Merge commit 'v2.6.34-rc1' into perf/urgent
[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 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
60
61 /*
62  * max perf event sample rate
63  */
64 int sysctl_perf_event_sample_rate __read_mostly = 100000;
65
66 static atomic64_t perf_event_id;
67
68 /*
69  * Lock for (sysadmin-configurable) event reservations:
70  */
71 static DEFINE_SPINLOCK(perf_resource_lock);
72
73 /*
74  * Architecture provided APIs - weak aliases:
75  */
76 extern __weak const struct pmu *hw_perf_event_init(struct perf_event *event)
77 {
78         return NULL;
79 }
80
81 void __weak hw_perf_disable(void)               { barrier(); }
82 void __weak hw_perf_enable(void)                { barrier(); }
83
84 void __weak hw_perf_event_setup(int cpu)        { barrier(); }
85 void __weak hw_perf_event_setup_online(int cpu) { barrier(); }
86 void __weak hw_perf_event_setup_offline(int cpu)        { barrier(); }
87
88 int __weak
89 hw_perf_group_sched_in(struct perf_event *group_leader,
90                struct perf_cpu_context *cpuctx,
91                struct perf_event_context *ctx)
92 {
93         return 0;
94 }
95
96 void __weak perf_event_print_debug(void)        { }
97
98 static DEFINE_PER_CPU(int, perf_disable_count);
99
100 void __perf_disable(void)
101 {
102         __get_cpu_var(perf_disable_count)++;
103 }
104
105 bool __perf_enable(void)
106 {
107         return !--__get_cpu_var(perf_disable_count);
108 }
109
110 void perf_disable(void)
111 {
112         __perf_disable();
113         hw_perf_disable();
114 }
115
116 void perf_enable(void)
117 {
118         if (__perf_enable())
119                 hw_perf_enable();
120 }
121
122 static void get_ctx(struct perf_event_context *ctx)
123 {
124         WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
125 }
126
127 static void free_ctx(struct rcu_head *head)
128 {
129         struct perf_event_context *ctx;
130
131         ctx = container_of(head, struct perf_event_context, rcu_head);
132         kfree(ctx);
133 }
134
135 static void put_ctx(struct perf_event_context *ctx)
136 {
137         if (atomic_dec_and_test(&ctx->refcount)) {
138                 if (ctx->parent_ctx)
139                         put_ctx(ctx->parent_ctx);
140                 if (ctx->task)
141                         put_task_struct(ctx->task);
142                 call_rcu(&ctx->rcu_head, free_ctx);
143         }
144 }
145
146 static void unclone_ctx(struct perf_event_context *ctx)
147 {
148         if (ctx->parent_ctx) {
149                 put_ctx(ctx->parent_ctx);
150                 ctx->parent_ctx = NULL;
151         }
152 }
153
154 /*
155  * If we inherit events we want to return the parent event id
156  * to userspace.
157  */
158 static u64 primary_event_id(struct perf_event *event)
159 {
160         u64 id = event->id;
161
162         if (event->parent)
163                 id = event->parent->id;
164
165         return id;
166 }
167
168 /*
169  * Get the perf_event_context for a task and lock it.
170  * This has to cope with with the fact that until it is locked,
171  * the context could get moved to another task.
172  */
173 static struct perf_event_context *
174 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
175 {
176         struct perf_event_context *ctx;
177
178         rcu_read_lock();
179  retry:
180         ctx = rcu_dereference(task->perf_event_ctxp);
181         if (ctx) {
182                 /*
183                  * If this context is a clone of another, it might
184                  * get swapped for another underneath us by
185                  * perf_event_task_sched_out, though the
186                  * rcu_read_lock() protects us from any context
187                  * getting freed.  Lock the context and check if it
188                  * got swapped before we could get the lock, and retry
189                  * if so.  If we locked the right context, then it
190                  * can't get swapped on us any more.
191                  */
192                 raw_spin_lock_irqsave(&ctx->lock, *flags);
193                 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
194                         raw_spin_unlock_irqrestore(&ctx->lock, *flags);
195                         goto retry;
196                 }
197
198                 if (!atomic_inc_not_zero(&ctx->refcount)) {
199                         raw_spin_unlock_irqrestore(&ctx->lock, *flags);
200                         ctx = NULL;
201                 }
202         }
203         rcu_read_unlock();
204         return ctx;
205 }
206
207 /*
208  * Get the context for a task and increment its pin_count so it
209  * can't get swapped to another task.  This also increments its
210  * reference count so that the context can't get freed.
211  */
212 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
213 {
214         struct perf_event_context *ctx;
215         unsigned long flags;
216
217         ctx = perf_lock_task_context(task, &flags);
218         if (ctx) {
219                 ++ctx->pin_count;
220                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
221         }
222         return ctx;
223 }
224
225 static void perf_unpin_context(struct perf_event_context *ctx)
226 {
227         unsigned long flags;
228
229         raw_spin_lock_irqsave(&ctx->lock, flags);
230         --ctx->pin_count;
231         raw_spin_unlock_irqrestore(&ctx->lock, flags);
232         put_ctx(ctx);
233 }
234
235 static inline u64 perf_clock(void)
236 {
237         return cpu_clock(raw_smp_processor_id());
238 }
239
240 /*
241  * Update the record of the current time in a context.
242  */
243 static void update_context_time(struct perf_event_context *ctx)
244 {
245         u64 now = perf_clock();
246
247         ctx->time += now - ctx->timestamp;
248         ctx->timestamp = now;
249 }
250
251 /*
252  * Update the total_time_enabled and total_time_running fields for a event.
253  */
254 static void update_event_times(struct perf_event *event)
255 {
256         struct perf_event_context *ctx = event->ctx;
257         u64 run_end;
258
259         if (event->state < PERF_EVENT_STATE_INACTIVE ||
260             event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
261                 return;
262
263         if (ctx->is_active)
264                 run_end = ctx->time;
265         else
266                 run_end = event->tstamp_stopped;
267
268         event->total_time_enabled = run_end - event->tstamp_enabled;
269
270         if (event->state == PERF_EVENT_STATE_INACTIVE)
271                 run_end = event->tstamp_stopped;
272         else
273                 run_end = ctx->time;
274
275         event->total_time_running = run_end - event->tstamp_running;
276 }
277
278 static struct list_head *
279 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
280 {
281         if (event->attr.pinned)
282                 return &ctx->pinned_groups;
283         else
284                 return &ctx->flexible_groups;
285 }
286
287 /*
288  * Add a event from the lists for its context.
289  * Must be called with ctx->mutex and ctx->lock held.
290  */
291 static void
292 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
293 {
294         struct perf_event *group_leader = event->group_leader;
295
296         /*
297          * Depending on whether it is a standalone or sibling event,
298          * add it straight to the context's event list, or to the group
299          * leader's sibling list:
300          */
301         if (group_leader == event) {
302                 struct list_head *list;
303
304                 if (is_software_event(event))
305                         event->group_flags |= PERF_GROUP_SOFTWARE;
306
307                 list = ctx_group_list(event, ctx);
308                 list_add_tail(&event->group_entry, list);
309         } else {
310                 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
311                     !is_software_event(event))
312                         group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
313
314                 list_add_tail(&event->group_entry, &group_leader->sibling_list);
315                 group_leader->nr_siblings++;
316         }
317
318         list_add_rcu(&event->event_entry, &ctx->event_list);
319         ctx->nr_events++;
320         if (event->attr.inherit_stat)
321                 ctx->nr_stat++;
322 }
323
324 /*
325  * Remove a event from the lists for its context.
326  * Must be called with ctx->mutex and ctx->lock held.
327  */
328 static void
329 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
330 {
331         struct perf_event *sibling, *tmp;
332
333         if (list_empty(&event->group_entry))
334                 return;
335         ctx->nr_events--;
336         if (event->attr.inherit_stat)
337                 ctx->nr_stat--;
338
339         list_del_init(&event->group_entry);
340         list_del_rcu(&event->event_entry);
341
342         if (event->group_leader != event)
343                 event->group_leader->nr_siblings--;
344
345         update_event_times(event);
346
347         /*
348          * If event was in error state, then keep it
349          * that way, otherwise bogus counts will be
350          * returned on read(). The only way to get out
351          * of error state is by explicit re-enabling
352          * of the event
353          */
354         if (event->state > PERF_EVENT_STATE_OFF)
355                 event->state = PERF_EVENT_STATE_OFF;
356
357         /*
358          * If this was a group event with sibling events then
359          * upgrade the siblings to singleton events by adding them
360          * to the context list directly:
361          */
362         list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
363                 struct list_head *list;
364
365                 list = ctx_group_list(event, ctx);
366                 list_move_tail(&sibling->group_entry, list);
367                 sibling->group_leader = sibling;
368
369                 /* Inherit group flags from the previous leader */
370                 sibling->group_flags = event->group_flags;
371         }
372 }
373
374 static void
375 event_sched_out(struct perf_event *event,
376                   struct perf_cpu_context *cpuctx,
377                   struct perf_event_context *ctx)
378 {
379         if (event->state != PERF_EVENT_STATE_ACTIVE)
380                 return;
381
382         event->state = PERF_EVENT_STATE_INACTIVE;
383         if (event->pending_disable) {
384                 event->pending_disable = 0;
385                 event->state = PERF_EVENT_STATE_OFF;
386         }
387         event->tstamp_stopped = ctx->time;
388         event->pmu->disable(event);
389         event->oncpu = -1;
390
391         if (!is_software_event(event))
392                 cpuctx->active_oncpu--;
393         ctx->nr_active--;
394         if (event->attr.exclusive || !cpuctx->active_oncpu)
395                 cpuctx->exclusive = 0;
396 }
397
398 static void
399 group_sched_out(struct perf_event *group_event,
400                 struct perf_cpu_context *cpuctx,
401                 struct perf_event_context *ctx)
402 {
403         struct perf_event *event;
404
405         if (group_event->state != PERF_EVENT_STATE_ACTIVE)
406                 return;
407
408         event_sched_out(group_event, cpuctx, ctx);
409
410         /*
411          * Schedule out siblings (if any):
412          */
413         list_for_each_entry(event, &group_event->sibling_list, group_entry)
414                 event_sched_out(event, cpuctx, ctx);
415
416         if (group_event->attr.exclusive)
417                 cpuctx->exclusive = 0;
418 }
419
420 /*
421  * Cross CPU call to remove a performance event
422  *
423  * We disable the event on the hardware level first. After that we
424  * remove it from the context list.
425  */
426 static void __perf_event_remove_from_context(void *info)
427 {
428         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
429         struct perf_event *event = info;
430         struct perf_event_context *ctx = event->ctx;
431
432         /*
433          * If this is a task context, we need to check whether it is
434          * the current task context of this cpu. If not it has been
435          * scheduled out before the smp call arrived.
436          */
437         if (ctx->task && cpuctx->task_ctx != ctx)
438                 return;
439
440         raw_spin_lock(&ctx->lock);
441         /*
442          * Protect the list operation against NMI by disabling the
443          * events on a global level.
444          */
445         perf_disable();
446
447         event_sched_out(event, cpuctx, ctx);
448
449         list_del_event(event, ctx);
450
451         if (!ctx->task) {
452                 /*
453                  * Allow more per task events with respect to the
454                  * reservation:
455                  */
456                 cpuctx->max_pertask =
457                         min(perf_max_events - ctx->nr_events,
458                             perf_max_events - perf_reserved_percpu);
459         }
460
461         perf_enable();
462         raw_spin_unlock(&ctx->lock);
463 }
464
465
466 /*
467  * Remove the event from a task's (or a CPU's) list of events.
468  *
469  * Must be called with ctx->mutex held.
470  *
471  * CPU events are removed with a smp call. For task events we only
472  * call when the task is on a CPU.
473  *
474  * If event->ctx is a cloned context, callers must make sure that
475  * every task struct that event->ctx->task could possibly point to
476  * remains valid.  This is OK when called from perf_release since
477  * that only calls us on the top-level context, which can't be a clone.
478  * When called from perf_event_exit_task, it's OK because the
479  * context has been detached from its task.
480  */
481 static void perf_event_remove_from_context(struct perf_event *event)
482 {
483         struct perf_event_context *ctx = event->ctx;
484         struct task_struct *task = ctx->task;
485
486         if (!task) {
487                 /*
488                  * Per cpu events are removed via an smp call and
489                  * the removal is always successful.
490                  */
491                 smp_call_function_single(event->cpu,
492                                          __perf_event_remove_from_context,
493                                          event, 1);
494                 return;
495         }
496
497 retry:
498         task_oncpu_function_call(task, __perf_event_remove_from_context,
499                                  event);
500
501         raw_spin_lock_irq(&ctx->lock);
502         /*
503          * If the context is active we need to retry the smp call.
504          */
505         if (ctx->nr_active && !list_empty(&event->group_entry)) {
506                 raw_spin_unlock_irq(&ctx->lock);
507                 goto retry;
508         }
509
510         /*
511          * The lock prevents that this context is scheduled in so we
512          * can remove the event safely, if the call above did not
513          * succeed.
514          */
515         if (!list_empty(&event->group_entry))
516                 list_del_event(event, ctx);
517         raw_spin_unlock_irq(&ctx->lock);
518 }
519
520 /*
521  * Update total_time_enabled and total_time_running for all events in a group.
522  */
523 static void update_group_times(struct perf_event *leader)
524 {
525         struct perf_event *event;
526
527         update_event_times(leader);
528         list_for_each_entry(event, &leader->sibling_list, group_entry)
529                 update_event_times(event);
530 }
531
532 /*
533  * Cross CPU call to disable a performance event
534  */
535 static void __perf_event_disable(void *info)
536 {
537         struct perf_event *event = info;
538         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
539         struct perf_event_context *ctx = event->ctx;
540
541         /*
542          * If this is a per-task event, need to check whether this
543          * event's task is the current task on this cpu.
544          */
545         if (ctx->task && cpuctx->task_ctx != ctx)
546                 return;
547
548         raw_spin_lock(&ctx->lock);
549
550         /*
551          * If the event is on, turn it off.
552          * If it is in error state, leave it in error state.
553          */
554         if (event->state >= PERF_EVENT_STATE_INACTIVE) {
555                 update_context_time(ctx);
556                 update_group_times(event);
557                 if (event == event->group_leader)
558                         group_sched_out(event, cpuctx, ctx);
559                 else
560                         event_sched_out(event, cpuctx, ctx);
561                 event->state = PERF_EVENT_STATE_OFF;
562         }
563
564         raw_spin_unlock(&ctx->lock);
565 }
566
567 /*
568  * Disable a event.
569  *
570  * If event->ctx is a cloned context, callers must make sure that
571  * every task struct that event->ctx->task could possibly point to
572  * remains valid.  This condition is satisifed when called through
573  * perf_event_for_each_child or perf_event_for_each because they
574  * hold the top-level event's child_mutex, so any descendant that
575  * goes to exit will block in sync_child_event.
576  * When called from perf_pending_event it's OK because event->ctx
577  * is the current context on this CPU and preemption is disabled,
578  * hence we can't get into perf_event_task_sched_out for this context.
579  */
580 void perf_event_disable(struct perf_event *event)
581 {
582         struct perf_event_context *ctx = event->ctx;
583         struct task_struct *task = ctx->task;
584
585         if (!task) {
586                 /*
587                  * Disable the event on the cpu that it's on
588                  */
589                 smp_call_function_single(event->cpu, __perf_event_disable,
590                                          event, 1);
591                 return;
592         }
593
594  retry:
595         task_oncpu_function_call(task, __perf_event_disable, event);
596
597         raw_spin_lock_irq(&ctx->lock);
598         /*
599          * If the event is still active, we need to retry the cross-call.
600          */
601         if (event->state == PERF_EVENT_STATE_ACTIVE) {
602                 raw_spin_unlock_irq(&ctx->lock);
603                 goto retry;
604         }
605
606         /*
607          * Since we have the lock this context can't be scheduled
608          * in, so we can change the state safely.
609          */
610         if (event->state == PERF_EVENT_STATE_INACTIVE) {
611                 update_group_times(event);
612                 event->state = PERF_EVENT_STATE_OFF;
613         }
614
615         raw_spin_unlock_irq(&ctx->lock);
616 }
617
618 static int
619 event_sched_in(struct perf_event *event,
620                  struct perf_cpu_context *cpuctx,
621                  struct perf_event_context *ctx)
622 {
623         if (event->state <= PERF_EVENT_STATE_OFF)
624                 return 0;
625
626         event->state = PERF_EVENT_STATE_ACTIVE;
627         event->oncpu = smp_processor_id();
628         /*
629          * The new state must be visible before we turn it on in the hardware:
630          */
631         smp_wmb();
632
633         if (event->pmu->enable(event)) {
634                 event->state = PERF_EVENT_STATE_INACTIVE;
635                 event->oncpu = -1;
636                 return -EAGAIN;
637         }
638
639         event->tstamp_running += ctx->time - event->tstamp_stopped;
640
641         if (!is_software_event(event))
642                 cpuctx->active_oncpu++;
643         ctx->nr_active++;
644
645         if (event->attr.exclusive)
646                 cpuctx->exclusive = 1;
647
648         return 0;
649 }
650
651 static int
652 group_sched_in(struct perf_event *group_event,
653                struct perf_cpu_context *cpuctx,
654                struct perf_event_context *ctx)
655 {
656         struct perf_event *event, *partial_group;
657         int ret;
658
659         if (group_event->state == PERF_EVENT_STATE_OFF)
660                 return 0;
661
662         ret = hw_perf_group_sched_in(group_event, cpuctx, ctx);
663         if (ret)
664                 return ret < 0 ? ret : 0;
665
666         if (event_sched_in(group_event, cpuctx, ctx))
667                 return -EAGAIN;
668
669         /*
670          * Schedule in siblings as one group (if any):
671          */
672         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
673                 if (event_sched_in(event, cpuctx, ctx)) {
674                         partial_group = event;
675                         goto group_error;
676                 }
677         }
678
679         return 0;
680
681 group_error:
682         /*
683          * Groups can be scheduled in as one unit only, so undo any
684          * partial group before returning:
685          */
686         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
687                 if (event == partial_group)
688                         break;
689                 event_sched_out(event, cpuctx, ctx);
690         }
691         event_sched_out(group_event, cpuctx, ctx);
692
693         return -EAGAIN;
694 }
695
696 /*
697  * Work out whether we can put this event group on the CPU now.
698  */
699 static int group_can_go_on(struct perf_event *event,
700                            struct perf_cpu_context *cpuctx,
701                            int can_add_hw)
702 {
703         /*
704          * Groups consisting entirely of software events can always go on.
705          */
706         if (event->group_flags & PERF_GROUP_SOFTWARE)
707                 return 1;
708         /*
709          * If an exclusive group is already on, no other hardware
710          * events can go on.
711          */
712         if (cpuctx->exclusive)
713                 return 0;
714         /*
715          * If this group is exclusive and there are already
716          * events on the CPU, it can't go on.
717          */
718         if (event->attr.exclusive && cpuctx->active_oncpu)
719                 return 0;
720         /*
721          * Otherwise, try to add it if all previous groups were able
722          * to go on.
723          */
724         return can_add_hw;
725 }
726
727 static void add_event_to_ctx(struct perf_event *event,
728                                struct perf_event_context *ctx)
729 {
730         list_add_event(event, ctx);
731         event->tstamp_enabled = ctx->time;
732         event->tstamp_running = ctx->time;
733         event->tstamp_stopped = ctx->time;
734 }
735
736 /*
737  * Cross CPU call to install and enable a performance event
738  *
739  * Must be called with ctx->mutex held
740  */
741 static void __perf_install_in_context(void *info)
742 {
743         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
744         struct perf_event *event = info;
745         struct perf_event_context *ctx = event->ctx;
746         struct perf_event *leader = event->group_leader;
747         int err;
748
749         /*
750          * If this is a task context, we need to check whether it is
751          * the current task context of this cpu. If not it has been
752          * scheduled out before the smp call arrived.
753          * Or possibly this is the right context but it isn't
754          * on this cpu because it had no events.
755          */
756         if (ctx->task && cpuctx->task_ctx != ctx) {
757                 if (cpuctx->task_ctx || ctx->task != current)
758                         return;
759                 cpuctx->task_ctx = ctx;
760         }
761
762         raw_spin_lock(&ctx->lock);
763         ctx->is_active = 1;
764         update_context_time(ctx);
765
766         /*
767          * Protect the list operation against NMI by disabling the
768          * events on a global level. NOP for non NMI based events.
769          */
770         perf_disable();
771
772         add_event_to_ctx(event, ctx);
773
774         if (event->cpu != -1 && event->cpu != smp_processor_id())
775                 goto unlock;
776
777         /*
778          * Don't put the event on if it is disabled or if
779          * it is in a group and the group isn't on.
780          */
781         if (event->state != PERF_EVENT_STATE_INACTIVE ||
782             (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
783                 goto unlock;
784
785         /*
786          * An exclusive event can't go on if there are already active
787          * hardware events, and no hardware event can go on if there
788          * is already an exclusive event on.
789          */
790         if (!group_can_go_on(event, cpuctx, 1))
791                 err = -EEXIST;
792         else
793                 err = event_sched_in(event, cpuctx, ctx);
794
795         if (err) {
796                 /*
797                  * This event couldn't go on.  If it is in a group
798                  * then we have to pull the whole group off.
799                  * If the event group is pinned then put it in error state.
800                  */
801                 if (leader != event)
802                         group_sched_out(leader, cpuctx, ctx);
803                 if (leader->attr.pinned) {
804                         update_group_times(leader);
805                         leader->state = PERF_EVENT_STATE_ERROR;
806                 }
807         }
808
809         if (!err && !ctx->task && cpuctx->max_pertask)
810                 cpuctx->max_pertask--;
811
812  unlock:
813         perf_enable();
814
815         raw_spin_unlock(&ctx->lock);
816 }
817
818 /*
819  * Attach a performance event to a context
820  *
821  * First we add the event to the list with the hardware enable bit
822  * in event->hw_config cleared.
823  *
824  * If the event is attached to a task which is on a CPU we use a smp
825  * call to enable it in the task context. The task might have been
826  * scheduled away, but we check this in the smp call again.
827  *
828  * Must be called with ctx->mutex held.
829  */
830 static void
831 perf_install_in_context(struct perf_event_context *ctx,
832                         struct perf_event *event,
833                         int cpu)
834 {
835         struct task_struct *task = ctx->task;
836
837         if (!task) {
838                 /*
839                  * Per cpu events are installed via an smp call and
840                  * the install is always successful.
841                  */
842                 smp_call_function_single(cpu, __perf_install_in_context,
843                                          event, 1);
844                 return;
845         }
846
847 retry:
848         task_oncpu_function_call(task, __perf_install_in_context,
849                                  event);
850
851         raw_spin_lock_irq(&ctx->lock);
852         /*
853          * we need to retry the smp call.
854          */
855         if (ctx->is_active && list_empty(&event->group_entry)) {
856                 raw_spin_unlock_irq(&ctx->lock);
857                 goto retry;
858         }
859
860         /*
861          * The lock prevents that this context is scheduled in so we
862          * can add the event safely, if it the call above did not
863          * succeed.
864          */
865         if (list_empty(&event->group_entry))
866                 add_event_to_ctx(event, ctx);
867         raw_spin_unlock_irq(&ctx->lock);
868 }
869
870 /*
871  * Put a event into inactive state and update time fields.
872  * Enabling the leader of a group effectively enables all
873  * the group members that aren't explicitly disabled, so we
874  * have to update their ->tstamp_enabled also.
875  * Note: this works for group members as well as group leaders
876  * since the non-leader members' sibling_lists will be empty.
877  */
878 static void __perf_event_mark_enabled(struct perf_event *event,
879                                         struct perf_event_context *ctx)
880 {
881         struct perf_event *sub;
882
883         event->state = PERF_EVENT_STATE_INACTIVE;
884         event->tstamp_enabled = ctx->time - event->total_time_enabled;
885         list_for_each_entry(sub, &event->sibling_list, group_entry)
886                 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
887                         sub->tstamp_enabled =
888                                 ctx->time - sub->total_time_enabled;
889 }
890
891 /*
892  * Cross CPU call to enable a performance event
893  */
894 static void __perf_event_enable(void *info)
895 {
896         struct perf_event *event = info;
897         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
898         struct perf_event_context *ctx = event->ctx;
899         struct perf_event *leader = event->group_leader;
900         int err;
901
902         /*
903          * If this is a per-task event, need to check whether this
904          * event's task is the current task on this cpu.
905          */
906         if (ctx->task && cpuctx->task_ctx != ctx) {
907                 if (cpuctx->task_ctx || ctx->task != current)
908                         return;
909                 cpuctx->task_ctx = ctx;
910         }
911
912         raw_spin_lock(&ctx->lock);
913         ctx->is_active = 1;
914         update_context_time(ctx);
915
916         if (event->state >= PERF_EVENT_STATE_INACTIVE)
917                 goto unlock;
918         __perf_event_mark_enabled(event, ctx);
919
920         if (event->cpu != -1 && event->cpu != smp_processor_id())
921                 goto unlock;
922
923         /*
924          * If the event is in a group and isn't the group leader,
925          * then don't put it on unless the group is on.
926          */
927         if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
928                 goto unlock;
929
930         if (!group_can_go_on(event, cpuctx, 1)) {
931                 err = -EEXIST;
932         } else {
933                 perf_disable();
934                 if (event == leader)
935                         err = group_sched_in(event, cpuctx, ctx);
936                 else
937                         err = event_sched_in(event, cpuctx, ctx);
938                 perf_enable();
939         }
940
941         if (err) {
942                 /*
943                  * If this event can't go on and it's part of a
944                  * group, then the whole group has to come off.
945                  */
946                 if (leader != event)
947                         group_sched_out(leader, cpuctx, ctx);
948                 if (leader->attr.pinned) {
949                         update_group_times(leader);
950                         leader->state = PERF_EVENT_STATE_ERROR;
951                 }
952         }
953
954  unlock:
955         raw_spin_unlock(&ctx->lock);
956 }
957
958 /*
959  * Enable a event.
960  *
961  * If event->ctx is a cloned context, callers must make sure that
962  * every task struct that event->ctx->task could possibly point to
963  * remains valid.  This condition is satisfied when called through
964  * perf_event_for_each_child or perf_event_for_each as described
965  * for perf_event_disable.
966  */
967 void perf_event_enable(struct perf_event *event)
968 {
969         struct perf_event_context *ctx = event->ctx;
970         struct task_struct *task = ctx->task;
971
972         if (!task) {
973                 /*
974                  * Enable the event on the cpu that it's on
975                  */
976                 smp_call_function_single(event->cpu, __perf_event_enable,
977                                          event, 1);
978                 return;
979         }
980
981         raw_spin_lock_irq(&ctx->lock);
982         if (event->state >= PERF_EVENT_STATE_INACTIVE)
983                 goto out;
984
985         /*
986          * If the event is in error state, clear that first.
987          * That way, if we see the event in error state below, we
988          * know that it has gone back into error state, as distinct
989          * from the task having been scheduled away before the
990          * cross-call arrived.
991          */
992         if (event->state == PERF_EVENT_STATE_ERROR)
993                 event->state = PERF_EVENT_STATE_OFF;
994
995  retry:
996         raw_spin_unlock_irq(&ctx->lock);
997         task_oncpu_function_call(task, __perf_event_enable, event);
998
999         raw_spin_lock_irq(&ctx->lock);
1000
1001         /*
1002          * If the context is active and the event is still off,
1003          * we need to retry the cross-call.
1004          */
1005         if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1006                 goto retry;
1007
1008         /*
1009          * Since we have the lock this context can't be scheduled
1010          * in, so we can change the state safely.
1011          */
1012         if (event->state == PERF_EVENT_STATE_OFF)
1013                 __perf_event_mark_enabled(event, ctx);
1014
1015  out:
1016         raw_spin_unlock_irq(&ctx->lock);
1017 }
1018
1019 static int perf_event_refresh(struct perf_event *event, int refresh)
1020 {
1021         /*
1022          * not supported on inherited events
1023          */
1024         if (event->attr.inherit)
1025                 return -EINVAL;
1026
1027         atomic_add(refresh, &event->event_limit);
1028         perf_event_enable(event);
1029
1030         return 0;
1031 }
1032
1033 enum event_type_t {
1034         EVENT_FLEXIBLE = 0x1,
1035         EVENT_PINNED = 0x2,
1036         EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1037 };
1038
1039 static void ctx_sched_out(struct perf_event_context *ctx,
1040                           struct perf_cpu_context *cpuctx,
1041                           enum event_type_t event_type)
1042 {
1043         struct perf_event *event;
1044
1045         raw_spin_lock(&ctx->lock);
1046         ctx->is_active = 0;
1047         if (likely(!ctx->nr_events))
1048                 goto out;
1049         update_context_time(ctx);
1050
1051         perf_disable();
1052         if (!ctx->nr_active)
1053                 goto out_enable;
1054
1055         if (event_type & EVENT_PINNED)
1056                 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1057                         group_sched_out(event, cpuctx, ctx);
1058
1059         if (event_type & EVENT_FLEXIBLE)
1060                 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1061                         group_sched_out(event, cpuctx, ctx);
1062
1063  out_enable:
1064         perf_enable();
1065  out:
1066         raw_spin_unlock(&ctx->lock);
1067 }
1068
1069 /*
1070  * Test whether two contexts are equivalent, i.e. whether they
1071  * have both been cloned from the same version of the same context
1072  * and they both have the same number of enabled events.
1073  * If the number of enabled events is the same, then the set
1074  * of enabled events should be the same, because these are both
1075  * inherited contexts, therefore we can't access individual events
1076  * in them directly with an fd; we can only enable/disable all
1077  * events via prctl, or enable/disable all events in a family
1078  * via ioctl, which will have the same effect on both contexts.
1079  */
1080 static int context_equiv(struct perf_event_context *ctx1,
1081                          struct perf_event_context *ctx2)
1082 {
1083         return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1084                 && ctx1->parent_gen == ctx2->parent_gen
1085                 && !ctx1->pin_count && !ctx2->pin_count;
1086 }
1087
1088 static void __perf_event_sync_stat(struct perf_event *event,
1089                                      struct perf_event *next_event)
1090 {
1091         u64 value;
1092
1093         if (!event->attr.inherit_stat)
1094                 return;
1095
1096         /*
1097          * Update the event value, we cannot use perf_event_read()
1098          * because we're in the middle of a context switch and have IRQs
1099          * disabled, which upsets smp_call_function_single(), however
1100          * we know the event must be on the current CPU, therefore we
1101          * don't need to use it.
1102          */
1103         switch (event->state) {
1104         case PERF_EVENT_STATE_ACTIVE:
1105                 event->pmu->read(event);
1106                 /* fall-through */
1107
1108         case PERF_EVENT_STATE_INACTIVE:
1109                 update_event_times(event);
1110                 break;
1111
1112         default:
1113                 break;
1114         }
1115
1116         /*
1117          * In order to keep per-task stats reliable we need to flip the event
1118          * values when we flip the contexts.
1119          */
1120         value = atomic64_read(&next_event->count);
1121         value = atomic64_xchg(&event->count, value);
1122         atomic64_set(&next_event->count, value);
1123
1124         swap(event->total_time_enabled, next_event->total_time_enabled);
1125         swap(event->total_time_running, next_event->total_time_running);
1126
1127         /*
1128          * Since we swizzled the values, update the user visible data too.
1129          */
1130         perf_event_update_userpage(event);
1131         perf_event_update_userpage(next_event);
1132 }
1133
1134 #define list_next_entry(pos, member) \
1135         list_entry(pos->member.next, typeof(*pos), member)
1136
1137 static void perf_event_sync_stat(struct perf_event_context *ctx,
1138                                    struct perf_event_context *next_ctx)
1139 {
1140         struct perf_event *event, *next_event;
1141
1142         if (!ctx->nr_stat)
1143                 return;
1144
1145         update_context_time(ctx);
1146
1147         event = list_first_entry(&ctx->event_list,
1148                                    struct perf_event, event_entry);
1149
1150         next_event = list_first_entry(&next_ctx->event_list,
1151                                         struct perf_event, event_entry);
1152
1153         while (&event->event_entry != &ctx->event_list &&
1154                &next_event->event_entry != &next_ctx->event_list) {
1155
1156                 __perf_event_sync_stat(event, next_event);
1157
1158                 event = list_next_entry(event, event_entry);
1159                 next_event = list_next_entry(next_event, event_entry);
1160         }
1161 }
1162
1163 /*
1164  * Called from scheduler to remove the events of the current task,
1165  * with interrupts disabled.
1166  *
1167  * We stop each event and update the event value in event->count.
1168  *
1169  * This does not protect us against NMI, but disable()
1170  * sets the disabled bit in the control field of event _before_
1171  * accessing the event control register. If a NMI hits, then it will
1172  * not restart the event.
1173  */
1174 void perf_event_task_sched_out(struct task_struct *task,
1175                                  struct task_struct *next)
1176 {
1177         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1178         struct perf_event_context *ctx = task->perf_event_ctxp;
1179         struct perf_event_context *next_ctx;
1180         struct perf_event_context *parent;
1181         struct pt_regs *regs;
1182         int do_switch = 1;
1183
1184         regs = task_pt_regs(task);
1185         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1186
1187         if (likely(!ctx || !cpuctx->task_ctx))
1188                 return;
1189
1190         rcu_read_lock();
1191         parent = rcu_dereference(ctx->parent_ctx);
1192         next_ctx = next->perf_event_ctxp;
1193         if (parent && next_ctx &&
1194             rcu_dereference(next_ctx->parent_ctx) == parent) {
1195                 /*
1196                  * Looks like the two contexts are clones, so we might be
1197                  * able to optimize the context switch.  We lock both
1198                  * contexts and check that they are clones under the
1199                  * lock (including re-checking that neither has been
1200                  * uncloned in the meantime).  It doesn't matter which
1201                  * order we take the locks because no other cpu could
1202                  * be trying to lock both of these tasks.
1203                  */
1204                 raw_spin_lock(&ctx->lock);
1205                 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1206                 if (context_equiv(ctx, next_ctx)) {
1207                         /*
1208                          * XXX do we need a memory barrier of sorts
1209                          * wrt to rcu_dereference() of perf_event_ctxp
1210                          */
1211                         task->perf_event_ctxp = next_ctx;
1212                         next->perf_event_ctxp = ctx;
1213                         ctx->task = next;
1214                         next_ctx->task = task;
1215                         do_switch = 0;
1216
1217                         perf_event_sync_stat(ctx, next_ctx);
1218                 }
1219                 raw_spin_unlock(&next_ctx->lock);
1220                 raw_spin_unlock(&ctx->lock);
1221         }
1222         rcu_read_unlock();
1223
1224         if (do_switch) {
1225                 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1226                 cpuctx->task_ctx = NULL;
1227         }
1228 }
1229
1230 static void task_ctx_sched_out(struct perf_event_context *ctx,
1231                                enum event_type_t event_type)
1232 {
1233         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1234
1235         if (!cpuctx->task_ctx)
1236                 return;
1237
1238         if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1239                 return;
1240
1241         ctx_sched_out(ctx, cpuctx, event_type);
1242         cpuctx->task_ctx = NULL;
1243 }
1244
1245 /*
1246  * Called with IRQs disabled
1247  */
1248 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1249 {
1250         task_ctx_sched_out(ctx, EVENT_ALL);
1251 }
1252
1253 /*
1254  * Called with IRQs disabled
1255  */
1256 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1257                               enum event_type_t event_type)
1258 {
1259         ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1260 }
1261
1262 static void
1263 ctx_pinned_sched_in(struct perf_event_context *ctx,
1264                     struct perf_cpu_context *cpuctx)
1265 {
1266         struct perf_event *event;
1267
1268         list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1269                 if (event->state <= PERF_EVENT_STATE_OFF)
1270                         continue;
1271                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1272                         continue;
1273
1274                 if (group_can_go_on(event, cpuctx, 1))
1275                         group_sched_in(event, cpuctx, ctx);
1276
1277                 /*
1278                  * If this pinned group hasn't been scheduled,
1279                  * put it in error state.
1280                  */
1281                 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1282                         update_group_times(event);
1283                         event->state = PERF_EVENT_STATE_ERROR;
1284                 }
1285         }
1286 }
1287
1288 static void
1289 ctx_flexible_sched_in(struct perf_event_context *ctx,
1290                       struct perf_cpu_context *cpuctx)
1291 {
1292         struct perf_event *event;
1293         int can_add_hw = 1;
1294
1295         list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1296                 /* Ignore events in OFF or ERROR state */
1297                 if (event->state <= PERF_EVENT_STATE_OFF)
1298                         continue;
1299                 /*
1300                  * Listen to the 'cpu' scheduling filter constraint
1301                  * of events:
1302                  */
1303                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1304                         continue;
1305
1306                 if (group_can_go_on(event, cpuctx, can_add_hw))
1307                         if (group_sched_in(event, cpuctx, ctx))
1308                                 can_add_hw = 0;
1309         }
1310 }
1311
1312 static void
1313 ctx_sched_in(struct perf_event_context *ctx,
1314              struct perf_cpu_context *cpuctx,
1315              enum event_type_t event_type)
1316 {
1317         raw_spin_lock(&ctx->lock);
1318         ctx->is_active = 1;
1319         if (likely(!ctx->nr_events))
1320                 goto out;
1321
1322         ctx->timestamp = perf_clock();
1323
1324         perf_disable();
1325
1326         /*
1327          * First go through the list and put on any pinned groups
1328          * in order to give them the best chance of going on.
1329          */
1330         if (event_type & EVENT_PINNED)
1331                 ctx_pinned_sched_in(ctx, cpuctx);
1332
1333         /* Then walk through the lower prio flexible groups */
1334         if (event_type & EVENT_FLEXIBLE)
1335                 ctx_flexible_sched_in(ctx, cpuctx);
1336
1337         perf_enable();
1338  out:
1339         raw_spin_unlock(&ctx->lock);
1340 }
1341
1342 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1343                              enum event_type_t event_type)
1344 {
1345         struct perf_event_context *ctx = &cpuctx->ctx;
1346
1347         ctx_sched_in(ctx, cpuctx, event_type);
1348 }
1349
1350 static void task_ctx_sched_in(struct task_struct *task,
1351                               enum event_type_t event_type)
1352 {
1353         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1354         struct perf_event_context *ctx = task->perf_event_ctxp;
1355
1356         if (likely(!ctx))
1357                 return;
1358         if (cpuctx->task_ctx == ctx)
1359                 return;
1360         ctx_sched_in(ctx, cpuctx, event_type);
1361         cpuctx->task_ctx = ctx;
1362 }
1363 /*
1364  * Called from scheduler to add the events of the current task
1365  * with interrupts disabled.
1366  *
1367  * We restore the event value and then enable it.
1368  *
1369  * This does not protect us against NMI, but enable()
1370  * sets the enabled bit in the control field of event _before_
1371  * accessing the event control register. If a NMI hits, then it will
1372  * keep the event running.
1373  */
1374 void perf_event_task_sched_in(struct task_struct *task)
1375 {
1376         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1377         struct perf_event_context *ctx = task->perf_event_ctxp;
1378
1379         if (likely(!ctx))
1380                 return;
1381
1382         if (cpuctx->task_ctx == ctx)
1383                 return;
1384
1385         /*
1386          * We want to keep the following priority order:
1387          * cpu pinned (that don't need to move), task pinned,
1388          * cpu flexible, task flexible.
1389          */
1390         cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1391
1392         ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1393         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1394         ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1395
1396         cpuctx->task_ctx = ctx;
1397 }
1398
1399 #define MAX_INTERRUPTS (~0ULL)
1400
1401 static void perf_log_throttle(struct perf_event *event, int enable);
1402
1403 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1404 {
1405         u64 frequency = event->attr.sample_freq;
1406         u64 sec = NSEC_PER_SEC;
1407         u64 divisor, dividend;
1408
1409         int count_fls, nsec_fls, frequency_fls, sec_fls;
1410
1411         count_fls = fls64(count);
1412         nsec_fls = fls64(nsec);
1413         frequency_fls = fls64(frequency);
1414         sec_fls = 30;
1415
1416         /*
1417          * We got @count in @nsec, with a target of sample_freq HZ
1418          * the target period becomes:
1419          *
1420          *             @count * 10^9
1421          * period = -------------------
1422          *          @nsec * sample_freq
1423          *
1424          */
1425
1426         /*
1427          * Reduce accuracy by one bit such that @a and @b converge
1428          * to a similar magnitude.
1429          */
1430 #define REDUCE_FLS(a, b)                \
1431 do {                                    \
1432         if (a##_fls > b##_fls) {        \
1433                 a >>= 1;                \
1434                 a##_fls--;              \
1435         } else {                        \
1436                 b >>= 1;                \
1437                 b##_fls--;              \
1438         }                               \
1439 } while (0)
1440
1441         /*
1442          * Reduce accuracy until either term fits in a u64, then proceed with
1443          * the other, so that finally we can do a u64/u64 division.
1444          */
1445         while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1446                 REDUCE_FLS(nsec, frequency);
1447                 REDUCE_FLS(sec, count);
1448         }
1449
1450         if (count_fls + sec_fls > 64) {
1451                 divisor = nsec * frequency;
1452
1453                 while (count_fls + sec_fls > 64) {
1454                         REDUCE_FLS(count, sec);
1455                         divisor >>= 1;
1456                 }
1457
1458                 dividend = count * sec;
1459         } else {
1460                 dividend = count * sec;
1461
1462                 while (nsec_fls + frequency_fls > 64) {
1463                         REDUCE_FLS(nsec, frequency);
1464                         dividend >>= 1;
1465                 }
1466
1467                 divisor = nsec * frequency;
1468         }
1469
1470         return div64_u64(dividend, divisor);
1471 }
1472
1473 static void perf_event_stop(struct perf_event *event)
1474 {
1475         if (!event->pmu->stop)
1476                 return event->pmu->disable(event);
1477
1478         return event->pmu->stop(event);
1479 }
1480
1481 static int perf_event_start(struct perf_event *event)
1482 {
1483         if (!event->pmu->start)
1484                 return event->pmu->enable(event);
1485
1486         return event->pmu->start(event);
1487 }
1488
1489 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1490 {
1491         struct hw_perf_event *hwc = &event->hw;
1492         u64 period, sample_period;
1493         s64 delta;
1494
1495         period = perf_calculate_period(event, nsec, count);
1496
1497         delta = (s64)(period - hwc->sample_period);
1498         delta = (delta + 7) / 8; /* low pass filter */
1499
1500         sample_period = hwc->sample_period + delta;
1501
1502         if (!sample_period)
1503                 sample_period = 1;
1504
1505         hwc->sample_period = sample_period;
1506
1507         if (atomic64_read(&hwc->period_left) > 8*sample_period) {
1508                 perf_disable();
1509                 perf_event_stop(event);
1510                 atomic64_set(&hwc->period_left, 0);
1511                 perf_event_start(event);
1512                 perf_enable();
1513         }
1514 }
1515
1516 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1517 {
1518         struct perf_event *event;
1519         struct hw_perf_event *hwc;
1520         u64 interrupts, now;
1521         s64 delta;
1522
1523         raw_spin_lock(&ctx->lock);
1524         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1525                 if (event->state != PERF_EVENT_STATE_ACTIVE)
1526                         continue;
1527
1528                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1529                         continue;
1530
1531                 hwc = &event->hw;
1532
1533                 interrupts = hwc->interrupts;
1534                 hwc->interrupts = 0;
1535
1536                 /*
1537                  * unthrottle events on the tick
1538                  */
1539                 if (interrupts == MAX_INTERRUPTS) {
1540                         perf_log_throttle(event, 1);
1541                         event->pmu->unthrottle(event);
1542                 }
1543
1544                 if (!event->attr.freq || !event->attr.sample_freq)
1545                         continue;
1546
1547                 event->pmu->read(event);
1548                 now = atomic64_read(&event->count);
1549                 delta = now - hwc->freq_count_stamp;
1550                 hwc->freq_count_stamp = now;
1551
1552                 if (delta > 0)
1553                         perf_adjust_period(event, TICK_NSEC, delta);
1554         }
1555         raw_spin_unlock(&ctx->lock);
1556 }
1557
1558 /*
1559  * Round-robin a context's events:
1560  */
1561 static void rotate_ctx(struct perf_event_context *ctx)
1562 {
1563         if (!ctx->nr_events)
1564                 return;
1565
1566         raw_spin_lock(&ctx->lock);
1567
1568         /* Rotate the first entry last of non-pinned groups */
1569         list_rotate_left(&ctx->flexible_groups);
1570
1571         raw_spin_unlock(&ctx->lock);
1572 }
1573
1574 void perf_event_task_tick(struct task_struct *curr)
1575 {
1576         struct perf_cpu_context *cpuctx;
1577         struct perf_event_context *ctx;
1578
1579         if (!atomic_read(&nr_events))
1580                 return;
1581
1582         cpuctx = &__get_cpu_var(perf_cpu_context);
1583         ctx = curr->perf_event_ctxp;
1584
1585         perf_disable();
1586
1587         perf_ctx_adjust_freq(&cpuctx->ctx);
1588         if (ctx)
1589                 perf_ctx_adjust_freq(ctx);
1590
1591         cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1592         if (ctx)
1593                 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1594
1595         rotate_ctx(&cpuctx->ctx);
1596         if (ctx)
1597                 rotate_ctx(ctx);
1598
1599         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1600         if (ctx)
1601                 task_ctx_sched_in(curr, EVENT_FLEXIBLE);
1602
1603         perf_enable();
1604 }
1605
1606 static int event_enable_on_exec(struct perf_event *event,
1607                                 struct perf_event_context *ctx)
1608 {
1609         if (!event->attr.enable_on_exec)
1610                 return 0;
1611
1612         event->attr.enable_on_exec = 0;
1613         if (event->state >= PERF_EVENT_STATE_INACTIVE)
1614                 return 0;
1615
1616         __perf_event_mark_enabled(event, ctx);
1617
1618         return 1;
1619 }
1620
1621 /*
1622  * Enable all of a task's events that have been marked enable-on-exec.
1623  * This expects task == current.
1624  */
1625 static void perf_event_enable_on_exec(struct task_struct *task)
1626 {
1627         struct perf_event_context *ctx;
1628         struct perf_event *event;
1629         unsigned long flags;
1630         int enabled = 0;
1631         int ret;
1632
1633         local_irq_save(flags);
1634         ctx = task->perf_event_ctxp;
1635         if (!ctx || !ctx->nr_events)
1636                 goto out;
1637
1638         __perf_event_task_sched_out(ctx);
1639
1640         raw_spin_lock(&ctx->lock);
1641
1642         list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1643                 ret = event_enable_on_exec(event, ctx);
1644                 if (ret)
1645                         enabled = 1;
1646         }
1647
1648         list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1649                 ret = event_enable_on_exec(event, ctx);
1650                 if (ret)
1651                         enabled = 1;
1652         }
1653
1654         /*
1655          * Unclone this context if we enabled any event.
1656          */
1657         if (enabled)
1658                 unclone_ctx(ctx);
1659
1660         raw_spin_unlock(&ctx->lock);
1661
1662         perf_event_task_sched_in(task);
1663  out:
1664         local_irq_restore(flags);
1665 }
1666
1667 /*
1668  * Cross CPU call to read the hardware event
1669  */
1670 static void __perf_event_read(void *info)
1671 {
1672         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1673         struct perf_event *event = info;
1674         struct perf_event_context *ctx = event->ctx;
1675
1676         /*
1677          * If this is a task context, we need to check whether it is
1678          * the current task context of this cpu.  If not it has been
1679          * scheduled out before the smp call arrived.  In that case
1680          * event->count would have been updated to a recent sample
1681          * when the event was scheduled out.
1682          */
1683         if (ctx->task && cpuctx->task_ctx != ctx)
1684                 return;
1685
1686         raw_spin_lock(&ctx->lock);
1687         update_context_time(ctx);
1688         update_event_times(event);
1689         raw_spin_unlock(&ctx->lock);
1690
1691         event->pmu->read(event);
1692 }
1693
1694 static u64 perf_event_read(struct perf_event *event)
1695 {
1696         /*
1697          * If event is enabled and currently active on a CPU, update the
1698          * value in the event structure:
1699          */
1700         if (event->state == PERF_EVENT_STATE_ACTIVE) {
1701                 smp_call_function_single(event->oncpu,
1702                                          __perf_event_read, event, 1);
1703         } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1704                 struct perf_event_context *ctx = event->ctx;
1705                 unsigned long flags;
1706
1707                 raw_spin_lock_irqsave(&ctx->lock, flags);
1708                 update_context_time(ctx);
1709                 update_event_times(event);
1710                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1711         }
1712
1713         return atomic64_read(&event->count);
1714 }
1715
1716 /*
1717  * Initialize the perf_event context in a task_struct:
1718  */
1719 static void
1720 __perf_event_init_context(struct perf_event_context *ctx,
1721                             struct task_struct *task)
1722 {
1723         raw_spin_lock_init(&ctx->lock);
1724         mutex_init(&ctx->mutex);
1725         INIT_LIST_HEAD(&ctx->pinned_groups);
1726         INIT_LIST_HEAD(&ctx->flexible_groups);
1727         INIT_LIST_HEAD(&ctx->event_list);
1728         atomic_set(&ctx->refcount, 1);
1729         ctx->task = task;
1730 }
1731
1732 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1733 {
1734         struct perf_event_context *ctx;
1735         struct perf_cpu_context *cpuctx;
1736         struct task_struct *task;
1737         unsigned long flags;
1738         int err;
1739
1740         if (pid == -1 && cpu != -1) {
1741                 /* Must be root to operate on a CPU event: */
1742                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1743                         return ERR_PTR(-EACCES);
1744
1745                 if (cpu < 0 || cpu >= nr_cpumask_bits)
1746                         return ERR_PTR(-EINVAL);
1747
1748                 /*
1749                  * We could be clever and allow to attach a event to an
1750                  * offline CPU and activate it when the CPU comes up, but
1751                  * that's for later.
1752                  */
1753                 if (!cpu_online(cpu))
1754                         return ERR_PTR(-ENODEV);
1755
1756                 cpuctx = &per_cpu(perf_cpu_context, cpu);
1757                 ctx = &cpuctx->ctx;
1758                 get_ctx(ctx);
1759
1760                 return ctx;
1761         }
1762
1763         rcu_read_lock();
1764         if (!pid)
1765                 task = current;
1766         else
1767                 task = find_task_by_vpid(pid);
1768         if (task)
1769                 get_task_struct(task);
1770         rcu_read_unlock();
1771
1772         if (!task)
1773                 return ERR_PTR(-ESRCH);
1774
1775         /*
1776          * Can't attach events to a dying task.
1777          */
1778         err = -ESRCH;
1779         if (task->flags & PF_EXITING)
1780                 goto errout;
1781
1782         /* Reuse ptrace permission checks for now. */
1783         err = -EACCES;
1784         if (!ptrace_may_access(task, PTRACE_MODE_READ))
1785                 goto errout;
1786
1787  retry:
1788         ctx = perf_lock_task_context(task, &flags);
1789         if (ctx) {
1790                 unclone_ctx(ctx);
1791                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1792         }
1793
1794         if (!ctx) {
1795                 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1796                 err = -ENOMEM;
1797                 if (!ctx)
1798                         goto errout;
1799                 __perf_event_init_context(ctx, task);
1800                 get_ctx(ctx);
1801                 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1802                         /*
1803                          * We raced with some other task; use
1804                          * the context they set.
1805                          */
1806                         kfree(ctx);
1807                         goto retry;
1808                 }
1809                 get_task_struct(task);
1810         }
1811
1812         put_task_struct(task);
1813         return ctx;
1814
1815  errout:
1816         put_task_struct(task);
1817         return ERR_PTR(err);
1818 }
1819
1820 static void perf_event_free_filter(struct perf_event *event);
1821
1822 static void free_event_rcu(struct rcu_head *head)
1823 {
1824         struct perf_event *event;
1825
1826         event = container_of(head, struct perf_event, rcu_head);
1827         if (event->ns)
1828                 put_pid_ns(event->ns);
1829         perf_event_free_filter(event);
1830         kfree(event);
1831 }
1832
1833 static void perf_pending_sync(struct perf_event *event);
1834
1835 static void free_event(struct perf_event *event)
1836 {
1837         perf_pending_sync(event);
1838
1839         if (!event->parent) {
1840                 atomic_dec(&nr_events);
1841                 if (event->attr.mmap)
1842                         atomic_dec(&nr_mmap_events);
1843                 if (event->attr.comm)
1844                         atomic_dec(&nr_comm_events);
1845                 if (event->attr.task)
1846                         atomic_dec(&nr_task_events);
1847         }
1848
1849         if (event->output) {
1850                 fput(event->output->filp);
1851                 event->output = NULL;
1852         }
1853
1854         if (event->destroy)
1855                 event->destroy(event);
1856
1857         put_ctx(event->ctx);
1858         call_rcu(&event->rcu_head, free_event_rcu);
1859 }
1860
1861 int perf_event_release_kernel(struct perf_event *event)
1862 {
1863         struct perf_event_context *ctx = event->ctx;
1864
1865         WARN_ON_ONCE(ctx->parent_ctx);
1866         mutex_lock(&ctx->mutex);
1867         perf_event_remove_from_context(event);
1868         mutex_unlock(&ctx->mutex);
1869
1870         mutex_lock(&event->owner->perf_event_mutex);
1871         list_del_init(&event->owner_entry);
1872         mutex_unlock(&event->owner->perf_event_mutex);
1873         put_task_struct(event->owner);
1874
1875         free_event(event);
1876
1877         return 0;
1878 }
1879 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1880
1881 /*
1882  * Called when the last reference to the file is gone.
1883  */
1884 static int perf_release(struct inode *inode, struct file *file)
1885 {
1886         struct perf_event *event = file->private_data;
1887
1888         file->private_data = NULL;
1889
1890         return perf_event_release_kernel(event);
1891 }
1892
1893 static int perf_event_read_size(struct perf_event *event)
1894 {
1895         int entry = sizeof(u64); /* value */
1896         int size = 0;
1897         int nr = 1;
1898
1899         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1900                 size += sizeof(u64);
1901
1902         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1903                 size += sizeof(u64);
1904
1905         if (event->attr.read_format & PERF_FORMAT_ID)
1906                 entry += sizeof(u64);
1907
1908         if (event->attr.read_format & PERF_FORMAT_GROUP) {
1909                 nr += event->group_leader->nr_siblings;
1910                 size += sizeof(u64);
1911         }
1912
1913         size += entry * nr;
1914
1915         return size;
1916 }
1917
1918 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1919 {
1920         struct perf_event *child;
1921         u64 total = 0;
1922
1923         *enabled = 0;
1924         *running = 0;
1925
1926         mutex_lock(&event->child_mutex);
1927         total += perf_event_read(event);
1928         *enabled += event->total_time_enabled +
1929                         atomic64_read(&event->child_total_time_enabled);
1930         *running += event->total_time_running +
1931                         atomic64_read(&event->child_total_time_running);
1932
1933         list_for_each_entry(child, &event->child_list, child_list) {
1934                 total += perf_event_read(child);
1935                 *enabled += child->total_time_enabled;
1936                 *running += child->total_time_running;
1937         }
1938         mutex_unlock(&event->child_mutex);
1939
1940         return total;
1941 }
1942 EXPORT_SYMBOL_GPL(perf_event_read_value);
1943
1944 static int perf_event_read_group(struct perf_event *event,
1945                                    u64 read_format, char __user *buf)
1946 {
1947         struct perf_event *leader = event->group_leader, *sub;
1948         int n = 0, size = 0, ret = -EFAULT;
1949         struct perf_event_context *ctx = leader->ctx;
1950         u64 values[5];
1951         u64 count, enabled, running;
1952
1953         mutex_lock(&ctx->mutex);
1954         count = perf_event_read_value(leader, &enabled, &running);
1955
1956         values[n++] = 1 + leader->nr_siblings;
1957         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1958                 values[n++] = enabled;
1959         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1960                 values[n++] = running;
1961         values[n++] = count;
1962         if (read_format & PERF_FORMAT_ID)
1963                 values[n++] = primary_event_id(leader);
1964
1965         size = n * sizeof(u64);
1966
1967         if (copy_to_user(buf, values, size))
1968                 goto unlock;
1969
1970         ret = size;
1971
1972         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1973                 n = 0;
1974
1975                 values[n++] = perf_event_read_value(sub, &enabled, &running);
1976                 if (read_format & PERF_FORMAT_ID)
1977                         values[n++] = primary_event_id(sub);
1978
1979                 size = n * sizeof(u64);
1980
1981                 if (copy_to_user(buf + ret, values, size)) {
1982                         ret = -EFAULT;
1983                         goto unlock;
1984                 }
1985
1986                 ret += size;
1987         }
1988 unlock:
1989         mutex_unlock(&ctx->mutex);
1990
1991         return ret;
1992 }
1993
1994 static int perf_event_read_one(struct perf_event *event,
1995                                  u64 read_format, char __user *buf)
1996 {
1997         u64 enabled, running;
1998         u64 values[4];
1999         int n = 0;
2000
2001         values[n++] = perf_event_read_value(event, &enabled, &running);
2002         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2003                 values[n++] = enabled;
2004         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2005                 values[n++] = running;
2006         if (read_format & PERF_FORMAT_ID)
2007                 values[n++] = primary_event_id(event);
2008
2009         if (copy_to_user(buf, values, n * sizeof(u64)))
2010                 return -EFAULT;
2011
2012         return n * sizeof(u64);
2013 }
2014
2015 /*
2016  * Read the performance event - simple non blocking version for now
2017  */
2018 static ssize_t
2019 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2020 {
2021         u64 read_format = event->attr.read_format;
2022         int ret;
2023
2024         /*
2025          * Return end-of-file for a read on a event that is in
2026          * error state (i.e. because it was pinned but it couldn't be
2027          * scheduled on to the CPU at some point).
2028          */
2029         if (event->state == PERF_EVENT_STATE_ERROR)
2030                 return 0;
2031
2032         if (count < perf_event_read_size(event))
2033                 return -ENOSPC;
2034
2035         WARN_ON_ONCE(event->ctx->parent_ctx);
2036         if (read_format & PERF_FORMAT_GROUP)
2037                 ret = perf_event_read_group(event, read_format, buf);
2038         else
2039                 ret = perf_event_read_one(event, read_format, buf);
2040
2041         return ret;
2042 }
2043
2044 static ssize_t
2045 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2046 {
2047         struct perf_event *event = file->private_data;
2048
2049         return perf_read_hw(event, buf, count);
2050 }
2051
2052 static unsigned int perf_poll(struct file *file, poll_table *wait)
2053 {
2054         struct perf_event *event = file->private_data;
2055         struct perf_mmap_data *data;
2056         unsigned int events = POLL_HUP;
2057
2058         rcu_read_lock();
2059         data = rcu_dereference(event->data);
2060         if (data)
2061                 events = atomic_xchg(&data->poll, 0);
2062         rcu_read_unlock();
2063
2064         poll_wait(file, &event->waitq, wait);
2065
2066         return events;
2067 }
2068
2069 static void perf_event_reset(struct perf_event *event)
2070 {
2071         (void)perf_event_read(event);
2072         atomic64_set(&event->count, 0);
2073         perf_event_update_userpage(event);
2074 }
2075
2076 /*
2077  * Holding the top-level event's child_mutex means that any
2078  * descendant process that has inherited this event will block
2079  * in sync_child_event if it goes to exit, thus satisfying the
2080  * task existence requirements of perf_event_enable/disable.
2081  */
2082 static void perf_event_for_each_child(struct perf_event *event,
2083                                         void (*func)(struct perf_event *))
2084 {
2085         struct perf_event *child;
2086
2087         WARN_ON_ONCE(event->ctx->parent_ctx);
2088         mutex_lock(&event->child_mutex);
2089         func(event);
2090         list_for_each_entry(child, &event->child_list, child_list)
2091                 func(child);
2092         mutex_unlock(&event->child_mutex);
2093 }
2094
2095 static void perf_event_for_each(struct perf_event *event,
2096                                   void (*func)(struct perf_event *))
2097 {
2098         struct perf_event_context *ctx = event->ctx;
2099         struct perf_event *sibling;
2100
2101         WARN_ON_ONCE(ctx->parent_ctx);
2102         mutex_lock(&ctx->mutex);
2103         event = event->group_leader;
2104
2105         perf_event_for_each_child(event, func);
2106         func(event);
2107         list_for_each_entry(sibling, &event->sibling_list, group_entry)
2108                 perf_event_for_each_child(event, func);
2109         mutex_unlock(&ctx->mutex);
2110 }
2111
2112 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2113 {
2114         struct perf_event_context *ctx = event->ctx;
2115         unsigned long size;
2116         int ret = 0;
2117         u64 value;
2118
2119         if (!event->attr.sample_period)
2120                 return -EINVAL;
2121
2122         size = copy_from_user(&value, arg, sizeof(value));
2123         if (size != sizeof(value))
2124                 return -EFAULT;
2125
2126         if (!value)
2127                 return -EINVAL;
2128
2129         raw_spin_lock_irq(&ctx->lock);
2130         if (event->attr.freq) {
2131                 if (value > sysctl_perf_event_sample_rate) {
2132                         ret = -EINVAL;
2133                         goto unlock;
2134                 }
2135
2136                 event->attr.sample_freq = value;
2137         } else {
2138                 event->attr.sample_period = value;
2139                 event->hw.sample_period = value;
2140         }
2141 unlock:
2142         raw_spin_unlock_irq(&ctx->lock);
2143
2144         return ret;
2145 }
2146
2147 static int perf_event_set_output(struct perf_event *event, int output_fd);
2148 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2149
2150 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2151 {
2152         struct perf_event *event = file->private_data;
2153         void (*func)(struct perf_event *);
2154         u32 flags = arg;
2155
2156         switch (cmd) {
2157         case PERF_EVENT_IOC_ENABLE:
2158                 func = perf_event_enable;
2159                 break;
2160         case PERF_EVENT_IOC_DISABLE:
2161                 func = perf_event_disable;
2162                 break;
2163         case PERF_EVENT_IOC_RESET:
2164                 func = perf_event_reset;
2165                 break;
2166
2167         case PERF_EVENT_IOC_REFRESH:
2168                 return perf_event_refresh(event, arg);
2169
2170         case PERF_EVENT_IOC_PERIOD:
2171                 return perf_event_period(event, (u64 __user *)arg);
2172
2173         case PERF_EVENT_IOC_SET_OUTPUT:
2174                 return perf_event_set_output(event, arg);
2175
2176         case PERF_EVENT_IOC_SET_FILTER:
2177                 return perf_event_set_filter(event, (void __user *)arg);
2178
2179         default:
2180                 return -ENOTTY;
2181         }
2182
2183         if (flags & PERF_IOC_FLAG_GROUP)
2184                 perf_event_for_each(event, func);
2185         else
2186                 perf_event_for_each_child(event, func);
2187
2188         return 0;
2189 }
2190
2191 int perf_event_task_enable(void)
2192 {
2193         struct perf_event *event;
2194
2195         mutex_lock(&current->perf_event_mutex);
2196         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2197                 perf_event_for_each_child(event, perf_event_enable);
2198         mutex_unlock(&current->perf_event_mutex);
2199
2200         return 0;
2201 }
2202
2203 int perf_event_task_disable(void)
2204 {
2205         struct perf_event *event;
2206
2207         mutex_lock(&current->perf_event_mutex);
2208         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2209                 perf_event_for_each_child(event, perf_event_disable);
2210         mutex_unlock(&current->perf_event_mutex);
2211
2212         return 0;
2213 }
2214
2215 #ifndef PERF_EVENT_INDEX_OFFSET
2216 # define PERF_EVENT_INDEX_OFFSET 0
2217 #endif
2218
2219 static int perf_event_index(struct perf_event *event)
2220 {
2221         if (event->state != PERF_EVENT_STATE_ACTIVE)
2222                 return 0;
2223
2224         return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2225 }
2226
2227 /*
2228  * Callers need to ensure there can be no nesting of this function, otherwise
2229  * the seqlock logic goes bad. We can not serialize this because the arch
2230  * code calls this from NMI context.
2231  */
2232 void perf_event_update_userpage(struct perf_event *event)
2233 {
2234         struct perf_event_mmap_page *userpg;
2235         struct perf_mmap_data *data;
2236
2237         rcu_read_lock();
2238         data = rcu_dereference(event->data);
2239         if (!data)
2240                 goto unlock;
2241
2242         userpg = data->user_page;
2243
2244         /*
2245          * Disable preemption so as to not let the corresponding user-space
2246          * spin too long if we get preempted.
2247          */
2248         preempt_disable();
2249         ++userpg->lock;
2250         barrier();
2251         userpg->index = perf_event_index(event);
2252         userpg->offset = atomic64_read(&event->count);
2253         if (event->state == PERF_EVENT_STATE_ACTIVE)
2254                 userpg->offset -= atomic64_read(&event->hw.prev_count);
2255
2256         userpg->time_enabled = event->total_time_enabled +
2257                         atomic64_read(&event->child_total_time_enabled);
2258
2259         userpg->time_running = event->total_time_running +
2260                         atomic64_read(&event->child_total_time_running);
2261
2262         barrier();
2263         ++userpg->lock;
2264         preempt_enable();
2265 unlock:
2266         rcu_read_unlock();
2267 }
2268
2269 static unsigned long perf_data_size(struct perf_mmap_data *data)
2270 {
2271         return data->nr_pages << (PAGE_SHIFT + data->data_order);
2272 }
2273
2274 #ifndef CONFIG_PERF_USE_VMALLOC
2275
2276 /*
2277  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2278  */
2279
2280 static struct page *
2281 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2282 {
2283         if (pgoff > data->nr_pages)
2284                 return NULL;
2285
2286         if (pgoff == 0)
2287                 return virt_to_page(data->user_page);
2288
2289         return virt_to_page(data->data_pages[pgoff - 1]);
2290 }
2291
2292 static struct perf_mmap_data *
2293 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2294 {
2295         struct perf_mmap_data *data;
2296         unsigned long size;
2297         int i;
2298
2299         WARN_ON(atomic_read(&event->mmap_count));
2300
2301         size = sizeof(struct perf_mmap_data);
2302         size += nr_pages * sizeof(void *);
2303
2304         data = kzalloc(size, GFP_KERNEL);
2305         if (!data)
2306                 goto fail;
2307
2308         data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2309         if (!data->user_page)
2310                 goto fail_user_page;
2311
2312         for (i = 0; i < nr_pages; i++) {
2313                 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2314                 if (!data->data_pages[i])
2315                         goto fail_data_pages;
2316         }
2317
2318         data->data_order = 0;
2319         data->nr_pages = nr_pages;
2320
2321         return data;
2322
2323 fail_data_pages:
2324         for (i--; i >= 0; i--)
2325                 free_page((unsigned long)data->data_pages[i]);
2326
2327         free_page((unsigned long)data->user_page);
2328
2329 fail_user_page:
2330         kfree(data);
2331
2332 fail:
2333         return NULL;
2334 }
2335
2336 static void perf_mmap_free_page(unsigned long addr)
2337 {
2338         struct page *page = virt_to_page((void *)addr);
2339
2340         page->mapping = NULL;
2341         __free_page(page);
2342 }
2343
2344 static void perf_mmap_data_free(struct perf_mmap_data *data)
2345 {
2346         int i;
2347
2348         perf_mmap_free_page((unsigned long)data->user_page);
2349         for (i = 0; i < data->nr_pages; i++)
2350                 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2351         kfree(data);
2352 }
2353
2354 #else
2355
2356 /*
2357  * Back perf_mmap() with vmalloc memory.
2358  *
2359  * Required for architectures that have d-cache aliasing issues.
2360  */
2361
2362 static struct page *
2363 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2364 {
2365         if (pgoff > (1UL << data->data_order))
2366                 return NULL;
2367
2368         return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2369 }
2370
2371 static void perf_mmap_unmark_page(void *addr)
2372 {
2373         struct page *page = vmalloc_to_page(addr);
2374
2375         page->mapping = NULL;
2376 }
2377
2378 static void perf_mmap_data_free_work(struct work_struct *work)
2379 {
2380         struct perf_mmap_data *data;
2381         void *base;
2382         int i, nr;
2383
2384         data = container_of(work, struct perf_mmap_data, work);
2385         nr = 1 << data->data_order;
2386
2387         base = data->user_page;
2388         for (i = 0; i < nr + 1; i++)
2389                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2390
2391         vfree(base);
2392         kfree(data);
2393 }
2394
2395 static void perf_mmap_data_free(struct perf_mmap_data *data)
2396 {
2397         schedule_work(&data->work);
2398 }
2399
2400 static struct perf_mmap_data *
2401 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2402 {
2403         struct perf_mmap_data *data;
2404         unsigned long size;
2405         void *all_buf;
2406
2407         WARN_ON(atomic_read(&event->mmap_count));
2408
2409         size = sizeof(struct perf_mmap_data);
2410         size += sizeof(void *);
2411
2412         data = kzalloc(size, GFP_KERNEL);
2413         if (!data)
2414                 goto fail;
2415
2416         INIT_WORK(&data->work, perf_mmap_data_free_work);
2417
2418         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2419         if (!all_buf)
2420                 goto fail_all_buf;
2421
2422         data->user_page = all_buf;
2423         data->data_pages[0] = all_buf + PAGE_SIZE;
2424         data->data_order = ilog2(nr_pages);
2425         data->nr_pages = 1;
2426
2427         return data;
2428
2429 fail_all_buf:
2430         kfree(data);
2431
2432 fail:
2433         return NULL;
2434 }
2435
2436 #endif
2437
2438 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2439 {
2440         struct perf_event *event = vma->vm_file->private_data;
2441         struct perf_mmap_data *data;
2442         int ret = VM_FAULT_SIGBUS;
2443
2444         if (vmf->flags & FAULT_FLAG_MKWRITE) {
2445                 if (vmf->pgoff == 0)
2446                         ret = 0;
2447                 return ret;
2448         }
2449
2450         rcu_read_lock();
2451         data = rcu_dereference(event->data);
2452         if (!data)
2453                 goto unlock;
2454
2455         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2456                 goto unlock;
2457
2458         vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2459         if (!vmf->page)
2460                 goto unlock;
2461
2462         get_page(vmf->page);
2463         vmf->page->mapping = vma->vm_file->f_mapping;
2464         vmf->page->index   = vmf->pgoff;
2465
2466         ret = 0;
2467 unlock:
2468         rcu_read_unlock();
2469
2470         return ret;
2471 }
2472
2473 static void
2474 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2475 {
2476         long max_size = perf_data_size(data);
2477
2478         atomic_set(&data->lock, -1);
2479
2480         if (event->attr.watermark) {
2481                 data->watermark = min_t(long, max_size,
2482                                         event->attr.wakeup_watermark);
2483         }
2484
2485         if (!data->watermark)
2486                 data->watermark = max_size / 2;
2487
2488
2489         rcu_assign_pointer(event->data, data);
2490 }
2491
2492 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2493 {
2494         struct perf_mmap_data *data;
2495
2496         data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2497         perf_mmap_data_free(data);
2498 }
2499
2500 static void perf_mmap_data_release(struct perf_event *event)
2501 {
2502         struct perf_mmap_data *data = event->data;
2503
2504         WARN_ON(atomic_read(&event->mmap_count));
2505
2506         rcu_assign_pointer(event->data, NULL);
2507         call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2508 }
2509
2510 static void perf_mmap_open(struct vm_area_struct *vma)
2511 {
2512         struct perf_event *event = vma->vm_file->private_data;
2513
2514         atomic_inc(&event->mmap_count);
2515 }
2516
2517 static void perf_mmap_close(struct vm_area_struct *vma)
2518 {
2519         struct perf_event *event = vma->vm_file->private_data;
2520
2521         WARN_ON_ONCE(event->ctx->parent_ctx);
2522         if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2523                 unsigned long size = perf_data_size(event->data);
2524                 struct user_struct *user = current_user();
2525
2526                 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2527                 vma->vm_mm->locked_vm -= event->data->nr_locked;
2528                 perf_mmap_data_release(event);
2529                 mutex_unlock(&event->mmap_mutex);
2530         }
2531 }
2532
2533 static const struct vm_operations_struct perf_mmap_vmops = {
2534         .open           = perf_mmap_open,
2535         .close          = perf_mmap_close,
2536         .fault          = perf_mmap_fault,
2537         .page_mkwrite   = perf_mmap_fault,
2538 };
2539
2540 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2541 {
2542         struct perf_event *event = file->private_data;
2543         unsigned long user_locked, user_lock_limit;
2544         struct user_struct *user = current_user();
2545         unsigned long locked, lock_limit;
2546         struct perf_mmap_data *data;
2547         unsigned long vma_size;
2548         unsigned long nr_pages;
2549         long user_extra, extra;
2550         int ret = 0;
2551
2552         if (!(vma->vm_flags & VM_SHARED))
2553                 return -EINVAL;
2554
2555         vma_size = vma->vm_end - vma->vm_start;
2556         nr_pages = (vma_size / PAGE_SIZE) - 1;
2557
2558         /*
2559          * If we have data pages ensure they're a power-of-two number, so we
2560          * can do bitmasks instead of modulo.
2561          */
2562         if (nr_pages != 0 && !is_power_of_2(nr_pages))
2563                 return -EINVAL;
2564
2565         if (vma_size != PAGE_SIZE * (1 + nr_pages))
2566                 return -EINVAL;
2567
2568         if (vma->vm_pgoff != 0)
2569                 return -EINVAL;
2570
2571         WARN_ON_ONCE(event->ctx->parent_ctx);
2572         mutex_lock(&event->mmap_mutex);
2573         if (event->output) {
2574                 ret = -EINVAL;
2575                 goto unlock;
2576         }
2577
2578         if (atomic_inc_not_zero(&event->mmap_count)) {
2579                 if (nr_pages != event->data->nr_pages)
2580                         ret = -EINVAL;
2581                 goto unlock;
2582         }
2583
2584         user_extra = nr_pages + 1;
2585         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2586
2587         /*
2588          * Increase the limit linearly with more CPUs:
2589          */
2590         user_lock_limit *= num_online_cpus();
2591
2592         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2593
2594         extra = 0;
2595         if (user_locked > user_lock_limit)
2596                 extra = user_locked - user_lock_limit;
2597
2598         lock_limit = rlimit(RLIMIT_MEMLOCK);
2599         lock_limit >>= PAGE_SHIFT;
2600         locked = vma->vm_mm->locked_vm + extra;
2601
2602         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2603                 !capable(CAP_IPC_LOCK)) {
2604                 ret = -EPERM;
2605                 goto unlock;
2606         }
2607
2608         WARN_ON(event->data);
2609
2610         data = perf_mmap_data_alloc(event, nr_pages);
2611         ret = -ENOMEM;
2612         if (!data)
2613                 goto unlock;
2614
2615         ret = 0;
2616         perf_mmap_data_init(event, data);
2617
2618         atomic_set(&event->mmap_count, 1);
2619         atomic_long_add(user_extra, &user->locked_vm);
2620         vma->vm_mm->locked_vm += extra;
2621         event->data->nr_locked = extra;
2622         if (vma->vm_flags & VM_WRITE)
2623                 event->data->writable = 1;
2624
2625 unlock:
2626         mutex_unlock(&event->mmap_mutex);
2627
2628         vma->vm_flags |= VM_RESERVED;
2629         vma->vm_ops = &perf_mmap_vmops;
2630
2631         return ret;
2632 }
2633
2634 static int perf_fasync(int fd, struct file *filp, int on)
2635 {
2636         struct inode *inode = filp->f_path.dentry->d_inode;
2637         struct perf_event *event = filp->private_data;
2638         int retval;
2639
2640         mutex_lock(&inode->i_mutex);
2641         retval = fasync_helper(fd, filp, on, &event->fasync);
2642         mutex_unlock(&inode->i_mutex);
2643
2644         if (retval < 0)
2645                 return retval;
2646
2647         return 0;
2648 }
2649
2650 static const struct file_operations perf_fops = {
2651         .release                = perf_release,
2652         .read                   = perf_read,
2653         .poll                   = perf_poll,
2654         .unlocked_ioctl         = perf_ioctl,
2655         .compat_ioctl           = perf_ioctl,
2656         .mmap                   = perf_mmap,
2657         .fasync                 = perf_fasync,
2658 };
2659
2660 /*
2661  * Perf event wakeup
2662  *
2663  * If there's data, ensure we set the poll() state and publish everything
2664  * to user-space before waking everybody up.
2665  */
2666
2667 void perf_event_wakeup(struct perf_event *event)
2668 {
2669         wake_up_all(&event->waitq);
2670
2671         if (event->pending_kill) {
2672                 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2673                 event->pending_kill = 0;
2674         }
2675 }
2676
2677 /*
2678  * Pending wakeups
2679  *
2680  * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2681  *
2682  * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2683  * single linked list and use cmpxchg() to add entries lockless.
2684  */
2685
2686 static void perf_pending_event(struct perf_pending_entry *entry)
2687 {
2688         struct perf_event *event = container_of(entry,
2689                         struct perf_event, pending);
2690
2691         if (event->pending_disable) {
2692                 event->pending_disable = 0;
2693                 __perf_event_disable(event);
2694         }
2695
2696         if (event->pending_wakeup) {
2697                 event->pending_wakeup = 0;
2698                 perf_event_wakeup(event);
2699         }
2700 }
2701
2702 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2703
2704 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2705         PENDING_TAIL,
2706 };
2707
2708 static void perf_pending_queue(struct perf_pending_entry *entry,
2709                                void (*func)(struct perf_pending_entry *))
2710 {
2711         struct perf_pending_entry **head;
2712
2713         if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2714                 return;
2715
2716         entry->func = func;
2717
2718         head = &get_cpu_var(perf_pending_head);
2719
2720         do {
2721                 entry->next = *head;
2722         } while (cmpxchg(head, entry->next, entry) != entry->next);
2723
2724         set_perf_event_pending();
2725
2726         put_cpu_var(perf_pending_head);
2727 }
2728
2729 static int __perf_pending_run(void)
2730 {
2731         struct perf_pending_entry *list;
2732         int nr = 0;
2733
2734         list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2735         while (list != PENDING_TAIL) {
2736                 void (*func)(struct perf_pending_entry *);
2737                 struct perf_pending_entry *entry = list;
2738
2739                 list = list->next;
2740
2741                 func = entry->func;
2742                 entry->next = NULL;
2743                 /*
2744                  * Ensure we observe the unqueue before we issue the wakeup,
2745                  * so that we won't be waiting forever.
2746                  * -- see perf_not_pending().
2747                  */
2748                 smp_wmb();
2749
2750                 func(entry);
2751                 nr++;
2752         }
2753
2754         return nr;
2755 }
2756
2757 static inline int perf_not_pending(struct perf_event *event)
2758 {
2759         /*
2760          * If we flush on whatever cpu we run, there is a chance we don't
2761          * need to wait.
2762          */
2763         get_cpu();
2764         __perf_pending_run();
2765         put_cpu();
2766
2767         /*
2768          * Ensure we see the proper queue state before going to sleep
2769          * so that we do not miss the wakeup. -- see perf_pending_handle()
2770          */
2771         smp_rmb();
2772         return event->pending.next == NULL;
2773 }
2774
2775 static void perf_pending_sync(struct perf_event *event)
2776 {
2777         wait_event(event->waitq, perf_not_pending(event));
2778 }
2779
2780 void perf_event_do_pending(void)
2781 {
2782         __perf_pending_run();
2783 }
2784
2785 /*
2786  * Callchain support -- arch specific
2787  */
2788
2789 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2790 {
2791         return NULL;
2792 }
2793
2794 /*
2795  * Output
2796  */
2797 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2798                               unsigned long offset, unsigned long head)
2799 {
2800         unsigned long mask;
2801
2802         if (!data->writable)
2803                 return true;
2804
2805         mask = perf_data_size(data) - 1;
2806
2807         offset = (offset - tail) & mask;
2808         head   = (head   - tail) & mask;
2809
2810         if ((int)(head - offset) < 0)
2811                 return false;
2812
2813         return true;
2814 }
2815
2816 static void perf_output_wakeup(struct perf_output_handle *handle)
2817 {
2818         atomic_set(&handle->data->poll, POLL_IN);
2819
2820         if (handle->nmi) {
2821                 handle->event->pending_wakeup = 1;
2822                 perf_pending_queue(&handle->event->pending,
2823                                    perf_pending_event);
2824         } else
2825                 perf_event_wakeup(handle->event);
2826 }
2827
2828 /*
2829  * Curious locking construct.
2830  *
2831  * We need to ensure a later event_id doesn't publish a head when a former
2832  * event_id isn't done writing. However since we need to deal with NMIs we
2833  * cannot fully serialize things.
2834  *
2835  * What we do is serialize between CPUs so we only have to deal with NMI
2836  * nesting on a single CPU.
2837  *
2838  * We only publish the head (and generate a wakeup) when the outer-most
2839  * event_id completes.
2840  */
2841 static void perf_output_lock(struct perf_output_handle *handle)
2842 {
2843         struct perf_mmap_data *data = handle->data;
2844         int cur, cpu = get_cpu();
2845
2846         handle->locked = 0;
2847
2848         for (;;) {
2849                 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2850                 if (cur == -1) {
2851                         handle->locked = 1;
2852                         break;
2853                 }
2854                 if (cur == cpu)
2855                         break;
2856
2857                 cpu_relax();
2858         }
2859 }
2860
2861 static void perf_output_unlock(struct perf_output_handle *handle)
2862 {
2863         struct perf_mmap_data *data = handle->data;
2864         unsigned long head;
2865         int cpu;
2866
2867         data->done_head = data->head;
2868
2869         if (!handle->locked)
2870                 goto out;
2871
2872 again:
2873         /*
2874          * The xchg implies a full barrier that ensures all writes are done
2875          * before we publish the new head, matched by a rmb() in userspace when
2876          * reading this position.
2877          */
2878         while ((head = atomic_long_xchg(&data->done_head, 0)))
2879                 data->user_page->data_head = head;
2880
2881         /*
2882          * NMI can happen here, which means we can miss a done_head update.
2883          */
2884
2885         cpu = atomic_xchg(&data->lock, -1);
2886         WARN_ON_ONCE(cpu != smp_processor_id());
2887
2888         /*
2889          * Therefore we have to validate we did not indeed do so.
2890          */
2891         if (unlikely(atomic_long_read(&data->done_head))) {
2892                 /*
2893                  * Since we had it locked, we can lock it again.
2894                  */
2895                 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2896                         cpu_relax();
2897
2898                 goto again;
2899         }
2900
2901         if (atomic_xchg(&data->wakeup, 0))
2902                 perf_output_wakeup(handle);
2903 out:
2904         put_cpu();
2905 }
2906
2907 void perf_output_copy(struct perf_output_handle *handle,
2908                       const void *buf, unsigned int len)
2909 {
2910         unsigned int pages_mask;
2911         unsigned long offset;
2912         unsigned int size;
2913         void **pages;
2914
2915         offset          = handle->offset;
2916         pages_mask      = handle->data->nr_pages - 1;
2917         pages           = handle->data->data_pages;
2918
2919         do {
2920                 unsigned long page_offset;
2921                 unsigned long page_size;
2922                 int nr;
2923
2924                 nr          = (offset >> PAGE_SHIFT) & pages_mask;
2925                 page_size   = 1UL << (handle->data->data_order + PAGE_SHIFT);
2926                 page_offset = offset & (page_size - 1);
2927                 size        = min_t(unsigned int, page_size - page_offset, len);
2928
2929                 memcpy(pages[nr] + page_offset, buf, size);
2930
2931                 len         -= size;
2932                 buf         += size;
2933                 offset      += size;
2934         } while (len);
2935
2936         handle->offset = offset;
2937
2938         /*
2939          * Check we didn't copy past our reservation window, taking the
2940          * possible unsigned int wrap into account.
2941          */
2942         WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2943 }
2944
2945 int perf_output_begin(struct perf_output_handle *handle,
2946                       struct perf_event *event, unsigned int size,
2947                       int nmi, int sample)
2948 {
2949         struct perf_event *output_event;
2950         struct perf_mmap_data *data;
2951         unsigned long tail, offset, head;
2952         int have_lost;
2953         struct {
2954                 struct perf_event_header header;
2955                 u64                      id;
2956                 u64                      lost;
2957         } lost_event;
2958
2959         rcu_read_lock();
2960         /*
2961          * For inherited events we send all the output towards the parent.
2962          */
2963         if (event->parent)
2964                 event = event->parent;
2965
2966         output_event = rcu_dereference(event->output);
2967         if (output_event)
2968                 event = output_event;
2969
2970         data = rcu_dereference(event->data);
2971         if (!data)
2972                 goto out;
2973
2974         handle->data    = data;
2975         handle->event   = event;
2976         handle->nmi     = nmi;
2977         handle->sample  = sample;
2978
2979         if (!data->nr_pages)
2980                 goto fail;
2981
2982         have_lost = atomic_read(&data->lost);
2983         if (have_lost)
2984                 size += sizeof(lost_event);
2985
2986         perf_output_lock(handle);
2987
2988         do {
2989                 /*
2990                  * Userspace could choose to issue a mb() before updating the
2991                  * tail pointer. So that all reads will be completed before the
2992                  * write is issued.
2993                  */
2994                 tail = ACCESS_ONCE(data->user_page->data_tail);
2995                 smp_rmb();
2996                 offset = head = atomic_long_read(&data->head);
2997                 head += size;
2998                 if (unlikely(!perf_output_space(data, tail, offset, head)))
2999                         goto fail;
3000         } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
3001
3002         handle->offset  = offset;
3003         handle->head    = head;
3004
3005         if (head - tail > data->watermark)
3006                 atomic_set(&data->wakeup, 1);
3007
3008         if (have_lost) {
3009                 lost_event.header.type = PERF_RECORD_LOST;
3010                 lost_event.header.misc = 0;
3011                 lost_event.header.size = sizeof(lost_event);
3012                 lost_event.id          = event->id;
3013                 lost_event.lost        = atomic_xchg(&data->lost, 0);
3014
3015                 perf_output_put(handle, lost_event);
3016         }
3017
3018         return 0;
3019
3020 fail:
3021         atomic_inc(&data->lost);
3022         perf_output_unlock(handle);
3023 out:
3024         rcu_read_unlock();
3025
3026         return -ENOSPC;
3027 }
3028
3029 void perf_output_end(struct perf_output_handle *handle)
3030 {
3031         struct perf_event *event = handle->event;
3032         struct perf_mmap_data *data = handle->data;
3033
3034         int wakeup_events = event->attr.wakeup_events;
3035
3036         if (handle->sample && wakeup_events) {
3037                 int events = atomic_inc_return(&data->events);
3038                 if (events >= wakeup_events) {
3039                         atomic_sub(wakeup_events, &data->events);
3040                         atomic_set(&data->wakeup, 1);
3041                 }
3042         }
3043
3044         perf_output_unlock(handle);
3045         rcu_read_unlock();
3046 }
3047
3048 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3049 {
3050         /*
3051          * only top level events have the pid namespace they were created in
3052          */
3053         if (event->parent)
3054                 event = event->parent;
3055
3056         return task_tgid_nr_ns(p, event->ns);
3057 }
3058
3059 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3060 {
3061         /*
3062          * only top level events have the pid namespace they were created in
3063          */
3064         if (event->parent)
3065                 event = event->parent;
3066
3067         return task_pid_nr_ns(p, event->ns);
3068 }
3069
3070 static void perf_output_read_one(struct perf_output_handle *handle,
3071                                  struct perf_event *event)
3072 {
3073         u64 read_format = event->attr.read_format;
3074         u64 values[4];
3075         int n = 0;
3076
3077         values[n++] = atomic64_read(&event->count);
3078         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3079                 values[n++] = event->total_time_enabled +
3080                         atomic64_read(&event->child_total_time_enabled);
3081         }
3082         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3083                 values[n++] = event->total_time_running +
3084                         atomic64_read(&event->child_total_time_running);
3085         }
3086         if (read_format & PERF_FORMAT_ID)
3087                 values[n++] = primary_event_id(event);
3088
3089         perf_output_copy(handle, values, n * sizeof(u64));
3090 }
3091
3092 /*
3093  * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3094  */
3095 static void perf_output_read_group(struct perf_output_handle *handle,
3096                             struct perf_event *event)
3097 {
3098         struct perf_event *leader = event->group_leader, *sub;
3099         u64 read_format = event->attr.read_format;
3100         u64 values[5];
3101         int n = 0;
3102
3103         values[n++] = 1 + leader->nr_siblings;
3104
3105         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3106                 values[n++] = leader->total_time_enabled;
3107
3108         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3109                 values[n++] = leader->total_time_running;
3110
3111         if (leader != event)
3112                 leader->pmu->read(leader);
3113
3114         values[n++] = atomic64_read(&leader->count);
3115         if (read_format & PERF_FORMAT_ID)
3116                 values[n++] = primary_event_id(leader);
3117
3118         perf_output_copy(handle, values, n * sizeof(u64));
3119
3120         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3121                 n = 0;
3122
3123                 if (sub != event)
3124                         sub->pmu->read(sub);
3125
3126                 values[n++] = atomic64_read(&sub->count);
3127                 if (read_format & PERF_FORMAT_ID)
3128                         values[n++] = primary_event_id(sub);
3129
3130                 perf_output_copy(handle, values, n * sizeof(u64));
3131         }
3132 }
3133
3134 static void perf_output_read(struct perf_output_handle *handle,
3135                              struct perf_event *event)
3136 {
3137         if (event->attr.read_format & PERF_FORMAT_GROUP)
3138                 perf_output_read_group(handle, event);
3139         else
3140                 perf_output_read_one(handle, event);
3141 }
3142
3143 void perf_output_sample(struct perf_output_handle *handle,
3144                         struct perf_event_header *header,
3145                         struct perf_sample_data *data,
3146                         struct perf_event *event)
3147 {
3148         u64 sample_type = data->type;
3149
3150         perf_output_put(handle, *header);
3151
3152         if (sample_type & PERF_SAMPLE_IP)
3153                 perf_output_put(handle, data->ip);
3154
3155         if (sample_type & PERF_SAMPLE_TID)
3156                 perf_output_put(handle, data->tid_entry);
3157
3158         if (sample_type & PERF_SAMPLE_TIME)
3159                 perf_output_put(handle, data->time);
3160
3161         if (sample_type & PERF_SAMPLE_ADDR)
3162                 perf_output_put(handle, data->addr);
3163
3164         if (sample_type & PERF_SAMPLE_ID)
3165                 perf_output_put(handle, data->id);
3166
3167         if (sample_type & PERF_SAMPLE_STREAM_ID)
3168                 perf_output_put(handle, data->stream_id);
3169
3170         if (sample_type & PERF_SAMPLE_CPU)
3171                 perf_output_put(handle, data->cpu_entry);
3172
3173         if (sample_type & PERF_SAMPLE_PERIOD)
3174                 perf_output_put(handle, data->period);
3175
3176         if (sample_type & PERF_SAMPLE_READ)
3177                 perf_output_read(handle, event);
3178
3179         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3180                 if (data->callchain) {
3181                         int size = 1;
3182
3183                         if (data->callchain)
3184                                 size += data->callchain->nr;
3185
3186                         size *= sizeof(u64);
3187
3188                         perf_output_copy(handle, data->callchain, size);
3189                 } else {
3190                         u64 nr = 0;
3191                         perf_output_put(handle, nr);
3192                 }
3193         }
3194
3195         if (sample_type & PERF_SAMPLE_RAW) {
3196                 if (data->raw) {
3197                         perf_output_put(handle, data->raw->size);
3198                         perf_output_copy(handle, data->raw->data,
3199                                          data->raw->size);
3200                 } else {
3201                         struct {
3202                                 u32     size;
3203                                 u32     data;
3204                         } raw = {
3205                                 .size = sizeof(u32),
3206                                 .data = 0,
3207                         };
3208                         perf_output_put(handle, raw);
3209                 }
3210         }
3211 }
3212
3213 void perf_prepare_sample(struct perf_event_header *header,
3214                          struct perf_sample_data *data,
3215                          struct perf_event *event,
3216                          struct pt_regs *regs)
3217 {
3218         u64 sample_type = event->attr.sample_type;
3219
3220         data->type = sample_type;
3221
3222         header->type = PERF_RECORD_SAMPLE;
3223         header->size = sizeof(*header);
3224
3225         header->misc = 0;
3226         header->misc |= perf_misc_flags(regs);
3227
3228         if (sample_type & PERF_SAMPLE_IP) {
3229                 data->ip = perf_instruction_pointer(regs);
3230
3231                 header->size += sizeof(data->ip);
3232         }
3233
3234         if (sample_type & PERF_SAMPLE_TID) {
3235                 /* namespace issues */
3236                 data->tid_entry.pid = perf_event_pid(event, current);
3237                 data->tid_entry.tid = perf_event_tid(event, current);
3238
3239                 header->size += sizeof(data->tid_entry);
3240         }
3241
3242         if (sample_type & PERF_SAMPLE_TIME) {
3243                 data->time = perf_clock();
3244
3245                 header->size += sizeof(data->time);
3246         }
3247
3248         if (sample_type & PERF_SAMPLE_ADDR)
3249                 header->size += sizeof(data->addr);
3250
3251         if (sample_type & PERF_SAMPLE_ID) {
3252                 data->id = primary_event_id(event);
3253
3254                 header->size += sizeof(data->id);
3255         }
3256
3257         if (sample_type & PERF_SAMPLE_STREAM_ID) {
3258                 data->stream_id = event->id;
3259
3260                 header->size += sizeof(data->stream_id);
3261         }
3262
3263         if (sample_type & PERF_SAMPLE_CPU) {
3264                 data->cpu_entry.cpu             = raw_smp_processor_id();
3265                 data->cpu_entry.reserved        = 0;
3266
3267                 header->size += sizeof(data->cpu_entry);
3268         }
3269
3270         if (sample_type & PERF_SAMPLE_PERIOD)
3271                 header->size += sizeof(data->period);
3272
3273         if (sample_type & PERF_SAMPLE_READ)
3274                 header->size += perf_event_read_size(event);
3275
3276         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3277                 int size = 1;
3278
3279                 data->callchain = perf_callchain(regs);
3280
3281                 if (data->callchain)
3282                         size += data->callchain->nr;
3283
3284                 header->size += size * sizeof(u64);
3285         }
3286
3287         if (sample_type & PERF_SAMPLE_RAW) {
3288                 int size = sizeof(u32);
3289
3290                 if (data->raw)
3291                         size += data->raw->size;
3292                 else
3293                         size += sizeof(u32);
3294
3295                 WARN_ON_ONCE(size & (sizeof(u64)-1));
3296                 header->size += size;
3297         }
3298 }
3299
3300 static void perf_event_output(struct perf_event *event, int nmi,
3301                                 struct perf_sample_data *data,
3302                                 struct pt_regs *regs)
3303 {
3304         struct perf_output_handle handle;
3305         struct perf_event_header header;
3306
3307         perf_prepare_sample(&header, data, event, regs);
3308
3309         if (perf_output_begin(&handle, event, header.size, nmi, 1))
3310                 return;
3311
3312         perf_output_sample(&handle, &header, data, event);
3313
3314         perf_output_end(&handle);
3315 }
3316
3317 /*
3318  * read event_id
3319  */
3320
3321 struct perf_read_event {
3322         struct perf_event_header        header;
3323
3324         u32                             pid;
3325         u32                             tid;
3326 };
3327
3328 static void
3329 perf_event_read_event(struct perf_event *event,
3330                         struct task_struct *task)
3331 {
3332         struct perf_output_handle handle;
3333         struct perf_read_event read_event = {
3334                 .header = {
3335                         .type = PERF_RECORD_READ,
3336                         .misc = 0,
3337                         .size = sizeof(read_event) + perf_event_read_size(event),
3338                 },
3339                 .pid = perf_event_pid(event, task),
3340                 .tid = perf_event_tid(event, task),
3341         };
3342         int ret;
3343
3344         ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3345         if (ret)
3346                 return;
3347
3348         perf_output_put(&handle, read_event);
3349         perf_output_read(&handle, event);
3350
3351         perf_output_end(&handle);
3352 }
3353
3354 /*
3355  * task tracking -- fork/exit
3356  *
3357  * enabled by: attr.comm | attr.mmap | attr.task
3358  */
3359
3360 struct perf_task_event {
3361         struct task_struct              *task;
3362         struct perf_event_context       *task_ctx;
3363
3364         struct {
3365                 struct perf_event_header        header;
3366
3367                 u32                             pid;
3368                 u32                             ppid;
3369                 u32                             tid;
3370                 u32                             ptid;
3371                 u64                             time;
3372         } event_id;
3373 };
3374
3375 static void perf_event_task_output(struct perf_event *event,
3376                                      struct perf_task_event *task_event)
3377 {
3378         struct perf_output_handle handle;
3379         int size;
3380         struct task_struct *task = task_event->task;
3381         int ret;
3382
3383         size  = task_event->event_id.header.size;
3384         ret = perf_output_begin(&handle, event, size, 0, 0);
3385
3386         if (ret)
3387                 return;
3388
3389         task_event->event_id.pid = perf_event_pid(event, task);
3390         task_event->event_id.ppid = perf_event_pid(event, current);
3391
3392         task_event->event_id.tid = perf_event_tid(event, task);
3393         task_event->event_id.ptid = perf_event_tid(event, current);
3394
3395         perf_output_put(&handle, task_event->event_id);
3396
3397         perf_output_end(&handle);
3398 }
3399
3400 static int perf_event_task_match(struct perf_event *event)
3401 {
3402         if (event->state < PERF_EVENT_STATE_INACTIVE)
3403                 return 0;
3404
3405         if (event->cpu != -1 && event->cpu != smp_processor_id())
3406                 return 0;
3407
3408         if (event->attr.comm || event->attr.mmap || event->attr.task)
3409                 return 1;
3410
3411         return 0;
3412 }
3413
3414 static void perf_event_task_ctx(struct perf_event_context *ctx,
3415                                   struct perf_task_event *task_event)
3416 {
3417         struct perf_event *event;
3418
3419         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3420                 if (perf_event_task_match(event))
3421                         perf_event_task_output(event, task_event);
3422         }
3423 }
3424
3425 static void perf_event_task_event(struct perf_task_event *task_event)
3426 {
3427         struct perf_cpu_context *cpuctx;
3428         struct perf_event_context *ctx = task_event->task_ctx;
3429
3430         rcu_read_lock();
3431         cpuctx = &get_cpu_var(perf_cpu_context);
3432         perf_event_task_ctx(&cpuctx->ctx, task_event);
3433         if (!ctx)
3434                 ctx = rcu_dereference(current->perf_event_ctxp);
3435         if (ctx)
3436                 perf_event_task_ctx(ctx, task_event);
3437         put_cpu_var(perf_cpu_context);
3438         rcu_read_unlock();
3439 }
3440
3441 static void perf_event_task(struct task_struct *task,
3442                               struct perf_event_context *task_ctx,
3443                               int new)
3444 {
3445         struct perf_task_event task_event;
3446
3447         if (!atomic_read(&nr_comm_events) &&
3448             !atomic_read(&nr_mmap_events) &&
3449             !atomic_read(&nr_task_events))
3450                 return;
3451
3452         task_event = (struct perf_task_event){
3453                 .task     = task,
3454                 .task_ctx = task_ctx,
3455                 .event_id    = {
3456                         .header = {
3457                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3458                                 .misc = 0,
3459                                 .size = sizeof(task_event.event_id),
3460                         },
3461                         /* .pid  */
3462                         /* .ppid */
3463                         /* .tid  */
3464                         /* .ptid */
3465                         .time = perf_clock(),
3466                 },
3467         };
3468
3469         perf_event_task_event(&task_event);
3470 }
3471
3472 void perf_event_fork(struct task_struct *task)
3473 {
3474         perf_event_task(task, NULL, 1);
3475 }
3476
3477 /*
3478  * comm tracking
3479  */
3480
3481 struct perf_comm_event {
3482         struct task_struct      *task;
3483         char                    *comm;
3484         int                     comm_size;
3485
3486         struct {
3487                 struct perf_event_header        header;
3488
3489                 u32                             pid;
3490                 u32                             tid;
3491         } event_id;
3492 };
3493
3494 static void perf_event_comm_output(struct perf_event *event,
3495                                      struct perf_comm_event *comm_event)
3496 {
3497         struct perf_output_handle handle;
3498         int size = comm_event->event_id.header.size;
3499         int ret = perf_output_begin(&handle, event, size, 0, 0);
3500
3501         if (ret)
3502                 return;
3503
3504         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3505         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3506
3507         perf_output_put(&handle, comm_event->event_id);
3508         perf_output_copy(&handle, comm_event->comm,
3509                                    comm_event->comm_size);
3510         perf_output_end(&handle);
3511 }
3512
3513 static int perf_event_comm_match(struct perf_event *event)
3514 {
3515         if (event->state < PERF_EVENT_STATE_INACTIVE)
3516                 return 0;
3517
3518         if (event->cpu != -1 && event->cpu != smp_processor_id())
3519                 return 0;
3520
3521         if (event->attr.comm)
3522                 return 1;
3523
3524         return 0;
3525 }
3526
3527 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3528                                   struct perf_comm_event *comm_event)
3529 {
3530         struct perf_event *event;
3531
3532         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3533                 if (perf_event_comm_match(event))
3534                         perf_event_comm_output(event, comm_event);
3535         }
3536 }
3537
3538 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3539 {
3540         struct perf_cpu_context *cpuctx;
3541         struct perf_event_context *ctx;
3542         unsigned int size;
3543         char comm[TASK_COMM_LEN];
3544
3545         memset(comm, 0, sizeof(comm));
3546         strlcpy(comm, comm_event->task->comm, sizeof(comm));
3547         size = ALIGN(strlen(comm)+1, sizeof(u64));
3548
3549         comm_event->comm = comm;
3550         comm_event->comm_size = size;
3551
3552         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3553
3554         rcu_read_lock();
3555         cpuctx = &get_cpu_var(perf_cpu_context);
3556         perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3557         ctx = rcu_dereference(current->perf_event_ctxp);
3558         if (ctx)
3559                 perf_event_comm_ctx(ctx, comm_event);
3560         put_cpu_var(perf_cpu_context);
3561         rcu_read_unlock();
3562 }
3563
3564 void perf_event_comm(struct task_struct *task)
3565 {
3566         struct perf_comm_event comm_event;
3567
3568         if (task->perf_event_ctxp)
3569                 perf_event_enable_on_exec(task);
3570
3571         if (!atomic_read(&nr_comm_events))
3572                 return;
3573
3574         comm_event = (struct perf_comm_event){
3575                 .task   = task,
3576                 /* .comm      */
3577                 /* .comm_size */
3578                 .event_id  = {
3579                         .header = {
3580                                 .type = PERF_RECORD_COMM,
3581                                 .misc = 0,
3582                                 /* .size */
3583                         },
3584                         /* .pid */
3585                         /* .tid */
3586                 },
3587         };
3588
3589         perf_event_comm_event(&comm_event);
3590 }
3591
3592 /*
3593  * mmap tracking
3594  */
3595
3596 struct perf_mmap_event {
3597         struct vm_area_struct   *vma;
3598
3599         const char              *file_name;
3600         int                     file_size;
3601
3602         struct {
3603                 struct perf_event_header        header;
3604
3605                 u32                             pid;
3606                 u32                             tid;
3607                 u64                             start;
3608                 u64                             len;
3609                 u64                             pgoff;
3610         } event_id;
3611 };
3612
3613 static void perf_event_mmap_output(struct perf_event *event,
3614                                      struct perf_mmap_event *mmap_event)
3615 {
3616         struct perf_output_handle handle;
3617         int size = mmap_event->event_id.header.size;
3618         int ret = perf_output_begin(&handle, event, size, 0, 0);
3619
3620         if (ret)
3621                 return;
3622
3623         mmap_event->event_id.pid = perf_event_pid(event, current);
3624         mmap_event->event_id.tid = perf_event_tid(event, current);
3625
3626         perf_output_put(&handle, mmap_event->event_id);
3627         perf_output_copy(&handle, mmap_event->file_name,
3628                                    mmap_event->file_size);
3629         perf_output_end(&handle);
3630 }
3631
3632 static int perf_event_mmap_match(struct perf_event *event,
3633                                    struct perf_mmap_event *mmap_event)
3634 {
3635         if (event->state < PERF_EVENT_STATE_INACTIVE)
3636                 return 0;
3637
3638         if (event->cpu != -1 && event->cpu != smp_processor_id())
3639                 return 0;
3640
3641         if (event->attr.mmap)
3642                 return 1;
3643
3644         return 0;
3645 }
3646
3647 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3648                                   struct perf_mmap_event *mmap_event)
3649 {
3650         struct perf_event *event;
3651
3652         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3653                 if (perf_event_mmap_match(event, mmap_event))
3654                         perf_event_mmap_output(event, mmap_event);
3655         }
3656 }
3657
3658 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3659 {
3660         struct perf_cpu_context *cpuctx;
3661         struct perf_event_context *ctx;
3662         struct vm_area_struct *vma = mmap_event->vma;
3663         struct file *file = vma->vm_file;
3664         unsigned int size;
3665         char tmp[16];
3666         char *buf = NULL;
3667         const char *name;
3668
3669         memset(tmp, 0, sizeof(tmp));
3670
3671         if (file) {
3672                 /*
3673                  * d_path works from the end of the buffer backwards, so we
3674                  * need to add enough zero bytes after the string to handle
3675                  * the 64bit alignment we do later.
3676                  */
3677                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3678                 if (!buf) {
3679                         name = strncpy(tmp, "//enomem", sizeof(tmp));
3680                         goto got_name;
3681                 }
3682                 name = d_path(&file->f_path, buf, PATH_MAX);
3683                 if (IS_ERR(name)) {
3684                         name = strncpy(tmp, "//toolong", sizeof(tmp));
3685                         goto got_name;
3686                 }
3687         } else {
3688                 if (arch_vma_name(mmap_event->vma)) {
3689                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3690                                        sizeof(tmp));
3691                         goto got_name;
3692                 }
3693
3694                 if (!vma->vm_mm) {
3695                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
3696                         goto got_name;
3697                 }
3698
3699                 name = strncpy(tmp, "//anon", sizeof(tmp));
3700                 goto got_name;
3701         }
3702
3703 got_name:
3704         size = ALIGN(strlen(name)+1, sizeof(u64));
3705
3706         mmap_event->file_name = name;
3707         mmap_event->file_size = size;
3708
3709         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3710
3711         rcu_read_lock();
3712         cpuctx = &get_cpu_var(perf_cpu_context);
3713         perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3714         ctx = rcu_dereference(current->perf_event_ctxp);
3715         if (ctx)
3716                 perf_event_mmap_ctx(ctx, mmap_event);
3717         put_cpu_var(perf_cpu_context);
3718         rcu_read_unlock();
3719
3720         kfree(buf);
3721 }
3722
3723 void __perf_event_mmap(struct vm_area_struct *vma)
3724 {
3725         struct perf_mmap_event mmap_event;
3726
3727         if (!atomic_read(&nr_mmap_events))
3728                 return;
3729
3730         mmap_event = (struct perf_mmap_event){
3731                 .vma    = vma,
3732                 /* .file_name */
3733                 /* .file_size */
3734                 .event_id  = {
3735                         .header = {
3736                                 .type = PERF_RECORD_MMAP,
3737                                 .misc = 0,
3738                                 /* .size */
3739                         },
3740                         /* .pid */
3741                         /* .tid */
3742                         .start  = vma->vm_start,
3743                         .len    = vma->vm_end - vma->vm_start,
3744                         .pgoff  = (u64)vma->vm_pgoff << PAGE_SHIFT,
3745                 },
3746         };
3747
3748         perf_event_mmap_event(&mmap_event);
3749 }
3750
3751 /*
3752  * IRQ throttle logging
3753  */
3754
3755 static void perf_log_throttle(struct perf_event *event, int enable)
3756 {
3757         struct perf_output_handle handle;
3758         int ret;
3759
3760         struct {
3761                 struct perf_event_header        header;
3762                 u64                             time;
3763                 u64                             id;
3764                 u64                             stream_id;
3765         } throttle_event = {
3766                 .header = {
3767                         .type = PERF_RECORD_THROTTLE,
3768                         .misc = 0,
3769                         .size = sizeof(throttle_event),
3770                 },
3771                 .time           = perf_clock(),
3772                 .id             = primary_event_id(event),
3773                 .stream_id      = event->id,
3774         };
3775
3776         if (enable)
3777                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3778
3779         ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3780         if (ret)
3781                 return;
3782
3783         perf_output_put(&handle, throttle_event);
3784         perf_output_end(&handle);
3785 }
3786
3787 /*
3788  * Generic event overflow handling, sampling.
3789  */
3790
3791 static int __perf_event_overflow(struct perf_event *event, int nmi,
3792                                    int throttle, struct perf_sample_data *data,
3793                                    struct pt_regs *regs)
3794 {
3795         int events = atomic_read(&event->event_limit);
3796         struct hw_perf_event *hwc = &event->hw;
3797         int ret = 0;
3798
3799         throttle = (throttle && event->pmu->unthrottle != NULL);
3800
3801         if (!throttle) {
3802                 hwc->interrupts++;
3803         } else {
3804                 if (hwc->interrupts != MAX_INTERRUPTS) {
3805                         hwc->interrupts++;
3806                         if (HZ * hwc->interrupts >
3807                                         (u64)sysctl_perf_event_sample_rate) {
3808                                 hwc->interrupts = MAX_INTERRUPTS;
3809                                 perf_log_throttle(event, 0);
3810                                 ret = 1;
3811                         }
3812                 } else {
3813                         /*
3814                          * Keep re-disabling events even though on the previous
3815                          * pass we disabled it - just in case we raced with a
3816                          * sched-in and the event got enabled again:
3817                          */
3818                         ret = 1;
3819                 }
3820         }
3821
3822         if (event->attr.freq) {
3823                 u64 now = perf_clock();
3824                 s64 delta = now - hwc->freq_time_stamp;
3825
3826                 hwc->freq_time_stamp = now;
3827
3828                 if (delta > 0 && delta < 2*TICK_NSEC)
3829                         perf_adjust_period(event, delta, hwc->last_period);
3830         }
3831
3832         /*
3833          * XXX event_limit might not quite work as expected on inherited
3834          * events
3835          */
3836
3837         event->pending_kill = POLL_IN;
3838         if (events && atomic_dec_and_test(&event->event_limit)) {
3839                 ret = 1;
3840                 event->pending_kill = POLL_HUP;
3841                 if (nmi) {
3842                         event->pending_disable = 1;
3843                         perf_pending_queue(&event->pending,
3844                                            perf_pending_event);
3845                 } else
3846                         perf_event_disable(event);
3847         }
3848
3849         if (event->overflow_handler)
3850                 event->overflow_handler(event, nmi, data, regs);
3851         else
3852                 perf_event_output(event, nmi, data, regs);
3853
3854         return ret;
3855 }
3856
3857 int perf_event_overflow(struct perf_event *event, int nmi,
3858                           struct perf_sample_data *data,
3859                           struct pt_regs *regs)
3860 {
3861         return __perf_event_overflow(event, nmi, 1, data, regs);
3862 }
3863
3864 /*
3865  * Generic software event infrastructure
3866  */
3867
3868 /*
3869  * We directly increment event->count and keep a second value in
3870  * event->hw.period_left to count intervals. This period event
3871  * is kept in the range [-sample_period, 0] so that we can use the
3872  * sign as trigger.
3873  */
3874
3875 static u64 perf_swevent_set_period(struct perf_event *event)
3876 {
3877         struct hw_perf_event *hwc = &event->hw;
3878         u64 period = hwc->last_period;
3879         u64 nr, offset;
3880         s64 old, val;
3881
3882         hwc->last_period = hwc->sample_period;
3883
3884 again:
3885         old = val = atomic64_read(&hwc->period_left);
3886         if (val < 0)
3887                 return 0;
3888
3889         nr = div64_u64(period + val, period);
3890         offset = nr * period;
3891         val -= offset;
3892         if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3893                 goto again;
3894
3895         return nr;
3896 }
3897
3898 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3899                                     int nmi, struct perf_sample_data *data,
3900                                     struct pt_regs *regs)
3901 {
3902         struct hw_perf_event *hwc = &event->hw;
3903         int throttle = 0;
3904
3905         data->period = event->hw.last_period;
3906         if (!overflow)
3907                 overflow = perf_swevent_set_period(event);
3908
3909         if (hwc->interrupts == MAX_INTERRUPTS)
3910                 return;
3911
3912         for (; overflow; overflow--) {
3913                 if (__perf_event_overflow(event, nmi, throttle,
3914                                             data, regs)) {
3915                         /*
3916                          * We inhibit the overflow from happening when
3917                          * hwc->interrupts == MAX_INTERRUPTS.
3918                          */
3919                         break;
3920                 }
3921                 throttle = 1;
3922         }
3923 }
3924
3925 static void perf_swevent_unthrottle(struct perf_event *event)
3926 {
3927         /*
3928          * Nothing to do, we already reset hwc->interrupts.
3929          */
3930 }
3931
3932 static void perf_swevent_add(struct perf_event *event, u64 nr,
3933                                int nmi, struct perf_sample_data *data,
3934                                struct pt_regs *regs)
3935 {
3936         struct hw_perf_event *hwc = &event->hw;
3937
3938         atomic64_add(nr, &event->count);
3939
3940         if (!regs)
3941                 return;
3942
3943         if (!hwc->sample_period)
3944                 return;
3945
3946         if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3947                 return perf_swevent_overflow(event, 1, nmi, data, regs);
3948
3949         if (atomic64_add_negative(nr, &hwc->period_left))
3950                 return;
3951
3952         perf_swevent_overflow(event, 0, nmi, data, regs);
3953 }
3954
3955 static int perf_swevent_is_counting(struct perf_event *event)
3956 {
3957         /*
3958          * The event is active, we're good!
3959          */
3960         if (event->state == PERF_EVENT_STATE_ACTIVE)
3961                 return 1;
3962
3963         /*
3964          * The event is off/error, not counting.
3965          */
3966         if (event->state != PERF_EVENT_STATE_INACTIVE)
3967                 return 0;
3968
3969         /*
3970          * The event is inactive, if the context is active
3971          * we're part of a group that didn't make it on the 'pmu',
3972          * not counting.
3973          */
3974         if (event->ctx->is_active)
3975                 return 0;
3976
3977         /*
3978          * We're inactive and the context is too, this means the
3979          * task is scheduled out, we're counting events that happen
3980          * to us, like migration events.
3981          */
3982         return 1;
3983 }
3984
3985 static int perf_tp_event_match(struct perf_event *event,
3986                                 struct perf_sample_data *data);
3987
3988 static int perf_exclude_event(struct perf_event *event,
3989                               struct pt_regs *regs)
3990 {
3991         if (regs) {
3992                 if (event->attr.exclude_user && user_mode(regs))
3993                         return 1;
3994
3995                 if (event->attr.exclude_kernel && !user_mode(regs))
3996                         return 1;
3997         }
3998
3999         return 0;
4000 }
4001
4002 static int perf_swevent_match(struct perf_event *event,
4003                                 enum perf_type_id type,
4004                                 u32 event_id,
4005                                 struct perf_sample_data *data,
4006                                 struct pt_regs *regs)
4007 {
4008         if (event->cpu != -1 && event->cpu != smp_processor_id())
4009                 return 0;
4010
4011         if (!perf_swevent_is_counting(event))
4012                 return 0;
4013
4014         if (event->attr.type != type)
4015                 return 0;
4016
4017         if (event->attr.config != event_id)
4018                 return 0;
4019
4020         if (perf_exclude_event(event, regs))
4021                 return 0;
4022
4023         if (event->attr.type == PERF_TYPE_TRACEPOINT &&
4024             !perf_tp_event_match(event, data))
4025                 return 0;
4026
4027         return 1;
4028 }
4029
4030 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
4031                                      enum perf_type_id type,
4032                                      u32 event_id, u64 nr, int nmi,
4033                                      struct perf_sample_data *data,
4034                                      struct pt_regs *regs)
4035 {
4036         struct perf_event *event;
4037
4038         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4039                 if (perf_swevent_match(event, type, event_id, data, regs))
4040                         perf_swevent_add(event, nr, nmi, data, regs);
4041         }
4042 }
4043
4044 int perf_swevent_get_recursion_context(void)
4045 {
4046         struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
4047         int rctx;
4048
4049         if (in_nmi())
4050                 rctx = 3;
4051         else if (in_irq())
4052                 rctx = 2;
4053         else if (in_softirq())
4054                 rctx = 1;
4055         else
4056                 rctx = 0;
4057
4058         if (cpuctx->recursion[rctx]) {
4059                 put_cpu_var(perf_cpu_context);
4060                 return -1;
4061         }
4062
4063         cpuctx->recursion[rctx]++;
4064         barrier();
4065
4066         return rctx;
4067 }
4068 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4069
4070 void perf_swevent_put_recursion_context(int rctx)
4071 {
4072         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4073         barrier();
4074         cpuctx->recursion[rctx]--;
4075         put_cpu_var(perf_cpu_context);
4076 }
4077 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
4078
4079 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4080                                     u64 nr, int nmi,
4081                                     struct perf_sample_data *data,
4082                                     struct pt_regs *regs)
4083 {
4084         struct perf_cpu_context *cpuctx;
4085         struct perf_event_context *ctx;
4086
4087         cpuctx = &__get_cpu_var(perf_cpu_context);
4088         rcu_read_lock();
4089         perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
4090                                  nr, nmi, data, regs);
4091         /*
4092          * doesn't really matter which of the child contexts the
4093          * events ends up in.
4094          */
4095         ctx = rcu_dereference(current->perf_event_ctxp);
4096         if (ctx)
4097                 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
4098         rcu_read_unlock();
4099 }
4100
4101 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4102                             struct pt_regs *regs, u64 addr)
4103 {
4104         struct perf_sample_data data;
4105         int rctx;
4106
4107         rctx = perf_swevent_get_recursion_context();
4108         if (rctx < 0)
4109                 return;
4110
4111         data.addr = addr;
4112         data.raw  = NULL;
4113
4114         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4115
4116         perf_swevent_put_recursion_context(rctx);
4117 }
4118
4119 static void perf_swevent_read(struct perf_event *event)
4120 {
4121 }
4122
4123 static int perf_swevent_enable(struct perf_event *event)
4124 {
4125         struct hw_perf_event *hwc = &event->hw;
4126
4127         if (hwc->sample_period) {
4128                 hwc->last_period = hwc->sample_period;
4129                 perf_swevent_set_period(event);
4130         }
4131         return 0;
4132 }
4133
4134 static void perf_swevent_disable(struct perf_event *event)
4135 {
4136 }
4137
4138 static const struct pmu perf_ops_generic = {
4139         .enable         = perf_swevent_enable,
4140         .disable        = perf_swevent_disable,
4141         .read           = perf_swevent_read,
4142         .unthrottle     = perf_swevent_unthrottle,
4143 };
4144
4145 /*
4146  * hrtimer based swevent callback
4147  */
4148
4149 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4150 {
4151         enum hrtimer_restart ret = HRTIMER_RESTART;
4152         struct perf_sample_data data;
4153         struct pt_regs *regs;
4154         struct perf_event *event;
4155         u64 period;
4156
4157         event   = container_of(hrtimer, struct perf_event, hw.hrtimer);
4158         event->pmu->read(event);
4159
4160         data.addr = 0;
4161         data.raw = NULL;
4162         data.period = event->hw.last_period;
4163         regs = get_irq_regs();
4164         /*
4165          * In case we exclude kernel IPs or are somehow not in interrupt
4166          * context, provide the next best thing, the user IP.
4167          */
4168         if ((event->attr.exclude_kernel || !regs) &&
4169                         !event->attr.exclude_user)
4170                 regs = task_pt_regs(current);
4171
4172         if (regs) {
4173                 if (!(event->attr.exclude_idle && current->pid == 0))
4174                         if (perf_event_overflow(event, 0, &data, regs))
4175                                 ret = HRTIMER_NORESTART;
4176         }
4177
4178         period = max_t(u64, 10000, event->hw.sample_period);
4179         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4180
4181         return ret;
4182 }
4183
4184 static void perf_swevent_start_hrtimer(struct perf_event *event)
4185 {
4186         struct hw_perf_event *hwc = &event->hw;
4187
4188         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4189         hwc->hrtimer.function = perf_swevent_hrtimer;
4190         if (hwc->sample_period) {
4191                 u64 period;
4192
4193                 if (hwc->remaining) {
4194                         if (hwc->remaining < 0)
4195                                 period = 10000;
4196                         else
4197                                 period = hwc->remaining;
4198                         hwc->remaining = 0;
4199                 } else {
4200                         period = max_t(u64, 10000, hwc->sample_period);
4201                 }
4202                 __hrtimer_start_range_ns(&hwc->hrtimer,
4203                                 ns_to_ktime(period), 0,
4204                                 HRTIMER_MODE_REL, 0);
4205         }
4206 }
4207
4208 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4209 {
4210         struct hw_perf_event *hwc = &event->hw;
4211
4212         if (hwc->sample_period) {
4213                 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4214                 hwc->remaining = ktime_to_ns(remaining);
4215
4216                 hrtimer_cancel(&hwc->hrtimer);
4217         }
4218 }
4219
4220 /*
4221  * Software event: cpu wall time clock
4222  */
4223
4224 static void cpu_clock_perf_event_update(struct perf_event *event)
4225 {
4226         int cpu = raw_smp_processor_id();
4227         s64 prev;
4228         u64 now;
4229
4230         now = cpu_clock(cpu);
4231         prev = atomic64_xchg(&event->hw.prev_count, now);
4232         atomic64_add(now - prev, &event->count);
4233 }
4234
4235 static int cpu_clock_perf_event_enable(struct perf_event *event)
4236 {
4237         struct hw_perf_event *hwc = &event->hw;
4238         int cpu = raw_smp_processor_id();
4239
4240         atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4241         perf_swevent_start_hrtimer(event);
4242
4243         return 0;
4244 }
4245
4246 static void cpu_clock_perf_event_disable(struct perf_event *event)
4247 {
4248         perf_swevent_cancel_hrtimer(event);
4249         cpu_clock_perf_event_update(event);
4250 }
4251
4252 static void cpu_clock_perf_event_read(struct perf_event *event)
4253 {
4254         cpu_clock_perf_event_update(event);
4255 }
4256
4257 static const struct pmu perf_ops_cpu_clock = {
4258         .enable         = cpu_clock_perf_event_enable,
4259         .disable        = cpu_clock_perf_event_disable,
4260         .read           = cpu_clock_perf_event_read,
4261 };
4262
4263 /*
4264  * Software event: task time clock
4265  */
4266
4267 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4268 {
4269         u64 prev;
4270         s64 delta;
4271
4272         prev = atomic64_xchg(&event->hw.prev_count, now);
4273         delta = now - prev;
4274         atomic64_add(delta, &event->count);
4275 }
4276
4277 static int task_clock_perf_event_enable(struct perf_event *event)
4278 {
4279         struct hw_perf_event *hwc = &event->hw;
4280         u64 now;
4281
4282         now = event->ctx->time;
4283
4284         atomic64_set(&hwc->prev_count, now);
4285
4286         perf_swevent_start_hrtimer(event);
4287
4288         return 0;
4289 }
4290
4291 static void task_clock_perf_event_disable(struct perf_event *event)
4292 {
4293         perf_swevent_cancel_hrtimer(event);
4294         task_clock_perf_event_update(event, event->ctx->time);
4295
4296 }
4297
4298 static void task_clock_perf_event_read(struct perf_event *event)
4299 {
4300         u64 time;
4301
4302         if (!in_nmi()) {
4303                 update_context_time(event->ctx);
4304                 time = event->ctx->time;
4305         } else {
4306                 u64 now = perf_clock();
4307                 u64 delta = now - event->ctx->timestamp;
4308                 time = event->ctx->time + delta;
4309         }
4310
4311         task_clock_perf_event_update(event, time);
4312 }
4313
4314 static const struct pmu perf_ops_task_clock = {
4315         .enable         = task_clock_perf_event_enable,
4316         .disable        = task_clock_perf_event_disable,
4317         .read           = task_clock_perf_event_read,
4318 };
4319
4320 #ifdef CONFIG_EVENT_TRACING
4321
4322 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4323                           int entry_size)
4324 {
4325         struct perf_raw_record raw = {
4326                 .size = entry_size,
4327                 .data = record,
4328         };
4329
4330         struct perf_sample_data data = {
4331                 .addr = addr,
4332                 .raw = &raw,
4333         };
4334
4335         struct pt_regs *regs = get_irq_regs();
4336
4337         if (!regs)
4338                 regs = task_pt_regs(current);
4339
4340         /* Trace events already protected against recursion */
4341         do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4342                                 &data, regs);
4343 }
4344 EXPORT_SYMBOL_GPL(perf_tp_event);
4345
4346 static int perf_tp_event_match(struct perf_event *event,
4347                                 struct perf_sample_data *data)
4348 {
4349         void *record = data->raw->data;
4350
4351         if (likely(!event->filter) || filter_match_preds(event->filter, record))
4352                 return 1;
4353         return 0;
4354 }
4355
4356 static void tp_perf_event_destroy(struct perf_event *event)
4357 {
4358         ftrace_profile_disable(event->attr.config);
4359 }
4360
4361 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4362 {
4363         /*
4364          * Raw tracepoint data is a severe data leak, only allow root to
4365          * have these.
4366          */
4367         if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4368                         perf_paranoid_tracepoint_raw() &&
4369                         !capable(CAP_SYS_ADMIN))
4370                 return ERR_PTR(-EPERM);
4371
4372         if (ftrace_profile_enable(event->attr.config))
4373                 return NULL;
4374
4375         event->destroy = tp_perf_event_destroy;
4376
4377         return &perf_ops_generic;
4378 }
4379
4380 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4381 {
4382         char *filter_str;
4383         int ret;
4384
4385         if (event->attr.type != PERF_TYPE_TRACEPOINT)
4386                 return -EINVAL;
4387
4388         filter_str = strndup_user(arg, PAGE_SIZE);
4389         if (IS_ERR(filter_str))
4390                 return PTR_ERR(filter_str);
4391
4392         ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4393
4394         kfree(filter_str);
4395         return ret;
4396 }
4397
4398 static void perf_event_free_filter(struct perf_event *event)
4399 {
4400         ftrace_profile_free_filter(event);
4401 }
4402
4403 #else
4404
4405 static int perf_tp_event_match(struct perf_event *event,
4406                                 struct perf_sample_data *data)
4407 {
4408         return 1;
4409 }
4410
4411 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4412 {
4413         return NULL;
4414 }
4415
4416 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4417 {
4418         return -ENOENT;
4419 }
4420
4421 static void perf_event_free_filter(struct perf_event *event)
4422 {
4423 }
4424
4425 #endif /* CONFIG_EVENT_TRACING */
4426
4427 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4428 static void bp_perf_event_destroy(struct perf_event *event)
4429 {
4430         release_bp_slot(event);
4431 }
4432
4433 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4434 {
4435         int err;
4436
4437         err = register_perf_hw_breakpoint(bp);
4438         if (err)
4439                 return ERR_PTR(err);
4440
4441         bp->destroy = bp_perf_event_destroy;
4442
4443         return &perf_ops_bp;
4444 }
4445
4446 void perf_bp_event(struct perf_event *bp, void *data)
4447 {
4448         struct perf_sample_data sample;
4449         struct pt_regs *regs = data;
4450
4451         sample.raw = NULL;
4452         sample.addr = bp->attr.bp_addr;
4453
4454         if (!perf_exclude_event(bp, regs))
4455                 perf_swevent_add(bp, 1, 1, &sample, regs);
4456 }
4457 #else
4458 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4459 {
4460         return NULL;
4461 }
4462
4463 void perf_bp_event(struct perf_event *bp, void *regs)
4464 {
4465 }
4466 #endif
4467
4468 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4469
4470 static void sw_perf_event_destroy(struct perf_event *event)
4471 {
4472         u64 event_id = event->attr.config;
4473
4474         WARN_ON(event->parent);
4475
4476         atomic_dec(&perf_swevent_enabled[event_id]);
4477 }
4478
4479 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4480 {
4481         const struct pmu *pmu = NULL;
4482         u64 event_id = event->attr.config;
4483
4484         /*
4485          * Software events (currently) can't in general distinguish
4486          * between user, kernel and hypervisor events.
4487          * However, context switches and cpu migrations are considered
4488          * to be kernel events, and page faults are never hypervisor
4489          * events.
4490          */
4491         switch (event_id) {
4492         case PERF_COUNT_SW_CPU_CLOCK:
4493                 pmu = &perf_ops_cpu_clock;
4494
4495                 break;
4496         case PERF_COUNT_SW_TASK_CLOCK:
4497                 /*
4498                  * If the user instantiates this as a per-cpu event,
4499                  * use the cpu_clock event instead.
4500                  */
4501                 if (event->ctx->task)
4502                         pmu = &perf_ops_task_clock;
4503                 else
4504                         pmu = &perf_ops_cpu_clock;
4505
4506                 break;
4507         case PERF_COUNT_SW_PAGE_FAULTS:
4508         case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4509         case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4510         case PERF_COUNT_SW_CONTEXT_SWITCHES:
4511         case PERF_COUNT_SW_CPU_MIGRATIONS:
4512         case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4513         case PERF_COUNT_SW_EMULATION_FAULTS:
4514                 if (!event->parent) {
4515                         atomic_inc(&perf_swevent_enabled[event_id]);
4516                         event->destroy = sw_perf_event_destroy;
4517                 }
4518                 pmu = &perf_ops_generic;
4519                 break;
4520         }
4521
4522         return pmu;
4523 }
4524
4525 /*
4526  * Allocate and initialize a event structure
4527  */
4528 static struct perf_event *
4529 perf_event_alloc(struct perf_event_attr *attr,
4530                    int cpu,
4531                    struct perf_event_context *ctx,
4532                    struct perf_event *group_leader,
4533                    struct perf_event *parent_event,
4534                    perf_overflow_handler_t overflow_handler,
4535                    gfp_t gfpflags)
4536 {
4537         const struct pmu *pmu;
4538         struct perf_event *event;
4539         struct hw_perf_event *hwc;
4540         long err;
4541
4542         event = kzalloc(sizeof(*event), gfpflags);
4543         if (!event)
4544                 return ERR_PTR(-ENOMEM);
4545
4546         /*
4547          * Single events are their own group leaders, with an
4548          * empty sibling list:
4549          */
4550         if (!group_leader)
4551                 group_leader = event;
4552
4553         mutex_init(&event->child_mutex);
4554         INIT_LIST_HEAD(&event->child_list);
4555
4556         INIT_LIST_HEAD(&event->group_entry);
4557         INIT_LIST_HEAD(&event->event_entry);
4558         INIT_LIST_HEAD(&event->sibling_list);
4559         init_waitqueue_head(&event->waitq);
4560
4561         mutex_init(&event->mmap_mutex);
4562
4563         event->cpu              = cpu;
4564         event->attr             = *attr;
4565         event->group_leader     = group_leader;
4566         event->pmu              = NULL;
4567         event->ctx              = ctx;
4568         event->oncpu            = -1;
4569
4570         event->parent           = parent_event;
4571
4572         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
4573         event->id               = atomic64_inc_return(&perf_event_id);
4574
4575         event->state            = PERF_EVENT_STATE_INACTIVE;
4576
4577         if (!overflow_handler && parent_event)
4578                 overflow_handler = parent_event->overflow_handler;
4579         
4580         event->overflow_handler = overflow_handler;
4581
4582         if (attr->disabled)
4583                 event->state = PERF_EVENT_STATE_OFF;
4584
4585         pmu = NULL;
4586
4587         hwc = &event->hw;
4588         hwc->sample_period = attr->sample_period;
4589         if (attr->freq && attr->sample_freq)
4590                 hwc->sample_period = 1;
4591         hwc->last_period = hwc->sample_period;
4592
4593         atomic64_set(&hwc->period_left, hwc->sample_period);
4594
4595         /*
4596          * we currently do not support PERF_FORMAT_GROUP on inherited events
4597          */
4598         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4599                 goto done;
4600
4601         switch (attr->type) {
4602         case PERF_TYPE_RAW:
4603         case PERF_TYPE_HARDWARE:
4604         case PERF_TYPE_HW_CACHE:
4605                 pmu = hw_perf_event_init(event);
4606                 break;
4607
4608         case PERF_TYPE_SOFTWARE:
4609                 pmu = sw_perf_event_init(event);
4610                 break;
4611
4612         case PERF_TYPE_TRACEPOINT:
4613                 pmu = tp_perf_event_init(event);
4614                 break;
4615
4616         case PERF_TYPE_BREAKPOINT:
4617                 pmu = bp_perf_event_init(event);
4618                 break;
4619
4620
4621         default:
4622                 break;
4623         }
4624 done:
4625         err = 0;
4626         if (!pmu)
4627                 err = -EINVAL;
4628         else if (IS_ERR(pmu))
4629                 err = PTR_ERR(pmu);
4630
4631         if (err) {
4632                 if (event->ns)
4633                         put_pid_ns(event->ns);
4634                 kfree(event);
4635                 return ERR_PTR(err);
4636         }
4637
4638         event->pmu = pmu;
4639
4640         if (!event->parent) {
4641                 atomic_inc(&nr_events);
4642                 if (event->attr.mmap)
4643                         atomic_inc(&nr_mmap_events);
4644                 if (event->attr.comm)
4645                         atomic_inc(&nr_comm_events);
4646                 if (event->attr.task)
4647                         atomic_inc(&nr_task_events);
4648         }
4649
4650         return event;
4651 }
4652
4653 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4654                           struct perf_event_attr *attr)
4655 {
4656         u32 size;
4657         int ret;
4658
4659         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4660                 return -EFAULT;
4661
4662         /*
4663          * zero the full structure, so that a short copy will be nice.
4664          */
4665         memset(attr, 0, sizeof(*attr));
4666
4667         ret = get_user(size, &uattr->size);
4668         if (ret)
4669                 return ret;
4670
4671         if (size > PAGE_SIZE)   /* silly large */
4672                 goto err_size;
4673
4674         if (!size)              /* abi compat */
4675                 size = PERF_ATTR_SIZE_VER0;
4676
4677         if (size < PERF_ATTR_SIZE_VER0)
4678                 goto err_size;
4679
4680         /*
4681          * If we're handed a bigger struct than we know of,
4682          * ensure all the unknown bits are 0 - i.e. new
4683          * user-space does not rely on any kernel feature
4684          * extensions we dont know about yet.
4685          */
4686         if (size > sizeof(*attr)) {
4687                 unsigned char __user *addr;
4688                 unsigned char __user *end;
4689                 unsigned char val;
4690
4691                 addr = (void __user *)uattr + sizeof(*attr);
4692                 end  = (void __user *)uattr + size;
4693
4694                 for (; addr < end; addr++) {
4695                         ret = get_user(val, addr);
4696                         if (ret)
4697                                 return ret;
4698                         if (val)
4699                                 goto err_size;
4700                 }
4701                 size = sizeof(*attr);
4702         }
4703
4704         ret = copy_from_user(attr, uattr, size);
4705         if (ret)
4706                 return -EFAULT;
4707
4708         /*
4709          * If the type exists, the corresponding creation will verify
4710          * the attr->config.
4711          */
4712         if (attr->type >= PERF_TYPE_MAX)
4713                 return -EINVAL;
4714
4715         if (attr->__reserved_1)
4716                 return -EINVAL;
4717
4718         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4719                 return -EINVAL;
4720
4721         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4722                 return -EINVAL;
4723
4724 out:
4725         return ret;
4726
4727 err_size:
4728         put_user(sizeof(*attr), &uattr->size);
4729         ret = -E2BIG;
4730         goto out;
4731 }
4732
4733 static int perf_event_set_output(struct perf_event *event, int output_fd)
4734 {
4735         struct perf_event *output_event = NULL;
4736         struct file *output_file = NULL;
4737         struct perf_event *old_output;
4738         int fput_needed = 0;
4739         int ret = -EINVAL;
4740
4741         if (!output_fd)
4742                 goto set;
4743
4744         output_file = fget_light(output_fd, &fput_needed);
4745         if (!output_file)
4746                 return -EBADF;
4747
4748         if (output_file->f_op != &perf_fops)
4749                 goto out;
4750
4751         output_event = output_file->private_data;
4752
4753         /* Don't chain output fds */
4754         if (output_event->output)
4755                 goto out;
4756
4757         /* Don't set an output fd when we already have an output channel */
4758         if (event->data)
4759                 goto out;
4760
4761         atomic_long_inc(&output_file->f_count);
4762
4763 set:
4764         mutex_lock(&event->mmap_mutex);
4765         old_output = event->output;
4766         rcu_assign_pointer(event->output, output_event);
4767         mutex_unlock(&event->mmap_mutex);
4768
4769         if (old_output) {
4770                 /*
4771                  * we need to make sure no existing perf_output_*()
4772                  * is still referencing this event.
4773                  */
4774                 synchronize_rcu();
4775                 fput(old_output->filp);
4776         }
4777
4778         ret = 0;
4779 out:
4780         fput_light(output_file, fput_needed);
4781         return ret;
4782 }
4783
4784 /**
4785  * sys_perf_event_open - open a performance event, associate it to a task/cpu
4786  *
4787  * @attr_uptr:  event_id type attributes for monitoring/sampling
4788  * @pid:                target pid
4789  * @cpu:                target cpu
4790  * @group_fd:           group leader event fd
4791  */
4792 SYSCALL_DEFINE5(perf_event_open,
4793                 struct perf_event_attr __user *, attr_uptr,
4794                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4795 {
4796         struct perf_event *event, *group_leader;
4797         struct perf_event_attr attr;
4798         struct perf_event_context *ctx;
4799         struct file *event_file = NULL;
4800         struct file *group_file = NULL;
4801         int fput_needed = 0;
4802         int fput_needed2 = 0;
4803         int err;
4804
4805         /* for future expandability... */
4806         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4807                 return -EINVAL;
4808
4809         err = perf_copy_attr(attr_uptr, &attr);
4810         if (err)
4811                 return err;
4812
4813         if (!attr.exclude_kernel) {
4814                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4815                         return -EACCES;
4816         }
4817
4818         if (attr.freq) {
4819                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4820                         return -EINVAL;
4821         }
4822
4823         /*
4824          * Get the target context (task or percpu):
4825          */
4826         ctx = find_get_context(pid, cpu);
4827         if (IS_ERR(ctx))
4828                 return PTR_ERR(ctx);
4829
4830         /*
4831          * Look up the group leader (we will attach this event to it):
4832          */
4833         group_leader = NULL;
4834         if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4835                 err = -EINVAL;
4836                 group_file = fget_light(group_fd, &fput_needed);
4837                 if (!group_file)
4838                         goto err_put_context;
4839                 if (group_file->f_op != &perf_fops)
4840                         goto err_put_context;
4841
4842                 group_leader = group_file->private_data;
4843                 /*
4844                  * Do not allow a recursive hierarchy (this new sibling
4845                  * becoming part of another group-sibling):
4846                  */
4847                 if (group_leader->group_leader != group_leader)
4848                         goto err_put_context;
4849                 /*
4850                  * Do not allow to attach to a group in a different
4851                  * task or CPU context:
4852                  */
4853                 if (group_leader->ctx != ctx)
4854                         goto err_put_context;
4855                 /*
4856                  * Only a group leader can be exclusive or pinned
4857                  */
4858                 if (attr.exclusive || attr.pinned)
4859                         goto err_put_context;
4860         }
4861
4862         event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4863                                      NULL, NULL, GFP_KERNEL);
4864         err = PTR_ERR(event);
4865         if (IS_ERR(event))
4866                 goto err_put_context;
4867
4868         err = anon_inode_getfd("[perf_event]", &perf_fops, event, O_RDWR);
4869         if (err < 0)
4870                 goto err_free_put_context;
4871
4872         event_file = fget_light(err, &fput_needed2);
4873         if (!event_file)
4874                 goto err_free_put_context;
4875
4876         if (flags & PERF_FLAG_FD_OUTPUT) {
4877                 err = perf_event_set_output(event, group_fd);
4878                 if (err)
4879                         goto err_fput_free_put_context;
4880         }
4881
4882         event->filp = event_file;
4883         WARN_ON_ONCE(ctx->parent_ctx);
4884         mutex_lock(&ctx->mutex);
4885         perf_install_in_context(ctx, event, cpu);
4886         ++ctx->generation;
4887         mutex_unlock(&ctx->mutex);
4888
4889         event->owner = current;
4890         get_task_struct(current);
4891         mutex_lock(&current->perf_event_mutex);
4892         list_add_tail(&event->owner_entry, &current->perf_event_list);
4893         mutex_unlock(&current->perf_event_mutex);
4894
4895 err_fput_free_put_context:
4896         fput_light(event_file, fput_needed2);
4897
4898 err_free_put_context:
4899         if (err < 0)
4900                 kfree(event);
4901
4902 err_put_context:
4903         if (err < 0)
4904                 put_ctx(ctx);
4905
4906         fput_light(group_file, fput_needed);
4907
4908         return err;
4909 }
4910
4911 /**
4912  * perf_event_create_kernel_counter
4913  *
4914  * @attr: attributes of the counter to create
4915  * @cpu: cpu in which the counter is bound
4916  * @pid: task to profile
4917  */
4918 struct perf_event *
4919 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4920                                  pid_t pid,
4921                                  perf_overflow_handler_t overflow_handler)
4922 {
4923         struct perf_event *event;
4924         struct perf_event_context *ctx;
4925         int err;
4926
4927         /*
4928          * Get the target context (task or percpu):
4929          */
4930
4931         ctx = find_get_context(pid, cpu);
4932         if (IS_ERR(ctx)) {
4933                 err = PTR_ERR(ctx);
4934                 goto err_exit;
4935         }
4936
4937         event = perf_event_alloc(attr, cpu, ctx, NULL,
4938                                  NULL, overflow_handler, GFP_KERNEL);
4939         if (IS_ERR(event)) {
4940                 err = PTR_ERR(event);
4941                 goto err_put_context;
4942         }
4943
4944         event->filp = NULL;
4945         WARN_ON_ONCE(ctx->parent_ctx);
4946         mutex_lock(&ctx->mutex);
4947         perf_install_in_context(ctx, event, cpu);
4948         ++ctx->generation;
4949         mutex_unlock(&ctx->mutex);
4950
4951         event->owner = current;
4952         get_task_struct(current);
4953         mutex_lock(&current->perf_event_mutex);
4954         list_add_tail(&event->owner_entry, &current->perf_event_list);
4955         mutex_unlock(&current->perf_event_mutex);
4956
4957         return event;
4958
4959  err_put_context:
4960         put_ctx(ctx);
4961  err_exit:
4962         return ERR_PTR(err);
4963 }
4964 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4965
4966 /*
4967  * inherit a event from parent task to child task:
4968  */
4969 static struct perf_event *
4970 inherit_event(struct perf_event *parent_event,
4971               struct task_struct *parent,
4972               struct perf_event_context *parent_ctx,
4973               struct task_struct *child,
4974               struct perf_event *group_leader,
4975               struct perf_event_context *child_ctx)
4976 {
4977         struct perf_event *child_event;
4978
4979         /*
4980          * Instead of creating recursive hierarchies of events,
4981          * we link inherited events back to the original parent,
4982          * which has a filp for sure, which we use as the reference
4983          * count:
4984          */
4985         if (parent_event->parent)
4986                 parent_event = parent_event->parent;
4987
4988         child_event = perf_event_alloc(&parent_event->attr,
4989                                            parent_event->cpu, child_ctx,
4990                                            group_leader, parent_event,
4991                                            NULL, GFP_KERNEL);
4992         if (IS_ERR(child_event))
4993                 return child_event;
4994         get_ctx(child_ctx);
4995
4996         /*
4997          * Make the child state follow the state of the parent event,
4998          * not its attr.disabled bit.  We hold the parent's mutex,
4999          * so we won't race with perf_event_{en, dis}able_family.
5000          */
5001         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
5002                 child_event->state = PERF_EVENT_STATE_INACTIVE;
5003         else
5004                 child_event->state = PERF_EVENT_STATE_OFF;
5005
5006         if (parent_event->attr.freq) {
5007                 u64 sample_period = parent_event->hw.sample_period;
5008                 struct hw_perf_event *hwc = &child_event->hw;
5009
5010                 hwc->sample_period = sample_period;
5011                 hwc->last_period   = sample_period;
5012
5013                 atomic64_set(&hwc->period_left, sample_period);
5014         }
5015
5016         child_event->overflow_handler = parent_event->overflow_handler;
5017
5018         /*
5019          * Link it up in the child's context:
5020          */
5021         add_event_to_ctx(child_event, child_ctx);
5022
5023         /*
5024          * Get a reference to the parent filp - we will fput it
5025          * when the child event exits. This is safe to do because
5026          * we are in the parent and we know that the filp still
5027          * exists and has a nonzero count:
5028          */
5029         atomic_long_inc(&parent_event->filp->f_count);
5030
5031         /*
5032          * Link this into the parent event's child list
5033          */
5034         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5035         mutex_lock(&parent_event->child_mutex);
5036         list_add_tail(&child_event->child_list, &parent_event->child_list);
5037         mutex_unlock(&parent_event->child_mutex);
5038
5039         return child_event;
5040 }
5041
5042 static int inherit_group(struct perf_event *parent_event,
5043               struct task_struct *parent,
5044               struct perf_event_context *parent_ctx,
5045               struct task_struct *child,
5046               struct perf_event_context *child_ctx)
5047 {
5048         struct perf_event *leader;
5049         struct perf_event *sub;
5050         struct perf_event *child_ctr;
5051
5052         leader = inherit_event(parent_event, parent, parent_ctx,
5053                                  child, NULL, child_ctx);
5054         if (IS_ERR(leader))
5055                 return PTR_ERR(leader);
5056         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5057                 child_ctr = inherit_event(sub, parent, parent_ctx,
5058                                             child, leader, child_ctx);
5059                 if (IS_ERR(child_ctr))
5060                         return PTR_ERR(child_ctr);
5061         }
5062         return 0;
5063 }
5064
5065 static void sync_child_event(struct perf_event *child_event,
5066                                struct task_struct *child)
5067 {
5068         struct perf_event *parent_event = child_event->parent;
5069         u64 child_val;
5070
5071         if (child_event->attr.inherit_stat)
5072                 perf_event_read_event(child_event, child);
5073
5074         child_val = atomic64_read(&child_event->count);
5075
5076         /*
5077          * Add back the child's count to the parent's count:
5078          */
5079         atomic64_add(child_val, &parent_event->count);
5080         atomic64_add(child_event->total_time_enabled,
5081                      &parent_event->child_total_time_enabled);
5082         atomic64_add(child_event->total_time_running,
5083                      &parent_event->child_total_time_running);
5084
5085         /*
5086          * Remove this event from the parent's list
5087          */
5088         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5089         mutex_lock(&parent_event->child_mutex);
5090         list_del_init(&child_event->child_list);
5091         mutex_unlock(&parent_event->child_mutex);
5092
5093         /*
5094          * Release the parent event, if this was the last
5095          * reference to it.
5096          */
5097         fput(parent_event->filp);
5098 }
5099
5100 static void
5101 __perf_event_exit_task(struct perf_event *child_event,
5102                          struct perf_event_context *child_ctx,
5103                          struct task_struct *child)
5104 {
5105         struct perf_event *parent_event;
5106
5107         perf_event_remove_from_context(child_event);
5108
5109         parent_event = child_event->parent;
5110         /*
5111          * It can happen that parent exits first, and has events
5112          * that are still around due to the child reference. These
5113          * events need to be zapped - but otherwise linger.
5114          */
5115         if (parent_event) {
5116                 sync_child_event(child_event, child);
5117                 free_event(child_event);
5118         }
5119 }
5120
5121 /*
5122  * When a child task exits, feed back event values to parent events.
5123  */
5124 void perf_event_exit_task(struct task_struct *child)
5125 {
5126         struct perf_event *child_event, *tmp;
5127         struct perf_event_context *child_ctx;
5128         unsigned long flags;
5129
5130         if (likely(!child->perf_event_ctxp)) {
5131                 perf_event_task(child, NULL, 0);
5132                 return;
5133         }
5134
5135         local_irq_save(flags);
5136         /*
5137          * We can't reschedule here because interrupts are disabled,
5138          * and either child is current or it is a task that can't be
5139          * scheduled, so we are now safe from rescheduling changing
5140          * our context.
5141          */
5142         child_ctx = child->perf_event_ctxp;
5143         __perf_event_task_sched_out(child_ctx);
5144
5145         /*
5146          * Take the context lock here so that if find_get_context is
5147          * reading child->perf_event_ctxp, we wait until it has
5148          * incremented the context's refcount before we do put_ctx below.
5149          */
5150         raw_spin_lock(&child_ctx->lock);
5151         child->perf_event_ctxp = NULL;
5152         /*
5153          * If this context is a clone; unclone it so it can't get
5154          * swapped to another process while we're removing all
5155          * the events from it.
5156          */
5157         unclone_ctx(child_ctx);
5158         update_context_time(child_ctx);
5159         raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5160
5161         /*
5162          * Report the task dead after unscheduling the events so that we
5163          * won't get any samples after PERF_RECORD_EXIT. We can however still
5164          * get a few PERF_RECORD_READ events.
5165          */
5166         perf_event_task(child, child_ctx, 0);
5167
5168         /*
5169          * We can recurse on the same lock type through:
5170          *
5171          *   __perf_event_exit_task()
5172          *     sync_child_event()
5173          *       fput(parent_event->filp)
5174          *         perf_release()
5175          *           mutex_lock(&ctx->mutex)
5176          *
5177          * But since its the parent context it won't be the same instance.
5178          */
5179         mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
5180
5181 again:
5182         list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5183                                  group_entry)
5184                 __perf_event_exit_task(child_event, child_ctx, child);
5185
5186         list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5187                                  group_entry)
5188                 __perf_event_exit_task(child_event, child_ctx, child);
5189
5190         /*
5191          * If the last event was a group event, it will have appended all
5192          * its siblings to the list, but we obtained 'tmp' before that which
5193          * will still point to the list head terminating the iteration.
5194          */
5195         if (!list_empty(&child_ctx->pinned_groups) ||
5196             !list_empty(&child_ctx->flexible_groups))
5197                 goto again;
5198
5199         mutex_unlock(&child_ctx->mutex);
5200
5201         put_ctx(child_ctx);
5202 }
5203
5204 static void perf_free_event(struct perf_event *event,
5205                             struct perf_event_context *ctx)
5206 {
5207         struct perf_event *parent = event->parent;
5208
5209         if (WARN_ON_ONCE(!parent))
5210                 return;
5211
5212         mutex_lock(&parent->child_mutex);
5213         list_del_init(&event->child_list);
5214         mutex_unlock(&parent->child_mutex);
5215
5216         fput(parent->filp);
5217
5218         list_del_event(event, ctx);
5219         free_event(event);
5220 }
5221
5222 /*
5223  * free an unexposed, unused context as created by inheritance by
5224  * init_task below, used by fork() in case of fail.
5225  */
5226 void perf_event_free_task(struct task_struct *task)
5227 {
5228         struct perf_event_context *ctx = task->perf_event_ctxp;
5229         struct perf_event *event, *tmp;
5230
5231         if (!ctx)
5232                 return;
5233
5234         mutex_lock(&ctx->mutex);
5235 again:
5236         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5237                 perf_free_event(event, ctx);
5238
5239         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5240                                  group_entry)
5241                 perf_free_event(event, ctx);
5242
5243         if (!list_empty(&ctx->pinned_groups) ||
5244             !list_empty(&ctx->flexible_groups))
5245                 goto again;
5246
5247         mutex_unlock(&ctx->mutex);
5248
5249         put_ctx(ctx);
5250 }
5251
5252 static int
5253 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5254                    struct perf_event_context *parent_ctx,
5255                    struct task_struct *child,
5256                    int *inherited_all)
5257 {
5258         int ret;
5259         struct perf_event_context *child_ctx = child->perf_event_ctxp;
5260
5261         if (!event->attr.inherit) {
5262                 *inherited_all = 0;
5263                 return 0;
5264         }
5265
5266         if (!child_ctx) {
5267                 /*
5268                  * This is executed from the parent task context, so
5269                  * inherit events that have been marked for cloning.
5270                  * First allocate and initialize a context for the
5271                  * child.
5272                  */
5273
5274                 child_ctx = kzalloc(sizeof(struct perf_event_context),
5275                                     GFP_KERNEL);
5276                 if (!child_ctx)
5277                         return -ENOMEM;
5278
5279                 __perf_event_init_context(child_ctx, child);
5280                 child->perf_event_ctxp = child_ctx;
5281                 get_task_struct(child);
5282         }
5283
5284         ret = inherit_group(event, parent, parent_ctx,
5285                             child, child_ctx);
5286
5287         if (ret)
5288                 *inherited_all = 0;
5289
5290         return ret;
5291 }
5292
5293
5294 /*
5295  * Initialize the perf_event context in task_struct
5296  */
5297 int perf_event_init_task(struct task_struct *child)
5298 {
5299         struct perf_event_context *child_ctx, *parent_ctx;
5300         struct perf_event_context *cloned_ctx;
5301         struct perf_event *event;
5302         struct task_struct *parent = current;
5303         int inherited_all = 1;
5304         int ret = 0;
5305
5306         child->perf_event_ctxp = NULL;
5307
5308         mutex_init(&child->perf_event_mutex);
5309         INIT_LIST_HEAD(&child->perf_event_list);
5310
5311         if (likely(!parent->perf_event_ctxp))
5312                 return 0;
5313
5314         /*
5315          * If the parent's context is a clone, pin it so it won't get
5316          * swapped under us.
5317          */
5318         parent_ctx = perf_pin_task_context(parent);
5319
5320         /*
5321          * No need to check if parent_ctx != NULL here; since we saw
5322          * it non-NULL earlier, the only reason for it to become NULL
5323          * is if we exit, and since we're currently in the middle of
5324          * a fork we can't be exiting at the same time.
5325          */
5326
5327         /*
5328          * Lock the parent list. No need to lock the child - not PID
5329          * hashed yet and not running, so nobody can access it.
5330          */
5331         mutex_lock(&parent_ctx->mutex);
5332
5333         /*
5334          * We dont have to disable NMIs - we are only looking at
5335          * the list, not manipulating it:
5336          */
5337         list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5338                 ret = inherit_task_group(event, parent, parent_ctx, child,
5339                                          &inherited_all);
5340                 if (ret)
5341                         break;
5342         }
5343
5344         list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5345                 ret = inherit_task_group(event, parent, parent_ctx, child,
5346                                          &inherited_all);
5347                 if (ret)
5348                         break;
5349         }
5350
5351         child_ctx = child->perf_event_ctxp;
5352
5353         if (child_ctx && inherited_all) {
5354                 /*
5355                  * Mark the child context as a clone of the parent
5356                  * context, or of whatever the parent is a clone of.
5357                  * Note that if the parent is a clone, it could get
5358                  * uncloned at any point, but that doesn't matter
5359                  * because the list of events and the generation
5360                  * count can't have changed since we took the mutex.
5361                  */
5362                 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5363                 if (cloned_ctx) {
5364                         child_ctx->parent_ctx = cloned_ctx;
5365                         child_ctx->parent_gen = parent_ctx->parent_gen;
5366                 } else {
5367                         child_ctx->parent_ctx = parent_ctx;
5368                         child_ctx->parent_gen = parent_ctx->generation;
5369                 }
5370                 get_ctx(child_ctx->parent_ctx);
5371         }
5372
5373         mutex_unlock(&parent_ctx->mutex);
5374
5375         perf_unpin_context(parent_ctx);
5376
5377         return ret;
5378 }
5379
5380 static void __cpuinit perf_event_init_cpu(int cpu)
5381 {
5382         struct perf_cpu_context *cpuctx;
5383
5384         cpuctx = &per_cpu(perf_cpu_context, cpu);
5385         __perf_event_init_context(&cpuctx->ctx, NULL);
5386
5387         spin_lock(&perf_resource_lock);
5388         cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5389         spin_unlock(&perf_resource_lock);
5390
5391         hw_perf_event_setup(cpu);
5392 }
5393
5394 #ifdef CONFIG_HOTPLUG_CPU
5395 static void __perf_event_exit_cpu(void *info)
5396 {
5397         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5398         struct perf_event_context *ctx = &cpuctx->ctx;
5399         struct perf_event *event, *tmp;
5400
5401         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5402                 __perf_event_remove_from_context(event);
5403         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5404                 __perf_event_remove_from_context(event);
5405 }
5406 static void perf_event_exit_cpu(int cpu)
5407 {
5408         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5409         struct perf_event_context *ctx = &cpuctx->ctx;
5410
5411         mutex_lock(&ctx->mutex);
5412         smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5413         mutex_unlock(&ctx->mutex);
5414 }
5415 #else
5416 static inline void perf_event_exit_cpu(int cpu) { }
5417 #endif
5418
5419 static int __cpuinit
5420 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5421 {
5422         unsigned int cpu = (long)hcpu;
5423
5424         switch (action) {
5425
5426         case CPU_UP_PREPARE:
5427         case CPU_UP_PREPARE_FROZEN:
5428                 perf_event_init_cpu(cpu);
5429                 break;
5430
5431         case CPU_ONLINE:
5432         case CPU_ONLINE_FROZEN:
5433                 hw_perf_event_setup_online(cpu);
5434                 break;
5435
5436         case CPU_DOWN_PREPARE:
5437         case CPU_DOWN_PREPARE_FROZEN:
5438                 perf_event_exit_cpu(cpu);
5439                 break;
5440
5441         case CPU_DEAD:
5442                 hw_perf_event_setup_offline(cpu);
5443                 break;
5444
5445         default:
5446                 break;
5447         }
5448
5449         return NOTIFY_OK;
5450 }
5451
5452 /*
5453  * This has to have a higher priority than migration_notifier in sched.c.
5454  */
5455 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5456         .notifier_call          = perf_cpu_notify,
5457         .priority               = 20,
5458 };
5459
5460 void __init perf_event_init(void)
5461 {
5462         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5463                         (void *)(long)smp_processor_id());
5464         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5465                         (void *)(long)smp_processor_id());
5466         register_cpu_notifier(&perf_cpu_nb);
5467 }
5468
5469 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class,
5470                                         struct sysdev_class_attribute *attr,
5471                                         char *buf)
5472 {
5473         return sprintf(buf, "%d\n", perf_reserved_percpu);
5474 }
5475
5476 static ssize_t
5477 perf_set_reserve_percpu(struct sysdev_class *class,
5478                         struct sysdev_class_attribute *attr,
5479                         const char *buf,
5480                         size_t count)
5481 {
5482         struct perf_cpu_context *cpuctx;
5483         unsigned long val;
5484         int err, cpu, mpt;
5485
5486         err = strict_strtoul(buf, 10, &val);
5487         if (err)
5488                 return err;
5489         if (val > perf_max_events)
5490                 return -EINVAL;
5491
5492         spin_lock(&perf_resource_lock);
5493         perf_reserved_percpu = val;
5494         for_each_online_cpu(cpu) {
5495                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5496                 raw_spin_lock_irq(&cpuctx->ctx.lock);
5497                 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5498                           perf_max_events - perf_reserved_percpu);
5499                 cpuctx->max_pertask = mpt;
5500                 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5501         }
5502         spin_unlock(&perf_resource_lock);
5503
5504         return count;
5505 }
5506
5507 static ssize_t perf_show_overcommit(struct sysdev_class *class,
5508                                     struct sysdev_class_attribute *attr,
5509                                     char *buf)
5510 {
5511         return sprintf(buf, "%d\n", perf_overcommit);
5512 }
5513
5514 static ssize_t
5515 perf_set_overcommit(struct sysdev_class *class,
5516                     struct sysdev_class_attribute *attr,
5517                     const char *buf, size_t count)
5518 {
5519         unsigned long val;
5520         int err;
5521
5522         err = strict_strtoul(buf, 10, &val);
5523         if (err)
5524                 return err;
5525         if (val > 1)
5526                 return -EINVAL;
5527
5528         spin_lock(&perf_resource_lock);
5529         perf_overcommit = val;
5530         spin_unlock(&perf_resource_lock);
5531
5532         return count;
5533 }
5534
5535 static SYSDEV_CLASS_ATTR(
5536                                 reserve_percpu,
5537                                 0644,
5538                                 perf_show_reserve_percpu,
5539                                 perf_set_reserve_percpu
5540                         );
5541
5542 static SYSDEV_CLASS_ATTR(
5543                                 overcommit,
5544                                 0644,
5545                                 perf_show_overcommit,
5546                                 perf_set_overcommit
5547                         );
5548
5549 static struct attribute *perfclass_attrs[] = {
5550         &attr_reserve_percpu.attr,
5551         &attr_overcommit.attr,
5552         NULL
5553 };
5554
5555 static struct attribute_group perfclass_attr_group = {
5556         .attrs                  = perfclass_attrs,
5557         .name                   = "perf_events",
5558 };
5559
5560 static int __init perf_event_sysfs_init(void)
5561 {
5562         return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5563                                   &perfclass_attr_group);
5564 }
5565 device_initcall(perf_event_sysfs_init);