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