perf: Fix event scaling for inherited counters
[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         spin_lock(&ctx->lock);
1530         update_context_time(ctx);
1531         update_event_times(event);
1532         spin_unlock(&ctx->lock);
1533
1534         event->pmu->read(event);
1535 }
1536
1537 static u64 perf_event_read(struct perf_event *event)
1538 {
1539         /*
1540          * If event is enabled and currently active on a CPU, update the
1541          * value in the event structure:
1542          */
1543         if (event->state == PERF_EVENT_STATE_ACTIVE) {
1544                 smp_call_function_single(event->oncpu,
1545                                          __perf_event_read, event, 1);
1546         } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1547                 struct perf_event_context *ctx = event->ctx;
1548                 unsigned long flags;
1549
1550                 spin_lock_irqsave(&ctx->lock, flags);
1551                 update_context_time(ctx);
1552                 update_event_times(event);
1553                 spin_unlock_irqrestore(&ctx->lock, flags);
1554         }
1555
1556         return atomic64_read(&event->count);
1557 }
1558
1559 /*
1560  * Initialize the perf_event context in a task_struct:
1561  */
1562 static void
1563 __perf_event_init_context(struct perf_event_context *ctx,
1564                             struct task_struct *task)
1565 {
1566         memset(ctx, 0, sizeof(*ctx));
1567         spin_lock_init(&ctx->lock);
1568         mutex_init(&ctx->mutex);
1569         INIT_LIST_HEAD(&ctx->group_list);
1570         INIT_LIST_HEAD(&ctx->event_list);
1571         atomic_set(&ctx->refcount, 1);
1572         ctx->task = task;
1573 }
1574
1575 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1576 {
1577         struct perf_event_context *ctx;
1578         struct perf_cpu_context *cpuctx;
1579         struct task_struct *task;
1580         unsigned long flags;
1581         int err;
1582
1583         /*
1584          * If cpu is not a wildcard then this is a percpu event:
1585          */
1586         if (cpu != -1) {
1587                 /* Must be root to operate on a CPU event: */
1588                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1589                         return ERR_PTR(-EACCES);
1590
1591                 if (cpu < 0 || cpu > num_possible_cpus())
1592                         return ERR_PTR(-EINVAL);
1593
1594                 /*
1595                  * We could be clever and allow to attach a event to an
1596                  * offline CPU and activate it when the CPU comes up, but
1597                  * that's for later.
1598                  */
1599                 if (!cpu_isset(cpu, cpu_online_map))
1600                         return ERR_PTR(-ENODEV);
1601
1602                 cpuctx = &per_cpu(perf_cpu_context, cpu);
1603                 ctx = &cpuctx->ctx;
1604                 get_ctx(ctx);
1605
1606                 return ctx;
1607         }
1608
1609         rcu_read_lock();
1610         if (!pid)
1611                 task = current;
1612         else
1613                 task = find_task_by_vpid(pid);
1614         if (task)
1615                 get_task_struct(task);
1616         rcu_read_unlock();
1617
1618         if (!task)
1619                 return ERR_PTR(-ESRCH);
1620
1621         /*
1622          * Can't attach events to a dying task.
1623          */
1624         err = -ESRCH;
1625         if (task->flags & PF_EXITING)
1626                 goto errout;
1627
1628         /* Reuse ptrace permission checks for now. */
1629         err = -EACCES;
1630         if (!ptrace_may_access(task, PTRACE_MODE_READ))
1631                 goto errout;
1632
1633  retry:
1634         ctx = perf_lock_task_context(task, &flags);
1635         if (ctx) {
1636                 unclone_ctx(ctx);
1637                 spin_unlock_irqrestore(&ctx->lock, flags);
1638         }
1639
1640         if (!ctx) {
1641                 ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1642                 err = -ENOMEM;
1643                 if (!ctx)
1644                         goto errout;
1645                 __perf_event_init_context(ctx, task);
1646                 get_ctx(ctx);
1647                 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1648                         /*
1649                          * We raced with some other task; use
1650                          * the context they set.
1651                          */
1652                         kfree(ctx);
1653                         goto retry;
1654                 }
1655                 get_task_struct(task);
1656         }
1657
1658         put_task_struct(task);
1659         return ctx;
1660
1661  errout:
1662         put_task_struct(task);
1663         return ERR_PTR(err);
1664 }
1665
1666 static void perf_event_free_filter(struct perf_event *event);
1667
1668 static void free_event_rcu(struct rcu_head *head)
1669 {
1670         struct perf_event *event;
1671
1672         event = container_of(head, struct perf_event, rcu_head);
1673         if (event->ns)
1674                 put_pid_ns(event->ns);
1675         perf_event_free_filter(event);
1676         kfree(event);
1677 }
1678
1679 static void perf_pending_sync(struct perf_event *event);
1680
1681 static void free_event(struct perf_event *event)
1682 {
1683         perf_pending_sync(event);
1684
1685         if (!event->parent) {
1686                 atomic_dec(&nr_events);
1687                 if (event->attr.mmap)
1688                         atomic_dec(&nr_mmap_events);
1689                 if (event->attr.comm)
1690                         atomic_dec(&nr_comm_events);
1691                 if (event->attr.task)
1692                         atomic_dec(&nr_task_events);
1693         }
1694
1695         if (event->output) {
1696                 fput(event->output->filp);
1697                 event->output = NULL;
1698         }
1699
1700         if (event->destroy)
1701                 event->destroy(event);
1702
1703         put_ctx(event->ctx);
1704         call_rcu(&event->rcu_head, free_event_rcu);
1705 }
1706
1707 /*
1708  * Called when the last reference to the file is gone.
1709  */
1710 static int perf_release(struct inode *inode, struct file *file)
1711 {
1712         struct perf_event *event = file->private_data;
1713         struct perf_event_context *ctx = event->ctx;
1714
1715         file->private_data = NULL;
1716
1717         WARN_ON_ONCE(ctx->parent_ctx);
1718         mutex_lock(&ctx->mutex);
1719         perf_event_remove_from_context(event);
1720         mutex_unlock(&ctx->mutex);
1721
1722         mutex_lock(&event->owner->perf_event_mutex);
1723         list_del_init(&event->owner_entry);
1724         mutex_unlock(&event->owner->perf_event_mutex);
1725         put_task_struct(event->owner);
1726
1727         free_event(event);
1728
1729         return 0;
1730 }
1731
1732 int perf_event_release_kernel(struct perf_event *event)
1733 {
1734         struct perf_event_context *ctx = event->ctx;
1735
1736         WARN_ON_ONCE(ctx->parent_ctx);
1737         mutex_lock(&ctx->mutex);
1738         perf_event_remove_from_context(event);
1739         mutex_unlock(&ctx->mutex);
1740
1741         mutex_lock(&event->owner->perf_event_mutex);
1742         list_del_init(&event->owner_entry);
1743         mutex_unlock(&event->owner->perf_event_mutex);
1744         put_task_struct(event->owner);
1745
1746         free_event(event);
1747
1748         return 0;
1749 }
1750 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1751
1752 static int perf_event_read_size(struct perf_event *event)
1753 {
1754         int entry = sizeof(u64); /* value */
1755         int size = 0;
1756         int nr = 1;
1757
1758         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1759                 size += sizeof(u64);
1760
1761         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1762                 size += sizeof(u64);
1763
1764         if (event->attr.read_format & PERF_FORMAT_ID)
1765                 entry += sizeof(u64);
1766
1767         if (event->attr.read_format & PERF_FORMAT_GROUP) {
1768                 nr += event->group_leader->nr_siblings;
1769                 size += sizeof(u64);
1770         }
1771
1772         size += entry * nr;
1773
1774         return size;
1775 }
1776
1777 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1778 {
1779         struct perf_event *child;
1780         u64 total = 0;
1781
1782         *enabled = 0;
1783         *running = 0;
1784
1785         total += perf_event_read(event);
1786         *enabled += event->total_time_enabled +
1787                         atomic64_read(&event->child_total_time_enabled);
1788         *running += event->total_time_running +
1789                         atomic64_read(&event->child_total_time_running);
1790
1791         list_for_each_entry(child, &event->child_list, child_list) {
1792                 total += perf_event_read(child);
1793                 *enabled += child->total_time_enabled;
1794                 *running += child->total_time_running;
1795         }
1796
1797         return total;
1798 }
1799 EXPORT_SYMBOL_GPL(perf_event_read_value);
1800
1801 static int perf_event_read_group(struct perf_event *event,
1802                                    u64 read_format, char __user *buf)
1803 {
1804         struct perf_event *leader = event->group_leader, *sub;
1805         int n = 0, size = 0, ret = 0;
1806         u64 values[5];
1807         u64 count, enabled, running;
1808
1809         count = perf_event_read_value(leader, &enabled, &running);
1810
1811         values[n++] = 1 + leader->nr_siblings;
1812         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1813                 values[n++] = enabled;
1814         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1815                 values[n++] = running;
1816         values[n++] = count;
1817         if (read_format & PERF_FORMAT_ID)
1818                 values[n++] = primary_event_id(leader);
1819
1820         size = n * sizeof(u64);
1821
1822         if (copy_to_user(buf, values, size))
1823                 return -EFAULT;
1824
1825         ret += size;
1826
1827         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1828                 n = 0;
1829
1830                 values[n++] = perf_event_read_value(sub, &enabled, &running);
1831                 if (read_format & PERF_FORMAT_ID)
1832                         values[n++] = primary_event_id(sub);
1833
1834                 size = n * sizeof(u64);
1835
1836                 if (copy_to_user(buf + size, values, size))
1837                         return -EFAULT;
1838
1839                 ret += size;
1840         }
1841
1842         return ret;
1843 }
1844
1845 static int perf_event_read_one(struct perf_event *event,
1846                                  u64 read_format, char __user *buf)
1847 {
1848         u64 enabled, running;
1849         u64 values[4];
1850         int n = 0;
1851
1852         values[n++] = perf_event_read_value(event, &enabled, &running);
1853         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1854                 values[n++] = enabled;
1855         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1856                 values[n++] = running;
1857         if (read_format & PERF_FORMAT_ID)
1858                 values[n++] = primary_event_id(event);
1859
1860         if (copy_to_user(buf, values, n * sizeof(u64)))
1861                 return -EFAULT;
1862
1863         return n * sizeof(u64);
1864 }
1865
1866 /*
1867  * Read the performance event - simple non blocking version for now
1868  */
1869 static ssize_t
1870 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
1871 {
1872         u64 read_format = event->attr.read_format;
1873         int ret;
1874
1875         /*
1876          * Return end-of-file for a read on a event that is in
1877          * error state (i.e. because it was pinned but it couldn't be
1878          * scheduled on to the CPU at some point).
1879          */
1880         if (event->state == PERF_EVENT_STATE_ERROR)
1881                 return 0;
1882
1883         if (count < perf_event_read_size(event))
1884                 return -ENOSPC;
1885
1886         WARN_ON_ONCE(event->ctx->parent_ctx);
1887         mutex_lock(&event->child_mutex);
1888         if (read_format & PERF_FORMAT_GROUP)
1889                 ret = perf_event_read_group(event, read_format, buf);
1890         else
1891                 ret = perf_event_read_one(event, read_format, buf);
1892         mutex_unlock(&event->child_mutex);
1893
1894         return ret;
1895 }
1896
1897 static ssize_t
1898 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1899 {
1900         struct perf_event *event = file->private_data;
1901
1902         return perf_read_hw(event, buf, count);
1903 }
1904
1905 static unsigned int perf_poll(struct file *file, poll_table *wait)
1906 {
1907         struct perf_event *event = file->private_data;
1908         struct perf_mmap_data *data;
1909         unsigned int events = POLL_HUP;
1910
1911         rcu_read_lock();
1912         data = rcu_dereference(event->data);
1913         if (data)
1914                 events = atomic_xchg(&data->poll, 0);
1915         rcu_read_unlock();
1916
1917         poll_wait(file, &event->waitq, wait);
1918
1919         return events;
1920 }
1921
1922 static void perf_event_reset(struct perf_event *event)
1923 {
1924         (void)perf_event_read(event);
1925         atomic64_set(&event->count, 0);
1926         perf_event_update_userpage(event);
1927 }
1928
1929 /*
1930  * Holding the top-level event's child_mutex means that any
1931  * descendant process that has inherited this event will block
1932  * in sync_child_event if it goes to exit, thus satisfying the
1933  * task existence requirements of perf_event_enable/disable.
1934  */
1935 static void perf_event_for_each_child(struct perf_event *event,
1936                                         void (*func)(struct perf_event *))
1937 {
1938         struct perf_event *child;
1939
1940         WARN_ON_ONCE(event->ctx->parent_ctx);
1941         mutex_lock(&event->child_mutex);
1942         func(event);
1943         list_for_each_entry(child, &event->child_list, child_list)
1944                 func(child);
1945         mutex_unlock(&event->child_mutex);
1946 }
1947
1948 static void perf_event_for_each(struct perf_event *event,
1949                                   void (*func)(struct perf_event *))
1950 {
1951         struct perf_event_context *ctx = event->ctx;
1952         struct perf_event *sibling;
1953
1954         WARN_ON_ONCE(ctx->parent_ctx);
1955         mutex_lock(&ctx->mutex);
1956         event = event->group_leader;
1957
1958         perf_event_for_each_child(event, func);
1959         func(event);
1960         list_for_each_entry(sibling, &event->sibling_list, group_entry)
1961                 perf_event_for_each_child(event, func);
1962         mutex_unlock(&ctx->mutex);
1963 }
1964
1965 static int perf_event_period(struct perf_event *event, u64 __user *arg)
1966 {
1967         struct perf_event_context *ctx = event->ctx;
1968         unsigned long size;
1969         int ret = 0;
1970         u64 value;
1971
1972         if (!event->attr.sample_period)
1973                 return -EINVAL;
1974
1975         size = copy_from_user(&value, arg, sizeof(value));
1976         if (size != sizeof(value))
1977                 return -EFAULT;
1978
1979         if (!value)
1980                 return -EINVAL;
1981
1982         spin_lock_irq(&ctx->lock);
1983         if (event->attr.freq) {
1984                 if (value > sysctl_perf_event_sample_rate) {
1985                         ret = -EINVAL;
1986                         goto unlock;
1987                 }
1988
1989                 event->attr.sample_freq = value;
1990         } else {
1991                 event->attr.sample_period = value;
1992                 event->hw.sample_period = value;
1993         }
1994 unlock:
1995         spin_unlock_irq(&ctx->lock);
1996
1997         return ret;
1998 }
1999
2000 static int perf_event_set_output(struct perf_event *event, int output_fd);
2001 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2002
2003 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2004 {
2005         struct perf_event *event = file->private_data;
2006         void (*func)(struct perf_event *);
2007         u32 flags = arg;
2008
2009         switch (cmd) {
2010         case PERF_EVENT_IOC_ENABLE:
2011                 func = perf_event_enable;
2012                 break;
2013         case PERF_EVENT_IOC_DISABLE:
2014                 func = perf_event_disable;
2015                 break;
2016         case PERF_EVENT_IOC_RESET:
2017                 func = perf_event_reset;
2018                 break;
2019
2020         case PERF_EVENT_IOC_REFRESH:
2021                 return perf_event_refresh(event, arg);
2022
2023         case PERF_EVENT_IOC_PERIOD:
2024                 return perf_event_period(event, (u64 __user *)arg);
2025
2026         case PERF_EVENT_IOC_SET_OUTPUT:
2027                 return perf_event_set_output(event, arg);
2028
2029         case PERF_EVENT_IOC_SET_FILTER:
2030                 return perf_event_set_filter(event, (void __user *)arg);
2031
2032         default:
2033                 return -ENOTTY;
2034         }
2035
2036         if (flags & PERF_IOC_FLAG_GROUP)
2037                 perf_event_for_each(event, func);
2038         else
2039                 perf_event_for_each_child(event, func);
2040
2041         return 0;
2042 }
2043
2044 int perf_event_task_enable(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_enable);
2051         mutex_unlock(&current->perf_event_mutex);
2052
2053         return 0;
2054 }
2055
2056 int perf_event_task_disable(void)
2057 {
2058         struct perf_event *event;
2059
2060         mutex_lock(&current->perf_event_mutex);
2061         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2062                 perf_event_for_each_child(event, perf_event_disable);
2063         mutex_unlock(&current->perf_event_mutex);
2064
2065         return 0;
2066 }
2067
2068 #ifndef PERF_EVENT_INDEX_OFFSET
2069 # define PERF_EVENT_INDEX_OFFSET 0
2070 #endif
2071
2072 static int perf_event_index(struct perf_event *event)
2073 {
2074         if (event->state != PERF_EVENT_STATE_ACTIVE)
2075                 return 0;
2076
2077         return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2078 }
2079
2080 /*
2081  * Callers need to ensure there can be no nesting of this function, otherwise
2082  * the seqlock logic goes bad. We can not serialize this because the arch
2083  * code calls this from NMI context.
2084  */
2085 void perf_event_update_userpage(struct perf_event *event)
2086 {
2087         struct perf_event_mmap_page *userpg;
2088         struct perf_mmap_data *data;
2089
2090         rcu_read_lock();
2091         data = rcu_dereference(event->data);
2092         if (!data)
2093                 goto unlock;
2094
2095         userpg = data->user_page;
2096
2097         /*
2098          * Disable preemption so as to not let the corresponding user-space
2099          * spin too long if we get preempted.
2100          */
2101         preempt_disable();
2102         ++userpg->lock;
2103         barrier();
2104         userpg->index = perf_event_index(event);
2105         userpg->offset = atomic64_read(&event->count);
2106         if (event->state == PERF_EVENT_STATE_ACTIVE)
2107                 userpg->offset -= atomic64_read(&event->hw.prev_count);
2108
2109         userpg->time_enabled = event->total_time_enabled +
2110                         atomic64_read(&event->child_total_time_enabled);
2111
2112         userpg->time_running = event->total_time_running +
2113                         atomic64_read(&event->child_total_time_running);
2114
2115         barrier();
2116         ++userpg->lock;
2117         preempt_enable();
2118 unlock:
2119         rcu_read_unlock();
2120 }
2121
2122 static unsigned long perf_data_size(struct perf_mmap_data *data)
2123 {
2124         return data->nr_pages << (PAGE_SHIFT + data->data_order);
2125 }
2126
2127 #ifndef CONFIG_PERF_USE_VMALLOC
2128
2129 /*
2130  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2131  */
2132
2133 static struct page *
2134 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2135 {
2136         if (pgoff > data->nr_pages)
2137                 return NULL;
2138
2139         if (pgoff == 0)
2140                 return virt_to_page(data->user_page);
2141
2142         return virt_to_page(data->data_pages[pgoff - 1]);
2143 }
2144
2145 static struct perf_mmap_data *
2146 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2147 {
2148         struct perf_mmap_data *data;
2149         unsigned long size;
2150         int i;
2151
2152         WARN_ON(atomic_read(&event->mmap_count));
2153
2154         size = sizeof(struct perf_mmap_data);
2155         size += nr_pages * sizeof(void *);
2156
2157         data = kzalloc(size, GFP_KERNEL);
2158         if (!data)
2159                 goto fail;
2160
2161         data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2162         if (!data->user_page)
2163                 goto fail_user_page;
2164
2165         for (i = 0; i < nr_pages; i++) {
2166                 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2167                 if (!data->data_pages[i])
2168                         goto fail_data_pages;
2169         }
2170
2171         data->data_order = 0;
2172         data->nr_pages = nr_pages;
2173
2174         return data;
2175
2176 fail_data_pages:
2177         for (i--; i >= 0; i--)
2178                 free_page((unsigned long)data->data_pages[i]);
2179
2180         free_page((unsigned long)data->user_page);
2181
2182 fail_user_page:
2183         kfree(data);
2184
2185 fail:
2186         return NULL;
2187 }
2188
2189 static void perf_mmap_free_page(unsigned long addr)
2190 {
2191         struct page *page = virt_to_page((void *)addr);
2192
2193         page->mapping = NULL;
2194         __free_page(page);
2195 }
2196
2197 static void perf_mmap_data_free(struct perf_mmap_data *data)
2198 {
2199         int i;
2200
2201         perf_mmap_free_page((unsigned long)data->user_page);
2202         for (i = 0; i < data->nr_pages; i++)
2203                 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2204 }
2205
2206 #else
2207
2208 /*
2209  * Back perf_mmap() with vmalloc memory.
2210  *
2211  * Required for architectures that have d-cache aliasing issues.
2212  */
2213
2214 static struct page *
2215 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2216 {
2217         if (pgoff > (1UL << data->data_order))
2218                 return NULL;
2219
2220         return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2221 }
2222
2223 static void perf_mmap_unmark_page(void *addr)
2224 {
2225         struct page *page = vmalloc_to_page(addr);
2226
2227         page->mapping = NULL;
2228 }
2229
2230 static void perf_mmap_data_free_work(struct work_struct *work)
2231 {
2232         struct perf_mmap_data *data;
2233         void *base;
2234         int i, nr;
2235
2236         data = container_of(work, struct perf_mmap_data, work);
2237         nr = 1 << data->data_order;
2238
2239         base = data->user_page;
2240         for (i = 0; i < nr + 1; i++)
2241                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2242
2243         vfree(base);
2244 }
2245
2246 static void perf_mmap_data_free(struct perf_mmap_data *data)
2247 {
2248         schedule_work(&data->work);
2249 }
2250
2251 static struct perf_mmap_data *
2252 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2253 {
2254         struct perf_mmap_data *data;
2255         unsigned long size;
2256         void *all_buf;
2257
2258         WARN_ON(atomic_read(&event->mmap_count));
2259
2260         size = sizeof(struct perf_mmap_data);
2261         size += sizeof(void *);
2262
2263         data = kzalloc(size, GFP_KERNEL);
2264         if (!data)
2265                 goto fail;
2266
2267         INIT_WORK(&data->work, perf_mmap_data_free_work);
2268
2269         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2270         if (!all_buf)
2271                 goto fail_all_buf;
2272
2273         data->user_page = all_buf;
2274         data->data_pages[0] = all_buf + PAGE_SIZE;
2275         data->data_order = ilog2(nr_pages);
2276         data->nr_pages = 1;
2277
2278         return data;
2279
2280 fail_all_buf:
2281         kfree(data);
2282
2283 fail:
2284         return NULL;
2285 }
2286
2287 #endif
2288
2289 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2290 {
2291         struct perf_event *event = vma->vm_file->private_data;
2292         struct perf_mmap_data *data;
2293         int ret = VM_FAULT_SIGBUS;
2294
2295         if (vmf->flags & FAULT_FLAG_MKWRITE) {
2296                 if (vmf->pgoff == 0)
2297                         ret = 0;
2298                 return ret;
2299         }
2300
2301         rcu_read_lock();
2302         data = rcu_dereference(event->data);
2303         if (!data)
2304                 goto unlock;
2305
2306         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2307                 goto unlock;
2308
2309         vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2310         if (!vmf->page)
2311                 goto unlock;
2312
2313         get_page(vmf->page);
2314         vmf->page->mapping = vma->vm_file->f_mapping;
2315         vmf->page->index   = vmf->pgoff;
2316
2317         ret = 0;
2318 unlock:
2319         rcu_read_unlock();
2320
2321         return ret;
2322 }
2323
2324 static void
2325 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2326 {
2327         long max_size = perf_data_size(data);
2328
2329         atomic_set(&data->lock, -1);
2330
2331         if (event->attr.watermark) {
2332                 data->watermark = min_t(long, max_size,
2333                                         event->attr.wakeup_watermark);
2334         }
2335
2336         if (!data->watermark)
2337                 data->watermark = max_t(long, PAGE_SIZE, max_size / 2);
2338
2339
2340         rcu_assign_pointer(event->data, data);
2341 }
2342
2343 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2344 {
2345         struct perf_mmap_data *data;
2346
2347         data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2348         perf_mmap_data_free(data);
2349         kfree(data);
2350 }
2351
2352 static void perf_mmap_data_release(struct perf_event *event)
2353 {
2354         struct perf_mmap_data *data = event->data;
2355
2356         WARN_ON(atomic_read(&event->mmap_count));
2357
2358         rcu_assign_pointer(event->data, NULL);
2359         call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2360 }
2361
2362 static void perf_mmap_open(struct vm_area_struct *vma)
2363 {
2364         struct perf_event *event = vma->vm_file->private_data;
2365
2366         atomic_inc(&event->mmap_count);
2367 }
2368
2369 static void perf_mmap_close(struct vm_area_struct *vma)
2370 {
2371         struct perf_event *event = vma->vm_file->private_data;
2372
2373         WARN_ON_ONCE(event->ctx->parent_ctx);
2374         if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2375                 unsigned long size = perf_data_size(event->data);
2376                 struct user_struct *user = current_user();
2377
2378                 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2379                 vma->vm_mm->locked_vm -= event->data->nr_locked;
2380                 perf_mmap_data_release(event);
2381                 mutex_unlock(&event->mmap_mutex);
2382         }
2383 }
2384
2385 static const struct vm_operations_struct perf_mmap_vmops = {
2386         .open           = perf_mmap_open,
2387         .close          = perf_mmap_close,
2388         .fault          = perf_mmap_fault,
2389         .page_mkwrite   = perf_mmap_fault,
2390 };
2391
2392 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2393 {
2394         struct perf_event *event = file->private_data;
2395         unsigned long user_locked, user_lock_limit;
2396         struct user_struct *user = current_user();
2397         unsigned long locked, lock_limit;
2398         struct perf_mmap_data *data;
2399         unsigned long vma_size;
2400         unsigned long nr_pages;
2401         long user_extra, extra;
2402         int ret = 0;
2403
2404         if (!(vma->vm_flags & VM_SHARED))
2405                 return -EINVAL;
2406
2407         vma_size = vma->vm_end - vma->vm_start;
2408         nr_pages = (vma_size / PAGE_SIZE) - 1;
2409
2410         /*
2411          * If we have data pages ensure they're a power-of-two number, so we
2412          * can do bitmasks instead of modulo.
2413          */
2414         if (nr_pages != 0 && !is_power_of_2(nr_pages))
2415                 return -EINVAL;
2416
2417         if (vma_size != PAGE_SIZE * (1 + nr_pages))
2418                 return -EINVAL;
2419
2420         if (vma->vm_pgoff != 0)
2421                 return -EINVAL;
2422
2423         WARN_ON_ONCE(event->ctx->parent_ctx);
2424         mutex_lock(&event->mmap_mutex);
2425         if (event->output) {
2426                 ret = -EINVAL;
2427                 goto unlock;
2428         }
2429
2430         if (atomic_inc_not_zero(&event->mmap_count)) {
2431                 if (nr_pages != event->data->nr_pages)
2432                         ret = -EINVAL;
2433                 goto unlock;
2434         }
2435
2436         user_extra = nr_pages + 1;
2437         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2438
2439         /*
2440          * Increase the limit linearly with more CPUs:
2441          */
2442         user_lock_limit *= num_online_cpus();
2443
2444         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2445
2446         extra = 0;
2447         if (user_locked > user_lock_limit)
2448                 extra = user_locked - user_lock_limit;
2449
2450         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2451         lock_limit >>= PAGE_SHIFT;
2452         locked = vma->vm_mm->locked_vm + extra;
2453
2454         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2455                 !capable(CAP_IPC_LOCK)) {
2456                 ret = -EPERM;
2457                 goto unlock;
2458         }
2459
2460         WARN_ON(event->data);
2461
2462         data = perf_mmap_data_alloc(event, nr_pages);
2463         ret = -ENOMEM;
2464         if (!data)
2465                 goto unlock;
2466
2467         ret = 0;
2468         perf_mmap_data_init(event, data);
2469
2470         atomic_set(&event->mmap_count, 1);
2471         atomic_long_add(user_extra, &user->locked_vm);
2472         vma->vm_mm->locked_vm += extra;
2473         event->data->nr_locked = extra;
2474         if (vma->vm_flags & VM_WRITE)
2475                 event->data->writable = 1;
2476
2477 unlock:
2478         mutex_unlock(&event->mmap_mutex);
2479
2480         vma->vm_flags |= VM_RESERVED;
2481         vma->vm_ops = &perf_mmap_vmops;
2482
2483         return ret;
2484 }
2485
2486 static int perf_fasync(int fd, struct file *filp, int on)
2487 {
2488         struct inode *inode = filp->f_path.dentry->d_inode;
2489         struct perf_event *event = filp->private_data;
2490         int retval;
2491
2492         mutex_lock(&inode->i_mutex);
2493         retval = fasync_helper(fd, filp, on, &event->fasync);
2494         mutex_unlock(&inode->i_mutex);
2495
2496         if (retval < 0)
2497                 return retval;
2498
2499         return 0;
2500 }
2501
2502 static const struct file_operations perf_fops = {
2503         .release                = perf_release,
2504         .read                   = perf_read,
2505         .poll                   = perf_poll,
2506         .unlocked_ioctl         = perf_ioctl,
2507         .compat_ioctl           = perf_ioctl,
2508         .mmap                   = perf_mmap,
2509         .fasync                 = perf_fasync,
2510 };
2511
2512 /*
2513  * Perf event wakeup
2514  *
2515  * If there's data, ensure we set the poll() state and publish everything
2516  * to user-space before waking everybody up.
2517  */
2518
2519 void perf_event_wakeup(struct perf_event *event)
2520 {
2521         wake_up_all(&event->waitq);
2522
2523         if (event->pending_kill) {
2524                 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2525                 event->pending_kill = 0;
2526         }
2527 }
2528
2529 /*
2530  * Pending wakeups
2531  *
2532  * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2533  *
2534  * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2535  * single linked list and use cmpxchg() to add entries lockless.
2536  */
2537
2538 static void perf_pending_event(struct perf_pending_entry *entry)
2539 {
2540         struct perf_event *event = container_of(entry,
2541                         struct perf_event, pending);
2542
2543         if (event->pending_disable) {
2544                 event->pending_disable = 0;
2545                 __perf_event_disable(event);
2546         }
2547
2548         if (event->pending_wakeup) {
2549                 event->pending_wakeup = 0;
2550                 perf_event_wakeup(event);
2551         }
2552 }
2553
2554 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2555
2556 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2557         PENDING_TAIL,
2558 };
2559
2560 static void perf_pending_queue(struct perf_pending_entry *entry,
2561                                void (*func)(struct perf_pending_entry *))
2562 {
2563         struct perf_pending_entry **head;
2564
2565         if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2566                 return;
2567
2568         entry->func = func;
2569
2570         head = &get_cpu_var(perf_pending_head);
2571
2572         do {
2573                 entry->next = *head;
2574         } while (cmpxchg(head, entry->next, entry) != entry->next);
2575
2576         set_perf_event_pending();
2577
2578         put_cpu_var(perf_pending_head);
2579 }
2580
2581 static int __perf_pending_run(void)
2582 {
2583         struct perf_pending_entry *list;
2584         int nr = 0;
2585
2586         list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2587         while (list != PENDING_TAIL) {
2588                 void (*func)(struct perf_pending_entry *);
2589                 struct perf_pending_entry *entry = list;
2590
2591                 list = list->next;
2592
2593                 func = entry->func;
2594                 entry->next = NULL;
2595                 /*
2596                  * Ensure we observe the unqueue before we issue the wakeup,
2597                  * so that we won't be waiting forever.
2598                  * -- see perf_not_pending().
2599                  */
2600                 smp_wmb();
2601
2602                 func(entry);
2603                 nr++;
2604         }
2605
2606         return nr;
2607 }
2608
2609 static inline int perf_not_pending(struct perf_event *event)
2610 {
2611         /*
2612          * If we flush on whatever cpu we run, there is a chance we don't
2613          * need to wait.
2614          */
2615         get_cpu();
2616         __perf_pending_run();
2617         put_cpu();
2618
2619         /*
2620          * Ensure we see the proper queue state before going to sleep
2621          * so that we do not miss the wakeup. -- see perf_pending_handle()
2622          */
2623         smp_rmb();
2624         return event->pending.next == NULL;
2625 }
2626
2627 static void perf_pending_sync(struct perf_event *event)
2628 {
2629         wait_event(event->waitq, perf_not_pending(event));
2630 }
2631
2632 void perf_event_do_pending(void)
2633 {
2634         __perf_pending_run();
2635 }
2636
2637 /*
2638  * Callchain support -- arch specific
2639  */
2640
2641 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2642 {
2643         return NULL;
2644 }
2645
2646 /*
2647  * Output
2648  */
2649 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2650                               unsigned long offset, unsigned long head)
2651 {
2652         unsigned long mask;
2653
2654         if (!data->writable)
2655                 return true;
2656
2657         mask = perf_data_size(data) - 1;
2658
2659         offset = (offset - tail) & mask;
2660         head   = (head   - tail) & mask;
2661
2662         if ((int)(head - offset) < 0)
2663                 return false;
2664
2665         return true;
2666 }
2667
2668 static void perf_output_wakeup(struct perf_output_handle *handle)
2669 {
2670         atomic_set(&handle->data->poll, POLL_IN);
2671
2672         if (handle->nmi) {
2673                 handle->event->pending_wakeup = 1;
2674                 perf_pending_queue(&handle->event->pending,
2675                                    perf_pending_event);
2676         } else
2677                 perf_event_wakeup(handle->event);
2678 }
2679
2680 /*
2681  * Curious locking construct.
2682  *
2683  * We need to ensure a later event_id doesn't publish a head when a former
2684  * event_id isn't done writing. However since we need to deal with NMIs we
2685  * cannot fully serialize things.
2686  *
2687  * What we do is serialize between CPUs so we only have to deal with NMI
2688  * nesting on a single CPU.
2689  *
2690  * We only publish the head (and generate a wakeup) when the outer-most
2691  * event_id completes.
2692  */
2693 static void perf_output_lock(struct perf_output_handle *handle)
2694 {
2695         struct perf_mmap_data *data = handle->data;
2696         int cur, cpu = get_cpu();
2697
2698         handle->locked = 0;
2699
2700         for (;;) {
2701                 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2702                 if (cur == -1) {
2703                         handle->locked = 1;
2704                         break;
2705                 }
2706                 if (cur == cpu)
2707                         break;
2708
2709                 cpu_relax();
2710         }
2711 }
2712
2713 static void perf_output_unlock(struct perf_output_handle *handle)
2714 {
2715         struct perf_mmap_data *data = handle->data;
2716         unsigned long head;
2717         int cpu;
2718
2719         data->done_head = data->head;
2720
2721         if (!handle->locked)
2722                 goto out;
2723
2724 again:
2725         /*
2726          * The xchg implies a full barrier that ensures all writes are done
2727          * before we publish the new head, matched by a rmb() in userspace when
2728          * reading this position.
2729          */
2730         while ((head = atomic_long_xchg(&data->done_head, 0)))
2731                 data->user_page->data_head = head;
2732
2733         /*
2734          * NMI can happen here, which means we can miss a done_head update.
2735          */
2736
2737         cpu = atomic_xchg(&data->lock, -1);
2738         WARN_ON_ONCE(cpu != smp_processor_id());
2739
2740         /*
2741          * Therefore we have to validate we did not indeed do so.
2742          */
2743         if (unlikely(atomic_long_read(&data->done_head))) {
2744                 /*
2745                  * Since we had it locked, we can lock it again.
2746                  */
2747                 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2748                         cpu_relax();
2749
2750                 goto again;
2751         }
2752
2753         if (atomic_xchg(&data->wakeup, 0))
2754                 perf_output_wakeup(handle);
2755 out:
2756         put_cpu();
2757 }
2758
2759 void perf_output_copy(struct perf_output_handle *handle,
2760                       const void *buf, unsigned int len)
2761 {
2762         unsigned int pages_mask;
2763         unsigned long offset;
2764         unsigned int size;
2765         void **pages;
2766
2767         offset          = handle->offset;
2768         pages_mask      = handle->data->nr_pages - 1;
2769         pages           = handle->data->data_pages;
2770
2771         do {
2772                 unsigned long page_offset;
2773                 unsigned long page_size;
2774                 int nr;
2775
2776                 nr          = (offset >> PAGE_SHIFT) & pages_mask;
2777                 page_size   = 1UL << (handle->data->data_order + PAGE_SHIFT);
2778                 page_offset = offset & (page_size - 1);
2779                 size        = min_t(unsigned int, page_size - page_offset, len);
2780
2781                 memcpy(pages[nr] + page_offset, buf, size);
2782
2783                 len         -= size;
2784                 buf         += size;
2785                 offset      += size;
2786         } while (len);
2787
2788         handle->offset = offset;
2789
2790         /*
2791          * Check we didn't copy past our reservation window, taking the
2792          * possible unsigned int wrap into account.
2793          */
2794         WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2795 }
2796
2797 int perf_output_begin(struct perf_output_handle *handle,
2798                       struct perf_event *event, unsigned int size,
2799                       int nmi, int sample)
2800 {
2801         struct perf_event *output_event;
2802         struct perf_mmap_data *data;
2803         unsigned long tail, offset, head;
2804         int have_lost;
2805         struct {
2806                 struct perf_event_header header;
2807                 u64                      id;
2808                 u64                      lost;
2809         } lost_event;
2810
2811         rcu_read_lock();
2812         /*
2813          * For inherited events we send all the output towards the parent.
2814          */
2815         if (event->parent)
2816                 event = event->parent;
2817
2818         output_event = rcu_dereference(event->output);
2819         if (output_event)
2820                 event = output_event;
2821
2822         data = rcu_dereference(event->data);
2823         if (!data)
2824                 goto out;
2825
2826         handle->data    = data;
2827         handle->event   = event;
2828         handle->nmi     = nmi;
2829         handle->sample  = sample;
2830
2831         if (!data->nr_pages)
2832                 goto fail;
2833
2834         have_lost = atomic_read(&data->lost);
2835         if (have_lost)
2836                 size += sizeof(lost_event);
2837
2838         perf_output_lock(handle);
2839
2840         do {
2841                 /*
2842                  * Userspace could choose to issue a mb() before updating the
2843                  * tail pointer. So that all reads will be completed before the
2844                  * write is issued.
2845                  */
2846                 tail = ACCESS_ONCE(data->user_page->data_tail);
2847                 smp_rmb();
2848                 offset = head = atomic_long_read(&data->head);
2849                 head += size;
2850                 if (unlikely(!perf_output_space(data, tail, offset, head)))
2851                         goto fail;
2852         } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2853
2854         handle->offset  = offset;
2855         handle->head    = head;
2856
2857         if (head - tail > data->watermark)
2858                 atomic_set(&data->wakeup, 1);
2859
2860         if (have_lost) {
2861                 lost_event.header.type = PERF_RECORD_LOST;
2862                 lost_event.header.misc = 0;
2863                 lost_event.header.size = sizeof(lost_event);
2864                 lost_event.id          = event->id;
2865                 lost_event.lost        = atomic_xchg(&data->lost, 0);
2866
2867                 perf_output_put(handle, lost_event);
2868         }
2869
2870         return 0;
2871
2872 fail:
2873         atomic_inc(&data->lost);
2874         perf_output_unlock(handle);
2875 out:
2876         rcu_read_unlock();
2877
2878         return -ENOSPC;
2879 }
2880
2881 void perf_output_end(struct perf_output_handle *handle)
2882 {
2883         struct perf_event *event = handle->event;
2884         struct perf_mmap_data *data = handle->data;
2885
2886         int wakeup_events = event->attr.wakeup_events;
2887
2888         if (handle->sample && wakeup_events) {
2889                 int events = atomic_inc_return(&data->events);
2890                 if (events >= wakeup_events) {
2891                         atomic_sub(wakeup_events, &data->events);
2892                         atomic_set(&data->wakeup, 1);
2893                 }
2894         }
2895
2896         perf_output_unlock(handle);
2897         rcu_read_unlock();
2898 }
2899
2900 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
2901 {
2902         /*
2903          * only top level events have the pid namespace they were created in
2904          */
2905         if (event->parent)
2906                 event = event->parent;
2907
2908         return task_tgid_nr_ns(p, event->ns);
2909 }
2910
2911 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
2912 {
2913         /*
2914          * only top level events have the pid namespace they were created in
2915          */
2916         if (event->parent)
2917                 event = event->parent;
2918
2919         return task_pid_nr_ns(p, event->ns);
2920 }
2921
2922 static void perf_output_read_one(struct perf_output_handle *handle,
2923                                  struct perf_event *event)
2924 {
2925         u64 read_format = event->attr.read_format;
2926         u64 values[4];
2927         int n = 0;
2928
2929         values[n++] = atomic64_read(&event->count);
2930         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2931                 values[n++] = event->total_time_enabled +
2932                         atomic64_read(&event->child_total_time_enabled);
2933         }
2934         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2935                 values[n++] = event->total_time_running +
2936                         atomic64_read(&event->child_total_time_running);
2937         }
2938         if (read_format & PERF_FORMAT_ID)
2939                 values[n++] = primary_event_id(event);
2940
2941         perf_output_copy(handle, values, n * sizeof(u64));
2942 }
2943
2944 /*
2945  * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
2946  */
2947 static void perf_output_read_group(struct perf_output_handle *handle,
2948                             struct perf_event *event)
2949 {
2950         struct perf_event *leader = event->group_leader, *sub;
2951         u64 read_format = event->attr.read_format;
2952         u64 values[5];
2953         int n = 0;
2954
2955         values[n++] = 1 + leader->nr_siblings;
2956
2957         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2958                 values[n++] = leader->total_time_enabled;
2959
2960         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2961                 values[n++] = leader->total_time_running;
2962
2963         if (leader != event)
2964                 leader->pmu->read(leader);
2965
2966         values[n++] = atomic64_read(&leader->count);
2967         if (read_format & PERF_FORMAT_ID)
2968                 values[n++] = primary_event_id(leader);
2969
2970         perf_output_copy(handle, values, n * sizeof(u64));
2971
2972         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2973                 n = 0;
2974
2975                 if (sub != event)
2976                         sub->pmu->read(sub);
2977
2978                 values[n++] = atomic64_read(&sub->count);
2979                 if (read_format & PERF_FORMAT_ID)
2980                         values[n++] = primary_event_id(sub);
2981
2982                 perf_output_copy(handle, values, n * sizeof(u64));
2983         }
2984 }
2985
2986 static void perf_output_read(struct perf_output_handle *handle,
2987                              struct perf_event *event)
2988 {
2989         if (event->attr.read_format & PERF_FORMAT_GROUP)
2990                 perf_output_read_group(handle, event);
2991         else
2992                 perf_output_read_one(handle, event);
2993 }
2994
2995 void perf_output_sample(struct perf_output_handle *handle,
2996                         struct perf_event_header *header,
2997                         struct perf_sample_data *data,
2998                         struct perf_event *event)
2999 {
3000         u64 sample_type = data->type;
3001
3002         perf_output_put(handle, *header);
3003
3004         if (sample_type & PERF_SAMPLE_IP)
3005                 perf_output_put(handle, data->ip);
3006
3007         if (sample_type & PERF_SAMPLE_TID)
3008                 perf_output_put(handle, data->tid_entry);
3009
3010         if (sample_type & PERF_SAMPLE_TIME)
3011                 perf_output_put(handle, data->time);
3012
3013         if (sample_type & PERF_SAMPLE_ADDR)
3014                 perf_output_put(handle, data->addr);
3015
3016         if (sample_type & PERF_SAMPLE_ID)
3017                 perf_output_put(handle, data->id);
3018
3019         if (sample_type & PERF_SAMPLE_STREAM_ID)
3020                 perf_output_put(handle, data->stream_id);
3021
3022         if (sample_type & PERF_SAMPLE_CPU)
3023                 perf_output_put(handle, data->cpu_entry);
3024
3025         if (sample_type & PERF_SAMPLE_PERIOD)
3026                 perf_output_put(handle, data->period);
3027
3028         if (sample_type & PERF_SAMPLE_READ)
3029                 perf_output_read(handle, event);
3030
3031         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3032                 if (data->callchain) {
3033                         int size = 1;
3034
3035                         if (data->callchain)
3036                                 size += data->callchain->nr;
3037
3038                         size *= sizeof(u64);
3039
3040                         perf_output_copy(handle, data->callchain, size);
3041                 } else {
3042                         u64 nr = 0;
3043                         perf_output_put(handle, nr);
3044                 }
3045         }
3046
3047         if (sample_type & PERF_SAMPLE_RAW) {
3048                 if (data->raw) {
3049                         perf_output_put(handle, data->raw->size);
3050                         perf_output_copy(handle, data->raw->data,
3051                                          data->raw->size);
3052                 } else {
3053                         struct {
3054                                 u32     size;
3055                                 u32     data;
3056                         } raw = {
3057                                 .size = sizeof(u32),
3058                                 .data = 0,
3059                         };
3060                         perf_output_put(handle, raw);
3061                 }
3062         }
3063 }
3064
3065 void perf_prepare_sample(struct perf_event_header *header,
3066                          struct perf_sample_data *data,
3067                          struct perf_event *event,
3068                          struct pt_regs *regs)
3069 {
3070         u64 sample_type = event->attr.sample_type;
3071
3072         data->type = sample_type;
3073
3074         header->type = PERF_RECORD_SAMPLE;
3075         header->size = sizeof(*header);
3076
3077         header->misc = 0;
3078         header->misc |= perf_misc_flags(regs);
3079
3080         if (sample_type & PERF_SAMPLE_IP) {
3081                 data->ip = perf_instruction_pointer(regs);
3082
3083                 header->size += sizeof(data->ip);
3084         }
3085
3086         if (sample_type & PERF_SAMPLE_TID) {
3087                 /* namespace issues */
3088                 data->tid_entry.pid = perf_event_pid(event, current);
3089                 data->tid_entry.tid = perf_event_tid(event, current);
3090
3091                 header->size += sizeof(data->tid_entry);
3092         }
3093
3094         if (sample_type & PERF_SAMPLE_TIME) {
3095                 data->time = perf_clock();
3096
3097                 header->size += sizeof(data->time);
3098         }
3099
3100         if (sample_type & PERF_SAMPLE_ADDR)
3101                 header->size += sizeof(data->addr);
3102
3103         if (sample_type & PERF_SAMPLE_ID) {
3104                 data->id = primary_event_id(event);
3105
3106                 header->size += sizeof(data->id);
3107         }
3108
3109         if (sample_type & PERF_SAMPLE_STREAM_ID) {
3110                 data->stream_id = event->id;
3111
3112                 header->size += sizeof(data->stream_id);
3113         }
3114
3115         if (sample_type & PERF_SAMPLE_CPU) {
3116                 data->cpu_entry.cpu             = raw_smp_processor_id();
3117                 data->cpu_entry.reserved        = 0;
3118
3119                 header->size += sizeof(data->cpu_entry);
3120         }
3121
3122         if (sample_type & PERF_SAMPLE_PERIOD)
3123                 header->size += sizeof(data->period);
3124
3125         if (sample_type & PERF_SAMPLE_READ)
3126                 header->size += perf_event_read_size(event);
3127
3128         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3129                 int size = 1;
3130
3131                 data->callchain = perf_callchain(regs);
3132
3133                 if (data->callchain)
3134                         size += data->callchain->nr;
3135
3136                 header->size += size * sizeof(u64);
3137         }
3138
3139         if (sample_type & PERF_SAMPLE_RAW) {
3140                 int size = sizeof(u32);
3141
3142                 if (data->raw)
3143                         size += data->raw->size;
3144                 else
3145                         size += sizeof(u32);
3146
3147                 WARN_ON_ONCE(size & (sizeof(u64)-1));
3148                 header->size += size;
3149         }
3150 }
3151
3152 static void perf_event_output(struct perf_event *event, int nmi,
3153                                 struct perf_sample_data *data,
3154                                 struct pt_regs *regs)
3155 {
3156         struct perf_output_handle handle;
3157         struct perf_event_header header;
3158
3159         perf_prepare_sample(&header, data, event, regs);
3160
3161         if (perf_output_begin(&handle, event, header.size, nmi, 1))
3162                 return;
3163
3164         perf_output_sample(&handle, &header, data, event);
3165
3166         perf_output_end(&handle);
3167 }
3168
3169 /*
3170  * read event_id
3171  */
3172
3173 struct perf_read_event {
3174         struct perf_event_header        header;
3175
3176         u32                             pid;
3177         u32                             tid;
3178 };
3179
3180 static void
3181 perf_event_read_event(struct perf_event *event,
3182                         struct task_struct *task)
3183 {
3184         struct perf_output_handle handle;
3185         struct perf_read_event read_event = {
3186                 .header = {
3187                         .type = PERF_RECORD_READ,
3188                         .misc = 0,
3189                         .size = sizeof(read_event) + perf_event_read_size(event),
3190                 },
3191                 .pid = perf_event_pid(event, task),
3192                 .tid = perf_event_tid(event, task),
3193         };
3194         int ret;
3195
3196         ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3197         if (ret)
3198                 return;
3199
3200         perf_output_put(&handle, read_event);
3201         perf_output_read(&handle, event);
3202
3203         perf_output_end(&handle);
3204 }
3205
3206 /*
3207  * task tracking -- fork/exit
3208  *
3209  * enabled by: attr.comm | attr.mmap | attr.task
3210  */
3211
3212 struct perf_task_event {
3213         struct task_struct              *task;
3214         struct perf_event_context       *task_ctx;
3215
3216         struct {
3217                 struct perf_event_header        header;
3218
3219                 u32                             pid;
3220                 u32                             ppid;
3221                 u32                             tid;
3222                 u32                             ptid;
3223                 u64                             time;
3224         } event_id;
3225 };
3226
3227 static void perf_event_task_output(struct perf_event *event,
3228                                      struct perf_task_event *task_event)
3229 {
3230         struct perf_output_handle handle;
3231         int size;
3232         struct task_struct *task = task_event->task;
3233         int ret;
3234
3235         size  = task_event->event_id.header.size;
3236         ret = perf_output_begin(&handle, event, size, 0, 0);
3237
3238         if (ret)
3239                 return;
3240
3241         task_event->event_id.pid = perf_event_pid(event, task);
3242         task_event->event_id.ppid = perf_event_pid(event, current);
3243
3244         task_event->event_id.tid = perf_event_tid(event, task);
3245         task_event->event_id.ptid = perf_event_tid(event, current);
3246
3247         task_event->event_id.time = perf_clock();
3248
3249         perf_output_put(&handle, task_event->event_id);
3250
3251         perf_output_end(&handle);
3252 }
3253
3254 static int perf_event_task_match(struct perf_event *event)
3255 {
3256         if (event->attr.comm || event->attr.mmap || event->attr.task)
3257                 return 1;
3258
3259         return 0;
3260 }
3261
3262 static void perf_event_task_ctx(struct perf_event_context *ctx,
3263                                   struct perf_task_event *task_event)
3264 {
3265         struct perf_event *event;
3266
3267         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3268                 if (perf_event_task_match(event))
3269                         perf_event_task_output(event, task_event);
3270         }
3271 }
3272
3273 static void perf_event_task_event(struct perf_task_event *task_event)
3274 {
3275         struct perf_cpu_context *cpuctx;
3276         struct perf_event_context *ctx = task_event->task_ctx;
3277
3278         rcu_read_lock();
3279         cpuctx = &get_cpu_var(perf_cpu_context);
3280         perf_event_task_ctx(&cpuctx->ctx, task_event);
3281         put_cpu_var(perf_cpu_context);
3282
3283         if (!ctx)
3284                 ctx = rcu_dereference(task_event->task->perf_event_ctxp);
3285         if (ctx)
3286                 perf_event_task_ctx(ctx, task_event);
3287         rcu_read_unlock();
3288 }
3289
3290 static void perf_event_task(struct task_struct *task,
3291                               struct perf_event_context *task_ctx,
3292                               int new)
3293 {
3294         struct perf_task_event task_event;
3295
3296         if (!atomic_read(&nr_comm_events) &&
3297             !atomic_read(&nr_mmap_events) &&
3298             !atomic_read(&nr_task_events))
3299                 return;
3300
3301         task_event = (struct perf_task_event){
3302                 .task     = task,
3303                 .task_ctx = task_ctx,
3304                 .event_id    = {
3305                         .header = {
3306                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3307                                 .misc = 0,
3308                                 .size = sizeof(task_event.event_id),
3309                         },
3310                         /* .pid  */
3311                         /* .ppid */
3312                         /* .tid  */
3313                         /* .ptid */
3314                 },
3315         };
3316
3317         perf_event_task_event(&task_event);
3318 }
3319
3320 void perf_event_fork(struct task_struct *task)
3321 {
3322         perf_event_task(task, NULL, 1);
3323 }
3324
3325 /*
3326  * comm tracking
3327  */
3328
3329 struct perf_comm_event {
3330         struct task_struct      *task;
3331         char                    *comm;
3332         int                     comm_size;
3333
3334         struct {
3335                 struct perf_event_header        header;
3336
3337                 u32                             pid;
3338                 u32                             tid;
3339         } event_id;
3340 };
3341
3342 static void perf_event_comm_output(struct perf_event *event,
3343                                      struct perf_comm_event *comm_event)
3344 {
3345         struct perf_output_handle handle;
3346         int size = comm_event->event_id.header.size;
3347         int ret = perf_output_begin(&handle, event, size, 0, 0);
3348
3349         if (ret)
3350                 return;
3351
3352         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3353         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3354
3355         perf_output_put(&handle, comm_event->event_id);
3356         perf_output_copy(&handle, comm_event->comm,
3357                                    comm_event->comm_size);
3358         perf_output_end(&handle);
3359 }
3360
3361 static int perf_event_comm_match(struct perf_event *event)
3362 {
3363         if (event->attr.comm)
3364                 return 1;
3365
3366         return 0;
3367 }
3368
3369 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3370                                   struct perf_comm_event *comm_event)
3371 {
3372         struct perf_event *event;
3373
3374         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3375                 if (perf_event_comm_match(event))
3376                         perf_event_comm_output(event, comm_event);
3377         }
3378 }
3379
3380 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3381 {
3382         struct perf_cpu_context *cpuctx;
3383         struct perf_event_context *ctx;
3384         unsigned int size;
3385         char comm[TASK_COMM_LEN];
3386
3387         memset(comm, 0, sizeof(comm));
3388         strncpy(comm, comm_event->task->comm, sizeof(comm));
3389         size = ALIGN(strlen(comm)+1, sizeof(u64));
3390
3391         comm_event->comm = comm;
3392         comm_event->comm_size = size;
3393
3394         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3395
3396         rcu_read_lock();
3397         cpuctx = &get_cpu_var(perf_cpu_context);
3398         perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3399         put_cpu_var(perf_cpu_context);
3400
3401         /*
3402          * doesn't really matter which of the child contexts the
3403          * events ends up in.
3404          */
3405         ctx = rcu_dereference(current->perf_event_ctxp);
3406         if (ctx)
3407                 perf_event_comm_ctx(ctx, comm_event);
3408         rcu_read_unlock();
3409 }
3410
3411 void perf_event_comm(struct task_struct *task)
3412 {
3413         struct perf_comm_event comm_event;
3414
3415         if (task->perf_event_ctxp)
3416                 perf_event_enable_on_exec(task);
3417
3418         if (!atomic_read(&nr_comm_events))
3419                 return;
3420
3421         comm_event = (struct perf_comm_event){
3422                 .task   = task,
3423                 /* .comm      */
3424                 /* .comm_size */
3425                 .event_id  = {
3426                         .header = {
3427                                 .type = PERF_RECORD_COMM,
3428                                 .misc = 0,
3429                                 /* .size */
3430                         },
3431                         /* .pid */
3432                         /* .tid */
3433                 },
3434         };
3435
3436         perf_event_comm_event(&comm_event);
3437 }
3438
3439 /*
3440  * mmap tracking
3441  */
3442
3443 struct perf_mmap_event {
3444         struct vm_area_struct   *vma;
3445
3446         const char              *file_name;
3447         int                     file_size;
3448
3449         struct {
3450                 struct perf_event_header        header;
3451
3452                 u32                             pid;
3453                 u32                             tid;
3454                 u64                             start;
3455                 u64                             len;
3456                 u64                             pgoff;
3457         } event_id;
3458 };
3459
3460 static void perf_event_mmap_output(struct perf_event *event,
3461                                      struct perf_mmap_event *mmap_event)
3462 {
3463         struct perf_output_handle handle;
3464         int size = mmap_event->event_id.header.size;
3465         int ret = perf_output_begin(&handle, event, size, 0, 0);
3466
3467         if (ret)
3468                 return;
3469
3470         mmap_event->event_id.pid = perf_event_pid(event, current);
3471         mmap_event->event_id.tid = perf_event_tid(event, current);
3472
3473         perf_output_put(&handle, mmap_event->event_id);
3474         perf_output_copy(&handle, mmap_event->file_name,
3475                                    mmap_event->file_size);
3476         perf_output_end(&handle);
3477 }
3478
3479 static int perf_event_mmap_match(struct perf_event *event,
3480                                    struct perf_mmap_event *mmap_event)
3481 {
3482         if (event->attr.mmap)
3483                 return 1;
3484
3485         return 0;
3486 }
3487
3488 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3489                                   struct perf_mmap_event *mmap_event)
3490 {
3491         struct perf_event *event;
3492
3493         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3494                 if (perf_event_mmap_match(event, mmap_event))
3495                         perf_event_mmap_output(event, mmap_event);
3496         }
3497 }
3498
3499 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3500 {
3501         struct perf_cpu_context *cpuctx;
3502         struct perf_event_context *ctx;
3503         struct vm_area_struct *vma = mmap_event->vma;
3504         struct file *file = vma->vm_file;
3505         unsigned int size;
3506         char tmp[16];
3507         char *buf = NULL;
3508         const char *name;
3509
3510         memset(tmp, 0, sizeof(tmp));
3511
3512         if (file) {
3513                 /*
3514                  * d_path works from the end of the buffer backwards, so we
3515                  * need to add enough zero bytes after the string to handle
3516                  * the 64bit alignment we do later.
3517                  */
3518                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3519                 if (!buf) {
3520                         name = strncpy(tmp, "//enomem", sizeof(tmp));
3521                         goto got_name;
3522                 }
3523                 name = d_path(&file->f_path, buf, PATH_MAX);
3524                 if (IS_ERR(name)) {
3525                         name = strncpy(tmp, "//toolong", sizeof(tmp));
3526                         goto got_name;
3527                 }
3528         } else {
3529                 if (arch_vma_name(mmap_event->vma)) {
3530                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3531                                        sizeof(tmp));
3532                         goto got_name;
3533                 }
3534
3535                 if (!vma->vm_mm) {
3536                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
3537                         goto got_name;
3538                 }
3539
3540                 name = strncpy(tmp, "//anon", sizeof(tmp));
3541                 goto got_name;
3542         }
3543
3544 got_name:
3545         size = ALIGN(strlen(name)+1, sizeof(u64));
3546
3547         mmap_event->file_name = name;
3548         mmap_event->file_size = size;
3549
3550         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3551
3552         rcu_read_lock();
3553         cpuctx = &get_cpu_var(perf_cpu_context);
3554         perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3555         put_cpu_var(perf_cpu_context);
3556
3557         /*
3558          * doesn't really matter which of the child contexts the
3559          * events ends up in.
3560          */
3561         ctx = rcu_dereference(current->perf_event_ctxp);
3562         if (ctx)
3563                 perf_event_mmap_ctx(ctx, mmap_event);
3564         rcu_read_unlock();
3565
3566         kfree(buf);
3567 }
3568
3569 void __perf_event_mmap(struct vm_area_struct *vma)
3570 {
3571         struct perf_mmap_event mmap_event;
3572
3573         if (!atomic_read(&nr_mmap_events))
3574                 return;
3575
3576         mmap_event = (struct perf_mmap_event){
3577                 .vma    = vma,
3578                 /* .file_name */
3579                 /* .file_size */
3580                 .event_id  = {
3581                         .header = {
3582                                 .type = PERF_RECORD_MMAP,
3583                                 .misc = 0,
3584                                 /* .size */
3585                         },
3586                         /* .pid */
3587                         /* .tid */
3588                         .start  = vma->vm_start,
3589                         .len    = vma->vm_end - vma->vm_start,
3590                         .pgoff  = vma->vm_pgoff,
3591                 },
3592         };
3593
3594         perf_event_mmap_event(&mmap_event);
3595 }
3596
3597 /*
3598  * IRQ throttle logging
3599  */
3600
3601 static void perf_log_throttle(struct perf_event *event, int enable)
3602 {
3603         struct perf_output_handle handle;
3604         int ret;
3605
3606         struct {
3607                 struct perf_event_header        header;
3608                 u64                             time;
3609                 u64                             id;
3610                 u64                             stream_id;
3611         } throttle_event = {
3612                 .header = {
3613                         .type = PERF_RECORD_THROTTLE,
3614                         .misc = 0,
3615                         .size = sizeof(throttle_event),
3616                 },
3617                 .time           = perf_clock(),
3618                 .id             = primary_event_id(event),
3619                 .stream_id      = event->id,
3620         };
3621
3622         if (enable)
3623                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3624
3625         ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3626         if (ret)
3627                 return;
3628
3629         perf_output_put(&handle, throttle_event);
3630         perf_output_end(&handle);
3631 }
3632
3633 /*
3634  * Generic event overflow handling, sampling.
3635  */
3636
3637 static int __perf_event_overflow(struct perf_event *event, int nmi,
3638                                    int throttle, struct perf_sample_data *data,
3639                                    struct pt_regs *regs)
3640 {
3641         int events = atomic_read(&event->event_limit);
3642         struct hw_perf_event *hwc = &event->hw;
3643         int ret = 0;
3644
3645         throttle = (throttle && event->pmu->unthrottle != NULL);
3646
3647         if (!throttle) {
3648                 hwc->interrupts++;
3649         } else {
3650                 if (hwc->interrupts != MAX_INTERRUPTS) {
3651                         hwc->interrupts++;
3652                         if (HZ * hwc->interrupts >
3653                                         (u64)sysctl_perf_event_sample_rate) {
3654                                 hwc->interrupts = MAX_INTERRUPTS;
3655                                 perf_log_throttle(event, 0);
3656                                 ret = 1;
3657                         }
3658                 } else {
3659                         /*
3660                          * Keep re-disabling events even though on the previous
3661                          * pass we disabled it - just in case we raced with a
3662                          * sched-in and the event got enabled again:
3663                          */
3664                         ret = 1;
3665                 }
3666         }
3667
3668         if (event->attr.freq) {
3669                 u64 now = perf_clock();
3670                 s64 delta = now - hwc->freq_stamp;
3671
3672                 hwc->freq_stamp = now;
3673
3674                 if (delta > 0 && delta < TICK_NSEC)
3675                         perf_adjust_period(event, NSEC_PER_SEC / (int)delta);
3676         }
3677
3678         /*
3679          * XXX event_limit might not quite work as expected on inherited
3680          * events
3681          */
3682
3683         event->pending_kill = POLL_IN;
3684         if (events && atomic_dec_and_test(&event->event_limit)) {
3685                 ret = 1;
3686                 event->pending_kill = POLL_HUP;
3687                 if (nmi) {
3688                         event->pending_disable = 1;
3689                         perf_pending_queue(&event->pending,
3690                                            perf_pending_event);
3691                 } else
3692                         perf_event_disable(event);
3693         }
3694
3695         if (event->overflow_handler)
3696                 event->overflow_handler(event, nmi, data, regs);
3697         else
3698                 perf_event_output(event, nmi, data, regs);
3699
3700         return ret;
3701 }
3702
3703 int perf_event_overflow(struct perf_event *event, int nmi,
3704                           struct perf_sample_data *data,
3705                           struct pt_regs *regs)
3706 {
3707         return __perf_event_overflow(event, nmi, 1, data, regs);
3708 }
3709
3710 /*
3711  * Generic software event infrastructure
3712  */
3713
3714 /*
3715  * We directly increment event->count and keep a second value in
3716  * event->hw.period_left to count intervals. This period event
3717  * is kept in the range [-sample_period, 0] so that we can use the
3718  * sign as trigger.
3719  */
3720
3721 static u64 perf_swevent_set_period(struct perf_event *event)
3722 {
3723         struct hw_perf_event *hwc = &event->hw;
3724         u64 period = hwc->last_period;
3725         u64 nr, offset;
3726         s64 old, val;
3727
3728         hwc->last_period = hwc->sample_period;
3729
3730 again:
3731         old = val = atomic64_read(&hwc->period_left);
3732         if (val < 0)
3733                 return 0;
3734
3735         nr = div64_u64(period + val, period);
3736         offset = nr * period;
3737         val -= offset;
3738         if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3739                 goto again;
3740
3741         return nr;
3742 }
3743
3744 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3745                                     int nmi, struct perf_sample_data *data,
3746                                     struct pt_regs *regs)
3747 {
3748         struct hw_perf_event *hwc = &event->hw;
3749         int throttle = 0;
3750
3751         data->period = event->hw.last_period;
3752         if (!overflow)
3753                 overflow = perf_swevent_set_period(event);
3754
3755         if (hwc->interrupts == MAX_INTERRUPTS)
3756                 return;
3757
3758         for (; overflow; overflow--) {
3759                 if (__perf_event_overflow(event, nmi, throttle,
3760                                             data, regs)) {
3761                         /*
3762                          * We inhibit the overflow from happening when
3763                          * hwc->interrupts == MAX_INTERRUPTS.
3764                          */
3765                         break;
3766                 }
3767                 throttle = 1;
3768         }
3769 }
3770
3771 static void perf_swevent_unthrottle(struct perf_event *event)
3772 {
3773         /*
3774          * Nothing to do, we already reset hwc->interrupts.
3775          */
3776 }
3777
3778 static void perf_swevent_add(struct perf_event *event, u64 nr,
3779                                int nmi, struct perf_sample_data *data,
3780                                struct pt_regs *regs)
3781 {
3782         struct hw_perf_event *hwc = &event->hw;
3783
3784         atomic64_add(nr, &event->count);
3785
3786         if (!regs)
3787                 return;
3788
3789         if (!hwc->sample_period)
3790                 return;
3791
3792         if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3793                 return perf_swevent_overflow(event, 1, nmi, data, regs);
3794
3795         if (atomic64_add_negative(nr, &hwc->period_left))
3796                 return;
3797
3798         perf_swevent_overflow(event, 0, nmi, data, regs);
3799 }
3800
3801 static int perf_swevent_is_counting(struct perf_event *event)
3802 {
3803         /*
3804          * The event is active, we're good!
3805          */
3806         if (event->state == PERF_EVENT_STATE_ACTIVE)
3807                 return 1;
3808
3809         /*
3810          * The event is off/error, not counting.
3811          */
3812         if (event->state != PERF_EVENT_STATE_INACTIVE)
3813                 return 0;
3814
3815         /*
3816          * The event is inactive, if the context is active
3817          * we're part of a group that didn't make it on the 'pmu',
3818          * not counting.
3819          */
3820         if (event->ctx->is_active)
3821                 return 0;
3822
3823         /*
3824          * We're inactive and the context is too, this means the
3825          * task is scheduled out, we're counting events that happen
3826          * to us, like migration events.
3827          */
3828         return 1;
3829 }
3830
3831 static int perf_tp_event_match(struct perf_event *event,
3832                                 struct perf_sample_data *data);
3833
3834 static int perf_swevent_match(struct perf_event *event,
3835                                 enum perf_type_id type,
3836                                 u32 event_id,
3837                                 struct perf_sample_data *data,
3838                                 struct pt_regs *regs)
3839 {
3840         if (!perf_swevent_is_counting(event))
3841                 return 0;
3842
3843         if (event->attr.type != type)
3844                 return 0;
3845         if (event->attr.config != event_id)
3846                 return 0;
3847
3848         if (regs) {
3849                 if (event->attr.exclude_user && user_mode(regs))
3850                         return 0;
3851
3852                 if (event->attr.exclude_kernel && !user_mode(regs))
3853                         return 0;
3854         }
3855
3856         if (event->attr.type == PERF_TYPE_TRACEPOINT &&
3857             !perf_tp_event_match(event, data))
3858                 return 0;
3859
3860         return 1;
3861 }
3862
3863 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
3864                                      enum perf_type_id type,
3865                                      u32 event_id, u64 nr, int nmi,
3866                                      struct perf_sample_data *data,
3867                                      struct pt_regs *regs)
3868 {
3869         struct perf_event *event;
3870
3871         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3872                 if (perf_swevent_match(event, type, event_id, data, regs))
3873                         perf_swevent_add(event, nr, nmi, data, regs);
3874         }
3875 }
3876
3877 static int *perf_swevent_recursion_context(struct perf_cpu_context *cpuctx)
3878 {
3879         if (in_nmi())
3880                 return &cpuctx->recursion[3];
3881
3882         if (in_irq())
3883                 return &cpuctx->recursion[2];
3884
3885         if (in_softirq())
3886                 return &cpuctx->recursion[1];
3887
3888         return &cpuctx->recursion[0];
3889 }
3890
3891 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
3892                                     u64 nr, int nmi,
3893                                     struct perf_sample_data *data,
3894                                     struct pt_regs *regs)
3895 {
3896         struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3897         int *recursion = perf_swevent_recursion_context(cpuctx);
3898         struct perf_event_context *ctx;
3899
3900         if (*recursion)
3901                 goto out;
3902
3903         (*recursion)++;
3904         barrier();
3905
3906         rcu_read_lock();
3907         perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
3908                                  nr, nmi, data, regs);
3909         /*
3910          * doesn't really matter which of the child contexts the
3911          * events ends up in.
3912          */
3913         ctx = rcu_dereference(current->perf_event_ctxp);
3914         if (ctx)
3915                 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
3916         rcu_read_unlock();
3917
3918         barrier();
3919         (*recursion)--;
3920
3921 out:
3922         put_cpu_var(perf_cpu_context);
3923 }
3924
3925 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
3926                             struct pt_regs *regs, u64 addr)
3927 {
3928         struct perf_sample_data data = {
3929                 .addr = addr,
3930         };
3931
3932         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi,
3933                                 &data, regs);
3934 }
3935
3936 static void perf_swevent_read(struct perf_event *event)
3937 {
3938 }
3939
3940 static int perf_swevent_enable(struct perf_event *event)
3941 {
3942         struct hw_perf_event *hwc = &event->hw;
3943
3944         if (hwc->sample_period) {
3945                 hwc->last_period = hwc->sample_period;
3946                 perf_swevent_set_period(event);
3947         }
3948         return 0;
3949 }
3950
3951 static void perf_swevent_disable(struct perf_event *event)
3952 {
3953 }
3954
3955 static const struct pmu perf_ops_generic = {
3956         .enable         = perf_swevent_enable,
3957         .disable        = perf_swevent_disable,
3958         .read           = perf_swevent_read,
3959         .unthrottle     = perf_swevent_unthrottle,
3960 };
3961
3962 /*
3963  * hrtimer based swevent callback
3964  */
3965
3966 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
3967 {
3968         enum hrtimer_restart ret = HRTIMER_RESTART;
3969         struct perf_sample_data data;
3970         struct pt_regs *regs;
3971         struct perf_event *event;
3972         u64 period;
3973
3974         event   = container_of(hrtimer, struct perf_event, hw.hrtimer);
3975         event->pmu->read(event);
3976
3977         data.addr = 0;
3978         regs = get_irq_regs();
3979         /*
3980          * In case we exclude kernel IPs or are somehow not in interrupt
3981          * context, provide the next best thing, the user IP.
3982          */
3983         if ((event->attr.exclude_kernel || !regs) &&
3984                         !event->attr.exclude_user)
3985                 regs = task_pt_regs(current);
3986
3987         if (regs) {
3988                 if (!(event->attr.exclude_idle && current->pid == 0))
3989                         if (perf_event_overflow(event, 0, &data, regs))
3990                                 ret = HRTIMER_NORESTART;
3991         }
3992
3993         period = max_t(u64, 10000, event->hw.sample_period);
3994         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
3995
3996         return ret;
3997 }
3998
3999 static void perf_swevent_start_hrtimer(struct perf_event *event)
4000 {
4001         struct hw_perf_event *hwc = &event->hw;
4002
4003         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4004         hwc->hrtimer.function = perf_swevent_hrtimer;
4005         if (hwc->sample_period) {
4006                 u64 period;
4007
4008                 if (hwc->remaining) {
4009                         if (hwc->remaining < 0)
4010                                 period = 10000;
4011                         else
4012                                 period = hwc->remaining;
4013                         hwc->remaining = 0;
4014                 } else {
4015                         period = max_t(u64, 10000, hwc->sample_period);
4016                 }
4017                 __hrtimer_start_range_ns(&hwc->hrtimer,
4018                                 ns_to_ktime(period), 0,
4019                                 HRTIMER_MODE_REL, 0);
4020         }
4021 }
4022
4023 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4024 {
4025         struct hw_perf_event *hwc = &event->hw;
4026
4027         if (hwc->sample_period) {
4028                 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4029                 hwc->remaining = ktime_to_ns(remaining);
4030
4031                 hrtimer_cancel(&hwc->hrtimer);
4032         }
4033 }
4034
4035 /*
4036  * Software event: cpu wall time clock
4037  */
4038
4039 static void cpu_clock_perf_event_update(struct perf_event *event)
4040 {
4041         int cpu = raw_smp_processor_id();
4042         s64 prev;
4043         u64 now;
4044
4045         now = cpu_clock(cpu);
4046         prev = atomic64_read(&event->hw.prev_count);
4047         atomic64_set(&event->hw.prev_count, now);
4048         atomic64_add(now - prev, &event->count);
4049 }
4050
4051 static int cpu_clock_perf_event_enable(struct perf_event *event)
4052 {
4053         struct hw_perf_event *hwc = &event->hw;
4054         int cpu = raw_smp_processor_id();
4055
4056         atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4057         perf_swevent_start_hrtimer(event);
4058
4059         return 0;
4060 }
4061
4062 static void cpu_clock_perf_event_disable(struct perf_event *event)
4063 {
4064         perf_swevent_cancel_hrtimer(event);
4065         cpu_clock_perf_event_update(event);
4066 }
4067
4068 static void cpu_clock_perf_event_read(struct perf_event *event)
4069 {
4070         cpu_clock_perf_event_update(event);
4071 }
4072
4073 static const struct pmu perf_ops_cpu_clock = {
4074         .enable         = cpu_clock_perf_event_enable,
4075         .disable        = cpu_clock_perf_event_disable,
4076         .read           = cpu_clock_perf_event_read,
4077 };
4078
4079 /*
4080  * Software event: task time clock
4081  */
4082
4083 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4084 {
4085         u64 prev;
4086         s64 delta;
4087
4088         prev = atomic64_xchg(&event->hw.prev_count, now);
4089         delta = now - prev;
4090         atomic64_add(delta, &event->count);
4091 }
4092
4093 static int task_clock_perf_event_enable(struct perf_event *event)
4094 {
4095         struct hw_perf_event *hwc = &event->hw;
4096         u64 now;
4097
4098         now = event->ctx->time;
4099
4100         atomic64_set(&hwc->prev_count, now);
4101
4102         perf_swevent_start_hrtimer(event);
4103
4104         return 0;
4105 }
4106
4107 static void task_clock_perf_event_disable(struct perf_event *event)
4108 {
4109         perf_swevent_cancel_hrtimer(event);
4110         task_clock_perf_event_update(event, event->ctx->time);
4111
4112 }
4113
4114 static void task_clock_perf_event_read(struct perf_event *event)
4115 {
4116         u64 time;
4117
4118         if (!in_nmi()) {
4119                 update_context_time(event->ctx);
4120                 time = event->ctx->time;
4121         } else {
4122                 u64 now = perf_clock();
4123                 u64 delta = now - event->ctx->timestamp;
4124                 time = event->ctx->time + delta;
4125         }
4126
4127         task_clock_perf_event_update(event, time);
4128 }
4129
4130 static const struct pmu perf_ops_task_clock = {
4131         .enable         = task_clock_perf_event_enable,
4132         .disable        = task_clock_perf_event_disable,
4133         .read           = task_clock_perf_event_read,
4134 };
4135
4136 #ifdef CONFIG_EVENT_PROFILE
4137
4138 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4139                           int entry_size)
4140 {
4141         struct perf_raw_record raw = {
4142                 .size = entry_size,
4143                 .data = record,
4144         };
4145
4146         struct perf_sample_data data = {
4147                 .addr = addr,
4148                 .raw = &raw,
4149         };
4150
4151         struct pt_regs *regs = get_irq_regs();
4152
4153         if (!regs)
4154                 regs = task_pt_regs(current);
4155
4156         do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4157                                 &data, regs);
4158 }
4159 EXPORT_SYMBOL_GPL(perf_tp_event);
4160
4161 static int perf_tp_event_match(struct perf_event *event,
4162                                 struct perf_sample_data *data)
4163 {
4164         void *record = data->raw->data;
4165
4166         if (likely(!event->filter) || filter_match_preds(event->filter, record))
4167                 return 1;
4168         return 0;
4169 }
4170
4171 static void tp_perf_event_destroy(struct perf_event *event)
4172 {
4173         ftrace_profile_disable(event->attr.config);
4174 }
4175
4176 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4177 {
4178         /*
4179          * Raw tracepoint data is a severe data leak, only allow root to
4180          * have these.
4181          */
4182         if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4183                         perf_paranoid_tracepoint_raw() &&
4184                         !capable(CAP_SYS_ADMIN))
4185                 return ERR_PTR(-EPERM);
4186
4187         if (ftrace_profile_enable(event->attr.config))
4188                 return NULL;
4189
4190         event->destroy = tp_perf_event_destroy;
4191
4192         return &perf_ops_generic;
4193 }
4194
4195 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4196 {
4197         char *filter_str;
4198         int ret;
4199
4200         if (event->attr.type != PERF_TYPE_TRACEPOINT)
4201                 return -EINVAL;
4202
4203         filter_str = strndup_user(arg, PAGE_SIZE);
4204         if (IS_ERR(filter_str))
4205                 return PTR_ERR(filter_str);
4206
4207         ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4208
4209         kfree(filter_str);
4210         return ret;
4211 }
4212
4213 static void perf_event_free_filter(struct perf_event *event)
4214 {
4215         ftrace_profile_free_filter(event);
4216 }
4217
4218 #else
4219
4220 static int perf_tp_event_match(struct perf_event *event,
4221                                 struct perf_sample_data *data)
4222 {
4223         return 1;
4224 }
4225
4226 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4227 {
4228         return NULL;
4229 }
4230
4231 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4232 {
4233         return -ENOENT;
4234 }
4235
4236 static void perf_event_free_filter(struct perf_event *event)
4237 {
4238 }
4239
4240 #endif /* CONFIG_EVENT_PROFILE */
4241
4242 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4243 static void bp_perf_event_destroy(struct perf_event *event)
4244 {
4245         release_bp_slot(event);
4246 }
4247
4248 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4249 {
4250         int err;
4251         /*
4252          * The breakpoint is already filled if we haven't created the counter
4253          * through perf syscall
4254          * FIXME: manage to get trigerred to NULL if it comes from syscalls
4255          */
4256         if (!bp->callback)
4257                 err = register_perf_hw_breakpoint(bp);
4258         else
4259                 err = __register_perf_hw_breakpoint(bp);
4260         if (err)
4261                 return ERR_PTR(err);
4262
4263         bp->destroy = bp_perf_event_destroy;
4264
4265         return &perf_ops_bp;
4266 }
4267
4268 void perf_bp_event(struct perf_event *bp, void *regs)
4269 {
4270         /* TODO */
4271 }
4272 #else
4273 static void bp_perf_event_destroy(struct perf_event *event)
4274 {
4275 }
4276
4277 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4278 {
4279         return NULL;
4280 }
4281
4282 void perf_bp_event(struct perf_event *bp, void *regs)
4283 {
4284 }
4285 #endif
4286
4287 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4288
4289 static void sw_perf_event_destroy(struct perf_event *event)
4290 {
4291         u64 event_id = event->attr.config;
4292
4293         WARN_ON(event->parent);
4294
4295         atomic_dec(&perf_swevent_enabled[event_id]);
4296 }
4297
4298 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4299 {
4300         const struct pmu *pmu = NULL;
4301         u64 event_id = event->attr.config;
4302
4303         /*
4304          * Software events (currently) can't in general distinguish
4305          * between user, kernel and hypervisor events.
4306          * However, context switches and cpu migrations are considered
4307          * to be kernel events, and page faults are never hypervisor
4308          * events.
4309          */
4310         switch (event_id) {
4311         case PERF_COUNT_SW_CPU_CLOCK:
4312                 pmu = &perf_ops_cpu_clock;
4313
4314                 break;
4315         case PERF_COUNT_SW_TASK_CLOCK:
4316                 /*
4317                  * If the user instantiates this as a per-cpu event,
4318                  * use the cpu_clock event instead.
4319                  */
4320                 if (event->ctx->task)
4321                         pmu = &perf_ops_task_clock;
4322                 else
4323                         pmu = &perf_ops_cpu_clock;
4324
4325                 break;
4326         case PERF_COUNT_SW_PAGE_FAULTS:
4327         case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4328         case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4329         case PERF_COUNT_SW_CONTEXT_SWITCHES:
4330         case PERF_COUNT_SW_CPU_MIGRATIONS:
4331         case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4332         case PERF_COUNT_SW_EMULATION_FAULTS:
4333                 if (!event->parent) {
4334                         atomic_inc(&perf_swevent_enabled[event_id]);
4335                         event->destroy = sw_perf_event_destroy;
4336                 }
4337                 pmu = &perf_ops_generic;
4338                 break;
4339         }
4340
4341         return pmu;
4342 }
4343
4344 /*
4345  * Allocate and initialize a event structure
4346  */
4347 static struct perf_event *
4348 perf_event_alloc(struct perf_event_attr *attr,
4349                    int cpu,
4350                    struct perf_event_context *ctx,
4351                    struct perf_event *group_leader,
4352                    struct perf_event *parent_event,
4353                    perf_callback_t callback,
4354                    gfp_t gfpflags)
4355 {
4356         const struct pmu *pmu;
4357         struct perf_event *event;
4358         struct hw_perf_event *hwc;
4359         long err;
4360
4361         event = kzalloc(sizeof(*event), gfpflags);
4362         if (!event)
4363                 return ERR_PTR(-ENOMEM);
4364
4365         /*
4366          * Single events are their own group leaders, with an
4367          * empty sibling list:
4368          */
4369         if (!group_leader)
4370                 group_leader = event;
4371
4372         mutex_init(&event->child_mutex);
4373         INIT_LIST_HEAD(&event->child_list);
4374
4375         INIT_LIST_HEAD(&event->group_entry);
4376         INIT_LIST_HEAD(&event->event_entry);
4377         INIT_LIST_HEAD(&event->sibling_list);
4378         init_waitqueue_head(&event->waitq);
4379
4380         mutex_init(&event->mmap_mutex);
4381
4382         event->cpu              = cpu;
4383         event->attr             = *attr;
4384         event->group_leader     = group_leader;
4385         event->pmu              = NULL;
4386         event->ctx              = ctx;
4387         event->oncpu            = -1;
4388
4389         event->parent           = parent_event;
4390
4391         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
4392         event->id               = atomic64_inc_return(&perf_event_id);
4393
4394         event->state            = PERF_EVENT_STATE_INACTIVE;
4395
4396         if (!callback && parent_event)
4397                 callback = parent_event->callback;
4398         
4399         event->callback = callback;
4400
4401         if (attr->disabled)
4402                 event->state = PERF_EVENT_STATE_OFF;
4403
4404         pmu = NULL;
4405
4406         hwc = &event->hw;
4407         hwc->sample_period = attr->sample_period;
4408         if (attr->freq && attr->sample_freq)
4409                 hwc->sample_period = 1;
4410         hwc->last_period = hwc->sample_period;
4411
4412         atomic64_set(&hwc->period_left, hwc->sample_period);
4413
4414         /*
4415          * we currently do not support PERF_FORMAT_GROUP on inherited events
4416          */
4417         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4418                 goto done;
4419
4420         switch (attr->type) {
4421         case PERF_TYPE_RAW:
4422         case PERF_TYPE_HARDWARE:
4423         case PERF_TYPE_HW_CACHE:
4424                 pmu = hw_perf_event_init(event);
4425                 break;
4426
4427         case PERF_TYPE_SOFTWARE:
4428                 pmu = sw_perf_event_init(event);
4429                 break;
4430
4431         case PERF_TYPE_TRACEPOINT:
4432                 pmu = tp_perf_event_init(event);
4433                 break;
4434
4435         case PERF_TYPE_BREAKPOINT:
4436                 pmu = bp_perf_event_init(event);
4437                 break;
4438
4439
4440         default:
4441                 break;
4442         }
4443 done:
4444         err = 0;
4445         if (!pmu)
4446                 err = -EINVAL;
4447         else if (IS_ERR(pmu))
4448                 err = PTR_ERR(pmu);
4449
4450         if (err) {
4451                 if (event->ns)
4452                         put_pid_ns(event->ns);
4453                 kfree(event);
4454                 return ERR_PTR(err);
4455         }
4456
4457         event->pmu = pmu;
4458
4459         if (!event->parent) {
4460                 atomic_inc(&nr_events);
4461                 if (event->attr.mmap)
4462                         atomic_inc(&nr_mmap_events);
4463                 if (event->attr.comm)
4464                         atomic_inc(&nr_comm_events);
4465                 if (event->attr.task)
4466                         atomic_inc(&nr_task_events);
4467         }
4468
4469         return event;
4470 }
4471
4472 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4473                           struct perf_event_attr *attr)
4474 {
4475         u32 size;
4476         int ret;
4477
4478         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4479                 return -EFAULT;
4480
4481         /*
4482          * zero the full structure, so that a short copy will be nice.
4483          */
4484         memset(attr, 0, sizeof(*attr));
4485
4486         ret = get_user(size, &uattr->size);
4487         if (ret)
4488                 return ret;
4489
4490         if (size > PAGE_SIZE)   /* silly large */
4491                 goto err_size;
4492
4493         if (!size)              /* abi compat */
4494                 size = PERF_ATTR_SIZE_VER0;
4495
4496         if (size < PERF_ATTR_SIZE_VER0)
4497                 goto err_size;
4498
4499         /*
4500          * If we're handed a bigger struct than we know of,
4501          * ensure all the unknown bits are 0 - i.e. new
4502          * user-space does not rely on any kernel feature
4503          * extensions we dont know about yet.
4504          */
4505         if (size > sizeof(*attr)) {
4506                 unsigned char __user *addr;
4507                 unsigned char __user *end;
4508                 unsigned char val;
4509
4510                 addr = (void __user *)uattr + sizeof(*attr);
4511                 end  = (void __user *)uattr + size;
4512
4513                 for (; addr < end; addr++) {
4514                         ret = get_user(val, addr);
4515                         if (ret)
4516                                 return ret;
4517                         if (val)
4518                                 goto err_size;
4519                 }
4520                 size = sizeof(*attr);
4521         }
4522
4523         ret = copy_from_user(attr, uattr, size);
4524         if (ret)
4525                 return -EFAULT;
4526
4527         /*
4528          * If the type exists, the corresponding creation will verify
4529          * the attr->config.
4530          */
4531         if (attr->type >= PERF_TYPE_MAX)
4532                 return -EINVAL;
4533
4534         if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3)
4535                 return -EINVAL;
4536
4537         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4538                 return -EINVAL;
4539
4540         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4541                 return -EINVAL;
4542
4543 out:
4544         return ret;
4545
4546 err_size:
4547         put_user(sizeof(*attr), &uattr->size);
4548         ret = -E2BIG;
4549         goto out;
4550 }
4551
4552 static int perf_event_set_output(struct perf_event *event, int output_fd)
4553 {
4554         struct perf_event *output_event = NULL;
4555         struct file *output_file = NULL;
4556         struct perf_event *old_output;
4557         int fput_needed = 0;
4558         int ret = -EINVAL;
4559
4560         if (!output_fd)
4561                 goto set;
4562
4563         output_file = fget_light(output_fd, &fput_needed);
4564         if (!output_file)
4565                 return -EBADF;
4566
4567         if (output_file->f_op != &perf_fops)
4568                 goto out;
4569
4570         output_event = output_file->private_data;
4571
4572         /* Don't chain output fds */
4573         if (output_event->output)
4574                 goto out;
4575
4576         /* Don't set an output fd when we already have an output channel */
4577         if (event->data)
4578                 goto out;
4579
4580         atomic_long_inc(&output_file->f_count);
4581
4582 set:
4583         mutex_lock(&event->mmap_mutex);
4584         old_output = event->output;
4585         rcu_assign_pointer(event->output, output_event);
4586         mutex_unlock(&event->mmap_mutex);
4587
4588         if (old_output) {
4589                 /*
4590                  * we need to make sure no existing perf_output_*()
4591                  * is still referencing this event.
4592                  */
4593                 synchronize_rcu();
4594                 fput(old_output->filp);
4595         }
4596
4597         ret = 0;
4598 out:
4599         fput_light(output_file, fput_needed);
4600         return ret;
4601 }
4602
4603 /**
4604  * sys_perf_event_open - open a performance event, associate it to a task/cpu
4605  *
4606  * @attr_uptr:  event_id type attributes for monitoring/sampling
4607  * @pid:                target pid
4608  * @cpu:                target cpu
4609  * @group_fd:           group leader event fd
4610  */
4611 SYSCALL_DEFINE5(perf_event_open,
4612                 struct perf_event_attr __user *, attr_uptr,
4613                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4614 {
4615         struct perf_event *event, *group_leader;
4616         struct perf_event_attr attr;
4617         struct perf_event_context *ctx;
4618         struct file *event_file = NULL;
4619         struct file *group_file = NULL;
4620         int fput_needed = 0;
4621         int fput_needed2 = 0;
4622         int err;
4623
4624         /* for future expandability... */
4625         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4626                 return -EINVAL;
4627
4628         err = perf_copy_attr(attr_uptr, &attr);
4629         if (err)
4630                 return err;
4631
4632         if (!attr.exclude_kernel) {
4633                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4634                         return -EACCES;
4635         }
4636
4637         if (attr.freq) {
4638                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4639                         return -EINVAL;
4640         }
4641
4642         /*
4643          * Get the target context (task or percpu):
4644          */
4645         ctx = find_get_context(pid, cpu);
4646         if (IS_ERR(ctx))
4647                 return PTR_ERR(ctx);
4648
4649         /*
4650          * Look up the group leader (we will attach this event to it):
4651          */
4652         group_leader = NULL;
4653         if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4654                 err = -EINVAL;
4655                 group_file = fget_light(group_fd, &fput_needed);
4656                 if (!group_file)
4657                         goto err_put_context;
4658                 if (group_file->f_op != &perf_fops)
4659                         goto err_put_context;
4660
4661                 group_leader = group_file->private_data;
4662                 /*
4663                  * Do not allow a recursive hierarchy (this new sibling
4664                  * becoming part of another group-sibling):
4665                  */
4666                 if (group_leader->group_leader != group_leader)
4667                         goto err_put_context;
4668                 /*
4669                  * Do not allow to attach to a group in a different
4670                  * task or CPU context:
4671                  */
4672                 if (group_leader->ctx != ctx)
4673                         goto err_put_context;
4674                 /*
4675                  * Only a group leader can be exclusive or pinned
4676                  */
4677                 if (attr.exclusive || attr.pinned)
4678                         goto err_put_context;
4679         }
4680
4681         event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4682                                      NULL, NULL, GFP_KERNEL);
4683         err = PTR_ERR(event);
4684         if (IS_ERR(event))
4685                 goto err_put_context;
4686
4687         err = anon_inode_getfd("[perf_event]", &perf_fops, event, 0);
4688         if (err < 0)
4689                 goto err_free_put_context;
4690
4691         event_file = fget_light(err, &fput_needed2);
4692         if (!event_file)
4693                 goto err_free_put_context;
4694
4695         if (flags & PERF_FLAG_FD_OUTPUT) {
4696                 err = perf_event_set_output(event, group_fd);
4697                 if (err)
4698                         goto err_fput_free_put_context;
4699         }
4700
4701         event->filp = event_file;
4702         WARN_ON_ONCE(ctx->parent_ctx);
4703         mutex_lock(&ctx->mutex);
4704         perf_install_in_context(ctx, event, cpu);
4705         ++ctx->generation;
4706         mutex_unlock(&ctx->mutex);
4707
4708         event->owner = current;
4709         get_task_struct(current);
4710         mutex_lock(&current->perf_event_mutex);
4711         list_add_tail(&event->owner_entry, &current->perf_event_list);
4712         mutex_unlock(&current->perf_event_mutex);
4713
4714 err_fput_free_put_context:
4715         fput_light(event_file, fput_needed2);
4716
4717 err_free_put_context:
4718         if (err < 0)
4719                 kfree(event);
4720
4721 err_put_context:
4722         if (err < 0)
4723                 put_ctx(ctx);
4724
4725         fput_light(group_file, fput_needed);
4726
4727         return err;
4728 }
4729
4730 /**
4731  * perf_event_create_kernel_counter
4732  *
4733  * @attr: attributes of the counter to create
4734  * @cpu: cpu in which the counter is bound
4735  * @pid: task to profile
4736  */
4737 struct perf_event *
4738 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4739                                  pid_t pid, perf_callback_t callback)
4740 {
4741         struct perf_event *event;
4742         struct perf_event_context *ctx;
4743         int err;
4744
4745         /*
4746          * Get the target context (task or percpu):
4747          */
4748
4749         ctx = find_get_context(pid, cpu);
4750         if (IS_ERR(ctx))
4751                 return NULL;
4752
4753         event = perf_event_alloc(attr, cpu, ctx, NULL,
4754                                      NULL, callback, GFP_KERNEL);
4755         err = PTR_ERR(event);
4756         if (IS_ERR(event))
4757                 goto err_put_context;
4758
4759         event->filp = NULL;
4760         WARN_ON_ONCE(ctx->parent_ctx);
4761         mutex_lock(&ctx->mutex);
4762         perf_install_in_context(ctx, event, cpu);
4763         ++ctx->generation;
4764         mutex_unlock(&ctx->mutex);
4765
4766         event->owner = current;
4767         get_task_struct(current);
4768         mutex_lock(&current->perf_event_mutex);
4769         list_add_tail(&event->owner_entry, &current->perf_event_list);
4770         mutex_unlock(&current->perf_event_mutex);
4771
4772         return event;
4773
4774 err_put_context:
4775         if (err < 0)
4776                 put_ctx(ctx);
4777
4778         return NULL;
4779 }
4780 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4781
4782 /*
4783  * inherit a event from parent task to child task:
4784  */
4785 static struct perf_event *
4786 inherit_event(struct perf_event *parent_event,
4787               struct task_struct *parent,
4788               struct perf_event_context *parent_ctx,
4789               struct task_struct *child,
4790               struct perf_event *group_leader,
4791               struct perf_event_context *child_ctx)
4792 {
4793         struct perf_event *child_event;
4794
4795         /*
4796          * Instead of creating recursive hierarchies of events,
4797          * we link inherited events back to the original parent,
4798          * which has a filp for sure, which we use as the reference
4799          * count:
4800          */
4801         if (parent_event->parent)
4802                 parent_event = parent_event->parent;
4803
4804         child_event = perf_event_alloc(&parent_event->attr,
4805                                            parent_event->cpu, child_ctx,
4806                                            group_leader, parent_event,
4807                                            NULL, GFP_KERNEL);
4808         if (IS_ERR(child_event))
4809                 return child_event;
4810         get_ctx(child_ctx);
4811
4812         /*
4813          * Make the child state follow the state of the parent event,
4814          * not its attr.disabled bit.  We hold the parent's mutex,
4815          * so we won't race with perf_event_{en, dis}able_family.
4816          */
4817         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
4818                 child_event->state = PERF_EVENT_STATE_INACTIVE;
4819         else
4820                 child_event->state = PERF_EVENT_STATE_OFF;
4821
4822         if (parent_event->attr.freq)
4823                 child_event->hw.sample_period = parent_event->hw.sample_period;
4824
4825         child_event->overflow_handler = parent_event->overflow_handler;
4826
4827         /*
4828          * Link it up in the child's context:
4829          */
4830         add_event_to_ctx(child_event, child_ctx);
4831
4832         /*
4833          * Get a reference to the parent filp - we will fput it
4834          * when the child event exits. This is safe to do because
4835          * we are in the parent and we know that the filp still
4836          * exists and has a nonzero count:
4837          */
4838         atomic_long_inc(&parent_event->filp->f_count);
4839
4840         /*
4841          * Link this into the parent event's child list
4842          */
4843         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4844         mutex_lock(&parent_event->child_mutex);
4845         list_add_tail(&child_event->child_list, &parent_event->child_list);
4846         mutex_unlock(&parent_event->child_mutex);
4847
4848         return child_event;
4849 }
4850
4851 static int inherit_group(struct perf_event *parent_event,
4852               struct task_struct *parent,
4853               struct perf_event_context *parent_ctx,
4854               struct task_struct *child,
4855               struct perf_event_context *child_ctx)
4856 {
4857         struct perf_event *leader;
4858         struct perf_event *sub;
4859         struct perf_event *child_ctr;
4860
4861         leader = inherit_event(parent_event, parent, parent_ctx,
4862                                  child, NULL, child_ctx);
4863         if (IS_ERR(leader))
4864                 return PTR_ERR(leader);
4865         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
4866                 child_ctr = inherit_event(sub, parent, parent_ctx,
4867                                             child, leader, child_ctx);
4868                 if (IS_ERR(child_ctr))
4869                         return PTR_ERR(child_ctr);
4870         }
4871         return 0;
4872 }
4873
4874 static void sync_child_event(struct perf_event *child_event,
4875                                struct task_struct *child)
4876 {
4877         struct perf_event *parent_event = child_event->parent;
4878         u64 child_val;
4879
4880         if (child_event->attr.inherit_stat)
4881                 perf_event_read_event(child_event, child);
4882
4883         child_val = atomic64_read(&child_event->count);
4884
4885         /*
4886          * Add back the child's count to the parent's count:
4887          */
4888         atomic64_add(child_val, &parent_event->count);
4889         atomic64_add(child_event->total_time_enabled,
4890                      &parent_event->child_total_time_enabled);
4891         atomic64_add(child_event->total_time_running,
4892                      &parent_event->child_total_time_running);
4893
4894         /*
4895          * Remove this event from the parent's list
4896          */
4897         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4898         mutex_lock(&parent_event->child_mutex);
4899         list_del_init(&child_event->child_list);
4900         mutex_unlock(&parent_event->child_mutex);
4901
4902         /*
4903          * Release the parent event, if this was the last
4904          * reference to it.
4905          */
4906         fput(parent_event->filp);
4907 }
4908
4909 static void
4910 __perf_event_exit_task(struct perf_event *child_event,
4911                          struct perf_event_context *child_ctx,
4912                          struct task_struct *child)
4913 {
4914         struct perf_event *parent_event;
4915
4916         update_event_times(child_event);
4917         perf_event_remove_from_context(child_event);
4918
4919         parent_event = child_event->parent;
4920         /*
4921          * It can happen that parent exits first, and has events
4922          * that are still around due to the child reference. These
4923          * events need to be zapped - but otherwise linger.
4924          */
4925         if (parent_event) {
4926                 sync_child_event(child_event, child);
4927                 free_event(child_event);
4928         }
4929 }
4930
4931 /*
4932  * When a child task exits, feed back event values to parent events.
4933  */
4934 void perf_event_exit_task(struct task_struct *child)
4935 {
4936         struct perf_event *child_event, *tmp;
4937         struct perf_event_context *child_ctx;
4938         unsigned long flags;
4939
4940         if (likely(!child->perf_event_ctxp)) {
4941                 perf_event_task(child, NULL, 0);
4942                 return;
4943         }
4944
4945         local_irq_save(flags);
4946         /*
4947          * We can't reschedule here because interrupts are disabled,
4948          * and either child is current or it is a task that can't be
4949          * scheduled, so we are now safe from rescheduling changing
4950          * our context.
4951          */
4952         child_ctx = child->perf_event_ctxp;
4953         __perf_event_task_sched_out(child_ctx);
4954
4955         /*
4956          * Take the context lock here so that if find_get_context is
4957          * reading child->perf_event_ctxp, we wait until it has
4958          * incremented the context's refcount before we do put_ctx below.
4959          */
4960         spin_lock(&child_ctx->lock);
4961         child->perf_event_ctxp = NULL;
4962         /*
4963          * If this context is a clone; unclone it so it can't get
4964          * swapped to another process while we're removing all
4965          * the events from it.
4966          */
4967         unclone_ctx(child_ctx);
4968         spin_unlock_irqrestore(&child_ctx->lock, flags);
4969
4970         /*
4971          * Report the task dead after unscheduling the events so that we
4972          * won't get any samples after PERF_RECORD_EXIT. We can however still
4973          * get a few PERF_RECORD_READ events.
4974          */
4975         perf_event_task(child, child_ctx, 0);
4976
4977         /*
4978          * We can recurse on the same lock type through:
4979          *
4980          *   __perf_event_exit_task()
4981          *     sync_child_event()
4982          *       fput(parent_event->filp)
4983          *         perf_release()
4984          *           mutex_lock(&ctx->mutex)
4985          *
4986          * But since its the parent context it won't be the same instance.
4987          */
4988         mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
4989
4990 again:
4991         list_for_each_entry_safe(child_event, tmp, &child_ctx->group_list,
4992                                  group_entry)
4993                 __perf_event_exit_task(child_event, child_ctx, child);
4994
4995         /*
4996          * If the last event was a group event, it will have appended all
4997          * its siblings to the list, but we obtained 'tmp' before that which
4998          * will still point to the list head terminating the iteration.
4999          */
5000         if (!list_empty(&child_ctx->group_list))
5001                 goto again;
5002
5003         mutex_unlock(&child_ctx->mutex);
5004
5005         put_ctx(child_ctx);
5006 }
5007
5008 /*
5009  * free an unexposed, unused context as created by inheritance by
5010  * init_task below, used by fork() in case of fail.
5011  */
5012 void perf_event_free_task(struct task_struct *task)
5013 {
5014         struct perf_event_context *ctx = task->perf_event_ctxp;
5015         struct perf_event *event, *tmp;
5016
5017         if (!ctx)
5018                 return;
5019
5020         mutex_lock(&ctx->mutex);
5021 again:
5022         list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry) {
5023                 struct perf_event *parent = event->parent;
5024
5025                 if (WARN_ON_ONCE(!parent))
5026                         continue;
5027
5028                 mutex_lock(&parent->child_mutex);
5029                 list_del_init(&event->child_list);
5030                 mutex_unlock(&parent->child_mutex);
5031
5032                 fput(parent->filp);
5033
5034                 list_del_event(event, ctx);
5035                 free_event(event);
5036         }
5037
5038         if (!list_empty(&ctx->group_list))
5039                 goto again;
5040
5041         mutex_unlock(&ctx->mutex);
5042
5043         put_ctx(ctx);
5044 }
5045
5046 /*
5047  * Initialize the perf_event context in task_struct
5048  */
5049 int perf_event_init_task(struct task_struct *child)
5050 {
5051         struct perf_event_context *child_ctx, *parent_ctx;
5052         struct perf_event_context *cloned_ctx;
5053         struct perf_event *event;
5054         struct task_struct *parent = current;
5055         int inherited_all = 1;
5056         int ret = 0;
5057
5058         child->perf_event_ctxp = NULL;
5059
5060         mutex_init(&child->perf_event_mutex);
5061         INIT_LIST_HEAD(&child->perf_event_list);
5062
5063         if (likely(!parent->perf_event_ctxp))
5064                 return 0;
5065
5066         /*
5067          * This is executed from the parent task context, so inherit
5068          * events that have been marked for cloning.
5069          * First allocate and initialize a context for the child.
5070          */
5071
5072         child_ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
5073         if (!child_ctx)
5074                 return -ENOMEM;
5075
5076         __perf_event_init_context(child_ctx, child);
5077         child->perf_event_ctxp = child_ctx;
5078         get_task_struct(child);
5079
5080         /*
5081          * If the parent's context is a clone, pin it so it won't get
5082          * swapped under us.
5083          */
5084         parent_ctx = perf_pin_task_context(parent);
5085
5086         /*
5087          * No need to check if parent_ctx != NULL here; since we saw
5088          * it non-NULL earlier, the only reason for it to become NULL
5089          * is if we exit, and since we're currently in the middle of
5090          * a fork we can't be exiting at the same time.
5091          */
5092
5093         /*
5094          * Lock the parent list. No need to lock the child - not PID
5095          * hashed yet and not running, so nobody can access it.
5096          */
5097         mutex_lock(&parent_ctx->mutex);
5098
5099         /*
5100          * We dont have to disable NMIs - we are only looking at
5101          * the list, not manipulating it:
5102          */
5103         list_for_each_entry(event, &parent_ctx->group_list, group_entry) {
5104
5105                 if (!event->attr.inherit) {
5106                         inherited_all = 0;
5107                         continue;
5108                 }
5109
5110                 ret = inherit_group(event, parent, parent_ctx,
5111                                              child, child_ctx);
5112                 if (ret) {
5113                         inherited_all = 0;
5114                         break;
5115                 }
5116         }
5117
5118         if (inherited_all) {
5119                 /*
5120                  * Mark the child context as a clone of the parent
5121                  * context, or of whatever the parent is a clone of.
5122                  * Note that if the parent is a clone, it could get
5123                  * uncloned at any point, but that doesn't matter
5124                  * because the list of events and the generation
5125                  * count can't have changed since we took the mutex.
5126                  */
5127                 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5128                 if (cloned_ctx) {
5129                         child_ctx->parent_ctx = cloned_ctx;
5130                         child_ctx->parent_gen = parent_ctx->parent_gen;
5131                 } else {
5132                         child_ctx->parent_ctx = parent_ctx;
5133                         child_ctx->parent_gen = parent_ctx->generation;
5134                 }
5135                 get_ctx(child_ctx->parent_ctx);
5136         }
5137
5138         mutex_unlock(&parent_ctx->mutex);
5139
5140         perf_unpin_context(parent_ctx);
5141
5142         return ret;
5143 }
5144
5145 static void __cpuinit perf_event_init_cpu(int cpu)
5146 {
5147         struct perf_cpu_context *cpuctx;
5148
5149         cpuctx = &per_cpu(perf_cpu_context, cpu);
5150         __perf_event_init_context(&cpuctx->ctx, NULL);
5151
5152         spin_lock(&perf_resource_lock);
5153         cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5154         spin_unlock(&perf_resource_lock);
5155
5156         hw_perf_event_setup(cpu);
5157 }
5158
5159 #ifdef CONFIG_HOTPLUG_CPU
5160 static void __perf_event_exit_cpu(void *info)
5161 {
5162         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5163         struct perf_event_context *ctx = &cpuctx->ctx;
5164         struct perf_event *event, *tmp;
5165
5166         list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry)
5167                 __perf_event_remove_from_context(event);
5168 }
5169 static void perf_event_exit_cpu(int cpu)
5170 {
5171         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5172         struct perf_event_context *ctx = &cpuctx->ctx;
5173
5174         mutex_lock(&ctx->mutex);
5175         smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5176         mutex_unlock(&ctx->mutex);
5177 }
5178 #else
5179 static inline void perf_event_exit_cpu(int cpu) { }
5180 #endif
5181
5182 static int __cpuinit
5183 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5184 {
5185         unsigned int cpu = (long)hcpu;
5186
5187         switch (action) {
5188
5189         case CPU_UP_PREPARE:
5190         case CPU_UP_PREPARE_FROZEN:
5191                 perf_event_init_cpu(cpu);
5192                 break;
5193
5194         case CPU_ONLINE:
5195         case CPU_ONLINE_FROZEN:
5196                 hw_perf_event_setup_online(cpu);
5197                 break;
5198
5199         case CPU_DOWN_PREPARE:
5200         case CPU_DOWN_PREPARE_FROZEN:
5201                 perf_event_exit_cpu(cpu);
5202                 break;
5203
5204         default:
5205                 break;
5206         }
5207
5208         return NOTIFY_OK;
5209 }
5210
5211 /*
5212  * This has to have a higher priority than migration_notifier in sched.c.
5213  */
5214 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5215         .notifier_call          = perf_cpu_notify,
5216         .priority               = 20,
5217 };
5218
5219 void __init perf_event_init(void)
5220 {
5221         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5222                         (void *)(long)smp_processor_id());
5223         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5224                         (void *)(long)smp_processor_id());
5225         register_cpu_notifier(&perf_cpu_nb);
5226 }
5227
5228 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
5229 {
5230         return sprintf(buf, "%d\n", perf_reserved_percpu);
5231 }
5232
5233 static ssize_t
5234 perf_set_reserve_percpu(struct sysdev_class *class,
5235                         const char *buf,
5236                         size_t count)
5237 {
5238         struct perf_cpu_context *cpuctx;
5239         unsigned long val;
5240         int err, cpu, mpt;
5241
5242         err = strict_strtoul(buf, 10, &val);
5243         if (err)
5244                 return err;
5245         if (val > perf_max_events)
5246                 return -EINVAL;
5247
5248         spin_lock(&perf_resource_lock);
5249         perf_reserved_percpu = val;
5250         for_each_online_cpu(cpu) {
5251                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5252                 spin_lock_irq(&cpuctx->ctx.lock);
5253                 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5254                           perf_max_events - perf_reserved_percpu);
5255                 cpuctx->max_pertask = mpt;
5256                 spin_unlock_irq(&cpuctx->ctx.lock);
5257         }
5258         spin_unlock(&perf_resource_lock);
5259
5260         return count;
5261 }
5262
5263 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
5264 {
5265         return sprintf(buf, "%d\n", perf_overcommit);
5266 }
5267
5268 static ssize_t
5269 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
5270 {
5271         unsigned long val;
5272         int err;
5273
5274         err = strict_strtoul(buf, 10, &val);
5275         if (err)
5276                 return err;
5277         if (val > 1)
5278                 return -EINVAL;
5279
5280         spin_lock(&perf_resource_lock);
5281         perf_overcommit = val;
5282         spin_unlock(&perf_resource_lock);
5283
5284         return count;
5285 }
5286
5287 static SYSDEV_CLASS_ATTR(
5288                                 reserve_percpu,
5289                                 0644,
5290                                 perf_show_reserve_percpu,
5291                                 perf_set_reserve_percpu
5292                         );
5293
5294 static SYSDEV_CLASS_ATTR(
5295                                 overcommit,
5296                                 0644,
5297                                 perf_show_overcommit,
5298                                 perf_set_overcommit
5299                         );
5300
5301 static struct attribute *perfclass_attrs[] = {
5302         &attr_reserve_percpu.attr,
5303         &attr_overcommit.attr,
5304         NULL
5305 };
5306
5307 static struct attribute_group perfclass_attr_group = {
5308         .attrs                  = perfclass_attrs,
5309         .name                   = "perf_events",
5310 };
5311
5312 static int __init perf_event_sysfs_init(void)
5313 {
5314         return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5315                                   &perfclass_attr_group);
5316 }
5317 device_initcall(perf_event_sysfs_init);