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