perf: Simplify __perf_event_sync_stat
[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 DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
40
41 int perf_max_events __read_mostly = 1;
42 static int perf_reserved_percpu __read_mostly;
43 static int perf_overcommit __read_mostly = 1;
44
45 static atomic_t nr_events __read_mostly;
46 static atomic_t nr_mmap_events __read_mostly;
47 static atomic_t nr_comm_events __read_mostly;
48 static atomic_t nr_task_events __read_mostly;
49
50 /*
51  * perf event paranoia level:
52  *  -1 - not paranoid at all
53  *   0 - disallow raw tracepoint access for unpriv
54  *   1 - disallow cpu events for unpriv
55  *   2 - disallow kernel profiling for unpriv
56  */
57 int sysctl_perf_event_paranoid __read_mostly = 1;
58
59 static inline bool perf_paranoid_tracepoint_raw(void)
60 {
61         return sysctl_perf_event_paranoid > -1;
62 }
63
64 static inline bool perf_paranoid_cpu(void)
65 {
66         return sysctl_perf_event_paranoid > 0;
67 }
68
69 static inline bool perf_paranoid_kernel(void)
70 {
71         return sysctl_perf_event_paranoid > 1;
72 }
73
74 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
75
76 /*
77  * max perf event sample rate
78  */
79 int sysctl_perf_event_sample_rate __read_mostly = 100000;
80
81 static atomic64_t perf_event_id;
82
83 /*
84  * Lock for (sysadmin-configurable) event reservations:
85  */
86 static DEFINE_SPINLOCK(perf_resource_lock);
87
88 /*
89  * Architecture provided APIs - weak aliases:
90  */
91 extern __weak const struct pmu *hw_perf_event_init(struct perf_event *event)
92 {
93         return NULL;
94 }
95
96 void __weak hw_perf_disable(void)               { barrier(); }
97 void __weak hw_perf_enable(void)                { barrier(); }
98
99 void __weak hw_perf_event_setup(int cpu)        { barrier(); }
100 void __weak hw_perf_event_setup_online(int cpu) { barrier(); }
101
102 int __weak
103 hw_perf_group_sched_in(struct perf_event *group_leader,
104                struct perf_cpu_context *cpuctx,
105                struct perf_event_context *ctx, int cpu)
106 {
107         return 0;
108 }
109
110 void __weak perf_event_print_debug(void)        { }
111
112 static DEFINE_PER_CPU(int, perf_disable_count);
113
114 void __perf_disable(void)
115 {
116         __get_cpu_var(perf_disable_count)++;
117 }
118
119 bool __perf_enable(void)
120 {
121         return !--__get_cpu_var(perf_disable_count);
122 }
123
124 void perf_disable(void)
125 {
126         __perf_disable();
127         hw_perf_disable();
128 }
129
130 void perf_enable(void)
131 {
132         if (__perf_enable())
133                 hw_perf_enable();
134 }
135
136 static void get_ctx(struct perf_event_context *ctx)
137 {
138         WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
139 }
140
141 static void free_ctx(struct rcu_head *head)
142 {
143         struct perf_event_context *ctx;
144
145         ctx = container_of(head, struct perf_event_context, rcu_head);
146         kfree(ctx);
147 }
148
149 static void put_ctx(struct perf_event_context *ctx)
150 {
151         if (atomic_dec_and_test(&ctx->refcount)) {
152                 if (ctx->parent_ctx)
153                         put_ctx(ctx->parent_ctx);
154                 if (ctx->task)
155                         put_task_struct(ctx->task);
156                 call_rcu(&ctx->rcu_head, free_ctx);
157         }
158 }
159
160 static void unclone_ctx(struct perf_event_context *ctx)
161 {
162         if (ctx->parent_ctx) {
163                 put_ctx(ctx->parent_ctx);
164                 ctx->parent_ctx = NULL;
165         }
166 }
167
168 /*
169  * If we inherit events we want to return the parent event id
170  * to userspace.
171  */
172 static u64 primary_event_id(struct perf_event *event)
173 {
174         u64 id = event->id;
175
176         if (event->parent)
177                 id = event->parent->id;
178
179         return id;
180 }
181
182 /*
183  * Get the perf_event_context for a task and lock it.
184  * This has to cope with with the fact that until it is locked,
185  * the context could get moved to another task.
186  */
187 static struct perf_event_context *
188 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
189 {
190         struct perf_event_context *ctx;
191
192         rcu_read_lock();
193  retry:
194         ctx = rcu_dereference(task->perf_event_ctxp);
195         if (ctx) {
196                 /*
197                  * If this context is a clone of another, it might
198                  * get swapped for another underneath us by
199                  * perf_event_task_sched_out, though the
200                  * rcu_read_lock() protects us from any context
201                  * getting freed.  Lock the context and check if it
202                  * got swapped before we could get the lock, and retry
203                  * if so.  If we locked the right context, then it
204                  * can't get swapped on us any more.
205                  */
206                 spin_lock_irqsave(&ctx->lock, *flags);
207                 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
208                         spin_unlock_irqrestore(&ctx->lock, *flags);
209                         goto retry;
210                 }
211
212                 if (!atomic_inc_not_zero(&ctx->refcount)) {
213                         spin_unlock_irqrestore(&ctx->lock, *flags);
214                         ctx = NULL;
215                 }
216         }
217         rcu_read_unlock();
218         return ctx;
219 }
220
221 /*
222  * Get the context for a task and increment its pin_count so it
223  * can't get swapped to another task.  This also increments its
224  * reference count so that the context can't get freed.
225  */
226 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
227 {
228         struct perf_event_context *ctx;
229         unsigned long flags;
230
231         ctx = perf_lock_task_context(task, &flags);
232         if (ctx) {
233                 ++ctx->pin_count;
234                 spin_unlock_irqrestore(&ctx->lock, flags);
235         }
236         return ctx;
237 }
238
239 static void perf_unpin_context(struct perf_event_context *ctx)
240 {
241         unsigned long flags;
242
243         spin_lock_irqsave(&ctx->lock, flags);
244         --ctx->pin_count;
245         spin_unlock_irqrestore(&ctx->lock, flags);
246         put_ctx(ctx);
247 }
248
249 /*
250  * Add a event from the lists for its context.
251  * Must be called with ctx->mutex and ctx->lock held.
252  */
253 static void
254 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
255 {
256         struct perf_event *group_leader = event->group_leader;
257
258         /*
259          * Depending on whether it is a standalone or sibling event,
260          * add it straight to the context's event list, or to the group
261          * leader's sibling list:
262          */
263         if (group_leader == event)
264                 list_add_tail(&event->group_entry, &ctx->group_list);
265         else {
266                 list_add_tail(&event->group_entry, &group_leader->sibling_list);
267                 group_leader->nr_siblings++;
268         }
269
270         list_add_rcu(&event->event_entry, &ctx->event_list);
271         ctx->nr_events++;
272         if (event->attr.inherit_stat)
273                 ctx->nr_stat++;
274 }
275
276 /*
277  * Remove a event from the lists for its context.
278  * Must be called with ctx->mutex and ctx->lock held.
279  */
280 static void
281 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
282 {
283         struct perf_event *sibling, *tmp;
284
285         if (list_empty(&event->group_entry))
286                 return;
287         ctx->nr_events--;
288         if (event->attr.inherit_stat)
289                 ctx->nr_stat--;
290
291         list_del_init(&event->group_entry);
292         list_del_rcu(&event->event_entry);
293
294         if (event->group_leader != event)
295                 event->group_leader->nr_siblings--;
296
297         /*
298          * If this was a group event with sibling events then
299          * upgrade the siblings to singleton events by adding them
300          * to the context list directly:
301          */
302         list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
303
304                 list_move_tail(&sibling->group_entry, &ctx->group_list);
305                 sibling->group_leader = sibling;
306         }
307 }
308
309 static void
310 event_sched_out(struct perf_event *event,
311                   struct perf_cpu_context *cpuctx,
312                   struct perf_event_context *ctx)
313 {
314         if (event->state != PERF_EVENT_STATE_ACTIVE)
315                 return;
316
317         event->state = PERF_EVENT_STATE_INACTIVE;
318         if (event->pending_disable) {
319                 event->pending_disable = 0;
320                 event->state = PERF_EVENT_STATE_OFF;
321         }
322         event->tstamp_stopped = ctx->time;
323         event->pmu->disable(event);
324         event->oncpu = -1;
325
326         if (!is_software_event(event))
327                 cpuctx->active_oncpu--;
328         ctx->nr_active--;
329         if (event->attr.exclusive || !cpuctx->active_oncpu)
330                 cpuctx->exclusive = 0;
331 }
332
333 static void
334 group_sched_out(struct perf_event *group_event,
335                 struct perf_cpu_context *cpuctx,
336                 struct perf_event_context *ctx)
337 {
338         struct perf_event *event;
339
340         if (group_event->state != PERF_EVENT_STATE_ACTIVE)
341                 return;
342
343         event_sched_out(group_event, cpuctx, ctx);
344
345         /*
346          * Schedule out siblings (if any):
347          */
348         list_for_each_entry(event, &group_event->sibling_list, group_entry)
349                 event_sched_out(event, cpuctx, ctx);
350
351         if (group_event->attr.exclusive)
352                 cpuctx->exclusive = 0;
353 }
354
355 /*
356  * Cross CPU call to remove a performance event
357  *
358  * We disable the event on the hardware level first. After that we
359  * remove it from the context list.
360  */
361 static void __perf_event_remove_from_context(void *info)
362 {
363         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
364         struct perf_event *event = info;
365         struct perf_event_context *ctx = event->ctx;
366
367         /*
368          * If this is a task context, we need to check whether it is
369          * the current task context of this cpu. If not it has been
370          * scheduled out before the smp call arrived.
371          */
372         if (ctx->task && cpuctx->task_ctx != ctx)
373                 return;
374
375         spin_lock(&ctx->lock);
376         /*
377          * Protect the list operation against NMI by disabling the
378          * events on a global level.
379          */
380         perf_disable();
381
382         event_sched_out(event, cpuctx, ctx);
383
384         list_del_event(event, ctx);
385
386         if (!ctx->task) {
387                 /*
388                  * Allow more per task events with respect to the
389                  * reservation:
390                  */
391                 cpuctx->max_pertask =
392                         min(perf_max_events - ctx->nr_events,
393                             perf_max_events - perf_reserved_percpu);
394         }
395
396         perf_enable();
397         spin_unlock(&ctx->lock);
398 }
399
400
401 /*
402  * Remove the event from a task's (or a CPU's) list of events.
403  *
404  * Must be called with ctx->mutex held.
405  *
406  * CPU events are removed with a smp call. For task events we only
407  * call when the task is on a CPU.
408  *
409  * If event->ctx is a cloned context, callers must make sure that
410  * every task struct that event->ctx->task could possibly point to
411  * remains valid.  This is OK when called from perf_release since
412  * that only calls us on the top-level context, which can't be a clone.
413  * When called from perf_event_exit_task, it's OK because the
414  * context has been detached from its task.
415  */
416 static void perf_event_remove_from_context(struct perf_event *event)
417 {
418         struct perf_event_context *ctx = event->ctx;
419         struct task_struct *task = ctx->task;
420
421         if (!task) {
422                 /*
423                  * Per cpu events are removed via an smp call and
424                  * the removal is always sucessful.
425                  */
426                 smp_call_function_single(event->cpu,
427                                          __perf_event_remove_from_context,
428                                          event, 1);
429                 return;
430         }
431
432 retry:
433         task_oncpu_function_call(task, __perf_event_remove_from_context,
434                                  event);
435
436         spin_lock_irq(&ctx->lock);
437         /*
438          * If the context is active we need to retry the smp call.
439          */
440         if (ctx->nr_active && !list_empty(&event->group_entry)) {
441                 spin_unlock_irq(&ctx->lock);
442                 goto retry;
443         }
444
445         /*
446          * The lock prevents that this context is scheduled in so we
447          * can remove the event safely, if the call above did not
448          * succeed.
449          */
450         if (!list_empty(&event->group_entry)) {
451                 list_del_event(event, ctx);
452         }
453         spin_unlock_irq(&ctx->lock);
454 }
455
456 static inline u64 perf_clock(void)
457 {
458         return cpu_clock(smp_processor_id());
459 }
460
461 /*
462  * Update the record of the current time in a context.
463  */
464 static void update_context_time(struct perf_event_context *ctx)
465 {
466         u64 now = perf_clock();
467
468         ctx->time += now - ctx->timestamp;
469         ctx->timestamp = now;
470 }
471
472 /*
473  * Update the total_time_enabled and total_time_running fields for a event.
474  */
475 static void update_event_times(struct perf_event *event)
476 {
477         struct perf_event_context *ctx = event->ctx;
478         u64 run_end;
479
480         if (event->state < PERF_EVENT_STATE_INACTIVE ||
481             event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
482                 return;
483
484         event->total_time_enabled = ctx->time - event->tstamp_enabled;
485
486         if (event->state == PERF_EVENT_STATE_INACTIVE)
487                 run_end = event->tstamp_stopped;
488         else
489                 run_end = ctx->time;
490
491         event->total_time_running = run_end - event->tstamp_running;
492 }
493
494 /*
495  * Update total_time_enabled and total_time_running for all events in a group.
496  */
497 static void update_group_times(struct perf_event *leader)
498 {
499         struct perf_event *event;
500
501         update_event_times(leader);
502         list_for_each_entry(event, &leader->sibling_list, group_entry)
503                 update_event_times(event);
504 }
505
506 /*
507  * Cross CPU call to disable a performance event
508  */
509 static void __perf_event_disable(void *info)
510 {
511         struct perf_event *event = info;
512         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
513         struct perf_event_context *ctx = event->ctx;
514
515         /*
516          * If this is a per-task event, need to check whether this
517          * event's task is the current task on this cpu.
518          */
519         if (ctx->task && cpuctx->task_ctx != ctx)
520                 return;
521
522         spin_lock(&ctx->lock);
523
524         /*
525          * If the event is on, turn it off.
526          * If it is in error state, leave it in error state.
527          */
528         if (event->state >= PERF_EVENT_STATE_INACTIVE) {
529                 update_context_time(ctx);
530                 update_group_times(event);
531                 if (event == event->group_leader)
532                         group_sched_out(event, cpuctx, ctx);
533                 else
534                         event_sched_out(event, cpuctx, ctx);
535                 event->state = PERF_EVENT_STATE_OFF;
536         }
537
538         spin_unlock(&ctx->lock);
539 }
540
541 /*
542  * Disable a event.
543  *
544  * If event->ctx is a cloned context, callers must make sure that
545  * every task struct that event->ctx->task could possibly point to
546  * remains valid.  This condition is satisifed when called through
547  * perf_event_for_each_child or perf_event_for_each because they
548  * hold the top-level event's child_mutex, so any descendant that
549  * goes to exit will block in sync_child_event.
550  * When called from perf_pending_event it's OK because event->ctx
551  * is the current context on this CPU and preemption is disabled,
552  * hence we can't get into perf_event_task_sched_out for this context.
553  */
554 static void perf_event_disable(struct perf_event *event)
555 {
556         struct perf_event_context *ctx = event->ctx;
557         struct task_struct *task = ctx->task;
558
559         if (!task) {
560                 /*
561                  * Disable the event on the cpu that it's on
562                  */
563                 smp_call_function_single(event->cpu, __perf_event_disable,
564                                          event, 1);
565                 return;
566         }
567
568  retry:
569         task_oncpu_function_call(task, __perf_event_disable, event);
570
571         spin_lock_irq(&ctx->lock);
572         /*
573          * If the event is still active, we need to retry the cross-call.
574          */
575         if (event->state == PERF_EVENT_STATE_ACTIVE) {
576                 spin_unlock_irq(&ctx->lock);
577                 goto retry;
578         }
579
580         /*
581          * Since we have the lock this context can't be scheduled
582          * in, so we can change the state safely.
583          */
584         if (event->state == PERF_EVENT_STATE_INACTIVE) {
585                 update_group_times(event);
586                 event->state = PERF_EVENT_STATE_OFF;
587         }
588
589         spin_unlock_irq(&ctx->lock);
590 }
591
592 static int
593 event_sched_in(struct perf_event *event,
594                  struct perf_cpu_context *cpuctx,
595                  struct perf_event_context *ctx,
596                  int cpu)
597 {
598         if (event->state <= PERF_EVENT_STATE_OFF)
599                 return 0;
600
601         event->state = PERF_EVENT_STATE_ACTIVE;
602         event->oncpu = cpu;     /* TODO: put 'cpu' into cpuctx->cpu */
603         /*
604          * The new state must be visible before we turn it on in the hardware:
605          */
606         smp_wmb();
607
608         if (event->pmu->enable(event)) {
609                 event->state = PERF_EVENT_STATE_INACTIVE;
610                 event->oncpu = -1;
611                 return -EAGAIN;
612         }
613
614         event->tstamp_running += ctx->time - event->tstamp_stopped;
615
616         if (!is_software_event(event))
617                 cpuctx->active_oncpu++;
618         ctx->nr_active++;
619
620         if (event->attr.exclusive)
621                 cpuctx->exclusive = 1;
622
623         return 0;
624 }
625
626 static int
627 group_sched_in(struct perf_event *group_event,
628                struct perf_cpu_context *cpuctx,
629                struct perf_event_context *ctx,
630                int cpu)
631 {
632         struct perf_event *event, *partial_group;
633         int ret;
634
635         if (group_event->state == PERF_EVENT_STATE_OFF)
636                 return 0;
637
638         ret = hw_perf_group_sched_in(group_event, cpuctx, ctx, cpu);
639         if (ret)
640                 return ret < 0 ? ret : 0;
641
642         if (event_sched_in(group_event, cpuctx, ctx, cpu))
643                 return -EAGAIN;
644
645         /*
646          * Schedule in siblings as one group (if any):
647          */
648         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
649                 if (event_sched_in(event, cpuctx, ctx, cpu)) {
650                         partial_group = event;
651                         goto group_error;
652                 }
653         }
654
655         return 0;
656
657 group_error:
658         /*
659          * Groups can be scheduled in as one unit only, so undo any
660          * partial group before returning:
661          */
662         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
663                 if (event == partial_group)
664                         break;
665                 event_sched_out(event, cpuctx, ctx);
666         }
667         event_sched_out(group_event, cpuctx, ctx);
668
669         return -EAGAIN;
670 }
671
672 /*
673  * Return 1 for a group consisting entirely of software events,
674  * 0 if the group contains any hardware events.
675  */
676 static int is_software_only_group(struct perf_event *leader)
677 {
678         struct perf_event *event;
679
680         if (!is_software_event(leader))
681                 return 0;
682
683         list_for_each_entry(event, &leader->sibling_list, group_entry)
684                 if (!is_software_event(event))
685                         return 0;
686
687         return 1;
688 }
689
690 /*
691  * Work out whether we can put this event group on the CPU now.
692  */
693 static int group_can_go_on(struct perf_event *event,
694                            struct perf_cpu_context *cpuctx,
695                            int can_add_hw)
696 {
697         /*
698          * Groups consisting entirely of software events can always go on.
699          */
700         if (is_software_only_group(event))
701                 return 1;
702         /*
703          * If an exclusive group is already on, no other hardware
704          * events can go on.
705          */
706         if (cpuctx->exclusive)
707                 return 0;
708         /*
709          * If this group is exclusive and there are already
710          * events on the CPU, it can't go on.
711          */
712         if (event->attr.exclusive && cpuctx->active_oncpu)
713                 return 0;
714         /*
715          * Otherwise, try to add it if all previous groups were able
716          * to go on.
717          */
718         return can_add_hw;
719 }
720
721 static void add_event_to_ctx(struct perf_event *event,
722                                struct perf_event_context *ctx)
723 {
724         list_add_event(event, ctx);
725         event->tstamp_enabled = ctx->time;
726         event->tstamp_running = ctx->time;
727         event->tstamp_stopped = ctx->time;
728 }
729
730 /*
731  * Cross CPU call to install and enable a performance event
732  *
733  * Must be called with ctx->mutex held
734  */
735 static void __perf_install_in_context(void *info)
736 {
737         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
738         struct perf_event *event = info;
739         struct perf_event_context *ctx = event->ctx;
740         struct perf_event *leader = event->group_leader;
741         int cpu = smp_processor_id();
742         int err;
743
744         /*
745          * If this is a task context, we need to check whether it is
746          * the current task context of this cpu. If not it has been
747          * scheduled out before the smp call arrived.
748          * Or possibly this is the right context but it isn't
749          * on this cpu because it had no events.
750          */
751         if (ctx->task && cpuctx->task_ctx != ctx) {
752                 if (cpuctx->task_ctx || ctx->task != current)
753                         return;
754                 cpuctx->task_ctx = ctx;
755         }
756
757         spin_lock(&ctx->lock);
758         ctx->is_active = 1;
759         update_context_time(ctx);
760
761         /*
762          * Protect the list operation against NMI by disabling the
763          * events on a global level. NOP for non NMI based events.
764          */
765         perf_disable();
766
767         add_event_to_ctx(event, ctx);
768
769         /*
770          * Don't put the event on if it is disabled or if
771          * it is in a group and the group isn't on.
772          */
773         if (event->state != PERF_EVENT_STATE_INACTIVE ||
774             (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
775                 goto unlock;
776
777         /*
778          * An exclusive event can't go on if there are already active
779          * hardware events, and no hardware event can go on if there
780          * is already an exclusive event on.
781          */
782         if (!group_can_go_on(event, cpuctx, 1))
783                 err = -EEXIST;
784         else
785                 err = event_sched_in(event, cpuctx, ctx, cpu);
786
787         if (err) {
788                 /*
789                  * This event couldn't go on.  If it is in a group
790                  * then we have to pull the whole group off.
791                  * If the event group is pinned then put it in error state.
792                  */
793                 if (leader != event)
794                         group_sched_out(leader, cpuctx, ctx);
795                 if (leader->attr.pinned) {
796                         update_group_times(leader);
797                         leader->state = PERF_EVENT_STATE_ERROR;
798                 }
799         }
800
801         if (!err && !ctx->task && cpuctx->max_pertask)
802                 cpuctx->max_pertask--;
803
804  unlock:
805         perf_enable();
806
807         spin_unlock(&ctx->lock);
808 }
809
810 /*
811  * Attach a performance event to a context
812  *
813  * First we add the event to the list with the hardware enable bit
814  * in event->hw_config cleared.
815  *
816  * If the event is attached to a task which is on a CPU we use a smp
817  * call to enable it in the task context. The task might have been
818  * scheduled away, but we check this in the smp call again.
819  *
820  * Must be called with ctx->mutex held.
821  */
822 static void
823 perf_install_in_context(struct perf_event_context *ctx,
824                         struct perf_event *event,
825                         int cpu)
826 {
827         struct task_struct *task = ctx->task;
828
829         if (!task) {
830                 /*
831                  * Per cpu events are installed via an smp call and
832                  * the install is always sucessful.
833                  */
834                 smp_call_function_single(cpu, __perf_install_in_context,
835                                          event, 1);
836                 return;
837         }
838
839 retry:
840         task_oncpu_function_call(task, __perf_install_in_context,
841                                  event);
842
843         spin_lock_irq(&ctx->lock);
844         /*
845          * we need to retry the smp call.
846          */
847         if (ctx->is_active && list_empty(&event->group_entry)) {
848                 spin_unlock_irq(&ctx->lock);
849                 goto retry;
850         }
851
852         /*
853          * The lock prevents that this context is scheduled in so we
854          * can add the event safely, if it the call above did not
855          * succeed.
856          */
857         if (list_empty(&event->group_entry))
858                 add_event_to_ctx(event, ctx);
859         spin_unlock_irq(&ctx->lock);
860 }
861
862 /*
863  * Put a event into inactive state and update time fields.
864  * Enabling the leader of a group effectively enables all
865  * the group members that aren't explicitly disabled, so we
866  * have to update their ->tstamp_enabled also.
867  * Note: this works for group members as well as group leaders
868  * since the non-leader members' sibling_lists will be empty.
869  */
870 static void __perf_event_mark_enabled(struct perf_event *event,
871                                         struct perf_event_context *ctx)
872 {
873         struct perf_event *sub;
874
875         event->state = PERF_EVENT_STATE_INACTIVE;
876         event->tstamp_enabled = ctx->time - event->total_time_enabled;
877         list_for_each_entry(sub, &event->sibling_list, group_entry)
878                 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
879                         sub->tstamp_enabled =
880                                 ctx->time - sub->total_time_enabled;
881 }
882
883 /*
884  * Cross CPU call to enable a performance event
885  */
886 static void __perf_event_enable(void *info)
887 {
888         struct perf_event *event = info;
889         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
890         struct perf_event_context *ctx = event->ctx;
891         struct perf_event *leader = event->group_leader;
892         int err;
893
894         /*
895          * If this is a per-task event, need to check whether this
896          * event's task is the current task on this cpu.
897          */
898         if (ctx->task && cpuctx->task_ctx != ctx) {
899                 if (cpuctx->task_ctx || ctx->task != current)
900                         return;
901                 cpuctx->task_ctx = ctx;
902         }
903
904         spin_lock(&ctx->lock);
905         ctx->is_active = 1;
906         update_context_time(ctx);
907
908         if (event->state >= PERF_EVENT_STATE_INACTIVE)
909                 goto unlock;
910         __perf_event_mark_enabled(event, ctx);
911
912         /*
913          * If the event is in a group and isn't the group leader,
914          * then don't put it on unless the group is on.
915          */
916         if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
917                 goto unlock;
918
919         if (!group_can_go_on(event, cpuctx, 1)) {
920                 err = -EEXIST;
921         } else {
922                 perf_disable();
923                 if (event == leader)
924                         err = group_sched_in(event, cpuctx, ctx,
925                                              smp_processor_id());
926                 else
927                         err = event_sched_in(event, cpuctx, ctx,
928                                                smp_processor_id());
929                 perf_enable();
930         }
931
932         if (err) {
933                 /*
934                  * If this event can't go on and it's part of a
935                  * group, then the whole group has to come off.
936                  */
937                 if (leader != event)
938                         group_sched_out(leader, cpuctx, ctx);
939                 if (leader->attr.pinned) {
940                         update_group_times(leader);
941                         leader->state = PERF_EVENT_STATE_ERROR;
942                 }
943         }
944
945  unlock:
946         spin_unlock(&ctx->lock);
947 }
948
949 /*
950  * Enable a event.
951  *
952  * If event->ctx is a cloned context, callers must make sure that
953  * every task struct that event->ctx->task could possibly point to
954  * remains valid.  This condition is satisfied when called through
955  * perf_event_for_each_child or perf_event_for_each as described
956  * for perf_event_disable.
957  */
958 static void perf_event_enable(struct perf_event *event)
959 {
960         struct perf_event_context *ctx = event->ctx;
961         struct task_struct *task = ctx->task;
962
963         if (!task) {
964                 /*
965                  * Enable the event on the cpu that it's on
966                  */
967                 smp_call_function_single(event->cpu, __perf_event_enable,
968                                          event, 1);
969                 return;
970         }
971
972         spin_lock_irq(&ctx->lock);
973         if (event->state >= PERF_EVENT_STATE_INACTIVE)
974                 goto out;
975
976         /*
977          * If the event is in error state, clear that first.
978          * That way, if we see the event in error state below, we
979          * know that it has gone back into error state, as distinct
980          * from the task having been scheduled away before the
981          * cross-call arrived.
982          */
983         if (event->state == PERF_EVENT_STATE_ERROR)
984                 event->state = PERF_EVENT_STATE_OFF;
985
986  retry:
987         spin_unlock_irq(&ctx->lock);
988         task_oncpu_function_call(task, __perf_event_enable, event);
989
990         spin_lock_irq(&ctx->lock);
991
992         /*
993          * If the context is active and the event is still off,
994          * we need to retry the cross-call.
995          */
996         if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
997                 goto retry;
998
999         /*
1000          * Since we have the lock this context can't be scheduled
1001          * in, so we can change the state safely.
1002          */
1003         if (event->state == PERF_EVENT_STATE_OFF)
1004                 __perf_event_mark_enabled(event, ctx);
1005
1006  out:
1007         spin_unlock_irq(&ctx->lock);
1008 }
1009
1010 static int perf_event_refresh(struct perf_event *event, int refresh)
1011 {
1012         /*
1013          * not supported on inherited events
1014          */
1015         if (event->attr.inherit)
1016                 return -EINVAL;
1017
1018         atomic_add(refresh, &event->event_limit);
1019         perf_event_enable(event);
1020
1021         return 0;
1022 }
1023
1024 void __perf_event_sched_out(struct perf_event_context *ctx,
1025                               struct perf_cpu_context *cpuctx)
1026 {
1027         struct perf_event *event;
1028
1029         spin_lock(&ctx->lock);
1030         ctx->is_active = 0;
1031         if (likely(!ctx->nr_events))
1032                 goto out;
1033         update_context_time(ctx);
1034
1035         perf_disable();
1036         if (ctx->nr_active)
1037                 list_for_each_entry(event, &ctx->group_list, group_entry)
1038                         group_sched_out(event, cpuctx, ctx);
1039
1040         perf_enable();
1041  out:
1042         spin_unlock(&ctx->lock);
1043 }
1044
1045 /*
1046  * Test whether two contexts are equivalent, i.e. whether they
1047  * have both been cloned from the same version of the same context
1048  * and they both have the same number of enabled events.
1049  * If the number of enabled events is the same, then the set
1050  * of enabled events should be the same, because these are both
1051  * inherited contexts, therefore we can't access individual events
1052  * in them directly with an fd; we can only enable/disable all
1053  * events via prctl, or enable/disable all events in a family
1054  * via ioctl, which will have the same effect on both contexts.
1055  */
1056 static int context_equiv(struct perf_event_context *ctx1,
1057                          struct perf_event_context *ctx2)
1058 {
1059         return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1060                 && ctx1->parent_gen == ctx2->parent_gen
1061                 && !ctx1->pin_count && !ctx2->pin_count;
1062 }
1063
1064 static void __perf_event_sync_stat(struct perf_event *event,
1065                                      struct perf_event *next_event)
1066 {
1067         u64 value;
1068
1069         if (!event->attr.inherit_stat)
1070                 return;
1071
1072         /*
1073          * Update the event value, we cannot use perf_event_read()
1074          * because we're in the middle of a context switch and have IRQs
1075          * disabled, which upsets smp_call_function_single(), however
1076          * we know the event must be on the current CPU, therefore we
1077          * don't need to use it.
1078          */
1079         switch (event->state) {
1080         case PERF_EVENT_STATE_ACTIVE:
1081                 event->pmu->read(event);
1082                 /* fall-through */
1083
1084         case PERF_EVENT_STATE_INACTIVE:
1085                 update_event_times(event);
1086                 break;
1087
1088         default:
1089                 break;
1090         }
1091
1092         /*
1093          * In order to keep per-task stats reliable we need to flip the event
1094          * values when we flip the contexts.
1095          */
1096         value = atomic64_read(&next_event->count);
1097         value = atomic64_xchg(&event->count, value);
1098         atomic64_set(&next_event->count, value);
1099
1100         swap(event->total_time_enabled, next_event->total_time_enabled);
1101         swap(event->total_time_running, next_event->total_time_running);
1102
1103         /*
1104          * Since we swizzled the values, update the user visible data too.
1105          */
1106         perf_event_update_userpage(event);
1107         perf_event_update_userpage(next_event);
1108 }
1109
1110 #define list_next_entry(pos, member) \
1111         list_entry(pos->member.next, typeof(*pos), member)
1112
1113 static void perf_event_sync_stat(struct perf_event_context *ctx,
1114                                    struct perf_event_context *next_ctx)
1115 {
1116         struct perf_event *event, *next_event;
1117
1118         if (!ctx->nr_stat)
1119                 return;
1120
1121         update_context_time(ctx);
1122
1123         event = list_first_entry(&ctx->event_list,
1124                                    struct perf_event, event_entry);
1125
1126         next_event = list_first_entry(&next_ctx->event_list,
1127                                         struct perf_event, event_entry);
1128
1129         while (&event->event_entry != &ctx->event_list &&
1130                &next_event->event_entry != &next_ctx->event_list) {
1131
1132                 __perf_event_sync_stat(event, next_event);
1133
1134                 event = list_next_entry(event, event_entry);
1135                 next_event = list_next_entry(next_event, event_entry);
1136         }
1137 }
1138
1139 /*
1140  * Called from scheduler to remove the events of the current task,
1141  * with interrupts disabled.
1142  *
1143  * We stop each event and update the event value in event->count.
1144  *
1145  * This does not protect us against NMI, but disable()
1146  * sets the disabled bit in the control field of event _before_
1147  * accessing the event control register. If a NMI hits, then it will
1148  * not restart the event.
1149  */
1150 void perf_event_task_sched_out(struct task_struct *task,
1151                                  struct task_struct *next, int cpu)
1152 {
1153         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1154         struct perf_event_context *ctx = task->perf_event_ctxp;
1155         struct perf_event_context *next_ctx;
1156         struct perf_event_context *parent;
1157         struct pt_regs *regs;
1158         int do_switch = 1;
1159
1160         regs = task_pt_regs(task);
1161         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1162
1163         if (likely(!ctx || !cpuctx->task_ctx))
1164                 return;
1165
1166         rcu_read_lock();
1167         parent = rcu_dereference(ctx->parent_ctx);
1168         next_ctx = next->perf_event_ctxp;
1169         if (parent && next_ctx &&
1170             rcu_dereference(next_ctx->parent_ctx) == parent) {
1171                 /*
1172                  * Looks like the two contexts are clones, so we might be
1173                  * able to optimize the context switch.  We lock both
1174                  * contexts and check that they are clones under the
1175                  * lock (including re-checking that neither has been
1176                  * uncloned in the meantime).  It doesn't matter which
1177                  * order we take the locks because no other cpu could
1178                  * be trying to lock both of these tasks.
1179                  */
1180                 spin_lock(&ctx->lock);
1181                 spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1182                 if (context_equiv(ctx, next_ctx)) {
1183                         /*
1184                          * XXX do we need a memory barrier of sorts
1185                          * wrt to rcu_dereference() of perf_event_ctxp
1186                          */
1187                         task->perf_event_ctxp = next_ctx;
1188                         next->perf_event_ctxp = ctx;
1189                         ctx->task = next;
1190                         next_ctx->task = task;
1191                         do_switch = 0;
1192
1193                         perf_event_sync_stat(ctx, next_ctx);
1194                 }
1195                 spin_unlock(&next_ctx->lock);
1196                 spin_unlock(&ctx->lock);
1197         }
1198         rcu_read_unlock();
1199
1200         if (do_switch) {
1201                 __perf_event_sched_out(ctx, cpuctx);
1202                 cpuctx->task_ctx = NULL;
1203         }
1204 }
1205
1206 /*
1207  * Called with IRQs disabled
1208  */
1209 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1210 {
1211         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1212
1213         if (!cpuctx->task_ctx)
1214                 return;
1215
1216         if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1217                 return;
1218
1219         __perf_event_sched_out(ctx, cpuctx);
1220         cpuctx->task_ctx = NULL;
1221 }
1222
1223 /*
1224  * Called with IRQs disabled
1225  */
1226 static void perf_event_cpu_sched_out(struct perf_cpu_context *cpuctx)
1227 {
1228         __perf_event_sched_out(&cpuctx->ctx, cpuctx);
1229 }
1230
1231 static void
1232 __perf_event_sched_in(struct perf_event_context *ctx,
1233                         struct perf_cpu_context *cpuctx, int cpu)
1234 {
1235         struct perf_event *event;
1236         int can_add_hw = 1;
1237
1238         spin_lock(&ctx->lock);
1239         ctx->is_active = 1;
1240         if (likely(!ctx->nr_events))
1241                 goto out;
1242
1243         ctx->timestamp = perf_clock();
1244
1245         perf_disable();
1246
1247         /*
1248          * First go through the list and put on any pinned groups
1249          * in order to give them the best chance of going on.
1250          */
1251         list_for_each_entry(event, &ctx->group_list, group_entry) {
1252                 if (event->state <= PERF_EVENT_STATE_OFF ||
1253                     !event->attr.pinned)
1254                         continue;
1255                 if (event->cpu != -1 && event->cpu != cpu)
1256                         continue;
1257
1258                 if (group_can_go_on(event, cpuctx, 1))
1259                         group_sched_in(event, cpuctx, ctx, cpu);
1260
1261                 /*
1262                  * If this pinned group hasn't been scheduled,
1263                  * put it in error state.
1264                  */
1265                 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1266                         update_group_times(event);
1267                         event->state = PERF_EVENT_STATE_ERROR;
1268                 }
1269         }
1270
1271         list_for_each_entry(event, &ctx->group_list, group_entry) {
1272                 /*
1273                  * Ignore events in OFF or ERROR state, and
1274                  * ignore pinned events since we did them already.
1275                  */
1276                 if (event->state <= PERF_EVENT_STATE_OFF ||
1277                     event->attr.pinned)
1278                         continue;
1279
1280                 /*
1281                  * Listen to the 'cpu' scheduling filter constraint
1282                  * of events:
1283                  */
1284                 if (event->cpu != -1 && event->cpu != cpu)
1285                         continue;
1286
1287                 if (group_can_go_on(event, cpuctx, can_add_hw))
1288                         if (group_sched_in(event, cpuctx, ctx, cpu))
1289                                 can_add_hw = 0;
1290         }
1291         perf_enable();
1292  out:
1293         spin_unlock(&ctx->lock);
1294 }
1295
1296 /*
1297  * Called from scheduler to add the events of the current task
1298  * with interrupts disabled.
1299  *
1300  * We restore the event value and then enable it.
1301  *
1302  * This does not protect us against NMI, but enable()
1303  * sets the enabled bit in the control field of event _before_
1304  * accessing the event control register. If a NMI hits, then it will
1305  * keep the event running.
1306  */
1307 void perf_event_task_sched_in(struct task_struct *task, int cpu)
1308 {
1309         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1310         struct perf_event_context *ctx = task->perf_event_ctxp;
1311
1312         if (likely(!ctx))
1313                 return;
1314         if (cpuctx->task_ctx == ctx)
1315                 return;
1316         __perf_event_sched_in(ctx, cpuctx, cpu);
1317         cpuctx->task_ctx = ctx;
1318 }
1319
1320 static void perf_event_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
1321 {
1322         struct perf_event_context *ctx = &cpuctx->ctx;
1323
1324         __perf_event_sched_in(ctx, cpuctx, cpu);
1325 }
1326
1327 #define MAX_INTERRUPTS (~0ULL)
1328
1329 static void perf_log_throttle(struct perf_event *event, int enable);
1330
1331 static void perf_adjust_period(struct perf_event *event, u64 events)
1332 {
1333         struct hw_perf_event *hwc = &event->hw;
1334         u64 period, sample_period;
1335         s64 delta;
1336
1337         events *= hwc->sample_period;
1338         period = div64_u64(events, event->attr.sample_freq);
1339
1340         delta = (s64)(period - hwc->sample_period);
1341         delta = (delta + 7) / 8; /* low pass filter */
1342
1343         sample_period = hwc->sample_period + delta;
1344
1345         if (!sample_period)
1346                 sample_period = 1;
1347
1348         hwc->sample_period = sample_period;
1349 }
1350
1351 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1352 {
1353         struct perf_event *event;
1354         struct hw_perf_event *hwc;
1355         u64 interrupts, freq;
1356
1357         spin_lock(&ctx->lock);
1358         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1359                 if (event->state != PERF_EVENT_STATE_ACTIVE)
1360                         continue;
1361
1362                 hwc = &event->hw;
1363
1364                 interrupts = hwc->interrupts;
1365                 hwc->interrupts = 0;
1366
1367                 /*
1368                  * unthrottle events on the tick
1369                  */
1370                 if (interrupts == MAX_INTERRUPTS) {
1371                         perf_log_throttle(event, 1);
1372                         event->pmu->unthrottle(event);
1373                         interrupts = 2*sysctl_perf_event_sample_rate/HZ;
1374                 }
1375
1376                 if (!event->attr.freq || !event->attr.sample_freq)
1377                         continue;
1378
1379                 /*
1380                  * if the specified freq < HZ then we need to skip ticks
1381                  */
1382                 if (event->attr.sample_freq < HZ) {
1383                         freq = event->attr.sample_freq;
1384
1385                         hwc->freq_count += freq;
1386                         hwc->freq_interrupts += interrupts;
1387
1388                         if (hwc->freq_count < HZ)
1389                                 continue;
1390
1391                         interrupts = hwc->freq_interrupts;
1392                         hwc->freq_interrupts = 0;
1393                         hwc->freq_count -= HZ;
1394                 } else
1395                         freq = HZ;
1396
1397                 perf_adjust_period(event, freq * interrupts);
1398
1399                 /*
1400                  * In order to avoid being stalled by an (accidental) huge
1401                  * sample period, force reset the sample period if we didn't
1402                  * get any events in this freq period.
1403                  */
1404                 if (!interrupts) {
1405                         perf_disable();
1406                         event->pmu->disable(event);
1407                         atomic64_set(&hwc->period_left, 0);
1408                         event->pmu->enable(event);
1409                         perf_enable();
1410                 }
1411         }
1412         spin_unlock(&ctx->lock);
1413 }
1414
1415 /*
1416  * Round-robin a context's events:
1417  */
1418 static void rotate_ctx(struct perf_event_context *ctx)
1419 {
1420         struct perf_event *event;
1421
1422         if (!ctx->nr_events)
1423                 return;
1424
1425         spin_lock(&ctx->lock);
1426         /*
1427          * Rotate the first entry last (works just fine for group events too):
1428          */
1429         perf_disable();
1430         list_for_each_entry(event, &ctx->group_list, group_entry) {
1431                 list_move_tail(&event->group_entry, &ctx->group_list);
1432                 break;
1433         }
1434         perf_enable();
1435
1436         spin_unlock(&ctx->lock);
1437 }
1438
1439 void perf_event_task_tick(struct task_struct *curr, int cpu)
1440 {
1441         struct perf_cpu_context *cpuctx;
1442         struct perf_event_context *ctx;
1443
1444         if (!atomic_read(&nr_events))
1445                 return;
1446
1447         cpuctx = &per_cpu(perf_cpu_context, cpu);
1448         ctx = curr->perf_event_ctxp;
1449
1450         perf_ctx_adjust_freq(&cpuctx->ctx);
1451         if (ctx)
1452                 perf_ctx_adjust_freq(ctx);
1453
1454         perf_event_cpu_sched_out(cpuctx);
1455         if (ctx)
1456                 __perf_event_task_sched_out(ctx);
1457
1458         rotate_ctx(&cpuctx->ctx);
1459         if (ctx)
1460                 rotate_ctx(ctx);
1461
1462         perf_event_cpu_sched_in(cpuctx, cpu);
1463         if (ctx)
1464                 perf_event_task_sched_in(curr, cpu);
1465 }
1466
1467 /*
1468  * Enable all of a task's events that have been marked enable-on-exec.
1469  * This expects task == current.
1470  */
1471 static void perf_event_enable_on_exec(struct task_struct *task)
1472 {
1473         struct perf_event_context *ctx;
1474         struct perf_event *event;
1475         unsigned long flags;
1476         int enabled = 0;
1477
1478         local_irq_save(flags);
1479         ctx = task->perf_event_ctxp;
1480         if (!ctx || !ctx->nr_events)
1481                 goto out;
1482
1483         __perf_event_task_sched_out(ctx);
1484
1485         spin_lock(&ctx->lock);
1486
1487         list_for_each_entry(event, &ctx->group_list, group_entry) {
1488                 if (!event->attr.enable_on_exec)
1489                         continue;
1490                 event->attr.enable_on_exec = 0;
1491                 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1492                         continue;
1493                 __perf_event_mark_enabled(event, ctx);
1494                 enabled = 1;
1495         }
1496
1497         /*
1498          * Unclone this context if we enabled any event.
1499          */
1500         if (enabled)
1501                 unclone_ctx(ctx);
1502
1503         spin_unlock(&ctx->lock);
1504
1505         perf_event_task_sched_in(task, smp_processor_id());
1506  out:
1507         local_irq_restore(flags);
1508 }
1509
1510 /*
1511  * Cross CPU call to read the hardware event
1512  */
1513 static void __perf_event_read(void *info)
1514 {
1515         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1516         struct perf_event *event = info;
1517         struct perf_event_context *ctx = event->ctx;
1518
1519         /*
1520          * If this is a task context, we need to check whether it is
1521          * the current task context of this cpu.  If not it has been
1522          * scheduled out before the smp call arrived.  In that case
1523          * event->count would have been updated to a recent sample
1524          * when the event was scheduled out.
1525          */
1526         if (ctx->task && cpuctx->task_ctx != ctx)
1527                 return;
1528
1529         if (ctx->is_active)
1530                 update_context_time(ctx);
1531         event->pmu->read(event);
1532         update_event_times(event);
1533 }
1534
1535 static u64 perf_event_read(struct perf_event *event)
1536 {
1537         /*
1538          * If event is enabled and currently active on a CPU, update the
1539          * value in the event structure:
1540          */
1541         if (event->state == PERF_EVENT_STATE_ACTIVE) {
1542                 smp_call_function_single(event->oncpu,
1543                                          __perf_event_read, event, 1);
1544         } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1545                 update_event_times(event);
1546         }
1547
1548         return atomic64_read(&event->count);
1549 }
1550
1551 /*
1552  * Initialize the perf_event context in a task_struct:
1553  */
1554 static void
1555 __perf_event_init_context(struct perf_event_context *ctx,
1556                             struct task_struct *task)
1557 {
1558         memset(ctx, 0, sizeof(*ctx));
1559         spin_lock_init(&ctx->lock);
1560         mutex_init(&ctx->mutex);
1561         INIT_LIST_HEAD(&ctx->group_list);
1562         INIT_LIST_HEAD(&ctx->event_list);
1563         atomic_set(&ctx->refcount, 1);
1564         ctx->task = task;
1565 }
1566
1567 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1568 {
1569         struct perf_event_context *ctx;
1570         struct perf_cpu_context *cpuctx;
1571         struct task_struct *task;
1572         unsigned long flags;
1573         int err;
1574
1575         /*
1576          * If cpu is not a wildcard then this is a percpu event:
1577          */
1578         if (cpu != -1) {
1579                 /* Must be root to operate on a CPU event: */
1580                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1581                         return ERR_PTR(-EACCES);
1582
1583                 if (cpu < 0 || cpu > num_possible_cpus())
1584                         return ERR_PTR(-EINVAL);
1585
1586                 /*
1587                  * We could be clever and allow to attach a event to an
1588                  * offline CPU and activate it when the CPU comes up, but
1589                  * that's for later.
1590                  */
1591                 if (!cpu_isset(cpu, cpu_online_map))
1592                         return ERR_PTR(-ENODEV);
1593
1594                 cpuctx = &per_cpu(perf_cpu_context, cpu);
1595                 ctx = &cpuctx->ctx;
1596                 get_ctx(ctx);
1597
1598                 return ctx;
1599         }
1600
1601         rcu_read_lock();
1602         if (!pid)
1603                 task = current;
1604         else
1605                 task = find_task_by_vpid(pid);
1606         if (task)
1607                 get_task_struct(task);
1608         rcu_read_unlock();
1609
1610         if (!task)
1611                 return ERR_PTR(-ESRCH);
1612
1613         /*
1614          * Can't attach events to a dying task.
1615          */
1616         err = -ESRCH;
1617         if (task->flags & PF_EXITING)
1618                 goto errout;
1619
1620         /* Reuse ptrace permission checks for now. */
1621         err = -EACCES;
1622         if (!ptrace_may_access(task, PTRACE_MODE_READ))
1623                 goto errout;
1624
1625  retry:
1626         ctx = perf_lock_task_context(task, &flags);
1627         if (ctx) {
1628                 unclone_ctx(ctx);
1629                 spin_unlock_irqrestore(&ctx->lock, flags);
1630         }
1631
1632         if (!ctx) {
1633                 ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1634                 err = -ENOMEM;
1635                 if (!ctx)
1636                         goto errout;
1637                 __perf_event_init_context(ctx, task);
1638                 get_ctx(ctx);
1639                 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1640                         /*
1641                          * We raced with some other task; use
1642                          * the context they set.
1643                          */
1644                         kfree(ctx);
1645                         goto retry;
1646                 }
1647                 get_task_struct(task);
1648         }
1649
1650         put_task_struct(task);
1651         return ctx;
1652
1653  errout:
1654         put_task_struct(task);
1655         return ERR_PTR(err);
1656 }
1657
1658 static void perf_event_free_filter(struct perf_event *event);
1659
1660 static void free_event_rcu(struct rcu_head *head)
1661 {
1662         struct perf_event *event;
1663
1664         event = container_of(head, struct perf_event, rcu_head);
1665         if (event->ns)
1666                 put_pid_ns(event->ns);
1667         perf_event_free_filter(event);
1668         kfree(event);
1669 }
1670
1671 static void perf_pending_sync(struct perf_event *event);
1672
1673 static void free_event(struct perf_event *event)
1674 {
1675         perf_pending_sync(event);
1676
1677         if (!event->parent) {
1678                 atomic_dec(&nr_events);
1679                 if (event->attr.mmap)
1680                         atomic_dec(&nr_mmap_events);
1681                 if (event->attr.comm)
1682                         atomic_dec(&nr_comm_events);
1683                 if (event->attr.task)
1684                         atomic_dec(&nr_task_events);
1685         }
1686
1687         if (event->output) {
1688                 fput(event->output->filp);
1689                 event->output = NULL;
1690         }
1691
1692         if (event->destroy)
1693                 event->destroy(event);
1694
1695         put_ctx(event->ctx);
1696         call_rcu(&event->rcu_head, free_event_rcu);
1697 }
1698
1699 /*
1700  * Called when the last reference to the file is gone.
1701  */
1702 static int perf_release(struct inode *inode, struct file *file)
1703 {
1704         struct perf_event *event = file->private_data;
1705         struct perf_event_context *ctx = event->ctx;
1706
1707         file->private_data = NULL;
1708
1709         WARN_ON_ONCE(ctx->parent_ctx);
1710         mutex_lock(&ctx->mutex);
1711         perf_event_remove_from_context(event);
1712         mutex_unlock(&ctx->mutex);
1713
1714         mutex_lock(&event->owner->perf_event_mutex);
1715         list_del_init(&event->owner_entry);
1716         mutex_unlock(&event->owner->perf_event_mutex);
1717         put_task_struct(event->owner);
1718
1719         free_event(event);
1720
1721         return 0;
1722 }
1723
1724 int perf_event_release_kernel(struct perf_event *event)
1725 {
1726         struct perf_event_context *ctx = event->ctx;
1727
1728         WARN_ON_ONCE(ctx->parent_ctx);
1729         mutex_lock(&ctx->mutex);
1730         perf_event_remove_from_context(event);
1731         mutex_unlock(&ctx->mutex);
1732
1733         mutex_lock(&event->owner->perf_event_mutex);
1734         list_del_init(&event->owner_entry);
1735         mutex_unlock(&event->owner->perf_event_mutex);
1736         put_task_struct(event->owner);
1737
1738         free_event(event);
1739
1740         return 0;
1741 }
1742 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1743
1744 static int perf_event_read_size(struct perf_event *event)
1745 {
1746         int entry = sizeof(u64); /* value */
1747         int size = 0;
1748         int nr = 1;
1749
1750         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1751                 size += sizeof(u64);
1752
1753         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1754                 size += sizeof(u64);
1755
1756         if (event->attr.read_format & PERF_FORMAT_ID)
1757                 entry += sizeof(u64);
1758
1759         if (event->attr.read_format & PERF_FORMAT_GROUP) {
1760                 nr += event->group_leader->nr_siblings;
1761                 size += sizeof(u64);
1762         }
1763
1764         size += entry * nr;
1765
1766         return size;
1767 }
1768
1769 u64 perf_event_read_value(struct perf_event *event)
1770 {
1771         struct perf_event *child;
1772         u64 total = 0;
1773
1774         total += perf_event_read(event);
1775         list_for_each_entry(child, &event->child_list, child_list)
1776                 total += perf_event_read(child);
1777
1778         return total;
1779 }
1780 EXPORT_SYMBOL_GPL(perf_event_read_value);
1781
1782 static int perf_event_read_group(struct perf_event *event,
1783                                    u64 read_format, char __user *buf)
1784 {
1785         struct perf_event *leader = event->group_leader, *sub;
1786         int n = 0, size = 0, ret = 0;
1787         u64 values[5];
1788         u64 count;
1789
1790         count = perf_event_read_value(leader);
1791
1792         values[n++] = 1 + leader->nr_siblings;
1793         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1794                 values[n++] = leader->total_time_enabled +
1795                         atomic64_read(&leader->child_total_time_enabled);
1796         }
1797         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1798                 values[n++] = leader->total_time_running +
1799                         atomic64_read(&leader->child_total_time_running);
1800         }
1801         values[n++] = count;
1802         if (read_format & PERF_FORMAT_ID)
1803                 values[n++] = primary_event_id(leader);
1804
1805         size = n * sizeof(u64);
1806
1807         if (copy_to_user(buf, values, size))
1808                 return -EFAULT;
1809
1810         ret += size;
1811
1812         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1813                 n = 0;
1814
1815                 values[n++] = perf_event_read_value(sub);
1816                 if (read_format & PERF_FORMAT_ID)
1817                         values[n++] = primary_event_id(sub);
1818
1819                 size = n * sizeof(u64);
1820
1821                 if (copy_to_user(buf + size, values, size))
1822                         return -EFAULT;
1823
1824                 ret += size;
1825         }
1826
1827         return ret;
1828 }
1829
1830 static int perf_event_read_one(struct perf_event *event,
1831                                  u64 read_format, char __user *buf)
1832 {
1833         u64 values[4];
1834         int n = 0;
1835
1836         values[n++] = perf_event_read_value(event);
1837         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1838                 values[n++] = event->total_time_enabled +
1839                         atomic64_read(&event->child_total_time_enabled);
1840         }
1841         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1842                 values[n++] = event->total_time_running +
1843                         atomic64_read(&event->child_total_time_running);
1844         }
1845         if (read_format & PERF_FORMAT_ID)
1846                 values[n++] = primary_event_id(event);
1847
1848         if (copy_to_user(buf, values, n * sizeof(u64)))
1849                 return -EFAULT;
1850
1851         return n * sizeof(u64);
1852 }
1853
1854 /*
1855  * Read the performance event - simple non blocking version for now
1856  */
1857 static ssize_t
1858 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
1859 {
1860         u64 read_format = event->attr.read_format;
1861         int ret;
1862
1863         /*
1864          * Return end-of-file for a read on a event that is in
1865          * error state (i.e. because it was pinned but it couldn't be
1866          * scheduled on to the CPU at some point).
1867          */
1868         if (event->state == PERF_EVENT_STATE_ERROR)
1869                 return 0;
1870
1871         if (count < perf_event_read_size(event))
1872                 return -ENOSPC;
1873
1874         WARN_ON_ONCE(event->ctx->parent_ctx);
1875         mutex_lock(&event->child_mutex);
1876         if (read_format & PERF_FORMAT_GROUP)
1877                 ret = perf_event_read_group(event, read_format, buf);
1878         else
1879                 ret = perf_event_read_one(event, read_format, buf);
1880         mutex_unlock(&event->child_mutex);
1881
1882         return ret;
1883 }
1884
1885 static ssize_t
1886 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1887 {
1888         struct perf_event *event = file->private_data;
1889
1890         return perf_read_hw(event, buf, count);
1891 }
1892
1893 static unsigned int perf_poll(struct file *file, poll_table *wait)
1894 {
1895         struct perf_event *event = file->private_data;
1896         struct perf_mmap_data *data;
1897         unsigned int events = POLL_HUP;
1898
1899         rcu_read_lock();
1900         data = rcu_dereference(event->data);
1901         if (data)
1902                 events = atomic_xchg(&data->poll, 0);
1903         rcu_read_unlock();
1904
1905         poll_wait(file, &event->waitq, wait);
1906
1907         return events;
1908 }
1909
1910 static void perf_event_reset(struct perf_event *event)
1911 {
1912         (void)perf_event_read(event);
1913         atomic64_set(&event->count, 0);
1914         perf_event_update_userpage(event);
1915 }
1916
1917 /*
1918  * Holding the top-level event's child_mutex means that any
1919  * descendant process that has inherited this event will block
1920  * in sync_child_event if it goes to exit, thus satisfying the
1921  * task existence requirements of perf_event_enable/disable.
1922  */
1923 static void perf_event_for_each_child(struct perf_event *event,
1924                                         void (*func)(struct perf_event *))
1925 {
1926         struct perf_event *child;
1927
1928         WARN_ON_ONCE(event->ctx->parent_ctx);
1929         mutex_lock(&event->child_mutex);
1930         func(event);
1931         list_for_each_entry(child, &event->child_list, child_list)
1932                 func(child);
1933         mutex_unlock(&event->child_mutex);
1934 }
1935
1936 static void perf_event_for_each(struct perf_event *event,
1937                                   void (*func)(struct perf_event *))
1938 {
1939         struct perf_event_context *ctx = event->ctx;
1940         struct perf_event *sibling;
1941
1942         WARN_ON_ONCE(ctx->parent_ctx);
1943         mutex_lock(&ctx->mutex);
1944         event = event->group_leader;
1945
1946         perf_event_for_each_child(event, func);
1947         func(event);
1948         list_for_each_entry(sibling, &event->sibling_list, group_entry)
1949                 perf_event_for_each_child(event, func);
1950         mutex_unlock(&ctx->mutex);
1951 }
1952
1953 static int perf_event_period(struct perf_event *event, u64 __user *arg)
1954 {
1955         struct perf_event_context *ctx = event->ctx;
1956         unsigned long size;
1957         int ret = 0;
1958         u64 value;
1959
1960         if (!event->attr.sample_period)
1961                 return -EINVAL;
1962
1963         size = copy_from_user(&value, arg, sizeof(value));
1964         if (size != sizeof(value))
1965                 return -EFAULT;
1966
1967         if (!value)
1968                 return -EINVAL;
1969
1970         spin_lock_irq(&ctx->lock);
1971         if (event->attr.freq) {
1972                 if (value > sysctl_perf_event_sample_rate) {
1973                         ret = -EINVAL;
1974                         goto unlock;
1975                 }
1976
1977                 event->attr.sample_freq = value;
1978         } else {
1979                 event->attr.sample_period = value;
1980                 event->hw.sample_period = value;
1981         }
1982 unlock:
1983         spin_unlock_irq(&ctx->lock);
1984
1985         return ret;
1986 }
1987
1988 static int perf_event_set_output(struct perf_event *event, int output_fd);
1989 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
1990
1991 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1992 {
1993         struct perf_event *event = file->private_data;
1994         void (*func)(struct perf_event *);
1995         u32 flags = arg;
1996
1997         switch (cmd) {
1998         case PERF_EVENT_IOC_ENABLE:
1999                 func = perf_event_enable;
2000                 break;
2001         case PERF_EVENT_IOC_DISABLE:
2002                 func = perf_event_disable;
2003                 break;
2004         case PERF_EVENT_IOC_RESET:
2005                 func = perf_event_reset;
2006                 break;
2007
2008         case PERF_EVENT_IOC_REFRESH:
2009                 return perf_event_refresh(event, arg);
2010
2011         case PERF_EVENT_IOC_PERIOD:
2012                 return perf_event_period(event, (u64 __user *)arg);
2013
2014         case PERF_EVENT_IOC_SET_OUTPUT:
2015                 return perf_event_set_output(event, arg);
2016
2017         case PERF_EVENT_IOC_SET_FILTER:
2018                 return perf_event_set_filter(event, (void __user *)arg);
2019
2020         default:
2021                 return -ENOTTY;
2022         }
2023
2024         if (flags & PERF_IOC_FLAG_GROUP)
2025                 perf_event_for_each(event, func);
2026         else
2027                 perf_event_for_each_child(event, func);
2028
2029         return 0;
2030 }
2031
2032 int perf_event_task_enable(void)
2033 {
2034         struct perf_event *event;
2035
2036         mutex_lock(&current->perf_event_mutex);
2037         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2038                 perf_event_for_each_child(event, perf_event_enable);
2039         mutex_unlock(&current->perf_event_mutex);
2040
2041         return 0;
2042 }
2043
2044 int perf_event_task_disable(void)
2045 {
2046         struct perf_event *event;
2047
2048         mutex_lock(&current->perf_event_mutex);
2049         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2050                 perf_event_for_each_child(event, perf_event_disable);
2051         mutex_unlock(&current->perf_event_mutex);
2052
2053         return 0;
2054 }
2055
2056 #ifndef PERF_EVENT_INDEX_OFFSET
2057 # define PERF_EVENT_INDEX_OFFSET 0
2058 #endif
2059
2060 static int perf_event_index(struct perf_event *event)
2061 {
2062         if (event->state != PERF_EVENT_STATE_ACTIVE)
2063                 return 0;
2064
2065         return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2066 }
2067
2068 /*
2069  * Callers need to ensure there can be no nesting of this function, otherwise
2070  * the seqlock logic goes bad. We can not serialize this because the arch
2071  * code calls this from NMI context.
2072  */
2073 void perf_event_update_userpage(struct perf_event *event)
2074 {
2075         struct perf_event_mmap_page *userpg;
2076         struct perf_mmap_data *data;
2077
2078         rcu_read_lock();
2079         data = rcu_dereference(event->data);
2080         if (!data)
2081                 goto unlock;
2082
2083         userpg = data->user_page;
2084
2085         /*
2086          * Disable preemption so as to not let the corresponding user-space
2087          * spin too long if we get preempted.
2088          */
2089         preempt_disable();
2090         ++userpg->lock;
2091         barrier();
2092         userpg->index = perf_event_index(event);
2093         userpg->offset = atomic64_read(&event->count);
2094         if (event->state == PERF_EVENT_STATE_ACTIVE)
2095                 userpg->offset -= atomic64_read(&event->hw.prev_count);
2096
2097         userpg->time_enabled = event->total_time_enabled +
2098                         atomic64_read(&event->child_total_time_enabled);
2099
2100         userpg->time_running = event->total_time_running +
2101                         atomic64_read(&event->child_total_time_running);
2102
2103         barrier();
2104         ++userpg->lock;
2105         preempt_enable();
2106 unlock:
2107         rcu_read_unlock();
2108 }
2109
2110 static unsigned long perf_data_size(struct perf_mmap_data *data)
2111 {
2112         return data->nr_pages << (PAGE_SHIFT + data->data_order);
2113 }
2114
2115 #ifndef CONFIG_PERF_USE_VMALLOC
2116
2117 /*
2118  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2119  */
2120
2121 static struct page *
2122 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2123 {
2124         if (pgoff > data->nr_pages)
2125                 return NULL;
2126
2127         if (pgoff == 0)
2128                 return virt_to_page(data->user_page);
2129
2130         return virt_to_page(data->data_pages[pgoff - 1]);
2131 }
2132
2133 static struct perf_mmap_data *
2134 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2135 {
2136         struct perf_mmap_data *data;
2137         unsigned long size;
2138         int i;
2139
2140         WARN_ON(atomic_read(&event->mmap_count));
2141
2142         size = sizeof(struct perf_mmap_data);
2143         size += nr_pages * sizeof(void *);
2144
2145         data = kzalloc(size, GFP_KERNEL);
2146         if (!data)
2147                 goto fail;
2148
2149         data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2150         if (!data->user_page)
2151                 goto fail_user_page;
2152
2153         for (i = 0; i < nr_pages; i++) {
2154                 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2155                 if (!data->data_pages[i])
2156                         goto fail_data_pages;
2157         }
2158
2159         data->data_order = 0;
2160         data->nr_pages = nr_pages;
2161
2162         return data;
2163
2164 fail_data_pages:
2165         for (i--; i >= 0; i--)
2166                 free_page((unsigned long)data->data_pages[i]);
2167
2168         free_page((unsigned long)data->user_page);
2169
2170 fail_user_page:
2171         kfree(data);
2172
2173 fail:
2174         return NULL;
2175 }
2176
2177 static void perf_mmap_free_page(unsigned long addr)
2178 {
2179         struct page *page = virt_to_page((void *)addr);
2180
2181         page->mapping = NULL;
2182         __free_page(page);
2183 }
2184
2185 static void perf_mmap_data_free(struct perf_mmap_data *data)
2186 {
2187         int i;
2188
2189         perf_mmap_free_page((unsigned long)data->user_page);
2190         for (i = 0; i < data->nr_pages; i++)
2191                 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2192 }
2193
2194 #else
2195
2196 /*
2197  * Back perf_mmap() with vmalloc memory.
2198  *
2199  * Required for architectures that have d-cache aliasing issues.
2200  */
2201
2202 static struct page *
2203 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2204 {
2205         if (pgoff > (1UL << data->data_order))
2206                 return NULL;
2207
2208         return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2209 }
2210
2211 static void perf_mmap_unmark_page(void *addr)
2212 {
2213         struct page *page = vmalloc_to_page(addr);
2214
2215         page->mapping = NULL;
2216 }
2217
2218 static void perf_mmap_data_free_work(struct work_struct *work)
2219 {
2220         struct perf_mmap_data *data;
2221         void *base;
2222         int i, nr;
2223
2224         data = container_of(work, struct perf_mmap_data, work);
2225         nr = 1 << data->data_order;
2226
2227         base = data->user_page;
2228         for (i = 0; i < nr + 1; i++)
2229                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2230
2231         vfree(base);
2232 }
2233
2234 static void perf_mmap_data_free(struct perf_mmap_data *data)
2235 {
2236         schedule_work(&data->work);
2237 }
2238
2239 static struct perf_mmap_data *
2240 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2241 {
2242         struct perf_mmap_data *data;
2243         unsigned long size;
2244         void *all_buf;
2245
2246         WARN_ON(atomic_read(&event->mmap_count));
2247
2248         size = sizeof(struct perf_mmap_data);
2249         size += sizeof(void *);
2250
2251         data = kzalloc(size, GFP_KERNEL);
2252         if (!data)
2253                 goto fail;
2254
2255         INIT_WORK(&data->work, perf_mmap_data_free_work);
2256
2257         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2258         if (!all_buf)
2259                 goto fail_all_buf;
2260
2261         data->user_page = all_buf;
2262         data->data_pages[0] = all_buf + PAGE_SIZE;
2263         data->data_order = ilog2(nr_pages);
2264         data->nr_pages = 1;
2265
2266         return data;
2267
2268 fail_all_buf:
2269         kfree(data);
2270
2271 fail:
2272         return NULL;
2273 }
2274
2275 #endif
2276
2277 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2278 {
2279         struct perf_event *event = vma->vm_file->private_data;
2280         struct perf_mmap_data *data;
2281         int ret = VM_FAULT_SIGBUS;
2282
2283         if (vmf->flags & FAULT_FLAG_MKWRITE) {
2284                 if (vmf->pgoff == 0)
2285                         ret = 0;
2286                 return ret;
2287         }
2288
2289         rcu_read_lock();
2290         data = rcu_dereference(event->data);
2291         if (!data)
2292                 goto unlock;
2293
2294         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2295                 goto unlock;
2296
2297         vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2298         if (!vmf->page)
2299                 goto unlock;
2300
2301         get_page(vmf->page);
2302         vmf->page->mapping = vma->vm_file->f_mapping;
2303         vmf->page->index   = vmf->pgoff;
2304
2305         ret = 0;
2306 unlock:
2307         rcu_read_unlock();
2308
2309         return ret;
2310 }
2311
2312 static void
2313 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2314 {
2315         long max_size = perf_data_size(data);
2316
2317         atomic_set(&data->lock, -1);
2318
2319         if (event->attr.watermark) {
2320                 data->watermark = min_t(long, max_size,
2321                                         event->attr.wakeup_watermark);
2322         }
2323
2324         if (!data->watermark)
2325                 data->watermark = max_t(long, PAGE_SIZE, max_size / 2);
2326
2327
2328         rcu_assign_pointer(event->data, data);
2329 }
2330
2331 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2332 {
2333         struct perf_mmap_data *data;
2334
2335         data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2336         perf_mmap_data_free(data);
2337         kfree(data);
2338 }
2339
2340 static void perf_mmap_data_release(struct perf_event *event)
2341 {
2342         struct perf_mmap_data *data = event->data;
2343
2344         WARN_ON(atomic_read(&event->mmap_count));
2345
2346         rcu_assign_pointer(event->data, NULL);
2347         call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2348 }
2349
2350 static void perf_mmap_open(struct vm_area_struct *vma)
2351 {
2352         struct perf_event *event = vma->vm_file->private_data;
2353
2354         atomic_inc(&event->mmap_count);
2355 }
2356
2357 static void perf_mmap_close(struct vm_area_struct *vma)
2358 {
2359         struct perf_event *event = vma->vm_file->private_data;
2360
2361         WARN_ON_ONCE(event->ctx->parent_ctx);
2362         if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2363                 unsigned long size = perf_data_size(event->data);
2364                 struct user_struct *user = current_user();
2365
2366                 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2367                 vma->vm_mm->locked_vm -= event->data->nr_locked;
2368                 perf_mmap_data_release(event);
2369                 mutex_unlock(&event->mmap_mutex);
2370         }
2371 }
2372
2373 static const struct vm_operations_struct perf_mmap_vmops = {
2374         .open           = perf_mmap_open,
2375         .close          = perf_mmap_close,
2376         .fault          = perf_mmap_fault,
2377         .page_mkwrite   = perf_mmap_fault,
2378 };
2379
2380 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2381 {
2382         struct perf_event *event = file->private_data;
2383         unsigned long user_locked, user_lock_limit;
2384         struct user_struct *user = current_user();
2385         unsigned long locked, lock_limit;
2386         struct perf_mmap_data *data;
2387         unsigned long vma_size;
2388         unsigned long nr_pages;
2389         long user_extra, extra;
2390         int ret = 0;
2391
2392         if (!(vma->vm_flags & VM_SHARED))
2393                 return -EINVAL;
2394
2395         vma_size = vma->vm_end - vma->vm_start;
2396         nr_pages = (vma_size / PAGE_SIZE) - 1;
2397
2398         /*
2399          * If we have data pages ensure they're a power-of-two number, so we
2400          * can do bitmasks instead of modulo.
2401          */
2402         if (nr_pages != 0 && !is_power_of_2(nr_pages))
2403                 return -EINVAL;
2404
2405         if (vma_size != PAGE_SIZE * (1 + nr_pages))
2406                 return -EINVAL;
2407
2408         if (vma->vm_pgoff != 0)
2409                 return -EINVAL;
2410
2411         WARN_ON_ONCE(event->ctx->parent_ctx);
2412         mutex_lock(&event->mmap_mutex);
2413         if (event->output) {
2414                 ret = -EINVAL;
2415                 goto unlock;
2416         }
2417
2418         if (atomic_inc_not_zero(&event->mmap_count)) {
2419                 if (nr_pages != event->data->nr_pages)
2420                         ret = -EINVAL;
2421                 goto unlock;
2422         }
2423
2424         user_extra = nr_pages + 1;
2425         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2426
2427         /*
2428          * Increase the limit linearly with more CPUs:
2429          */
2430         user_lock_limit *= num_online_cpus();
2431
2432         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2433
2434         extra = 0;
2435         if (user_locked > user_lock_limit)
2436                 extra = user_locked - user_lock_limit;
2437
2438         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2439         lock_limit >>= PAGE_SHIFT;
2440         locked = vma->vm_mm->locked_vm + extra;
2441
2442         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2443                 !capable(CAP_IPC_LOCK)) {
2444                 ret = -EPERM;
2445                 goto unlock;
2446         }
2447
2448         WARN_ON(event->data);
2449
2450         data = perf_mmap_data_alloc(event, nr_pages);
2451         ret = -ENOMEM;
2452         if (!data)
2453                 goto unlock;
2454
2455         ret = 0;
2456         perf_mmap_data_init(event, data);
2457
2458         atomic_set(&event->mmap_count, 1);
2459         atomic_long_add(user_extra, &user->locked_vm);
2460         vma->vm_mm->locked_vm += extra;
2461         event->data->nr_locked = extra;
2462         if (vma->vm_flags & VM_WRITE)
2463                 event->data->writable = 1;
2464
2465 unlock:
2466         mutex_unlock(&event->mmap_mutex);
2467
2468         vma->vm_flags |= VM_RESERVED;
2469         vma->vm_ops = &perf_mmap_vmops;
2470
2471         return ret;
2472 }
2473
2474 static int perf_fasync(int fd, struct file *filp, int on)
2475 {
2476         struct inode *inode = filp->f_path.dentry->d_inode;
2477         struct perf_event *event = filp->private_data;
2478         int retval;
2479
2480         mutex_lock(&inode->i_mutex);
2481         retval = fasync_helper(fd, filp, on, &event->fasync);
2482         mutex_unlock(&inode->i_mutex);
2483
2484         if (retval < 0)
2485                 return retval;
2486
2487         return 0;
2488 }
2489
2490 static const struct file_operations perf_fops = {
2491         .release                = perf_release,
2492         .read                   = perf_read,
2493         .poll                   = perf_poll,
2494         .unlocked_ioctl         = perf_ioctl,
2495         .compat_ioctl           = perf_ioctl,
2496         .mmap                   = perf_mmap,
2497         .fasync                 = perf_fasync,
2498 };
2499
2500 /*
2501  * Perf event wakeup
2502  *
2503  * If there's data, ensure we set the poll() state and publish everything
2504  * to user-space before waking everybody up.
2505  */
2506
2507 void perf_event_wakeup(struct perf_event *event)
2508 {
2509         wake_up_all(&event->waitq);
2510
2511         if (event->pending_kill) {
2512                 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2513                 event->pending_kill = 0;
2514         }
2515 }
2516
2517 /*
2518  * Pending wakeups
2519  *
2520  * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2521  *
2522  * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2523  * single linked list and use cmpxchg() to add entries lockless.
2524  */
2525
2526 static void perf_pending_event(struct perf_pending_entry *entry)
2527 {
2528         struct perf_event *event = container_of(entry,
2529                         struct perf_event, pending);
2530
2531         if (event->pending_disable) {
2532                 event->pending_disable = 0;
2533                 __perf_event_disable(event);
2534         }
2535
2536         if (event->pending_wakeup) {
2537                 event->pending_wakeup = 0;
2538                 perf_event_wakeup(event);
2539         }
2540 }
2541
2542 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2543
2544 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2545         PENDING_TAIL,
2546 };
2547
2548 static void perf_pending_queue(struct perf_pending_entry *entry,
2549                                void (*func)(struct perf_pending_entry *))
2550 {
2551         struct perf_pending_entry **head;
2552
2553         if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2554                 return;
2555
2556         entry->func = func;
2557
2558         head = &get_cpu_var(perf_pending_head);
2559
2560         do {
2561                 entry->next = *head;
2562         } while (cmpxchg(head, entry->next, entry) != entry->next);
2563
2564         set_perf_event_pending();
2565
2566         put_cpu_var(perf_pending_head);
2567 }
2568
2569 static int __perf_pending_run(void)
2570 {
2571         struct perf_pending_entry *list;
2572         int nr = 0;
2573
2574         list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2575         while (list != PENDING_TAIL) {
2576                 void (*func)(struct perf_pending_entry *);
2577                 struct perf_pending_entry *entry = list;
2578
2579                 list = list->next;
2580
2581                 func = entry->func;
2582                 entry->next = NULL;
2583                 /*
2584                  * Ensure we observe the unqueue before we issue the wakeup,
2585                  * so that we won't be waiting forever.
2586                  * -- see perf_not_pending().
2587                  */
2588                 smp_wmb();
2589
2590                 func(entry);
2591                 nr++;
2592         }
2593
2594         return nr;
2595 }
2596
2597 static inline int perf_not_pending(struct perf_event *event)
2598 {
2599         /*
2600          * If we flush on whatever cpu we run, there is a chance we don't
2601          * need to wait.
2602          */
2603         get_cpu();
2604         __perf_pending_run();
2605         put_cpu();
2606
2607         /*
2608          * Ensure we see the proper queue state before going to sleep
2609          * so that we do not miss the wakeup. -- see perf_pending_handle()
2610          */
2611         smp_rmb();
2612         return event->pending.next == NULL;
2613 }
2614
2615 static void perf_pending_sync(struct perf_event *event)
2616 {
2617         wait_event(event->waitq, perf_not_pending(event));
2618 }
2619
2620 void perf_event_do_pending(void)
2621 {
2622         __perf_pending_run();
2623 }
2624
2625 /*
2626  * Callchain support -- arch specific
2627  */
2628
2629 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2630 {
2631         return NULL;
2632 }
2633
2634 /*
2635  * Output
2636  */
2637 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2638                               unsigned long offset, unsigned long head)
2639 {
2640         unsigned long mask;
2641
2642         if (!data->writable)
2643                 return true;
2644
2645         mask = perf_data_size(data) - 1;
2646
2647         offset = (offset - tail) & mask;
2648         head   = (head   - tail) & mask;
2649
2650         if ((int)(head - offset) < 0)
2651                 return false;
2652
2653         return true;
2654 }
2655
2656 static void perf_output_wakeup(struct perf_output_handle *handle)
2657 {
2658         atomic_set(&handle->data->poll, POLL_IN);
2659
2660         if (handle->nmi) {
2661                 handle->event->pending_wakeup = 1;
2662                 perf_pending_queue(&handle->event->pending,
2663                                    perf_pending_event);
2664         } else
2665                 perf_event_wakeup(handle->event);
2666 }
2667
2668 /*
2669  * Curious locking construct.
2670  *
2671  * We need to ensure a later event_id doesn't publish a head when a former
2672  * event_id isn't done writing. However since we need to deal with NMIs we
2673  * cannot fully serialize things.
2674  *
2675  * What we do is serialize between CPUs so we only have to deal with NMI
2676  * nesting on a single CPU.
2677  *
2678  * We only publish the head (and generate a wakeup) when the outer-most
2679  * event_id completes.
2680  */
2681 static void perf_output_lock(struct perf_output_handle *handle)
2682 {
2683         struct perf_mmap_data *data = handle->data;
2684         int cur, cpu = get_cpu();
2685
2686         handle->locked = 0;
2687
2688         for (;;) {
2689                 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2690                 if (cur == -1) {
2691                         handle->locked = 1;
2692                         break;
2693                 }
2694                 if (cur == cpu)
2695                         break;
2696
2697                 cpu_relax();
2698         }
2699 }
2700
2701 static void perf_output_unlock(struct perf_output_handle *handle)
2702 {
2703         struct perf_mmap_data *data = handle->data;
2704         unsigned long head;
2705         int cpu;
2706
2707         data->done_head = data->head;
2708
2709         if (!handle->locked)
2710                 goto out;
2711
2712 again:
2713         /*
2714          * The xchg implies a full barrier that ensures all writes are done
2715          * before we publish the new head, matched by a rmb() in userspace when
2716          * reading this position.
2717          */
2718         while ((head = atomic_long_xchg(&data->done_head, 0)))
2719                 data->user_page->data_head = head;
2720
2721         /*
2722          * NMI can happen here, which means we can miss a done_head update.
2723          */
2724
2725         cpu = atomic_xchg(&data->lock, -1);
2726         WARN_ON_ONCE(cpu != smp_processor_id());
2727
2728         /*
2729          * Therefore we have to validate we did not indeed do so.
2730          */
2731         if (unlikely(atomic_long_read(&data->done_head))) {
2732                 /*
2733                  * Since we had it locked, we can lock it again.
2734                  */
2735                 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2736                         cpu_relax();
2737
2738                 goto again;
2739         }
2740
2741         if (atomic_xchg(&data->wakeup, 0))
2742                 perf_output_wakeup(handle);
2743 out:
2744         put_cpu();
2745 }
2746
2747 void perf_output_copy(struct perf_output_handle *handle,
2748                       const void *buf, unsigned int len)
2749 {
2750         unsigned int pages_mask;
2751         unsigned long offset;
2752         unsigned int size;
2753         void **pages;
2754
2755         offset          = handle->offset;
2756         pages_mask      = handle->data->nr_pages - 1;
2757         pages           = handle->data->data_pages;
2758
2759         do {
2760                 unsigned long page_offset;
2761                 unsigned long page_size;
2762                 int nr;
2763
2764                 nr          = (offset >> PAGE_SHIFT) & pages_mask;
2765                 page_size   = 1UL << (handle->data->data_order + PAGE_SHIFT);
2766                 page_offset = offset & (page_size - 1);
2767                 size        = min_t(unsigned int, page_size - page_offset, len);
2768
2769                 memcpy(pages[nr] + page_offset, buf, size);
2770
2771                 len         -= size;
2772                 buf         += size;
2773                 offset      += size;
2774         } while (len);
2775
2776         handle->offset = offset;
2777
2778         /*
2779          * Check we didn't copy past our reservation window, taking the
2780          * possible unsigned int wrap into account.
2781          */
2782         WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2783 }
2784
2785 int perf_output_begin(struct perf_output_handle *handle,
2786                       struct perf_event *event, unsigned int size,
2787                       int nmi, int sample)
2788 {
2789         struct perf_event *output_event;
2790         struct perf_mmap_data *data;
2791         unsigned long tail, offset, head;
2792         int have_lost;
2793         struct {
2794                 struct perf_event_header header;
2795                 u64                      id;
2796                 u64                      lost;
2797         } lost_event;
2798
2799         rcu_read_lock();
2800         /*
2801          * For inherited events we send all the output towards the parent.
2802          */
2803         if (event->parent)
2804                 event = event->parent;
2805
2806         output_event = rcu_dereference(event->output);
2807         if (output_event)
2808                 event = output_event;
2809
2810         data = rcu_dereference(event->data);
2811         if (!data)
2812                 goto out;
2813
2814         handle->data    = data;
2815         handle->event   = event;
2816         handle->nmi     = nmi;
2817         handle->sample  = sample;
2818
2819         if (!data->nr_pages)
2820                 goto fail;
2821
2822         have_lost = atomic_read(&data->lost);
2823         if (have_lost)
2824                 size += sizeof(lost_event);
2825
2826         perf_output_lock(handle);
2827
2828         do {
2829                 /*
2830                  * Userspace could choose to issue a mb() before updating the
2831                  * tail pointer. So that all reads will be completed before the
2832                  * write is issued.
2833                  */
2834                 tail = ACCESS_ONCE(data->user_page->data_tail);
2835                 smp_rmb();
2836                 offset = head = atomic_long_read(&data->head);
2837                 head += size;
2838                 if (unlikely(!perf_output_space(data, tail, offset, head)))
2839                         goto fail;
2840         } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2841
2842         handle->offset  = offset;
2843         handle->head    = head;
2844
2845         if (head - tail > data->watermark)
2846                 atomic_set(&data->wakeup, 1);
2847
2848         if (have_lost) {
2849                 lost_event.header.type = PERF_RECORD_LOST;
2850                 lost_event.header.misc = 0;
2851                 lost_event.header.size = sizeof(lost_event);
2852                 lost_event.id          = event->id;
2853                 lost_event.lost        = atomic_xchg(&data->lost, 0);
2854
2855                 perf_output_put(handle, lost_event);
2856         }
2857
2858         return 0;
2859
2860 fail:
2861         atomic_inc(&data->lost);
2862         perf_output_unlock(handle);
2863 out:
2864         rcu_read_unlock();
2865
2866         return -ENOSPC;
2867 }
2868
2869 void perf_output_end(struct perf_output_handle *handle)
2870 {
2871         struct perf_event *event = handle->event;
2872         struct perf_mmap_data *data = handle->data;
2873
2874         int wakeup_events = event->attr.wakeup_events;
2875
2876         if (handle->sample && wakeup_events) {
2877                 int events = atomic_inc_return(&data->events);
2878                 if (events >= wakeup_events) {
2879                         atomic_sub(wakeup_events, &data->events);
2880                         atomic_set(&data->wakeup, 1);
2881                 }
2882         }
2883
2884         perf_output_unlock(handle);
2885         rcu_read_unlock();
2886 }
2887
2888 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
2889 {
2890         /*
2891          * only top level events have the pid namespace they were created in
2892          */
2893         if (event->parent)
2894                 event = event->parent;
2895
2896         return task_tgid_nr_ns(p, event->ns);
2897 }
2898
2899 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
2900 {
2901         /*
2902          * only top level events have the pid namespace they were created in
2903          */
2904         if (event->parent)
2905                 event = event->parent;
2906
2907         return task_pid_nr_ns(p, event->ns);
2908 }
2909
2910 static void perf_output_read_one(struct perf_output_handle *handle,
2911                                  struct perf_event *event)
2912 {
2913         u64 read_format = event->attr.read_format;
2914         u64 values[4];
2915         int n = 0;
2916
2917         values[n++] = atomic64_read(&event->count);
2918         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2919                 values[n++] = event->total_time_enabled +
2920                         atomic64_read(&event->child_total_time_enabled);
2921         }
2922         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2923                 values[n++] = event->total_time_running +
2924                         atomic64_read(&event->child_total_time_running);
2925         }
2926         if (read_format & PERF_FORMAT_ID)
2927                 values[n++] = primary_event_id(event);
2928
2929         perf_output_copy(handle, values, n * sizeof(u64));
2930 }
2931
2932 /*
2933  * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
2934  */
2935 static void perf_output_read_group(struct perf_output_handle *handle,
2936                             struct perf_event *event)
2937 {
2938         struct perf_event *leader = event->group_leader, *sub;
2939         u64 read_format = event->attr.read_format;
2940         u64 values[5];
2941         int n = 0;
2942
2943         values[n++] = 1 + leader->nr_siblings;
2944
2945         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2946                 values[n++] = leader->total_time_enabled;
2947
2948         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2949                 values[n++] = leader->total_time_running;
2950
2951         if (leader != event)
2952                 leader->pmu->read(leader);
2953
2954         values[n++] = atomic64_read(&leader->count);
2955         if (read_format & PERF_FORMAT_ID)
2956                 values[n++] = primary_event_id(leader);
2957
2958         perf_output_copy(handle, values, n * sizeof(u64));
2959
2960         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2961                 n = 0;
2962
2963                 if (sub != event)
2964                         sub->pmu->read(sub);
2965
2966                 values[n++] = atomic64_read(&sub->count);
2967                 if (read_format & PERF_FORMAT_ID)
2968                         values[n++] = primary_event_id(sub);
2969
2970                 perf_output_copy(handle, values, n * sizeof(u64));
2971         }
2972 }
2973
2974 static void perf_output_read(struct perf_output_handle *handle,
2975                              struct perf_event *event)
2976 {
2977         if (event->attr.read_format & PERF_FORMAT_GROUP)
2978                 perf_output_read_group(handle, event);
2979         else
2980                 perf_output_read_one(handle, event);
2981 }
2982
2983 void perf_output_sample(struct perf_output_handle *handle,
2984                         struct perf_event_header *header,
2985                         struct perf_sample_data *data,
2986                         struct perf_event *event)
2987 {
2988         u64 sample_type = data->type;
2989
2990         perf_output_put(handle, *header);
2991
2992         if (sample_type & PERF_SAMPLE_IP)
2993                 perf_output_put(handle, data->ip);
2994
2995         if (sample_type & PERF_SAMPLE_TID)
2996                 perf_output_put(handle, data->tid_entry);
2997
2998         if (sample_type & PERF_SAMPLE_TIME)
2999                 perf_output_put(handle, data->time);
3000
3001         if (sample_type & PERF_SAMPLE_ADDR)
3002                 perf_output_put(handle, data->addr);
3003
3004         if (sample_type & PERF_SAMPLE_ID)
3005                 perf_output_put(handle, data->id);
3006
3007         if (sample_type & PERF_SAMPLE_STREAM_ID)
3008                 perf_output_put(handle, data->stream_id);
3009
3010         if (sample_type & PERF_SAMPLE_CPU)
3011                 perf_output_put(handle, data->cpu_entry);
3012
3013         if (sample_type & PERF_SAMPLE_PERIOD)
3014                 perf_output_put(handle, data->period);
3015
3016         if (sample_type & PERF_SAMPLE_READ)
3017                 perf_output_read(handle, event);
3018
3019         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3020                 if (data->callchain) {
3021                         int size = 1;
3022
3023                         if (data->callchain)
3024                                 size += data->callchain->nr;
3025
3026                         size *= sizeof(u64);
3027
3028                         perf_output_copy(handle, data->callchain, size);
3029                 } else {
3030                         u64 nr = 0;
3031                         perf_output_put(handle, nr);
3032                 }
3033         }
3034
3035         if (sample_type & PERF_SAMPLE_RAW) {
3036                 if (data->raw) {
3037                         perf_output_put(handle, data->raw->size);
3038                         perf_output_copy(handle, data->raw->data,
3039                                          data->raw->size);
3040                 } else {
3041                         struct {
3042                                 u32     size;
3043                                 u32     data;
3044                         } raw = {
3045                                 .size = sizeof(u32),
3046                                 .data = 0,
3047                         };
3048                         perf_output_put(handle, raw);
3049                 }
3050         }
3051 }
3052
3053 void perf_prepare_sample(struct perf_event_header *header,
3054                          struct perf_sample_data *data,
3055                          struct perf_event *event,
3056                          struct pt_regs *regs)
3057 {
3058         u64 sample_type = event->attr.sample_type;
3059
3060         data->type = sample_type;
3061
3062         header->type = PERF_RECORD_SAMPLE;
3063         header->size = sizeof(*header);
3064
3065         header->misc = 0;
3066         header->misc |= perf_misc_flags(regs);
3067
3068         if (sample_type & PERF_SAMPLE_IP) {
3069                 data->ip = perf_instruction_pointer(regs);
3070
3071                 header->size += sizeof(data->ip);
3072         }
3073
3074         if (sample_type & PERF_SAMPLE_TID) {
3075                 /* namespace issues */
3076                 data->tid_entry.pid = perf_event_pid(event, current);
3077                 data->tid_entry.tid = perf_event_tid(event, current);
3078
3079                 header->size += sizeof(data->tid_entry);
3080         }
3081
3082         if (sample_type & PERF_SAMPLE_TIME) {
3083                 data->time = perf_clock();
3084
3085                 header->size += sizeof(data->time);
3086         }
3087
3088         if (sample_type & PERF_SAMPLE_ADDR)
3089                 header->size += sizeof(data->addr);
3090
3091         if (sample_type & PERF_SAMPLE_ID) {
3092                 data->id = primary_event_id(event);
3093
3094                 header->size += sizeof(data->id);
3095         }
3096
3097         if (sample_type & PERF_SAMPLE_STREAM_ID) {
3098                 data->stream_id = event->id;
3099
3100                 header->size += sizeof(data->stream_id);
3101         }
3102
3103         if (sample_type & PERF_SAMPLE_CPU) {
3104                 data->cpu_entry.cpu             = raw_smp_processor_id();
3105                 data->cpu_entry.reserved        = 0;
3106
3107                 header->size += sizeof(data->cpu_entry);
3108         }
3109
3110         if (sample_type & PERF_SAMPLE_PERIOD)
3111                 header->size += sizeof(data->period);
3112
3113         if (sample_type & PERF_SAMPLE_READ)
3114                 header->size += perf_event_read_size(event);
3115
3116         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3117                 int size = 1;
3118
3119                 data->callchain = perf_callchain(regs);
3120
3121                 if (data->callchain)
3122                         size += data->callchain->nr;
3123
3124                 header->size += size * sizeof(u64);
3125         }
3126
3127         if (sample_type & PERF_SAMPLE_RAW) {
3128                 int size = sizeof(u32);
3129
3130                 if (data->raw)
3131                         size += data->raw->size;
3132                 else
3133                         size += sizeof(u32);
3134
3135                 WARN_ON_ONCE(size & (sizeof(u64)-1));
3136                 header->size += size;
3137         }
3138 }
3139
3140 static void perf_event_output(struct perf_event *event, int nmi,
3141                                 struct perf_sample_data *data,
3142                                 struct pt_regs *regs)
3143 {
3144         struct perf_output_handle handle;
3145         struct perf_event_header header;
3146
3147         perf_prepare_sample(&header, data, event, regs);
3148
3149         if (perf_output_begin(&handle, event, header.size, nmi, 1))
3150                 return;
3151
3152         perf_output_sample(&handle, &header, data, event);
3153
3154         perf_output_end(&handle);
3155 }
3156
3157 /*
3158  * read event_id
3159  */
3160
3161 struct perf_read_event {
3162         struct perf_event_header        header;
3163
3164         u32                             pid;
3165         u32                             tid;
3166 };
3167
3168 static void
3169 perf_event_read_event(struct perf_event *event,
3170                         struct task_struct *task)
3171 {
3172         struct perf_output_handle handle;
3173         struct perf_read_event read_event = {
3174                 .header = {
3175                         .type = PERF_RECORD_READ,
3176                         .misc = 0,
3177                         .size = sizeof(read_event) + perf_event_read_size(event),
3178                 },
3179                 .pid = perf_event_pid(event, task),
3180                 .tid = perf_event_tid(event, task),
3181         };
3182         int ret;
3183
3184         ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3185         if (ret)
3186                 return;
3187
3188         perf_output_put(&handle, read_event);
3189         perf_output_read(&handle, event);
3190
3191         perf_output_end(&handle);
3192 }
3193
3194 /*
3195  * task tracking -- fork/exit
3196  *
3197  * enabled by: attr.comm | attr.mmap | attr.task
3198  */
3199
3200 struct perf_task_event {
3201         struct task_struct              *task;
3202         struct perf_event_context       *task_ctx;
3203
3204         struct {
3205                 struct perf_event_header        header;
3206
3207                 u32                             pid;
3208                 u32                             ppid;
3209                 u32                             tid;
3210                 u32                             ptid;
3211                 u64                             time;
3212         } event_id;
3213 };
3214
3215 static void perf_event_task_output(struct perf_event *event,
3216                                      struct perf_task_event *task_event)
3217 {
3218         struct perf_output_handle handle;
3219         int size;
3220         struct task_struct *task = task_event->task;
3221         int ret;
3222
3223         size  = task_event->event_id.header.size;
3224         ret = perf_output_begin(&handle, event, size, 0, 0);
3225
3226         if (ret)
3227                 return;
3228
3229         task_event->event_id.pid = perf_event_pid(event, task);
3230         task_event->event_id.ppid = perf_event_pid(event, current);
3231
3232         task_event->event_id.tid = perf_event_tid(event, task);
3233         task_event->event_id.ptid = perf_event_tid(event, current);
3234
3235         task_event->event_id.time = perf_clock();
3236
3237         perf_output_put(&handle, task_event->event_id);
3238
3239         perf_output_end(&handle);
3240 }
3241
3242 static int perf_event_task_match(struct perf_event *event)
3243 {
3244         if (event->attr.comm || event->attr.mmap || event->attr.task)
3245                 return 1;
3246
3247         return 0;
3248 }
3249
3250 static void perf_event_task_ctx(struct perf_event_context *ctx,
3251                                   struct perf_task_event *task_event)
3252 {
3253         struct perf_event *event;
3254
3255         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3256                 if (perf_event_task_match(event))
3257                         perf_event_task_output(event, task_event);
3258         }
3259 }
3260
3261 static void perf_event_task_event(struct perf_task_event *task_event)
3262 {
3263         struct perf_cpu_context *cpuctx;
3264         struct perf_event_context *ctx = task_event->task_ctx;
3265
3266         rcu_read_lock();
3267         cpuctx = &get_cpu_var(perf_cpu_context);
3268         perf_event_task_ctx(&cpuctx->ctx, task_event);
3269         put_cpu_var(perf_cpu_context);
3270
3271         if (!ctx)
3272                 ctx = rcu_dereference(task_event->task->perf_event_ctxp);
3273         if (ctx)
3274                 perf_event_task_ctx(ctx, task_event);
3275         rcu_read_unlock();
3276 }
3277
3278 static void perf_event_task(struct task_struct *task,
3279                               struct perf_event_context *task_ctx,
3280                               int new)
3281 {
3282         struct perf_task_event task_event;
3283
3284         if (!atomic_read(&nr_comm_events) &&
3285             !atomic_read(&nr_mmap_events) &&
3286             !atomic_read(&nr_task_events))
3287                 return;
3288
3289         task_event = (struct perf_task_event){
3290                 .task     = task,
3291                 .task_ctx = task_ctx,
3292                 .event_id    = {
3293                         .header = {
3294                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3295                                 .misc = 0,
3296                                 .size = sizeof(task_event.event_id),
3297                         },
3298                         /* .pid  */
3299                         /* .ppid */
3300                         /* .tid  */
3301                         /* .ptid */
3302                 },
3303         };
3304
3305         perf_event_task_event(&task_event);
3306 }
3307
3308 void perf_event_fork(struct task_struct *task)
3309 {
3310         perf_event_task(task, NULL, 1);
3311 }
3312
3313 /*
3314  * comm tracking
3315  */
3316
3317 struct perf_comm_event {
3318         struct task_struct      *task;
3319         char                    *comm;
3320         int                     comm_size;
3321
3322         struct {
3323                 struct perf_event_header        header;
3324
3325                 u32                             pid;
3326                 u32                             tid;
3327         } event_id;
3328 };
3329
3330 static void perf_event_comm_output(struct perf_event *event,
3331                                      struct perf_comm_event *comm_event)
3332 {
3333         struct perf_output_handle handle;
3334         int size = comm_event->event_id.header.size;
3335         int ret = perf_output_begin(&handle, event, size, 0, 0);
3336
3337         if (ret)
3338                 return;
3339
3340         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3341         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3342
3343         perf_output_put(&handle, comm_event->event_id);
3344         perf_output_copy(&handle, comm_event->comm,
3345                                    comm_event->comm_size);
3346         perf_output_end(&handle);
3347 }
3348
3349 static int perf_event_comm_match(struct perf_event *event)
3350 {
3351         if (event->attr.comm)
3352                 return 1;
3353
3354         return 0;
3355 }
3356
3357 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3358                                   struct perf_comm_event *comm_event)
3359 {
3360         struct perf_event *event;
3361
3362         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3363                 if (perf_event_comm_match(event))
3364                         perf_event_comm_output(event, comm_event);
3365         }
3366 }
3367
3368 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3369 {
3370         struct perf_cpu_context *cpuctx;
3371         struct perf_event_context *ctx;
3372         unsigned int size;
3373         char comm[TASK_COMM_LEN];
3374
3375         memset(comm, 0, sizeof(comm));
3376         strncpy(comm, comm_event->task->comm, sizeof(comm));
3377         size = ALIGN(strlen(comm)+1, sizeof(u64));
3378
3379         comm_event->comm = comm;
3380         comm_event->comm_size = size;
3381
3382         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3383
3384         rcu_read_lock();
3385         cpuctx = &get_cpu_var(perf_cpu_context);
3386         perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3387         put_cpu_var(perf_cpu_context);
3388
3389         /*
3390          * doesn't really matter which of the child contexts the
3391          * events ends up in.
3392          */
3393         ctx = rcu_dereference(current->perf_event_ctxp);
3394         if (ctx)
3395                 perf_event_comm_ctx(ctx, comm_event);
3396         rcu_read_unlock();
3397 }
3398
3399 void perf_event_comm(struct task_struct *task)
3400 {
3401         struct perf_comm_event comm_event;
3402
3403         if (task->perf_event_ctxp)
3404                 perf_event_enable_on_exec(task);
3405
3406         if (!atomic_read(&nr_comm_events))
3407                 return;
3408
3409         comm_event = (struct perf_comm_event){
3410                 .task   = task,
3411                 /* .comm      */
3412                 /* .comm_size */
3413                 .event_id  = {
3414                         .header = {
3415                                 .type = PERF_RECORD_COMM,
3416                                 .misc = 0,
3417                                 /* .size */
3418                         },
3419                         /* .pid */
3420                         /* .tid */
3421                 },
3422         };
3423
3424         perf_event_comm_event(&comm_event);
3425 }
3426
3427 /*
3428  * mmap tracking
3429  */
3430
3431 struct perf_mmap_event {
3432         struct vm_area_struct   *vma;
3433
3434         const char              *file_name;
3435         int                     file_size;
3436
3437         struct {
3438                 struct perf_event_header        header;
3439
3440                 u32                             pid;
3441                 u32                             tid;
3442                 u64                             start;
3443                 u64                             len;
3444                 u64                             pgoff;
3445         } event_id;
3446 };
3447
3448 static void perf_event_mmap_output(struct perf_event *event,
3449                                      struct perf_mmap_event *mmap_event)
3450 {
3451         struct perf_output_handle handle;
3452         int size = mmap_event->event_id.header.size;
3453         int ret = perf_output_begin(&handle, event, size, 0, 0);
3454
3455         if (ret)
3456                 return;
3457
3458         mmap_event->event_id.pid = perf_event_pid(event, current);
3459         mmap_event->event_id.tid = perf_event_tid(event, current);
3460
3461         perf_output_put(&handle, mmap_event->event_id);
3462         perf_output_copy(&handle, mmap_event->file_name,
3463                                    mmap_event->file_size);
3464         perf_output_end(&handle);
3465 }
3466
3467 static int perf_event_mmap_match(struct perf_event *event,
3468                                    struct perf_mmap_event *mmap_event)
3469 {
3470         if (event->attr.mmap)
3471                 return 1;
3472
3473         return 0;
3474 }
3475
3476 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3477                                   struct perf_mmap_event *mmap_event)
3478 {
3479         struct perf_event *event;
3480
3481         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3482                 if (perf_event_mmap_match(event, mmap_event))
3483                         perf_event_mmap_output(event, mmap_event);
3484         }
3485 }
3486
3487 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3488 {
3489         struct perf_cpu_context *cpuctx;
3490         struct perf_event_context *ctx;
3491         struct vm_area_struct *vma = mmap_event->vma;
3492         struct file *file = vma->vm_file;
3493         unsigned int size;
3494         char tmp[16];
3495         char *buf = NULL;
3496         const char *name;
3497
3498         memset(tmp, 0, sizeof(tmp));
3499
3500         if (file) {
3501                 /*
3502                  * d_path works from the end of the buffer backwards, so we
3503                  * need to add enough zero bytes after the string to handle
3504                  * the 64bit alignment we do later.
3505                  */
3506                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3507                 if (!buf) {
3508                         name = strncpy(tmp, "//enomem", sizeof(tmp));
3509                         goto got_name;
3510                 }
3511                 name = d_path(&file->f_path, buf, PATH_MAX);
3512                 if (IS_ERR(name)) {
3513                         name = strncpy(tmp, "//toolong", sizeof(tmp));
3514                         goto got_name;
3515                 }
3516         } else {
3517                 if (arch_vma_name(mmap_event->vma)) {
3518                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3519                                        sizeof(tmp));
3520                         goto got_name;
3521                 }
3522
3523                 if (!vma->vm_mm) {
3524                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
3525                         goto got_name;
3526                 }
3527
3528                 name = strncpy(tmp, "//anon", sizeof(tmp));
3529                 goto got_name;
3530         }
3531
3532 got_name:
3533         size = ALIGN(strlen(name)+1, sizeof(u64));
3534
3535         mmap_event->file_name = name;
3536         mmap_event->file_size = size;
3537
3538         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3539
3540         rcu_read_lock();
3541         cpuctx = &get_cpu_var(perf_cpu_context);
3542         perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3543         put_cpu_var(perf_cpu_context);
3544
3545         /*
3546          * doesn't really matter which of the child contexts the
3547          * events ends up in.
3548          */
3549         ctx = rcu_dereference(current->perf_event_ctxp);
3550         if (ctx)
3551                 perf_event_mmap_ctx(ctx, mmap_event);
3552         rcu_read_unlock();
3553
3554         kfree(buf);
3555 }
3556
3557 void __perf_event_mmap(struct vm_area_struct *vma)
3558 {
3559         struct perf_mmap_event mmap_event;
3560
3561         if (!atomic_read(&nr_mmap_events))
3562                 return;
3563
3564         mmap_event = (struct perf_mmap_event){
3565                 .vma    = vma,
3566                 /* .file_name */
3567                 /* .file_size */
3568                 .event_id  = {
3569                         .header = {
3570                                 .type = PERF_RECORD_MMAP,
3571                                 .misc = 0,
3572                                 /* .size */
3573                         },
3574                         /* .pid */
3575                         /* .tid */
3576                         .start  = vma->vm_start,
3577                         .len    = vma->vm_end - vma->vm_start,
3578                         .pgoff  = vma->vm_pgoff,
3579                 },
3580         };
3581
3582         perf_event_mmap_event(&mmap_event);
3583 }
3584
3585 /*
3586  * IRQ throttle logging
3587  */
3588
3589 static void perf_log_throttle(struct perf_event *event, int enable)
3590 {
3591         struct perf_output_handle handle;
3592         int ret;
3593
3594         struct {
3595                 struct perf_event_header        header;
3596                 u64                             time;
3597                 u64                             id;
3598                 u64                             stream_id;
3599         } throttle_event = {
3600                 .header = {
3601                         .type = PERF_RECORD_THROTTLE,
3602                         .misc = 0,
3603                         .size = sizeof(throttle_event),
3604                 },
3605                 .time           = perf_clock(),
3606                 .id             = primary_event_id(event),
3607                 .stream_id      = event->id,
3608         };
3609
3610         if (enable)
3611                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3612
3613         ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3614         if (ret)
3615                 return;
3616
3617         perf_output_put(&handle, throttle_event);
3618         perf_output_end(&handle);
3619 }
3620
3621 /*
3622  * Generic event overflow handling, sampling.
3623  */
3624
3625 static int __perf_event_overflow(struct perf_event *event, int nmi,
3626                                    int throttle, struct perf_sample_data *data,
3627                                    struct pt_regs *regs)
3628 {
3629         int events = atomic_read(&event->event_limit);
3630         struct hw_perf_event *hwc = &event->hw;
3631         int ret = 0;
3632
3633         throttle = (throttle && event->pmu->unthrottle != NULL);
3634
3635         if (!throttle) {
3636                 hwc->interrupts++;
3637         } else {
3638                 if (hwc->interrupts != MAX_INTERRUPTS) {
3639                         hwc->interrupts++;
3640                         if (HZ * hwc->interrupts >
3641                                         (u64)sysctl_perf_event_sample_rate) {
3642                                 hwc->interrupts = MAX_INTERRUPTS;
3643                                 perf_log_throttle(event, 0);
3644                                 ret = 1;
3645                         }
3646                 } else {
3647                         /*
3648                          * Keep re-disabling events even though on the previous
3649                          * pass we disabled it - just in case we raced with a
3650                          * sched-in and the event got enabled again:
3651                          */
3652                         ret = 1;
3653                 }
3654         }
3655
3656         if (event->attr.freq) {
3657                 u64 now = perf_clock();
3658                 s64 delta = now - hwc->freq_stamp;
3659
3660                 hwc->freq_stamp = now;
3661
3662                 if (delta > 0 && delta < TICK_NSEC)
3663                         perf_adjust_period(event, NSEC_PER_SEC / (int)delta);
3664         }
3665
3666         /*
3667          * XXX event_limit might not quite work as expected on inherited
3668          * events
3669          */
3670
3671         event->pending_kill = POLL_IN;
3672         if (events && atomic_dec_and_test(&event->event_limit)) {
3673                 ret = 1;
3674                 event->pending_kill = POLL_HUP;
3675                 if (nmi) {
3676                         event->pending_disable = 1;
3677                         perf_pending_queue(&event->pending,
3678                                            perf_pending_event);
3679                 } else
3680                         perf_event_disable(event);
3681         }
3682
3683         if (event->overflow_handler)
3684                 event->overflow_handler(event, nmi, data, regs);
3685         else
3686                 perf_event_output(event, nmi, data, regs);
3687
3688         return ret;
3689 }
3690
3691 int perf_event_overflow(struct perf_event *event, int nmi,
3692                           struct perf_sample_data *data,
3693                           struct pt_regs *regs)
3694 {
3695         return __perf_event_overflow(event, nmi, 1, data, regs);
3696 }
3697
3698 /*
3699  * Generic software event infrastructure
3700  */
3701
3702 /*
3703  * We directly increment event->count and keep a second value in
3704  * event->hw.period_left to count intervals. This period event
3705  * is kept in the range [-sample_period, 0] so that we can use the
3706  * sign as trigger.
3707  */
3708
3709 static u64 perf_swevent_set_period(struct perf_event *event)
3710 {
3711         struct hw_perf_event *hwc = &event->hw;
3712         u64 period = hwc->last_period;
3713         u64 nr, offset;
3714         s64 old, val;
3715
3716         hwc->last_period = hwc->sample_period;
3717
3718 again:
3719         old = val = atomic64_read(&hwc->period_left);
3720         if (val < 0)
3721                 return 0;
3722
3723         nr = div64_u64(period + val, period);
3724         offset = nr * period;
3725         val -= offset;
3726         if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3727                 goto again;
3728
3729         return nr;
3730 }
3731
3732 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3733                                     int nmi, struct perf_sample_data *data,
3734                                     struct pt_regs *regs)
3735 {
3736         struct hw_perf_event *hwc = &event->hw;
3737         int throttle = 0;
3738
3739         data->period = event->hw.last_period;
3740         if (!overflow)
3741                 overflow = perf_swevent_set_period(event);
3742
3743         if (hwc->interrupts == MAX_INTERRUPTS)
3744                 return;
3745
3746         for (; overflow; overflow--) {
3747                 if (__perf_event_overflow(event, nmi, throttle,
3748                                             data, regs)) {
3749                         /*
3750                          * We inhibit the overflow from happening when
3751                          * hwc->interrupts == MAX_INTERRUPTS.
3752                          */
3753                         break;
3754                 }
3755                 throttle = 1;
3756         }
3757 }
3758
3759 static void perf_swevent_unthrottle(struct perf_event *event)
3760 {
3761         /*
3762          * Nothing to do, we already reset hwc->interrupts.
3763          */
3764 }
3765
3766 static void perf_swevent_add(struct perf_event *event, u64 nr,
3767                                int nmi, struct perf_sample_data *data,
3768                                struct pt_regs *regs)
3769 {
3770         struct hw_perf_event *hwc = &event->hw;
3771
3772         atomic64_add(nr, &event->count);
3773
3774         if (!regs)
3775                 return;
3776
3777         if (!hwc->sample_period)
3778                 return;
3779
3780         if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3781                 return perf_swevent_overflow(event, 1, nmi, data, regs);
3782
3783         if (atomic64_add_negative(nr, &hwc->period_left))
3784                 return;
3785
3786         perf_swevent_overflow(event, 0, nmi, data, regs);
3787 }
3788
3789 static int perf_swevent_is_counting(struct perf_event *event)
3790 {
3791         /*
3792          * The event is active, we're good!
3793          */
3794         if (event->state == PERF_EVENT_STATE_ACTIVE)
3795                 return 1;
3796
3797         /*
3798          * The event is off/error, not counting.
3799          */
3800         if (event->state != PERF_EVENT_STATE_INACTIVE)
3801                 return 0;
3802
3803         /*
3804          * The event is inactive, if the context is active
3805          * we're part of a group that didn't make it on the 'pmu',
3806          * not counting.
3807          */
3808         if (event->ctx->is_active)
3809                 return 0;
3810
3811         /*
3812          * We're inactive and the context is too, this means the
3813          * task is scheduled out, we're counting events that happen
3814          * to us, like migration events.
3815          */
3816         return 1;
3817 }
3818
3819 static int perf_tp_event_match(struct perf_event *event,
3820                                 struct perf_sample_data *data);
3821
3822 static int perf_swevent_match(struct perf_event *event,
3823                                 enum perf_type_id type,
3824                                 u32 event_id,
3825                                 struct perf_sample_data *data,
3826                                 struct pt_regs *regs)
3827 {
3828         if (!perf_swevent_is_counting(event))
3829                 return 0;
3830
3831         if (event->attr.type != type)
3832                 return 0;
3833         if (event->attr.config != event_id)
3834                 return 0;
3835
3836         if (regs) {
3837                 if (event->attr.exclude_user && user_mode(regs))
3838                         return 0;
3839
3840                 if (event->attr.exclude_kernel && !user_mode(regs))
3841                         return 0;
3842         }
3843
3844         if (event->attr.type == PERF_TYPE_TRACEPOINT &&
3845             !perf_tp_event_match(event, data))
3846                 return 0;
3847
3848         return 1;
3849 }
3850
3851 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
3852                                      enum perf_type_id type,
3853                                      u32 event_id, u64 nr, int nmi,
3854                                      struct perf_sample_data *data,
3855                                      struct pt_regs *regs)
3856 {
3857         struct perf_event *event;
3858
3859         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3860                 if (perf_swevent_match(event, type, event_id, data, regs))
3861                         perf_swevent_add(event, nr, nmi, data, regs);
3862         }
3863 }
3864
3865 static int *perf_swevent_recursion_context(struct perf_cpu_context *cpuctx)
3866 {
3867         if (in_nmi())
3868                 return &cpuctx->recursion[3];
3869
3870         if (in_irq())
3871                 return &cpuctx->recursion[2];
3872
3873         if (in_softirq())
3874                 return &cpuctx->recursion[1];
3875
3876         return &cpuctx->recursion[0];
3877 }
3878
3879 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
3880                                     u64 nr, int nmi,
3881                                     struct perf_sample_data *data,
3882                                     struct pt_regs *regs)
3883 {
3884         struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3885         int *recursion = perf_swevent_recursion_context(cpuctx);
3886         struct perf_event_context *ctx;
3887
3888         if (*recursion)
3889                 goto out;
3890
3891         (*recursion)++;
3892         barrier();
3893
3894         rcu_read_lock();
3895         perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
3896                                  nr, nmi, data, regs);
3897         /*
3898          * doesn't really matter which of the child contexts the
3899          * events ends up in.
3900          */
3901         ctx = rcu_dereference(current->perf_event_ctxp);
3902         if (ctx)
3903                 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
3904         rcu_read_unlock();
3905
3906         barrier();
3907         (*recursion)--;
3908
3909 out:
3910         put_cpu_var(perf_cpu_context);
3911 }
3912
3913 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
3914                             struct pt_regs *regs, u64 addr)
3915 {
3916         struct perf_sample_data data = {
3917                 .addr = addr,
3918         };
3919
3920         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi,
3921                                 &data, regs);
3922 }
3923
3924 static void perf_swevent_read(struct perf_event *event)
3925 {
3926 }
3927
3928 static int perf_swevent_enable(struct perf_event *event)
3929 {
3930         struct hw_perf_event *hwc = &event->hw;
3931
3932         if (hwc->sample_period) {
3933                 hwc->last_period = hwc->sample_period;
3934                 perf_swevent_set_period(event);
3935         }
3936         return 0;
3937 }
3938
3939 static void perf_swevent_disable(struct perf_event *event)
3940 {
3941 }
3942
3943 static const struct pmu perf_ops_generic = {
3944         .enable         = perf_swevent_enable,
3945         .disable        = perf_swevent_disable,
3946         .read           = perf_swevent_read,
3947         .unthrottle     = perf_swevent_unthrottle,
3948 };
3949
3950 /*
3951  * hrtimer based swevent callback
3952  */
3953
3954 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
3955 {
3956         enum hrtimer_restart ret = HRTIMER_RESTART;
3957         struct perf_sample_data data;
3958         struct pt_regs *regs;
3959         struct perf_event *event;
3960         u64 period;
3961
3962         event   = container_of(hrtimer, struct perf_event, hw.hrtimer);
3963         event->pmu->read(event);
3964
3965         data.addr = 0;
3966         regs = get_irq_regs();
3967         /*
3968          * In case we exclude kernel IPs or are somehow not in interrupt
3969          * context, provide the next best thing, the user IP.
3970          */
3971         if ((event->attr.exclude_kernel || !regs) &&
3972                         !event->attr.exclude_user)
3973                 regs = task_pt_regs(current);
3974
3975         if (regs) {
3976                 if (!(event->attr.exclude_idle && current->pid == 0))
3977                         if (perf_event_overflow(event, 0, &data, regs))
3978                                 ret = HRTIMER_NORESTART;
3979         }
3980
3981         period = max_t(u64, 10000, event->hw.sample_period);
3982         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
3983
3984         return ret;
3985 }
3986
3987 static void perf_swevent_start_hrtimer(struct perf_event *event)
3988 {
3989         struct hw_perf_event *hwc = &event->hw;
3990
3991         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3992         hwc->hrtimer.function = perf_swevent_hrtimer;
3993         if (hwc->sample_period) {
3994                 u64 period;
3995
3996                 if (hwc->remaining) {
3997                         if (hwc->remaining < 0)
3998                                 period = 10000;
3999                         else
4000                                 period = hwc->remaining;
4001                         hwc->remaining = 0;
4002                 } else {
4003                         period = max_t(u64, 10000, hwc->sample_period);
4004                 }
4005                 __hrtimer_start_range_ns(&hwc->hrtimer,
4006                                 ns_to_ktime(period), 0,
4007                                 HRTIMER_MODE_REL, 0);
4008         }
4009 }
4010
4011 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4012 {
4013         struct hw_perf_event *hwc = &event->hw;
4014
4015         if (hwc->sample_period) {
4016                 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4017                 hwc->remaining = ktime_to_ns(remaining);
4018
4019                 hrtimer_cancel(&hwc->hrtimer);
4020         }
4021 }
4022
4023 /*
4024  * Software event: cpu wall time clock
4025  */
4026
4027 static void cpu_clock_perf_event_update(struct perf_event *event)
4028 {
4029         int cpu = raw_smp_processor_id();
4030         s64 prev;
4031         u64 now;
4032
4033         now = cpu_clock(cpu);
4034         prev = atomic64_read(&event->hw.prev_count);
4035         atomic64_set(&event->hw.prev_count, now);
4036         atomic64_add(now - prev, &event->count);
4037 }
4038
4039 static int cpu_clock_perf_event_enable(struct perf_event *event)
4040 {
4041         struct hw_perf_event *hwc = &event->hw;
4042         int cpu = raw_smp_processor_id();
4043
4044         atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4045         perf_swevent_start_hrtimer(event);
4046
4047         return 0;
4048 }
4049
4050 static void cpu_clock_perf_event_disable(struct perf_event *event)
4051 {
4052         perf_swevent_cancel_hrtimer(event);
4053         cpu_clock_perf_event_update(event);
4054 }
4055
4056 static void cpu_clock_perf_event_read(struct perf_event *event)
4057 {
4058         cpu_clock_perf_event_update(event);
4059 }
4060
4061 static const struct pmu perf_ops_cpu_clock = {
4062         .enable         = cpu_clock_perf_event_enable,
4063         .disable        = cpu_clock_perf_event_disable,
4064         .read           = cpu_clock_perf_event_read,
4065 };
4066
4067 /*
4068  * Software event: task time clock
4069  */
4070
4071 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4072 {
4073         u64 prev;
4074         s64 delta;
4075
4076         prev = atomic64_xchg(&event->hw.prev_count, now);
4077         delta = now - prev;
4078         atomic64_add(delta, &event->count);
4079 }
4080
4081 static int task_clock_perf_event_enable(struct perf_event *event)
4082 {
4083         struct hw_perf_event *hwc = &event->hw;
4084         u64 now;
4085
4086         now = event->ctx->time;
4087
4088         atomic64_set(&hwc->prev_count, now);
4089
4090         perf_swevent_start_hrtimer(event);
4091
4092         return 0;
4093 }
4094
4095 static void task_clock_perf_event_disable(struct perf_event *event)
4096 {
4097         perf_swevent_cancel_hrtimer(event);
4098         task_clock_perf_event_update(event, event->ctx->time);
4099
4100 }
4101
4102 static void task_clock_perf_event_read(struct perf_event *event)
4103 {
4104         u64 time;
4105
4106         if (!in_nmi()) {
4107                 update_context_time(event->ctx);
4108                 time = event->ctx->time;
4109         } else {
4110                 u64 now = perf_clock();
4111                 u64 delta = now - event->ctx->timestamp;
4112                 time = event->ctx->time + delta;
4113         }
4114
4115         task_clock_perf_event_update(event, time);
4116 }
4117
4118 static const struct pmu perf_ops_task_clock = {
4119         .enable         = task_clock_perf_event_enable,
4120         .disable        = task_clock_perf_event_disable,
4121         .read           = task_clock_perf_event_read,
4122 };
4123
4124 #ifdef CONFIG_EVENT_PROFILE
4125
4126 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4127                           int entry_size)
4128 {
4129         struct perf_raw_record raw = {
4130                 .size = entry_size,
4131                 .data = record,
4132         };
4133
4134         struct perf_sample_data data = {
4135                 .addr = addr,
4136                 .raw = &raw,
4137         };
4138
4139         struct pt_regs *regs = get_irq_regs();
4140
4141         if (!regs)
4142                 regs = task_pt_regs(current);
4143
4144         do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4145                                 &data, regs);
4146 }
4147 EXPORT_SYMBOL_GPL(perf_tp_event);
4148
4149 static int perf_tp_event_match(struct perf_event *event,
4150                                 struct perf_sample_data *data)
4151 {
4152         void *record = data->raw->data;
4153
4154         if (likely(!event->filter) || filter_match_preds(event->filter, record))
4155                 return 1;
4156         return 0;
4157 }
4158
4159 static void tp_perf_event_destroy(struct perf_event *event)
4160 {
4161         ftrace_profile_disable(event->attr.config);
4162 }
4163
4164 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4165 {
4166         /*
4167          * Raw tracepoint data is a severe data leak, only allow root to
4168          * have these.
4169          */
4170         if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4171                         perf_paranoid_tracepoint_raw() &&
4172                         !capable(CAP_SYS_ADMIN))
4173                 return ERR_PTR(-EPERM);
4174
4175         if (ftrace_profile_enable(event->attr.config))
4176                 return NULL;
4177
4178         event->destroy = tp_perf_event_destroy;
4179
4180         return &perf_ops_generic;
4181 }
4182
4183 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4184 {
4185         char *filter_str;
4186         int ret;
4187
4188         if (event->attr.type != PERF_TYPE_TRACEPOINT)
4189                 return -EINVAL;
4190
4191         filter_str = strndup_user(arg, PAGE_SIZE);
4192         if (IS_ERR(filter_str))
4193                 return PTR_ERR(filter_str);
4194
4195         ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4196
4197         kfree(filter_str);
4198         return ret;
4199 }
4200
4201 static void perf_event_free_filter(struct perf_event *event)
4202 {
4203         ftrace_profile_free_filter(event);
4204 }
4205
4206 #else
4207
4208 static int perf_tp_event_match(struct perf_event *event,
4209                                 struct perf_sample_data *data)
4210 {
4211         return 1;
4212 }
4213
4214 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4215 {
4216         return NULL;
4217 }
4218
4219 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4220 {
4221         return -ENOENT;
4222 }
4223
4224 static void perf_event_free_filter(struct perf_event *event)
4225 {
4226 }
4227
4228 #endif /* CONFIG_EVENT_PROFILE */
4229
4230 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4231 static void bp_perf_event_destroy(struct perf_event *event)
4232 {
4233         release_bp_slot(event);
4234 }
4235
4236 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4237 {
4238         int err;
4239         /*
4240          * The breakpoint is already filled if we haven't created the counter
4241          * through perf syscall
4242          * FIXME: manage to get trigerred to NULL if it comes from syscalls
4243          */
4244         if (!bp->callback)
4245                 err = register_perf_hw_breakpoint(bp);
4246         else
4247                 err = __register_perf_hw_breakpoint(bp);
4248         if (err)
4249                 return ERR_PTR(err);
4250
4251         bp->destroy = bp_perf_event_destroy;
4252
4253         return &perf_ops_bp;
4254 }
4255
4256 void perf_bp_event(struct perf_event *bp, void *regs)
4257 {
4258         /* TODO */
4259 }
4260 #else
4261 static void bp_perf_event_destroy(struct perf_event *event)
4262 {
4263 }
4264
4265 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4266 {
4267         return NULL;
4268 }
4269
4270 void perf_bp_event(struct perf_event *bp, void *regs)
4271 {
4272 }
4273 #endif
4274
4275 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4276
4277 static void sw_perf_event_destroy(struct perf_event *event)
4278 {
4279         u64 event_id = event->attr.config;
4280
4281         WARN_ON(event->parent);
4282
4283         atomic_dec(&perf_swevent_enabled[event_id]);
4284 }
4285
4286 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4287 {
4288         const struct pmu *pmu = NULL;
4289         u64 event_id = event->attr.config;
4290
4291         /*
4292          * Software events (currently) can't in general distinguish
4293          * between user, kernel and hypervisor events.
4294          * However, context switches and cpu migrations are considered
4295          * to be kernel events, and page faults are never hypervisor
4296          * events.
4297          */
4298         switch (event_id) {
4299         case PERF_COUNT_SW_CPU_CLOCK:
4300                 pmu = &perf_ops_cpu_clock;
4301
4302                 break;
4303         case PERF_COUNT_SW_TASK_CLOCK:
4304                 /*
4305                  * If the user instantiates this as a per-cpu event,
4306                  * use the cpu_clock event instead.
4307                  */
4308                 if (event->ctx->task)
4309                         pmu = &perf_ops_task_clock;
4310                 else
4311                         pmu = &perf_ops_cpu_clock;
4312
4313                 break;
4314         case PERF_COUNT_SW_PAGE_FAULTS:
4315         case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4316         case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4317         case PERF_COUNT_SW_CONTEXT_SWITCHES:
4318         case PERF_COUNT_SW_CPU_MIGRATIONS:
4319         case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4320         case PERF_COUNT_SW_EMULATION_FAULTS:
4321                 if (!event->parent) {
4322                         atomic_inc(&perf_swevent_enabled[event_id]);
4323                         event->destroy = sw_perf_event_destroy;
4324                 }
4325                 pmu = &perf_ops_generic;
4326                 break;
4327         }
4328
4329         return pmu;
4330 }
4331
4332 /*
4333  * Allocate and initialize a event structure
4334  */
4335 static struct perf_event *
4336 perf_event_alloc(struct perf_event_attr *attr,
4337                    int cpu,
4338                    struct perf_event_context *ctx,
4339                    struct perf_event *group_leader,
4340                    struct perf_event *parent_event,
4341                    perf_callback_t callback,
4342                    gfp_t gfpflags)
4343 {
4344         const struct pmu *pmu;
4345         struct perf_event *event;
4346         struct hw_perf_event *hwc;
4347         long err;
4348
4349         event = kzalloc(sizeof(*event), gfpflags);
4350         if (!event)
4351                 return ERR_PTR(-ENOMEM);
4352
4353         /*
4354          * Single events are their own group leaders, with an
4355          * empty sibling list:
4356          */
4357         if (!group_leader)
4358                 group_leader = event;
4359
4360         mutex_init(&event->child_mutex);
4361         INIT_LIST_HEAD(&event->child_list);
4362
4363         INIT_LIST_HEAD(&event->group_entry);
4364         INIT_LIST_HEAD(&event->event_entry);
4365         INIT_LIST_HEAD(&event->sibling_list);
4366         init_waitqueue_head(&event->waitq);
4367
4368         mutex_init(&event->mmap_mutex);
4369
4370         event->cpu              = cpu;
4371         event->attr             = *attr;
4372         event->group_leader     = group_leader;
4373         event->pmu              = NULL;
4374         event->ctx              = ctx;
4375         event->oncpu            = -1;
4376
4377         event->parent           = parent_event;
4378
4379         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
4380         event->id               = atomic64_inc_return(&perf_event_id);
4381
4382         event->state            = PERF_EVENT_STATE_INACTIVE;
4383
4384         if (!callback && parent_event)
4385                 callback = parent_event->callback;
4386         
4387         event->callback = callback;
4388
4389         if (attr->disabled)
4390                 event->state = PERF_EVENT_STATE_OFF;
4391
4392         pmu = NULL;
4393
4394         hwc = &event->hw;
4395         hwc->sample_period = attr->sample_period;
4396         if (attr->freq && attr->sample_freq)
4397                 hwc->sample_period = 1;
4398         hwc->last_period = hwc->sample_period;
4399
4400         atomic64_set(&hwc->period_left, hwc->sample_period);
4401
4402         /*
4403          * we currently do not support PERF_FORMAT_GROUP on inherited events
4404          */
4405         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4406                 goto done;
4407
4408         switch (attr->type) {
4409         case PERF_TYPE_RAW:
4410         case PERF_TYPE_HARDWARE:
4411         case PERF_TYPE_HW_CACHE:
4412                 pmu = hw_perf_event_init(event);
4413                 break;
4414
4415         case PERF_TYPE_SOFTWARE:
4416                 pmu = sw_perf_event_init(event);
4417                 break;
4418
4419         case PERF_TYPE_TRACEPOINT:
4420                 pmu = tp_perf_event_init(event);
4421                 break;
4422
4423         case PERF_TYPE_BREAKPOINT:
4424                 pmu = bp_perf_event_init(event);
4425                 break;
4426
4427
4428         default:
4429                 break;
4430         }
4431 done:
4432         err = 0;
4433         if (!pmu)
4434                 err = -EINVAL;
4435         else if (IS_ERR(pmu))
4436                 err = PTR_ERR(pmu);
4437
4438         if (err) {
4439                 if (event->ns)
4440                         put_pid_ns(event->ns);
4441                 kfree(event);
4442                 return ERR_PTR(err);
4443         }
4444
4445         event->pmu = pmu;
4446
4447         if (!event->parent) {
4448                 atomic_inc(&nr_events);
4449                 if (event->attr.mmap)
4450                         atomic_inc(&nr_mmap_events);
4451                 if (event->attr.comm)
4452                         atomic_inc(&nr_comm_events);
4453                 if (event->attr.task)
4454                         atomic_inc(&nr_task_events);
4455         }
4456
4457         return event;
4458 }
4459
4460 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4461                           struct perf_event_attr *attr)
4462 {
4463         u32 size;
4464         int ret;
4465
4466         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4467                 return -EFAULT;
4468
4469         /*
4470          * zero the full structure, so that a short copy will be nice.
4471          */
4472         memset(attr, 0, sizeof(*attr));
4473
4474         ret = get_user(size, &uattr->size);
4475         if (ret)
4476                 return ret;
4477
4478         if (size > PAGE_SIZE)   /* silly large */
4479                 goto err_size;
4480
4481         if (!size)              /* abi compat */
4482                 size = PERF_ATTR_SIZE_VER0;
4483
4484         if (size < PERF_ATTR_SIZE_VER0)
4485                 goto err_size;
4486
4487         /*
4488          * If we're handed a bigger struct than we know of,
4489          * ensure all the unknown bits are 0 - i.e. new
4490          * user-space does not rely on any kernel feature
4491          * extensions we dont know about yet.
4492          */
4493         if (size > sizeof(*attr)) {
4494                 unsigned char __user *addr;
4495                 unsigned char __user *end;
4496                 unsigned char val;
4497
4498                 addr = (void __user *)uattr + sizeof(*attr);
4499                 end  = (void __user *)uattr + size;
4500
4501                 for (; addr < end; addr++) {
4502                         ret = get_user(val, addr);
4503                         if (ret)
4504                                 return ret;
4505                         if (val)
4506                                 goto err_size;
4507                 }
4508                 size = sizeof(*attr);
4509         }
4510
4511         ret = copy_from_user(attr, uattr, size);
4512         if (ret)
4513                 return -EFAULT;
4514
4515         /*
4516          * If the type exists, the corresponding creation will verify
4517          * the attr->config.
4518          */
4519         if (attr->type >= PERF_TYPE_MAX)
4520                 return -EINVAL;
4521
4522         if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3)
4523                 return -EINVAL;
4524
4525         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4526                 return -EINVAL;
4527
4528         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4529                 return -EINVAL;
4530
4531 out:
4532         return ret;
4533
4534 err_size:
4535         put_user(sizeof(*attr), &uattr->size);
4536         ret = -E2BIG;
4537         goto out;
4538 }
4539
4540 static int perf_event_set_output(struct perf_event *event, int output_fd)
4541 {
4542         struct perf_event *output_event = NULL;
4543         struct file *output_file = NULL;
4544         struct perf_event *old_output;
4545         int fput_needed = 0;
4546         int ret = -EINVAL;
4547
4548         if (!output_fd)
4549                 goto set;
4550
4551         output_file = fget_light(output_fd, &fput_needed);
4552         if (!output_file)
4553                 return -EBADF;
4554
4555         if (output_file->f_op != &perf_fops)
4556                 goto out;
4557
4558         output_event = output_file->private_data;
4559
4560         /* Don't chain output fds */
4561         if (output_event->output)
4562                 goto out;
4563
4564         /* Don't set an output fd when we already have an output channel */
4565         if (event->data)
4566                 goto out;
4567
4568         atomic_long_inc(&output_file->f_count);
4569
4570 set:
4571         mutex_lock(&event->mmap_mutex);
4572         old_output = event->output;
4573         rcu_assign_pointer(event->output, output_event);
4574         mutex_unlock(&event->mmap_mutex);
4575
4576         if (old_output) {
4577                 /*
4578                  * we need to make sure no existing perf_output_*()
4579                  * is still referencing this event.
4580                  */
4581                 synchronize_rcu();
4582                 fput(old_output->filp);
4583         }
4584
4585         ret = 0;
4586 out:
4587         fput_light(output_file, fput_needed);
4588         return ret;
4589 }
4590
4591 /**
4592  * sys_perf_event_open - open a performance event, associate it to a task/cpu
4593  *
4594  * @attr_uptr:  event_id type attributes for monitoring/sampling
4595  * @pid:                target pid
4596  * @cpu:                target cpu
4597  * @group_fd:           group leader event fd
4598  */
4599 SYSCALL_DEFINE5(perf_event_open,
4600                 struct perf_event_attr __user *, attr_uptr,
4601                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4602 {
4603         struct perf_event *event, *group_leader;
4604         struct perf_event_attr attr;
4605         struct perf_event_context *ctx;
4606         struct file *event_file = NULL;
4607         struct file *group_file = NULL;
4608         int fput_needed = 0;
4609         int fput_needed2 = 0;
4610         int err;
4611
4612         /* for future expandability... */
4613         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4614                 return -EINVAL;
4615
4616         err = perf_copy_attr(attr_uptr, &attr);
4617         if (err)
4618                 return err;
4619
4620         if (!attr.exclude_kernel) {
4621                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4622                         return -EACCES;
4623         }
4624
4625         if (attr.freq) {
4626                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4627                         return -EINVAL;
4628         }
4629
4630         /*
4631          * Get the target context (task or percpu):
4632          */
4633         ctx = find_get_context(pid, cpu);
4634         if (IS_ERR(ctx))
4635                 return PTR_ERR(ctx);
4636
4637         /*
4638          * Look up the group leader (we will attach this event to it):
4639          */
4640         group_leader = NULL;
4641         if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4642                 err = -EINVAL;
4643                 group_file = fget_light(group_fd, &fput_needed);
4644                 if (!group_file)
4645                         goto err_put_context;
4646                 if (group_file->f_op != &perf_fops)
4647                         goto err_put_context;
4648
4649                 group_leader = group_file->private_data;
4650                 /*
4651                  * Do not allow a recursive hierarchy (this new sibling
4652                  * becoming part of another group-sibling):
4653                  */
4654                 if (group_leader->group_leader != group_leader)
4655                         goto err_put_context;
4656                 /*
4657                  * Do not allow to attach to a group in a different
4658                  * task or CPU context:
4659                  */
4660                 if (group_leader->ctx != ctx)
4661                         goto err_put_context;
4662                 /*
4663                  * Only a group leader can be exclusive or pinned
4664                  */
4665                 if (attr.exclusive || attr.pinned)
4666                         goto err_put_context;
4667         }
4668
4669         event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4670                                      NULL, NULL, GFP_KERNEL);
4671         err = PTR_ERR(event);
4672         if (IS_ERR(event))
4673                 goto err_put_context;
4674
4675         err = anon_inode_getfd("[perf_event]", &perf_fops, event, 0);
4676         if (err < 0)
4677                 goto err_free_put_context;
4678
4679         event_file = fget_light(err, &fput_needed2);
4680         if (!event_file)
4681                 goto err_free_put_context;
4682
4683         if (flags & PERF_FLAG_FD_OUTPUT) {
4684                 err = perf_event_set_output(event, group_fd);
4685                 if (err)
4686                         goto err_fput_free_put_context;
4687         }
4688
4689         event->filp = event_file;
4690         WARN_ON_ONCE(ctx->parent_ctx);
4691         mutex_lock(&ctx->mutex);
4692         perf_install_in_context(ctx, event, cpu);
4693         ++ctx->generation;
4694         mutex_unlock(&ctx->mutex);
4695
4696         event->owner = current;
4697         get_task_struct(current);
4698         mutex_lock(&current->perf_event_mutex);
4699         list_add_tail(&event->owner_entry, &current->perf_event_list);
4700         mutex_unlock(&current->perf_event_mutex);
4701
4702 err_fput_free_put_context:
4703         fput_light(event_file, fput_needed2);
4704
4705 err_free_put_context:
4706         if (err < 0)
4707                 kfree(event);
4708
4709 err_put_context:
4710         if (err < 0)
4711                 put_ctx(ctx);
4712
4713         fput_light(group_file, fput_needed);
4714
4715         return err;
4716 }
4717
4718 /**
4719  * perf_event_create_kernel_counter
4720  *
4721  * @attr: attributes of the counter to create
4722  * @cpu: cpu in which the counter is bound
4723  * @pid: task to profile
4724  */
4725 struct perf_event *
4726 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4727                                  pid_t pid, perf_callback_t callback)
4728 {
4729         struct perf_event *event;
4730         struct perf_event_context *ctx;
4731         int err;
4732
4733         /*
4734          * Get the target context (task or percpu):
4735          */
4736
4737         ctx = find_get_context(pid, cpu);
4738         if (IS_ERR(ctx))
4739                 return NULL;
4740
4741         event = perf_event_alloc(attr, cpu, ctx, NULL,
4742                                      NULL, callback, GFP_KERNEL);
4743         err = PTR_ERR(event);
4744         if (IS_ERR(event))
4745                 goto err_put_context;
4746
4747         event->filp = NULL;
4748         WARN_ON_ONCE(ctx->parent_ctx);
4749         mutex_lock(&ctx->mutex);
4750         perf_install_in_context(ctx, event, cpu);
4751         ++ctx->generation;
4752         mutex_unlock(&ctx->mutex);
4753
4754         event->owner = current;
4755         get_task_struct(current);
4756         mutex_lock(&current->perf_event_mutex);
4757         list_add_tail(&event->owner_entry, &current->perf_event_list);
4758         mutex_unlock(&current->perf_event_mutex);
4759
4760         return event;
4761
4762 err_put_context:
4763         if (err < 0)
4764                 put_ctx(ctx);
4765
4766         return NULL;
4767 }
4768 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4769
4770 /*
4771  * inherit a event from parent task to child task:
4772  */
4773 static struct perf_event *
4774 inherit_event(struct perf_event *parent_event,
4775               struct task_struct *parent,
4776               struct perf_event_context *parent_ctx,
4777               struct task_struct *child,
4778               struct perf_event *group_leader,
4779               struct perf_event_context *child_ctx)
4780 {
4781         struct perf_event *child_event;
4782
4783         /*
4784          * Instead of creating recursive hierarchies of events,
4785          * we link inherited events back to the original parent,
4786          * which has a filp for sure, which we use as the reference
4787          * count:
4788          */
4789         if (parent_event->parent)
4790                 parent_event = parent_event->parent;
4791
4792         child_event = perf_event_alloc(&parent_event->attr,
4793                                            parent_event->cpu, child_ctx,
4794                                            group_leader, parent_event,
4795                                            NULL, GFP_KERNEL);
4796         if (IS_ERR(child_event))
4797                 return child_event;
4798         get_ctx(child_ctx);
4799
4800         /*
4801          * Make the child state follow the state of the parent event,
4802          * not its attr.disabled bit.  We hold the parent's mutex,
4803          * so we won't race with perf_event_{en, dis}able_family.
4804          */
4805         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
4806                 child_event->state = PERF_EVENT_STATE_INACTIVE;
4807         else
4808                 child_event->state = PERF_EVENT_STATE_OFF;
4809
4810         if (parent_event->attr.freq)
4811                 child_event->hw.sample_period = parent_event->hw.sample_period;
4812
4813         child_event->overflow_handler = parent_event->overflow_handler;
4814
4815         /*
4816          * Link it up in the child's context:
4817          */
4818         add_event_to_ctx(child_event, child_ctx);
4819
4820         /*
4821          * Get a reference to the parent filp - we will fput it
4822          * when the child event exits. This is safe to do because
4823          * we are in the parent and we know that the filp still
4824          * exists and has a nonzero count:
4825          */
4826         atomic_long_inc(&parent_event->filp->f_count);
4827
4828         /*
4829          * Link this into the parent event's child list
4830          */
4831         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4832         mutex_lock(&parent_event->child_mutex);
4833         list_add_tail(&child_event->child_list, &parent_event->child_list);
4834         mutex_unlock(&parent_event->child_mutex);
4835
4836         return child_event;
4837 }
4838
4839 static int inherit_group(struct perf_event *parent_event,
4840               struct task_struct *parent,
4841               struct perf_event_context *parent_ctx,
4842               struct task_struct *child,
4843               struct perf_event_context *child_ctx)
4844 {
4845         struct perf_event *leader;
4846         struct perf_event *sub;
4847         struct perf_event *child_ctr;
4848
4849         leader = inherit_event(parent_event, parent, parent_ctx,
4850                                  child, NULL, child_ctx);
4851         if (IS_ERR(leader))
4852                 return PTR_ERR(leader);
4853         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
4854                 child_ctr = inherit_event(sub, parent, parent_ctx,
4855                                             child, leader, child_ctx);
4856                 if (IS_ERR(child_ctr))
4857                         return PTR_ERR(child_ctr);
4858         }
4859         return 0;
4860 }
4861
4862 static void sync_child_event(struct perf_event *child_event,
4863                                struct task_struct *child)
4864 {
4865         struct perf_event *parent_event = child_event->parent;
4866         u64 child_val;
4867
4868         if (child_event->attr.inherit_stat)
4869                 perf_event_read_event(child_event, child);
4870
4871         child_val = atomic64_read(&child_event->count);
4872
4873         /*
4874          * Add back the child's count to the parent's count:
4875          */
4876         atomic64_add(child_val, &parent_event->count);
4877         atomic64_add(child_event->total_time_enabled,
4878                      &parent_event->child_total_time_enabled);
4879         atomic64_add(child_event->total_time_running,
4880                      &parent_event->child_total_time_running);
4881
4882         /*
4883          * Remove this event from the parent's list
4884          */
4885         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4886         mutex_lock(&parent_event->child_mutex);
4887         list_del_init(&child_event->child_list);
4888         mutex_unlock(&parent_event->child_mutex);
4889
4890         /*
4891          * Release the parent event, if this was the last
4892          * reference to it.
4893          */
4894         fput(parent_event->filp);
4895 }
4896
4897 static void
4898 __perf_event_exit_task(struct perf_event *child_event,
4899                          struct perf_event_context *child_ctx,
4900                          struct task_struct *child)
4901 {
4902         struct perf_event *parent_event;
4903
4904         update_event_times(child_event);
4905         perf_event_remove_from_context(child_event);
4906
4907         parent_event = child_event->parent;
4908         /*
4909          * It can happen that parent exits first, and has events
4910          * that are still around due to the child reference. These
4911          * events need to be zapped - but otherwise linger.
4912          */
4913         if (parent_event) {
4914                 sync_child_event(child_event, child);
4915                 free_event(child_event);
4916         }
4917 }
4918
4919 /*
4920  * When a child task exits, feed back event values to parent events.
4921  */
4922 void perf_event_exit_task(struct task_struct *child)
4923 {
4924         struct perf_event *child_event, *tmp;
4925         struct perf_event_context *child_ctx;
4926         unsigned long flags;
4927
4928         if (likely(!child->perf_event_ctxp)) {
4929                 perf_event_task(child, NULL, 0);
4930                 return;
4931         }
4932
4933         local_irq_save(flags);
4934         /*
4935          * We can't reschedule here because interrupts are disabled,
4936          * and either child is current or it is a task that can't be
4937          * scheduled, so we are now safe from rescheduling changing
4938          * our context.
4939          */
4940         child_ctx = child->perf_event_ctxp;
4941         __perf_event_task_sched_out(child_ctx);
4942
4943         /*
4944          * Take the context lock here so that if find_get_context is
4945          * reading child->perf_event_ctxp, we wait until it has
4946          * incremented the context's refcount before we do put_ctx below.
4947          */
4948         spin_lock(&child_ctx->lock);
4949         child->perf_event_ctxp = NULL;
4950         /*
4951          * If this context is a clone; unclone it so it can't get
4952          * swapped to another process while we're removing all
4953          * the events from it.
4954          */
4955         unclone_ctx(child_ctx);
4956         spin_unlock_irqrestore(&child_ctx->lock, flags);
4957
4958         /*
4959          * Report the task dead after unscheduling the events so that we
4960          * won't get any samples after PERF_RECORD_EXIT. We can however still
4961          * get a few PERF_RECORD_READ events.
4962          */
4963         perf_event_task(child, child_ctx, 0);
4964
4965         /*
4966          * We can recurse on the same lock type through:
4967          *
4968          *   __perf_event_exit_task()
4969          *     sync_child_event()
4970          *       fput(parent_event->filp)
4971          *         perf_release()
4972          *           mutex_lock(&ctx->mutex)
4973          *
4974          * But since its the parent context it won't be the same instance.
4975          */
4976         mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
4977
4978 again:
4979         list_for_each_entry_safe(child_event, tmp, &child_ctx->group_list,
4980                                  group_entry)
4981                 __perf_event_exit_task(child_event, child_ctx, child);
4982
4983         /*
4984          * If the last event was a group event, it will have appended all
4985          * its siblings to the list, but we obtained 'tmp' before that which
4986          * will still point to the list head terminating the iteration.
4987          */
4988         if (!list_empty(&child_ctx->group_list))
4989                 goto again;
4990
4991         mutex_unlock(&child_ctx->mutex);
4992
4993         put_ctx(child_ctx);
4994 }
4995
4996 /*
4997  * free an unexposed, unused context as created by inheritance by
4998  * init_task below, used by fork() in case of fail.
4999  */
5000 void perf_event_free_task(struct task_struct *task)
5001 {
5002         struct perf_event_context *ctx = task->perf_event_ctxp;
5003         struct perf_event *event, *tmp;
5004
5005         if (!ctx)
5006                 return;
5007
5008         mutex_lock(&ctx->mutex);
5009 again:
5010         list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry) {
5011                 struct perf_event *parent = event->parent;
5012
5013                 if (WARN_ON_ONCE(!parent))
5014                         continue;
5015
5016                 mutex_lock(&parent->child_mutex);
5017                 list_del_init(&event->child_list);
5018                 mutex_unlock(&parent->child_mutex);
5019
5020                 fput(parent->filp);
5021
5022                 list_del_event(event, ctx);
5023                 free_event(event);
5024         }
5025
5026         if (!list_empty(&ctx->group_list))
5027                 goto again;
5028
5029         mutex_unlock(&ctx->mutex);
5030
5031         put_ctx(ctx);
5032 }
5033
5034 /*
5035  * Initialize the perf_event context in task_struct
5036  */
5037 int perf_event_init_task(struct task_struct *child)
5038 {
5039         struct perf_event_context *child_ctx, *parent_ctx;
5040         struct perf_event_context *cloned_ctx;
5041         struct perf_event *event;
5042         struct task_struct *parent = current;
5043         int inherited_all = 1;
5044         int ret = 0;
5045
5046         child->perf_event_ctxp = NULL;
5047
5048         mutex_init(&child->perf_event_mutex);
5049         INIT_LIST_HEAD(&child->perf_event_list);
5050
5051         if (likely(!parent->perf_event_ctxp))
5052                 return 0;
5053
5054         /*
5055          * This is executed from the parent task context, so inherit
5056          * events that have been marked for cloning.
5057          * First allocate and initialize a context for the child.
5058          */
5059
5060         child_ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
5061         if (!child_ctx)
5062                 return -ENOMEM;
5063
5064         __perf_event_init_context(child_ctx, child);
5065         child->perf_event_ctxp = child_ctx;
5066         get_task_struct(child);
5067
5068         /*
5069          * If the parent's context is a clone, pin it so it won't get
5070          * swapped under us.
5071          */
5072         parent_ctx = perf_pin_task_context(parent);
5073
5074         /*
5075          * No need to check if parent_ctx != NULL here; since we saw
5076          * it non-NULL earlier, the only reason for it to become NULL
5077          * is if we exit, and since we're currently in the middle of
5078          * a fork we can't be exiting at the same time.
5079          */
5080
5081         /*
5082          * Lock the parent list. No need to lock the child - not PID
5083          * hashed yet and not running, so nobody can access it.
5084          */
5085         mutex_lock(&parent_ctx->mutex);
5086
5087         /*
5088          * We dont have to disable NMIs - we are only looking at
5089          * the list, not manipulating it:
5090          */
5091         list_for_each_entry(event, &parent_ctx->group_list, group_entry) {
5092
5093                 if (!event->attr.inherit) {
5094                         inherited_all = 0;
5095                         continue;
5096                 }
5097
5098                 ret = inherit_group(event, parent, parent_ctx,
5099                                              child, child_ctx);
5100                 if (ret) {
5101                         inherited_all = 0;
5102                         break;
5103                 }
5104         }
5105
5106         if (inherited_all) {
5107                 /*
5108                  * Mark the child context as a clone of the parent
5109                  * context, or of whatever the parent is a clone of.
5110                  * Note that if the parent is a clone, it could get
5111                  * uncloned at any point, but that doesn't matter
5112                  * because the list of events and the generation
5113                  * count can't have changed since we took the mutex.
5114                  */
5115                 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5116                 if (cloned_ctx) {
5117                         child_ctx->parent_ctx = cloned_ctx;
5118                         child_ctx->parent_gen = parent_ctx->parent_gen;
5119                 } else {
5120                         child_ctx->parent_ctx = parent_ctx;
5121                         child_ctx->parent_gen = parent_ctx->generation;
5122                 }
5123                 get_ctx(child_ctx->parent_ctx);
5124         }
5125
5126         mutex_unlock(&parent_ctx->mutex);
5127
5128         perf_unpin_context(parent_ctx);
5129
5130         return ret;
5131 }
5132
5133 static void __cpuinit perf_event_init_cpu(int cpu)
5134 {
5135         struct perf_cpu_context *cpuctx;
5136
5137         cpuctx = &per_cpu(perf_cpu_context, cpu);
5138         __perf_event_init_context(&cpuctx->ctx, NULL);
5139
5140         spin_lock(&perf_resource_lock);
5141         cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5142         spin_unlock(&perf_resource_lock);
5143
5144         hw_perf_event_setup(cpu);
5145 }
5146
5147 #ifdef CONFIG_HOTPLUG_CPU
5148 static void __perf_event_exit_cpu(void *info)
5149 {
5150         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5151         struct perf_event_context *ctx = &cpuctx->ctx;
5152         struct perf_event *event, *tmp;
5153
5154         list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry)
5155                 __perf_event_remove_from_context(event);
5156 }
5157 static void perf_event_exit_cpu(int cpu)
5158 {
5159         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5160         struct perf_event_context *ctx = &cpuctx->ctx;
5161
5162         mutex_lock(&ctx->mutex);
5163         smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5164         mutex_unlock(&ctx->mutex);
5165 }
5166 #else
5167 static inline void perf_event_exit_cpu(int cpu) { }
5168 #endif
5169
5170 static int __cpuinit
5171 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5172 {
5173         unsigned int cpu = (long)hcpu;
5174
5175         switch (action) {
5176
5177         case CPU_UP_PREPARE:
5178         case CPU_UP_PREPARE_FROZEN:
5179                 perf_event_init_cpu(cpu);
5180                 break;
5181
5182         case CPU_ONLINE:
5183         case CPU_ONLINE_FROZEN:
5184                 hw_perf_event_setup_online(cpu);
5185                 break;
5186
5187         case CPU_DOWN_PREPARE:
5188         case CPU_DOWN_PREPARE_FROZEN:
5189                 perf_event_exit_cpu(cpu);
5190                 break;
5191
5192         default:
5193                 break;
5194         }
5195
5196         return NOTIFY_OK;
5197 }
5198
5199 /*
5200  * This has to have a higher priority than migration_notifier in sched.c.
5201  */
5202 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5203         .notifier_call          = perf_cpu_notify,
5204         .priority               = 20,
5205 };
5206
5207 void __init perf_event_init(void)
5208 {
5209         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5210                         (void *)(long)smp_processor_id());
5211         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5212                         (void *)(long)smp_processor_id());
5213         register_cpu_notifier(&perf_cpu_nb);
5214 }
5215
5216 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
5217 {
5218         return sprintf(buf, "%d\n", perf_reserved_percpu);
5219 }
5220
5221 static ssize_t
5222 perf_set_reserve_percpu(struct sysdev_class *class,
5223                         const char *buf,
5224                         size_t count)
5225 {
5226         struct perf_cpu_context *cpuctx;
5227         unsigned long val;
5228         int err, cpu, mpt;
5229
5230         err = strict_strtoul(buf, 10, &val);
5231         if (err)
5232                 return err;
5233         if (val > perf_max_events)
5234                 return -EINVAL;
5235
5236         spin_lock(&perf_resource_lock);
5237         perf_reserved_percpu = val;
5238         for_each_online_cpu(cpu) {
5239                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5240                 spin_lock_irq(&cpuctx->ctx.lock);
5241                 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5242                           perf_max_events - perf_reserved_percpu);
5243                 cpuctx->max_pertask = mpt;
5244                 spin_unlock_irq(&cpuctx->ctx.lock);
5245         }
5246         spin_unlock(&perf_resource_lock);
5247
5248         return count;
5249 }
5250
5251 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
5252 {
5253         return sprintf(buf, "%d\n", perf_overcommit);
5254 }
5255
5256 static ssize_t
5257 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
5258 {
5259         unsigned long val;
5260         int err;
5261
5262         err = strict_strtoul(buf, 10, &val);
5263         if (err)
5264                 return err;
5265         if (val > 1)
5266                 return -EINVAL;
5267
5268         spin_lock(&perf_resource_lock);
5269         perf_overcommit = val;
5270         spin_unlock(&perf_resource_lock);
5271
5272         return count;
5273 }
5274
5275 static SYSDEV_CLASS_ATTR(
5276                                 reserve_percpu,
5277                                 0644,
5278                                 perf_show_reserve_percpu,
5279                                 perf_set_reserve_percpu
5280                         );
5281
5282 static SYSDEV_CLASS_ATTR(
5283                                 overcommit,
5284                                 0644,
5285                                 perf_show_overcommit,
5286                                 perf_set_overcommit
5287                         );
5288
5289 static struct attribute *perfclass_attrs[] = {
5290         &attr_reserve_percpu.attr,
5291         &attr_overcommit.attr,
5292         NULL
5293 };
5294
5295 static struct attribute_group perfclass_attr_group = {
5296         .attrs                  = perfclass_attrs,
5297         .name                   = "perf_events",
5298 };
5299
5300 static int __init perf_event_sysfs_init(void)
5301 {
5302         return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5303                                   &perfclass_attr_group);
5304 }
5305 device_initcall(perf_event_sysfs_init);