perf_event: Optimize perf_output_lock()
[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
33 #include <asm/irq_regs.h>
34
35 /*
36  * Each CPU has a list of per CPU events:
37  */
38 DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
39
40 int perf_max_events __read_mostly = 1;
41 static int perf_reserved_percpu __read_mostly;
42 static int perf_overcommit __read_mostly = 1;
43
44 static atomic_t nr_events __read_mostly;
45 static atomic_t nr_mmap_events __read_mostly;
46 static atomic_t nr_comm_events __read_mostly;
47 static atomic_t nr_task_events __read_mostly;
48
49 /*
50  * perf event paranoia level:
51  *  -1 - not paranoid at all
52  *   0 - disallow raw tracepoint access for unpriv
53  *   1 - disallow cpu events for unpriv
54  *   2 - disallow kernel profiling for unpriv
55  */
56 int sysctl_perf_event_paranoid __read_mostly = 1;
57
58 static inline bool perf_paranoid_tracepoint_raw(void)
59 {
60         return sysctl_perf_event_paranoid > -1;
61 }
62
63 static inline bool perf_paranoid_cpu(void)
64 {
65         return sysctl_perf_event_paranoid > 0;
66 }
67
68 static inline bool perf_paranoid_kernel(void)
69 {
70         return sysctl_perf_event_paranoid > 1;
71 }
72
73 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
74
75 /*
76  * max perf event sample rate
77  */
78 int sysctl_perf_event_sample_rate __read_mostly = 100000;
79
80 static atomic64_t perf_event_id;
81
82 /*
83  * Lock for (sysadmin-configurable) event reservations:
84  */
85 static DEFINE_SPINLOCK(perf_resource_lock);
86
87 /*
88  * Architecture provided APIs - weak aliases:
89  */
90 extern __weak const struct pmu *hw_perf_event_init(struct perf_event *event)
91 {
92         return NULL;
93 }
94
95 void __weak hw_perf_disable(void)               { barrier(); }
96 void __weak hw_perf_enable(void)                { barrier(); }
97
98 void __weak hw_perf_event_setup(int cpu)        { barrier(); }
99 void __weak hw_perf_event_setup_online(int cpu) { barrier(); }
100
101 int __weak
102 hw_perf_group_sched_in(struct perf_event *group_leader,
103                struct perf_cpu_context *cpuctx,
104                struct perf_event_context *ctx, int cpu)
105 {
106         return 0;
107 }
108
109 void __weak perf_event_print_debug(void)        { }
110
111 static DEFINE_PER_CPU(int, perf_disable_count);
112
113 void __perf_disable(void)
114 {
115         __get_cpu_var(perf_disable_count)++;
116 }
117
118 bool __perf_enable(void)
119 {
120         return !--__get_cpu_var(perf_disable_count);
121 }
122
123 void perf_disable(void)
124 {
125         __perf_disable();
126         hw_perf_disable();
127 }
128
129 void perf_enable(void)
130 {
131         if (__perf_enable())
132                 hw_perf_enable();
133 }
134
135 static void get_ctx(struct perf_event_context *ctx)
136 {
137         WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
138 }
139
140 static void free_ctx(struct rcu_head *head)
141 {
142         struct perf_event_context *ctx;
143
144         ctx = container_of(head, struct perf_event_context, rcu_head);
145         kfree(ctx);
146 }
147
148 static void put_ctx(struct perf_event_context *ctx)
149 {
150         if (atomic_dec_and_test(&ctx->refcount)) {
151                 if (ctx->parent_ctx)
152                         put_ctx(ctx->parent_ctx);
153                 if (ctx->task)
154                         put_task_struct(ctx->task);
155                 call_rcu(&ctx->rcu_head, free_ctx);
156         }
157 }
158
159 static void unclone_ctx(struct perf_event_context *ctx)
160 {
161         if (ctx->parent_ctx) {
162                 put_ctx(ctx->parent_ctx);
163                 ctx->parent_ctx = NULL;
164         }
165 }
166
167 /*
168  * If we inherit events we want to return the parent event id
169  * to userspace.
170  */
171 static u64 primary_event_id(struct perf_event *event)
172 {
173         u64 id = event->id;
174
175         if (event->parent)
176                 id = event->parent->id;
177
178         return id;
179 }
180
181 /*
182  * Get the perf_event_context for a task and lock it.
183  * This has to cope with with the fact that until it is locked,
184  * the context could get moved to another task.
185  */
186 static struct perf_event_context *
187 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
188 {
189         struct perf_event_context *ctx;
190
191         rcu_read_lock();
192  retry:
193         ctx = rcu_dereference(task->perf_event_ctxp);
194         if (ctx) {
195                 /*
196                  * If this context is a clone of another, it might
197                  * get swapped for another underneath us by
198                  * perf_event_task_sched_out, though the
199                  * rcu_read_lock() protects us from any context
200                  * getting freed.  Lock the context and check if it
201                  * got swapped before we could get the lock, and retry
202                  * if so.  If we locked the right context, then it
203                  * can't get swapped on us any more.
204                  */
205                 spin_lock_irqsave(&ctx->lock, *flags);
206                 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
207                         spin_unlock_irqrestore(&ctx->lock, *flags);
208                         goto retry;
209                 }
210
211                 if (!atomic_inc_not_zero(&ctx->refcount)) {
212                         spin_unlock_irqrestore(&ctx->lock, *flags);
213                         ctx = NULL;
214                 }
215         }
216         rcu_read_unlock();
217         return ctx;
218 }
219
220 /*
221  * Get the context for a task and increment its pin_count so it
222  * can't get swapped to another task.  This also increments its
223  * reference count so that the context can't get freed.
224  */
225 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
226 {
227         struct perf_event_context *ctx;
228         unsigned long flags;
229
230         ctx = perf_lock_task_context(task, &flags);
231         if (ctx) {
232                 ++ctx->pin_count;
233                 spin_unlock_irqrestore(&ctx->lock, flags);
234         }
235         return ctx;
236 }
237
238 static void perf_unpin_context(struct perf_event_context *ctx)
239 {
240         unsigned long flags;
241
242         spin_lock_irqsave(&ctx->lock, flags);
243         --ctx->pin_count;
244         spin_unlock_irqrestore(&ctx->lock, flags);
245         put_ctx(ctx);
246 }
247
248 /*
249  * Add a event from the lists for its context.
250  * Must be called with ctx->mutex and ctx->lock held.
251  */
252 static void
253 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
254 {
255         struct perf_event *group_leader = event->group_leader;
256
257         /*
258          * Depending on whether it is a standalone or sibling event,
259          * add it straight to the context's event list, or to the group
260          * leader's sibling list:
261          */
262         if (group_leader == event)
263                 list_add_tail(&event->group_entry, &ctx->group_list);
264         else {
265                 list_add_tail(&event->group_entry, &group_leader->sibling_list);
266                 group_leader->nr_siblings++;
267         }
268
269         list_add_rcu(&event->event_entry, &ctx->event_list);
270         ctx->nr_events++;
271         if (event->attr.inherit_stat)
272                 ctx->nr_stat++;
273 }
274
275 /*
276  * Remove a event from the lists for its context.
277  * Must be called with ctx->mutex and ctx->lock held.
278  */
279 static void
280 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
281 {
282         struct perf_event *sibling, *tmp;
283
284         if (list_empty(&event->group_entry))
285                 return;
286         ctx->nr_events--;
287         if (event->attr.inherit_stat)
288                 ctx->nr_stat--;
289
290         list_del_init(&event->group_entry);
291         list_del_rcu(&event->event_entry);
292
293         if (event->group_leader != event)
294                 event->group_leader->nr_siblings--;
295
296         /*
297          * If this was a group event with sibling events then
298          * upgrade the siblings to singleton events by adding them
299          * to the context list directly:
300          */
301         list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
302
303                 list_move_tail(&sibling->group_entry, &ctx->group_list);
304                 sibling->group_leader = sibling;
305         }
306 }
307
308 static void
309 event_sched_out(struct perf_event *event,
310                   struct perf_cpu_context *cpuctx,
311                   struct perf_event_context *ctx)
312 {
313         if (event->state != PERF_EVENT_STATE_ACTIVE)
314                 return;
315
316         event->state = PERF_EVENT_STATE_INACTIVE;
317         if (event->pending_disable) {
318                 event->pending_disable = 0;
319                 event->state = PERF_EVENT_STATE_OFF;
320         }
321         event->tstamp_stopped = ctx->time;
322         event->pmu->disable(event);
323         event->oncpu = -1;
324
325         if (!is_software_event(event))
326                 cpuctx->active_oncpu--;
327         ctx->nr_active--;
328         if (event->attr.exclusive || !cpuctx->active_oncpu)
329                 cpuctx->exclusive = 0;
330 }
331
332 static void
333 group_sched_out(struct perf_event *group_event,
334                 struct perf_cpu_context *cpuctx,
335                 struct perf_event_context *ctx)
336 {
337         struct perf_event *event;
338
339         if (group_event->state != PERF_EVENT_STATE_ACTIVE)
340                 return;
341
342         event_sched_out(group_event, cpuctx, ctx);
343
344         /*
345          * Schedule out siblings (if any):
346          */
347         list_for_each_entry(event, &group_event->sibling_list, group_entry)
348                 event_sched_out(event, cpuctx, ctx);
349
350         if (group_event->attr.exclusive)
351                 cpuctx->exclusive = 0;
352 }
353
354 /*
355  * Cross CPU call to remove a performance event
356  *
357  * We disable the event on the hardware level first. After that we
358  * remove it from the context list.
359  */
360 static void __perf_event_remove_from_context(void *info)
361 {
362         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
363         struct perf_event *event = info;
364         struct perf_event_context *ctx = event->ctx;
365
366         /*
367          * If this is a task context, we need to check whether it is
368          * the current task context of this cpu. If not it has been
369          * scheduled out before the smp call arrived.
370          */
371         if (ctx->task && cpuctx->task_ctx != ctx)
372                 return;
373
374         spin_lock(&ctx->lock);
375         /*
376          * Protect the list operation against NMI by disabling the
377          * events on a global level.
378          */
379         perf_disable();
380
381         event_sched_out(event, cpuctx, ctx);
382
383         list_del_event(event, ctx);
384
385         if (!ctx->task) {
386                 /*
387                  * Allow more per task events with respect to the
388                  * reservation:
389                  */
390                 cpuctx->max_pertask =
391                         min(perf_max_events - ctx->nr_events,
392                             perf_max_events - perf_reserved_percpu);
393         }
394
395         perf_enable();
396         spin_unlock(&ctx->lock);
397 }
398
399
400 /*
401  * Remove the event from a task's (or a CPU's) list of events.
402  *
403  * Must be called with ctx->mutex held.
404  *
405  * CPU events are removed with a smp call. For task events we only
406  * call when the task is on a CPU.
407  *
408  * If event->ctx is a cloned context, callers must make sure that
409  * every task struct that event->ctx->task could possibly point to
410  * remains valid.  This is OK when called from perf_release since
411  * that only calls us on the top-level context, which can't be a clone.
412  * When called from perf_event_exit_task, it's OK because the
413  * context has been detached from its task.
414  */
415 static void perf_event_remove_from_context(struct perf_event *event)
416 {
417         struct perf_event_context *ctx = event->ctx;
418         struct task_struct *task = ctx->task;
419
420         if (!task) {
421                 /*
422                  * Per cpu events are removed via an smp call and
423                  * the removal is always sucessful.
424                  */
425                 smp_call_function_single(event->cpu,
426                                          __perf_event_remove_from_context,
427                                          event, 1);
428                 return;
429         }
430
431 retry:
432         task_oncpu_function_call(task, __perf_event_remove_from_context,
433                                  event);
434
435         spin_lock_irq(&ctx->lock);
436         /*
437          * If the context is active we need to retry the smp call.
438          */
439         if (ctx->nr_active && !list_empty(&event->group_entry)) {
440                 spin_unlock_irq(&ctx->lock);
441                 goto retry;
442         }
443
444         /*
445          * The lock prevents that this context is scheduled in so we
446          * can remove the event safely, if the call above did not
447          * succeed.
448          */
449         if (!list_empty(&event->group_entry)) {
450                 list_del_event(event, ctx);
451         }
452         spin_unlock_irq(&ctx->lock);
453 }
454
455 static inline u64 perf_clock(void)
456 {
457         return cpu_clock(smp_processor_id());
458 }
459
460 /*
461  * Update the record of the current time in a context.
462  */
463 static void update_context_time(struct perf_event_context *ctx)
464 {
465         u64 now = perf_clock();
466
467         ctx->time += now - ctx->timestamp;
468         ctx->timestamp = now;
469 }
470
471 /*
472  * Update the total_time_enabled and total_time_running fields for a event.
473  */
474 static void update_event_times(struct perf_event *event)
475 {
476         struct perf_event_context *ctx = event->ctx;
477         u64 run_end;
478
479         if (event->state < PERF_EVENT_STATE_INACTIVE ||
480             event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
481                 return;
482
483         event->total_time_enabled = ctx->time - event->tstamp_enabled;
484
485         if (event->state == PERF_EVENT_STATE_INACTIVE)
486                 run_end = event->tstamp_stopped;
487         else
488                 run_end = ctx->time;
489
490         event->total_time_running = run_end - event->tstamp_running;
491 }
492
493 /*
494  * Update total_time_enabled and total_time_running for all events in a group.
495  */
496 static void update_group_times(struct perf_event *leader)
497 {
498         struct perf_event *event;
499
500         update_event_times(leader);
501         list_for_each_entry(event, &leader->sibling_list, group_entry)
502                 update_event_times(event);
503 }
504
505 /*
506  * Cross CPU call to disable a performance event
507  */
508 static void __perf_event_disable(void *info)
509 {
510         struct perf_event *event = info;
511         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
512         struct perf_event_context *ctx = event->ctx;
513
514         /*
515          * If this is a per-task event, need to check whether this
516          * event's task is the current task on this cpu.
517          */
518         if (ctx->task && cpuctx->task_ctx != ctx)
519                 return;
520
521         spin_lock(&ctx->lock);
522
523         /*
524          * If the event is on, turn it off.
525          * If it is in error state, leave it in error state.
526          */
527         if (event->state >= PERF_EVENT_STATE_INACTIVE) {
528                 update_context_time(ctx);
529                 update_group_times(event);
530                 if (event == event->group_leader)
531                         group_sched_out(event, cpuctx, ctx);
532                 else
533                         event_sched_out(event, cpuctx, ctx);
534                 event->state = PERF_EVENT_STATE_OFF;
535         }
536
537         spin_unlock(&ctx->lock);
538 }
539
540 /*
541  * Disable a event.
542  *
543  * If event->ctx is a cloned context, callers must make sure that
544  * every task struct that event->ctx->task could possibly point to
545  * remains valid.  This condition is satisifed when called through
546  * perf_event_for_each_child or perf_event_for_each because they
547  * hold the top-level event's child_mutex, so any descendant that
548  * goes to exit will block in sync_child_event.
549  * When called from perf_pending_event it's OK because event->ctx
550  * is the current context on this CPU and preemption is disabled,
551  * hence we can't get into perf_event_task_sched_out for this context.
552  */
553 static void perf_event_disable(struct perf_event *event)
554 {
555         struct perf_event_context *ctx = event->ctx;
556         struct task_struct *task = ctx->task;
557
558         if (!task) {
559                 /*
560                  * Disable the event on the cpu that it's on
561                  */
562                 smp_call_function_single(event->cpu, __perf_event_disable,
563                                          event, 1);
564                 return;
565         }
566
567  retry:
568         task_oncpu_function_call(task, __perf_event_disable, event);
569
570         spin_lock_irq(&ctx->lock);
571         /*
572          * If the event is still active, we need to retry the cross-call.
573          */
574         if (event->state == PERF_EVENT_STATE_ACTIVE) {
575                 spin_unlock_irq(&ctx->lock);
576                 goto retry;
577         }
578
579         /*
580          * Since we have the lock this context can't be scheduled
581          * in, so we can change the state safely.
582          */
583         if (event->state == PERF_EVENT_STATE_INACTIVE) {
584                 update_group_times(event);
585                 event->state = PERF_EVENT_STATE_OFF;
586         }
587
588         spin_unlock_irq(&ctx->lock);
589 }
590
591 static int
592 event_sched_in(struct perf_event *event,
593                  struct perf_cpu_context *cpuctx,
594                  struct perf_event_context *ctx,
595                  int cpu)
596 {
597         if (event->state <= PERF_EVENT_STATE_OFF)
598                 return 0;
599
600         event->state = PERF_EVENT_STATE_ACTIVE;
601         event->oncpu = cpu;     /* TODO: put 'cpu' into cpuctx->cpu */
602         /*
603          * The new state must be visible before we turn it on in the hardware:
604          */
605         smp_wmb();
606
607         if (event->pmu->enable(event)) {
608                 event->state = PERF_EVENT_STATE_INACTIVE;
609                 event->oncpu = -1;
610                 return -EAGAIN;
611         }
612
613         event->tstamp_running += ctx->time - event->tstamp_stopped;
614
615         if (!is_software_event(event))
616                 cpuctx->active_oncpu++;
617         ctx->nr_active++;
618
619         if (event->attr.exclusive)
620                 cpuctx->exclusive = 1;
621
622         return 0;
623 }
624
625 static int
626 group_sched_in(struct perf_event *group_event,
627                struct perf_cpu_context *cpuctx,
628                struct perf_event_context *ctx,
629                int cpu)
630 {
631         struct perf_event *event, *partial_group;
632         int ret;
633
634         if (group_event->state == PERF_EVENT_STATE_OFF)
635                 return 0;
636
637         ret = hw_perf_group_sched_in(group_event, cpuctx, ctx, cpu);
638         if (ret)
639                 return ret < 0 ? ret : 0;
640
641         if (event_sched_in(group_event, cpuctx, ctx, cpu))
642                 return -EAGAIN;
643
644         /*
645          * Schedule in siblings as one group (if any):
646          */
647         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
648                 if (event_sched_in(event, cpuctx, ctx, cpu)) {
649                         partial_group = event;
650                         goto group_error;
651                 }
652         }
653
654         return 0;
655
656 group_error:
657         /*
658          * Groups can be scheduled in as one unit only, so undo any
659          * partial group before returning:
660          */
661         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
662                 if (event == partial_group)
663                         break;
664                 event_sched_out(event, cpuctx, ctx);
665         }
666         event_sched_out(group_event, cpuctx, ctx);
667
668         return -EAGAIN;
669 }
670
671 /*
672  * Return 1 for a group consisting entirely of software events,
673  * 0 if the group contains any hardware events.
674  */
675 static int is_software_only_group(struct perf_event *leader)
676 {
677         struct perf_event *event;
678
679         if (!is_software_event(leader))
680                 return 0;
681
682         list_for_each_entry(event, &leader->sibling_list, group_entry)
683                 if (!is_software_event(event))
684                         return 0;
685
686         return 1;
687 }
688
689 /*
690  * Work out whether we can put this event group on the CPU now.
691  */
692 static int group_can_go_on(struct perf_event *event,
693                            struct perf_cpu_context *cpuctx,
694                            int can_add_hw)
695 {
696         /*
697          * Groups consisting entirely of software events can always go on.
698          */
699         if (is_software_only_group(event))
700                 return 1;
701         /*
702          * If an exclusive group is already on, no other hardware
703          * events can go on.
704          */
705         if (cpuctx->exclusive)
706                 return 0;
707         /*
708          * If this group is exclusive and there are already
709          * events on the CPU, it can't go on.
710          */
711         if (event->attr.exclusive && cpuctx->active_oncpu)
712                 return 0;
713         /*
714          * Otherwise, try to add it if all previous groups were able
715          * to go on.
716          */
717         return can_add_hw;
718 }
719
720 static void add_event_to_ctx(struct perf_event *event,
721                                struct perf_event_context *ctx)
722 {
723         list_add_event(event, ctx);
724         event->tstamp_enabled = ctx->time;
725         event->tstamp_running = ctx->time;
726         event->tstamp_stopped = ctx->time;
727 }
728
729 /*
730  * Cross CPU call to install and enable a performance event
731  *
732  * Must be called with ctx->mutex held
733  */
734 static void __perf_install_in_context(void *info)
735 {
736         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
737         struct perf_event *event = info;
738         struct perf_event_context *ctx = event->ctx;
739         struct perf_event *leader = event->group_leader;
740         int cpu = smp_processor_id();
741         int err;
742
743         /*
744          * If this is a task context, we need to check whether it is
745          * the current task context of this cpu. If not it has been
746          * scheduled out before the smp call arrived.
747          * Or possibly this is the right context but it isn't
748          * on this cpu because it had no events.
749          */
750         if (ctx->task && cpuctx->task_ctx != ctx) {
751                 if (cpuctx->task_ctx || ctx->task != current)
752                         return;
753                 cpuctx->task_ctx = ctx;
754         }
755
756         spin_lock(&ctx->lock);
757         ctx->is_active = 1;
758         update_context_time(ctx);
759
760         /*
761          * Protect the list operation against NMI by disabling the
762          * events on a global level. NOP for non NMI based events.
763          */
764         perf_disable();
765
766         add_event_to_ctx(event, ctx);
767
768         /*
769          * Don't put the event on if it is disabled or if
770          * it is in a group and the group isn't on.
771          */
772         if (event->state != PERF_EVENT_STATE_INACTIVE ||
773             (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
774                 goto unlock;
775
776         /*
777          * An exclusive event can't go on if there are already active
778          * hardware events, and no hardware event can go on if there
779          * is already an exclusive event on.
780          */
781         if (!group_can_go_on(event, cpuctx, 1))
782                 err = -EEXIST;
783         else
784                 err = event_sched_in(event, cpuctx, ctx, cpu);
785
786         if (err) {
787                 /*
788                  * This event couldn't go on.  If it is in a group
789                  * then we have to pull the whole group off.
790                  * If the event group is pinned then put it in error state.
791                  */
792                 if (leader != event)
793                         group_sched_out(leader, cpuctx, ctx);
794                 if (leader->attr.pinned) {
795                         update_group_times(leader);
796                         leader->state = PERF_EVENT_STATE_ERROR;
797                 }
798         }
799
800         if (!err && !ctx->task && cpuctx->max_pertask)
801                 cpuctx->max_pertask--;
802
803  unlock:
804         perf_enable();
805
806         spin_unlock(&ctx->lock);
807 }
808
809 /*
810  * Attach a performance event to a context
811  *
812  * First we add the event to the list with the hardware enable bit
813  * in event->hw_config cleared.
814  *
815  * If the event is attached to a task which is on a CPU we use a smp
816  * call to enable it in the task context. The task might have been
817  * scheduled away, but we check this in the smp call again.
818  *
819  * Must be called with ctx->mutex held.
820  */
821 static void
822 perf_install_in_context(struct perf_event_context *ctx,
823                         struct perf_event *event,
824                         int cpu)
825 {
826         struct task_struct *task = ctx->task;
827
828         if (!task) {
829                 /*
830                  * Per cpu events are installed via an smp call and
831                  * the install is always sucessful.
832                  */
833                 smp_call_function_single(cpu, __perf_install_in_context,
834                                          event, 1);
835                 return;
836         }
837
838 retry:
839         task_oncpu_function_call(task, __perf_install_in_context,
840                                  event);
841
842         spin_lock_irq(&ctx->lock);
843         /*
844          * we need to retry the smp call.
845          */
846         if (ctx->is_active && list_empty(&event->group_entry)) {
847                 spin_unlock_irq(&ctx->lock);
848                 goto retry;
849         }
850
851         /*
852          * The lock prevents that this context is scheduled in so we
853          * can add the event safely, if it the call above did not
854          * succeed.
855          */
856         if (list_empty(&event->group_entry))
857                 add_event_to_ctx(event, ctx);
858         spin_unlock_irq(&ctx->lock);
859 }
860
861 /*
862  * Put a event into inactive state and update time fields.
863  * Enabling the leader of a group effectively enables all
864  * the group members that aren't explicitly disabled, so we
865  * have to update their ->tstamp_enabled also.
866  * Note: this works for group members as well as group leaders
867  * since the non-leader members' sibling_lists will be empty.
868  */
869 static void __perf_event_mark_enabled(struct perf_event *event,
870                                         struct perf_event_context *ctx)
871 {
872         struct perf_event *sub;
873
874         event->state = PERF_EVENT_STATE_INACTIVE;
875         event->tstamp_enabled = ctx->time - event->total_time_enabled;
876         list_for_each_entry(sub, &event->sibling_list, group_entry)
877                 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
878                         sub->tstamp_enabled =
879                                 ctx->time - sub->total_time_enabled;
880 }
881
882 /*
883  * Cross CPU call to enable a performance event
884  */
885 static void __perf_event_enable(void *info)
886 {
887         struct perf_event *event = info;
888         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
889         struct perf_event_context *ctx = event->ctx;
890         struct perf_event *leader = event->group_leader;
891         int err;
892
893         /*
894          * If this is a per-task event, need to check whether this
895          * event's task is the current task on this cpu.
896          */
897         if (ctx->task && cpuctx->task_ctx != ctx) {
898                 if (cpuctx->task_ctx || ctx->task != current)
899                         return;
900                 cpuctx->task_ctx = ctx;
901         }
902
903         spin_lock(&ctx->lock);
904         ctx->is_active = 1;
905         update_context_time(ctx);
906
907         if (event->state >= PERF_EVENT_STATE_INACTIVE)
908                 goto unlock;
909         __perf_event_mark_enabled(event, ctx);
910
911         /*
912          * If the event is in a group and isn't the group leader,
913          * then don't put it on unless the group is on.
914          */
915         if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
916                 goto unlock;
917
918         if (!group_can_go_on(event, cpuctx, 1)) {
919                 err = -EEXIST;
920         } else {
921                 perf_disable();
922                 if (event == leader)
923                         err = group_sched_in(event, cpuctx, ctx,
924                                              smp_processor_id());
925                 else
926                         err = event_sched_in(event, cpuctx, ctx,
927                                                smp_processor_id());
928                 perf_enable();
929         }
930
931         if (err) {
932                 /*
933                  * If this event can't go on and it's part of a
934                  * group, then the whole group has to come off.
935                  */
936                 if (leader != event)
937                         group_sched_out(leader, cpuctx, ctx);
938                 if (leader->attr.pinned) {
939                         update_group_times(leader);
940                         leader->state = PERF_EVENT_STATE_ERROR;
941                 }
942         }
943
944  unlock:
945         spin_unlock(&ctx->lock);
946 }
947
948 /*
949  * Enable a event.
950  *
951  * If event->ctx is a cloned context, callers must make sure that
952  * every task struct that event->ctx->task could possibly point to
953  * remains valid.  This condition is satisfied when called through
954  * perf_event_for_each_child or perf_event_for_each as described
955  * for perf_event_disable.
956  */
957 static void perf_event_enable(struct perf_event *event)
958 {
959         struct perf_event_context *ctx = event->ctx;
960         struct task_struct *task = ctx->task;
961
962         if (!task) {
963                 /*
964                  * Enable the event on the cpu that it's on
965                  */
966                 smp_call_function_single(event->cpu, __perf_event_enable,
967                                          event, 1);
968                 return;
969         }
970
971         spin_lock_irq(&ctx->lock);
972         if (event->state >= PERF_EVENT_STATE_INACTIVE)
973                 goto out;
974
975         /*
976          * If the event is in error state, clear that first.
977          * That way, if we see the event in error state below, we
978          * know that it has gone back into error state, as distinct
979          * from the task having been scheduled away before the
980          * cross-call arrived.
981          */
982         if (event->state == PERF_EVENT_STATE_ERROR)
983                 event->state = PERF_EVENT_STATE_OFF;
984
985  retry:
986         spin_unlock_irq(&ctx->lock);
987         task_oncpu_function_call(task, __perf_event_enable, event);
988
989         spin_lock_irq(&ctx->lock);
990
991         /*
992          * If the context is active and the event is still off,
993          * we need to retry the cross-call.
994          */
995         if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
996                 goto retry;
997
998         /*
999          * Since we have the lock this context can't be scheduled
1000          * in, so we can change the state safely.
1001          */
1002         if (event->state == PERF_EVENT_STATE_OFF)
1003                 __perf_event_mark_enabled(event, ctx);
1004
1005  out:
1006         spin_unlock_irq(&ctx->lock);
1007 }
1008
1009 static int perf_event_refresh(struct perf_event *event, int refresh)
1010 {
1011         /*
1012          * not supported on inherited events
1013          */
1014         if (event->attr.inherit)
1015                 return -EINVAL;
1016
1017         atomic_add(refresh, &event->event_limit);
1018         perf_event_enable(event);
1019
1020         return 0;
1021 }
1022
1023 void __perf_event_sched_out(struct perf_event_context *ctx,
1024                               struct perf_cpu_context *cpuctx)
1025 {
1026         struct perf_event *event;
1027
1028         spin_lock(&ctx->lock);
1029         ctx->is_active = 0;
1030         if (likely(!ctx->nr_events))
1031                 goto out;
1032         update_context_time(ctx);
1033
1034         perf_disable();
1035         if (ctx->nr_active)
1036                 list_for_each_entry(event, &ctx->group_list, group_entry)
1037                         group_sched_out(event, cpuctx, ctx);
1038
1039         perf_enable();
1040  out:
1041         spin_unlock(&ctx->lock);
1042 }
1043
1044 /*
1045  * Test whether two contexts are equivalent, i.e. whether they
1046  * have both been cloned from the same version of the same context
1047  * and they both have the same number of enabled events.
1048  * If the number of enabled events is the same, then the set
1049  * of enabled events should be the same, because these are both
1050  * inherited contexts, therefore we can't access individual events
1051  * in them directly with an fd; we can only enable/disable all
1052  * events via prctl, or enable/disable all events in a family
1053  * via ioctl, which will have the same effect on both contexts.
1054  */
1055 static int context_equiv(struct perf_event_context *ctx1,
1056                          struct perf_event_context *ctx2)
1057 {
1058         return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1059                 && ctx1->parent_gen == ctx2->parent_gen
1060                 && !ctx1->pin_count && !ctx2->pin_count;
1061 }
1062
1063 static void __perf_event_read(void *event);
1064
1065 static void __perf_event_sync_stat(struct perf_event *event,
1066                                      struct perf_event *next_event)
1067 {
1068         u64 value;
1069
1070         if (!event->attr.inherit_stat)
1071                 return;
1072
1073         /*
1074          * Update the event value, we cannot use perf_event_read()
1075          * because we're in the middle of a context switch and have IRQs
1076          * disabled, which upsets smp_call_function_single(), however
1077          * we know the event must be on the current CPU, therefore we
1078          * don't need to use it.
1079          */
1080         switch (event->state) {
1081         case PERF_EVENT_STATE_ACTIVE:
1082                 __perf_event_read(event);
1083                 break;
1084
1085         case PERF_EVENT_STATE_INACTIVE:
1086                 update_event_times(event);
1087                 break;
1088
1089         default:
1090                 break;
1091         }
1092
1093         /*
1094          * In order to keep per-task stats reliable we need to flip the event
1095          * values when we flip the contexts.
1096          */
1097         value = atomic64_read(&next_event->count);
1098         value = atomic64_xchg(&event->count, value);
1099         atomic64_set(&next_event->count, value);
1100
1101         swap(event->total_time_enabled, next_event->total_time_enabled);
1102         swap(event->total_time_running, next_event->total_time_running);
1103
1104         /*
1105          * Since we swizzled the values, update the user visible data too.
1106          */
1107         perf_event_update_userpage(event);
1108         perf_event_update_userpage(next_event);
1109 }
1110
1111 #define list_next_entry(pos, member) \
1112         list_entry(pos->member.next, typeof(*pos), member)
1113
1114 static void perf_event_sync_stat(struct perf_event_context *ctx,
1115                                    struct perf_event_context *next_ctx)
1116 {
1117         struct perf_event *event, *next_event;
1118
1119         if (!ctx->nr_stat)
1120                 return;
1121
1122         event = list_first_entry(&ctx->event_list,
1123                                    struct perf_event, event_entry);
1124
1125         next_event = list_first_entry(&next_ctx->event_list,
1126                                         struct perf_event, event_entry);
1127
1128         while (&event->event_entry != &ctx->event_list &&
1129                &next_event->event_entry != &next_ctx->event_list) {
1130
1131                 __perf_event_sync_stat(event, next_event);
1132
1133                 event = list_next_entry(event, event_entry);
1134                 next_event = list_next_entry(next_event, event_entry);
1135         }
1136 }
1137
1138 /*
1139  * Called from scheduler to remove the events of the current task,
1140  * with interrupts disabled.
1141  *
1142  * We stop each event and update the event value in event->count.
1143  *
1144  * This does not protect us against NMI, but disable()
1145  * sets the disabled bit in the control field of event _before_
1146  * accessing the event control register. If a NMI hits, then it will
1147  * not restart the event.
1148  */
1149 void perf_event_task_sched_out(struct task_struct *task,
1150                                  struct task_struct *next, int cpu)
1151 {
1152         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1153         struct perf_event_context *ctx = task->perf_event_ctxp;
1154         struct perf_event_context *next_ctx;
1155         struct perf_event_context *parent;
1156         struct pt_regs *regs;
1157         int do_switch = 1;
1158
1159         regs = task_pt_regs(task);
1160         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1161
1162         if (likely(!ctx || !cpuctx->task_ctx))
1163                 return;
1164
1165         update_context_time(ctx);
1166
1167         rcu_read_lock();
1168         parent = rcu_dereference(ctx->parent_ctx);
1169         next_ctx = next->perf_event_ctxp;
1170         if (parent && next_ctx &&
1171             rcu_dereference(next_ctx->parent_ctx) == parent) {
1172                 /*
1173                  * Looks like the two contexts are clones, so we might be
1174                  * able to optimize the context switch.  We lock both
1175                  * contexts and check that they are clones under the
1176                  * lock (including re-checking that neither has been
1177                  * uncloned in the meantime).  It doesn't matter which
1178                  * order we take the locks because no other cpu could
1179                  * be trying to lock both of these tasks.
1180                  */
1181                 spin_lock(&ctx->lock);
1182                 spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1183                 if (context_equiv(ctx, next_ctx)) {
1184                         /*
1185                          * XXX do we need a memory barrier of sorts
1186                          * wrt to rcu_dereference() of perf_event_ctxp
1187                          */
1188                         task->perf_event_ctxp = next_ctx;
1189                         next->perf_event_ctxp = ctx;
1190                         ctx->task = next;
1191                         next_ctx->task = task;
1192                         do_switch = 0;
1193
1194                         perf_event_sync_stat(ctx, next_ctx);
1195                 }
1196                 spin_unlock(&next_ctx->lock);
1197                 spin_unlock(&ctx->lock);
1198         }
1199         rcu_read_unlock();
1200
1201         if (do_switch) {
1202                 __perf_event_sched_out(ctx, cpuctx);
1203                 cpuctx->task_ctx = NULL;
1204         }
1205 }
1206
1207 /*
1208  * Called with IRQs disabled
1209  */
1210 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1211 {
1212         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1213
1214         if (!cpuctx->task_ctx)
1215                 return;
1216
1217         if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1218                 return;
1219
1220         __perf_event_sched_out(ctx, cpuctx);
1221         cpuctx->task_ctx = NULL;
1222 }
1223
1224 /*
1225  * Called with IRQs disabled
1226  */
1227 static void perf_event_cpu_sched_out(struct perf_cpu_context *cpuctx)
1228 {
1229         __perf_event_sched_out(&cpuctx->ctx, cpuctx);
1230 }
1231
1232 static void
1233 __perf_event_sched_in(struct perf_event_context *ctx,
1234                         struct perf_cpu_context *cpuctx, int cpu)
1235 {
1236         struct perf_event *event;
1237         int can_add_hw = 1;
1238
1239         spin_lock(&ctx->lock);
1240         ctx->is_active = 1;
1241         if (likely(!ctx->nr_events))
1242                 goto out;
1243
1244         ctx->timestamp = perf_clock();
1245
1246         perf_disable();
1247
1248         /*
1249          * First go through the list and put on any pinned groups
1250          * in order to give them the best chance of going on.
1251          */
1252         list_for_each_entry(event, &ctx->group_list, group_entry) {
1253                 if (event->state <= PERF_EVENT_STATE_OFF ||
1254                     !event->attr.pinned)
1255                         continue;
1256                 if (event->cpu != -1 && event->cpu != cpu)
1257                         continue;
1258
1259                 if (group_can_go_on(event, cpuctx, 1))
1260                         group_sched_in(event, cpuctx, ctx, cpu);
1261
1262                 /*
1263                  * If this pinned group hasn't been scheduled,
1264                  * put it in error state.
1265                  */
1266                 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1267                         update_group_times(event);
1268                         event->state = PERF_EVENT_STATE_ERROR;
1269                 }
1270         }
1271
1272         list_for_each_entry(event, &ctx->group_list, group_entry) {
1273                 /*
1274                  * Ignore events in OFF or ERROR state, and
1275                  * ignore pinned events since we did them already.
1276                  */
1277                 if (event->state <= PERF_EVENT_STATE_OFF ||
1278                     event->attr.pinned)
1279                         continue;
1280
1281                 /*
1282                  * Listen to the 'cpu' scheduling filter constraint
1283                  * of events:
1284                  */
1285                 if (event->cpu != -1 && event->cpu != cpu)
1286                         continue;
1287
1288                 if (group_can_go_on(event, cpuctx, can_add_hw))
1289                         if (group_sched_in(event, cpuctx, ctx, cpu))
1290                                 can_add_hw = 0;
1291         }
1292         perf_enable();
1293  out:
1294         spin_unlock(&ctx->lock);
1295 }
1296
1297 /*
1298  * Called from scheduler to add the events of the current task
1299  * with interrupts disabled.
1300  *
1301  * We restore the event value and then enable it.
1302  *
1303  * This does not protect us against NMI, but enable()
1304  * sets the enabled bit in the control field of event _before_
1305  * accessing the event control register. If a NMI hits, then it will
1306  * keep the event running.
1307  */
1308 void perf_event_task_sched_in(struct task_struct *task, int cpu)
1309 {
1310         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1311         struct perf_event_context *ctx = task->perf_event_ctxp;
1312
1313         if (likely(!ctx))
1314                 return;
1315         if (cpuctx->task_ctx == ctx)
1316                 return;
1317         __perf_event_sched_in(ctx, cpuctx, cpu);
1318         cpuctx->task_ctx = ctx;
1319 }
1320
1321 static void perf_event_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
1322 {
1323         struct perf_event_context *ctx = &cpuctx->ctx;
1324
1325         __perf_event_sched_in(ctx, cpuctx, cpu);
1326 }
1327
1328 #define MAX_INTERRUPTS (~0ULL)
1329
1330 static void perf_log_throttle(struct perf_event *event, int enable);
1331
1332 static void perf_adjust_period(struct perf_event *event, u64 events)
1333 {
1334         struct hw_perf_event *hwc = &event->hw;
1335         u64 period, sample_period;
1336         s64 delta;
1337
1338         events *= hwc->sample_period;
1339         period = div64_u64(events, event->attr.sample_freq);
1340
1341         delta = (s64)(period - hwc->sample_period);
1342         delta = (delta + 7) / 8; /* low pass filter */
1343
1344         sample_period = hwc->sample_period + delta;
1345
1346         if (!sample_period)
1347                 sample_period = 1;
1348
1349         hwc->sample_period = sample_period;
1350 }
1351
1352 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1353 {
1354         struct perf_event *event;
1355         struct hw_perf_event *hwc;
1356         u64 interrupts, freq;
1357
1358         spin_lock(&ctx->lock);
1359         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1360                 if (event->state != PERF_EVENT_STATE_ACTIVE)
1361                         continue;
1362
1363                 hwc = &event->hw;
1364
1365                 interrupts = hwc->interrupts;
1366                 hwc->interrupts = 0;
1367
1368                 /*
1369                  * unthrottle events on the tick
1370                  */
1371                 if (interrupts == MAX_INTERRUPTS) {
1372                         perf_log_throttle(event, 1);
1373                         event->pmu->unthrottle(event);
1374                         interrupts = 2*sysctl_perf_event_sample_rate/HZ;
1375                 }
1376
1377                 if (!event->attr.freq || !event->attr.sample_freq)
1378                         continue;
1379
1380                 /*
1381                  * if the specified freq < HZ then we need to skip ticks
1382                  */
1383                 if (event->attr.sample_freq < HZ) {
1384                         freq = event->attr.sample_freq;
1385
1386                         hwc->freq_count += freq;
1387                         hwc->freq_interrupts += interrupts;
1388
1389                         if (hwc->freq_count < HZ)
1390                                 continue;
1391
1392                         interrupts = hwc->freq_interrupts;
1393                         hwc->freq_interrupts = 0;
1394                         hwc->freq_count -= HZ;
1395                 } else
1396                         freq = HZ;
1397
1398                 perf_adjust_period(event, freq * interrupts);
1399
1400                 /*
1401                  * In order to avoid being stalled by an (accidental) huge
1402                  * sample period, force reset the sample period if we didn't
1403                  * get any events in this freq period.
1404                  */
1405                 if (!interrupts) {
1406                         perf_disable();
1407                         event->pmu->disable(event);
1408                         atomic64_set(&hwc->period_left, 0);
1409                         event->pmu->enable(event);
1410                         perf_enable();
1411                 }
1412         }
1413         spin_unlock(&ctx->lock);
1414 }
1415
1416 /*
1417  * Round-robin a context's events:
1418  */
1419 static void rotate_ctx(struct perf_event_context *ctx)
1420 {
1421         struct perf_event *event;
1422
1423         if (!ctx->nr_events)
1424                 return;
1425
1426         spin_lock(&ctx->lock);
1427         /*
1428          * Rotate the first entry last (works just fine for group events too):
1429          */
1430         perf_disable();
1431         list_for_each_entry(event, &ctx->group_list, group_entry) {
1432                 list_move_tail(&event->group_entry, &ctx->group_list);
1433                 break;
1434         }
1435         perf_enable();
1436
1437         spin_unlock(&ctx->lock);
1438 }
1439
1440 void perf_event_task_tick(struct task_struct *curr, int cpu)
1441 {
1442         struct perf_cpu_context *cpuctx;
1443         struct perf_event_context *ctx;
1444
1445         if (!atomic_read(&nr_events))
1446                 return;
1447
1448         cpuctx = &per_cpu(perf_cpu_context, cpu);
1449         ctx = curr->perf_event_ctxp;
1450
1451         perf_ctx_adjust_freq(&cpuctx->ctx);
1452         if (ctx)
1453                 perf_ctx_adjust_freq(ctx);
1454
1455         perf_event_cpu_sched_out(cpuctx);
1456         if (ctx)
1457                 __perf_event_task_sched_out(ctx);
1458
1459         rotate_ctx(&cpuctx->ctx);
1460         if (ctx)
1461                 rotate_ctx(ctx);
1462
1463         perf_event_cpu_sched_in(cpuctx, cpu);
1464         if (ctx)
1465                 perf_event_task_sched_in(curr, cpu);
1466 }
1467
1468 /*
1469  * Enable all of a task's events that have been marked enable-on-exec.
1470  * This expects task == current.
1471  */
1472 static void perf_event_enable_on_exec(struct task_struct *task)
1473 {
1474         struct perf_event_context *ctx;
1475         struct perf_event *event;
1476         unsigned long flags;
1477         int enabled = 0;
1478
1479         local_irq_save(flags);
1480         ctx = task->perf_event_ctxp;
1481         if (!ctx || !ctx->nr_events)
1482                 goto out;
1483
1484         __perf_event_task_sched_out(ctx);
1485
1486         spin_lock(&ctx->lock);
1487
1488         list_for_each_entry(event, &ctx->group_list, group_entry) {
1489                 if (!event->attr.enable_on_exec)
1490                         continue;
1491                 event->attr.enable_on_exec = 0;
1492                 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1493                         continue;
1494                 __perf_event_mark_enabled(event, ctx);
1495                 enabled = 1;
1496         }
1497
1498         /*
1499          * Unclone this context if we enabled any event.
1500          */
1501         if (enabled)
1502                 unclone_ctx(ctx);
1503
1504         spin_unlock(&ctx->lock);
1505
1506         perf_event_task_sched_in(task, smp_processor_id());
1507  out:
1508         local_irq_restore(flags);
1509 }
1510
1511 /*
1512  * Cross CPU call to read the hardware event
1513  */
1514 static void __perf_event_read(void *info)
1515 {
1516         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1517         struct perf_event *event = info;
1518         struct perf_event_context *ctx = event->ctx;
1519         unsigned long flags;
1520
1521         /*
1522          * If this is a task context, we need to check whether it is
1523          * the current task context of this cpu.  If not it has been
1524          * scheduled out before the smp call arrived.  In that case
1525          * event->count would have been updated to a recent sample
1526          * when the event was scheduled out.
1527          */
1528         if (ctx->task && cpuctx->task_ctx != ctx)
1529                 return;
1530
1531         local_irq_save(flags);
1532         if (ctx->is_active)
1533                 update_context_time(ctx);
1534         event->pmu->read(event);
1535         update_event_times(event);
1536         local_irq_restore(flags);
1537 }
1538
1539 static u64 perf_event_read(struct perf_event *event)
1540 {
1541         /*
1542          * If event is enabled and currently active on a CPU, update the
1543          * value in the event structure:
1544          */
1545         if (event->state == PERF_EVENT_STATE_ACTIVE) {
1546                 smp_call_function_single(event->oncpu,
1547                                          __perf_event_read, event, 1);
1548         } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1549                 update_event_times(event);
1550         }
1551
1552         return atomic64_read(&event->count);
1553 }
1554
1555 /*
1556  * Initialize the perf_event context in a task_struct:
1557  */
1558 static void
1559 __perf_event_init_context(struct perf_event_context *ctx,
1560                             struct task_struct *task)
1561 {
1562         memset(ctx, 0, sizeof(*ctx));
1563         spin_lock_init(&ctx->lock);
1564         mutex_init(&ctx->mutex);
1565         INIT_LIST_HEAD(&ctx->group_list);
1566         INIT_LIST_HEAD(&ctx->event_list);
1567         atomic_set(&ctx->refcount, 1);
1568         ctx->task = task;
1569 }
1570
1571 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1572 {
1573         struct perf_event_context *ctx;
1574         struct perf_cpu_context *cpuctx;
1575         struct task_struct *task;
1576         unsigned long flags;
1577         int err;
1578
1579         /*
1580          * If cpu is not a wildcard then this is a percpu event:
1581          */
1582         if (cpu != -1) {
1583                 /* Must be root to operate on a CPU event: */
1584                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1585                         return ERR_PTR(-EACCES);
1586
1587                 if (cpu < 0 || cpu > num_possible_cpus())
1588                         return ERR_PTR(-EINVAL);
1589
1590                 /*
1591                  * We could be clever and allow to attach a event to an
1592                  * offline CPU and activate it when the CPU comes up, but
1593                  * that's for later.
1594                  */
1595                 if (!cpu_isset(cpu, cpu_online_map))
1596                         return ERR_PTR(-ENODEV);
1597
1598                 cpuctx = &per_cpu(perf_cpu_context, cpu);
1599                 ctx = &cpuctx->ctx;
1600                 get_ctx(ctx);
1601
1602                 return ctx;
1603         }
1604
1605         rcu_read_lock();
1606         if (!pid)
1607                 task = current;
1608         else
1609                 task = find_task_by_vpid(pid);
1610         if (task)
1611                 get_task_struct(task);
1612         rcu_read_unlock();
1613
1614         if (!task)
1615                 return ERR_PTR(-ESRCH);
1616
1617         /*
1618          * Can't attach events to a dying task.
1619          */
1620         err = -ESRCH;
1621         if (task->flags & PF_EXITING)
1622                 goto errout;
1623
1624         /* Reuse ptrace permission checks for now. */
1625         err = -EACCES;
1626         if (!ptrace_may_access(task, PTRACE_MODE_READ))
1627                 goto errout;
1628
1629  retry:
1630         ctx = perf_lock_task_context(task, &flags);
1631         if (ctx) {
1632                 unclone_ctx(ctx);
1633                 spin_unlock_irqrestore(&ctx->lock, flags);
1634         }
1635
1636         if (!ctx) {
1637                 ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1638                 err = -ENOMEM;
1639                 if (!ctx)
1640                         goto errout;
1641                 __perf_event_init_context(ctx, task);
1642                 get_ctx(ctx);
1643                 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1644                         /*
1645                          * We raced with some other task; use
1646                          * the context they set.
1647                          */
1648                         kfree(ctx);
1649                         goto retry;
1650                 }
1651                 get_task_struct(task);
1652         }
1653
1654         put_task_struct(task);
1655         return ctx;
1656
1657  errout:
1658         put_task_struct(task);
1659         return ERR_PTR(err);
1660 }
1661
1662 static void perf_event_free_filter(struct perf_event *event);
1663
1664 static void free_event_rcu(struct rcu_head *head)
1665 {
1666         struct perf_event *event;
1667
1668         event = container_of(head, struct perf_event, rcu_head);
1669         if (event->ns)
1670                 put_pid_ns(event->ns);
1671         perf_event_free_filter(event);
1672         kfree(event);
1673 }
1674
1675 static void perf_pending_sync(struct perf_event *event);
1676
1677 static void free_event(struct perf_event *event)
1678 {
1679         perf_pending_sync(event);
1680
1681         if (!event->parent) {
1682                 atomic_dec(&nr_events);
1683                 if (event->attr.mmap)
1684                         atomic_dec(&nr_mmap_events);
1685                 if (event->attr.comm)
1686                         atomic_dec(&nr_comm_events);
1687                 if (event->attr.task)
1688                         atomic_dec(&nr_task_events);
1689         }
1690
1691         if (event->output) {
1692                 fput(event->output->filp);
1693                 event->output = NULL;
1694         }
1695
1696         if (event->destroy)
1697                 event->destroy(event);
1698
1699         put_ctx(event->ctx);
1700         call_rcu(&event->rcu_head, free_event_rcu);
1701 }
1702
1703 /*
1704  * Called when the last reference to the file is gone.
1705  */
1706 static int perf_release(struct inode *inode, struct file *file)
1707 {
1708         struct perf_event *event = file->private_data;
1709         struct perf_event_context *ctx = event->ctx;
1710
1711         file->private_data = NULL;
1712
1713         WARN_ON_ONCE(ctx->parent_ctx);
1714         mutex_lock(&ctx->mutex);
1715         perf_event_remove_from_context(event);
1716         mutex_unlock(&ctx->mutex);
1717
1718         mutex_lock(&event->owner->perf_event_mutex);
1719         list_del_init(&event->owner_entry);
1720         mutex_unlock(&event->owner->perf_event_mutex);
1721         put_task_struct(event->owner);
1722
1723         free_event(event);
1724
1725         return 0;
1726 }
1727
1728 static int perf_event_read_size(struct perf_event *event)
1729 {
1730         int entry = sizeof(u64); /* value */
1731         int size = 0;
1732         int nr = 1;
1733
1734         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1735                 size += sizeof(u64);
1736
1737         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1738                 size += sizeof(u64);
1739
1740         if (event->attr.read_format & PERF_FORMAT_ID)
1741                 entry += sizeof(u64);
1742
1743         if (event->attr.read_format & PERF_FORMAT_GROUP) {
1744                 nr += event->group_leader->nr_siblings;
1745                 size += sizeof(u64);
1746         }
1747
1748         size += entry * nr;
1749
1750         return size;
1751 }
1752
1753 static u64 perf_event_read_value(struct perf_event *event)
1754 {
1755         struct perf_event *child;
1756         u64 total = 0;
1757
1758         total += perf_event_read(event);
1759         list_for_each_entry(child, &event->child_list, child_list)
1760                 total += perf_event_read(child);
1761
1762         return total;
1763 }
1764
1765 static int perf_event_read_entry(struct perf_event *event,
1766                                    u64 read_format, char __user *buf)
1767 {
1768         int n = 0, count = 0;
1769         u64 values[2];
1770
1771         values[n++] = perf_event_read_value(event);
1772         if (read_format & PERF_FORMAT_ID)
1773                 values[n++] = primary_event_id(event);
1774
1775         count = n * sizeof(u64);
1776
1777         if (copy_to_user(buf, values, count))
1778                 return -EFAULT;
1779
1780         return count;
1781 }
1782
1783 static int perf_event_read_group(struct perf_event *event,
1784                                    u64 read_format, char __user *buf)
1785 {
1786         struct perf_event *leader = event->group_leader, *sub;
1787         int n = 0, size = 0, err = -EFAULT;
1788         u64 values[3];
1789
1790         values[n++] = 1 + leader->nr_siblings;
1791         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1792                 values[n++] = leader->total_time_enabled +
1793                         atomic64_read(&leader->child_total_time_enabled);
1794         }
1795         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1796                 values[n++] = leader->total_time_running +
1797                         atomic64_read(&leader->child_total_time_running);
1798         }
1799
1800         size = n * sizeof(u64);
1801
1802         if (copy_to_user(buf, values, size))
1803                 return -EFAULT;
1804
1805         err = perf_event_read_entry(leader, read_format, buf + size);
1806         if (err < 0)
1807                 return err;
1808
1809         size += err;
1810
1811         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1812                 err = perf_event_read_entry(sub, read_format,
1813                                 buf + size);
1814                 if (err < 0)
1815                         return err;
1816
1817                 size += err;
1818         }
1819
1820         return size;
1821 }
1822
1823 static int perf_event_read_one(struct perf_event *event,
1824                                  u64 read_format, char __user *buf)
1825 {
1826         u64 values[4];
1827         int n = 0;
1828
1829         values[n++] = perf_event_read_value(event);
1830         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1831                 values[n++] = event->total_time_enabled +
1832                         atomic64_read(&event->child_total_time_enabled);
1833         }
1834         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1835                 values[n++] = event->total_time_running +
1836                         atomic64_read(&event->child_total_time_running);
1837         }
1838         if (read_format & PERF_FORMAT_ID)
1839                 values[n++] = primary_event_id(event);
1840
1841         if (copy_to_user(buf, values, n * sizeof(u64)))
1842                 return -EFAULT;
1843
1844         return n * sizeof(u64);
1845 }
1846
1847 /*
1848  * Read the performance event - simple non blocking version for now
1849  */
1850 static ssize_t
1851 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
1852 {
1853         u64 read_format = event->attr.read_format;
1854         int ret;
1855
1856         /*
1857          * Return end-of-file for a read on a event that is in
1858          * error state (i.e. because it was pinned but it couldn't be
1859          * scheduled on to the CPU at some point).
1860          */
1861         if (event->state == PERF_EVENT_STATE_ERROR)
1862                 return 0;
1863
1864         if (count < perf_event_read_size(event))
1865                 return -ENOSPC;
1866
1867         WARN_ON_ONCE(event->ctx->parent_ctx);
1868         mutex_lock(&event->child_mutex);
1869         if (read_format & PERF_FORMAT_GROUP)
1870                 ret = perf_event_read_group(event, read_format, buf);
1871         else
1872                 ret = perf_event_read_one(event, read_format, buf);
1873         mutex_unlock(&event->child_mutex);
1874
1875         return ret;
1876 }
1877
1878 static ssize_t
1879 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1880 {
1881         struct perf_event *event = file->private_data;
1882
1883         return perf_read_hw(event, buf, count);
1884 }
1885
1886 static unsigned int perf_poll(struct file *file, poll_table *wait)
1887 {
1888         struct perf_event *event = file->private_data;
1889         struct perf_mmap_data *data;
1890         unsigned int events = POLL_HUP;
1891
1892         rcu_read_lock();
1893         data = rcu_dereference(event->data);
1894         if (data)
1895                 events = atomic_xchg(&data->poll, 0);
1896         rcu_read_unlock();
1897
1898         poll_wait(file, &event->waitq, wait);
1899
1900         return events;
1901 }
1902
1903 static void perf_event_reset(struct perf_event *event)
1904 {
1905         (void)perf_event_read(event);
1906         atomic64_set(&event->count, 0);
1907         perf_event_update_userpage(event);
1908 }
1909
1910 /*
1911  * Holding the top-level event's child_mutex means that any
1912  * descendant process that has inherited this event will block
1913  * in sync_child_event if it goes to exit, thus satisfying the
1914  * task existence requirements of perf_event_enable/disable.
1915  */
1916 static void perf_event_for_each_child(struct perf_event *event,
1917                                         void (*func)(struct perf_event *))
1918 {
1919         struct perf_event *child;
1920
1921         WARN_ON_ONCE(event->ctx->parent_ctx);
1922         mutex_lock(&event->child_mutex);
1923         func(event);
1924         list_for_each_entry(child, &event->child_list, child_list)
1925                 func(child);
1926         mutex_unlock(&event->child_mutex);
1927 }
1928
1929 static void perf_event_for_each(struct perf_event *event,
1930                                   void (*func)(struct perf_event *))
1931 {
1932         struct perf_event_context *ctx = event->ctx;
1933         struct perf_event *sibling;
1934
1935         WARN_ON_ONCE(ctx->parent_ctx);
1936         mutex_lock(&ctx->mutex);
1937         event = event->group_leader;
1938
1939         perf_event_for_each_child(event, func);
1940         func(event);
1941         list_for_each_entry(sibling, &event->sibling_list, group_entry)
1942                 perf_event_for_each_child(event, func);
1943         mutex_unlock(&ctx->mutex);
1944 }
1945
1946 static int perf_event_period(struct perf_event *event, u64 __user *arg)
1947 {
1948         struct perf_event_context *ctx = event->ctx;
1949         unsigned long size;
1950         int ret = 0;
1951         u64 value;
1952
1953         if (!event->attr.sample_period)
1954                 return -EINVAL;
1955
1956         size = copy_from_user(&value, arg, sizeof(value));
1957         if (size != sizeof(value))
1958                 return -EFAULT;
1959
1960         if (!value)
1961                 return -EINVAL;
1962
1963         spin_lock_irq(&ctx->lock);
1964         if (event->attr.freq) {
1965                 if (value > sysctl_perf_event_sample_rate) {
1966                         ret = -EINVAL;
1967                         goto unlock;
1968                 }
1969
1970                 event->attr.sample_freq = value;
1971         } else {
1972                 event->attr.sample_period = value;
1973                 event->hw.sample_period = value;
1974         }
1975 unlock:
1976         spin_unlock_irq(&ctx->lock);
1977
1978         return ret;
1979 }
1980
1981 static int perf_event_set_output(struct perf_event *event, int output_fd);
1982 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
1983
1984 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1985 {
1986         struct perf_event *event = file->private_data;
1987         void (*func)(struct perf_event *);
1988         u32 flags = arg;
1989
1990         switch (cmd) {
1991         case PERF_EVENT_IOC_ENABLE:
1992                 func = perf_event_enable;
1993                 break;
1994         case PERF_EVENT_IOC_DISABLE:
1995                 func = perf_event_disable;
1996                 break;
1997         case PERF_EVENT_IOC_RESET:
1998                 func = perf_event_reset;
1999                 break;
2000
2001         case PERF_EVENT_IOC_REFRESH:
2002                 return perf_event_refresh(event, arg);
2003
2004         case PERF_EVENT_IOC_PERIOD:
2005                 return perf_event_period(event, (u64 __user *)arg);
2006
2007         case PERF_EVENT_IOC_SET_OUTPUT:
2008                 return perf_event_set_output(event, arg);
2009
2010         case PERF_EVENT_IOC_SET_FILTER:
2011                 return perf_event_set_filter(event, (void __user *)arg);
2012
2013         default:
2014                 return -ENOTTY;
2015         }
2016
2017         if (flags & PERF_IOC_FLAG_GROUP)
2018                 perf_event_for_each(event, func);
2019         else
2020                 perf_event_for_each_child(event, func);
2021
2022         return 0;
2023 }
2024
2025 int perf_event_task_enable(void)
2026 {
2027         struct perf_event *event;
2028
2029         mutex_lock(&current->perf_event_mutex);
2030         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2031                 perf_event_for_each_child(event, perf_event_enable);
2032         mutex_unlock(&current->perf_event_mutex);
2033
2034         return 0;
2035 }
2036
2037 int perf_event_task_disable(void)
2038 {
2039         struct perf_event *event;
2040
2041         mutex_lock(&current->perf_event_mutex);
2042         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2043                 perf_event_for_each_child(event, perf_event_disable);
2044         mutex_unlock(&current->perf_event_mutex);
2045
2046         return 0;
2047 }
2048
2049 #ifndef PERF_EVENT_INDEX_OFFSET
2050 # define PERF_EVENT_INDEX_OFFSET 0
2051 #endif
2052
2053 static int perf_event_index(struct perf_event *event)
2054 {
2055         if (event->state != PERF_EVENT_STATE_ACTIVE)
2056                 return 0;
2057
2058         return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2059 }
2060
2061 /*
2062  * Callers need to ensure there can be no nesting of this function, otherwise
2063  * the seqlock logic goes bad. We can not serialize this because the arch
2064  * code calls this from NMI context.
2065  */
2066 void perf_event_update_userpage(struct perf_event *event)
2067 {
2068         struct perf_event_mmap_page *userpg;
2069         struct perf_mmap_data *data;
2070
2071         rcu_read_lock();
2072         data = rcu_dereference(event->data);
2073         if (!data)
2074                 goto unlock;
2075
2076         userpg = data->user_page;
2077
2078         /*
2079          * Disable preemption so as to not let the corresponding user-space
2080          * spin too long if we get preempted.
2081          */
2082         preempt_disable();
2083         ++userpg->lock;
2084         barrier();
2085         userpg->index = perf_event_index(event);
2086         userpg->offset = atomic64_read(&event->count);
2087         if (event->state == PERF_EVENT_STATE_ACTIVE)
2088                 userpg->offset -= atomic64_read(&event->hw.prev_count);
2089
2090         userpg->time_enabled = event->total_time_enabled +
2091                         atomic64_read(&event->child_total_time_enabled);
2092
2093         userpg->time_running = event->total_time_running +
2094                         atomic64_read(&event->child_total_time_running);
2095
2096         barrier();
2097         ++userpg->lock;
2098         preempt_enable();
2099 unlock:
2100         rcu_read_unlock();
2101 }
2102
2103 static unsigned long perf_data_size(struct perf_mmap_data *data)
2104 {
2105         return data->nr_pages << (PAGE_SHIFT + data->data_order);
2106 }
2107
2108 #ifndef CONFIG_PERF_USE_VMALLOC
2109
2110 /*
2111  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2112  */
2113
2114 static struct page *
2115 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2116 {
2117         if (pgoff > data->nr_pages)
2118                 return NULL;
2119
2120         if (pgoff == 0)
2121                 return virt_to_page(data->user_page);
2122
2123         return virt_to_page(data->data_pages[pgoff - 1]);
2124 }
2125
2126 static struct perf_mmap_data *
2127 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2128 {
2129         struct perf_mmap_data *data;
2130         unsigned long size;
2131         int i;
2132
2133         WARN_ON(atomic_read(&event->mmap_count));
2134
2135         size = sizeof(struct perf_mmap_data);
2136         size += nr_pages * sizeof(void *);
2137
2138         data = kzalloc(size, GFP_KERNEL);
2139         if (!data)
2140                 goto fail;
2141
2142         data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2143         if (!data->user_page)
2144                 goto fail_user_page;
2145
2146         for (i = 0; i < nr_pages; i++) {
2147                 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2148                 if (!data->data_pages[i])
2149                         goto fail_data_pages;
2150         }
2151
2152         data->data_order = 0;
2153         data->nr_pages = nr_pages;
2154
2155         return data;
2156
2157 fail_data_pages:
2158         for (i--; i >= 0; i--)
2159                 free_page((unsigned long)data->data_pages[i]);
2160
2161         free_page((unsigned long)data->user_page);
2162
2163 fail_user_page:
2164         kfree(data);
2165
2166 fail:
2167         return NULL;
2168 }
2169
2170 static void perf_mmap_free_page(unsigned long addr)
2171 {
2172         struct page *page = virt_to_page((void *)addr);
2173
2174         page->mapping = NULL;
2175         __free_page(page);
2176 }
2177
2178 static void perf_mmap_data_free(struct perf_mmap_data *data)
2179 {
2180         int i;
2181
2182         perf_mmap_free_page((unsigned long)data->user_page);
2183         for (i = 0; i < data->nr_pages; i++)
2184                 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2185 }
2186
2187 #else
2188
2189 /*
2190  * Back perf_mmap() with vmalloc memory.
2191  *
2192  * Required for architectures that have d-cache aliasing issues.
2193  */
2194
2195 static struct page *
2196 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2197 {
2198         if (pgoff > (1UL << data->data_order))
2199                 return NULL;
2200
2201         return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2202 }
2203
2204 static void perf_mmap_unmark_page(void *addr)
2205 {
2206         struct page *page = vmalloc_to_page(addr);
2207
2208         page->mapping = NULL;
2209 }
2210
2211 static void perf_mmap_data_free_work(struct work_struct *work)
2212 {
2213         struct perf_mmap_data *data;
2214         void *base;
2215         int i, nr;
2216
2217         data = container_of(work, struct perf_mmap_data, work);
2218         nr = 1 << data->data_order;
2219
2220         base = data->user_page;
2221         for (i = 0; i < nr + 1; i++)
2222                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2223
2224         vfree(base);
2225 }
2226
2227 static void perf_mmap_data_free(struct perf_mmap_data *data)
2228 {
2229         schedule_work(&data->work);
2230 }
2231
2232 static struct perf_mmap_data *
2233 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2234 {
2235         struct perf_mmap_data *data;
2236         unsigned long size;
2237         void *all_buf;
2238
2239         WARN_ON(atomic_read(&event->mmap_count));
2240
2241         size = sizeof(struct perf_mmap_data);
2242         size += sizeof(void *);
2243
2244         data = kzalloc(size, GFP_KERNEL);
2245         if (!data)
2246                 goto fail;
2247
2248         INIT_WORK(&data->work, perf_mmap_data_free_work);
2249
2250         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2251         if (!all_buf)
2252                 goto fail_all_buf;
2253
2254         data->user_page = all_buf;
2255         data->data_pages[0] = all_buf + PAGE_SIZE;
2256         data->data_order = ilog2(nr_pages);
2257         data->nr_pages = 1;
2258
2259         return data;
2260
2261 fail_all_buf:
2262         kfree(data);
2263
2264 fail:
2265         return NULL;
2266 }
2267
2268 #endif
2269
2270 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2271 {
2272         struct perf_event *event = vma->vm_file->private_data;
2273         struct perf_mmap_data *data;
2274         int ret = VM_FAULT_SIGBUS;
2275
2276         if (vmf->flags & FAULT_FLAG_MKWRITE) {
2277                 if (vmf->pgoff == 0)
2278                         ret = 0;
2279                 return ret;
2280         }
2281
2282         rcu_read_lock();
2283         data = rcu_dereference(event->data);
2284         if (!data)
2285                 goto unlock;
2286
2287         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2288                 goto unlock;
2289
2290         vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2291         if (!vmf->page)
2292                 goto unlock;
2293
2294         get_page(vmf->page);
2295         vmf->page->mapping = vma->vm_file->f_mapping;
2296         vmf->page->index   = vmf->pgoff;
2297
2298         ret = 0;
2299 unlock:
2300         rcu_read_unlock();
2301
2302         return ret;
2303 }
2304
2305 static void
2306 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2307 {
2308         long max_size = perf_data_size(data);
2309
2310         atomic_set(&data->lock, -1);
2311
2312         if (event->attr.watermark) {
2313                 data->watermark = min_t(long, max_size,
2314                                         event->attr.wakeup_watermark);
2315         }
2316
2317         if (!data->watermark)
2318                 data->watermark = max_t(long, PAGE_SIZE, max_size / 2);
2319
2320
2321         rcu_assign_pointer(event->data, data);
2322 }
2323
2324 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2325 {
2326         struct perf_mmap_data *data;
2327
2328         data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2329         perf_mmap_data_free(data);
2330         kfree(data);
2331 }
2332
2333 static void perf_mmap_data_release(struct perf_event *event)
2334 {
2335         struct perf_mmap_data *data = event->data;
2336
2337         WARN_ON(atomic_read(&event->mmap_count));
2338
2339         rcu_assign_pointer(event->data, NULL);
2340         call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2341 }
2342
2343 static void perf_mmap_open(struct vm_area_struct *vma)
2344 {
2345         struct perf_event *event = vma->vm_file->private_data;
2346
2347         atomic_inc(&event->mmap_count);
2348 }
2349
2350 static void perf_mmap_close(struct vm_area_struct *vma)
2351 {
2352         struct perf_event *event = vma->vm_file->private_data;
2353
2354         WARN_ON_ONCE(event->ctx->parent_ctx);
2355         if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2356                 unsigned long size = perf_data_size(event->data);
2357                 struct user_struct *user = current_user();
2358
2359                 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2360                 vma->vm_mm->locked_vm -= event->data->nr_locked;
2361                 perf_mmap_data_release(event);
2362                 mutex_unlock(&event->mmap_mutex);
2363         }
2364 }
2365
2366 static const struct vm_operations_struct perf_mmap_vmops = {
2367         .open           = perf_mmap_open,
2368         .close          = perf_mmap_close,
2369         .fault          = perf_mmap_fault,
2370         .page_mkwrite   = perf_mmap_fault,
2371 };
2372
2373 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2374 {
2375         struct perf_event *event = file->private_data;
2376         unsigned long user_locked, user_lock_limit;
2377         struct user_struct *user = current_user();
2378         unsigned long locked, lock_limit;
2379         struct perf_mmap_data *data;
2380         unsigned long vma_size;
2381         unsigned long nr_pages;
2382         long user_extra, extra;
2383         int ret = 0;
2384
2385         if (!(vma->vm_flags & VM_SHARED))
2386                 return -EINVAL;
2387
2388         vma_size = vma->vm_end - vma->vm_start;
2389         nr_pages = (vma_size / PAGE_SIZE) - 1;
2390
2391         /*
2392          * If we have data pages ensure they're a power-of-two number, so we
2393          * can do bitmasks instead of modulo.
2394          */
2395         if (nr_pages != 0 && !is_power_of_2(nr_pages))
2396                 return -EINVAL;
2397
2398         if (vma_size != PAGE_SIZE * (1 + nr_pages))
2399                 return -EINVAL;
2400
2401         if (vma->vm_pgoff != 0)
2402                 return -EINVAL;
2403
2404         WARN_ON_ONCE(event->ctx->parent_ctx);
2405         mutex_lock(&event->mmap_mutex);
2406         if (event->output) {
2407                 ret = -EINVAL;
2408                 goto unlock;
2409         }
2410
2411         if (atomic_inc_not_zero(&event->mmap_count)) {
2412                 if (nr_pages != event->data->nr_pages)
2413                         ret = -EINVAL;
2414                 goto unlock;
2415         }
2416
2417         user_extra = nr_pages + 1;
2418         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2419
2420         /*
2421          * Increase the limit linearly with more CPUs:
2422          */
2423         user_lock_limit *= num_online_cpus();
2424
2425         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2426
2427         extra = 0;
2428         if (user_locked > user_lock_limit)
2429                 extra = user_locked - user_lock_limit;
2430
2431         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2432         lock_limit >>= PAGE_SHIFT;
2433         locked = vma->vm_mm->locked_vm + extra;
2434
2435         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2436                 !capable(CAP_IPC_LOCK)) {
2437                 ret = -EPERM;
2438                 goto unlock;
2439         }
2440
2441         WARN_ON(event->data);
2442
2443         data = perf_mmap_data_alloc(event, nr_pages);
2444         ret = -ENOMEM;
2445         if (!data)
2446                 goto unlock;
2447
2448         ret = 0;
2449         perf_mmap_data_init(event, data);
2450
2451         atomic_set(&event->mmap_count, 1);
2452         atomic_long_add(user_extra, &user->locked_vm);
2453         vma->vm_mm->locked_vm += extra;
2454         event->data->nr_locked = extra;
2455         if (vma->vm_flags & VM_WRITE)
2456                 event->data->writable = 1;
2457
2458 unlock:
2459         mutex_unlock(&event->mmap_mutex);
2460
2461         vma->vm_flags |= VM_RESERVED;
2462         vma->vm_ops = &perf_mmap_vmops;
2463
2464         return ret;
2465 }
2466
2467 static int perf_fasync(int fd, struct file *filp, int on)
2468 {
2469         struct inode *inode = filp->f_path.dentry->d_inode;
2470         struct perf_event *event = filp->private_data;
2471         int retval;
2472
2473         mutex_lock(&inode->i_mutex);
2474         retval = fasync_helper(fd, filp, on, &event->fasync);
2475         mutex_unlock(&inode->i_mutex);
2476
2477         if (retval < 0)
2478                 return retval;
2479
2480         return 0;
2481 }
2482
2483 static const struct file_operations perf_fops = {
2484         .release                = perf_release,
2485         .read                   = perf_read,
2486         .poll                   = perf_poll,
2487         .unlocked_ioctl         = perf_ioctl,
2488         .compat_ioctl           = perf_ioctl,
2489         .mmap                   = perf_mmap,
2490         .fasync                 = perf_fasync,
2491 };
2492
2493 /*
2494  * Perf event wakeup
2495  *
2496  * If there's data, ensure we set the poll() state and publish everything
2497  * to user-space before waking everybody up.
2498  */
2499
2500 void perf_event_wakeup(struct perf_event *event)
2501 {
2502         wake_up_all(&event->waitq);
2503
2504         if (event->pending_kill) {
2505                 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2506                 event->pending_kill = 0;
2507         }
2508 }
2509
2510 /*
2511  * Pending wakeups
2512  *
2513  * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2514  *
2515  * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2516  * single linked list and use cmpxchg() to add entries lockless.
2517  */
2518
2519 static void perf_pending_event(struct perf_pending_entry *entry)
2520 {
2521         struct perf_event *event = container_of(entry,
2522                         struct perf_event, pending);
2523
2524         if (event->pending_disable) {
2525                 event->pending_disable = 0;
2526                 __perf_event_disable(event);
2527         }
2528
2529         if (event->pending_wakeup) {
2530                 event->pending_wakeup = 0;
2531                 perf_event_wakeup(event);
2532         }
2533 }
2534
2535 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2536
2537 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2538         PENDING_TAIL,
2539 };
2540
2541 static void perf_pending_queue(struct perf_pending_entry *entry,
2542                                void (*func)(struct perf_pending_entry *))
2543 {
2544         struct perf_pending_entry **head;
2545
2546         if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2547                 return;
2548
2549         entry->func = func;
2550
2551         head = &get_cpu_var(perf_pending_head);
2552
2553         do {
2554                 entry->next = *head;
2555         } while (cmpxchg(head, entry->next, entry) != entry->next);
2556
2557         set_perf_event_pending();
2558
2559         put_cpu_var(perf_pending_head);
2560 }
2561
2562 static int __perf_pending_run(void)
2563 {
2564         struct perf_pending_entry *list;
2565         int nr = 0;
2566
2567         list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2568         while (list != PENDING_TAIL) {
2569                 void (*func)(struct perf_pending_entry *);
2570                 struct perf_pending_entry *entry = list;
2571
2572                 list = list->next;
2573
2574                 func = entry->func;
2575                 entry->next = NULL;
2576                 /*
2577                  * Ensure we observe the unqueue before we issue the wakeup,
2578                  * so that we won't be waiting forever.
2579                  * -- see perf_not_pending().
2580                  */
2581                 smp_wmb();
2582
2583                 func(entry);
2584                 nr++;
2585         }
2586
2587         return nr;
2588 }
2589
2590 static inline int perf_not_pending(struct perf_event *event)
2591 {
2592         /*
2593          * If we flush on whatever cpu we run, there is a chance we don't
2594          * need to wait.
2595          */
2596         get_cpu();
2597         __perf_pending_run();
2598         put_cpu();
2599
2600         /*
2601          * Ensure we see the proper queue state before going to sleep
2602          * so that we do not miss the wakeup. -- see perf_pending_handle()
2603          */
2604         smp_rmb();
2605         return event->pending.next == NULL;
2606 }
2607
2608 static void perf_pending_sync(struct perf_event *event)
2609 {
2610         wait_event(event->waitq, perf_not_pending(event));
2611 }
2612
2613 void perf_event_do_pending(void)
2614 {
2615         __perf_pending_run();
2616 }
2617
2618 /*
2619  * Callchain support -- arch specific
2620  */
2621
2622 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2623 {
2624         return NULL;
2625 }
2626
2627 /*
2628  * Output
2629  */
2630 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2631                               unsigned long offset, unsigned long head)
2632 {
2633         unsigned long mask;
2634
2635         if (!data->writable)
2636                 return true;
2637
2638         mask = perf_data_size(data) - 1;
2639
2640         offset = (offset - tail) & mask;
2641         head   = (head   - tail) & mask;
2642
2643         if ((int)(head - offset) < 0)
2644                 return false;
2645
2646         return true;
2647 }
2648
2649 static void perf_output_wakeup(struct perf_output_handle *handle)
2650 {
2651         atomic_set(&handle->data->poll, POLL_IN);
2652
2653         if (handle->nmi) {
2654                 handle->event->pending_wakeup = 1;
2655                 perf_pending_queue(&handle->event->pending,
2656                                    perf_pending_event);
2657         } else
2658                 perf_event_wakeup(handle->event);
2659 }
2660
2661 /*
2662  * Curious locking construct.
2663  *
2664  * We need to ensure a later event_id doesn't publish a head when a former
2665  * event_id isn't done writing. However since we need to deal with NMIs we
2666  * cannot fully serialize things.
2667  *
2668  * What we do is serialize between CPUs so we only have to deal with NMI
2669  * nesting on a single CPU.
2670  *
2671  * We only publish the head (and generate a wakeup) when the outer-most
2672  * event_id completes.
2673  */
2674 static void perf_output_lock(struct perf_output_handle *handle)
2675 {
2676         struct perf_mmap_data *data = handle->data;
2677         int cur, cpu = get_cpu();
2678
2679         handle->locked = 0;
2680
2681         for (;;) {
2682                 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2683                 if (cur == -1) {
2684                         handle->locked = 1;
2685                         break;
2686                 }
2687                 if (cur == cpu)
2688                         break;
2689
2690                 cpu_relax();
2691         }
2692 }
2693
2694 static void perf_output_unlock(struct perf_output_handle *handle)
2695 {
2696         struct perf_mmap_data *data = handle->data;
2697         unsigned long head;
2698         int cpu;
2699
2700         data->done_head = data->head;
2701
2702         if (!handle->locked)
2703                 goto out;
2704
2705 again:
2706         /*
2707          * The xchg implies a full barrier that ensures all writes are done
2708          * before we publish the new head, matched by a rmb() in userspace when
2709          * reading this position.
2710          */
2711         while ((head = atomic_long_xchg(&data->done_head, 0)))
2712                 data->user_page->data_head = head;
2713
2714         /*
2715          * NMI can happen here, which means we can miss a done_head update.
2716          */
2717
2718         cpu = atomic_xchg(&data->lock, -1);
2719         WARN_ON_ONCE(cpu != smp_processor_id());
2720
2721         /*
2722          * Therefore we have to validate we did not indeed do so.
2723          */
2724         if (unlikely(atomic_long_read(&data->done_head))) {
2725                 /*
2726                  * Since we had it locked, we can lock it again.
2727                  */
2728                 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2729                         cpu_relax();
2730
2731                 goto again;
2732         }
2733
2734         if (atomic_xchg(&data->wakeup, 0))
2735                 perf_output_wakeup(handle);
2736 out:
2737         put_cpu();
2738 }
2739
2740 void perf_output_copy(struct perf_output_handle *handle,
2741                       const void *buf, unsigned int len)
2742 {
2743         unsigned int pages_mask;
2744         unsigned long offset;
2745         unsigned int size;
2746         void **pages;
2747
2748         offset          = handle->offset;
2749         pages_mask      = handle->data->nr_pages - 1;
2750         pages           = handle->data->data_pages;
2751
2752         do {
2753                 unsigned long page_offset;
2754                 unsigned long page_size;
2755                 int nr;
2756
2757                 nr          = (offset >> PAGE_SHIFT) & pages_mask;
2758                 page_size   = 1UL << (handle->data->data_order + PAGE_SHIFT);
2759                 page_offset = offset & (page_size - 1);
2760                 size        = min_t(unsigned int, page_size - page_offset, len);
2761
2762                 memcpy(pages[nr] + page_offset, buf, size);
2763
2764                 len         -= size;
2765                 buf         += size;
2766                 offset      += size;
2767         } while (len);
2768
2769         handle->offset = offset;
2770
2771         /*
2772          * Check we didn't copy past our reservation window, taking the
2773          * possible unsigned int wrap into account.
2774          */
2775         WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2776 }
2777
2778 int perf_output_begin(struct perf_output_handle *handle,
2779                       struct perf_event *event, unsigned int size,
2780                       int nmi, int sample)
2781 {
2782         struct perf_event *output_event;
2783         struct perf_mmap_data *data;
2784         unsigned long tail, offset, head;
2785         int have_lost;
2786         struct {
2787                 struct perf_event_header header;
2788                 u64                      id;
2789                 u64                      lost;
2790         } lost_event;
2791
2792         rcu_read_lock();
2793         /*
2794          * For inherited events we send all the output towards the parent.
2795          */
2796         if (event->parent)
2797                 event = event->parent;
2798
2799         output_event = rcu_dereference(event->output);
2800         if (output_event)
2801                 event = output_event;
2802
2803         data = rcu_dereference(event->data);
2804         if (!data)
2805                 goto out;
2806
2807         handle->data    = data;
2808         handle->event   = event;
2809         handle->nmi     = nmi;
2810         handle->sample  = sample;
2811
2812         if (!data->nr_pages)
2813                 goto fail;
2814
2815         have_lost = atomic_read(&data->lost);
2816         if (have_lost)
2817                 size += sizeof(lost_event);
2818
2819         perf_output_lock(handle);
2820
2821         do {
2822                 /*
2823                  * Userspace could choose to issue a mb() before updating the
2824                  * tail pointer. So that all reads will be completed before the
2825                  * write is issued.
2826                  */
2827                 tail = ACCESS_ONCE(data->user_page->data_tail);
2828                 smp_rmb();
2829                 offset = head = atomic_long_read(&data->head);
2830                 head += size;
2831                 if (unlikely(!perf_output_space(data, tail, offset, head)))
2832                         goto fail;
2833         } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2834
2835         handle->offset  = offset;
2836         handle->head    = head;
2837
2838         if (head - tail > data->watermark)
2839                 atomic_set(&data->wakeup, 1);
2840
2841         if (have_lost) {
2842                 lost_event.header.type = PERF_RECORD_LOST;
2843                 lost_event.header.misc = 0;
2844                 lost_event.header.size = sizeof(lost_event);
2845                 lost_event.id          = event->id;
2846                 lost_event.lost        = atomic_xchg(&data->lost, 0);
2847
2848                 perf_output_put(handle, lost_event);
2849         }
2850
2851         return 0;
2852
2853 fail:
2854         atomic_inc(&data->lost);
2855         perf_output_unlock(handle);
2856 out:
2857         rcu_read_unlock();
2858
2859         return -ENOSPC;
2860 }
2861
2862 void perf_output_end(struct perf_output_handle *handle)
2863 {
2864         struct perf_event *event = handle->event;
2865         struct perf_mmap_data *data = handle->data;
2866
2867         int wakeup_events = event->attr.wakeup_events;
2868
2869         if (handle->sample && wakeup_events) {
2870                 int events = atomic_inc_return(&data->events);
2871                 if (events >= wakeup_events) {
2872                         atomic_sub(wakeup_events, &data->events);
2873                         atomic_set(&data->wakeup, 1);
2874                 }
2875         }
2876
2877         perf_output_unlock(handle);
2878         rcu_read_unlock();
2879 }
2880
2881 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
2882 {
2883         /*
2884          * only top level events have the pid namespace they were created in
2885          */
2886         if (event->parent)
2887                 event = event->parent;
2888
2889         return task_tgid_nr_ns(p, event->ns);
2890 }
2891
2892 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
2893 {
2894         /*
2895          * only top level events have the pid namespace they were created in
2896          */
2897         if (event->parent)
2898                 event = event->parent;
2899
2900         return task_pid_nr_ns(p, event->ns);
2901 }
2902
2903 static void perf_output_read_one(struct perf_output_handle *handle,
2904                                  struct perf_event *event)
2905 {
2906         u64 read_format = event->attr.read_format;
2907         u64 values[4];
2908         int n = 0;
2909
2910         values[n++] = atomic64_read(&event->count);
2911         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2912                 values[n++] = event->total_time_enabled +
2913                         atomic64_read(&event->child_total_time_enabled);
2914         }
2915         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2916                 values[n++] = event->total_time_running +
2917                         atomic64_read(&event->child_total_time_running);
2918         }
2919         if (read_format & PERF_FORMAT_ID)
2920                 values[n++] = primary_event_id(event);
2921
2922         perf_output_copy(handle, values, n * sizeof(u64));
2923 }
2924
2925 /*
2926  * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
2927  */
2928 static void perf_output_read_group(struct perf_output_handle *handle,
2929                             struct perf_event *event)
2930 {
2931         struct perf_event *leader = event->group_leader, *sub;
2932         u64 read_format = event->attr.read_format;
2933         u64 values[5];
2934         int n = 0;
2935
2936         values[n++] = 1 + leader->nr_siblings;
2937
2938         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2939                 values[n++] = leader->total_time_enabled;
2940
2941         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2942                 values[n++] = leader->total_time_running;
2943
2944         if (leader != event)
2945                 leader->pmu->read(leader);
2946
2947         values[n++] = atomic64_read(&leader->count);
2948         if (read_format & PERF_FORMAT_ID)
2949                 values[n++] = primary_event_id(leader);
2950
2951         perf_output_copy(handle, values, n * sizeof(u64));
2952
2953         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2954                 n = 0;
2955
2956                 if (sub != event)
2957                         sub->pmu->read(sub);
2958
2959                 values[n++] = atomic64_read(&sub->count);
2960                 if (read_format & PERF_FORMAT_ID)
2961                         values[n++] = primary_event_id(sub);
2962
2963                 perf_output_copy(handle, values, n * sizeof(u64));
2964         }
2965 }
2966
2967 static void perf_output_read(struct perf_output_handle *handle,
2968                              struct perf_event *event)
2969 {
2970         if (event->attr.read_format & PERF_FORMAT_GROUP)
2971                 perf_output_read_group(handle, event);
2972         else
2973                 perf_output_read_one(handle, event);
2974 }
2975
2976 void perf_output_sample(struct perf_output_handle *handle,
2977                         struct perf_event_header *header,
2978                         struct perf_sample_data *data,
2979                         struct perf_event *event)
2980 {
2981         u64 sample_type = data->type;
2982
2983         perf_output_put(handle, *header);
2984
2985         if (sample_type & PERF_SAMPLE_IP)
2986                 perf_output_put(handle, data->ip);
2987
2988         if (sample_type & PERF_SAMPLE_TID)
2989                 perf_output_put(handle, data->tid_entry);
2990
2991         if (sample_type & PERF_SAMPLE_TIME)
2992                 perf_output_put(handle, data->time);
2993
2994         if (sample_type & PERF_SAMPLE_ADDR)
2995                 perf_output_put(handle, data->addr);
2996
2997         if (sample_type & PERF_SAMPLE_ID)
2998                 perf_output_put(handle, data->id);
2999
3000         if (sample_type & PERF_SAMPLE_STREAM_ID)
3001                 perf_output_put(handle, data->stream_id);
3002
3003         if (sample_type & PERF_SAMPLE_CPU)
3004                 perf_output_put(handle, data->cpu_entry);
3005
3006         if (sample_type & PERF_SAMPLE_PERIOD)
3007                 perf_output_put(handle, data->period);
3008
3009         if (sample_type & PERF_SAMPLE_READ)
3010                 perf_output_read(handle, event);
3011
3012         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3013                 if (data->callchain) {
3014                         int size = 1;
3015
3016                         if (data->callchain)
3017                                 size += data->callchain->nr;
3018
3019                         size *= sizeof(u64);
3020
3021                         perf_output_copy(handle, data->callchain, size);
3022                 } else {
3023                         u64 nr = 0;
3024                         perf_output_put(handle, nr);
3025                 }
3026         }
3027
3028         if (sample_type & PERF_SAMPLE_RAW) {
3029                 if (data->raw) {
3030                         perf_output_put(handle, data->raw->size);
3031                         perf_output_copy(handle, data->raw->data,
3032                                          data->raw->size);
3033                 } else {
3034                         struct {
3035                                 u32     size;
3036                                 u32     data;
3037                         } raw = {
3038                                 .size = sizeof(u32),
3039                                 .data = 0,
3040                         };
3041                         perf_output_put(handle, raw);
3042                 }
3043         }
3044 }
3045
3046 void perf_prepare_sample(struct perf_event_header *header,
3047                          struct perf_sample_data *data,
3048                          struct perf_event *event,
3049                          struct pt_regs *regs)
3050 {
3051         u64 sample_type = event->attr.sample_type;
3052
3053         data->type = sample_type;
3054
3055         header->type = PERF_RECORD_SAMPLE;
3056         header->size = sizeof(*header);
3057
3058         header->misc = 0;
3059         header->misc |= perf_misc_flags(regs);
3060
3061         if (sample_type & PERF_SAMPLE_IP) {
3062                 data->ip = perf_instruction_pointer(regs);
3063
3064                 header->size += sizeof(data->ip);
3065         }
3066
3067         if (sample_type & PERF_SAMPLE_TID) {
3068                 /* namespace issues */
3069                 data->tid_entry.pid = perf_event_pid(event, current);
3070                 data->tid_entry.tid = perf_event_tid(event, current);
3071
3072                 header->size += sizeof(data->tid_entry);
3073         }
3074
3075         if (sample_type & PERF_SAMPLE_TIME) {
3076                 data->time = perf_clock();
3077
3078                 header->size += sizeof(data->time);
3079         }
3080
3081         if (sample_type & PERF_SAMPLE_ADDR)
3082                 header->size += sizeof(data->addr);
3083
3084         if (sample_type & PERF_SAMPLE_ID) {
3085                 data->id = primary_event_id(event);
3086
3087                 header->size += sizeof(data->id);
3088         }
3089
3090         if (sample_type & PERF_SAMPLE_STREAM_ID) {
3091                 data->stream_id = event->id;
3092
3093                 header->size += sizeof(data->stream_id);
3094         }
3095
3096         if (sample_type & PERF_SAMPLE_CPU) {
3097                 data->cpu_entry.cpu             = raw_smp_processor_id();
3098                 data->cpu_entry.reserved        = 0;
3099
3100                 header->size += sizeof(data->cpu_entry);
3101         }
3102
3103         if (sample_type & PERF_SAMPLE_PERIOD)
3104                 header->size += sizeof(data->period);
3105
3106         if (sample_type & PERF_SAMPLE_READ)
3107                 header->size += perf_event_read_size(event);
3108
3109         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3110                 int size = 1;
3111
3112                 data->callchain = perf_callchain(regs);
3113
3114                 if (data->callchain)
3115                         size += data->callchain->nr;
3116
3117                 header->size += size * sizeof(u64);
3118         }
3119
3120         if (sample_type & PERF_SAMPLE_RAW) {
3121                 int size = sizeof(u32);
3122
3123                 if (data->raw)
3124                         size += data->raw->size;
3125                 else
3126                         size += sizeof(u32);
3127
3128                 WARN_ON_ONCE(size & (sizeof(u64)-1));
3129                 header->size += size;
3130         }
3131 }
3132
3133 static void perf_event_output(struct perf_event *event, int nmi,
3134                                 struct perf_sample_data *data,
3135                                 struct pt_regs *regs)
3136 {
3137         struct perf_output_handle handle;
3138         struct perf_event_header header;
3139
3140         perf_prepare_sample(&header, data, event, regs);
3141
3142         if (perf_output_begin(&handle, event, header.size, nmi, 1))
3143                 return;
3144
3145         perf_output_sample(&handle, &header, data, event);
3146
3147         perf_output_end(&handle);
3148 }
3149
3150 /*
3151  * read event_id
3152  */
3153
3154 struct perf_read_event {
3155         struct perf_event_header        header;
3156
3157         u32                             pid;
3158         u32                             tid;
3159 };
3160
3161 static void
3162 perf_event_read_event(struct perf_event *event,
3163                         struct task_struct *task)
3164 {
3165         struct perf_output_handle handle;
3166         struct perf_read_event read_event = {
3167                 .header = {
3168                         .type = PERF_RECORD_READ,
3169                         .misc = 0,
3170                         .size = sizeof(read_event) + perf_event_read_size(event),
3171                 },
3172                 .pid = perf_event_pid(event, task),
3173                 .tid = perf_event_tid(event, task),
3174         };
3175         int ret;
3176
3177         ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3178         if (ret)
3179                 return;
3180
3181         perf_output_put(&handle, read_event);
3182         perf_output_read(&handle, event);
3183
3184         perf_output_end(&handle);
3185 }
3186
3187 /*
3188  * task tracking -- fork/exit
3189  *
3190  * enabled by: attr.comm | attr.mmap | attr.task
3191  */
3192
3193 struct perf_task_event {
3194         struct task_struct              *task;
3195         struct perf_event_context       *task_ctx;
3196
3197         struct {
3198                 struct perf_event_header        header;
3199
3200                 u32                             pid;
3201                 u32                             ppid;
3202                 u32                             tid;
3203                 u32                             ptid;
3204                 u64                             time;
3205         } event_id;
3206 };
3207
3208 static void perf_event_task_output(struct perf_event *event,
3209                                      struct perf_task_event *task_event)
3210 {
3211         struct perf_output_handle handle;
3212         int size;
3213         struct task_struct *task = task_event->task;
3214         int ret;
3215
3216         size  = task_event->event_id.header.size;
3217         ret = perf_output_begin(&handle, event, size, 0, 0);
3218
3219         if (ret)
3220                 return;
3221
3222         task_event->event_id.pid = perf_event_pid(event, task);
3223         task_event->event_id.ppid = perf_event_pid(event, current);
3224
3225         task_event->event_id.tid = perf_event_tid(event, task);
3226         task_event->event_id.ptid = perf_event_tid(event, current);
3227
3228         task_event->event_id.time = perf_clock();
3229
3230         perf_output_put(&handle, task_event->event_id);
3231
3232         perf_output_end(&handle);
3233 }
3234
3235 static int perf_event_task_match(struct perf_event *event)
3236 {
3237         if (event->attr.comm || event->attr.mmap || event->attr.task)
3238                 return 1;
3239
3240         return 0;
3241 }
3242
3243 static void perf_event_task_ctx(struct perf_event_context *ctx,
3244                                   struct perf_task_event *task_event)
3245 {
3246         struct perf_event *event;
3247
3248         if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3249                 return;
3250
3251         rcu_read_lock();
3252         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3253                 if (perf_event_task_match(event))
3254                         perf_event_task_output(event, task_event);
3255         }
3256         rcu_read_unlock();
3257 }
3258
3259 static void perf_event_task_event(struct perf_task_event *task_event)
3260 {
3261         struct perf_cpu_context *cpuctx;
3262         struct perf_event_context *ctx = task_event->task_ctx;
3263
3264         cpuctx = &get_cpu_var(perf_cpu_context);
3265         perf_event_task_ctx(&cpuctx->ctx, task_event);
3266         put_cpu_var(perf_cpu_context);
3267
3268         rcu_read_lock();
3269         if (!ctx)
3270                 ctx = rcu_dereference(task_event->task->perf_event_ctxp);
3271         if (ctx)
3272                 perf_event_task_ctx(ctx, task_event);
3273         rcu_read_unlock();
3274 }
3275
3276 static void perf_event_task(struct task_struct *task,
3277                               struct perf_event_context *task_ctx,
3278                               int new)
3279 {
3280         struct perf_task_event task_event;
3281
3282         if (!atomic_read(&nr_comm_events) &&
3283             !atomic_read(&nr_mmap_events) &&
3284             !atomic_read(&nr_task_events))
3285                 return;
3286
3287         task_event = (struct perf_task_event){
3288                 .task     = task,
3289                 .task_ctx = task_ctx,
3290                 .event_id    = {
3291                         .header = {
3292                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3293                                 .misc = 0,
3294                                 .size = sizeof(task_event.event_id),
3295                         },
3296                         /* .pid  */
3297                         /* .ppid */
3298                         /* .tid  */
3299                         /* .ptid */
3300                 },
3301         };
3302
3303         perf_event_task_event(&task_event);
3304 }
3305
3306 void perf_event_fork(struct task_struct *task)
3307 {
3308         perf_event_task(task, NULL, 1);
3309 }
3310
3311 /*
3312  * comm tracking
3313  */
3314
3315 struct perf_comm_event {
3316         struct task_struct      *task;
3317         char                    *comm;
3318         int                     comm_size;
3319
3320         struct {
3321                 struct perf_event_header        header;
3322
3323                 u32                             pid;
3324                 u32                             tid;
3325         } event_id;
3326 };
3327
3328 static void perf_event_comm_output(struct perf_event *event,
3329                                      struct perf_comm_event *comm_event)
3330 {
3331         struct perf_output_handle handle;
3332         int size = comm_event->event_id.header.size;
3333         int ret = perf_output_begin(&handle, event, size, 0, 0);
3334
3335         if (ret)
3336                 return;
3337
3338         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3339         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3340
3341         perf_output_put(&handle, comm_event->event_id);
3342         perf_output_copy(&handle, comm_event->comm,
3343                                    comm_event->comm_size);
3344         perf_output_end(&handle);
3345 }
3346
3347 static int perf_event_comm_match(struct perf_event *event)
3348 {
3349         if (event->attr.comm)
3350                 return 1;
3351
3352         return 0;
3353 }
3354
3355 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3356                                   struct perf_comm_event *comm_event)
3357 {
3358         struct perf_event *event;
3359
3360         if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3361                 return;
3362
3363         rcu_read_lock();
3364         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3365                 if (perf_event_comm_match(event))
3366                         perf_event_comm_output(event, comm_event);
3367         }
3368         rcu_read_unlock();
3369 }
3370
3371 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3372 {
3373         struct perf_cpu_context *cpuctx;
3374         struct perf_event_context *ctx;
3375         unsigned int size;
3376         char comm[TASK_COMM_LEN];
3377
3378         memset(comm, 0, sizeof(comm));
3379         strncpy(comm, comm_event->task->comm, sizeof(comm));
3380         size = ALIGN(strlen(comm)+1, sizeof(u64));
3381
3382         comm_event->comm = comm;
3383         comm_event->comm_size = size;
3384
3385         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3386
3387         cpuctx = &get_cpu_var(perf_cpu_context);
3388         perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3389         put_cpu_var(perf_cpu_context);
3390
3391         rcu_read_lock();
3392         /*
3393          * doesn't really matter which of the child contexts the
3394          * events ends up in.
3395          */
3396         ctx = rcu_dereference(current->perf_event_ctxp);
3397         if (ctx)
3398                 perf_event_comm_ctx(ctx, comm_event);
3399         rcu_read_unlock();
3400 }
3401
3402 void perf_event_comm(struct task_struct *task)
3403 {
3404         struct perf_comm_event comm_event;
3405
3406         if (task->perf_event_ctxp)
3407                 perf_event_enable_on_exec(task);
3408
3409         if (!atomic_read(&nr_comm_events))
3410                 return;
3411
3412         comm_event = (struct perf_comm_event){
3413                 .task   = task,
3414                 /* .comm      */
3415                 /* .comm_size */
3416                 .event_id  = {
3417                         .header = {
3418                                 .type = PERF_RECORD_COMM,
3419                                 .misc = 0,
3420                                 /* .size */
3421                         },
3422                         /* .pid */
3423                         /* .tid */
3424                 },
3425         };
3426
3427         perf_event_comm_event(&comm_event);
3428 }
3429
3430 /*
3431  * mmap tracking
3432  */
3433
3434 struct perf_mmap_event {
3435         struct vm_area_struct   *vma;
3436
3437         const char              *file_name;
3438         int                     file_size;
3439
3440         struct {
3441                 struct perf_event_header        header;
3442
3443                 u32                             pid;
3444                 u32                             tid;
3445                 u64                             start;
3446                 u64                             len;
3447                 u64                             pgoff;
3448         } event_id;
3449 };
3450
3451 static void perf_event_mmap_output(struct perf_event *event,
3452                                      struct perf_mmap_event *mmap_event)
3453 {
3454         struct perf_output_handle handle;
3455         int size = mmap_event->event_id.header.size;
3456         int ret = perf_output_begin(&handle, event, size, 0, 0);
3457
3458         if (ret)
3459                 return;
3460
3461         mmap_event->event_id.pid = perf_event_pid(event, current);
3462         mmap_event->event_id.tid = perf_event_tid(event, current);
3463
3464         perf_output_put(&handle, mmap_event->event_id);
3465         perf_output_copy(&handle, mmap_event->file_name,
3466                                    mmap_event->file_size);
3467         perf_output_end(&handle);
3468 }
3469
3470 static int perf_event_mmap_match(struct perf_event *event,
3471                                    struct perf_mmap_event *mmap_event)
3472 {
3473         if (event->attr.mmap)
3474                 return 1;
3475
3476         return 0;
3477 }
3478
3479 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3480                                   struct perf_mmap_event *mmap_event)
3481 {
3482         struct perf_event *event;
3483
3484         if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3485                 return;
3486
3487         rcu_read_lock();
3488         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3489                 if (perf_event_mmap_match(event, mmap_event))
3490                         perf_event_mmap_output(event, mmap_event);
3491         }
3492         rcu_read_unlock();
3493 }
3494
3495 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3496 {
3497         struct perf_cpu_context *cpuctx;
3498         struct perf_event_context *ctx;
3499         struct vm_area_struct *vma = mmap_event->vma;
3500         struct file *file = vma->vm_file;
3501         unsigned int size;
3502         char tmp[16];
3503         char *buf = NULL;
3504         const char *name;
3505
3506         memset(tmp, 0, sizeof(tmp));
3507
3508         if (file) {
3509                 /*
3510                  * d_path works from the end of the buffer backwards, so we
3511                  * need to add enough zero bytes after the string to handle
3512                  * the 64bit alignment we do later.
3513                  */
3514                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3515                 if (!buf) {
3516                         name = strncpy(tmp, "//enomem", sizeof(tmp));
3517                         goto got_name;
3518                 }
3519                 name = d_path(&file->f_path, buf, PATH_MAX);
3520                 if (IS_ERR(name)) {
3521                         name = strncpy(tmp, "//toolong", sizeof(tmp));
3522                         goto got_name;
3523                 }
3524         } else {
3525                 if (arch_vma_name(mmap_event->vma)) {
3526                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3527                                        sizeof(tmp));
3528                         goto got_name;
3529                 }
3530
3531                 if (!vma->vm_mm) {
3532                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
3533                         goto got_name;
3534                 }
3535
3536                 name = strncpy(tmp, "//anon", sizeof(tmp));
3537                 goto got_name;
3538         }
3539
3540 got_name:
3541         size = ALIGN(strlen(name)+1, sizeof(u64));
3542
3543         mmap_event->file_name = name;
3544         mmap_event->file_size = size;
3545
3546         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3547
3548         cpuctx = &get_cpu_var(perf_cpu_context);
3549         perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3550         put_cpu_var(perf_cpu_context);
3551
3552         rcu_read_lock();
3553         /*
3554          * doesn't really matter which of the child contexts the
3555          * events ends up in.
3556          */
3557         ctx = rcu_dereference(current->perf_event_ctxp);
3558         if (ctx)
3559                 perf_event_mmap_ctx(ctx, mmap_event);
3560         rcu_read_unlock();
3561
3562         kfree(buf);
3563 }
3564
3565 void __perf_event_mmap(struct vm_area_struct *vma)
3566 {
3567         struct perf_mmap_event mmap_event;
3568
3569         if (!atomic_read(&nr_mmap_events))
3570                 return;
3571
3572         mmap_event = (struct perf_mmap_event){
3573                 .vma    = vma,
3574                 /* .file_name */
3575                 /* .file_size */
3576                 .event_id  = {
3577                         .header = {
3578                                 .type = PERF_RECORD_MMAP,
3579                                 .misc = 0,
3580                                 /* .size */
3581                         },
3582                         /* .pid */
3583                         /* .tid */
3584                         .start  = vma->vm_start,
3585                         .len    = vma->vm_end - vma->vm_start,
3586                         .pgoff  = vma->vm_pgoff,
3587                 },
3588         };
3589
3590         perf_event_mmap_event(&mmap_event);
3591 }
3592
3593 /*
3594  * IRQ throttle logging
3595  */
3596
3597 static void perf_log_throttle(struct perf_event *event, int enable)
3598 {
3599         struct perf_output_handle handle;
3600         int ret;
3601
3602         struct {
3603                 struct perf_event_header        header;
3604                 u64                             time;
3605                 u64                             id;
3606                 u64                             stream_id;
3607         } throttle_event = {
3608                 .header = {
3609                         .type = PERF_RECORD_THROTTLE,
3610                         .misc = 0,
3611                         .size = sizeof(throttle_event),
3612                 },
3613                 .time           = perf_clock(),
3614                 .id             = primary_event_id(event),
3615                 .stream_id      = event->id,
3616         };
3617
3618         if (enable)
3619                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3620
3621         ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3622         if (ret)
3623                 return;
3624
3625         perf_output_put(&handle, throttle_event);
3626         perf_output_end(&handle);
3627 }
3628
3629 /*
3630  * Generic event overflow handling, sampling.
3631  */
3632
3633 static int __perf_event_overflow(struct perf_event *event, int nmi,
3634                                    int throttle, struct perf_sample_data *data,
3635                                    struct pt_regs *regs)
3636 {
3637         int events = atomic_read(&event->event_limit);
3638         struct hw_perf_event *hwc = &event->hw;
3639         int ret = 0;
3640
3641         throttle = (throttle && event->pmu->unthrottle != NULL);
3642
3643         if (!throttle) {
3644                 hwc->interrupts++;
3645         } else {
3646                 if (hwc->interrupts != MAX_INTERRUPTS) {
3647                         hwc->interrupts++;
3648                         if (HZ * hwc->interrupts >
3649                                         (u64)sysctl_perf_event_sample_rate) {
3650                                 hwc->interrupts = MAX_INTERRUPTS;
3651                                 perf_log_throttle(event, 0);
3652                                 ret = 1;
3653                         }
3654                 } else {
3655                         /*
3656                          * Keep re-disabling events even though on the previous
3657                          * pass we disabled it - just in case we raced with a
3658                          * sched-in and the event got enabled again:
3659                          */
3660                         ret = 1;
3661                 }
3662         }
3663
3664         if (event->attr.freq) {
3665                 u64 now = perf_clock();
3666                 s64 delta = now - hwc->freq_stamp;
3667
3668                 hwc->freq_stamp = now;
3669
3670                 if (delta > 0 && delta < TICK_NSEC)
3671                         perf_adjust_period(event, NSEC_PER_SEC / (int)delta);
3672         }
3673
3674         /*
3675          * XXX event_limit might not quite work as expected on inherited
3676          * events
3677          */
3678
3679         event->pending_kill = POLL_IN;
3680         if (events && atomic_dec_and_test(&event->event_limit)) {
3681                 ret = 1;
3682                 event->pending_kill = POLL_HUP;
3683                 if (nmi) {
3684                         event->pending_disable = 1;
3685                         perf_pending_queue(&event->pending,
3686                                            perf_pending_event);
3687                 } else
3688                         perf_event_disable(event);
3689         }
3690
3691         perf_event_output(event, nmi, data, regs);
3692         return ret;
3693 }
3694
3695 int perf_event_overflow(struct perf_event *event, int nmi,
3696                           struct perf_sample_data *data,
3697                           struct pt_regs *regs)
3698 {
3699         return __perf_event_overflow(event, nmi, 1, data, regs);
3700 }
3701
3702 /*
3703  * Generic software event infrastructure
3704  */
3705
3706 /*
3707  * We directly increment event->count and keep a second value in
3708  * event->hw.period_left to count intervals. This period event
3709  * is kept in the range [-sample_period, 0] so that we can use the
3710  * sign as trigger.
3711  */
3712
3713 static u64 perf_swevent_set_period(struct perf_event *event)
3714 {
3715         struct hw_perf_event *hwc = &event->hw;
3716         u64 period = hwc->last_period;
3717         u64 nr, offset;
3718         s64 old, val;
3719
3720         hwc->last_period = hwc->sample_period;
3721
3722 again:
3723         old = val = atomic64_read(&hwc->period_left);
3724         if (val < 0)
3725                 return 0;
3726
3727         nr = div64_u64(period + val, period);
3728         offset = nr * period;
3729         val -= offset;
3730         if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3731                 goto again;
3732
3733         return nr;
3734 }
3735
3736 static void perf_swevent_overflow(struct perf_event *event,
3737                                     int nmi, struct perf_sample_data *data,
3738                                     struct pt_regs *regs)
3739 {
3740         struct hw_perf_event *hwc = &event->hw;
3741         int throttle = 0;
3742         u64 overflow;
3743
3744         data->period = event->hw.last_period;
3745         overflow = perf_swevent_set_period(event);
3746
3747         if (hwc->interrupts == MAX_INTERRUPTS)
3748                 return;
3749
3750         for (; overflow; overflow--) {
3751                 if (__perf_event_overflow(event, nmi, throttle,
3752                                             data, regs)) {
3753                         /*
3754                          * We inhibit the overflow from happening when
3755                          * hwc->interrupts == MAX_INTERRUPTS.
3756                          */
3757                         break;
3758                 }
3759                 throttle = 1;
3760         }
3761 }
3762
3763 static void perf_swevent_unthrottle(struct perf_event *event)
3764 {
3765         /*
3766          * Nothing to do, we already reset hwc->interrupts.
3767          */
3768 }
3769
3770 static void perf_swevent_add(struct perf_event *event, u64 nr,
3771                                int nmi, struct perf_sample_data *data,
3772                                struct pt_regs *regs)
3773 {
3774         struct hw_perf_event *hwc = &event->hw;
3775
3776         atomic64_add(nr, &event->count);
3777
3778         if (!hwc->sample_period)
3779                 return;
3780
3781         if (!regs)
3782                 return;
3783
3784         if (!atomic64_add_negative(nr, &hwc->period_left))
3785                 perf_swevent_overflow(event, nmi, data, regs);
3786 }
3787
3788 static int perf_swevent_is_counting(struct perf_event *event)
3789 {
3790         /*
3791          * The event is active, we're good!
3792          */
3793         if (event->state == PERF_EVENT_STATE_ACTIVE)
3794                 return 1;
3795
3796         /*
3797          * The event is off/error, not counting.
3798          */
3799         if (event->state != PERF_EVENT_STATE_INACTIVE)
3800                 return 0;
3801
3802         /*
3803          * The event is inactive, if the context is active
3804          * we're part of a group that didn't make it on the 'pmu',
3805          * not counting.
3806          */
3807         if (event->ctx->is_active)
3808                 return 0;
3809
3810         /*
3811          * We're inactive and the context is too, this means the
3812          * task is scheduled out, we're counting events that happen
3813          * to us, like migration events.
3814          */
3815         return 1;
3816 }
3817
3818 static int perf_tp_event_match(struct perf_event *event,
3819                                 struct perf_sample_data *data);
3820
3821 static int perf_swevent_match(struct perf_event *event,
3822                                 enum perf_type_id type,
3823                                 u32 event_id,
3824                                 struct perf_sample_data *data,
3825                                 struct pt_regs *regs)
3826 {
3827         if (!perf_swevent_is_counting(event))
3828                 return 0;
3829
3830         if (event->attr.type != type)
3831                 return 0;
3832         if (event->attr.config != event_id)
3833                 return 0;
3834
3835         if (regs) {
3836                 if (event->attr.exclude_user && user_mode(regs))
3837                         return 0;
3838
3839                 if (event->attr.exclude_kernel && !user_mode(regs))
3840                         return 0;
3841         }
3842
3843         if (event->attr.type == PERF_TYPE_TRACEPOINT &&
3844             !perf_tp_event_match(event, data))
3845                 return 0;
3846
3847         return 1;
3848 }
3849
3850 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
3851                                      enum perf_type_id type,
3852                                      u32 event_id, u64 nr, int nmi,
3853                                      struct perf_sample_data *data,
3854                                      struct pt_regs *regs)
3855 {
3856         struct perf_event *event;
3857
3858         if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3859                 return;
3860
3861         rcu_read_lock();
3862         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3863                 if (perf_swevent_match(event, type, event_id, data, regs))
3864                         perf_swevent_add(event, nr, nmi, data, regs);
3865         }
3866         rcu_read_unlock();
3867 }
3868
3869 static int *perf_swevent_recursion_context(struct perf_cpu_context *cpuctx)
3870 {
3871         if (in_nmi())
3872                 return &cpuctx->recursion[3];
3873
3874         if (in_irq())
3875                 return &cpuctx->recursion[2];
3876
3877         if (in_softirq())
3878                 return &cpuctx->recursion[1];
3879
3880         return &cpuctx->recursion[0];
3881 }
3882
3883 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
3884                                     u64 nr, int nmi,
3885                                     struct perf_sample_data *data,
3886                                     struct pt_regs *regs)
3887 {
3888         struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3889         int *recursion = perf_swevent_recursion_context(cpuctx);
3890         struct perf_event_context *ctx;
3891
3892         if (*recursion)
3893                 goto out;
3894
3895         (*recursion)++;
3896         barrier();
3897
3898         perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
3899                                  nr, nmi, data, regs);
3900         rcu_read_lock();
3901         /*
3902          * doesn't really matter which of the child contexts the
3903          * events ends up in.
3904          */
3905         ctx = rcu_dereference(current->perf_event_ctxp);
3906         if (ctx)
3907                 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
3908         rcu_read_unlock();
3909
3910         barrier();
3911         (*recursion)--;
3912
3913 out:
3914         put_cpu_var(perf_cpu_context);
3915 }
3916
3917 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
3918                             struct pt_regs *regs, u64 addr)
3919 {
3920         struct perf_sample_data data = {
3921                 .addr = addr,
3922         };
3923
3924         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi,
3925                                 &data, regs);
3926 }
3927
3928 static void perf_swevent_read(struct perf_event *event)
3929 {
3930 }
3931
3932 static int perf_swevent_enable(struct perf_event *event)
3933 {
3934         struct hw_perf_event *hwc = &event->hw;
3935
3936         if (hwc->sample_period) {
3937                 hwc->last_period = hwc->sample_period;
3938                 perf_swevent_set_period(event);
3939         }
3940         return 0;
3941 }
3942
3943 static void perf_swevent_disable(struct perf_event *event)
3944 {
3945 }
3946
3947 static const struct pmu perf_ops_generic = {
3948         .enable         = perf_swevent_enable,
3949         .disable        = perf_swevent_disable,
3950         .read           = perf_swevent_read,
3951         .unthrottle     = perf_swevent_unthrottle,
3952 };
3953
3954 /*
3955  * hrtimer based swevent callback
3956  */
3957
3958 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
3959 {
3960         enum hrtimer_restart ret = HRTIMER_RESTART;
3961         struct perf_sample_data data;
3962         struct pt_regs *regs;
3963         struct perf_event *event;
3964         u64 period;
3965
3966         event   = container_of(hrtimer, struct perf_event, hw.hrtimer);
3967         event->pmu->read(event);
3968
3969         data.addr = 0;
3970         regs = get_irq_regs();
3971         /*
3972          * In case we exclude kernel IPs or are somehow not in interrupt
3973          * context, provide the next best thing, the user IP.
3974          */
3975         if ((event->attr.exclude_kernel || !regs) &&
3976                         !event->attr.exclude_user)
3977                 regs = task_pt_regs(current);
3978
3979         if (regs) {
3980                 if (!(event->attr.exclude_idle && current->pid == 0))
3981                         if (perf_event_overflow(event, 0, &data, regs))
3982                                 ret = HRTIMER_NORESTART;
3983         }
3984
3985         period = max_t(u64, 10000, event->hw.sample_period);
3986         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
3987
3988         return ret;
3989 }
3990
3991 static void perf_swevent_start_hrtimer(struct perf_event *event)
3992 {
3993         struct hw_perf_event *hwc = &event->hw;
3994
3995         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3996         hwc->hrtimer.function = perf_swevent_hrtimer;
3997         if (hwc->sample_period) {
3998                 u64 period;
3999
4000                 if (hwc->remaining) {
4001                         if (hwc->remaining < 0)
4002                                 period = 10000;
4003                         else
4004                                 period = hwc->remaining;
4005                         hwc->remaining = 0;
4006                 } else {
4007                         period = max_t(u64, 10000, hwc->sample_period);
4008                 }
4009                 __hrtimer_start_range_ns(&hwc->hrtimer,
4010                                 ns_to_ktime(period), 0,
4011                                 HRTIMER_MODE_REL, 0);
4012         }
4013 }
4014
4015 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4016 {
4017         struct hw_perf_event *hwc = &event->hw;
4018
4019         if (hwc->sample_period) {
4020                 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4021                 hwc->remaining = ktime_to_ns(remaining);
4022
4023                 hrtimer_cancel(&hwc->hrtimer);
4024         }
4025 }
4026
4027 /*
4028  * Software event: cpu wall time clock
4029  */
4030
4031 static void cpu_clock_perf_event_update(struct perf_event *event)
4032 {
4033         int cpu = raw_smp_processor_id();
4034         s64 prev;
4035         u64 now;
4036
4037         now = cpu_clock(cpu);
4038         prev = atomic64_read(&event->hw.prev_count);
4039         atomic64_set(&event->hw.prev_count, now);
4040         atomic64_add(now - prev, &event->count);
4041 }
4042
4043 static int cpu_clock_perf_event_enable(struct perf_event *event)
4044 {
4045         struct hw_perf_event *hwc = &event->hw;
4046         int cpu = raw_smp_processor_id();
4047
4048         atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4049         perf_swevent_start_hrtimer(event);
4050
4051         return 0;
4052 }
4053
4054 static void cpu_clock_perf_event_disable(struct perf_event *event)
4055 {
4056         perf_swevent_cancel_hrtimer(event);
4057         cpu_clock_perf_event_update(event);
4058 }
4059
4060 static void cpu_clock_perf_event_read(struct perf_event *event)
4061 {
4062         cpu_clock_perf_event_update(event);
4063 }
4064
4065 static const struct pmu perf_ops_cpu_clock = {
4066         .enable         = cpu_clock_perf_event_enable,
4067         .disable        = cpu_clock_perf_event_disable,
4068         .read           = cpu_clock_perf_event_read,
4069 };
4070
4071 /*
4072  * Software event: task time clock
4073  */
4074
4075 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4076 {
4077         u64 prev;
4078         s64 delta;
4079
4080         prev = atomic64_xchg(&event->hw.prev_count, now);
4081         delta = now - prev;
4082         atomic64_add(delta, &event->count);
4083 }
4084
4085 static int task_clock_perf_event_enable(struct perf_event *event)
4086 {
4087         struct hw_perf_event *hwc = &event->hw;
4088         u64 now;
4089
4090         now = event->ctx->time;
4091
4092         atomic64_set(&hwc->prev_count, now);
4093
4094         perf_swevent_start_hrtimer(event);
4095
4096         return 0;
4097 }
4098
4099 static void task_clock_perf_event_disable(struct perf_event *event)
4100 {
4101         perf_swevent_cancel_hrtimer(event);
4102         task_clock_perf_event_update(event, event->ctx->time);
4103
4104 }
4105
4106 static void task_clock_perf_event_read(struct perf_event *event)
4107 {
4108         u64 time;
4109
4110         if (!in_nmi()) {
4111                 update_context_time(event->ctx);
4112                 time = event->ctx->time;
4113         } else {
4114                 u64 now = perf_clock();
4115                 u64 delta = now - event->ctx->timestamp;
4116                 time = event->ctx->time + delta;
4117         }
4118
4119         task_clock_perf_event_update(event, time);
4120 }
4121
4122 static const struct pmu perf_ops_task_clock = {
4123         .enable         = task_clock_perf_event_enable,
4124         .disable        = task_clock_perf_event_disable,
4125         .read           = task_clock_perf_event_read,
4126 };
4127
4128 #ifdef CONFIG_EVENT_PROFILE
4129
4130 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4131                           int entry_size)
4132 {
4133         struct perf_raw_record raw = {
4134                 .size = entry_size,
4135                 .data = record,
4136         };
4137
4138         struct perf_sample_data data = {
4139                 .addr = addr,
4140                 .raw = &raw,
4141         };
4142
4143         struct pt_regs *regs = get_irq_regs();
4144
4145         if (!regs)
4146                 regs = task_pt_regs(current);
4147
4148         do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4149                                 &data, regs);
4150 }
4151 EXPORT_SYMBOL_GPL(perf_tp_event);
4152
4153 static int perf_tp_event_match(struct perf_event *event,
4154                                 struct perf_sample_data *data)
4155 {
4156         void *record = data->raw->data;
4157
4158         if (likely(!event->filter) || filter_match_preds(event->filter, record))
4159                 return 1;
4160         return 0;
4161 }
4162
4163 static void tp_perf_event_destroy(struct perf_event *event)
4164 {
4165         ftrace_profile_disable(event->attr.config);
4166 }
4167
4168 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4169 {
4170         /*
4171          * Raw tracepoint data is a severe data leak, only allow root to
4172          * have these.
4173          */
4174         if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4175                         perf_paranoid_tracepoint_raw() &&
4176                         !capable(CAP_SYS_ADMIN))
4177                 return ERR_PTR(-EPERM);
4178
4179         if (ftrace_profile_enable(event->attr.config))
4180                 return NULL;
4181
4182         event->destroy = tp_perf_event_destroy;
4183
4184         return &perf_ops_generic;
4185 }
4186
4187 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4188 {
4189         char *filter_str;
4190         int ret;
4191
4192         if (event->attr.type != PERF_TYPE_TRACEPOINT)
4193                 return -EINVAL;
4194
4195         filter_str = strndup_user(arg, PAGE_SIZE);
4196         if (IS_ERR(filter_str))
4197                 return PTR_ERR(filter_str);
4198
4199         ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4200
4201         kfree(filter_str);
4202         return ret;
4203 }
4204
4205 static void perf_event_free_filter(struct perf_event *event)
4206 {
4207         ftrace_profile_free_filter(event);
4208 }
4209
4210 #else
4211
4212 static int perf_tp_event_match(struct perf_event *event,
4213                                 struct perf_sample_data *data)
4214 {
4215         return 1;
4216 }
4217
4218 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4219 {
4220         return NULL;
4221 }
4222
4223 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4224 {
4225         return -ENOENT;
4226 }
4227
4228 static void perf_event_free_filter(struct perf_event *event)
4229 {
4230 }
4231
4232 #endif /* CONFIG_EVENT_PROFILE */
4233
4234 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4235
4236 static void sw_perf_event_destroy(struct perf_event *event)
4237 {
4238         u64 event_id = event->attr.config;
4239
4240         WARN_ON(event->parent);
4241
4242         atomic_dec(&perf_swevent_enabled[event_id]);
4243 }
4244
4245 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4246 {
4247         const struct pmu *pmu = NULL;
4248         u64 event_id = event->attr.config;
4249
4250         /*
4251          * Software events (currently) can't in general distinguish
4252          * between user, kernel and hypervisor events.
4253          * However, context switches and cpu migrations are considered
4254          * to be kernel events, and page faults are never hypervisor
4255          * events.
4256          */
4257         switch (event_id) {
4258         case PERF_COUNT_SW_CPU_CLOCK:
4259                 pmu = &perf_ops_cpu_clock;
4260
4261                 break;
4262         case PERF_COUNT_SW_TASK_CLOCK:
4263                 /*
4264                  * If the user instantiates this as a per-cpu event,
4265                  * use the cpu_clock event instead.
4266                  */
4267                 if (event->ctx->task)
4268                         pmu = &perf_ops_task_clock;
4269                 else
4270                         pmu = &perf_ops_cpu_clock;
4271
4272                 break;
4273         case PERF_COUNT_SW_PAGE_FAULTS:
4274         case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4275         case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4276         case PERF_COUNT_SW_CONTEXT_SWITCHES:
4277         case PERF_COUNT_SW_CPU_MIGRATIONS:
4278         case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4279         case PERF_COUNT_SW_EMULATION_FAULTS:
4280                 if (!event->parent) {
4281                         atomic_inc(&perf_swevent_enabled[event_id]);
4282                         event->destroy = sw_perf_event_destroy;
4283                 }
4284                 pmu = &perf_ops_generic;
4285                 break;
4286         }
4287
4288         return pmu;
4289 }
4290
4291 /*
4292  * Allocate and initialize a event structure
4293  */
4294 static struct perf_event *
4295 perf_event_alloc(struct perf_event_attr *attr,
4296                    int cpu,
4297                    struct perf_event_context *ctx,
4298                    struct perf_event *group_leader,
4299                    struct perf_event *parent_event,
4300                    gfp_t gfpflags)
4301 {
4302         const struct pmu *pmu;
4303         struct perf_event *event;
4304         struct hw_perf_event *hwc;
4305         long err;
4306
4307         event = kzalloc(sizeof(*event), gfpflags);
4308         if (!event)
4309                 return ERR_PTR(-ENOMEM);
4310
4311         /*
4312          * Single events are their own group leaders, with an
4313          * empty sibling list:
4314          */
4315         if (!group_leader)
4316                 group_leader = event;
4317
4318         mutex_init(&event->child_mutex);
4319         INIT_LIST_HEAD(&event->child_list);
4320
4321         INIT_LIST_HEAD(&event->group_entry);
4322         INIT_LIST_HEAD(&event->event_entry);
4323         INIT_LIST_HEAD(&event->sibling_list);
4324         init_waitqueue_head(&event->waitq);
4325
4326         mutex_init(&event->mmap_mutex);
4327
4328         event->cpu              = cpu;
4329         event->attr             = *attr;
4330         event->group_leader     = group_leader;
4331         event->pmu              = NULL;
4332         event->ctx              = ctx;
4333         event->oncpu            = -1;
4334
4335         event->parent           = parent_event;
4336
4337         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
4338         event->id               = atomic64_inc_return(&perf_event_id);
4339
4340         event->state            = PERF_EVENT_STATE_INACTIVE;
4341
4342         if (attr->disabled)
4343                 event->state = PERF_EVENT_STATE_OFF;
4344
4345         pmu = NULL;
4346
4347         hwc = &event->hw;
4348         hwc->sample_period = attr->sample_period;
4349         if (attr->freq && attr->sample_freq)
4350                 hwc->sample_period = 1;
4351         hwc->last_period = hwc->sample_period;
4352
4353         atomic64_set(&hwc->period_left, hwc->sample_period);
4354
4355         /*
4356          * we currently do not support PERF_FORMAT_GROUP on inherited events
4357          */
4358         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4359                 goto done;
4360
4361         switch (attr->type) {
4362         case PERF_TYPE_RAW:
4363         case PERF_TYPE_HARDWARE:
4364         case PERF_TYPE_HW_CACHE:
4365                 pmu = hw_perf_event_init(event);
4366                 break;
4367
4368         case PERF_TYPE_SOFTWARE:
4369                 pmu = sw_perf_event_init(event);
4370                 break;
4371
4372         case PERF_TYPE_TRACEPOINT:
4373                 pmu = tp_perf_event_init(event);
4374                 break;
4375
4376         default:
4377                 break;
4378         }
4379 done:
4380         err = 0;
4381         if (!pmu)
4382                 err = -EINVAL;
4383         else if (IS_ERR(pmu))
4384                 err = PTR_ERR(pmu);
4385
4386         if (err) {
4387                 if (event->ns)
4388                         put_pid_ns(event->ns);
4389                 kfree(event);
4390                 return ERR_PTR(err);
4391         }
4392
4393         event->pmu = pmu;
4394
4395         if (!event->parent) {
4396                 atomic_inc(&nr_events);
4397                 if (event->attr.mmap)
4398                         atomic_inc(&nr_mmap_events);
4399                 if (event->attr.comm)
4400                         atomic_inc(&nr_comm_events);
4401                 if (event->attr.task)
4402                         atomic_inc(&nr_task_events);
4403         }
4404
4405         return event;
4406 }
4407
4408 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4409                           struct perf_event_attr *attr)
4410 {
4411         u32 size;
4412         int ret;
4413
4414         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4415                 return -EFAULT;
4416
4417         /*
4418          * zero the full structure, so that a short copy will be nice.
4419          */
4420         memset(attr, 0, sizeof(*attr));
4421
4422         ret = get_user(size, &uattr->size);
4423         if (ret)
4424                 return ret;
4425
4426         if (size > PAGE_SIZE)   /* silly large */
4427                 goto err_size;
4428
4429         if (!size)              /* abi compat */
4430                 size = PERF_ATTR_SIZE_VER0;
4431
4432         if (size < PERF_ATTR_SIZE_VER0)
4433                 goto err_size;
4434
4435         /*
4436          * If we're handed a bigger struct than we know of,
4437          * ensure all the unknown bits are 0 - i.e. new
4438          * user-space does not rely on any kernel feature
4439          * extensions we dont know about yet.
4440          */
4441         if (size > sizeof(*attr)) {
4442                 unsigned char __user *addr;
4443                 unsigned char __user *end;
4444                 unsigned char val;
4445
4446                 addr = (void __user *)uattr + sizeof(*attr);
4447                 end  = (void __user *)uattr + size;
4448
4449                 for (; addr < end; addr++) {
4450                         ret = get_user(val, addr);
4451                         if (ret)
4452                                 return ret;
4453                         if (val)
4454                                 goto err_size;
4455                 }
4456                 size = sizeof(*attr);
4457         }
4458
4459         ret = copy_from_user(attr, uattr, size);
4460         if (ret)
4461                 return -EFAULT;
4462
4463         /*
4464          * If the type exists, the corresponding creation will verify
4465          * the attr->config.
4466          */
4467         if (attr->type >= PERF_TYPE_MAX)
4468                 return -EINVAL;
4469
4470         if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3)
4471                 return -EINVAL;
4472
4473         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4474                 return -EINVAL;
4475
4476         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4477                 return -EINVAL;
4478
4479 out:
4480         return ret;
4481
4482 err_size:
4483         put_user(sizeof(*attr), &uattr->size);
4484         ret = -E2BIG;
4485         goto out;
4486 }
4487
4488 static int perf_event_set_output(struct perf_event *event, int output_fd)
4489 {
4490         struct perf_event *output_event = NULL;
4491         struct file *output_file = NULL;
4492         struct perf_event *old_output;
4493         int fput_needed = 0;
4494         int ret = -EINVAL;
4495
4496         if (!output_fd)
4497                 goto set;
4498
4499         output_file = fget_light(output_fd, &fput_needed);
4500         if (!output_file)
4501                 return -EBADF;
4502
4503         if (output_file->f_op != &perf_fops)
4504                 goto out;
4505
4506         output_event = output_file->private_data;
4507
4508         /* Don't chain output fds */
4509         if (output_event->output)
4510                 goto out;
4511
4512         /* Don't set an output fd when we already have an output channel */
4513         if (event->data)
4514                 goto out;
4515
4516         atomic_long_inc(&output_file->f_count);
4517
4518 set:
4519         mutex_lock(&event->mmap_mutex);
4520         old_output = event->output;
4521         rcu_assign_pointer(event->output, output_event);
4522         mutex_unlock(&event->mmap_mutex);
4523
4524         if (old_output) {
4525                 /*
4526                  * we need to make sure no existing perf_output_*()
4527                  * is still referencing this event.
4528                  */
4529                 synchronize_rcu();
4530                 fput(old_output->filp);
4531         }
4532
4533         ret = 0;
4534 out:
4535         fput_light(output_file, fput_needed);
4536         return ret;
4537 }
4538
4539 /**
4540  * sys_perf_event_open - open a performance event, associate it to a task/cpu
4541  *
4542  * @attr_uptr:  event_id type attributes for monitoring/sampling
4543  * @pid:                target pid
4544  * @cpu:                target cpu
4545  * @group_fd:           group leader event fd
4546  */
4547 SYSCALL_DEFINE5(perf_event_open,
4548                 struct perf_event_attr __user *, attr_uptr,
4549                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4550 {
4551         struct perf_event *event, *group_leader;
4552         struct perf_event_attr attr;
4553         struct perf_event_context *ctx;
4554         struct file *event_file = NULL;
4555         struct file *group_file = NULL;
4556         int fput_needed = 0;
4557         int fput_needed2 = 0;
4558         int err;
4559
4560         /* for future expandability... */
4561         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4562                 return -EINVAL;
4563
4564         err = perf_copy_attr(attr_uptr, &attr);
4565         if (err)
4566                 return err;
4567
4568         if (!attr.exclude_kernel) {
4569                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4570                         return -EACCES;
4571         }
4572
4573         if (attr.freq) {
4574                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4575                         return -EINVAL;
4576         }
4577
4578         /*
4579          * Get the target context (task or percpu):
4580          */
4581         ctx = find_get_context(pid, cpu);
4582         if (IS_ERR(ctx))
4583                 return PTR_ERR(ctx);
4584
4585         /*
4586          * Look up the group leader (we will attach this event to it):
4587          */
4588         group_leader = NULL;
4589         if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4590                 err = -EINVAL;
4591                 group_file = fget_light(group_fd, &fput_needed);
4592                 if (!group_file)
4593                         goto err_put_context;
4594                 if (group_file->f_op != &perf_fops)
4595                         goto err_put_context;
4596
4597                 group_leader = group_file->private_data;
4598                 /*
4599                  * Do not allow a recursive hierarchy (this new sibling
4600                  * becoming part of another group-sibling):
4601                  */
4602                 if (group_leader->group_leader != group_leader)
4603                         goto err_put_context;
4604                 /*
4605                  * Do not allow to attach to a group in a different
4606                  * task or CPU context:
4607                  */
4608                 if (group_leader->ctx != ctx)
4609                         goto err_put_context;
4610                 /*
4611                  * Only a group leader can be exclusive or pinned
4612                  */
4613                 if (attr.exclusive || attr.pinned)
4614                         goto err_put_context;
4615         }
4616
4617         event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4618                                      NULL, GFP_KERNEL);
4619         err = PTR_ERR(event);
4620         if (IS_ERR(event))
4621                 goto err_put_context;
4622
4623         err = anon_inode_getfd("[perf_event]", &perf_fops, event, 0);
4624         if (err < 0)
4625                 goto err_free_put_context;
4626
4627         event_file = fget_light(err, &fput_needed2);
4628         if (!event_file)
4629                 goto err_free_put_context;
4630
4631         if (flags & PERF_FLAG_FD_OUTPUT) {
4632                 err = perf_event_set_output(event, group_fd);
4633                 if (err)
4634                         goto err_fput_free_put_context;
4635         }
4636
4637         event->filp = event_file;
4638         WARN_ON_ONCE(ctx->parent_ctx);
4639         mutex_lock(&ctx->mutex);
4640         perf_install_in_context(ctx, event, cpu);
4641         ++ctx->generation;
4642         mutex_unlock(&ctx->mutex);
4643
4644         event->owner = current;
4645         get_task_struct(current);
4646         mutex_lock(&current->perf_event_mutex);
4647         list_add_tail(&event->owner_entry, &current->perf_event_list);
4648         mutex_unlock(&current->perf_event_mutex);
4649
4650 err_fput_free_put_context:
4651         fput_light(event_file, fput_needed2);
4652
4653 err_free_put_context:
4654         if (err < 0)
4655                 kfree(event);
4656
4657 err_put_context:
4658         if (err < 0)
4659                 put_ctx(ctx);
4660
4661         fput_light(group_file, fput_needed);
4662
4663         return err;
4664 }
4665
4666 /*
4667  * inherit a event from parent task to child task:
4668  */
4669 static struct perf_event *
4670 inherit_event(struct perf_event *parent_event,
4671               struct task_struct *parent,
4672               struct perf_event_context *parent_ctx,
4673               struct task_struct *child,
4674               struct perf_event *group_leader,
4675               struct perf_event_context *child_ctx)
4676 {
4677         struct perf_event *child_event;
4678
4679         /*
4680          * Instead of creating recursive hierarchies of events,
4681          * we link inherited events back to the original parent,
4682          * which has a filp for sure, which we use as the reference
4683          * count:
4684          */
4685         if (parent_event->parent)
4686                 parent_event = parent_event->parent;
4687
4688         child_event = perf_event_alloc(&parent_event->attr,
4689                                            parent_event->cpu, child_ctx,
4690                                            group_leader, parent_event,
4691                                            GFP_KERNEL);
4692         if (IS_ERR(child_event))
4693                 return child_event;
4694         get_ctx(child_ctx);
4695
4696         /*
4697          * Make the child state follow the state of the parent event,
4698          * not its attr.disabled bit.  We hold the parent's mutex,
4699          * so we won't race with perf_event_{en, dis}able_family.
4700          */
4701         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
4702                 child_event->state = PERF_EVENT_STATE_INACTIVE;
4703         else
4704                 child_event->state = PERF_EVENT_STATE_OFF;
4705
4706         if (parent_event->attr.freq)
4707                 child_event->hw.sample_period = parent_event->hw.sample_period;
4708
4709         /*
4710          * Link it up in the child's context:
4711          */
4712         add_event_to_ctx(child_event, child_ctx);
4713
4714         /*
4715          * Get a reference to the parent filp - we will fput it
4716          * when the child event exits. This is safe to do because
4717          * we are in the parent and we know that the filp still
4718          * exists and has a nonzero count:
4719          */
4720         atomic_long_inc(&parent_event->filp->f_count);
4721
4722         /*
4723          * Link this into the parent event's child list
4724          */
4725         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4726         mutex_lock(&parent_event->child_mutex);
4727         list_add_tail(&child_event->child_list, &parent_event->child_list);
4728         mutex_unlock(&parent_event->child_mutex);
4729
4730         return child_event;
4731 }
4732
4733 static int inherit_group(struct perf_event *parent_event,
4734               struct task_struct *parent,
4735               struct perf_event_context *parent_ctx,
4736               struct task_struct *child,
4737               struct perf_event_context *child_ctx)
4738 {
4739         struct perf_event *leader;
4740         struct perf_event *sub;
4741         struct perf_event *child_ctr;
4742
4743         leader = inherit_event(parent_event, parent, parent_ctx,
4744                                  child, NULL, child_ctx);
4745         if (IS_ERR(leader))
4746                 return PTR_ERR(leader);
4747         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
4748                 child_ctr = inherit_event(sub, parent, parent_ctx,
4749                                             child, leader, child_ctx);
4750                 if (IS_ERR(child_ctr))
4751                         return PTR_ERR(child_ctr);
4752         }
4753         return 0;
4754 }
4755
4756 static void sync_child_event(struct perf_event *child_event,
4757                                struct task_struct *child)
4758 {
4759         struct perf_event *parent_event = child_event->parent;
4760         u64 child_val;
4761
4762         if (child_event->attr.inherit_stat)
4763                 perf_event_read_event(child_event, child);
4764
4765         child_val = atomic64_read(&child_event->count);
4766
4767         /*
4768          * Add back the child's count to the parent's count:
4769          */
4770         atomic64_add(child_val, &parent_event->count);
4771         atomic64_add(child_event->total_time_enabled,
4772                      &parent_event->child_total_time_enabled);
4773         atomic64_add(child_event->total_time_running,
4774                      &parent_event->child_total_time_running);
4775
4776         /*
4777          * Remove this event from the parent's list
4778          */
4779         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4780         mutex_lock(&parent_event->child_mutex);
4781         list_del_init(&child_event->child_list);
4782         mutex_unlock(&parent_event->child_mutex);
4783
4784         /*
4785          * Release the parent event, if this was the last
4786          * reference to it.
4787          */
4788         fput(parent_event->filp);
4789 }
4790
4791 static void
4792 __perf_event_exit_task(struct perf_event *child_event,
4793                          struct perf_event_context *child_ctx,
4794                          struct task_struct *child)
4795 {
4796         struct perf_event *parent_event;
4797
4798         update_event_times(child_event);
4799         perf_event_remove_from_context(child_event);
4800
4801         parent_event = child_event->parent;
4802         /*
4803          * It can happen that parent exits first, and has events
4804          * that are still around due to the child reference. These
4805          * events need to be zapped - but otherwise linger.
4806          */
4807         if (parent_event) {
4808                 sync_child_event(child_event, child);
4809                 free_event(child_event);
4810         }
4811 }
4812
4813 /*
4814  * When a child task exits, feed back event values to parent events.
4815  */
4816 void perf_event_exit_task(struct task_struct *child)
4817 {
4818         struct perf_event *child_event, *tmp;
4819         struct perf_event_context *child_ctx;
4820         unsigned long flags;
4821
4822         if (likely(!child->perf_event_ctxp)) {
4823                 perf_event_task(child, NULL, 0);
4824                 return;
4825         }
4826
4827         local_irq_save(flags);
4828         /*
4829          * We can't reschedule here because interrupts are disabled,
4830          * and either child is current or it is a task that can't be
4831          * scheduled, so we are now safe from rescheduling changing
4832          * our context.
4833          */
4834         child_ctx = child->perf_event_ctxp;
4835         __perf_event_task_sched_out(child_ctx);
4836
4837         /*
4838          * Take the context lock here so that if find_get_context is
4839          * reading child->perf_event_ctxp, we wait until it has
4840          * incremented the context's refcount before we do put_ctx below.
4841          */
4842         spin_lock(&child_ctx->lock);
4843         child->perf_event_ctxp = NULL;
4844         /*
4845          * If this context is a clone; unclone it so it can't get
4846          * swapped to another process while we're removing all
4847          * the events from it.
4848          */
4849         unclone_ctx(child_ctx);
4850         spin_unlock_irqrestore(&child_ctx->lock, flags);
4851
4852         /*
4853          * Report the task dead after unscheduling the events so that we
4854          * won't get any samples after PERF_RECORD_EXIT. We can however still
4855          * get a few PERF_RECORD_READ events.
4856          */
4857         perf_event_task(child, child_ctx, 0);
4858
4859         /*
4860          * We can recurse on the same lock type through:
4861          *
4862          *   __perf_event_exit_task()
4863          *     sync_child_event()
4864          *       fput(parent_event->filp)
4865          *         perf_release()
4866          *           mutex_lock(&ctx->mutex)
4867          *
4868          * But since its the parent context it won't be the same instance.
4869          */
4870         mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
4871
4872 again:
4873         list_for_each_entry_safe(child_event, tmp, &child_ctx->group_list,
4874                                  group_entry)
4875                 __perf_event_exit_task(child_event, child_ctx, child);
4876
4877         /*
4878          * If the last event was a group event, it will have appended all
4879          * its siblings to the list, but we obtained 'tmp' before that which
4880          * will still point to the list head terminating the iteration.
4881          */
4882         if (!list_empty(&child_ctx->group_list))
4883                 goto again;
4884
4885         mutex_unlock(&child_ctx->mutex);
4886
4887         put_ctx(child_ctx);
4888 }
4889
4890 /*
4891  * free an unexposed, unused context as created by inheritance by
4892  * init_task below, used by fork() in case of fail.
4893  */
4894 void perf_event_free_task(struct task_struct *task)
4895 {
4896         struct perf_event_context *ctx = task->perf_event_ctxp;
4897         struct perf_event *event, *tmp;
4898
4899         if (!ctx)
4900                 return;
4901
4902         mutex_lock(&ctx->mutex);
4903 again:
4904         list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry) {
4905                 struct perf_event *parent = event->parent;
4906
4907                 if (WARN_ON_ONCE(!parent))
4908                         continue;
4909
4910                 mutex_lock(&parent->child_mutex);
4911                 list_del_init(&event->child_list);
4912                 mutex_unlock(&parent->child_mutex);
4913
4914                 fput(parent->filp);
4915
4916                 list_del_event(event, ctx);
4917                 free_event(event);
4918         }
4919
4920         if (!list_empty(&ctx->group_list))
4921                 goto again;
4922
4923         mutex_unlock(&ctx->mutex);
4924
4925         put_ctx(ctx);
4926 }
4927
4928 /*
4929  * Initialize the perf_event context in task_struct
4930  */
4931 int perf_event_init_task(struct task_struct *child)
4932 {
4933         struct perf_event_context *child_ctx, *parent_ctx;
4934         struct perf_event_context *cloned_ctx;
4935         struct perf_event *event;
4936         struct task_struct *parent = current;
4937         int inherited_all = 1;
4938         int ret = 0;
4939
4940         child->perf_event_ctxp = NULL;
4941
4942         mutex_init(&child->perf_event_mutex);
4943         INIT_LIST_HEAD(&child->perf_event_list);
4944
4945         if (likely(!parent->perf_event_ctxp))
4946                 return 0;
4947
4948         /*
4949          * This is executed from the parent task context, so inherit
4950          * events that have been marked for cloning.
4951          * First allocate and initialize a context for the child.
4952          */
4953
4954         child_ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
4955         if (!child_ctx)
4956                 return -ENOMEM;
4957
4958         __perf_event_init_context(child_ctx, child);
4959         child->perf_event_ctxp = child_ctx;
4960         get_task_struct(child);
4961
4962         /*
4963          * If the parent's context is a clone, pin it so it won't get
4964          * swapped under us.
4965          */
4966         parent_ctx = perf_pin_task_context(parent);
4967
4968         /*
4969          * No need to check if parent_ctx != NULL here; since we saw
4970          * it non-NULL earlier, the only reason for it to become NULL
4971          * is if we exit, and since we're currently in the middle of
4972          * a fork we can't be exiting at the same time.
4973          */
4974
4975         /*
4976          * Lock the parent list. No need to lock the child - not PID
4977          * hashed yet and not running, so nobody can access it.
4978          */
4979         mutex_lock(&parent_ctx->mutex);
4980
4981         /*
4982          * We dont have to disable NMIs - we are only looking at
4983          * the list, not manipulating it:
4984          */
4985         list_for_each_entry(event, &parent_ctx->group_list, group_entry) {
4986
4987                 if (!event->attr.inherit) {
4988                         inherited_all = 0;
4989                         continue;
4990                 }
4991
4992                 ret = inherit_group(event, parent, parent_ctx,
4993                                              child, child_ctx);
4994                 if (ret) {
4995                         inherited_all = 0;
4996                         break;
4997                 }
4998         }
4999
5000         if (inherited_all) {
5001                 /*
5002                  * Mark the child context as a clone of the parent
5003                  * context, or of whatever the parent is a clone of.
5004                  * Note that if the parent is a clone, it could get
5005                  * uncloned at any point, but that doesn't matter
5006                  * because the list of events and the generation
5007                  * count can't have changed since we took the mutex.
5008                  */
5009                 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5010                 if (cloned_ctx) {
5011                         child_ctx->parent_ctx = cloned_ctx;
5012                         child_ctx->parent_gen = parent_ctx->parent_gen;
5013                 } else {
5014                         child_ctx->parent_ctx = parent_ctx;
5015                         child_ctx->parent_gen = parent_ctx->generation;
5016                 }
5017                 get_ctx(child_ctx->parent_ctx);
5018         }
5019
5020         mutex_unlock(&parent_ctx->mutex);
5021
5022         perf_unpin_context(parent_ctx);
5023
5024         return ret;
5025 }
5026
5027 static void __cpuinit perf_event_init_cpu(int cpu)
5028 {
5029         struct perf_cpu_context *cpuctx;
5030
5031         cpuctx = &per_cpu(perf_cpu_context, cpu);
5032         __perf_event_init_context(&cpuctx->ctx, NULL);
5033
5034         spin_lock(&perf_resource_lock);
5035         cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5036         spin_unlock(&perf_resource_lock);
5037
5038         hw_perf_event_setup(cpu);
5039 }
5040
5041 #ifdef CONFIG_HOTPLUG_CPU
5042 static void __perf_event_exit_cpu(void *info)
5043 {
5044         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5045         struct perf_event_context *ctx = &cpuctx->ctx;
5046         struct perf_event *event, *tmp;
5047
5048         list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry)
5049                 __perf_event_remove_from_context(event);
5050 }
5051 static void perf_event_exit_cpu(int cpu)
5052 {
5053         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5054         struct perf_event_context *ctx = &cpuctx->ctx;
5055
5056         mutex_lock(&ctx->mutex);
5057         smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5058         mutex_unlock(&ctx->mutex);
5059 }
5060 #else
5061 static inline void perf_event_exit_cpu(int cpu) { }
5062 #endif
5063
5064 static int __cpuinit
5065 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5066 {
5067         unsigned int cpu = (long)hcpu;
5068
5069         switch (action) {
5070
5071         case CPU_UP_PREPARE:
5072         case CPU_UP_PREPARE_FROZEN:
5073                 perf_event_init_cpu(cpu);
5074                 break;
5075
5076         case CPU_ONLINE:
5077         case CPU_ONLINE_FROZEN:
5078                 hw_perf_event_setup_online(cpu);
5079                 break;
5080
5081         case CPU_DOWN_PREPARE:
5082         case CPU_DOWN_PREPARE_FROZEN:
5083                 perf_event_exit_cpu(cpu);
5084                 break;
5085
5086         default:
5087                 break;
5088         }
5089
5090         return NOTIFY_OK;
5091 }
5092
5093 /*
5094  * This has to have a higher priority than migration_notifier in sched.c.
5095  */
5096 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5097         .notifier_call          = perf_cpu_notify,
5098         .priority               = 20,
5099 };
5100
5101 void __init perf_event_init(void)
5102 {
5103         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5104                         (void *)(long)smp_processor_id());
5105         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5106                         (void *)(long)smp_processor_id());
5107         register_cpu_notifier(&perf_cpu_nb);
5108 }
5109
5110 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
5111 {
5112         return sprintf(buf, "%d\n", perf_reserved_percpu);
5113 }
5114
5115 static ssize_t
5116 perf_set_reserve_percpu(struct sysdev_class *class,
5117                         const char *buf,
5118                         size_t count)
5119 {
5120         struct perf_cpu_context *cpuctx;
5121         unsigned long val;
5122         int err, cpu, mpt;
5123
5124         err = strict_strtoul(buf, 10, &val);
5125         if (err)
5126                 return err;
5127         if (val > perf_max_events)
5128                 return -EINVAL;
5129
5130         spin_lock(&perf_resource_lock);
5131         perf_reserved_percpu = val;
5132         for_each_online_cpu(cpu) {
5133                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5134                 spin_lock_irq(&cpuctx->ctx.lock);
5135                 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5136                           perf_max_events - perf_reserved_percpu);
5137                 cpuctx->max_pertask = mpt;
5138                 spin_unlock_irq(&cpuctx->ctx.lock);
5139         }
5140         spin_unlock(&perf_resource_lock);
5141
5142         return count;
5143 }
5144
5145 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
5146 {
5147         return sprintf(buf, "%d\n", perf_overcommit);
5148 }
5149
5150 static ssize_t
5151 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
5152 {
5153         unsigned long val;
5154         int err;
5155
5156         err = strict_strtoul(buf, 10, &val);
5157         if (err)
5158                 return err;
5159         if (val > 1)
5160                 return -EINVAL;
5161
5162         spin_lock(&perf_resource_lock);
5163         perf_overcommit = val;
5164         spin_unlock(&perf_resource_lock);
5165
5166         return count;
5167 }
5168
5169 static SYSDEV_CLASS_ATTR(
5170                                 reserve_percpu,
5171                                 0644,
5172                                 perf_show_reserve_percpu,
5173                                 perf_set_reserve_percpu
5174                         );
5175
5176 static SYSDEV_CLASS_ATTR(
5177                                 overcommit,
5178                                 0644,
5179                                 perf_show_overcommit,
5180                                 perf_set_overcommit
5181                         );
5182
5183 static struct attribute *perfclass_attrs[] = {
5184         &attr_reserve_percpu.attr,
5185         &attr_overcommit.attr,
5186         NULL
5187 };
5188
5189 static struct attribute_group perfclass_attr_group = {
5190         .attrs                  = perfclass_attrs,
5191         .name                   = "perf_events",
5192 };
5193
5194 static int __init perf_event_sysfs_init(void)
5195 {
5196         return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5197                                   &perfclass_attr_group);
5198 }
5199 device_initcall(perf_event_sysfs_init);