)
1421                 sample_period = 1;
1422
1423         hwc->sample_period = sample_period;
1424 }
1425
1426 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1427 {
1428         struct perf_event *event;
1429         struct hw_perf_event *hwc;
1430         u64 interrupts, freq;
1431
1432         raw_spin_lock(&ctx->lock);
1433         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1434                 if (event->state != PERF_EVENT_STATE_ACTIVE)
1435                         continue;
1436
1437                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1438                         continue;
1439
1440                 hwc = &event->hw;
1441
1442                 interrupts = hwc->interrupts;
1443                 hwc->interrupts = 0;
1444
1445                 /*
1446                  * unthrottle events on the tick
1447                  */
1448                 if (interrupts == MAX_INTERRUPTS) {
1449                         perf_log_throttle(event, 1);
1450                         event->pmu->unthrottle(event);
1451                         interrupts = 2*sysctl_perf_event_sample_rate/HZ;
1452                 }
1453
1454                 if (!event->attr.freq || !event->attr.sample_freq)
1455                         continue;
1456
1457                 /*
1458                  * if the specified freq < HZ then we need to skip ticks
1459                  */
1460                 if (event->attr.sample_freq < HZ) {
1461                         freq = event->attr.sample_freq;
1462
1463                         hwc->freq_count += freq;
1464                         hwc->freq_interrupts += interrupts;
1465
1466                         if (hwc->freq_count < HZ)
1467                                 continue;
1468
1469                         interrupts = hwc->freq_interrupts;
1470                         hwc->freq_interrupts = 0;
1471                         hwc->freq_count -= HZ;
1472                 } else
1473                         freq = HZ;
1474
1475                 perf_adjust_period(event, freq * interrupts);
1476
1477                 /*
1478                  * In order to avoid being stalled by an (accidental) huge
1479                  * sample period, force reset the sample period if we didn't
1480                  * get any events in this freq period.
1481                  */
1482                 if (!interrupts) {
1483                         perf_disable();
1484                         event->pmu->disable(event);
1485                         atomic64_set(&hwc->period_left, 0);
1486                         event->pmu->enable(event);
1487                         perf_enable();
1488                 }
1489         }
1490         raw_spin_unlock(&ctx->lock);
1491 }
1492
1493 /*
1494  * Round-robin a context's events:
1495  */
1496 static void rotate_ctx(struct perf_event_context *ctx)
1497 {
1498         if (!ctx->nr_events)
1499                 return;
1500
1501         raw_spin_lock(&ctx->lock);
1502
1503         /* Rotate the first entry last of non-pinned groups */
1504         perf_disable();
1505
1506         list_rotate_left(&ctx->flexible_groups);
1507
1508         perf_enable();
1509
1510         raw_spin_unlock(&ctx->lock);
1511 }
1512
1513 void perf_event_task_tick(struct task_struct *curr)
1514 {
1515         struct perf_cpu_context *cpuctx;
1516         struct perf_event_context *ctx;
1517
1518         if (!atomic_read(&nr_events))
1519                 return;
1520
1521         cpuctx = &__get_cpu_var(perf_cpu_context);
1522         ctx = curr->perf_event_ctxp;
1523
1524         perf_ctx_adjust_freq(&cpuctx->ctx);
1525         if (ctx)
1526                 perf_ctx_adjust_freq(ctx);
1527
1528         cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1529         if (ctx)
1530                 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1531
1532         rotate_ctx(&cpuctx->ctx);
1533         if (ctx)
1534                 rotate_ctx(ctx);
1535
1536         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1537         if (ctx)
1538                 task_ctx_sched_in(curr, EVENT_FLEXIBLE);
1539 }
1540
1541 static int event_enable_on_exec(struct perf_event *event,
1542                                 struct perf_event_context *ctx)
1543 {
1544         if (!event->attr.enable_on_exec)
1545                 return 0;
1546
1547         event->attr.enable_on_exec = 0;
1548         if (event->state >= PERF_EVENT_STATE_INACTIVE)
1549                 return 0;
1550
1551         __perf_event_mark_enabled(event, ctx);
1552
1553         return 1;
1554 }
1555
1556 /*
1557  * Enable all of a task's events that have been marked enable-on-exec.
1558  * This expects task == current.
1559  */
1560 static void perf_event_enable_on_exec(struct task_struct *task)
1561 {
1562         struct perf_event_context *ctx;
1563         struct perf_event *event;
1564         unsigned long flags;
1565         int enabled = 0;
1566         int ret;
1567
1568         local_irq_save(flags);
1569         ctx = task->perf_event_ctxp;
1570         if (!ctx || !ctx->nr_events)
1571                 goto out;
1572
1573         __perf_event_task_sched_out(ctx);
1574
1575         raw_spin_lock(&ctx->lock);
1576
1577         list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1578                 ret = event_enable_on_exec(event, ctx);
1579                 if (ret)
1580                         enabled = 1;
1581         }
1582
1583         list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1584                 ret = event_enable_on_exec(event, ctx);
1585                 if (ret)
1586                         enabled = 1;
1587         }
1588
1589         /*
1590          * Unclone this context if we enabled any event.
1591          */
1592         if (enabled)
1593                 unclone_ctx(ctx);
1594
1595         raw_spin_unlock(&ctx->lock);
1596
1597         perf_event_task_sched_in(task);
1598  out:
1599         local_irq_restore(flags);
1600 }
1601
1602 /*
1603  * Cross CPU call to read the hardware event
1604  */
1605 static void __perf_event_read(void *info)
1606 {
1607         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1608         struct perf_event *event = info;
1609         struct perf_event_context *ctx = event->ctx;
1610
1611         /*
1612          * If this is a task context, we need to check whether it is
1613          * the current task context of this cpu.  If not it has been
1614          * scheduled out before the smp call arrived.  In that case
1615          * event->count would have been updated to a recent sample
1616          * when the event was scheduled out.
1617          */
1618         if (ctx->task && cpuctx->task_ctx != ctx)
1619                 return;
1620
1621         raw_spin_lock(&ctx->lock);
1622         update_context_time(ctx);
1623         update_event_times(event);
1624         raw_spin_unlock(&ctx->lock);
1625
1626         event->pmu->read(event);
1627 }
1628
1629 static u64 perf_event_read(struct perf_event *event)
1630 {
1631         /*
1632          * If event is enabled and currently active on a CPU, update the
1633          * value in the event structure:
1634          */
1635         if (event->state == PERF_EVENT_STATE_ACTIVE) {
1636                 smp_call_function_single(event->oncpu,
1637                                          __perf_event_read, event, 1);
1638         } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1639                 struct perf_event_context *ctx = event->ctx;
1640                 unsigned long flags;
1641
1642                 raw_spin_lock_irqsave(&ctx->lock, flags);
1643                 update_context_time(ctx);
1644                 update_event_times(event);
1645                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1646         }
1647
1648         return atomic64_read(&event->count);
1649 }
1650
1651 /*
1652  * Initialize the perf_event context in a task_struct:
1653  */
1654 static void
1655 __perf_event_init_context(struct perf_event_context *ctx,
1656                             struct task_struct *task)
1657 {
1658         raw_spin_lock_init(&ctx->lock);
1659         mutex_init(&ctx->mutex);
1660         INIT_LIST_HEAD(&ctx->pinned_groups);
1661         INIT_LIST_HEAD(&ctx->flexible_groups);
1662         INIT_LIST_HEAD(&ctx->event_list);
1663         atomic_set(&ctx->refcount, 1);
1664         ctx->task = task;
1665 }
1666
1667 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1668 {
1669         struct perf_event_context *ctx;
1670         struct perf_cpu_context *cpuctx;
1671         struct task_struct *task;
1672         unsigned long flags;
1673         int err;
1674
1675         if (pid == -1 && cpu != -1) {
1676                 /* Must be root to operate on a CPU event: */
1677                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1678                         return ERR_PTR(-EACCES);
1679
1680                 if (cpu < 0 || cpu >= nr_cpumask_bits)
1681                         return ERR_PTR(-EINVAL);
1682
1683                 /*
1684                  * We could be clever and allow to attach a event to an
1685                  * offline CPU and activate it when the CPU comes up, but
1686                  * that's for later.
1687                  */
1688                 if (!cpu_online(cpu))
1689                         return ERR_PTR(-ENODEV);
1690
1691                 cpuctx = &per_cpu(perf_cpu_context, cpu);
1692                 ctx = &cpuctx->ctx;
1693                 get_ctx(ctx);
1694
1695                 return ctx;
1696         }
1697
1698         rcu_read_lock();
1699         if (!pid)
1700                 task = current;
1701         else
1702                 task = find_task_by_vpid(pid);
1703         if (task)
1704                 get_task_struct(task);
1705         rcu_read_unlock();
1706
1707         if (!task)
1708                 return ERR_PTR(-ESRCH);
1709
1710         /*
1711          * Can't attach events to a dying task.
1712          */
1713         err = -ESRCH;
1714         if (task->flags & PF_EXITING)
1715                 goto errout;
1716
1717         /* Reuse ptrace permission checks for now. */
1718         err = -EACCES;
1719         if (!ptrace_may_access(task, PTRACE_MODE_READ))
1720                 goto errout;
1721
1722  retry:
1723         ctx = perf_lock_task_context(task, &flags);
1724         if (ctx) {
1725                 unclone_ctx(ctx);
1726                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1727         }
1728
1729         if (!ctx) {
1730                 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1731                 err = -ENOMEM;
1732                 if (!ctx)
1733                         goto errout;
1734                 __perf_event_init_context(ctx, task);
1735                 get_ctx(ctx);
1736                 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1737                         /*
1738                          * We raced with some other task; use
1739                          * the context they set.
1740                          */
1741                         kfree(ctx);
1742                         goto retry;
1743                 }
1744                 get_task_struct(task);
1745         }
1746
1747         put_task_struct(task);
1748         return ctx;
1749
1750  errout:
1751         put_task_struct(task);
1752         return ERR_PTR(err);
1753 }
1754
1755 static void perf_event_free_filter(struct perf_event *event);
1756
1757 static void free_event_rcu(struct rcu_head *head)
1758 {
1759         struct perf_event *event;
1760
1761         event = container_of(head, struct perf_event, rcu_head);
1762         if (event->ns)
1763                 put_pid_ns(event->ns);
1764         perf_event_free_filter(event);
1765         kfree(event);
1766 }
1767
1768 static void perf_pending_sync(struct perf_event *event);
1769
1770 static void free_event(struct perf_event *event)
1771 {
1772         perf_pending_sync(event);
1773
1774         if (!event->parent) {
1775                 atomic_dec(&nr_events);
1776                 if (event->attr.mmap)
1777                         atomic_dec(&nr_mmap_events);
1778                 if (event->attr.comm)
1779                         atomic_dec(&nr_comm_events);
1780                 if (event->attr.task)
1781                         atomic_dec(&nr_task_events);
1782         }
1783
1784         if (event->output) {
1785                 fput(event->output->filp);
1786                 event->output = NULL;
1787         }
1788
1789         if (event->destroy)
1790                 event->destroy(event);
1791
1792         put_ctx(event->ctx);
1793         call_rcu(&event->rcu_head, free_event_rcu);
1794 }
1795
1796 int perf_event_release_kernel(struct perf_event *event)
1797 {
1798         struct perf_event_context *ctx = event->ctx;
1799
1800         WARN_ON_ONCE(ctx->parent_ctx);
1801         mutex_lock(&ctx->mutex);
1802         perf_event_remove_from_context(event);
1803         mutex_unlock(&ctx->mutex);
1804
1805         mutex_lock(&event->owner->perf_event_mutex);
1806         list_del_init(&event->owner_entry);
1807         mutex_unlock(&event->owner->perf_event_mutex);
1808         put_task_struct(event->owner);
1809
1810         free_event(event);
1811
1812         return 0;
1813 }
1814 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1815
1816 /*
1817  * Called when the last reference to the file is gone.
1818  */
1819 static int perf_release(struct inode *inode, struct file *file)
1820 {
1821         struct perf_event *event = file->private_data;
1822
1823         file->private_data = NULL;
1824
1825         return perf_event_release_kernel(event);
1826 }
1827
1828 static int perf_event_read_size(struct perf_event *event)
1829 {
1830         int entry = sizeof(u64); /* value */
1831         int size = 0;
1832         int nr = 1;
1833
1834         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1835                 size += sizeof(u64);
1836
1837         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1838                 size += sizeof(u64);
1839
1840         if (event->attr.read_format & PERF_FORMAT_ID)
1841                 entry += sizeof(u64);
1842
1843         if (event->attr.read_format & PERF_FORMAT_GROUP) {
1844                 nr += event->group_leader->nr_siblings;
1845                 size += sizeof(u64);
1846         }
1847
1848         size += entry * nr;
1849
1850         return size;
1851 }
1852
1853 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1854 {
1855         struct perf_event *child;
1856         u64 total = 0;
1857
1858         *enabled = 0;
1859         *running = 0;
1860
1861         mutex_lock(&event->child_mutex);
1862         total += perf_event_read(event);
1863         *enabled += event->total_time_enabled +
1864                         atomic64_read(&event->child_total_time_enabled);
1865         *running += event->total_time_running +
1866                         atomic64_read(&event->child_total_time_running);
1867
1868         list_for_each_entry(child, &event->child_list, child_list) {
1869                 total += perf_event_read(child);
1870                 *enabled += child->total_time_enabled;
1871                 *running += child->total_time_running;
1872         }
1873         mutex_unlock(&event->child_mutex);
1874
1875         return total;
1876 }
1877 EXPORT_SYMBOL_GPL(perf_event_read_value);
1878
1879 static int perf_event_read_group(struct perf_event *event,
1880                                    u64 read_format, char __user *buf)
1881 {
1882         struct perf_event *leader = event->group_leader, *sub;
1883         int n = 0, size = 0, ret = -EFAULT;
1884         struct perf_event_context *ctx = leader->ctx;
1885         u64 values[5];
1886         u64 count, enabled, running;
1887
1888         mutex_lock(&ctx->mutex);
1889         count = perf_event_read_value(leader, &enabled, &running);
1890
1891         values[n++] = 1 + leader->nr_siblings;
1892         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1893                 values[n++] = enabled;
1894         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1895                 values[n++] = running;
1896         values[n++] = count;
1897         if (read_format & PERF_FORMAT_ID)
1898                 values[n++] = primary_event_id(leader);
1899
1900         size = n * sizeof(u64);
1901
1902         if (copy_to_user(buf, values, size))
1903                 goto unlock;
1904
1905         ret = size;
1906
1907         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1908                 n = 0;
1909
1910                 values[n++] = perf_event_read_value(sub, &enabled, &running);
1911                 if (read_format & PERF_FORMAT_ID)
1912                         values[n++] = primary_event_id(sub);
1913
1914                 size = n * sizeof(u64);
1915
1916                 if (copy_to_user(buf + ret, values, size)) {
1917                         ret = -EFAULT;
1918                         goto unlock;
1919                 }
1920
1921                 ret += size;
1922         }
1923 unlock:
1924         mutex_unlock(&ctx->mutex);
1925
1926         return ret;
1927 }
1928
1929 static int perf_event_read_one(struct perf_event *event,
1930                                  u64 read_format, char __user *buf)
1931 {
1932         u64 enabled, running;
1933         u64 values[4];
1934         int n = 0;
1935
1936         values[n++] = perf_event_read_value(event, &enabled, &running);
1937         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1938                 values[n++] = enabled;
1939         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1940                 values[n++] = running;
1941         if (read_format & PERF_FORMAT_ID)
1942                 values[n++] = primary_event_id(event);
1943
1944         if (copy_to_user(buf, values, n * sizeof(u64)))
1945                 return -EFAULT;
1946
1947         return n * sizeof(u64);
1948 }
1949
1950 /*
1951  * Read the performance event - simple non blocking version for now
1952  */
1953 static ssize_t
1954 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
1955 {
1956         u64 read_format = event->attr.read_format;
1957         int ret;
1958
1959         /*
1960          * Return end-of-file for a read on a event that is in
1961          * error state (i.e. because it was pinned but it couldn't be
1962          * scheduled on to the CPU at some point).
1963          */
1964         if (event->state == PERF_EVENT_STATE_ERROR)
1965                 return 0;
1966
1967         if (count < perf_event_read_size(event))
1968                 return -ENOSPC;
1969
1970         WARN_ON_ONCE(event->ctx->parent_ctx);
1971         if (read_format & PERF_FORMAT_GROUP)
1972                 ret = perf_event_read_group(event, read_format, buf);
1973         else
1974                 ret = perf_event_read_one(event, read_format, buf);
1975
1976         return ret;
1977 }
1978
1979 static ssize_t
1980 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1981 {
1982         struct perf_event *event = file->private_data;
1983
1984         return perf_read_hw(event, buf, count);
1985 }
1986
1987 static unsigned int perf_poll(struct file *file, poll_table *wait)
1988 {
1989         struct perf_event *event = file->private_data;
1990         struct perf_mmap_data *data;
1991         unsigned int events = POLL_HUP;
1992
1993         rcu_read_lock();
1994         data = rcu_dereference(event->data);
1995         if (data)
1996                 events = atomic_xchg(&data->poll, 0);
1997         rcu_read_unlock();
1998
1999         poll_wait(file, &event->waitq, wait);
2000
2001         return events;
2002 }
2003
2004 static void perf_event_reset(struct perf_event *event)
2005 {
2006         (void)perf_event_read(event);
2007         atomic64_set(&event->count, 0);
2008         perf_event_update_userpage(event);
2009 }
2010
2011 /*
2012  * Holding the top-level event's child_mutex means that any
2013  * descendant process that has inherited this event will block
2014  * in sync_child_event if it goes to exit, thus satisfying the
2015  * task existence requirements of perf_event_enable/disable.
2016  */
2017 static void perf_event_for_each_child(struct perf_event *event,
2018                                         void (*func)(struct perf_event *))
2019 {
2020         struct perf_event *child;
2021
2022         WARN_ON_ONCE(event->ctx->parent_ctx);
2023         mutex_lock(&event->child_mutex);
2024         func(event);
2025         list_for_each_entry(child, &event->child_list, child_list)
2026                 func(child);
2027         mutex_unlock(&event->child_mutex);
2028 }
2029
2030 static void perf_event_for_each(struct perf_event *event,
2031                                   void (*func)(struct perf_event *))
2032 {
2033         struct perf_event_context *ctx = event->ctx;
2034         struct perf_event *sibling;
2035
2036         WARN_ON_ONCE(ctx->parent_ctx);
2037         mutex_lock(&ctx->mutex);
2038         event = event->group_leader;
2039
2040         perf_event_for_each_child(event, func);
2041         func(event);
2042         list_for_each_entry(sibling, &event->sibling_list, group_entry)
2043                 perf_event_for_each_child(event, func);
2044         mutex_unlock(&ctx->mutex);
2045 }
2046
2047 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2048 {
2049         struct perf_event_context *ctx = event->ctx;
2050         unsigned long size;
2051         int ret = 0;
2052         u64 value;
2053
2054         if (!event->attr.sample_period)
2055                 return -EINVAL;
2056
2057         size = copy_from_user(&value, arg, sizeof(value));
2058         if (size != sizeof(value))
2059                 return -EFAULT;
2060
2061         if (!value)
2062                 return -EINVAL;
2063
2064         raw_spin_lock_irq(&ctx->lock);
2065         if (event->attr.freq) {
2066                 if (value > sysctl_perf_event_sample_rate) {
2067                         ret = -EINVAL;
2068                         goto unlock;
2069                 }
2070
2071                 event->attr.sample_freq = value;
2072         } else {
2073                 event->attr.sample_period = value;
2074                 event->hw.sample_period = value;
2075         }
2076 unlock:
2077         raw_spin_unlock_irq(&ctx->lock);
2078
2079         return ret;
2080 }
2081
2082 static int perf_event_set_output(struct perf_event *event, int output_fd);
2083 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2084
2085 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2086 {
2087         struct perf_event *event = file->private_data;
2088         void (*func)(struct perf_event *);
2089         u32 flags = arg;
2090
2091         switch (cmd) {
2092         case PERF_EVENT_IOC_ENABLE:
2093                 func = perf_event_enable;
2094                 break;
2095         case PERF_EVENT_IOC_DISABLE:
2096                 func = perf_event_disable;
2097                 break;
2098         case PERF_EVENT_IOC_RESET:
2099                 func = perf_event_reset;
2100                 break;
2101
2102         case PERF_EVENT_IOC_REFRESH:
2103                 return perf_event_refresh(event, arg);
2104
2105         case PERF_EVENT_IOC_PERIOD:
2106                 return perf_event_period(event, (u64 __user *)arg);
2107
2108         case PERF_EVENT_IOC_SET_OUTPUT:
2109                 return perf_event_set_output(event, arg);
2110
2111         case PERF_EVENT_IOC_SET_FILTER:
2112                 return perf_event_set_filter(event, (void __user *)arg);
2113
2114         default:
2115                 return -ENOTTY;
2116         }
2117
2118         if (flags & PERF_IOC_FLAG_GROUP)
2119                 perf_event_for_each(event, func);
2120         else
2121                 perf_event_for_each_child(event, func);
2122
2123         return 0;
2124 }
2125
2126 int perf_event_task_enable(void)
2127 {
2128         struct perf_event *event;
2129
2130         mutex_lock(&current->perf_event_mutex);
2131         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2132                 perf_event_for_each_child(event, perf_event_enable);
2133         mutex_unlock(&current->perf_event_mutex);
2134
2135         return 0;
2136 }
2137
2138 int perf_event_task_disable(void)
2139 {
2140         struct perf_event *event;
2141
2142         mutex_lock(&current->perf_event_mutex);
2143         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2144                 perf_event_for_each_child(event, perf_event_disable);
2145         mutex_unlock(&current->perf_event_mutex);
2146
2147         return 0;
2148 }
2149
2150 #ifndef PERF_EVENT_INDEX_OFFSET
2151 # define PERF_EVENT_INDEX_OFFSET 0
2152 #endif
2153
2154 static int perf_event_index(struct perf_event *event)
2155 {
2156         if (event->state != PERF_EVENT_STATE_ACTIVE)
2157                 return 0;
2158
2159         return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2160 }
2161
2162 /*
2163  * Callers need to ensure there can be no nesting of this function, otherwise
2164  * the seqlock logic goes bad. We can not serialize this because the arch
2165  * code calls this from NMI context.
2166  */
2167 void perf_event_update_userpage(struct perf_event *event)
2168 {
2169         struct perf_event_mmap_page *userpg;
2170         struct perf_mmap_data *data;
2171
2172         rcu_read_lock();
2173         data = rcu_dereference(event->data);
2174         if (!data)
2175                 goto unlock;
2176
2177         userpg = data->user_page;
2178
2179         /*
2180          * Disable preemption so as to not let the corresponding user-space
2181          * spin too long if we get preempted.
2182          */
2183         preempt_disable();
2184         ++userpg->lock;
2185         barrier();
2186         userpg->index = perf_event_index(event);
2187         userpg->offset = atomic64_read(&event->count);
2188         if (event->state == PERF_EVENT_STATE_ACTIVE)
2189                 userpg->offset -= atomic64_read(&event->hw.prev_count);
2190
2191         userpg->time_enabled = event->total_time_enabled +
2192                         atomic64_read(&event->child_total_time_enabled);
2193
2194         userpg->time_running = event->total_time_running +
2195                         atomic64_read(&event->child_total_time_running);
2196
2197         barrier();
2198         ++userpg->lock;
2199         preempt_enable();
2200 unlock:
2201         rcu_read_unlock();
2202 }
2203
2204 static unsigned long perf_data_size(struct perf_mmap_data *data)
2205 {
2206         return data->nr_pages << (PAGE_SHIFT + data->data_order);
2207 }
2208
2209 #ifndef CONFIG_PERF_USE_VMALLOC
2210
2211 /*
2212  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2213  */
2214
2215 static struct page *
2216 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2217 {
2218         if (pgoff > data->nr_pages)
2219                 return NULL;
2220
2221         if (pgoff == 0)
2222                 return virt_to_page(data->user_page);
2223
2224         return virt_to_page(data->data_pages[pgoff - 1]);
2225 }
2226
2227 static struct perf_mmap_data *
2228 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2229 {
2230         struct perf_mmap_data *data;
2231         unsigned long size;
2232         int i;
2233
2234         WARN_ON(atomic_read(&event->mmap_count));
2235
2236         size = sizeof(struct perf_mmap_data);
2237         size += nr_pages * sizeof(void *);
2238
2239         data = kzalloc(size, GFP_KERNEL);
2240         if (!data)
2241                 goto fail;
2242
2243         data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2244         if (!data->user_page)
2245                 goto fail_user_page;
2246
2247         for (i = 0; i < nr_pages; i++) {
2248                 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2249                 if (!data->data_pages[i])
2250                         goto fail_data_pages;
2251         }
2252
2253         data->data_order = 0;
2254         data->nr_pages = nr_pages;
2255
2256         return data;
2257
2258 fail_data_pages:
2259         for (i--; i >= 0; i--)
2260                 free_page((unsigned long)data->data_pages[i]);
2261
2262         free_page((unsigned long)data->user_page);
2263
2264 fail_user_page:
2265         kfree(data);
2266
2267 fail:
2268         return NULL;
2269 }
2270
2271 static void perf_mmap_free_page(unsigned long addr)
2272 {
2273         struct page *page = virt_to_page((void *)addr);
2274
2275         page->mapping = NULL;
2276         __free_page(page);
2277 }
2278
2279 static void perf_mmap_data_free(struct perf_mmap_data *data)
2280 {
2281         int i;
2282
2283         perf_mmap_free_page((unsigned long)data->user_page);
2284         for (i = 0; i < data->nr_pages; i++)
2285                 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2286         kfree(data);
2287 }
2288
2289 #else
2290
2291 /*
2292  * Back perf_mmap() with vmalloc memory.
2293  *
2294  * Required for architectures that have d-cache aliasing issues.
2295  */
2296
2297 static struct page *
2298 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2299 {
2300         if (pgoff > (1UL << data->data_order))
2301                 return NULL;
2302
2303         return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2304 }
2305
2306 static void perf_mmap_unmark_page(void *addr)
2307 {
2308         struct page *page = vmalloc_to_page(addr);
2309
2310         page->mapping = NULL;
2311 }
2312
2313 static void perf_mmap_data_free_work(struct work_struct *work)
2314 {
2315         struct perf_mmap_data *data;
2316         void *base;
2317         int i, nr;
2318
2319         data = container_of(work, struct perf_mmap_data, work);
2320         nr = 1 << data->data_order;
2321
2322         base = data->user_page;
2323         for (i = 0; i < nr + 1; i++)
2324                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2325
2326         vfree(base);
2327         kfree(data);
2328 }
2329
2330 static void perf_mmap_data_free(struct perf_mmap_data *data)
2331 {
2332         schedule_work(&data->work);
2333 }
2334
2335 static struct perf_mmap_data *
2336 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2337 {
2338         struct perf_mmap_data *data;
2339         unsigned long size;
2340         void *all_buf;
2341
2342         WARN_ON(atomic_read(&event->mmap_count));
2343
2344         size = sizeof(struct perf_mmap_data);
2345         size += sizeof(void *);
2346
2347         data = kzalloc(size, GFP_KERNEL);
2348         if (!data)
2349                 goto fail;
2350
2351         INIT_WORK(&data->work, perf_mmap_data_free_work);
2352
2353         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2354         if (!all_buf)
2355                 goto fail_all_buf;
2356
2357         data->user_page = all_buf;
2358         data->data_pages[0] = all_buf + PAGE_SIZE;
2359         data->data_order = ilog2(nr_pages);
2360         data->nr_pages = 1;
2361
2362         return data;
2363
2364 fail_all_buf:
2365         kfree(data);
2366
2367 fail:
2368         return NULL;
2369 }
2370
2371 #endif
2372
2373 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2374 {
2375         struct perf_event *event = vma->vm_file->private_data;
2376         struct perf_mmap_data *data;
2377         int ret = VM_FAULT_SIGBUS;
2378
2379         if (vmf->flags & FAULT_FLAG_MKWRITE) {
2380                 if (vmf->pgoff == 0)
2381                         ret = 0;
2382                 return ret;
2383         }
2384
2385         rcu_read_lock();
2386         data = rcu_dereference(event->data);
2387         if (!data)
2388                 goto unlock;
2389
2390         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2391                 goto unlock;
2392
2393         vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2394         if (!vmf->page)
2395                 goto unlock;
2396
2397         get_page(vmf->page);
2398         vmf->page->mapping = vma->vm_file->f_mapping;
2399         vmf->page->index   = vmf->pgoff;
2400
2401         ret = 0;
2402 unlock:
2403         rcu_read_unlock();
2404
2405         return ret;
2406 }
2407
2408 static void
2409 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2410 {
2411         long max_size = perf_data_size(data);
2412
2413         atomic_set(&data->lock, -1);
2414
2415         if (event->attr.watermark) {
2416                 data->watermark = min_t(long, max_size,
2417                                         event->attr.wakeup_watermark);
2418         }
2419
2420         if (!data->watermark)
2421                 data->watermark = max_size / 2;
2422
2423
2424         rcu_assign_pointer(event->data, data);
2425 }
2426
2427 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2428 {
2429         struct perf_mmap_data *data;
2430
2431         data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2432         perf_mmap_data_free(data);
2433 }
2434
2435 static void perf_mmap_data_release(struct perf_event *event)
2436 {
2437         struct perf_mmap_data *data = event->data;
2438
2439         WARN_ON(atomic_read(&event->mmap_count));
2440
2441         rcu_assign_pointer(event->data, NULL);
2442         call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2443 }
2444
2445 static void perf_mmap_open(struct vm_area_struct *vma)
2446 {
2447         struct perf_event *event = vma->vm_file->private_data;
2448
2449         atomic_inc(&event->mmap_count);
2450 }
2451
2452 static void perf_mmap_close(struct vm_area_struct *vma)
2453 {
2454         struct perf_event *event = vma->vm_file->private_data;
2455
2456         WARN_ON_ONCE(event->ctx->parent_ctx);
2457         if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2458                 unsigned long size = perf_data_size(event->data);
2459                 struct user_struct *user = current_user();
2460
2461                 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2462                 vma->vm_mm->locked_vm -= event->data->nr_locked;
2463                 perf_mmap_data_release(event);
2464                 mutex_unlock(&event->mmap_mutex);
2465         }
2466 }
2467
2468 static const struct vm_operations_struct perf_mmap_vmops = {
2469         .open           = perf_mmap_open,
2470         .close          = perf_mmap_close,
2471         .fault          = perf_mmap_fault,
2472         .page_mkwrite   = perf_mmap_fault,
2473 };
2474
2475 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2476 {
2477         struct perf_event *event = file->private_data;
2478         unsigned long user_locked, user_lock_limit;
2479         struct user_struct *user = current_user();
2480         unsigned long locked, lock_limit;
2481         struct perf_mmap_data *data;
2482         unsigned long vma_size;
2483         unsigned long nr_pages;
2484         long user_extra, extra;
2485         int ret = 0;
2486
2487         if (!(vma->vm_flags & VM_SHARED))
2488                 return -EINVAL;
2489
2490         vma_size = vma->vm_end - vma->vm_start;
2491         nr_pages = (vma_size / PAGE_SIZE) - 1;
2492
2493         /*
2494          * If we have data pages ensure they're a power-of-two number, so we
2495          * can do bitmasks instead of modulo.
2496          */
2497         if (nr_pages != 0 && !is_power_of_2(nr_pages))
2498                 return -EINVAL;
2499
2500         if (vma_size != PAGE_SIZE * (1 + nr_pages))
2501                 return -EINVAL;
2502
2503         if (vma->vm_pgoff != 0)
2504                 return -EINVAL;
2505
2506         WARN_ON_ONCE(event->ctx->parent_ctx);
2507         mutex_lock(&event->mmap_mutex);
2508         if (event->output) {
2509                 ret = -EINVAL;
2510                 goto unlock;
2511         }
2512
2513         if (atomic_inc_not_zero(&event->mmap_count)) {
2514                 if (nr_pages != event->data->nr_pages)
2515                         ret = -EINVAL;
2516                 goto unlock;
2517         }
2518
2519         user_extra = nr_pages + 1;
2520         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2521
2522         /*
2523          * Increase the limit linearly with more CPUs:
2524          */
2525         user_lock_limit *= num_online_cpus();
2526
2527         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2528
2529         extra = 0;
2530         if (user_locked > user_lock_limit)
2531                 extra = user_locked - user_lock_limit;
2532
2533         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2534         lock_limit >>= PAGE_SHIFT;
2535         locked = vma->vm_mm->locked_vm + extra;
2536
2537         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2538                 !capable(CAP_IPC_LOCK)) {
2539                 ret = -EPERM;
2540                 goto unlock;
2541         }
2542
2543         WARN_ON(event->data);
2544
2545         data = perf_mmap_data_alloc(event, nr_pages);
2546         ret = -ENOMEM;
2547         if (!data)
2548                 goto unlock;
2549
2550         ret = 0;
2551         perf_mmap_data_init(event, data);
2552
2553         atomic_set(&event->mmap_count, 1);
2554         atomic_long_add(user_extra, &user->locked_vm);
2555         vma->vm_mm->locked_vm += extra;
2556         event->data->nr_locked = extra;
2557         if (vma->vm_flags & VM_WRITE)
2558                 event->data->writable = 1;
2559
2560 unlock:
2561         mutex_unlock(&event->mmap_mutex);
2562
2563         vma->vm_flags |= VM_RESERVED;
2564         vma->vm_ops = &perf_mmap_vmops;
2565
2566         return ret;
2567 }
2568
2569 static int perf_fasync(int fd, struct file *filp, int on)
2570 {
2571         struct inode *inode = filp->f_path.dentry->d_inode;
2572         struct perf_event *event = filp->private_data;
2573         int retval;
2574
2575         mutex_lock(&inode->i_mutex);
2576         retval = fasync_helper(fd, filp, on, &event->fasync);
2577         mutex_unlock(&inode->i_mutex);
2578
2579         if (retval < 0)
2580                 return retval;
2581
2582         return 0;
2583 }
2584
2585 static const struct file_operations perf_fops = {
2586         .release                = perf_release,
2587         .read                   = perf_read,
2588         .poll                   = perf_poll,
2589         .unlocked_ioctl         = perf_ioctl,
2590         .compat_ioctl           = perf_ioctl,
2591         .mmap                   = perf_mmap,
2592         .fasync                 = perf_fasync,
2593 };
2594
2595 /*
2596  * Perf event wakeup
2597  *
2598  * If there's data, ensure we set the poll() state and publish everything
2599  * to user-space before waking everybody up.
2600  */
2601
2602 void perf_event_wakeup(struct perf_event *event)
2603 {
2604         wake_up_all(&event->waitq);
2605
2606         if (event->pending_kill) {
2607                 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2608                 event->pending_kill = 0;
2609         }
2610 }
2611
2612 /*
2613  * Pending wakeups
2614  *
2615  * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2616  *
2617  * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2618  * single linked list and use cmpxchg() to add entries lockless.
2619  */
2620
2621 static void perf_pending_event(struct perf_pending_entry *entry)
2622 {
2623         struct perf_event *event = container_of(entry,
2624                         struct perf_event, pending);
2625
2626         if (event->pending_disable) {
2627                 event->pending_disable = 0;
2628                 __perf_event_disable(event);
2629         }
2630
2631         if (event->pending_wakeup) {
2632                 event->pending_wakeup = 0;
2633                 perf_event_wakeup(event);
2634         }
2635 }
2636
2637 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2638
2639 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2640         PENDING_TAIL,
2641 };
2642
2643 static void perf_pending_queue(struct perf_pending_entry *entry,
2644                                void (*func)(struct perf_pending_entry *))
2645 {
2646         struct perf_pending_entry **head;
2647
2648         if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2649                 return;
2650
2651         entry->func = func;
2652
2653         head = &get_cpu_var(perf_pending_head);
2654
2655         do {
2656                 entry->next = *head;
2657         } while (cmpxchg(head, entry->next, entry) != entry->next);
2658
2659         set_perf_event_pending();
2660
2661         put_cpu_var(perf_pending_head);
2662 }
2663
2664 static int __perf_pending_run(void)
2665 {
2666         struct perf_pending_entry *list;
2667         int nr = 0;
2668
2669         list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2670         while (list != PENDING_TAIL) {
2671                 void (*func)(struct perf_pending_entry *);
2672                 struct perf_pending_entry *entry = list;
2673
2674                 list = list->next;
2675
2676                 func = entry->func;
2677                 entry->next = NULL;
2678                 /*
2679                  * Ensure we observe the unqueue before we issue the wakeup,
2680                  * so that we won't be waiting forever.
2681                  * -- see perf_not_pending().
2682                  */
2683                 smp_wmb();
2684
2685                 func(entry);
2686                 nr++;
2687         }
2688
2689         return nr;
2690 }
2691
2692 static inline int perf_not_pending(struct perf_event *event)
2693 {
2694         /*
2695          * If we flush on whatever cpu we run, there is a chance we don't
2696          * need to wait.
2697          */
2698         get_cpu();
2699         __perf_pending_run();
2700         put_cpu();
2701
2702         /*
2703          * Ensure we see the proper queue state before going to sleep
2704          * so that we do not miss the wakeup. -- see perf_pending_handle()
2705          */
2706         smp_rmb();
2707         return event->pending.next == NULL;
2708 }
2709
2710 static void perf_pending_sync(struct perf_event *event)
2711 {
2712         wait_event(event->waitq, perf_not_pending(event));
2713 }
2714
2715 void perf_event_do_pending(void)
2716 {
2717         __perf_pending_run();
2718 }
2719
2720 /*
2721  * Callchain support -- arch specific
2722  */
2723
2724 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2725 {
2726         return NULL;
2727 }
2728
2729 /*
2730  * Output
2731  */
2732 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2733                               unsigned long offset, unsigned long head)
2734 {
2735         unsigned long mask;
2736
2737         if (!data->writable)
2738                 return true;
2739
2740         mask = perf_data_size(data) - 1;
2741
2742         offset = (offset - tail) & mask;
2743         head   = (head   - tail) & mask;
2744
2745         if ((int)(head - offset) < 0)
2746                 return false;
2747
2748         return true;
2749 }
2750
2751 static void perf_output_wakeup(struct perf_output_handle *handle)
2752 {
2753         atomic_set(&handle->data->poll, POLL_IN);
2754
2755         if (handle->nmi) {
2756                 handle->event->pending_wakeup = 1;
2757                 perf_pending_queue(&handle->event->pending,
2758                                    perf_pending_event);
2759         } else
2760                 perf_event_wakeup(handle->event);
2761 }
2762
2763 /*
2764  * Curious locking construct.
2765  *
2766  * We need to ensure a later event_id doesn't publish a head when a former
2767  * event_id isn't done writing. However since we need to deal with NMIs we
2768  * cannot fully serialize things.
2769  *
2770  * What we do is serialize between CPUs so we only have to deal with NMI
2771  * nesting on a single CPU.
2772  *
2773  * We only publish the head (and generate a wakeup) when the outer-most
2774  * event_id completes.
2775  */
2776 static void perf_output_lock(struct perf_output_handle *handle)
2777 {
2778         struct perf_mmap_data *data = handle->data;
2779         int cur, cpu = get_cpu();
2780
2781         handle->locked = 0;
2782
2783         for (;;) {
2784                 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2785                 if (cur == -1) {
2786                         handle->locked = 1;
2787                         break;
2788                 }
2789                 if (cur == cpu)
2790                         break;
2791
2792                 cpu_relax();
2793         }
2794 }
2795
2796 static void perf_output_unlock(struct perf_output_handle *handle)
2797 {
2798         struct perf_mmap_data *data = handle->data;
2799         unsigned long head;
2800         int cpu;
2801
2802         data->done_head = data->head;
2803
2804         if (!handle->locked)
2805                 goto out;
2806
2807 again:
2808         /*
2809          * The xchg implies a full barrier that ensures all writes are done
2810          * before we publish the new head, matched by a rmb() in userspace when
2811          * reading this position.
2812          */
2813         while ((head = atomic_long_xchg(&data->done_head, 0)))
2814                 data->user_page->data_head = head;
2815
2816         /*
2817          * NMI can happen here, which means we can miss a done_head update.
2818          */
2819
2820         cpu = atomic_xchg(&data->lock, -1);
2821         WARN_ON_ONCE(cpu != smp_processor_id());
2822
2823         /*
2824          * Therefore we have to validate we did not indeed do so.
2825          */
2826         if (unlikely(atomic_long_read(&data->done_head))) {
2827                 /*
2828                  * Since we had it locked, we can lock it again.
2829                  */
2830                 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2831                         cpu_relax();
2832
2833                 goto again;
2834         }
2835
2836         if (atomic_xchg(&data->wakeup, 0))
2837                 perf_output_wakeup(handle);
2838 out:
2839         put_cpu();
2840 }
2841
2842 void perf_output_copy(struct perf_output_handle *handle,
2843                       const void *buf, unsigned int len)
2844 {
2845         unsigned int pages_mask;
2846         unsigned long offset;
2847         unsigned int size;
2848         void **pages;
2849
2850         offset          = handle->offset;
2851         pages_mask      = handle->data->nr_pages - 1;
2852         pages           = handle->data->data_pages;
2853
2854         do {
2855                 unsigned long page_offset;
2856                 unsigned long page_size;
2857                 int nr;
2858
2859                 nr          = (offset >> PAGE_SHIFT) & pages_mask;
2860                 page_size   = 1UL << (handle->data->data_order + PAGE_SHIFT);
2861                 page_offset = offset & (page_size - 1);
2862                 size        = min_t(unsigned int, page_size - page_offset, len);
2863
2864                 memcpy(pages[nr] + page_offset, buf, size);
2865
2866                 len         -= size;
2867                 buf         += size;
2868                 offset      += size;
2869         } while (len);
2870
2871         handle->offset = offset;
2872
2873         /*
2874          * Check we didn't copy past our reservation window, taking the
2875          * possible unsigned int wrap into account.
2876          */
2877         WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2878 }
2879
2880 int perf_output_begin(struct perf_output_handle *handle,
2881                       struct perf_event *event, unsigned int size,
2882                       int nmi, int sample)
2883 {
2884         struct perf_event *output_event;
2885         struct perf_mmap_data *data;
2886         unsigned long tail, offset, head;
2887         int have_lost;
2888         struct {
2889                 struct perf_event_header header;
2890                 u64                      id;
2891                 u64                      lost;
2892         } lost_event;
2893
2894         rcu_read_lock();
2895         /*
2896          * For inherited events we send all the output towards the parent.
2897          */
2898         if (event->parent)
2899                 event = event->parent;
2900
2901         output_event = rcu_dereference(event->output);
2902         if (output_event)
2903                 event = output_event;
2904
2905         data = rcu_dereference(event->data);
2906         if (!data)
2907                 goto out;
2908
2909         handle->data    = data;
2910         handle->event   = event;
2911         handle->nmi     = nmi;
2912         handle->sample  = sample;
2913
2914         if (!data->nr_pages)
2915                 goto fail;
2916
2917         have_lost = atomic_read(&data->lost);
2918         if (have_lost)
2919                 size += sizeof(lost_event);
2920
2921         perf_output_lock(handle);
2922
2923         do {
2924                 /*
2925                  * Userspace could choose to issue a mb() before updating the
2926                  * tail pointer. So that all reads will be completed before the
2927                  * write is issued.
2928                  */
2929                 tail = ACCESS_ONCE(data->user_page->data_tail);
2930                 smp_rmb();
2931                 offset = head = atomic_long_read(&data->head);
2932                 head += size;
2933                 if (unlikely(!perf_output_space(data, tail, offset, head)))
2934                         goto fail;
2935         } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2936
2937         handle->offset  = offset;
2938         handle->head    = head;
2939
2940         if (head - tail > data->watermark)
2941                 atomic_set(&data->wakeup, 1);
2942
2943         if (have_lost) {
2944                 lost_event.header.type = PERF_RECORD_LOST;
2945                 lost_event.header.misc = 0;
2946                 lost_event.header.size = sizeof(lost_event);
2947                 lost_event.id          = event->id;
2948                 lost_event.lost        = atomic_xchg(&data->lost, 0);
2949
2950                 perf_output_put(handle, lost_event);
2951         }
2952
2953         return 0;
2954
2955 fail:
2956         atomic_inc(&data->lost);
2957         perf_output_unlock(handle);
2958 out:
2959         rcu_read_unlock();
2960
2961         return -ENOSPC;
2962 }
2963
2964 void perf_output_end(struct perf_output_handle *handle)
2965 {
2966         struct perf_event *event = handle->event;
2967         struct perf_mmap_data *data = handle->data;
2968
2969         int wakeup_events = event->attr.wakeup_events;
2970
2971         if (handle->sample && wakeup_events) {
2972                 int events = atomic_inc_return(&data->events);
2973                 if (events >= wakeup_events) {
2974                         atomic_sub(wakeup_events, &data->events);
2975                         atomic_set(&data->wakeup, 1);
2976                 }
2977         }
2978
2979         perf_output_unlock(handle);
2980         rcu_read_unlock();
2981 }
2982
2983 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
2984 {
2985         /*
2986          * only top level events have the pid namespace they were created in
2987          */
2988         if (event->parent)
2989                 event = event->parent;
2990
2991         return task_tgid_nr_ns(p, event->ns);
2992 }
2993
2994 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
2995 {
2996         /*
2997          * only top level events have the pid namespace they were created in
2998          */
2999         if (event->parent)
3000                 event = event->parent;
3001
3002         return task_pid_nr_ns(p, event->ns);
3003 }
3004
3005 static void perf_output_read_one(struct perf_output_handle *handle,
3006                                  struct perf_event *event)
3007 {
3008         u64 read_format = event->attr.read_format;
3009         u64 values[4];
3010         int n = 0;
3011
3012         values[n++] = atomic64_read(&event->count);
3013         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3014                 values[n++] = event->total_time_enabled +
3015                         atomic64_read(&event->child_total_time_enabled);
3016         }
3017         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3018                 values[n++] = event->total_time_running +
3019                         atomic64_read(&event->child_total_time_running);
3020         }
3021         if (read_format & PERF_FORMAT_ID)
3022                 values[n++] = primary_event_id(event);
3023
3024         perf_output_copy(handle, values, n * sizeof(u64));
3025 }
3026
3027 /*
3028  * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3029  */
3030 static void perf_output_read_group(struct perf_output_handle *handle,
3031                             struct perf_event *event)
3032 {
3033         struct perf_event *leader = event->group_leader, *sub;
3034         u64 read_format = event->attr.read_format;
3035         u64 values[5];
3036         int n = 0;
3037
3038         values[n++] = 1 + leader->nr_siblings;
3039
3040         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3041                 values[n++] = leader->total_time_enabled;
3042
3043         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3044                 values[n++] = leader->total_time_running;
3045
3046         if (leader != event)
3047                 leader->pmu->read(leader);
3048
3049         values[n++] = atomic64_read(&leader->count);
3050         if (read_format & PERF_FORMAT_ID)
3051                 values[n++] = primary_event_id(leader);
3052
3053         perf_output_copy(handle, values, n * sizeof(u64));
3054
3055         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3056                 n = 0;
3057
3058                 if (sub != event)
3059                         sub->pmu->read(sub);
3060
3061                 values[n++] = atomic64_read(&sub->count);
3062                 if (read_format & PERF_FORMAT_ID)
3063                         values[n++] = primary_event_id(sub);
3064
3065                 perf_output_copy(handle, values, n * sizeof(u64));
3066         }
3067 }
3068
3069 static void perf_output_read(struct perf_output_handle *handle,
3070                              struct perf_event *event)
3071 {
3072         if (event->attr.read_format & PERF_FORMAT_GROUP)
3073                 perf_output_read_group(handle, event);
3074         else
3075                 perf_output_read_one(handle, event);
3076 }
3077
3078 void perf_output_sample(struct perf_output_handle *handle,
3079                         struct perf_event_header *header,
3080                         struct perf_sample_data *data,
3081                         struct perf_event *event)
3082 {
3083         u64 sample_type = data->type;
3084
3085         perf_output_put(handle, *header);
3086
3087         if (sample_type & PERF_SAMPLE_IP)
3088                 perf_output_put(handle, data->ip);
3089
3090         if (sample_type & PERF_SAMPLE_TID)
3091                 perf_output_put(handle, data->tid_entry);
3092
3093         if (sample_type & PERF_SAMPLE_TIME)
3094                 perf_output_put(handle, data->time);
3095
3096         if (sample_type & PERF_SAMPLE_ADDR)
3097                 perf_output_put(handle, data->addr);
3098
3099         if (sample_type & PERF_SAMPLE_ID)
3100                 perf_output_put(handle, data->id);
3101
3102         if (sample_type & PERF_SAMPLE_STREAM_ID)
3103                 perf_output_put(handle, data->stream_id);
3104
3105         if (sample_type & PERF_SAMPLE_CPU)
3106                 perf_output_put(handle, data->cpu_entry);
3107
3108         if (sample_type & PERF_SAMPLE_PERIOD)
3109                 perf_output_put(handle, data->period);
3110
3111         if (sample_type & PERF_SAMPLE_READ)
3112                 perf_output_read(handle, event);
3113
3114         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3115                 if (data->callchain) {
3116                         int size = 1;
3117
3118                         if (data->callchain)
3119                                 size += data->callchain->nr;
3120
3121                         size *= sizeof(u64);
3122
3123                         perf_output_copy(handle, data->callchain, size);
3124                 } else {
3125                         u64 nr = 0;
3126                         perf_output_put(handle, nr);
3127                 }
3128         }
3129
3130         if (sample_type & PERF_SAMPLE_RAW) {
3131                 if (data->raw) {
3132                         perf_output_put(handle, data->raw->size);
3133                         perf_output_copy(handle, data->raw->data,
3134                                          data->raw->size);
3135                 } else {
3136                         struct {
3137                                 u32     size;
3138                                 u32     data;
3139                         } raw = {
3140                                 .size = sizeof(u32),
3141                                 .data = 0,
3142                         };
3143                         perf_output_put(handle, raw);
3144                 }
3145         }
3146 }
3147
3148 void perf_prepare_sample(struct perf_event_header *header,
3149                          struct perf_sample_data *data,
3150                          struct perf_event *event,
3151                          struct pt_regs *regs)
3152 {
3153         u64 sample_type = event->attr.sample_type;
3154
3155         data->type = sample_type;
3156
3157         header->type = PERF_RECORD_SAMPLE;
3158         header->size = sizeof(*header);
3159
3160         header->misc = 0;
3161         header->misc |= perf_misc_flags(regs);
3162
3163         if (sample_type & PERF_SAMPLE_IP) {
3164                 data->ip = perf_instruction_pointer(regs);
3165
3166                 header->size += sizeof(data->ip);
3167         }
3168
3169         if (sample_type & PERF_SAMPLE_TID) {
3170                 /* namespace issues */
3171                 data->tid_entry.pid = perf_event_pid(event, current);
3172                 data->tid_entry.tid = perf_event_tid(event, current);
3173
3174                 header->size += sizeof(data->tid_entry);
3175         }
3176
3177         if (sample_type & PERF_SAMPLE_TIME) {
3178                 data->time = perf_clock();
3179
3180                 header->size += sizeof(data->time);
3181         }
3182
3183         if (sample_type & PERF_SAMPLE_ADDR)
3184                 header->size += sizeof(data->addr);
3185
3186         if (sample_type & PERF_SAMPLE_ID) {
3187                 data->id = primary_event_id(event);
3188
3189                 header->size += sizeof(data->id);
3190         }
3191
3192         if (sample_type & PERF_SAMPLE_STREAM_ID) {
3193                 data->stream_id = event->id;
3194
3195                 header->size += sizeof(data->stream_id);
3196         }
3197
3198         if (sample_type & PERF_SAMPLE_CPU) {
3199                 data->cpu_entry.cpu             = raw_smp_processor_id();
3200                 data->cpu_entry.reserved        = 0;
3201
3202                 header->size += sizeof(data->cpu_entry);
3203         }
3204
3205         if (sample_type & PERF_SAMPLE_PERIOD)
3206                 header->size += sizeof(data->period);
3207
3208         if (sample_type & PERF_SAMPLE_READ)
3209                 header->size += perf_event_read_size(event);
3210
3211         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3212                 int size = 1;
3213
3214                 data->callchain = perf_callchain(regs);
3215
3216                 if (data->callchain)
3217                         size += data->callchain->nr;
3218
3219                 header->size += size * sizeof(u64);
3220         }
3221
3222         if (sample_type & PERF_SAMPLE_RAW) {
3223                 int size = sizeof(u32);
3224
3225                 if (data->raw)
3226                         size += data->raw->size;
3227                 else
3228                         size += sizeof(u32);
3229
3230                 WARN_ON_ONCE(size & (sizeof(u64)-1));
3231                 header->size += size;
3232         }
3233 }
3234
3235 static void perf_event_output(struct perf_event *event, int nmi,
3236                                 struct perf_sample_data *data,
3237                                 struct pt_regs *regs)
3238 {
3239         struct perf_output_handle handle;
3240         struct perf_event_header header;
3241
3242         perf_prepare_sample(&header, data, event, regs);
3243
3244         if (perf_output_begin(&handle, event, header.size, nmi, 1))
3245                 return;
3246
3247         perf_output_sample(&handle, &header, data, event);
3248
3249         perf_output_end(&handle);
3250 }
3251
3252 /*
3253  * read event_id
3254  */
3255
3256 struct perf_read_event {
3257         struct perf_event_header        header;
3258
3259         u32                             pid;
3260         u32                             tid;
3261 };
3262
3263 static void
3264 perf_event_read_event(struct perf_event *event,
3265                         struct task_struct *task)
3266 {
3267         struct perf_output_handle handle;
3268         struct perf_read_event read_event = {
3269                 .header = {
3270                         .type = PERF_RECORD_READ,
3271                         .misc = 0,
3272                         .size = sizeof(read_event) + perf_event_read_size(event),
3273                 },
3274                 .pid = perf_event_pid(event, task),
3275                 .tid = perf_event_tid(event, task),
3276         };
3277         int ret;
3278
3279         ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3280         if (ret)
3281                 return;
3282
3283         perf_output_put(&handle, read_event);
3284         perf_output_read(&handle, event);
3285
3286         perf_output_end(&handle);
3287 }
3288
3289 /*
3290  * task tracking -- fork/exit
3291  *
3292  * enabled by: attr.comm | attr.mmap | attr.task
3293  */
3294
3295 struct perf_task_event {
3296         struct task_struct              *task;
3297         struct perf_event_context       *task_ctx;
3298
3299         struct {
3300                 struct perf_event_header        header;
3301
3302                 u32                             pid;
3303                 u32                             ppid;
3304                 u32                             tid;
3305                 u32                             ptid;
3306                 u64                             time;
3307         } event_id;
3308 };
3309
3310 static void perf_event_task_output(struct perf_event *event,
3311                                      struct perf_task_event *task_event)
3312 {
3313         struct perf_output_handle handle;
3314         int size;
3315         struct task_struct *task = task_event->task;
3316         int ret;
3317
3318         size  = task_event->event_id.header.size;
3319         ret = perf_output_begin(&handle, event, size, 0, 0);
3320
3321         if (ret)
3322                 return;
3323
3324         task_event->event_id.pid = perf_event_pid(event, task);
3325         task_event->event_id.ppid = perf_event_pid(event, current);
3326
3327         task_event->event_id.tid = perf_event_tid(event, task);
3328         task_event->event_id.ptid = perf_event_tid(event, current);
3329
3330         task_event->event_id.time = perf_clock();
3331
3332         perf_output_put(&handle, task_event->event_id);
3333
3334         perf_output_end(&handle);
3335 }
3336
3337 static int perf_event_task_match(struct perf_event *event)
3338 {
3339         if (event->cpu != -1 && event->cpu != smp_processor_id())
3340                 return 0;
3341
3342         if (event->attr.comm || event->attr.mmap || event->attr.task)
3343                 return 1;
3344
3345         return 0;
3346 }
3347
3348 static void perf_event_task_ctx(struct perf_event_context *ctx,
3349                                   struct perf_task_event *task_event)
3350 {
3351         struct perf_event *event;
3352
3353         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3354                 if (perf_event_task_match(event))
3355                         perf_event_task_output(event, task_event);
3356         }
3357 }
3358
3359 static void perf_event_task_event(struct perf_task_event *task_event)
3360 {
3361         struct perf_cpu_context *cpuctx;
3362         struct perf_event_context *ctx = task_event->task_ctx;
3363
3364         rcu_read_lock();
3365         cpuctx = &get_cpu_var(perf_cpu_context);
3366         perf_event_task_ctx(&cpuctx->ctx, task_event);
3367         if (!ctx)
3368                 ctx = rcu_dereference(task_event->task->perf_event_ctxp);
3369         if (ctx)
3370                 perf_event_task_ctx(ctx, task_event);
3371         put_cpu_var(perf_cpu_context);
3372         rcu_read_unlock();
3373 }
3374
3375 static void perf_event_task(struct task_struct *task,
3376                               struct perf_event_context *task_ctx,
3377                               int new)
3378 {
3379         struct perf_task_event task_event;
3380
3381         if (!atomic_read(&nr_comm_events) &&
3382             !atomic_read(&nr_mmap_events) &&
3383             !atomic_read(&nr_task_events))
3384                 return;
3385
3386         task_event = (struct perf_task_event){
3387                 .task     = task,
3388                 .task_ctx = task_ctx,
3389                 .event_id    = {
3390                         .header = {
3391                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3392                                 .misc = 0,
3393                                 .size = sizeof(task_event.event_id),
3394                         },
3395                         /* .pid  */
3396                         /* .ppid */
3397                         /* .tid  */
3398                         /* .ptid */
3399                 },
3400         };
3401
3402         perf_event_task_event(&task_event);
3403 }
3404
3405 void perf_event_fork(struct task_struct *task)
3406 {
3407         perf_event_task(task, NULL, 1);
3408 }
3409
3410 /*
3411  * comm tracking
3412  */
3413
3414 struct perf_comm_event {
3415         struct task_struct      *task;
3416         char                    *comm;
3417         int                     comm_size;
3418
3419         struct {
3420                 struct perf_event_header        header;
3421
3422                 u32                             pid;
3423                 u32                             tid;
3424         } event_id;
3425 };
3426
3427 static void perf_event_comm_output(struct perf_event *event,
3428                                      struct perf_comm_event *comm_event)
3429 {
3430         struct perf_output_handle handle;
3431         int size = comm_event->event_id.header.size;
3432         int ret = perf_output_begin(&handle, event, size, 0, 0);
3433
3434         if (ret)
3435                 return;
3436
3437         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3438         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3439
3440         perf_output_put(&handle, comm_event->event_id);
3441         perf_output_copy(&handle, comm_event->comm,
3442                                    comm_event->comm_size);
3443         perf_output_end(&handle);
3444 }
3445
3446 static int perf_event_comm_match(struct perf_event *event)
3447 {
3448         if (event->cpu != -1 && event->cpu != smp_processor_id())
3449                 return 0;
3450
3451         if (event->attr.comm)
3452                 return 1;
3453
3454         return 0;
3455 }
3456
3457 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3458                                   struct perf_comm_event *comm_event)
3459 {
3460         struct perf_event *event;
3461
3462         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3463                 if (perf_event_comm_match(event))
3464                         perf_event_comm_output(event, comm_event);
3465         }
3466 }
3467
3468 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3469 {
3470         struct perf_cpu_context *cpuctx;
3471         struct perf_event_context *ctx;
3472         unsigned int size;
3473         char comm[TASK_COMM_LEN];
3474
3475         memset(comm, 0, sizeof(comm));
3476         strlcpy(comm, comm_event->task->comm, sizeof(comm));
3477         size = ALIGN(strlen(comm)+1, sizeof(u64));
3478
3479         comm_event->comm = comm;
3480         comm_event->comm_size = size;
3481
3482         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3483
3484         rcu_read_lock();
3485         cpuctx = &get_cpu_var(perf_cpu_context);
3486         perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3487         ctx = rcu_dereference(current->perf_event_ctxp);
3488         if (ctx)
3489                 perf_event_comm_ctx(ctx, comm_event);
3490         put_cpu_var(perf_cpu_context);
3491         rcu_read_unlock();
3492 }
3493
3494 void perf_event_comm(struct task_struct *task)
3495 {
3496         struct perf_comm_event comm_event;
3497
3498         if (task->perf_event_ctxp)
3499                 perf_event_enable_on_exec(task);
3500
3501         if (!atomic_read(&nr_comm_events))
3502                 return;
3503
3504         comm_event = (struct perf_comm_event){
3505                 .task   = task,
3506                 /* .comm      */
3507                 /* .comm_size */
3508                 .event_id  = {
3509                         .header = {
3510                                 .type = PERF_RECORD_COMM,
3511                                 .misc = 0,
3512                                 /* .size */
3513                         },
3514                         /* .pid */
3515                         /* .tid */
3516                 },
3517         };
3518
3519         perf_event_comm_event(&comm_event);
3520 }
3521
3522 /*
3523  * mmap tracking
3524  */
3525
3526 struct perf_mmap_event {
3527         struct vm_area_struct   *vma;
3528
3529         const char              *file_name;
3530         int                     file_size;
3531
3532         struct {
3533                 struct perf_event_header        header;
3534
3535                 u32                             pid;
3536                 u32                             tid;
3537                 u64                             start;
3538                 u64                             len;
3539                 u64                             pgoff;
3540         } event_id;
3541 };
3542
3543 static void perf_event_mmap_output(struct perf_event *event,
3544                                      struct perf_mmap_event *mmap_event)
3545 {
3546         struct perf_output_handle handle;
3547         int size = mmap_event->event_id.header.size;
3548         int ret = perf_output_begin(&handle, event, size, 0, 0);
3549
3550         if (ret)
3551                 return;
3552
3553         mmap_event->event_id.pid = perf_event_pid(event, current);
3554         mmap_event->event_id.tid = perf_event_tid(event, current);
3555
3556         perf_output_put(&handle, mmap_event->event_id);
3557         perf_output_copy(&handle, mmap_event->file_name,
3558                                    mmap_event->file_size);
3559         perf_output_end(&handle);
3560 }
3561
3562 static int perf_event_mmap_match(struct perf_event *event,
3563                                    struct perf_mmap_event *mmap_event)
3564 {
3565         if (event->cpu != -1 && event->cpu != smp_processor_id())
3566                 return 0;
3567
3568         if (event->attr.mmap)
3569                 return 1;
3570
3571         return 0;
3572 }
3573
3574 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3575                                   struct perf_mmap_event *mmap_event)
3576 {
3577         struct perf_event *event;
3578
3579         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3580                 if (perf_event_mmap_match(event, mmap_event))
3581                         perf_event_mmap_output(event, mmap_event);
3582         }
3583 }
3584
3585 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3586 {
3587         struct perf_cpu_context *cpuctx;
3588         struct perf_event_context *ctx;
3589         struct vm_area_struct *vma = mmap_event->vma;
3590         struct file *file = vma->vm_file;
3591         unsigned int size;
3592         char tmp[16];
3593         char *buf = NULL;
3594         const char *name;
3595
3596         memset(tmp, 0, sizeof(tmp));
3597
3598         if (file) {
3599                 /*
3600                  * d_path works from the end of the buffer backwards, so we
3601                  * need to add enough zero bytes after the string to handle
3602                  * the 64bit alignment we do later.
3603                  */
3604                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3605                 if (!buf) {
3606                         name = strncpy(tmp, "//enomem", sizeof(tmp));
3607                         goto got_name;
3608                 }
3609                 name = d_path(&file->f_path, buf, PATH_MAX);
3610                 if (IS_ERR(name)) {
3611                         name = strncpy(tmp, "//toolong", sizeof(tmp));
3612                         goto got_name;
3613                 }
3614         } else {
3615                 if (arch_vma_name(mmap_event->vma)) {
3616                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3617                                        sizeof(tmp));
3618                         goto got_name;
3619                 }
3620
3621                 if (!vma->vm_mm) {
3622                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
3623                         goto got_name;
3624                 }
3625
3626                 name = strncpy(tmp, "//anon", sizeof(tmp));
3627                 goto got_name;
3628         }
3629
3630 got_name:
3631         size = ALIGN(strlen(name)+1, sizeof(u64));
3632
3633         mmap_event->file_name = name;
3634         mmap_event->file_size = size;
3635
3636         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3637
3638         rcu_read_lock();
3639         cpuctx = &get_cpu_var(perf_cpu_context);
3640         perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3641         ctx = rcu_dereference(current->perf_event_ctxp);
3642         if (ctx)
3643                 perf_event_mmap_ctx(ctx, mmap_event);
3644         put_cpu_var(perf_cpu_context);
3645         rcu_read_unlock();
3646
3647         kfree(buf);
3648 }
3649
3650 void __perf_event_mmap(struct vm_area_struct *vma)
3651 {
3652         struct perf_mmap_event mmap_event;
3653
3654         if (!atomic_read(&nr_mmap_events))
3655                 return;
3656
3657         mmap_event = (struct perf_mmap_event){
3658                 .vma    = vma,
3659                 /* .file_name */
3660                 /* .file_size */
3661                 .event_id  = {
3662                         .header = {
3663                                 .type = PERF_RECORD_MMAP,
3664                                 .misc = 0,
3665                                 /* .size */
3666                         },
3667                         /* .pid */
3668                         /* .tid */
3669                         .start  = vma->vm_start,
3670                         .len    = vma->vm_end - vma->vm_start,
3671                         .pgoff  = vma->vm_pgoff,
3672                 },
3673         };
3674
3675         perf_event_mmap_event(&mmap_event);
3676 }
3677
3678 /*
3679  * IRQ throttle logging
3680  */
3681
3682 static void perf_log_throttle(struct perf_event *event, int enable)
3683 {
3684         struct perf_output_handle handle;
3685         int ret;
3686
3687         struct {
3688                 struct perf_event_header        header;
3689                 u64                             time;
3690                 u64                             id;
3691                 u64                             stream_id;
3692         } throttle_event = {
3693                 .header = {
3694                         .type = PERF_RECORD_THROTTLE,
3695                         .misc = 0,
3696                         .size = sizeof(throttle_event),
3697                 },
3698                 .time           = perf_clock(),
3699                 .id             = primary_event_id(event),
3700                 .stream_id      = event->id,
3701         };
3702
3703         if (enable)
3704                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3705
3706         ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3707         if (ret)
3708                 return;
3709
3710         perf_output_put(&handle, throttle_event);
3711         perf_output_end(&handle);
3712 }
3713
3714 /*
3715  * Generic event overflow handling, sampling.
3716  */
3717
3718 static int __perf_event_overflow(struct perf_event *event, int nmi,
3719                                    int throttle, struct perf_sample_data *data,
3720                                    struct pt_regs *regs)
3721 {
3722         int events = atomic_read(&event->event_limit);
3723         struct hw_perf_event *hwc = &event->hw;
3724         int ret = 0;
3725
3726         throttle = (throttle && event->pmu->unthrottle != NULL);
3727
3728         if (!throttle) {
3729                 hwc->interrupts++;
3730         } else {
3731                 if (hwc->interrupts != MAX_INTERRUPTS) {
3732                         hwc->interrupts++;
3733                         if (HZ * hwc->interrupts >
3734                                         (u64)sysctl_perf_event_sample_rate) {
3735                                 hwc->interrupts = MAX_INTERRUPTS;
3736                                 perf_log_throttle(event, 0);
3737                                 ret = 1;
3738                         }
3739                 } else {
3740                         /*
3741                          * Keep re-disabling events even though on the previous
3742                          * pass we disabled it - just in case we raced with a
3743                          * sched-in and the event got enabled again:
3744                          */
3745                         ret = 1;
3746                 }
3747         }
3748
3749         if (event->attr.freq) {
3750                 u64 now = perf_clock();
3751                 s64 delta = now - hwc->freq_stamp;
3752
3753                 hwc->freq_stamp = now;
3754
3755                 if (delta > 0 && delta < TICK_NSEC)
3756                         perf_adjust_period(event, NSEC_PER_SEC / (int)delta);
3757         }
3758
3759         /*
3760          * XXX event_limit might not quite work as expected on inherited
3761          * events
3762          */
3763
3764         event->pending_kill = POLL_IN;
3765         if (events && atomic_dec_and_test(&event->event_limit)) {
3766                 ret = 1;
3767                 event->pending_kill = POLL_HUP;
3768                 if (nmi) {
3769                         event->pending_disable = 1;
3770                         perf_pending_queue(&event->pending,
3771                                            perf_pending_event);
3772                 } else
3773                         perf_event_disable(event);
3774         }
3775
3776         if (event->overflow_handler)
3777                 event->overflow_handler(event, nmi, data, regs);
3778         else
3779                 perf_event_output(event, nmi, data, regs);
3780
3781         return ret;
3782 }
3783
3784 int perf_event_overflow(struct perf_event *event, int nmi,
3785                           struct perf_sample_data *data,
3786                           struct pt_regs *regs)
3787 {
3788         return __perf_event_overflow(event, nmi, 1, data, regs);
3789 }
3790
3791 /*
3792  * Generic software event infrastructure
3793  */
3794
3795 /*
3796  * We directly increment event->count and keep a second value in
3797  * event->hw.period_left to count intervals. This period event
3798  * is kept in the range [-sample_period, 0] so that we can use the
3799  * sign as trigger.
3800  */
3801
3802 static u64 perf_swevent_set_period(struct perf_event *event)
3803 {
3804         struct hw_perf_event *hwc = &event->hw;
3805         u64 period = hwc->last_period;
3806         u64 nr, offset;
3807         s64 old, val;
3808
3809         hwc->last_period = hwc->sample_period;
3810
3811 again:
3812         old = val = atomic64_read(&hwc->period_left);
3813         if (val < 0)
3814                 return 0;
3815
3816         nr = div64_u64(period + val, period);
3817         offset = nr * period;
3818         val -= offset;
3819         if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3820                 goto again;
3821
3822         return nr;
3823 }
3824
3825 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3826                                     int nmi, struct perf_sample_data *data,
3827                                     struct pt_regs *regs)
3828 {
3829         struct hw_perf_event *hwc = &event->hw;
3830         int throttle = 0;
3831
3832         data->period = event->hw.last_period;
3833         if (!overflow)
3834                 overflow = perf_swevent_set_period(event);
3835
3836         if (hwc->interrupts == MAX_INTERRUPTS)
3837                 return;
3838
3839         for (; overflow; overflow--) {
3840                 if (__perf_event_overflow(event, nmi, throttle,
3841                                             data, regs)) {
3842                         /*
3843                          * We inhibit the overflow from happening when
3844                          * hwc->interrupts == MAX_INTERRUPTS.
3845                          */
3846                         break;
3847                 }
3848                 throttle = 1;
3849         }
3850 }
3851
3852 static void perf_swevent_unthrottle(struct perf_event *event)
3853 {
3854         /*
3855          * Nothing to do, we already reset hwc->interrupts.
3856          */
3857 }
3858
3859 static void perf_swevent_add(struct perf_event *event, u64 nr,
3860                                int nmi, struct perf_sample_data *data,
3861                                struct pt_regs *regs)
3862 {
3863         struct hw_perf_event *hwc = &event->hw;
3864
3865         atomic64_add(nr, &event->count);
3866
3867         if (!regs)
3868                 return;
3869
3870         if (!hwc->sample_period)
3871                 return;
3872
3873         if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3874                 return perf_swevent_overflow(event, 1, nmi, data, regs);
3875
3876         if (atomic64_add_negative(nr, &hwc->period_left))
3877                 return;
3878
3879         perf_swevent_overflow(event, 0, nmi, data, regs);
3880 }
3881
3882 static int perf_swevent_is_counting(struct perf_event *event)
3883 {
3884         /*
3885          * The event is active, we're good!
3886          */
3887         if (event->state == PERF_EVENT_STATE_ACTIVE)
3888                 return 1;
3889
3890         /*
3891          * The event is off/error, not counting.
3892          */
3893         if (event->state != PERF_EVENT_STATE_INACTIVE)
3894                 return 0;
3895
3896         /*
3897          * The event is inactive, if the context is active
3898          * we're part of a group that didn't make it on the 'pmu',
3899          * not counting.
3900          */
3901         if (event->ctx->is_active)
3902                 return 0;
3903
3904         /*
3905          * We're inactive and the context is too, this means the
3906          * task is scheduled out, we're counting events that happen
3907          * to us, like migration events.
3908          */
3909         return 1;
3910 }
3911
3912 static int perf_tp_event_match(struct perf_event *event,
3913                                 struct perf_sample_data *data);
3914
3915 static int perf_exclude_event(struct perf_event *event,
3916                               struct pt_regs *regs)
3917 {
3918         if (regs) {
3919                 if (event->attr.exclude_user && user_mode(regs))
3920                         return 1;
3921
3922                 if (event->attr.exclude_kernel && !user_mode(regs))
3923                         return 1;
3924         }
3925
3926         return 0;
3927 }
3928
3929 static int perf_swevent_match(struct perf_event *event,
3930                                 enum perf_type_id type,
3931                                 u32 event_id,
3932                                 struct perf_sample_data *data,
3933                                 struct pt_regs *regs)
3934 {
3935         if (event->cpu != -1 && event->cpu != smp_processor_id())
3936                 return 0;
3937
3938         if (!perf_swevent_is_counting(event))
3939                 return 0;
3940
3941         if (event->attr.type != type)
3942                 return 0;
3943
3944         if (event->attr.config != event_id)
3945                 return 0;
3946
3947         if (perf_exclude_event(event, regs))
3948                 return 0;
3949
3950         if (event->attr.type == PERF_TYPE_TRACEPOINT &&
3951             !perf_tp_event_match(event, data))
3952                 return 0;
3953
3954         return 1;
3955 }
3956
3957 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
3958                                      enum perf_type_id type,
3959                                      u32 event_id, u64 nr, int nmi,
3960                                      struct perf_sample_data *data,
3961                                      struct pt_regs *regs)
3962 {
3963         struct perf_event *event;
3964
3965         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3966                 if (perf_swevent_match(event, type, event_id, data, regs))
3967                         perf_swevent_add(event, nr, nmi, data, regs);
3968         }
3969 }
3970
3971 int perf_swevent_get_recursion_context(void)
3972 {
3973         struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3974         int rctx;
3975
3976         if (in_nmi())
3977                 rctx = 3;
3978         else if (in_irq())
3979                 rctx = 2;
3980         else if (in_softirq())
3981                 rctx = 1;
3982         else
3983                 rctx = 0;
3984
3985         if (cpuctx->recursion[rctx]) {
3986                 put_cpu_var(perf_cpu_context);
3987                 return -1;
3988         }
3989
3990         cpuctx->recursion[rctx]++;
3991         barrier();
3992
3993         return rctx;
3994 }
3995 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
3996
3997 void perf_swevent_put_recursion_context(int rctx)
3998 {
3999         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4000         barrier();
4001         cpuctx->recursion[rctx]--;
4002         put_cpu_var(perf_cpu_context);
4003 }
4004 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
4005
4006 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4007                                     u64 nr, int nmi,
4008                                     struct perf_sample_data *data,
4009                                     struct pt_regs *regs)
4010 {
4011         struct perf_cpu_context *cpuctx;
4012         struct perf_event_context *ctx;
4013
4014         cpuctx = &__get_cpu_var(perf_cpu_context);
4015         rcu_read_lock();
4016         perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
4017                                  nr, nmi, data, regs);
4018         /*
4019          * doesn't really matter which of the child contexts the
4020          * events ends up in.
4021          */
4022         ctx = rcu_dereference(current->perf_event_ctxp);
4023         if (ctx)
4024                 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
4025         rcu_read_unlock();
4026 }
4027
4028 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4029                             struct pt_regs *regs, u64 addr)
4030 {
4031         struct perf_sample_data data;
4032         int rctx;
4033
4034         rctx = perf_swevent_get_recursion_context();
4035         if (rctx < 0)
4036                 return;
4037
4038         data.addr = addr;
4039         data.raw  = NULL;
4040
4041         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4042
4043         perf_swevent_put_recursion_context(rctx);
4044 }
4045
4046 static void perf_swevent_read(struct perf_event *event)
4047 {
4048 }
4049
4050 static int perf_swevent_enable(struct perf_event *event)
4051 {
4052         struct hw_perf_event *hwc = &event->hw;
4053
4054         if (hwc->sample_period) {
4055                 hwc->last_period = hwc->sample_period;
4056                 perf_swevent_set_period(event);
4057         }
4058         return 0;
4059 }
4060
4061 static void perf_swevent_disable(struct perf_event *event)
4062 {
4063 }
4064
4065 static const struct pmu perf_ops_generic = {
4066         .enable         = perf_swevent_enable,
4067         .disable        = perf_swevent_disable,
4068         .read           = perf_swevent_read,
4069         .unthrottle     = perf_swevent_unthrottle,
4070 };
4071
4072 /*
4073  * hrtimer based swevent callback
4074  */
4075
4076 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4077 {
4078         enum hrtimer_restart ret = HRTIMER_RESTART;
4079         struct perf_sample_data data;
4080         struct pt_regs *regs;
4081         struct perf_event *event;
4082         u64 period;
4083
4084         event   = container_of(hrtimer, struct perf_event, hw.hrtimer);
4085         event->pmu->read(event);
4086
4087         data.addr = 0;
4088         data.raw = NULL;
4089         data.period = event->hw.last_period;
4090         regs = get_irq_regs();
4091         /*
4092          * In case we exclude kernel IPs or are somehow not in interrupt
4093          * context, provide the next best thing, the user IP.
4094          */
4095         if ((event->attr.exclude_kernel || !regs) &&
4096                         !event->attr.exclude_user)
4097                 regs = task_pt_regs(current);
4098
4099         if (regs) {
4100                 if (!(event->attr.exclude_idle && current->pid == 0))
4101                         if (perf_event_overflow(event, 0, &data, regs))
4102                                 ret = HRTIMER_NORESTART;
4103         }
4104
4105         period = max_t(u64, 10000, event->hw.sample_period);
4106         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4107
4108         return ret;
4109 }
4110
4111 static void perf_swevent_start_hrtimer(struct perf_event *event)
4112 {
4113         struct hw_perf_event *hwc = &event->hw;
4114
4115         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4116         hwc->hrtimer.function = perf_swevent_hrtimer;
4117         if (hwc->sample_period) {
4118                 u64 period;
4119
4120                 if (hwc->remaining) {
4121                         if (hwc->remaining < 0)
4122                                 period = 10000;
4123                         else
4124                                 period = hwc->remaining;
4125                         hwc->remaining = 0;
4126                 } else {
4127                         period = max_t(u64, 10000, hwc->sample_period);
4128                 }
4129                 __hrtimer_start_range_ns(&hwc->hrtimer,
4130                                 ns_to_ktime(period), 0,
4131                                 HRTIMER_MODE_REL, 0);
4132         }
4133 }
4134
4135 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4136 {
4137         struct hw_perf_event *hwc = &event->hw;
4138
4139         if (hwc->sample_period) {
4140                 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4141                 hwc->remaining = ktime_to_ns(remaining);
4142
4143                 hrtimer_cancel(&hwc->hrtimer);
4144         }
4145 }
4146
4147 /*
4148  * Software event: cpu wall time clock
4149  */
4150
4151 static void cpu_clock_perf_event_update(struct perf_event *event)
4152 {
4153         int cpu = raw_smp_processor_id();
4154         s64 prev;
4155         u64 now;
4156
4157         now = cpu_clock(cpu);
4158         prev = atomic64_xchg(&event->hw.prev_count, now);
4159         atomic64_add(now - prev, &event->count);
4160 }
4161
4162 static int cpu_clock_perf_event_enable(struct perf_event *event)
4163 {
4164         struct hw_perf_event *hwc = &event->hw;
4165         int cpu = raw_smp_processor_id();
4166
4167         atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4168         perf_swevent_start_hrtimer(event);
4169
4170         return 0;
4171 }
4172
4173 static void cpu_clock_perf_event_disable(struct perf_event *event)
4174 {
4175         perf_swevent_cancel_hrtimer(event);
4176         cpu_clock_perf_event_update(event);
4177 }
4178
4179 static void cpu_clock_perf_event_read(struct perf_event *event)
4180 {
4181         cpu_clock_perf_event_update(event);
4182 }
4183
4184 static const struct pmu perf_ops_cpu_clock = {
4185         .enable         = cpu_clock_perf_event_enable,
4186         .disable        = cpu_clock_perf_event_disable,
4187         .read           = cpu_clock_perf_event_read,
4188 };
4189
4190 /*
4191  * Software event: task time clock
4192  */
4193
4194 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4195 {
4196         u64 prev;
4197         s64 delta;
4198
4199         prev = atomic64_xchg(&event->hw.prev_count, now);
4200         delta = now - prev;
4201         atomic64_add(delta, &event->count);
4202 }
4203
4204 static int task_clock_perf_event_enable(struct perf_event *event)
4205 {
4206         struct hw_perf_event *hwc = &event->hw;
4207         u64 now;
4208
4209         now = event->ctx->time;
4210
4211         atomic64_set(&hwc->prev_count, now);
4212
4213         perf_swevent_start_hrtimer(event);
4214
4215         return 0;
4216 }
4217
4218 static void task_clock_perf_event_disable(struct perf_event *event)
4219 {
4220         perf_swevent_cancel_hrtimer(event);
4221         task_clock_perf_event_update(event, event->ctx->time);
4222
4223 }
4224
4225 static void task_clock_perf_event_read(struct perf_event *event)
4226 {
4227         u64 time;
4228
4229         if (!in_nmi()) {
4230                 update_context_time(event->ctx);
4231                 time = event->ctx->time;
4232         } else {
4233                 u64 now = perf_clock();
4234                 u64 delta = now - event->ctx->timestamp;
4235                 time = event->ctx->time + delta;
4236         }
4237
4238         task_clock_perf_event_update(event, time);
4239 }
4240
4241 static const struct pmu perf_ops_task_clock = {
4242         .enable         = task_clock_perf_event_enable,
4243         .disable        = task_clock_perf_event_disable,
4244         .read           = task_clock_perf_event_read,
4245 };
4246
4247 #ifdef CONFIG_EVENT_TRACING
4248
4249 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4250                           int entry_size)
4251 {
4252         struct perf_raw_record raw = {
4253                 .size = entry_size,
4254                 .data = record,
4255         };
4256
4257         struct perf_sample_data data = {
4258                 .addr = addr,
4259                 .raw = &raw,
4260         };
4261
4262         struct pt_regs *regs = get_irq_regs();
4263
4264         if (!regs)
4265                 regs = task_pt_regs(current);
4266
4267         /* Trace events already protected against recursion */
4268         do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4269                                 &data, regs);
4270 }
4271 EXPORT_SYMBOL_GPL(perf_tp_event);
4272
4273 static int perf_tp_event_match(struct perf_event *event,
4274                                 struct perf_sample_data *data)
4275 {
4276         void *record = data->raw->data;
4277
4278         if (likely(!event->filter) || filter_match_preds(event->filter, record))
4279                 return 1;
4280         return 0;
4281 }
4282
4283 static void tp_perf_event_destroy(struct perf_event *event)
4284 {
4285         ftrace_profile_disable(event->attr.config);
4286 }
4287
4288 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4289 {
4290         /*
4291          * Raw tracepoint data is a severe data leak, only allow root to
4292          * have these.
4293          */
4294         if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4295                         perf_paranoid_tracepoint_raw() &&
4296                         !capable(CAP_SYS_ADMIN))
4297                 return ERR_PTR(-EPERM);
4298
4299         if (ftrace_profile_enable(event->attr.config))
4300                 return NULL;
4301
4302         event->destroy = tp_perf_event_destroy;
4303
4304         return &perf_ops_generic;
4305 }
4306
4307 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4308 {
4309         char *filter_str;
4310         int ret;
4311
4312         if (event->attr.type != PERF_TYPE_TRACEPOINT)
4313                 return -EINVAL;
4314
4315         filter_str = strndup_user(arg, PAGE_SIZE);
4316         if (IS_ERR(filter_str))
4317                 return PTR_ERR(filter_str);
4318
4319         ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4320
4321         kfree(filter_str);
4322         return ret;
4323 }
4324
4325 static void perf_event_free_filter(struct perf_event *event)
4326 {
4327         ftrace_profile_free_filter(event);
4328 }
4329
4330 #else
4331
4332 static int perf_tp_event_match(struct perf_event *event,
4333                                 struct perf_sample_data *data)
4334 {
4335         return 1;
4336 }
4337
4338 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4339 {
4340         return NULL;
4341 }
4342
4343 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4344 {
4345         return -ENOENT;
4346 }
4347
4348 static void perf_event_free_filter(struct perf_event *event)
4349 {
4350 }
4351
4352 #endif /* CONFIG_EVENT_TRACING */
4353
4354 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4355 static void bp_perf_event_destroy(struct perf_event *event)
4356 {
4357         release_bp_slot(event);
4358 }
4359
4360 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4361 {
4362         int err;
4363
4364         err = register_perf_hw_breakpoint(bp);
4365         if (err)
4366                 return ERR_PTR(err);
4367
4368         bp->destroy = bp_perf_event_destroy;
4369
4370         return &perf_ops_bp;
4371 }
4372
4373 void perf_bp_event(struct perf_event *bp, void *data)
4374 {
4375         struct perf_sample_data sample;
4376         struct pt_regs *regs = data;
4377
4378         sample.raw = NULL;
4379         sample.addr = bp->attr.bp_addr;
4380
4381         if (!perf_exclude_event(bp, regs))
4382                 perf_swevent_add(bp, 1, 1, &sample, regs);
4383 }
4384 #else
4385 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4386 {
4387         return NULL;
4388 }
4389
4390 void perf_bp_event(struct perf_event *bp, void *regs)
4391 {
4392 }
4393 #endif
4394
4395 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4396
4397 static void sw_perf_event_destroy(struct perf_event *event)
4398 {
4399         u64 event_id = event->attr.config;
4400
4401         WARN_ON(event->parent);
4402
4403         atomic_dec(&perf_swevent_enabled[event_id]);
4404 }
4405
4406 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4407 {
4408         const struct pmu *pmu = NULL;
4409         u64 event_id = event->attr.config;
4410
4411         /*
4412          * Software events (currently) can't in general distinguish
4413          * between user, kernel and hypervisor events.
4414          * However, context switches and cpu migrations are considered
4415          * to be kernel events, and page faults are never hypervisor
4416          * events.
4417          */
4418         switch (event_id) {
4419         case PERF_COUNT_SW_CPU_CLOCK:
4420                 pmu = &perf_ops_cpu_clock;
4421
4422                 break;
4423         case PERF_COUNT_SW_TASK_CLOCK:
4424                 /*
4425                  * If the user instantiates this as a per-cpu event,
4426                  * use the cpu_clock event instead.
4427                  */
4428                 if (event->ctx->task)
4429                         pmu = &perf_ops_task_clock;
4430                 else
4431                         pmu = &perf_ops_cpu_clock;
4432
4433                 break;
4434         case PERF_COUNT_SW_PAGE_FAULTS:
4435         case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4436         case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4437         case PERF_COUNT_SW_CONTEXT_SWITCHES:
4438         case PERF_COUNT_SW_CPU_MIGRATIONS:
4439         case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4440         case PERF_COUNT_SW_EMULATION_FAULTS:
4441                 if (!event->parent) {
4442                         atomic_inc(&perf_swevent_enabled[event_id]);
4443                         event->destroy = sw_perf_event_destroy;
4444                 }
4445                 pmu = &perf_ops_generic;
4446                 break;
4447         }
4448
4449         return pmu;
4450 }
4451
4452 /*
4453  * Allocate and initialize a event structure
4454  */
4455 static struct perf_event *
4456 perf_event_alloc(struct perf_event_attr *attr,
4457                    int cpu,
4458                    struct perf_event_context *ctx,
4459                    struct perf_event *group_leader,
4460                    struct perf_event *parent_event,
4461                    perf_overflow_handler_t overflow_handler,
4462                    gfp_t gfpflags)
4463 {
4464         const struct pmu *pmu;
4465         struct perf_event *event;
4466         struct hw_perf_event *hwc;
4467         long err;
4468
4469         event = kzalloc(sizeof(*event), gfpflags);
4470         if (!event)
4471                 return ERR_PTR(-ENOMEM);
4472
4473         /*
4474          * Single events are their own group leaders, with an
4475          * empty sibling list:
4476          */
4477         if (!group_leader)
4478                 group_leader = event;
4479
4480         mutex_init(&event->child_mutex);
4481         INIT_LIST_HEAD(&event->child_list);
4482
4483         INIT_LIST_HEAD(&event->group_entry);
4484         INIT_LIST_HEAD(&event->event_entry);
4485         INIT_LIST_HEAD(&event->sibling_list);
4486         init_waitqueue_head(&event->waitq);
4487
4488         mutex_init(&event->mmap_mutex);
4489
4490         event->cpu              = cpu;
4491         event->attr             = *attr;
4492         event->group_leader     = group_leader;
4493         event->pmu              = NULL;
4494         event->ctx              = ctx;
4495         event->oncpu            = -1;
4496
4497         event->parent           = parent_event;
4498
4499         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
4500         event->id               = atomic64_inc_return(&perf_event_id);
4501
4502         event->state            = PERF_EVENT_STATE_INACTIVE;
4503
4504         if (!overflow_handler && parent_event)
4505                 overflow_handler = parent_event->overflow_handler;
4506         
4507         event->overflow_handler = overflow_handler;
4508
4509         if (attr->disabled)
4510                 event->state = PERF_EVENT_STATE_OFF;
4511
4512         pmu = NULL;
4513
4514         hwc = &event->hw;
4515         hwc->sample_period = attr->sample_period;
4516         if (attr->freq && attr->sample_freq)
4517                 hwc->sample_period = 1;
4518         hwc->last_period = hwc->sample_period;
4519
4520         atomic64_set(&hwc->period_left, hwc->sample_period);
4521
4522         /*
4523          * we currently do not support PERF_FORMAT_GROUP on inherited events
4524          */
4525         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4526                 goto done;
4527
4528         switch (attr->type) {
4529         case PERF_TYPE_RAW:
4530         case PERF_TYPE_HARDWARE:
4531         case PERF_TYPE_HW_CACHE:
4532                 pmu = hw_perf_event_init(event);
4533                 break;
4534
4535         case PERF_TYPE_SOFTWARE:
4536                 pmu = sw_perf_event_init(event);
4537                 break;
4538
4539         case PERF_TYPE_TRACEPOINT:
4540                 pmu = tp_perf_event_init(event);
4541                 break;
4542
4543         case PERF_TYPE_BREAKPOINT:
4544                 pmu = bp_perf_event_init(event);
4545                 break;
4546
4547
4548         default:
4549                 break;
4550         }
4551 done:
4552         err = 0;
4553         if (!pmu)
4554                 err = -EINVAL;
4555         else if (IS_ERR(pmu))
4556                 err = PTR_ERR(pmu);
4557
4558         if (err) {
4559                 if (event->ns)
4560                         put_pid_ns(event->ns);
4561                 kfree(event);
4562                 return ERR_PTR(err);
4563         }
4564
4565         event->pmu = pmu;
4566
4567         if (!event->parent) {
4568                 atomic_inc(&nr_events);
4569                 if (event->attr.mmap)
4570                         atomic_inc(&nr_mmap_events);
4571                 if (event->attr.comm)
4572                         atomic_inc(&nr_comm_events);
4573                 if (event->attr.task)
4574                         atomic_inc(&nr_task_events);
4575         }
4576
4577         return event;
4578 }
4579
4580 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4581                           struct perf_event_attr *attr)
4582 {
4583         u32 size;
4584         int ret;
4585
4586         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4587                 return -EFAULT;
4588
4589         /*
4590          * zero the full structure, so that a short copy will be nice.
4591          */
4592         memset(attr, 0, sizeof(*attr));
4593
4594         ret = get_user(size, &uattr->size);
4595         if (ret)
4596                 return ret;
4597
4598         if (size > PAGE_SIZE)   /* silly large */
4599                 goto err_size;
4600
4601         if (!size)              /* abi compat */
4602                 size = PERF_ATTR_SIZE_VER0;
4603
4604         if (size < PERF_ATTR_SIZE_VER0)
4605                 goto err_size;
4606
4607         /*
4608          * If we're handed a bigger struct than we know of,
4609          * ensure all the unknown bits are 0 - i.e. new
4610          * user-space does not rely on any kernel feature
4611          * extensions we dont know about yet.
4612          */
4613         if (size > sizeof(*attr)) {
4614                 unsigned char __user *addr;
4615                 unsigned char __user *end;
4616                 unsigned char val;
4617
4618                 addr = (void __user *)uattr + sizeof(*attr);
4619                 end  = (void __user *)uattr + size;
4620
4621                 for (; addr < end; addr++) {
4622                         ret = get_user(val, addr);
4623                         if (ret)
4624                                 return ret;
4625                         if (val)
4626                                 goto err_size;
4627                 }
4628                 size = sizeof(*attr);
4629         }
4630
4631         ret = copy_from_user(attr, uattr, size);
4632         if (ret)
4633                 return -EFAULT;
4634
4635         /*
4636          * If the type exists, the corresponding creation will verify
4637          * the attr->config.
4638          */
4639         if (attr->type >= PERF_TYPE_MAX)
4640                 return -EINVAL;
4641
4642         if (attr->__reserved_1 || attr->__reserved_2)
4643                 return -EINVAL;
4644
4645         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4646                 return -EINVAL;
4647
4648         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4649                 return -EINVAL;
4650
4651 out:
4652         return ret;
4653
4654 err_size:
4655         put_user(sizeof(*attr), &uattr->size);
4656         ret = -E2BIG;
4657         goto out;
4658 }
4659
4660 static int perf_event_set_output(struct perf_event *event, int output_fd)
4661 {
4662         struct perf_event *output_event = NULL;
4663         struct file *output_file = NULL;
4664         struct perf_event *old_output;
4665         int fput_needed = 0;
4666         int ret = -EINVAL;
4667
4668         if (!output_fd)
4669                 goto set;
4670
4671         output_file = fget_light(output_fd, &fput_needed);
4672         if (!output_file)
4673                 return -EBADF;
4674
4675         if (output_file->f_op != &perf_fops)
4676                 goto out;
4677
4678         output_event = output_file->private_data;
4679
4680         /* Don't chain output fds */
4681         if (output_event->output)
4682                 goto out;
4683
4684         /* Don't set an output fd when we already have an output channel */
4685         if (event->data)
4686                 goto out;
4687
4688         atomic_long_inc(&output_file->f_count);
4689
4690 set:
4691         mutex_lock(&event->mmap_mutex);
4692         old_output = event->output;
4693         rcu_assign_pointer(event->output, output_event);
4694         mutex_unlock(&event->mmap_mutex);
4695
4696         if (old_output) {
4697                 /*
4698                  * we need to make sure no existing perf_output_*()
4699                  * is still referencing this event.
4700                  */
4701                 synchronize_rcu();
4702                 fput(old_output->filp);
4703         }
4704
4705         ret = 0;
4706 out:
4707         fput_light(output_file, fput_needed);
4708         return ret;
4709 }
4710
4711 /**
4712  * sys_perf_event_open - open a performance event, associate it to a task/cpu
4713  *
4714  * @attr_uptr:  event_id type attributes for monitoring/sampling
4715  * @pid:                target pid
4716  * @cpu:                target cpu
4717  * @group_fd:           group leader event fd
4718  */
4719 SYSCALL_DEFINE5(perf_event_open,
4720                 struct perf_event_attr __user *, attr_uptr,
4721                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4722 {
4723         struct perf_event *event, *group_leader;
4724         struct perf_event_attr attr;
4725         struct perf_event_context *ctx;
4726         struct file *event_file = NULL;
4727         struct file *group_file = NULL;
4728         int fput_needed = 0;
4729         int fput_needed2 = 0;
4730         int err;
4731
4732         /* for future expandability... */
4733         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4734                 return -EINVAL;
4735
4736         err = perf_copy_attr(attr_uptr, &attr);
4737         if (err)
4738                 return err;
4739
4740         if (!attr.exclude_kernel) {
4741                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4742                         return -EACCES;
4743         }
4744
4745         if (attr.freq) {
4746                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4747                         return -EINVAL;
4748         }
4749
4750         /*
4751          * Get the target context (task or percpu):
4752          */
4753         ctx = find_get_context(pid, cpu);
4754         if (IS_ERR(ctx))
4755                 return PTR_ERR(ctx);
4756
4757         /*
4758          * Look up the group leader (we will attach this event to it):
4759          */
4760         group_leader = NULL;
4761         if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4762                 err = -EINVAL;
4763                 group_file = fget_light(group_fd, &fput_needed);
4764                 if (!group_file)
4765                         goto err_put_context;
4766                 if (group_file->f_op != &perf_fops)
4767                         goto err_put_context;
4768
4769                 group_leader = group_file->private_data;
4770                 /*
4771                  * Do not allow a recursive hierarchy (this new sibling
4772                  * becoming part of another group-sibling):
4773                  */
4774                 if (group_leader->group_leader != group_leader)
4775                         goto err_put_context;
4776                 /*
4777                  * Do not allow to attach to a group in a different
4778                  * task or CPU context:
4779                  */
4780                 if (group_leader->ctx != ctx)
4781                         goto err_put_context;
4782                 /*
4783                  * Only a group leader can be exclusive or pinned
4784                  */
4785                 if (attr.exclusive || attr.pinned)
4786                         goto err_put_context;
4787         }
4788
4789         event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4790                                      NULL, NULL, GFP_KERNEL);
4791         err = PTR_ERR(event);
4792         if (IS_ERR(event))
4793                 goto err_put_context;
4794
4795         err = anon_inode_getfd("[perf_event]", &perf_fops, event, O_RDWR);
4796         if (err < 0)
4797                 goto err_free_put_context;
4798
4799         event_file = fget_light(err, &fput_needed2);
4800         if (!event_file)
4801                 goto err_free_put_context;
4802
4803         if (flags & PERF_FLAG_FD_OUTPUT) {
4804                 err = perf_event_set_output(event, group_fd);
4805                 if (err)
4806                         goto err_fput_free_put_context;
4807         }
4808
4809         event->filp = event_file;
4810         WARN_ON_ONCE(ctx->parent_ctx);
4811         mutex_lock(&ctx->mutex);
4812         perf_install_in_context(ctx, event, cpu);
4813         ++ctx->generation;
4814         mutex_unlock(&ctx->mutex);
4815
4816         event->owner = current;
4817         get_task_struct(current);
4818         mutex_lock(&current->perf_event_mutex);
4819         list_add_tail(&event->owner_entry, &current->perf_event_list);
4820         mutex_unlock(&current->perf_event_mutex);
4821
4822 err_fput_free_put_context:
4823         fput_light(event_file, fput_needed2);
4824
4825 err_free_put_context:
4826         if (err < 0)
4827                 kfree(event);
4828
4829 err_put_context:
4830         if (err < 0)
4831                 put_ctx(ctx);
4832
4833         fput_light(group_file, fput_needed);
4834
4835         return err;
4836 }
4837
4838 /**
4839  * perf_event_create_kernel_counter
4840  *
4841  * @attr: attributes of the counter to create
4842  * @cpu: cpu in which the counter is bound
4843  * @pid: task to profile
4844  */
4845 struct perf_event *
4846 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4847                                  pid_t pid,
4848                                  perf_overflow_handler_t overflow_handler)
4849 {
4850         struct perf_event *event;
4851         struct perf_event_context *ctx;
4852         int err;
4853
4854         /*
4855          * Get the target context (task or percpu):
4856          */
4857
4858         ctx = find_get_context(pid, cpu);
4859         if (IS_ERR(ctx)) {
4860                 err = PTR_ERR(ctx);
4861                 goto err_exit;
4862         }
4863
4864         event = perf_event_alloc(attr, cpu, ctx, NULL,
4865                                  NULL, overflow_handler, GFP_KERNEL);
4866         if (IS_ERR(event)) {
4867                 err = PTR_ERR(event);
4868                 goto err_put_context;
4869         }
4870
4871         event->filp = NULL;
4872         WARN_ON_ONCE(ctx->parent_ctx);
4873         mutex_lock(&ctx->mutex);
4874         perf_install_in_context(ctx, event, cpu);
4875         ++ctx->generation;
4876         mutex_unlock(&ctx->mutex);
4877
4878         event->owner = current;
4879         get_task_struct(current);
4880         mutex_lock(&current->perf_event_mutex);
4881         list_add_tail(&event->owner_entry, &current->perf_event_list);
4882         mutex_unlock(&current->perf_event_mutex);
4883
4884         return event;
4885
4886  err_put_context:
4887         put_ctx(ctx);
4888  err_exit:
4889         return ERR_PTR(err);
4890 }
4891 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4892
4893 /*
4894  * inherit a event from parent task to child task:
4895  */
4896 static struct perf_event *
4897 inherit_event(struct perf_event *parent_event,
4898               struct task_struct *parent,
4899               struct perf_event_context *parent_ctx,
4900               struct task_struct *child,
4901               struct perf_event *group_leader,
4902               struct perf_event_context *child_ctx)
4903 {
4904         struct perf_event *child_event;
4905
4906         /*
4907          * Instead of creating recursive hierarchies of events,
4908          * we link inherited events back to the original parent,
4909          * which has a filp for sure, which we use as the reference
4910          * count:
4911          */
4912         if (parent_event->parent)
4913                 parent_event = parent_event->parent;
4914
4915         child_event = perf_event_alloc(&parent_event->attr,
4916                                            parent_event->cpu, child_ctx,
4917                                            group_leader, parent_event,
4918                                            NULL, GFP_KERNEL);
4919         if (IS_ERR(child_event))
4920                 return child_event;
4921         get_ctx(child_ctx);
4922
4923         /*
4924          * Make the child state follow the state of the parent event,
4925          * not its attr.disabled bit.  We hold the parent's mutex,
4926          * so we won't race with perf_event_{en, dis}able_family.
4927          */
4928         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
4929                 child_event->state = PERF_EVENT_STATE_INACTIVE;
4930         else
4931                 child_event->state = PERF_EVENT_STATE_OFF;
4932
4933         if (parent_event->attr.freq)
4934                 child_event->hw.sample_period = parent_event->hw.sample_period;
4935
4936         child_event->overflow_handler = parent_event->overflow_handler;
4937
4938         /*
4939          * Link it up in the child's context:
4940          */
4941         add_event_to_ctx(child_event, child_ctx);
4942
4943         /*
4944          * Get a reference to the parent filp - we will fput it
4945          * when the child event exits. This is safe to do because
4946          * we are in the parent and we know that the filp still
4947          * exists and has a nonzero count:
4948          */
4949         atomic_long_inc(&parent_event->filp->f_count);
4950
4951         /*
4952          * Link this into the parent event's child list
4953          */
4954         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4955         mutex_lock(&parent_event->child_mutex);
4956         list_add_tail(&child_event->child_list, &parent_event->child_list);
4957         mutex_unlock(&parent_event->child_mutex);
4958
4959         return child_event;
4960 }
4961
4962 static int inherit_group(struct perf_event *parent_event,
4963               struct task_struct *parent,
4964               struct perf_event_context *parent_ctx,
4965               struct task_struct *child,
4966               struct perf_event_context *child_ctx)
4967 {
4968         struct perf_event *leader;
4969         struct perf_event *sub;
4970         struct perf_event *child_ctr;
4971
4972         leader = inherit_event(parent_event, parent, parent_ctx,
4973                                  child, NULL, child_ctx);
4974         if (IS_ERR(leader))
4975                 return PTR_ERR(leader);
4976         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
4977                 child_ctr = inherit_event(sub, parent, parent_ctx,
4978                                             child, leader, child_ctx);
4979                 if (IS_ERR(child_ctr))
4980                         return PTR_ERR(child_ctr);
4981         }
4982         return 0;
4983 }
4984
4985 static void sync_child_event(struct perf_event *child_event,
4986                                struct task_struct *child)
4987 {
4988         struct perf_event *parent_event = child_event->parent;
4989         u64 child_val;
4990
4991         if (child_event->attr.inherit_stat)
4992                 perf_event_read_event(child_event, child);
4993
4994         child_val = atomic64_read(&child_event->count);
4995
4996         /*
4997          * Add back the child's count to the parent's count:
4998          */
4999         atomic64_add(child_val, &parent_event->count);
5000         atomic64_add(child_event->total_time_enabled,
5001                      &parent_event->child_total_time_enabled);
5002         atomic64_add(child_event->total_time_running,
5003                      &parent_event->child_total_time_running);
5004
5005         /*
5006          * Remove this event from the parent's list
5007          */
5008         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5009         mutex_lock(&parent_event->child_mutex);
5010         list_del_init(&child_event->child_list);
5011         mutex_unlock(&parent_event->child_mutex);
5012
5013         /*
5014          * Release the parent event, if this was the last
5015          * reference to it.
5016          */
5017         fput(parent_event->filp);
5018 }
5019
5020 static void
5021 __perf_event_exit_task(struct perf_event *child_event,
5022                          struct perf_event_context *child_ctx,
5023                          struct task_struct *child)
5024 {
5025         struct perf_event *parent_event;
5026
5027         perf_event_remove_from_context(child_event);
5028
5029         parent_event = child_event->parent;
5030         /*
5031          * It can happen that parent exits first, and has events
5032          * that are still around due to the child reference. These
5033          * events need to be zapped - but otherwise linger.
5034          */
5035         if (parent_event) {
5036                 sync_child_event(child_event, child);
5037                 free_event(child_event);
5038         }
5039 }
5040
5041 /*
5042  * When a child task exits, feed back event values to parent events.
5043  */
5044 void perf_event_exit_task(struct task_struct *child)
5045 {
5046         struct perf_event *child_event, *tmp;
5047         struct perf_event_context *child_ctx;
5048         unsigned long flags;
5049
5050         if (likely(!child->perf_event_ctxp)) {
5051                 perf_event_task(child, NULL, 0);
5052                 return;
5053         }
5054
5055         local_irq_save(flags);
5056         /*
5057          * We can't reschedule here because interrupts are disabled,
5058          * and either child is current or it is a task that can't be
5059          * scheduled, so we are now safe from rescheduling changing
5060          * our context.
5061          */
5062         child_ctx = child->perf_event_ctxp;
5063         __perf_event_task_sched_out(child_ctx);
5064
5065         /*
5066          * Take the context lock here so that if find_get_context is
5067          * reading child->perf_event_ctxp, we wait until it has
5068          * incremented the context's refcount before we do put_ctx below.
5069          */
5070         raw_spin_lock(&child_ctx->lock);
5071         child->perf_event_ctxp = NULL;
5072         /*
5073          * If this context is a clone; unclone it so it can't get
5074          * swapped to another process while we're removing all
5075          * the events from it.
5076          */
5077         unclone_ctx(child_ctx);
5078         update_context_time(child_ctx);
5079         raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5080
5081         /*
5082          * Report the task dead after unscheduling the events so that we
5083          * won't get any samples after PERF_RECORD_EXIT. We can however still
5084          * get a few PERF_RECORD_READ events.
5085          */
5086         perf_event_task(child, child_ctx, 0);
5087
5088         /*
5089          * We can recurse on the same lock type through:
5090          *
5091          *   __perf_event_exit_task()
5092          *     sync_child_event()
5093          *       fput(parent_event->filp)
5094          *         perf_release()
5095          *           mutex_lock(&ctx->mutex)
5096          *
5097          * But since its the parent context it won't be the same instance.
5098          */
5099         mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
5100
5101 again:
5102         list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5103                                  group_entry)
5104                 __perf_event_exit_task(child_event, child_ctx, child);
5105
5106         list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5107                                  group_entry)
5108                 __perf_event_exit_task(child_event, child_ctx, child);
5109
5110         /*
5111          * If the last event was a group event, it will have appended all
5112          * its siblings to the list, but we obtained 'tmp' before that which
5113          * will still point to the list head terminating the iteration.
5114          */
5115         if (!list_empty(&child_ctx->pinned_groups) ||
5116             !list_empty(&child_ctx->flexible_groups))
5117                 goto again;
5118
5119         mutex_unlock(&child_ctx->mutex);
5120
5121         put_ctx(child_ctx);
5122 }
5123
5124 static void perf_free_event(struct perf_event *event,
5125                             struct perf_event_context *ctx)
5126 {
5127         struct perf_event *parent = event->parent;
5128
5129         if (WARN_ON_ONCE(!parent))
5130                 return;
5131
5132         mutex_lock(&parent->child_mutex);
5133         list_del_init(&event->child_list);
5134         mutex_unlock(&parent->child_mutex);
5135
5136         fput(parent->filp);
5137
5138         list_del_event(event, ctx);
5139         free_event(event);
5140 }
5141
5142 /*
5143  * free an unexposed, unused context as created by inheritance by
5144  * init_task below, used by fork() in case of fail.
5145  */
5146 void perf_event_free_task(struct task_struct *task)
5147 {
5148         struct perf_event_context *ctx = task->perf_event_ctxp;
5149         struct perf_event *event, *tmp;
5150
5151         if (!ctx)
5152                 return;
5153
5154         mutex_lock(&ctx->mutex);
5155 again:
5156         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5157                 perf_free_event(event, ctx);
5158
5159         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5160                                  group_entry)
5161                 perf_free_event(event, ctx);
5162
5163         if (!list_empty(&ctx->pinned_groups) ||
5164             !list_empty(&ctx->flexible_groups))
5165                 goto again;
5166
5167         mutex_unlock(&ctx->mutex);
5168
5169         put_ctx(ctx);
5170 }
5171
5172 static int
5173 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5174                    struct perf_event_context *parent_ctx,
5175                    struct task_struct *child,
5176                    int *inherited_all)
5177 {
5178         int ret;
5179         struct perf_event_context *child_ctx = child->perf_event_ctxp;
5180
5181         if (!event->attr.inherit) {
5182                 *inherited_all = 0;
5183                 return 0;
5184         }
5185
5186         if (!child_ctx) {
5187                 /*
5188                  * This is executed from the parent task context, so
5189                  * inherit events that have been marked for cloning.
5190                  * First allocate and initialize a context for the
5191                  * child.
5192                  */
5193
5194                 child_ctx = kzalloc(sizeof(struct perf_event_context),
5195                                     GFP_KERNEL);
5196                 if (!child_ctx)
5197                         return -ENOMEM;
5198
5199                 __perf_event_init_context(child_ctx, child);
5200                 child->perf_event_ctxp = child_ctx;
5201                 get_task_struct(child);
5202         }
5203
5204         ret = inherit_group(event, parent, parent_ctx,
5205                             child, child_ctx);
5206
5207         if (ret)
5208                 *inherited_all = 0;
5209
5210         return ret;
5211 }
5212
5213
5214 /*
5215  * Initialize the perf_event context in task_struct
5216  */
5217 int perf_event_init_task(struct task_struct *child)
5218 {
5219         struct perf_event_context *child_ctx, *parent_ctx;
5220         struct perf_event_context *cloned_ctx;
5221         struct perf_event *event;
5222         struct task_struct *parent = current;
5223         int inherited_all = 1;
5224         int ret = 0;
5225
5226         child->perf_event_ctxp = NULL;
5227
5228         mutex_init(&child->perf_event_mutex);
5229         INIT_LIST_HEAD(&child->perf_event_list);
5230
5231         if (likely(!parent->perf_event_ctxp))
5232                 return 0;
5233
5234         /*
5235          * If the parent's context is a clone, pin it so it won't get
5236          * swapped under us.
5237          */
5238         parent_ctx = perf_pin_task_context(parent);
5239
5240         /*
5241          * No need to check if parent_ctx != NULL here; since we saw
5242          * it non-NULL earlier, the only reason for it to become NULL
5243          * is if we exit, and since we're currently in the middle of
5244          * a fork we can't be exiting at the same time.
5245          */
5246
5247         /*
5248          * Lock the parent list. No need to lock the child - not PID
5249          * hashed yet and not running, so nobody can access it.
5250          */
5251         mutex_lock(&parent_ctx->mutex);
5252
5253         /*
5254          * We dont have to disable NMIs - we are only looking at
5255          * the list, not manipulating it:
5256          */
5257         list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5258                 ret = inherit_task_group(event, parent, parent_ctx, child,
5259                                          &inherited_all);
5260                 if (ret)
5261                         break;
5262         }
5263
5264         list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5265                 ret = inherit_task_group(event, parent, parent_ctx, child,
5266                                          &inherited_all);
5267                 if (ret)
5268                         break;
5269         }
5270
5271         child_ctx = child->perf_event_ctxp;
5272
5273         if (child_ctx && inherited_all) {
5274                 /*
5275                  * Mark the child context as a clone of the parent
5276                  * context, or of whatever the parent is a clone of.
5277                  * Note that if the parent is a clone, it could get
5278                  * uncloned at any point, but that doesn't matter
5279                  * because the list of events and the generation
5280                  * count can't have changed since we took the mutex.
5281                  */
5282                 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5283                 if (cloned_ctx) {
5284                         child_ctx->parent_ctx = cloned_ctx;
5285                         child_ctx->parent_gen = parent_ctx->parent_gen;
5286                 } else {
5287                         child_ctx->parent_ctx = parent_ctx;
5288                         child_ctx->parent_gen = parent_ctx->generation;
5289                 }
5290                 get_ctx(child_ctx->parent_ctx);
5291         }
5292
5293         mutex_unlock(&parent_ctx->mutex);
5294
5295         perf_unpin_context(parent_ctx);
5296
5297         return ret;
5298 }
5299
5300 static void __cpuinit perf_event_init_cpu(int cpu)
5301 {
5302         struct perf_cpu_context *cpuctx;
5303
5304         cpuctx = &per_cpu(perf_cpu_context, cpu);
5305         __perf_event_init_context(&cpuctx->ctx, NULL);
5306
5307         spin_lock(&perf_resource_lock);
5308         cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5309         spin_unlock(&perf_resource_lock);
5310
5311         hw_perf_event_setup(cpu);
5312 }
5313
5314 #ifdef CONFIG_HOTPLUG_CPU
5315 static void __perf_event_exit_cpu(void *info)
5316 {
5317         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5318         struct perf_event_context *ctx = &cpuctx->ctx;
5319         struct perf_event *event, *tmp;
5320
5321         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5322                 __perf_event_remove_from_context(event);
5323         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5324                 __perf_event_remove_from_context(event);
5325 }
5326 static void perf_event_exit_cpu(int cpu)
5327 {
5328         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5329         struct perf_event_context *ctx = &cpuctx->ctx;
5330
5331         mutex_lock(&ctx->mutex);
5332         smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5333         mutex_unlock(&ctx->mutex);
5334 }
5335 #else
5336 static inline void perf_event_exit_cpu(int cpu) { }
5337 #endif
5338
5339 static int __cpuinit
5340 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5341 {
5342         unsigned int cpu = (long)hcpu;
5343
5344         switch (action) {
5345
5346         case CPU_UP_PREPARE:
5347         case CPU_UP_PREPARE_FROZEN:
5348                 perf_event_init_cpu(cpu);
5349                 break;
5350
5351         case CPU_ONLINE:
5352         case CPU_ONLINE_FROZEN:
5353                 hw_perf_event_setup_online(cpu);
5354                 break;
5355
5356         case CPU_DOWN_PREPARE:
5357         case CPU_DOWN_PREPARE_FROZEN:
5358                 perf_event_exit_cpu(cpu);
5359                 break;
5360
5361         default:
5362                 break;
5363         }
5364
5365         return NOTIFY_OK;
5366 }
5367
5368 /*
5369  * This has to have a higher priority than migration_notifier in sched.c.
5370  */
5371 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5372         .notifier_call          = perf_cpu_notify,
5373         .priority               = 20,
5374 };
5375
5376 void __init perf_event_init(void)
5377 {
5378         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5379                         (void *)(long)smp_processor_id());
5380         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5381                         (void *)(long)smp_processor_id());
5382         register_cpu_notifier(&perf_cpu_nb);
5383 }
5384
5385 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
5386 {
5387         return sprintf(buf, "%d\n", perf_reserved_percpu);
5388 }
5389
5390 static ssize_t
5391 perf_set_reserve_percpu(struct sysdev_class *class,
5392                         const char *buf,
5393                         size_t count)
5394 {
5395         struct perf_cpu_context *cpuctx;
5396         unsigned long val;
5397         int err, cpu, mpt;
5398
5399         err = strict_strtoul(buf, 10, &val);
5400         if (err)
5401                 return err;
5402         if (val > perf_max_events)
5403                 return -EINVAL;
5404
5405         spin_lock(&perf_resource_lock);
5406         perf_reserved_percpu = val;
5407         for_each_online_cpu(cpu) {
5408                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5409                 raw_spin_lock_irq(&cpuctx->ctx.lock);
5410                 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5411                           perf_max_events - perf_reserved_percpu);
5412                 cpuctx->max_pertask = mpt;
5413                 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5414         }
5415         spin_unlock(&perf_resource_lock);
5416
5417         return count;
5418 }
5419
5420 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
5421 {
5422         return sprintf(buf, "%d\n", perf_overcommit);
5423 }
5424
5425 static ssize_t
5426 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
5427 {
5428         unsigned long val;
5429         int err;
5430
5431         err = strict_strtoul(buf, 10, &val);
5432         if (err)
5433                 return err;
5434         if (val > 1)
5435                 return -EINVAL;
5436
5437         spin_lock(&perf_resource_lock);
5438         perf_overcommit = val;
5439         spin_unlock(&perf_resource_lock);
5440
5441         return count;
5442 }
5443
5444 static SYSDEV_CLASS_ATTR(
5445                                 reserve_percpu,
5446                                 0644,
5447                                 perf_show_reserve_percpu,
5448                                 perf_set_reserve_percpu
5449                         );
5450
5451 static SYSDEV_CLASS_ATTR(
5452                                 overcommit,
5453                                 0644,
5454                                 perf_show_overcommit,
5455                                 perf_set_overcommit
5456                         );
5457
5458 static struct attribute *perfclass_attrs[] = {
5459         &attr_reserve_percpu.attr,
5460         &attr_overcommit.attr,
5461         NULL
5462 };
5463
5464 static struct attribute_group perfclass_attr_group = {
5465         .attrs                  = perfclass_attrs,
5466         .name                   = "perf_events",
5467 };
5468
5469 static int __init perf_event_sysfs_init(void)
5470 {
5471         return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5472                                   &perfclass_attr_group);
5473 }
5474 device_initcall(perf_event_sysfs_init);