perf_events: Fix default watermark calculation
[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         mutex_lock(&event->child_mutex);
1786         total += perf_event_read(event);
1787         *enabled += event->total_time_enabled +
1788                         atomic64_read(&event->child_total_time_enabled);
1789         *running += event->total_time_running +
1790                         atomic64_read(&event->child_total_time_running);
1791
1792         list_for_each_entry(child, &event->child_list, child_list) {
1793                 total += perf_event_read(child);
1794                 *enabled += child->total_time_enabled;
1795                 *running += child->total_time_running;
1796         }
1797         mutex_unlock(&event->child_mutex);
1798
1799         return total;
1800 }
1801 EXPORT_SYMBOL_GPL(perf_event_read_value);
1802
1803 static int perf_event_read_group(struct perf_event *event,
1804                                    u64 read_format, char __user *buf)
1805 {
1806         struct perf_event *leader = event->group_leader, *sub;
1807         int n = 0, size = 0, ret = -EFAULT;
1808         struct perf_event_context *ctx = leader->ctx;
1809         u64 values[5];
1810         u64 count, enabled, running;
1811
1812         mutex_lock(&ctx->mutex);
1813         count = perf_event_read_value(leader, &enabled, &running);
1814
1815         values[n++] = 1 + leader->nr_siblings;
1816         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1817                 values[n++] = enabled;
1818         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1819                 values[n++] = running;
1820         values[n++] = count;
1821         if (read_format & PERF_FORMAT_ID)
1822                 values[n++] = primary_event_id(leader);
1823
1824         size = n * sizeof(u64);
1825
1826         if (copy_to_user(buf, values, size))
1827                 goto unlock;
1828
1829         ret = size;
1830
1831         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1832                 n = 0;
1833
1834                 values[n++] = perf_event_read_value(sub, &enabled, &running);
1835                 if (read_format & PERF_FORMAT_ID)
1836                         values[n++] = primary_event_id(sub);
1837
1838                 size = n * sizeof(u64);
1839
1840                 if (copy_to_user(buf + size, values, size)) {
1841                         ret = -EFAULT;
1842                         goto unlock;
1843                 }
1844
1845                 ret += size;
1846         }
1847 unlock:
1848         mutex_unlock(&ctx->mutex);
1849
1850         return ret;
1851 }
1852
1853 static int perf_event_read_one(struct perf_event *event,
1854                                  u64 read_format, char __user *buf)
1855 {
1856         u64 enabled, running;
1857         u64 values[4];
1858         int n = 0;
1859
1860         values[n++] = perf_event_read_value(event, &enabled, &running);
1861         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1862                 values[n++] = enabled;
1863         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1864                 values[n++] = running;
1865         if (read_format & PERF_FORMAT_ID)
1866                 values[n++] = primary_event_id(event);
1867
1868         if (copy_to_user(buf, values, n * sizeof(u64)))
1869                 return -EFAULT;
1870
1871         return n * sizeof(u64);
1872 }
1873
1874 /*
1875  * Read the performance event - simple non blocking version for now
1876  */
1877 static ssize_t
1878 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
1879 {
1880         u64 read_format = event->attr.read_format;
1881         int ret;
1882
1883         /*
1884          * Return end-of-file for a read on a event that is in
1885          * error state (i.e. because it was pinned but it couldn't be
1886          * scheduled on to the CPU at some point).
1887          */
1888         if (event->state == PERF_EVENT_STATE_ERROR)
1889                 return 0;
1890
1891         if (count < perf_event_read_size(event))
1892                 return -ENOSPC;
1893
1894         WARN_ON_ONCE(event->ctx->parent_ctx);
1895         if (read_format & PERF_FORMAT_GROUP)
1896                 ret = perf_event_read_group(event, read_format, buf);
1897         else
1898                 ret = perf_event_read_one(event, read_format, buf);
1899
1900         return ret;
1901 }
1902
1903 static ssize_t
1904 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1905 {
1906         struct perf_event *event = file->private_data;
1907
1908         return perf_read_hw(event, buf, count);
1909 }
1910
1911 static unsigned int perf_poll(struct file *file, poll_table *wait)
1912 {
1913         struct perf_event *event = file->private_data;
1914         struct perf_mmap_data *data;
1915         unsigned int events = POLL_HUP;
1916
1917         rcu_read_lock();
1918         data = rcu_dereference(event->data);
1919         if (data)
1920                 events = atomic_xchg(&data->poll, 0);
1921         rcu_read_unlock();
1922
1923         poll_wait(file, &event->waitq, wait);
1924
1925         return events;
1926 }
1927
1928 static void perf_event_reset(struct perf_event *event)
1929 {
1930         (void)perf_event_read(event);
1931         atomic64_set(&event->count, 0);
1932         perf_event_update_userpage(event);
1933 }
1934
1935 /*
1936  * Holding the top-level event's child_mutex means that any
1937  * descendant process that has inherited this event will block
1938  * in sync_child_event if it goes to exit, thus satisfying the
1939  * task existence requirements of perf_event_enable/disable.
1940  */
1941 static void perf_event_for_each_child(struct perf_event *event,
1942                                         void (*func)(struct perf_event *))
1943 {
1944         struct perf_event *child;
1945
1946         WARN_ON_ONCE(event->ctx->parent_ctx);
1947         mutex_lock(&event->child_mutex);
1948         func(event);
1949         list_for_each_entry(child, &event->child_list, child_list)
1950                 func(child);
1951         mutex_unlock(&event->child_mutex);
1952 }
1953
1954 static void perf_event_for_each(struct perf_event *event,
1955                                   void (*func)(struct perf_event *))
1956 {
1957         struct perf_event_context *ctx = event->ctx;
1958         struct perf_event *sibling;
1959
1960         WARN_ON_ONCE(ctx->parent_ctx);
1961         mutex_lock(&ctx->mutex);
1962         event = event->group_leader;
1963
1964         perf_event_for_each_child(event, func);
1965         func(event);
1966         list_for_each_entry(sibling, &event->sibling_list, group_entry)
1967                 perf_event_for_each_child(event, func);
1968         mutex_unlock(&ctx->mutex);
1969 }
1970
1971 static int perf_event_period(struct perf_event *event, u64 __user *arg)
1972 {
1973         struct perf_event_context *ctx = event->ctx;
1974         unsigned long size;
1975         int ret = 0;
1976         u64 value;
1977
1978         if (!event->attr.sample_period)
1979                 return -EINVAL;
1980
1981         size = copy_from_user(&value, arg, sizeof(value));
1982         if (size != sizeof(value))
1983                 return -EFAULT;
1984
1985         if (!value)
1986                 return -EINVAL;
1987
1988         spin_lock_irq(&ctx->lock);
1989         if (event->attr.freq) {
1990                 if (value > sysctl_perf_event_sample_rate) {
1991                         ret = -EINVAL;
1992                         goto unlock;
1993                 }
1994
1995                 event->attr.sample_freq = value;
1996         } else {
1997                 event->attr.sample_period = value;
1998                 event->hw.sample_period = value;
1999         }
2000 unlock:
2001         spin_unlock_irq(&ctx->lock);
2002
2003         return ret;
2004 }
2005
2006 static int perf_event_set_output(struct perf_event *event, int output_fd);
2007 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2008
2009 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2010 {
2011         struct perf_event *event = file->private_data;
2012         void (*func)(struct perf_event *);
2013         u32 flags = arg;
2014
2015         switch (cmd) {
2016         case PERF_EVENT_IOC_ENABLE:
2017                 func = perf_event_enable;
2018                 break;
2019         case PERF_EVENT_IOC_DISABLE:
2020                 func = perf_event_disable;
2021                 break;
2022         case PERF_EVENT_IOC_RESET:
2023                 func = perf_event_reset;
2024                 break;
2025
2026         case PERF_EVENT_IOC_REFRESH:
2027                 return perf_event_refresh(event, arg);
2028
2029         case PERF_EVENT_IOC_PERIOD:
2030                 return perf_event_period(event, (u64 __user *)arg);
2031
2032         case PERF_EVENT_IOC_SET_OUTPUT:
2033                 return perf_event_set_output(event, arg);
2034
2035         case PERF_EVENT_IOC_SET_FILTER:
2036                 return perf_event_set_filter(event, (void __user *)arg);
2037
2038         default:
2039                 return -ENOTTY;
2040         }
2041
2042         if (flags & PERF_IOC_FLAG_GROUP)
2043                 perf_event_for_each(event, func);
2044         else
2045                 perf_event_for_each_child(event, func);
2046
2047         return 0;
2048 }
2049
2050 int perf_event_task_enable(void)
2051 {
2052         struct perf_event *event;
2053
2054         mutex_lock(&current->perf_event_mutex);
2055         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2056                 perf_event_for_each_child(event, perf_event_enable);
2057         mutex_unlock(&current->perf_event_mutex);
2058
2059         return 0;
2060 }
2061
2062 int perf_event_task_disable(void)
2063 {
2064         struct perf_event *event;
2065
2066         mutex_lock(&current->perf_event_mutex);
2067         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2068                 perf_event_for_each_child(event, perf_event_disable);
2069         mutex_unlock(&current->perf_event_mutex);
2070
2071         return 0;
2072 }
2073
2074 #ifndef PERF_EVENT_INDEX_OFFSET
2075 # define PERF_EVENT_INDEX_OFFSET 0
2076 #endif
2077
2078 static int perf_event_index(struct perf_event *event)
2079 {
2080         if (event->state != PERF_EVENT_STATE_ACTIVE)
2081                 return 0;
2082
2083         return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2084 }
2085
2086 /*
2087  * Callers need to ensure there can be no nesting of this function, otherwise
2088  * the seqlock logic goes bad. We can not serialize this because the arch
2089  * code calls this from NMI context.
2090  */
2091 void perf_event_update_userpage(struct perf_event *event)
2092 {
2093         struct perf_event_mmap_page *userpg;
2094         struct perf_mmap_data *data;
2095
2096         rcu_read_lock();
2097         data = rcu_dereference(event->data);
2098         if (!data)
2099                 goto unlock;
2100
2101         userpg = data->user_page;
2102
2103         /*
2104          * Disable preemption so as to not let the corresponding user-space
2105          * spin too long if we get preempted.
2106          */
2107         preempt_disable();
2108         ++userpg->lock;
2109         barrier();
2110         userpg->index = perf_event_index(event);
2111         userpg->offset = atomic64_read(&event->count);
2112         if (event->state == PERF_EVENT_STATE_ACTIVE)
2113                 userpg->offset -= atomic64_read(&event->hw.prev_count);
2114
2115         userpg->time_enabled = event->total_time_enabled +
2116                         atomic64_read(&event->child_total_time_enabled);
2117
2118         userpg->time_running = event->total_time_running +
2119                         atomic64_read(&event->child_total_time_running);
2120
2121         barrier();
2122         ++userpg->lock;
2123         preempt_enable();
2124 unlock:
2125         rcu_read_unlock();
2126 }
2127
2128 static unsigned long perf_data_size(struct perf_mmap_data *data)
2129 {
2130         return data->nr_pages << (PAGE_SHIFT + data->data_order);
2131 }
2132
2133 #ifndef CONFIG_PERF_USE_VMALLOC
2134
2135 /*
2136  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2137  */
2138
2139 static struct page *
2140 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2141 {
2142         if (pgoff > data->nr_pages)
2143                 return NULL;
2144
2145         if (pgoff == 0)
2146                 return virt_to_page(data->user_page);
2147
2148         return virt_to_page(data->data_pages[pgoff - 1]);
2149 }
2150
2151 static struct perf_mmap_data *
2152 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2153 {
2154         struct perf_mmap_data *data;
2155         unsigned long size;
2156         int i;
2157
2158         WARN_ON(atomic_read(&event->mmap_count));
2159
2160         size = sizeof(struct perf_mmap_data);
2161         size += nr_pages * sizeof(void *);
2162
2163         data = kzalloc(size, GFP_KERNEL);
2164         if (!data)
2165                 goto fail;
2166
2167         data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2168         if (!data->user_page)
2169                 goto fail_user_page;
2170
2171         for (i = 0; i < nr_pages; i++) {
2172                 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2173                 if (!data->data_pages[i])
2174                         goto fail_data_pages;
2175         }
2176
2177         data->data_order = 0;
2178         data->nr_pages = nr_pages;
2179
2180         return data;
2181
2182 fail_data_pages:
2183         for (i--; i >= 0; i--)
2184                 free_page((unsigned long)data->data_pages[i]);
2185
2186         free_page((unsigned long)data->user_page);
2187
2188 fail_user_page:
2189         kfree(data);
2190
2191 fail:
2192         return NULL;
2193 }
2194
2195 static void perf_mmap_free_page(unsigned long addr)
2196 {
2197         struct page *page = virt_to_page((void *)addr);
2198
2199         page->mapping = NULL;
2200         __free_page(page);
2201 }
2202
2203 static void perf_mmap_data_free(struct perf_mmap_data *data)
2204 {
2205         int i;
2206
2207         perf_mmap_free_page((unsigned long)data->user_page);
2208         for (i = 0; i < data->nr_pages; i++)
2209                 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2210 }
2211
2212 #else
2213
2214 /*
2215  * Back perf_mmap() with vmalloc memory.
2216  *
2217  * Required for architectures that have d-cache aliasing issues.
2218  */
2219
2220 static struct page *
2221 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2222 {
2223         if (pgoff > (1UL << data->data_order))
2224                 return NULL;
2225
2226         return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2227 }
2228
2229 static void perf_mmap_unmark_page(void *addr)
2230 {
2231         struct page *page = vmalloc_to_page(addr);
2232
2233         page->mapping = NULL;
2234 }
2235
2236 static void perf_mmap_data_free_work(struct work_struct *work)
2237 {
2238         struct perf_mmap_data *data;
2239         void *base;
2240         int i, nr;
2241
2242         data = container_of(work, struct perf_mmap_data, work);
2243         nr = 1 << data->data_order;
2244
2245         base = data->user_page;
2246         for (i = 0; i < nr + 1; i++)
2247                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2248
2249         vfree(base);
2250 }
2251
2252 static void perf_mmap_data_free(struct perf_mmap_data *data)
2253 {
2254         schedule_work(&data->work);
2255 }
2256
2257 static struct perf_mmap_data *
2258 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2259 {
2260         struct perf_mmap_data *data;
2261         unsigned long size;
2262         void *all_buf;
2263
2264         WARN_ON(atomic_read(&event->mmap_count));
2265
2266         size = sizeof(struct perf_mmap_data);
2267         size += sizeof(void *);
2268
2269         data = kzalloc(size, GFP_KERNEL);
2270         if (!data)
2271                 goto fail;
2272
2273         INIT_WORK(&data->work, perf_mmap_data_free_work);
2274
2275         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2276         if (!all_buf)
2277                 goto fail_all_buf;
2278
2279         data->user_page = all_buf;
2280         data->data_pages[0] = all_buf + PAGE_SIZE;
2281         data->data_order = ilog2(nr_pages);
2282         data->nr_pages = 1;
2283
2284         return data;
2285
2286 fail_all_buf:
2287         kfree(data);
2288
2289 fail:
2290         return NULL;
2291 }
2292
2293 #endif
2294
2295 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2296 {
2297         struct perf_event *event = vma->vm_file->private_data;
2298         struct perf_mmap_data *data;
2299         int ret = VM_FAULT_SIGBUS;
2300
2301         if (vmf->flags & FAULT_FLAG_MKWRITE) {
2302                 if (vmf->pgoff == 0)
2303                         ret = 0;
2304                 return ret;
2305         }
2306
2307         rcu_read_lock();
2308         data = rcu_dereference(event->data);
2309         if (!data)
2310                 goto unlock;
2311
2312         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2313                 goto unlock;
2314
2315         vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2316         if (!vmf->page)
2317                 goto unlock;
2318
2319         get_page(vmf->page);
2320         vmf->page->mapping = vma->vm_file->f_mapping;
2321         vmf->page->index   = vmf->pgoff;
2322
2323         ret = 0;
2324 unlock:
2325         rcu_read_unlock();
2326
2327         return ret;
2328 }
2329
2330 static void
2331 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2332 {
2333         long max_size = perf_data_size(data);
2334
2335         atomic_set(&data->lock, -1);
2336
2337         if (event->attr.watermark) {
2338                 data->watermark = min_t(long, max_size,
2339                                         event->attr.wakeup_watermark);
2340         }
2341
2342         if (!data->watermark)
2343                 data->watermark = max_size / 2;
2344
2345
2346         rcu_assign_pointer(event->data, data);
2347 }
2348
2349 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2350 {
2351         struct perf_mmap_data *data;
2352
2353         data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2354         perf_mmap_data_free(data);
2355         kfree(data);
2356 }
2357
2358 static void perf_mmap_data_release(struct perf_event *event)
2359 {
2360         struct perf_mmap_data *data = event->data;
2361
2362         WARN_ON(atomic_read(&event->mmap_count));
2363
2364         rcu_assign_pointer(event->data, NULL);
2365         call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2366 }
2367
2368 static void perf_mmap_open(struct vm_area_struct *vma)
2369 {
2370         struct perf_event *event = vma->vm_file->private_data;
2371
2372         atomic_inc(&event->mmap_count);
2373 }
2374
2375 static void perf_mmap_close(struct vm_area_struct *vma)
2376 {
2377         struct perf_event *event = vma->vm_file->private_data;
2378
2379         WARN_ON_ONCE(event->ctx->parent_ctx);
2380         if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2381                 unsigned long size = perf_data_size(event->data);
2382                 struct user_struct *user = current_user();
2383
2384                 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2385                 vma->vm_mm->locked_vm -= event->data->nr_locked;
2386                 perf_mmap_data_release(event);
2387                 mutex_unlock(&event->mmap_mutex);
2388         }
2389 }
2390
2391 static const struct vm_operations_struct perf_mmap_vmops = {
2392         .open           = perf_mmap_open,
2393         .close          = perf_mmap_close,
2394         .fault          = perf_mmap_fault,
2395         .page_mkwrite   = perf_mmap_fault,
2396 };
2397
2398 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2399 {
2400         struct perf_event *event = file->private_data;
2401         unsigned long user_locked, user_lock_limit;
2402         struct user_struct *user = current_user();
2403         unsigned long locked, lock_limit;
2404         struct perf_mmap_data *data;
2405         unsigned long vma_size;
2406         unsigned long nr_pages;
2407         long user_extra, extra;
2408         int ret = 0;
2409
2410         if (!(vma->vm_flags & VM_SHARED))
2411                 return -EINVAL;
2412
2413         vma_size = vma->vm_end - vma->vm_start;
2414         nr_pages = (vma_size / PAGE_SIZE) - 1;
2415
2416         /*
2417          * If we have data pages ensure they're a power-of-two number, so we
2418          * can do bitmasks instead of modulo.
2419          */
2420         if (nr_pages != 0 && !is_power_of_2(nr_pages))
2421                 return -EINVAL;
2422
2423         if (vma_size != PAGE_SIZE * (1 + nr_pages))
2424                 return -EINVAL;
2425
2426         if (vma->vm_pgoff != 0)
2427                 return -EINVAL;
2428
2429         WARN_ON_ONCE(event->ctx->parent_ctx);
2430         mutex_lock(&event->mmap_mutex);
2431         if (event->output) {
2432                 ret = -EINVAL;
2433                 goto unlock;
2434         }
2435
2436         if (atomic_inc_not_zero(&event->mmap_count)) {
2437                 if (nr_pages != event->data->nr_pages)
2438                         ret = -EINVAL;
2439                 goto unlock;
2440         }
2441
2442         user_extra = nr_pages + 1;
2443         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2444
2445         /*
2446          * Increase the limit linearly with more CPUs:
2447          */
2448         user_lock_limit *= num_online_cpus();
2449
2450         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2451
2452         extra = 0;
2453         if (user_locked > user_lock_limit)
2454                 extra = user_locked - user_lock_limit;
2455
2456         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2457         lock_limit >>= PAGE_SHIFT;
2458         locked = vma->vm_mm->locked_vm + extra;
2459
2460         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2461                 !capable(CAP_IPC_LOCK)) {
2462                 ret = -EPERM;
2463                 goto unlock;
2464         }
2465
2466         WARN_ON(event->data);
2467
2468         data = perf_mmap_data_alloc(event, nr_pages);
2469         ret = -ENOMEM;
2470         if (!data)
2471                 goto unlock;
2472
2473         ret = 0;
2474         perf_mmap_data_init(event, data);
2475
2476         atomic_set(&event->mmap_count, 1);
2477         atomic_long_add(user_extra, &user->locked_vm);
2478         vma->vm_mm->locked_vm += extra;
2479         event->data->nr_locked = extra;
2480         if (vma->vm_flags & VM_WRITE)
2481                 event->data->writable = 1;
2482
2483 unlock:
2484         mutex_unlock(&event->mmap_mutex);
2485
2486         vma->vm_flags |= VM_RESERVED;
2487         vma->vm_ops = &perf_mmap_vmops;
2488
2489         return ret;
2490 }
2491
2492 static int perf_fasync(int fd, struct file *filp, int on)
2493 {
2494         struct inode *inode = filp->f_path.dentry->d_inode;
2495         struct perf_event *event = filp->private_data;
2496         int retval;
2497
2498         mutex_lock(&inode->i_mutex);
2499         retval = fasync_helper(fd, filp, on, &event->fasync);
2500         mutex_unlock(&inode->i_mutex);
2501
2502         if (retval < 0)
2503                 return retval;
2504
2505         return 0;
2506 }
2507
2508 static const struct file_operations perf_fops = {
2509         .release                = perf_release,
2510         .read                   = perf_read,
2511         .poll                   = perf_poll,
2512         .unlocked_ioctl         = perf_ioctl,
2513         .compat_ioctl           = perf_ioctl,
2514         .mmap                   = perf_mmap,
2515         .fasync                 = perf_fasync,
2516 };
2517
2518 /*
2519  * Perf event wakeup
2520  *
2521  * If there's data, ensure we set the poll() state and publish everything
2522  * to user-space before waking everybody up.
2523  */
2524
2525 void perf_event_wakeup(struct perf_event *event)
2526 {
2527         wake_up_all(&event->waitq);
2528
2529         if (event->pending_kill) {
2530                 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2531                 event->pending_kill = 0;
2532         }
2533 }
2534
2535 /*
2536  * Pending wakeups
2537  *
2538  * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2539  *
2540  * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2541  * single linked list and use cmpxchg() to add entries lockless.
2542  */
2543
2544 static void perf_pending_event(struct perf_pending_entry *entry)
2545 {
2546         struct perf_event *event = container_of(entry,
2547                         struct perf_event, pending);
2548
2549         if (event->pending_disable) {
2550                 event->pending_disable = 0;
2551                 __perf_event_disable(event);
2552         }
2553
2554         if (event->pending_wakeup) {
2555                 event->pending_wakeup = 0;
2556                 perf_event_wakeup(event);
2557         }
2558 }
2559
2560 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2561
2562 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2563         PENDING_TAIL,
2564 };
2565
2566 static void perf_pending_queue(struct perf_pending_entry *entry,
2567                                void (*func)(struct perf_pending_entry *))
2568 {
2569         struct perf_pending_entry **head;
2570
2571         if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2572                 return;
2573
2574         entry->func = func;
2575
2576         head = &get_cpu_var(perf_pending_head);
2577
2578         do {
2579                 entry->next = *head;
2580         } while (cmpxchg(head, entry->next, entry) != entry->next);
2581
2582         set_perf_event_pending();
2583
2584         put_cpu_var(perf_pending_head);
2585 }
2586
2587 static int __perf_pending_run(void)
2588 {
2589         struct perf_pending_entry *list;
2590         int nr = 0;
2591
2592         list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2593         while (list != PENDING_TAIL) {
2594                 void (*func)(struct perf_pending_entry *);
2595                 struct perf_pending_entry *entry = list;
2596
2597                 list = list->next;
2598
2599                 func = entry->func;
2600                 entry->next = NULL;
2601                 /*
2602                  * Ensure we observe the unqueue before we issue the wakeup,
2603                  * so that we won't be waiting forever.
2604                  * -- see perf_not_pending().
2605                  */
2606                 smp_wmb();
2607
2608                 func(entry);
2609                 nr++;
2610         }
2611
2612         return nr;
2613 }
2614
2615 static inline int perf_not_pending(struct perf_event *event)
2616 {
2617         /*
2618          * If we flush on whatever cpu we run, there is a chance we don't
2619          * need to wait.
2620          */
2621         get_cpu();
2622         __perf_pending_run();
2623         put_cpu();
2624
2625         /*
2626          * Ensure we see the proper queue state before going to sleep
2627          * so that we do not miss the wakeup. -- see perf_pending_handle()
2628          */
2629         smp_rmb();
2630         return event->pending.next == NULL;
2631 }
2632
2633 static void perf_pending_sync(struct perf_event *event)
2634 {
2635         wait_event(event->waitq, perf_not_pending(event));
2636 }
2637
2638 void perf_event_do_pending(void)
2639 {
2640         __perf_pending_run();
2641 }
2642
2643 /*
2644  * Callchain support -- arch specific
2645  */
2646
2647 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2648 {
2649         return NULL;
2650 }
2651
2652 /*
2653  * Output
2654  */
2655 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2656                               unsigned long offset, unsigned long head)
2657 {
2658         unsigned long mask;
2659
2660         if (!data->writable)
2661                 return true;
2662
2663         mask = perf_data_size(data) - 1;
2664
2665         offset = (offset - tail) & mask;
2666         head   = (head   - tail) & mask;
2667
2668         if ((int)(head - offset) < 0)
2669                 return false;
2670
2671         return true;
2672 }
2673
2674 static void perf_output_wakeup(struct perf_output_handle *handle)
2675 {
2676         atomic_set(&handle->data->poll, POLL_IN);
2677
2678         if (handle->nmi) {
2679                 handle->event->pending_wakeup = 1;
2680                 perf_pending_queue(&handle->event->pending,
2681                                    perf_pending_event);
2682         } else
2683                 perf_event_wakeup(handle->event);
2684 }
2685
2686 /*
2687  * Curious locking construct.
2688  *
2689  * We need to ensure a later event_id doesn't publish a head when a former
2690  * event_id isn't done writing. However since we need to deal with NMIs we
2691  * cannot fully serialize things.
2692  *
2693  * What we do is serialize between CPUs so we only have to deal with NMI
2694  * nesting on a single CPU.
2695  *
2696  * We only publish the head (and generate a wakeup) when the outer-most
2697  * event_id completes.
2698  */
2699 static void perf_output_lock(struct perf_output_handle *handle)
2700 {
2701         struct perf_mmap_data *data = handle->data;
2702         int cur, cpu = get_cpu();
2703
2704         handle->locked = 0;
2705
2706         for (;;) {
2707                 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2708                 if (cur == -1) {
2709                         handle->locked = 1;
2710                         break;
2711                 }
2712                 if (cur == cpu)
2713                         break;
2714
2715                 cpu_relax();
2716         }
2717 }
2718
2719 static void perf_output_unlock(struct perf_output_handle *handle)
2720 {
2721         struct perf_mmap_data *data = handle->data;
2722         unsigned long head;
2723         int cpu;
2724
2725         data->done_head = data->head;
2726
2727         if (!handle->locked)
2728                 goto out;
2729
2730 again:
2731         /*
2732          * The xchg implies a full barrier that ensures all writes are done
2733          * before we publish the new head, matched by a rmb() in userspace when
2734          * reading this position.
2735          */
2736         while ((head = atomic_long_xchg(&data->done_head, 0)))
2737                 data->user_page->data_head = head;
2738
2739         /*
2740          * NMI can happen here, which means we can miss a done_head update.
2741          */
2742
2743         cpu = atomic_xchg(&data->lock, -1);
2744         WARN_ON_ONCE(cpu != smp_processor_id());
2745
2746         /*
2747          * Therefore we have to validate we did not indeed do so.
2748          */
2749         if (unlikely(atomic_long_read(&data->done_head))) {
2750                 /*
2751                  * Since we had it locked, we can lock it again.
2752                  */
2753                 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2754                         cpu_relax();
2755
2756                 goto again;
2757         }
2758
2759         if (atomic_xchg(&data->wakeup, 0))
2760                 perf_output_wakeup(handle);
2761 out:
2762         put_cpu();
2763 }
2764
2765 void perf_output_copy(struct perf_output_handle *handle,
2766                       const void *buf, unsigned int len)
2767 {
2768         unsigned int pages_mask;
2769         unsigned long offset;
2770         unsigned int size;
2771         void **pages;
2772
2773         offset          = handle->offset;
2774         pages_mask      = handle->data->nr_pages - 1;
2775         pages           = handle->data->data_pages;
2776
2777         do {
2778                 unsigned long page_offset;
2779                 unsigned long page_size;
2780                 int nr;
2781
2782                 nr          = (offset >> PAGE_SHIFT) & pages_mask;
2783                 page_size   = 1UL << (handle->data->data_order + PAGE_SHIFT);
2784                 page_offset = offset & (page_size - 1);
2785                 size        = min_t(unsigned int, page_size - page_offset, len);
2786
2787                 memcpy(pages[nr] + page_offset, buf, size);
2788
2789                 len         -= size;
2790                 buf         += size;
2791                 offset      += size;
2792         } while (len);
2793
2794         handle->offset = offset;
2795
2796         /*
2797          * Check we didn't copy past our reservation window, taking the
2798          * possible unsigned int wrap into account.
2799          */
2800         WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2801 }
2802
2803 int perf_output_begin(struct perf_output_handle *handle,
2804                       struct perf_event *event, unsigned int size,
2805                       int nmi, int sample)
2806 {
2807         struct perf_event *output_event;
2808         struct perf_mmap_data *data;
2809         unsigned long tail, offset, head;
2810         int have_lost;
2811         struct {
2812                 struct perf_event_header header;
2813                 u64                      id;
2814                 u64                      lost;
2815         } lost_event;
2816
2817         rcu_read_lock();
2818         /*
2819          * For inherited events we send all the output towards the parent.
2820          */
2821         if (event->parent)
2822                 event = event->parent;
2823
2824         output_event = rcu_dereference(event->output);
2825         if (output_event)
2826                 event = output_event;
2827
2828         data = rcu_dereference(event->data);
2829         if (!data)
2830                 goto out;
2831
2832         handle->data    = data;
2833         handle->event   = event;
2834         handle->nmi     = nmi;
2835         handle->sample  = sample;
2836
2837         if (!data->nr_pages)
2838                 goto fail;
2839
2840         have_lost = atomic_read(&data->lost);
2841         if (have_lost)
2842                 size += sizeof(lost_event);
2843
2844         perf_output_lock(handle);
2845
2846         do {
2847                 /*
2848                  * Userspace could choose to issue a mb() before updating the
2849                  * tail pointer. So that all reads will be completed before the
2850                  * write is issued.
2851                  */
2852                 tail = ACCESS_ONCE(data->user_page->data_tail);
2853                 smp_rmb();
2854                 offset = head = atomic_long_read(&data->head);
2855                 head += size;
2856                 if (unlikely(!perf_output_space(data, tail, offset, head)))
2857                         goto fail;
2858         } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2859
2860         handle->offset  = offset;
2861         handle->head    = head;
2862
2863         if (head - tail > data->watermark)
2864                 atomic_set(&data->wakeup, 1);
2865
2866         if (have_lost) {
2867                 lost_event.header.type = PERF_RECORD_LOST;
2868                 lost_event.header.misc = 0;
2869                 lost_event.header.size = sizeof(lost_event);
2870                 lost_event.id          = event->id;
2871                 lost_event.lost        = atomic_xchg(&data->lost, 0);
2872
2873                 perf_output_put(handle, lost_event);
2874         }
2875
2876         return 0;
2877
2878 fail:
2879         atomic_inc(&data->lost);
2880         perf_output_unlock(handle);
2881 out:
2882         rcu_read_unlock();
2883
2884         return -ENOSPC;
2885 }
2886
2887 void perf_output_end(struct perf_output_handle *handle)
2888 {
2889         struct perf_event *event = handle->event;
2890         struct perf_mmap_data *data = handle->data;
2891
2892         int wakeup_events = event->attr.wakeup_events;
2893
2894         if (handle->sample && wakeup_events) {
2895                 int events = atomic_inc_return(&data->events);
2896                 if (events >= wakeup_events) {
2897                         atomic_sub(wakeup_events, &data->events);
2898                         atomic_set(&data->wakeup, 1);
2899                 }
2900         }
2901
2902         perf_output_unlock(handle);
2903         rcu_read_unlock();
2904 }
2905
2906 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
2907 {
2908         /*
2909          * only top level events have the pid namespace they were created in
2910          */
2911         if (event->parent)
2912                 event = event->parent;
2913
2914         return task_tgid_nr_ns(p, event->ns);
2915 }
2916
2917 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
2918 {
2919         /*
2920          * only top level events have the pid namespace they were created in
2921          */
2922         if (event->parent)
2923                 event = event->parent;
2924
2925         return task_pid_nr_ns(p, event->ns);
2926 }
2927
2928 static void perf_output_read_one(struct perf_output_handle *handle,
2929                                  struct perf_event *event)
2930 {
2931         u64 read_format = event->attr.read_format;
2932         u64 values[4];
2933         int n = 0;
2934
2935         values[n++] = atomic64_read(&event->count);
2936         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2937                 values[n++] = event->total_time_enabled +
2938                         atomic64_read(&event->child_total_time_enabled);
2939         }
2940         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2941                 values[n++] = event->total_time_running +
2942                         atomic64_read(&event->child_total_time_running);
2943         }
2944         if (read_format & PERF_FORMAT_ID)
2945                 values[n++] = primary_event_id(event);
2946
2947         perf_output_copy(handle, values, n * sizeof(u64));
2948 }
2949
2950 /*
2951  * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
2952  */
2953 static void perf_output_read_group(struct perf_output_handle *handle,
2954                             struct perf_event *event)
2955 {
2956         struct perf_event *leader = event->group_leader, *sub;
2957         u64 read_format = event->attr.read_format;
2958         u64 values[5];
2959         int n = 0;
2960
2961         values[n++] = 1 + leader->nr_siblings;
2962
2963         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2964                 values[n++] = leader->total_time_enabled;
2965
2966         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2967                 values[n++] = leader->total_time_running;
2968
2969         if (leader != event)
2970                 leader->pmu->read(leader);
2971
2972         values[n++] = atomic64_read(&leader->count);
2973         if (read_format & PERF_FORMAT_ID)
2974                 values[n++] = primary_event_id(leader);
2975
2976         perf_output_copy(handle, values, n * sizeof(u64));
2977
2978         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2979                 n = 0;
2980
2981                 if (sub != event)
2982                         sub->pmu->read(sub);
2983
2984                 values[n++] = atomic64_read(&sub->count);
2985                 if (read_format & PERF_FORMAT_ID)
2986                         values[n++] = primary_event_id(sub);
2987
2988                 perf_output_copy(handle, values, n * sizeof(u64));
2989         }
2990 }
2991
2992 static void perf_output_read(struct perf_output_handle *handle,
2993                              struct perf_event *event)
2994 {
2995         if (event->attr.read_format & PERF_FORMAT_GROUP)
2996                 perf_output_read_group(handle, event);
2997         else
2998                 perf_output_read_one(handle, event);
2999 }
3000
3001 void perf_output_sample(struct perf_output_handle *handle,
3002                         struct perf_event_header *header,
3003                         struct perf_sample_data *data,
3004                         struct perf_event *event)
3005 {
3006         u64 sample_type = data->type;
3007
3008         perf_output_put(handle, *header);
3009
3010         if (sample_type & PERF_SAMPLE_IP)
3011                 perf_output_put(handle, data->ip);
3012
3013         if (sample_type & PERF_SAMPLE_TID)
3014                 perf_output_put(handle, data->tid_entry);
3015
3016         if (sample_type & PERF_SAMPLE_TIME)
3017                 perf_output_put(handle, data->time);
3018
3019         if (sample_type & PERF_SAMPLE_ADDR)
3020                 perf_output_put(handle, data->addr);
3021
3022         if (sample_type & PERF_SAMPLE_ID)
3023                 perf_output_put(handle, data->id);
3024
3025         if (sample_type & PERF_SAMPLE_STREAM_ID)
3026                 perf_output_put(handle, data->stream_id);
3027
3028         if (sample_type & PERF_SAMPLE_CPU)
3029                 perf_output_put(handle, data->cpu_entry);
3030
3031         if (sample_type & PERF_SAMPLE_PERIOD)
3032                 perf_output_put(handle, data->period);
3033
3034         if (sample_type & PERF_SAMPLE_READ)
3035                 perf_output_read(handle, event);
3036
3037         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3038                 if (data->callchain) {
3039                         int size = 1;
3040
3041                         if (data->callchain)
3042                                 size += data->callchain->nr;
3043
3044                         size *= sizeof(u64);
3045
3046                         perf_output_copy(handle, data->callchain, size);
3047                 } else {
3048                         u64 nr = 0;
3049                         perf_output_put(handle, nr);
3050                 }
3051         }
3052
3053         if (sample_type & PERF_SAMPLE_RAW) {
3054                 if (data->raw) {
3055                         perf_output_put(handle, data->raw->size);
3056                         perf_output_copy(handle, data->raw->data,
3057                                          data->raw->size);
3058                 } else {
3059                         struct {
3060                                 u32     size;
3061                                 u32     data;
3062                         } raw = {
3063                                 .size = sizeof(u32),
3064                                 .data = 0,
3065                         };
3066                         perf_output_put(handle, raw);
3067                 }
3068         }
3069 }
3070
3071 void perf_prepare_sample(struct perf_event_header *header,
3072                          struct perf_sample_data *data,
3073                          struct perf_event *event,
3074                          struct pt_regs *regs)
3075 {
3076         u64 sample_type = event->attr.sample_type;
3077
3078         data->type = sample_type;
3079
3080         header->type = PERF_RECORD_SAMPLE;
3081         header->size = sizeof(*header);
3082
3083         header->misc = 0;
3084         header->misc |= perf_misc_flags(regs);
3085
3086         if (sample_type & PERF_SAMPLE_IP) {
3087                 data->ip = perf_instruction_pointer(regs);
3088
3089                 header->size += sizeof(data->ip);
3090         }
3091
3092         if (sample_type & PERF_SAMPLE_TID) {
3093                 /* namespace issues */
3094                 data->tid_entry.pid = perf_event_pid(event, current);
3095                 data->tid_entry.tid = perf_event_tid(event, current);
3096
3097                 header->size += sizeof(data->tid_entry);
3098         }
3099
3100         if (sample_type & PERF_SAMPLE_TIME) {
3101                 data->time = perf_clock();
3102
3103                 header->size += sizeof(data->time);
3104         }
3105
3106         if (sample_type & PERF_SAMPLE_ADDR)
3107                 header->size += sizeof(data->addr);
3108
3109         if (sample_type & PERF_SAMPLE_ID) {
3110                 data->id = primary_event_id(event);
3111
3112                 header->size += sizeof(data->id);
3113         }
3114
3115         if (sample_type & PERF_SAMPLE_STREAM_ID) {
3116                 data->stream_id = event->id;
3117
3118                 header->size += sizeof(data->stream_id);
3119         }
3120
3121         if (sample_type & PERF_SAMPLE_CPU) {
3122                 data->cpu_entry.cpu             = raw_smp_processor_id();
3123                 data->cpu_entry.reserved        = 0;
3124
3125                 header->size += sizeof(data->cpu_entry);
3126         }
3127
3128         if (sample_type & PERF_SAMPLE_PERIOD)
3129                 header->size += sizeof(data->period);
3130
3131         if (sample_type & PERF_SAMPLE_READ)
3132                 header->size += perf_event_read_size(event);
3133
3134         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3135                 int size = 1;
3136
3137                 data->callchain = perf_callchain(regs);
3138
3139                 if (data->callchain)
3140                         size += data->callchain->nr;
3141
3142                 header->size += size * sizeof(u64);
3143         }
3144
3145         if (sample_type & PERF_SAMPLE_RAW) {
3146                 int size = sizeof(u32);
3147
3148                 if (data->raw)
3149                         size += data->raw->size;
3150                 else
3151                         size += sizeof(u32);
3152
3153                 WARN_ON_ONCE(size & (sizeof(u64)-1));
3154                 header->size += size;
3155         }
3156 }
3157
3158 static void perf_event_output(struct perf_event *event, int nmi,
3159                                 struct perf_sample_data *data,
3160                                 struct pt_regs *regs)
3161 {
3162         struct perf_output_handle handle;
3163         struct perf_event_header header;
3164
3165         perf_prepare_sample(&header, data, event, regs);
3166
3167         if (perf_output_begin(&handle, event, header.size, nmi, 1))
3168                 return;
3169
3170         perf_output_sample(&handle, &header, data, event);
3171
3172         perf_output_end(&handle);
3173 }
3174
3175 /*
3176  * read event_id
3177  */
3178
3179 struct perf_read_event {
3180         struct perf_event_header        header;
3181
3182         u32                             pid;
3183         u32                             tid;
3184 };
3185
3186 static void
3187 perf_event_read_event(struct perf_event *event,
3188                         struct task_struct *task)
3189 {
3190         struct perf_output_handle handle;
3191         struct perf_read_event read_event = {
3192                 .header = {
3193                         .type = PERF_RECORD_READ,
3194                         .misc = 0,
3195                         .size = sizeof(read_event) + perf_event_read_size(event),
3196                 },
3197                 .pid = perf_event_pid(event, task),
3198                 .tid = perf_event_tid(event, task),
3199         };
3200         int ret;
3201
3202         ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3203         if (ret)
3204                 return;
3205
3206         perf_output_put(&handle, read_event);
3207         perf_output_read(&handle, event);
3208
3209         perf_output_end(&handle);
3210 }
3211
3212 /*
3213  * task tracking -- fork/exit
3214  *
3215  * enabled by: attr.comm | attr.mmap | attr.task
3216  */
3217
3218 struct perf_task_event {
3219         struct task_struct              *task;
3220         struct perf_event_context       *task_ctx;
3221
3222         struct {
3223                 struct perf_event_header        header;
3224
3225                 u32                             pid;
3226                 u32                             ppid;
3227                 u32                             tid;
3228                 u32                             ptid;
3229                 u64                             time;
3230         } event_id;
3231 };
3232
3233 static void perf_event_task_output(struct perf_event *event,
3234                                      struct perf_task_event *task_event)
3235 {
3236         struct perf_output_handle handle;
3237         int size;
3238         struct task_struct *task = task_event->task;
3239         int ret;
3240
3241         size  = task_event->event_id.header.size;
3242         ret = perf_output_begin(&handle, event, size, 0, 0);
3243
3244         if (ret)
3245                 return;
3246
3247         task_event->event_id.pid = perf_event_pid(event, task);
3248         task_event->event_id.ppid = perf_event_pid(event, current);
3249
3250         task_event->event_id.tid = perf_event_tid(event, task);
3251         task_event->event_id.ptid = perf_event_tid(event, current);
3252
3253         task_event->event_id.time = perf_clock();
3254
3255         perf_output_put(&handle, task_event->event_id);
3256
3257         perf_output_end(&handle);
3258 }
3259
3260 static int perf_event_task_match(struct perf_event *event)
3261 {
3262         if (event->attr.comm || event->attr.mmap || event->attr.task)
3263                 return 1;
3264
3265         return 0;
3266 }
3267
3268 static void perf_event_task_ctx(struct perf_event_context *ctx,
3269                                   struct perf_task_event *task_event)
3270 {
3271         struct perf_event *event;
3272
3273         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3274                 if (perf_event_task_match(event))
3275                         perf_event_task_output(event, task_event);
3276         }
3277 }
3278
3279 static void perf_event_task_event(struct perf_task_event *task_event)
3280 {
3281         struct perf_cpu_context *cpuctx;
3282         struct perf_event_context *ctx = task_event->task_ctx;
3283
3284         rcu_read_lock();
3285         cpuctx = &get_cpu_var(perf_cpu_context);
3286         perf_event_task_ctx(&cpuctx->ctx, task_event);
3287         put_cpu_var(perf_cpu_context);
3288
3289         if (!ctx)
3290                 ctx = rcu_dereference(task_event->task->perf_event_ctxp);
3291         if (ctx)
3292                 perf_event_task_ctx(ctx, task_event);
3293         rcu_read_unlock();
3294 }
3295
3296 static void perf_event_task(struct task_struct *task,
3297                               struct perf_event_context *task_ctx,
3298                               int new)
3299 {
3300         struct perf_task_event task_event;
3301
3302         if (!atomic_read(&nr_comm_events) &&
3303             !atomic_read(&nr_mmap_events) &&
3304             !atomic_read(&nr_task_events))
3305                 return;
3306
3307         task_event = (struct perf_task_event){
3308                 .task     = task,
3309                 .task_ctx = task_ctx,
3310                 .event_id    = {
3311                         .header = {
3312                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3313                                 .misc = 0,
3314                                 .size = sizeof(task_event.event_id),
3315                         },
3316                         /* .pid  */
3317                         /* .ppid */
3318                         /* .tid  */
3319                         /* .ptid */
3320                 },
3321         };
3322
3323         perf_event_task_event(&task_event);
3324 }
3325
3326 void perf_event_fork(struct task_struct *task)
3327 {
3328         perf_event_task(task, NULL, 1);
3329 }
3330
3331 /*
3332  * comm tracking
3333  */
3334
3335 struct perf_comm_event {
3336         struct task_struct      *task;
3337         char                    *comm;
3338         int                     comm_size;
3339
3340         struct {
3341                 struct perf_event_header        header;
3342
3343                 u32                             pid;
3344                 u32                             tid;
3345         } event_id;
3346 };
3347
3348 static void perf_event_comm_output(struct perf_event *event,
3349                                      struct perf_comm_event *comm_event)
3350 {
3351         struct perf_output_handle handle;
3352         int size = comm_event->event_id.header.size;
3353         int ret = perf_output_begin(&handle, event, size, 0, 0);
3354
3355         if (ret)
3356                 return;
3357
3358         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3359         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3360
3361         perf_output_put(&handle, comm_event->event_id);
3362         perf_output_copy(&handle, comm_event->comm,
3363                                    comm_event->comm_size);
3364         perf_output_end(&handle);
3365 }
3366
3367 static int perf_event_comm_match(struct perf_event *event)
3368 {
3369         if (event->attr.comm)
3370                 return 1;
3371
3372         return 0;
3373 }
3374
3375 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3376                                   struct perf_comm_event *comm_event)
3377 {
3378         struct perf_event *event;
3379
3380         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3381                 if (perf_event_comm_match(event))
3382                         perf_event_comm_output(event, comm_event);
3383         }
3384 }
3385
3386 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3387 {
3388         struct perf_cpu_context *cpuctx;
3389         struct perf_event_context *ctx;
3390         unsigned int size;
3391         char comm[TASK_COMM_LEN];
3392
3393         memset(comm, 0, sizeof(comm));
3394         strncpy(comm, comm_event->task->comm, sizeof(comm));
3395         size = ALIGN(strlen(comm)+1, sizeof(u64));
3396
3397         comm_event->comm = comm;
3398         comm_event->comm_size = size;
3399
3400         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3401
3402         rcu_read_lock();
3403         cpuctx = &get_cpu_var(perf_cpu_context);
3404         perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3405         put_cpu_var(perf_cpu_context);
3406
3407         /*
3408          * doesn't really matter which of the child contexts the
3409          * events ends up in.
3410          */
3411         ctx = rcu_dereference(current->perf_event_ctxp);
3412         if (ctx)
3413                 perf_event_comm_ctx(ctx, comm_event);
3414         rcu_read_unlock();
3415 }
3416
3417 void perf_event_comm(struct task_struct *task)
3418 {
3419         struct perf_comm_event comm_event;
3420
3421         if (task->perf_event_ctxp)
3422                 perf_event_enable_on_exec(task);
3423
3424         if (!atomic_read(&nr_comm_events))
3425                 return;
3426
3427         comm_event = (struct perf_comm_event){
3428                 .task   = task,
3429                 /* .comm      */
3430                 /* .comm_size */
3431                 .event_id  = {
3432                         .header = {
3433                                 .type = PERF_RECORD_COMM,
3434                                 .misc = 0,
3435                                 /* .size */
3436                         },
3437                         /* .pid */
3438                         /* .tid */
3439                 },
3440         };
3441
3442         perf_event_comm_event(&comm_event);
3443 }
3444
3445 /*
3446  * mmap tracking
3447  */
3448
3449 struct perf_mmap_event {
3450         struct vm_area_struct   *vma;
3451
3452         const char              *file_name;
3453         int                     file_size;
3454
3455         struct {
3456                 struct perf_event_header        header;
3457
3458                 u32                             pid;
3459                 u32                             tid;
3460                 u64                             start;
3461                 u64                             len;
3462                 u64                             pgoff;
3463         } event_id;
3464 };
3465
3466 static void perf_event_mmap_output(struct perf_event *event,
3467                                      struct perf_mmap_event *mmap_event)
3468 {
3469         struct perf_output_handle handle;
3470         int size = mmap_event->event_id.header.size;
3471         int ret = perf_output_begin(&handle, event, size, 0, 0);
3472
3473         if (ret)
3474                 return;
3475
3476         mmap_event->event_id.pid = perf_event_pid(event, current);
3477         mmap_event->event_id.tid = perf_event_tid(event, current);
3478
3479         perf_output_put(&handle, mmap_event->event_id);
3480         perf_output_copy(&handle, mmap_event->file_name,
3481                                    mmap_event->file_size);
3482         perf_output_end(&handle);
3483 }
3484
3485 static int perf_event_mmap_match(struct perf_event *event,
3486                                    struct perf_mmap_event *mmap_event)
3487 {
3488         if (event->attr.mmap)
3489                 return 1;
3490
3491         return 0;
3492 }
3493
3494 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3495                                   struct perf_mmap_event *mmap_event)
3496 {
3497         struct perf_event *event;
3498
3499         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3500                 if (perf_event_mmap_match(event, mmap_event))
3501                         perf_event_mmap_output(event, mmap_event);
3502         }
3503 }
3504
3505 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3506 {
3507         struct perf_cpu_context *cpuctx;
3508         struct perf_event_context *ctx;
3509         struct vm_area_struct *vma = mmap_event->vma;
3510         struct file *file = vma->vm_file;
3511         unsigned int size;
3512         char tmp[16];
3513         char *buf = NULL;
3514         const char *name;
3515
3516         memset(tmp, 0, sizeof(tmp));
3517
3518         if (file) {
3519                 /*
3520                  * d_path works from the end of the buffer backwards, so we
3521                  * need to add enough zero bytes after the string to handle
3522                  * the 64bit alignment we do later.
3523                  */
3524                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3525                 if (!buf) {
3526                         name = strncpy(tmp, "//enomem", sizeof(tmp));
3527                         goto got_name;
3528                 }
3529                 name = d_path(&file->f_path, buf, PATH_MAX);
3530                 if (IS_ERR(name)) {
3531                         name = strncpy(tmp, "//toolong", sizeof(tmp));
3532                         goto got_name;
3533                 }
3534         } else {
3535                 if (arch_vma_name(mmap_event->vma)) {
3536                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3537                                        sizeof(tmp));
3538                         goto got_name;
3539                 }
3540
3541                 if (!vma->vm_mm) {
3542                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
3543                         goto got_name;
3544                 }
3545
3546                 name = strncpy(tmp, "//anon", sizeof(tmp));
3547                 goto got_name;
3548         }
3549
3550 got_name:
3551         size = ALIGN(strlen(name)+1, sizeof(u64));
3552
3553         mmap_event->file_name = name;
3554         mmap_event->file_size = size;
3555
3556         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3557
3558         rcu_read_lock();
3559         cpuctx = &get_cpu_var(perf_cpu_context);
3560         perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3561         put_cpu_var(perf_cpu_context);
3562
3563         /*
3564          * doesn't really matter which of the child contexts the
3565          * events ends up in.
3566          */
3567         ctx = rcu_dereference(current->perf_event_ctxp);
3568         if (ctx)
3569                 perf_event_mmap_ctx(ctx, mmap_event);
3570         rcu_read_unlock();
3571
3572         kfree(buf);
3573 }
3574
3575 void __perf_event_mmap(struct vm_area_struct *vma)
3576 {
3577         struct perf_mmap_event mmap_event;
3578
3579         if (!atomic_read(&nr_mmap_events))
3580                 return;
3581
3582         mmap_event = (struct perf_mmap_event){
3583                 .vma    = vma,
3584                 /* .file_name */
3585                 /* .file_size */
3586                 .event_id  = {
3587                         .header = {
3588                                 .type = PERF_RECORD_MMAP,
3589                                 .misc = 0,
3590                                 /* .size */
3591                         },
3592                         /* .pid */
3593                         /* .tid */
3594                         .start  = vma->vm_start,
3595                         .len    = vma->vm_end - vma->vm_start,
3596                         .pgoff  = vma->vm_pgoff,
3597                 },
3598         };
3599
3600         perf_event_mmap_event(&mmap_event);
3601 }
3602
3603 /*
3604  * IRQ throttle logging
3605  */
3606
3607 static void perf_log_throttle(struct perf_event *event, int enable)
3608 {
3609         struct perf_output_handle handle;
3610         int ret;
3611
3612         struct {
3613                 struct perf_event_header        header;
3614                 u64                             time;
3615                 u64                             id;
3616                 u64                             stream_id;
3617         } throttle_event = {
3618                 .header = {
3619                         .type = PERF_RECORD_THROTTLE,
3620                         .misc = 0,
3621                         .size = sizeof(throttle_event),
3622                 },
3623                 .time           = perf_clock(),
3624                 .id             = primary_event_id(event),
3625                 .stream_id      = event->id,
3626         };
3627
3628         if (enable)
3629                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3630
3631         ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3632         if (ret)
3633                 return;
3634
3635         perf_output_put(&handle, throttle_event);
3636         perf_output_end(&handle);
3637 }
3638
3639 /*
3640  * Generic event overflow handling, sampling.
3641  */
3642
3643 static int __perf_event_overflow(struct perf_event *event, int nmi,
3644                                    int throttle, struct perf_sample_data *data,
3645                                    struct pt_regs *regs)
3646 {
3647         int events = atomic_read(&event->event_limit);
3648         struct hw_perf_event *hwc = &event->hw;
3649         int ret = 0;
3650
3651         throttle = (throttle && event->pmu->unthrottle != NULL);
3652
3653         if (!throttle) {
3654                 hwc->interrupts++;
3655         } else {
3656                 if (hwc->interrupts != MAX_INTERRUPTS) {
3657                         hwc->interrupts++;
3658                         if (HZ * hwc->interrupts >
3659                                         (u64)sysctl_perf_event_sample_rate) {
3660                                 hwc->interrupts = MAX_INTERRUPTS;
3661                                 perf_log_throttle(event, 0);
3662                                 ret = 1;
3663                         }
3664                 } else {
3665                         /*
3666                          * Keep re-disabling events even though on the previous
3667                          * pass we disabled it - just in case we raced with a
3668                          * sched-in and the event got enabled again:
3669                          */
3670                         ret = 1;
3671                 }
3672         }
3673
3674         if (event->attr.freq) {
3675                 u64 now = perf_clock();
3676                 s64 delta = now - hwc->freq_stamp;
3677
3678                 hwc->freq_stamp = now;
3679
3680                 if (delta > 0 && delta < TICK_NSEC)
3681                         perf_adjust_period(event, NSEC_PER_SEC / (int)delta);
3682         }
3683
3684         /*
3685          * XXX event_limit might not quite work as expected on inherited
3686          * events
3687          */
3688
3689         event->pending_kill = POLL_IN;
3690         if (events && atomic_dec_and_test(&event->event_limit)) {
3691                 ret = 1;
3692                 event->pending_kill = POLL_HUP;
3693                 if (nmi) {
3694                         event->pending_disable = 1;
3695                         perf_pending_queue(&event->pending,
3696                                            perf_pending_event);
3697                 } else
3698                         perf_event_disable(event);
3699         }
3700
3701         if (event->overflow_handler)
3702                 event->overflow_handler(event, nmi, data, regs);
3703         else
3704                 perf_event_output(event, nmi, data, regs);
3705
3706         return ret;
3707 }
3708
3709 int perf_event_overflow(struct perf_event *event, int nmi,
3710                           struct perf_sample_data *data,
3711                           struct pt_regs *regs)
3712 {
3713         return __perf_event_overflow(event, nmi, 1, data, regs);
3714 }
3715
3716 /*
3717  * Generic software event infrastructure
3718  */
3719
3720 /*
3721  * We directly increment event->count and keep a second value in
3722  * event->hw.period_left to count intervals. This period event
3723  * is kept in the range [-sample_period, 0] so that we can use the
3724  * sign as trigger.
3725  */
3726
3727 static u64 perf_swevent_set_period(struct perf_event *event)
3728 {
3729         struct hw_perf_event *hwc = &event->hw;
3730         u64 period = hwc->last_period;
3731         u64 nr, offset;
3732         s64 old, val;
3733
3734         hwc->last_period = hwc->sample_period;
3735
3736 again:
3737         old = val = atomic64_read(&hwc->period_left);
3738         if (val < 0)
3739                 return 0;
3740
3741         nr = div64_u64(period + val, period);
3742         offset = nr * period;
3743         val -= offset;
3744         if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3745                 goto again;
3746
3747         return nr;
3748 }
3749
3750 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3751                                     int nmi, struct perf_sample_data *data,
3752                                     struct pt_regs *regs)
3753 {
3754         struct hw_perf_event *hwc = &event->hw;
3755         int throttle = 0;
3756
3757         data->period = event->hw.last_period;
3758         if (!overflow)
3759                 overflow = perf_swevent_set_period(event);
3760
3761         if (hwc->interrupts == MAX_INTERRUPTS)
3762                 return;
3763
3764         for (; overflow; overflow--) {
3765                 if (__perf_event_overflow(event, nmi, throttle,
3766                                             data, regs)) {
3767                         /*
3768                          * We inhibit the overflow from happening when
3769                          * hwc->interrupts == MAX_INTERRUPTS.
3770                          */
3771                         break;
3772                 }
3773                 throttle = 1;
3774         }
3775 }
3776
3777 static void perf_swevent_unthrottle(struct perf_event *event)
3778 {
3779         /*
3780          * Nothing to do, we already reset hwc->interrupts.
3781          */
3782 }
3783
3784 static void perf_swevent_add(struct perf_event *event, u64 nr,
3785                                int nmi, struct perf_sample_data *data,
3786                                struct pt_regs *regs)
3787 {
3788         struct hw_perf_event *hwc = &event->hw;
3789
3790         atomic64_add(nr, &event->count);
3791
3792         if (!regs)
3793                 return;
3794
3795         if (!hwc->sample_period)
3796                 return;
3797
3798         if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3799                 return perf_swevent_overflow(event, 1, nmi, data, regs);
3800
3801         if (atomic64_add_negative(nr, &hwc->period_left))
3802                 return;
3803
3804         perf_swevent_overflow(event, 0, nmi, data, regs);
3805 }
3806
3807 static int perf_swevent_is_counting(struct perf_event *event)
3808 {
3809         /*
3810          * The event is active, we're good!
3811          */
3812         if (event->state == PERF_EVENT_STATE_ACTIVE)
3813                 return 1;
3814
3815         /*
3816          * The event is off/error, not counting.
3817          */
3818         if (event->state != PERF_EVENT_STATE_INACTIVE)
3819                 return 0;
3820
3821         /*
3822          * The event is inactive, if the context is active
3823          * we're part of a group that didn't make it on the 'pmu',
3824          * not counting.
3825          */
3826         if (event->ctx->is_active)
3827                 return 0;
3828
3829         /*
3830          * We're inactive and the context is too, this means the
3831          * task is scheduled out, we're counting events that happen
3832          * to us, like migration events.
3833          */
3834         return 1;
3835 }
3836
3837 static int perf_tp_event_match(struct perf_event *event,
3838                                 struct perf_sample_data *data);
3839
3840 static int perf_swevent_match(struct perf_event *event,
3841                                 enum perf_type_id type,
3842                                 u32 event_id,
3843                                 struct perf_sample_data *data,
3844                                 struct pt_regs *regs)
3845 {
3846         if (!perf_swevent_is_counting(event))
3847                 return 0;
3848
3849         if (event->attr.type != type)
3850                 return 0;
3851         if (event->attr.config != event_id)
3852                 return 0;
3853
3854         if (regs) {
3855                 if (event->attr.exclude_user && user_mode(regs))
3856                         return 0;
3857
3858                 if (event->attr.exclude_kernel && !user_mode(regs))
3859                         return 0;
3860         }
3861
3862         if (event->attr.type == PERF_TYPE_TRACEPOINT &&
3863             !perf_tp_event_match(event, data))
3864                 return 0;
3865
3866         return 1;
3867 }
3868
3869 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
3870                                      enum perf_type_id type,
3871                                      u32 event_id, u64 nr, int nmi,
3872                                      struct perf_sample_data *data,
3873                                      struct pt_regs *regs)
3874 {
3875         struct perf_event *event;
3876
3877         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3878                 if (perf_swevent_match(event, type, event_id, data, regs))
3879                         perf_swevent_add(event, nr, nmi, data, regs);
3880         }
3881 }
3882
3883 static int *perf_swevent_recursion_context(struct perf_cpu_context *cpuctx)
3884 {
3885         if (in_nmi())
3886                 return &cpuctx->recursion[3];
3887
3888         if (in_irq())
3889                 return &cpuctx->recursion[2];
3890
3891         if (in_softirq())
3892                 return &cpuctx->recursion[1];
3893
3894         return &cpuctx->recursion[0];
3895 }
3896
3897 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
3898                                     u64 nr, int nmi,
3899                                     struct perf_sample_data *data,
3900                                     struct pt_regs *regs)
3901 {
3902         struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3903         int *recursion = perf_swevent_recursion_context(cpuctx);
3904         struct perf_event_context *ctx;
3905
3906         if (*recursion)
3907                 goto out;
3908
3909         (*recursion)++;
3910         barrier();
3911
3912         rcu_read_lock();
3913         perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
3914                                  nr, nmi, data, regs);
3915         /*
3916          * doesn't really matter which of the child contexts the
3917          * events ends up in.
3918          */
3919         ctx = rcu_dereference(current->perf_event_ctxp);
3920         if (ctx)
3921                 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
3922         rcu_read_unlock();
3923
3924         barrier();
3925         (*recursion)--;
3926
3927 out:
3928         put_cpu_var(perf_cpu_context);
3929 }
3930
3931 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
3932                             struct pt_regs *regs, u64 addr)
3933 {
3934         struct perf_sample_data data = {
3935                 .addr = addr,
3936         };
3937
3938         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi,
3939                                 &data, regs);
3940 }
3941
3942 static void perf_swevent_read(struct perf_event *event)
3943 {
3944 }
3945
3946 static int perf_swevent_enable(struct perf_event *event)
3947 {
3948         struct hw_perf_event *hwc = &event->hw;
3949
3950         if (hwc->sample_period) {
3951                 hwc->last_period = hwc->sample_period;
3952                 perf_swevent_set_period(event);
3953         }
3954         return 0;
3955 }
3956
3957 static void perf_swevent_disable(struct perf_event *event)
3958 {
3959 }
3960
3961 static const struct pmu perf_ops_generic = {
3962         .enable         = perf_swevent_enable,
3963         .disable        = perf_swevent_disable,
3964         .read           = perf_swevent_read,
3965         .unthrottle     = perf_swevent_unthrottle,
3966 };
3967
3968 /*
3969  * hrtimer based swevent callback
3970  */
3971
3972 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
3973 {
3974         enum hrtimer_restart ret = HRTIMER_RESTART;
3975         struct perf_sample_data data;
3976         struct pt_regs *regs;
3977         struct perf_event *event;
3978         u64 period;
3979
3980         event   = container_of(hrtimer, struct perf_event, hw.hrtimer);
3981         event->pmu->read(event);
3982
3983         data.addr = 0;
3984         regs = get_irq_regs();
3985         /*
3986          * In case we exclude kernel IPs or are somehow not in interrupt
3987          * context, provide the next best thing, the user IP.
3988          */
3989         if ((event->attr.exclude_kernel || !regs) &&
3990                         !event->attr.exclude_user)
3991                 regs = task_pt_regs(current);
3992
3993         if (regs) {
3994                 if (!(event->attr.exclude_idle && current->pid == 0))
3995                         if (perf_event_overflow(event, 0, &data, regs))
3996                                 ret = HRTIMER_NORESTART;
3997         }
3998
3999         period = max_t(u64, 10000, event->hw.sample_period);
4000         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4001
4002         return ret;
4003 }
4004
4005 static void perf_swevent_start_hrtimer(struct perf_event *event)
4006 {
4007         struct hw_perf_event *hwc = &event->hw;
4008
4009         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4010         hwc->hrtimer.function = perf_swevent_hrtimer;
4011         if (hwc->sample_period) {
4012                 u64 period;
4013
4014                 if (hwc->remaining) {
4015                         if (hwc->remaining < 0)
4016                                 period = 10000;
4017                         else
4018                                 period = hwc->remaining;
4019                         hwc->remaining = 0;
4020                 } else {
4021                         period = max_t(u64, 10000, hwc->sample_period);
4022                 }
4023                 __hrtimer_start_range_ns(&hwc->hrtimer,
4024                                 ns_to_ktime(period), 0,
4025                                 HRTIMER_MODE_REL, 0);
4026         }
4027 }
4028
4029 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4030 {
4031         struct hw_perf_event *hwc = &event->hw;
4032
4033         if (hwc->sample_period) {
4034                 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4035                 hwc->remaining = ktime_to_ns(remaining);
4036
4037                 hrtimer_cancel(&hwc->hrtimer);
4038         }
4039 }
4040
4041 /*
4042  * Software event: cpu wall time clock
4043  */
4044
4045 static void cpu_clock_perf_event_update(struct perf_event *event)
4046 {
4047         int cpu = raw_smp_processor_id();
4048         s64 prev;
4049         u64 now;
4050
4051         now = cpu_clock(cpu);
4052         prev = atomic64_read(&event->hw.prev_count);
4053         atomic64_set(&event->hw.prev_count, now);
4054         atomic64_add(now - prev, &event->count);
4055 }
4056
4057 static int cpu_clock_perf_event_enable(struct perf_event *event)
4058 {
4059         struct hw_perf_event *hwc = &event->hw;
4060         int cpu = raw_smp_processor_id();
4061
4062         atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4063         perf_swevent_start_hrtimer(event);
4064
4065         return 0;
4066 }
4067
4068 static void cpu_clock_perf_event_disable(struct perf_event *event)
4069 {
4070         perf_swevent_cancel_hrtimer(event);
4071         cpu_clock_perf_event_update(event);
4072 }
4073
4074 static void cpu_clock_perf_event_read(struct perf_event *event)
4075 {
4076         cpu_clock_perf_event_update(event);
4077 }
4078
4079 static const struct pmu perf_ops_cpu_clock = {
4080         .enable         = cpu_clock_perf_event_enable,
4081         .disable        = cpu_clock_perf_event_disable,
4082         .read           = cpu_clock_perf_event_read,
4083 };
4084
4085 /*
4086  * Software event: task time clock
4087  */
4088
4089 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4090 {
4091         u64 prev;
4092         s64 delta;
4093
4094         prev = atomic64_xchg(&event->hw.prev_count, now);
4095         delta = now - prev;
4096         atomic64_add(delta, &event->count);
4097 }
4098
4099 static int task_clock_perf_event_enable(struct perf_event *event)
4100 {
4101         struct hw_perf_event *hwc = &event->hw;
4102         u64 now;
4103
4104         now = event->ctx->time;
4105
4106         atomic64_set(&hwc->prev_count, now);
4107
4108         perf_swevent_start_hrtimer(event);
4109
4110         return 0;
4111 }
4112
4113 static void task_clock_perf_event_disable(struct perf_event *event)
4114 {
4115         perf_swevent_cancel_hrtimer(event);
4116         task_clock_perf_event_update(event, event->ctx->time);
4117
4118 }
4119
4120 static void task_clock_perf_event_read(struct perf_event *event)
4121 {
4122         u64 time;
4123
4124         if (!in_nmi()) {
4125                 update_context_time(event->ctx);
4126                 time = event->ctx->time;
4127         } else {
4128                 u64 now = perf_clock();
4129                 u64 delta = now - event->ctx->timestamp;
4130                 time = event->ctx->time + delta;
4131         }
4132
4133         task_clock_perf_event_update(event, time);
4134 }
4135
4136 static const struct pmu perf_ops_task_clock = {
4137         .enable         = task_clock_perf_event_enable,
4138         .disable        = task_clock_perf_event_disable,
4139         .read           = task_clock_perf_event_read,
4140 };
4141
4142 #ifdef CONFIG_EVENT_PROFILE
4143
4144 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4145                           int entry_size)
4146 {
4147         struct perf_raw_record raw = {
4148                 .size = entry_size,
4149                 .data = record,
4150         };
4151
4152         struct perf_sample_data data = {
4153                 .addr = addr,
4154                 .raw = &raw,
4155         };
4156
4157         struct pt_regs *regs = get_irq_regs();
4158
4159         if (!regs)
4160                 regs = task_pt_regs(current);
4161
4162         do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4163                                 &data, regs);
4164 }
4165 EXPORT_SYMBOL_GPL(perf_tp_event);
4166
4167 static int perf_tp_event_match(struct perf_event *event,
4168                                 struct perf_sample_data *data)
4169 {
4170         void *record = data->raw->data;
4171
4172         if (likely(!event->filter) || filter_match_preds(event->filter, record))
4173                 return 1;
4174         return 0;
4175 }
4176
4177 static void tp_perf_event_destroy(struct perf_event *event)
4178 {
4179         ftrace_profile_disable(event->attr.config);
4180 }
4181
4182 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4183 {
4184         /*
4185          * Raw tracepoint data is a severe data leak, only allow root to
4186          * have these.
4187          */
4188         if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4189                         perf_paranoid_tracepoint_raw() &&
4190                         !capable(CAP_SYS_ADMIN))
4191                 return ERR_PTR(-EPERM);
4192
4193         if (ftrace_profile_enable(event->attr.config))
4194                 return NULL;
4195
4196         event->destroy = tp_perf_event_destroy;
4197
4198         return &perf_ops_generic;
4199 }
4200
4201 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4202 {
4203         char *filter_str;
4204         int ret;
4205
4206         if (event->attr.type != PERF_TYPE_TRACEPOINT)
4207                 return -EINVAL;
4208
4209         filter_str = strndup_user(arg, PAGE_SIZE);
4210         if (IS_ERR(filter_str))
4211                 return PTR_ERR(filter_str);
4212
4213         ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4214
4215         kfree(filter_str);
4216         return ret;
4217 }
4218
4219 static void perf_event_free_filter(struct perf_event *event)
4220 {
4221         ftrace_profile_free_filter(event);
4222 }
4223
4224 #else
4225
4226 static int perf_tp_event_match(struct perf_event *event,
4227                                 struct perf_sample_data *data)
4228 {
4229         return 1;
4230 }
4231
4232 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4233 {
4234         return NULL;
4235 }
4236
4237 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4238 {
4239         return -ENOENT;
4240 }
4241
4242 static void perf_event_free_filter(struct perf_event *event)
4243 {
4244 }
4245
4246 #endif /* CONFIG_EVENT_PROFILE */
4247
4248 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4249 static void bp_perf_event_destroy(struct perf_event *event)
4250 {
4251         release_bp_slot(event);
4252 }
4253
4254 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4255 {
4256         int err;
4257         /*
4258          * The breakpoint is already filled if we haven't created the counter
4259          * through perf syscall
4260          * FIXME: manage to get trigerred to NULL if it comes from syscalls
4261          */
4262         if (!bp->callback)
4263                 err = register_perf_hw_breakpoint(bp);
4264         else
4265                 err = __register_perf_hw_breakpoint(bp);
4266         if (err)
4267                 return ERR_PTR(err);
4268
4269         bp->destroy = bp_perf_event_destroy;
4270
4271         return &perf_ops_bp;
4272 }
4273
4274 void perf_bp_event(struct perf_event *bp, void *regs)
4275 {
4276         /* TODO */
4277 }
4278 #else
4279 static void bp_perf_event_destroy(struct perf_event *event)
4280 {
4281 }
4282
4283 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4284 {
4285         return NULL;
4286 }
4287
4288 void perf_bp_event(struct perf_event *bp, void *regs)
4289 {
4290 }
4291 #endif
4292
4293 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4294
4295 static void sw_perf_event_destroy(struct perf_event *event)
4296 {
4297         u64 event_id = event->attr.config;
4298
4299         WARN_ON(event->parent);
4300
4301         atomic_dec(&perf_swevent_enabled[event_id]);
4302 }
4303
4304 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4305 {
4306         const struct pmu *pmu = NULL;
4307         u64 event_id = event->attr.config;
4308
4309         /*
4310          * Software events (currently) can't in general distinguish
4311          * between user, kernel and hypervisor events.
4312          * However, context switches and cpu migrations are considered
4313          * to be kernel events, and page faults are never hypervisor
4314          * events.
4315          */
4316         switch (event_id) {
4317         case PERF_COUNT_SW_CPU_CLOCK:
4318                 pmu = &perf_ops_cpu_clock;
4319
4320                 break;
4321         case PERF_COUNT_SW_TASK_CLOCK:
4322                 /*
4323                  * If the user instantiates this as a per-cpu event,
4324                  * use the cpu_clock event instead.
4325                  */
4326                 if (event->ctx->task)
4327                         pmu = &perf_ops_task_clock;
4328                 else
4329                         pmu = &perf_ops_cpu_clock;
4330
4331                 break;
4332         case PERF_COUNT_SW_PAGE_FAULTS:
4333         case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4334         case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4335         case PERF_COUNT_SW_CONTEXT_SWITCHES:
4336         case PERF_COUNT_SW_CPU_MIGRATIONS:
4337         case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4338         case PERF_COUNT_SW_EMULATION_FAULTS:
4339                 if (!event->parent) {
4340                         atomic_inc(&perf_swevent_enabled[event_id]);
4341                         event->destroy = sw_perf_event_destroy;
4342                 }
4343                 pmu = &perf_ops_generic;
4344                 break;
4345         }
4346
4347         return pmu;
4348 }
4349
4350 /*
4351  * Allocate and initialize a event structure
4352  */
4353 static struct perf_event *
4354 perf_event_alloc(struct perf_event_attr *attr,
4355                    int cpu,
4356                    struct perf_event_context *ctx,
4357                    struct perf_event *group_leader,
4358                    struct perf_event *parent_event,
4359                    perf_callback_t callback,
4360                    gfp_t gfpflags)
4361 {
4362         const struct pmu *pmu;
4363         struct perf_event *event;
4364         struct hw_perf_event *hwc;
4365         long err;
4366
4367         event = kzalloc(sizeof(*event), gfpflags);
4368         if (!event)
4369                 return ERR_PTR(-ENOMEM);
4370
4371         /*
4372          * Single events are their own group leaders, with an
4373          * empty sibling list:
4374          */
4375         if (!group_leader)
4376                 group_leader = event;
4377
4378         mutex_init(&event->child_mutex);
4379         INIT_LIST_HEAD(&event->child_list);
4380
4381         INIT_LIST_HEAD(&event->group_entry);
4382         INIT_LIST_HEAD(&event->event_entry);
4383         INIT_LIST_HEAD(&event->sibling_list);
4384         init_waitqueue_head(&event->waitq);
4385
4386         mutex_init(&event->mmap_mutex);
4387
4388         event->cpu              = cpu;
4389         event->attr             = *attr;
4390         event->group_leader     = group_leader;
4391         event->pmu              = NULL;
4392         event->ctx              = ctx;
4393         event->oncpu            = -1;
4394
4395         event->parent           = parent_event;
4396
4397         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
4398         event->id               = atomic64_inc_return(&perf_event_id);
4399
4400         event->state            = PERF_EVENT_STATE_INACTIVE;
4401
4402         if (!callback && parent_event)
4403                 callback = parent_event->callback;
4404         
4405         event->callback = callback;
4406
4407         if (attr->disabled)
4408                 event->state = PERF_EVENT_STATE_OFF;
4409
4410         pmu = NULL;
4411
4412         hwc = &event->hw;
4413         hwc->sample_period = attr->sample_period;
4414         if (attr->freq && attr->sample_freq)
4415                 hwc->sample_period = 1;
4416         hwc->last_period = hwc->sample_period;
4417
4418         atomic64_set(&hwc->period_left, hwc->sample_period);
4419
4420         /*
4421          * we currently do not support PERF_FORMAT_GROUP on inherited events
4422          */
4423         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4424                 goto done;
4425
4426         switch (attr->type) {
4427         case PERF_TYPE_RAW:
4428         case PERF_TYPE_HARDWARE:
4429         case PERF_TYPE_HW_CACHE:
4430                 pmu = hw_perf_event_init(event);
4431                 break;
4432
4433         case PERF_TYPE_SOFTWARE:
4434                 pmu = sw_perf_event_init(event);
4435                 break;
4436
4437         case PERF_TYPE_TRACEPOINT:
4438                 pmu = tp_perf_event_init(event);
4439                 break;
4440
4441         case PERF_TYPE_BREAKPOINT:
4442                 pmu = bp_perf_event_init(event);
4443                 break;
4444
4445
4446         default:
4447                 break;
4448         }
4449 done:
4450         err = 0;
4451         if (!pmu)
4452                 err = -EINVAL;
4453         else if (IS_ERR(pmu))
4454                 err = PTR_ERR(pmu);
4455
4456         if (err) {
4457                 if (event->ns)
4458                         put_pid_ns(event->ns);
4459                 kfree(event);
4460                 return ERR_PTR(err);
4461         }
4462
4463         event->pmu = pmu;
4464
4465         if (!event->parent) {
4466                 atomic_inc(&nr_events);
4467                 if (event->attr.mmap)
4468                         atomic_inc(&nr_mmap_events);
4469                 if (event->attr.comm)
4470                         atomic_inc(&nr_comm_events);
4471                 if (event->attr.task)
4472                         atomic_inc(&nr_task_events);
4473         }
4474
4475         return event;
4476 }
4477
4478 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4479                           struct perf_event_attr *attr)
4480 {
4481         u32 size;
4482         int ret;
4483
4484         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4485                 return -EFAULT;
4486
4487         /*
4488          * zero the full structure, so that a short copy will be nice.
4489          */
4490         memset(attr, 0, sizeof(*attr));
4491
4492         ret = get_user(size, &uattr->size);
4493         if (ret)
4494                 return ret;
4495
4496         if (size > PAGE_SIZE)   /* silly large */
4497                 goto err_size;
4498
4499         if (!size)              /* abi compat */
4500                 size = PERF_ATTR_SIZE_VER0;
4501
4502         if (size < PERF_ATTR_SIZE_VER0)
4503                 goto err_size;
4504
4505         /*
4506          * If we're handed a bigger struct than we know of,
4507          * ensure all the unknown bits are 0 - i.e. new
4508          * user-space does not rely on any kernel feature
4509          * extensions we dont know about yet.
4510          */
4511         if (size > sizeof(*attr)) {
4512                 unsigned char __user *addr;
4513                 unsigned char __user *end;
4514                 unsigned char val;
4515
4516                 addr = (void __user *)uattr + sizeof(*attr);
4517                 end  = (void __user *)uattr + size;
4518
4519                 for (; addr < end; addr++) {
4520                         ret = get_user(val, addr);
4521                         if (ret)
4522                                 return ret;
4523                         if (val)
4524                                 goto err_size;
4525                 }
4526                 size = sizeof(*attr);
4527         }
4528
4529         ret = copy_from_user(attr, uattr, size);
4530         if (ret)
4531                 return -EFAULT;
4532
4533         /*
4534          * If the type exists, the corresponding creation will verify
4535          * the attr->config.
4536          */
4537         if (attr->type >= PERF_TYPE_MAX)
4538                 return -EINVAL;
4539
4540         if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3)
4541                 return -EINVAL;
4542
4543         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4544                 return -EINVAL;
4545
4546         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4547                 return -EINVAL;
4548
4549 out:
4550         return ret;
4551
4552 err_size:
4553         put_user(sizeof(*attr), &uattr->size);
4554         ret = -E2BIG;
4555         goto out;
4556 }
4557
4558 static int perf_event_set_output(struct perf_event *event, int output_fd)
4559 {
4560         struct perf_event *output_event = NULL;
4561         struct file *output_file = NULL;
4562         struct perf_event *old_output;
4563         int fput_needed = 0;
4564         int ret = -EINVAL;
4565
4566         if (!output_fd)
4567                 goto set;
4568
4569         output_file = fget_light(output_fd, &fput_needed);
4570         if (!output_file)
4571                 return -EBADF;
4572
4573         if (output_file->f_op != &perf_fops)
4574                 goto out;
4575
4576         output_event = output_file->private_data;
4577
4578         /* Don't chain output fds */
4579         if (output_event->output)
4580                 goto out;
4581
4582         /* Don't set an output fd when we already have an output channel */
4583         if (event->data)
4584                 goto out;
4585
4586         atomic_long_inc(&output_file->f_count);
4587
4588 set:
4589         mutex_lock(&event->mmap_mutex);
4590         old_output = event->output;
4591         rcu_assign_pointer(event->output, output_event);
4592         mutex_unlock(&event->mmap_mutex);
4593
4594         if (old_output) {
4595                 /*
4596                  * we need to make sure no existing perf_output_*()
4597                  * is still referencing this event.
4598                  */
4599                 synchronize_rcu();
4600                 fput(old_output->filp);
4601         }
4602
4603         ret = 0;
4604 out:
4605         fput_light(output_file, fput_needed);
4606         return ret;
4607 }
4608
4609 /**
4610  * sys_perf_event_open - open a performance event, associate it to a task/cpu
4611  *
4612  * @attr_uptr:  event_id type attributes for monitoring/sampling
4613  * @pid:                target pid
4614  * @cpu:                target cpu
4615  * @group_fd:           group leader event fd
4616  */
4617 SYSCALL_DEFINE5(perf_event_open,
4618                 struct perf_event_attr __user *, attr_uptr,
4619                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4620 {
4621         struct perf_event *event, *group_leader;
4622         struct perf_event_attr attr;
4623         struct perf_event_context *ctx;
4624         struct file *event_file = NULL;
4625         struct file *group_file = NULL;
4626         int fput_needed = 0;
4627         int fput_needed2 = 0;
4628         int err;
4629
4630         /* for future expandability... */
4631         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4632                 return -EINVAL;
4633
4634         err = perf_copy_attr(attr_uptr, &attr);
4635         if (err)
4636                 return err;
4637
4638         if (!attr.exclude_kernel) {
4639                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4640                         return -EACCES;
4641         }
4642
4643         if (attr.freq) {
4644                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4645                         return -EINVAL;
4646         }
4647
4648         /*
4649          * Get the target context (task or percpu):
4650          */
4651         ctx = find_get_context(pid, cpu);
4652         if (IS_ERR(ctx))
4653                 return PTR_ERR(ctx);
4654
4655         /*
4656          * Look up the group leader (we will attach this event to it):
4657          */
4658         group_leader = NULL;
4659         if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4660                 err = -EINVAL;
4661                 group_file = fget_light(group_fd, &fput_needed);
4662                 if (!group_file)
4663                         goto err_put_context;
4664                 if (group_file->f_op != &perf_fops)
4665                         goto err_put_context;
4666
4667                 group_leader = group_file->private_data;
4668                 /*
4669                  * Do not allow a recursive hierarchy (this new sibling
4670                  * becoming part of another group-sibling):
4671                  */
4672                 if (group_leader->group_leader != group_leader)
4673                         goto err_put_context;
4674                 /*
4675                  * Do not allow to attach to a group in a different
4676                  * task or CPU context:
4677                  */
4678                 if (group_leader->ctx != ctx)
4679                         goto err_put_context;
4680                 /*
4681                  * Only a group leader can be exclusive or pinned
4682                  */
4683                 if (attr.exclusive || attr.pinned)
4684                         goto err_put_context;
4685         }
4686
4687         event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4688                                      NULL, NULL, GFP_KERNEL);
4689         err = PTR_ERR(event);
4690         if (IS_ERR(event))
4691                 goto err_put_context;
4692
4693         err = anon_inode_getfd("[perf_event]", &perf_fops, event, 0);
4694         if (err < 0)
4695                 goto err_free_put_context;
4696
4697         event_file = fget_light(err, &fput_needed2);
4698         if (!event_file)
4699                 goto err_free_put_context;
4700
4701         if (flags & PERF_FLAG_FD_OUTPUT) {
4702                 err = perf_event_set_output(event, group_fd);
4703                 if (err)
4704                         goto err_fput_free_put_context;
4705         }
4706
4707         event->filp = event_file;
4708         WARN_ON_ONCE(ctx->parent_ctx);
4709         mutex_lock(&ctx->mutex);
4710         perf_install_in_context(ctx, event, cpu);
4711         ++ctx->generation;
4712         mutex_unlock(&ctx->mutex);
4713
4714         event->owner = current;
4715         get_task_struct(current);
4716         mutex_lock(&current->perf_event_mutex);
4717         list_add_tail(&event->owner_entry, &current->perf_event_list);
4718         mutex_unlock(&current->perf_event_mutex);
4719
4720 err_fput_free_put_context:
4721         fput_light(event_file, fput_needed2);
4722
4723 err_free_put_context:
4724         if (err < 0)
4725                 kfree(event);
4726
4727 err_put_context:
4728         if (err < 0)
4729                 put_ctx(ctx);
4730
4731         fput_light(group_file, fput_needed);
4732
4733         return err;
4734 }
4735
4736 /**
4737  * perf_event_create_kernel_counter
4738  *
4739  * @attr: attributes of the counter to create
4740  * @cpu: cpu in which the counter is bound
4741  * @pid: task to profile
4742  */
4743 struct perf_event *
4744 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4745                                  pid_t pid, perf_callback_t callback)
4746 {
4747         struct perf_event *event;
4748         struct perf_event_context *ctx;
4749         int err;
4750
4751         /*
4752          * Get the target context (task or percpu):
4753          */
4754
4755         ctx = find_get_context(pid, cpu);
4756         if (IS_ERR(ctx))
4757                 return NULL;
4758
4759         event = perf_event_alloc(attr, cpu, ctx, NULL,
4760                                      NULL, callback, GFP_KERNEL);
4761         err = PTR_ERR(event);
4762         if (IS_ERR(event))
4763                 goto err_put_context;
4764
4765         event->filp = NULL;
4766         WARN_ON_ONCE(ctx->parent_ctx);
4767         mutex_lock(&ctx->mutex);
4768         perf_install_in_context(ctx, event, cpu);
4769         ++ctx->generation;
4770         mutex_unlock(&ctx->mutex);
4771
4772         event->owner = current;
4773         get_task_struct(current);
4774         mutex_lock(&current->perf_event_mutex);
4775         list_add_tail(&event->owner_entry, &current->perf_event_list);
4776         mutex_unlock(&current->perf_event_mutex);
4777
4778         return event;
4779
4780 err_put_context:
4781         if (err < 0)
4782                 put_ctx(ctx);
4783
4784         return NULL;
4785 }
4786 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4787
4788 /*
4789  * inherit a event from parent task to child task:
4790  */
4791 static struct perf_event *
4792 inherit_event(struct perf_event *parent_event,
4793               struct task_struct *parent,
4794               struct perf_event_context *parent_ctx,
4795               struct task_struct *child,
4796               struct perf_event *group_leader,
4797               struct perf_event_context *child_ctx)
4798 {
4799         struct perf_event *child_event;
4800
4801         /*
4802          * Instead of creating recursive hierarchies of events,
4803          * we link inherited events back to the original parent,
4804          * which has a filp for sure, which we use as the reference
4805          * count:
4806          */
4807         if (parent_event->parent)
4808                 parent_event = parent_event->parent;
4809
4810         child_event = perf_event_alloc(&parent_event->attr,
4811                                            parent_event->cpu, child_ctx,
4812                                            group_leader, parent_event,
4813                                            NULL, GFP_KERNEL);
4814         if (IS_ERR(child_event))
4815                 return child_event;
4816         get_ctx(child_ctx);
4817
4818         /*
4819          * Make the child state follow the state of the parent event,
4820          * not its attr.disabled bit.  We hold the parent's mutex,
4821          * so we won't race with perf_event_{en, dis}able_family.
4822          */
4823         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
4824                 child_event->state = PERF_EVENT_STATE_INACTIVE;
4825         else
4826                 child_event->state = PERF_EVENT_STATE_OFF;
4827
4828         if (parent_event->attr.freq)
4829                 child_event->hw.sample_period = parent_event->hw.sample_period;
4830
4831         child_event->overflow_handler = parent_event->overflow_handler;
4832
4833         /*
4834          * Link it up in the child's context:
4835          */
4836         add_event_to_ctx(child_event, child_ctx);
4837
4838         /*
4839          * Get a reference to the parent filp - we will fput it
4840          * when the child event exits. This is safe to do because
4841          * we are in the parent and we know that the filp still
4842          * exists and has a nonzero count:
4843          */
4844         atomic_long_inc(&parent_event->filp->f_count);
4845
4846         /*
4847          * Link this into the parent event's child list
4848          */
4849         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4850         mutex_lock(&parent_event->child_mutex);
4851         list_add_tail(&child_event->child_list, &parent_event->child_list);
4852         mutex_unlock(&parent_event->child_mutex);
4853
4854         return child_event;
4855 }
4856
4857 static int inherit_group(struct perf_event *parent_event,
4858               struct task_struct *parent,
4859               struct perf_event_context *parent_ctx,
4860               struct task_struct *child,
4861               struct perf_event_context *child_ctx)
4862 {
4863         struct perf_event *leader;
4864         struct perf_event *sub;
4865         struct perf_event *child_ctr;
4866
4867         leader = inherit_event(parent_event, parent, parent_ctx,
4868                                  child, NULL, child_ctx);
4869         if (IS_ERR(leader))
4870                 return PTR_ERR(leader);
4871         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
4872                 child_ctr = inherit_event(sub, parent, parent_ctx,
4873                                             child, leader, child_ctx);
4874                 if (IS_ERR(child_ctr))
4875                         return PTR_ERR(child_ctr);
4876         }
4877         return 0;
4878 }
4879
4880 static void sync_child_event(struct perf_event *child_event,
4881                                struct task_struct *child)
4882 {
4883         struct perf_event *parent_event = child_event->parent;
4884         u64 child_val;
4885
4886         if (child_event->attr.inherit_stat)
4887                 perf_event_read_event(child_event, child);
4888
4889         child_val = atomic64_read(&child_event->count);
4890
4891         /*
4892          * Add back the child's count to the parent's count:
4893          */
4894         atomic64_add(child_val, &parent_event->count);
4895         atomic64_add(child_event->total_time_enabled,
4896                      &parent_event->child_total_time_enabled);
4897         atomic64_add(child_event->total_time_running,
4898                      &parent_event->child_total_time_running);
4899
4900         /*
4901          * Remove this event from the parent's list
4902          */
4903         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4904         mutex_lock(&parent_event->child_mutex);
4905         list_del_init(&child_event->child_list);
4906         mutex_unlock(&parent_event->child_mutex);
4907
4908         /*
4909          * Release the parent event, if this was the last
4910          * reference to it.
4911          */
4912         fput(parent_event->filp);
4913 }
4914
4915 static void
4916 __perf_event_exit_task(struct perf_event *child_event,
4917                          struct perf_event_context *child_ctx,
4918                          struct task_struct *child)
4919 {
4920         struct perf_event *parent_event;
4921
4922         update_event_times(child_event);
4923         perf_event_remove_from_context(child_event);
4924
4925         parent_event = child_event->parent;
4926         /*
4927          * It can happen that parent exits first, and has events
4928          * that are still around due to the child reference. These
4929          * events need to be zapped - but otherwise linger.
4930          */
4931         if (parent_event) {
4932                 sync_child_event(child_event, child);
4933                 free_event(child_event);
4934         }
4935 }
4936
4937 /*
4938  * When a child task exits, feed back event values to parent events.
4939  */
4940 void perf_event_exit_task(struct task_struct *child)
4941 {
4942         struct perf_event *child_event, *tmp;
4943         struct perf_event_context *child_ctx;
4944         unsigned long flags;
4945
4946         if (likely(!child->perf_event_ctxp)) {
4947                 perf_event_task(child, NULL, 0);
4948                 return;
4949         }
4950
4951         local_irq_save(flags);
4952         /*
4953          * We can't reschedule here because interrupts are disabled,
4954          * and either child is current or it is a task that can't be
4955          * scheduled, so we are now safe from rescheduling changing
4956          * our context.
4957          */
4958         child_ctx = child->perf_event_ctxp;
4959         __perf_event_task_sched_out(child_ctx);
4960
4961         /*
4962          * Take the context lock here so that if find_get_context is
4963          * reading child->perf_event_ctxp, we wait until it has
4964          * incremented the context's refcount before we do put_ctx below.
4965          */
4966         spin_lock(&child_ctx->lock);
4967         child->perf_event_ctxp = NULL;
4968         /*
4969          * If this context is a clone; unclone it so it can't get
4970          * swapped to another process while we're removing all
4971          * the events from it.
4972          */
4973         unclone_ctx(child_ctx);
4974         spin_unlock_irqrestore(&child_ctx->lock, flags);
4975
4976         /*
4977          * Report the task dead after unscheduling the events so that we
4978          * won't get any samples after PERF_RECORD_EXIT. We can however still
4979          * get a few PERF_RECORD_READ events.
4980          */
4981         perf_event_task(child, child_ctx, 0);
4982
4983         /*
4984          * We can recurse on the same lock type through:
4985          *
4986          *   __perf_event_exit_task()
4987          *     sync_child_event()
4988          *       fput(parent_event->filp)
4989          *         perf_release()
4990          *           mutex_lock(&ctx->mutex)
4991          *
4992          * But since its the parent context it won't be the same instance.
4993          */
4994         mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
4995
4996 again:
4997         list_for_each_entry_safe(child_event, tmp, &child_ctx->group_list,
4998                                  group_entry)
4999                 __perf_event_exit_task(child_event, child_ctx, child);
5000
5001         /*
5002          * If the last event was a group event, it will have appended all
5003          * its siblings to the list, but we obtained 'tmp' before that which
5004          * will still point to the list head terminating the iteration.
5005          */
5006         if (!list_empty(&child_ctx->group_list))
5007                 goto again;
5008
5009         mutex_unlock(&child_ctx->mutex);
5010
5011         put_ctx(child_ctx);
5012 }
5013
5014 /*
5015  * free an unexposed, unused context as created by inheritance by
5016  * init_task below, used by fork() in case of fail.
5017  */
5018 void perf_event_free_task(struct task_struct *task)
5019 {
5020         struct perf_event_context *ctx = task->perf_event_ctxp;
5021         struct perf_event *event, *tmp;
5022
5023         if (!ctx)
5024                 return;
5025
5026         mutex_lock(&ctx->mutex);
5027 again:
5028         list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry) {
5029                 struct perf_event *parent = event->parent;
5030
5031                 if (WARN_ON_ONCE(!parent))
5032                         continue;
5033
5034                 mutex_lock(&parent->child_mutex);
5035                 list_del_init(&event->child_list);
5036                 mutex_unlock(&parent->child_mutex);
5037
5038                 fput(parent->filp);
5039
5040                 list_del_event(event, ctx);
5041                 free_event(event);
5042         }
5043
5044         if (!list_empty(&ctx->group_list))
5045                 goto again;
5046
5047         mutex_unlock(&ctx->mutex);
5048
5049         put_ctx(ctx);
5050 }
5051
5052 /*
5053  * Initialize the perf_event context in task_struct
5054  */
5055 int perf_event_init_task(struct task_struct *child)
5056 {
5057         struct perf_event_context *child_ctx, *parent_ctx;
5058         struct perf_event_context *cloned_ctx;
5059         struct perf_event *event;
5060         struct task_struct *parent = current;
5061         int inherited_all = 1;
5062         int ret = 0;
5063
5064         child->perf_event_ctxp = NULL;
5065
5066         mutex_init(&child->perf_event_mutex);
5067         INIT_LIST_HEAD(&child->perf_event_list);
5068
5069         if (likely(!parent->perf_event_ctxp))
5070                 return 0;
5071
5072         /*
5073          * This is executed from the parent task context, so inherit
5074          * events that have been marked for cloning.
5075          * First allocate and initialize a context for the child.
5076          */
5077
5078         child_ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
5079         if (!child_ctx)
5080                 return -ENOMEM;
5081
5082         __perf_event_init_context(child_ctx, child);
5083         child->perf_event_ctxp = child_ctx;
5084         get_task_struct(child);
5085
5086         /*
5087          * If the parent's context is a clone, pin it so it won't get
5088          * swapped under us.
5089          */
5090         parent_ctx = perf_pin_task_context(parent);
5091
5092         /*
5093          * No need to check if parent_ctx != NULL here; since we saw
5094          * it non-NULL earlier, the only reason for it to become NULL
5095          * is if we exit, and since we're currently in the middle of
5096          * a fork we can't be exiting at the same time.
5097          */
5098
5099         /*
5100          * Lock the parent list. No need to lock the child - not PID
5101          * hashed yet and not running, so nobody can access it.
5102          */
5103         mutex_lock(&parent_ctx->mutex);
5104
5105         /*
5106          * We dont have to disable NMIs - we are only looking at
5107          * the list, not manipulating it:
5108          */
5109         list_for_each_entry(event, &parent_ctx->group_list, group_entry) {
5110
5111                 if (!event->attr.inherit) {
5112                         inherited_all = 0;
5113                         continue;
5114                 }
5115
5116                 ret = inherit_group(event, parent, parent_ctx,
5117                                              child, child_ctx);
5118                 if (ret) {
5119                         inherited_all = 0;
5120                         break;
5121                 }
5122         }
5123
5124         if (inherited_all) {
5125                 /*
5126                  * Mark the child context as a clone of the parent
5127                  * context, or of whatever the parent is a clone of.
5128                  * Note that if the parent is a clone, it could get
5129                  * uncloned at any point, but that doesn't matter
5130                  * because the list of events and the generation
5131                  * count can't have changed since we took the mutex.
5132                  */
5133                 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5134                 if (cloned_ctx) {
5135                         child_ctx->parent_ctx = cloned_ctx;
5136                         child_ctx->parent_gen = parent_ctx->parent_gen;
5137                 } else {
5138                         child_ctx->parent_ctx = parent_ctx;
5139                         child_ctx->parent_gen = parent_ctx->generation;
5140                 }
5141                 get_ctx(child_ctx->parent_ctx);
5142         }
5143
5144         mutex_unlock(&parent_ctx->mutex);
5145
5146         perf_unpin_context(parent_ctx);
5147
5148         return ret;
5149 }
5150
5151 static void __cpuinit perf_event_init_cpu(int cpu)
5152 {
5153         struct perf_cpu_context *cpuctx;
5154
5155         cpuctx = &per_cpu(perf_cpu_context, cpu);
5156         __perf_event_init_context(&cpuctx->ctx, NULL);
5157
5158         spin_lock(&perf_resource_lock);
5159         cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5160         spin_unlock(&perf_resource_lock);
5161
5162         hw_perf_event_setup(cpu);
5163 }
5164
5165 #ifdef CONFIG_HOTPLUG_CPU
5166 static void __perf_event_exit_cpu(void *info)
5167 {
5168         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5169         struct perf_event_context *ctx = &cpuctx->ctx;
5170         struct perf_event *event, *tmp;
5171
5172         list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry)
5173                 __perf_event_remove_from_context(event);
5174 }
5175 static void perf_event_exit_cpu(int cpu)
5176 {
5177         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5178         struct perf_event_context *ctx = &cpuctx->ctx;
5179
5180         mutex_lock(&ctx->mutex);
5181         smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5182         mutex_unlock(&ctx->mutex);
5183 }
5184 #else
5185 static inline void perf_event_exit_cpu(int cpu) { }
5186 #endif
5187
5188 static int __cpuinit
5189 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5190 {
5191         unsigned int cpu = (long)hcpu;
5192
5193         switch (action) {
5194
5195         case CPU_UP_PREPARE:
5196         case CPU_UP_PREPARE_FROZEN:
5197                 perf_event_init_cpu(cpu);
5198                 break;
5199
5200         case CPU_ONLINE:
5201         case CPU_ONLINE_FROZEN:
5202                 hw_perf_event_setup_online(cpu);
5203                 break;
5204
5205         case CPU_DOWN_PREPARE:
5206         case CPU_DOWN_PREPARE_FROZEN:
5207                 perf_event_exit_cpu(cpu);
5208                 break;
5209
5210         default:
5211                 break;
5212         }
5213
5214         return NOTIFY_OK;
5215 }
5216
5217 /*
5218  * This has to have a higher priority than migration_notifier in sched.c.
5219  */
5220 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5221         .notifier_call          = perf_cpu_notify,
5222         .priority               = 20,
5223 };
5224
5225 void __init perf_event_init(void)
5226 {
5227         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5228                         (void *)(long)smp_processor_id());
5229         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5230                         (void *)(long)smp_processor_id());
5231         register_cpu_notifier(&perf_cpu_nb);
5232 }
5233
5234 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
5235 {
5236         return sprintf(buf, "%d\n", perf_reserved_percpu);
5237 }
5238
5239 static ssize_t
5240 perf_set_reserve_percpu(struct sysdev_class *class,
5241                         const char *buf,
5242                         size_t count)
5243 {
5244         struct perf_cpu_context *cpuctx;
5245         unsigned long val;
5246         int err, cpu, mpt;
5247
5248         err = strict_strtoul(buf, 10, &val);
5249         if (err)
5250                 return err;
5251         if (val > perf_max_events)
5252                 return -EINVAL;
5253
5254         spin_lock(&perf_resource_lock);
5255         perf_reserved_percpu = val;
5256         for_each_online_cpu(cpu) {
5257                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5258                 spin_lock_irq(&cpuctx->ctx.lock);
5259                 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5260                           perf_max_events - perf_reserved_percpu);
5261                 cpuctx->max_pertask = mpt;
5262                 spin_unlock_irq(&cpuctx->ctx.lock);
5263         }
5264         spin_unlock(&perf_resource_lock);
5265
5266         return count;
5267 }
5268
5269 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
5270 {
5271         return sprintf(buf, "%d\n", perf_overcommit);
5272 }
5273
5274 static ssize_t
5275 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
5276 {
5277         unsigned long val;
5278         int err;
5279
5280         err = strict_strtoul(buf, 10, &val);
5281         if (err)
5282                 return err;
5283         if (val > 1)
5284                 return -EINVAL;
5285
5286         spin_lock(&perf_resource_lock);
5287         perf_overcommit = val;
5288         spin_unlock(&perf_resource_lock);
5289
5290         return count;
5291 }
5292
5293 static SYSDEV_CLASS_ATTR(
5294                                 reserve_percpu,
5295                                 0644,
5296                                 perf_show_reserve_percpu,
5297                                 perf_set_reserve_percpu
5298                         );
5299
5300 static SYSDEV_CLASS_ATTR(
5301                                 overcommit,
5302                                 0644,
5303                                 perf_show_overcommit,
5304                                 perf_set_overcommit
5305                         );
5306
5307 static struct attribute *perfclass_attrs[] = {
5308         &attr_reserve_percpu.attr,
5309         &attr_overcommit.attr,
5310         NULL
5311 };
5312
5313 static struct attribute_group perfclass_attr_group = {
5314         .attrs                  = perfclass_attrs,
5315         .name                   = "perf_events",
5316 };
5317
5318 static int __init perf_event_sysfs_init(void)
5319 {
5320         return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5321                                   &perfclass_attr_group);
5322 }
5323 device_initcall(perf_event_sysfs_init);