perf_events: Restore sanity to scaling land
[linux-2.6.git] / kernel / perf_event.c
1 /*
2  * Performance events core code:
3  *
4  *  Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5  *  Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6  *  Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7  *  Copyright  ©  2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
8  *
9  * For licensing details see kernel-base/COPYING
10  */
11
12 #include <linux/fs.h>
13 #include <linux/mm.h>
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/file.h>
17 #include <linux/poll.h>
18 #include <linux/sysfs.h>
19 #include <linux/dcache.h>
20 #include <linux/percpu.h>
21 #include <linux/ptrace.h>
22 #include <linux/vmstat.h>
23 #include <linux/vmalloc.h>
24 #include <linux/hardirq.h>
25 #include <linux/rculist.h>
26 #include <linux/uaccess.h>
27 #include <linux/syscalls.h>
28 #include <linux/anon_inodes.h>
29 #include <linux/kernel_stat.h>
30 #include <linux/perf_event.h>
31 #include <linux/ftrace_event.h>
32 #include <linux/hw_breakpoint.h>
33
34 #include <asm/irq_regs.h>
35
36 /*
37  * Each CPU has a list of per CPU events:
38  */
39 DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
40
41 int perf_max_events __read_mostly = 1;
42 static int perf_reserved_percpu __read_mostly;
43 static int perf_overcommit __read_mostly = 1;
44
45 static atomic_t nr_events __read_mostly;
46 static atomic_t nr_mmap_events __read_mostly;
47 static atomic_t nr_comm_events __read_mostly;
48 static atomic_t nr_task_events __read_mostly;
49
50 /*
51  * perf event paranoia level:
52  *  -1 - not paranoid at all
53  *   0 - disallow raw tracepoint access for unpriv
54  *   1 - disallow cpu events for unpriv
55  *   2 - disallow kernel profiling for unpriv
56  */
57 int sysctl_perf_event_paranoid __read_mostly = 1;
58
59 static inline bool perf_paranoid_tracepoint_raw(void)
60 {
61         return sysctl_perf_event_paranoid > -1;
62 }
63
64 static inline bool perf_paranoid_cpu(void)
65 {
66         return sysctl_perf_event_paranoid > 0;
67 }
68
69 static inline bool perf_paranoid_kernel(void)
70 {
71         return sysctl_perf_event_paranoid > 1;
72 }
73
74 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
75
76 /*
77  * max perf event sample rate
78  */
79 int sysctl_perf_event_sample_rate __read_mostly = 100000;
80
81 static atomic64_t perf_event_id;
82
83 /*
84  * Lock for (sysadmin-configurable) event reservations:
85  */
86 static DEFINE_SPINLOCK(perf_resource_lock);
87
88 /*
89  * Architecture provided APIs - weak aliases:
90  */
91 extern __weak const struct pmu *hw_perf_event_init(struct perf_event *event)
92 {
93         return NULL;
94 }
95
96 void __weak hw_perf_disable(void)               { barrier(); }
97 void __weak hw_perf_enable(void)                { barrier(); }
98
99 void __weak hw_perf_event_setup(int cpu)        { barrier(); }
100 void __weak hw_perf_event_setup_online(int cpu) { barrier(); }
101
102 int __weak
103 hw_perf_group_sched_in(struct perf_event *group_leader,
104                struct perf_cpu_context *cpuctx,
105                struct perf_event_context *ctx, int cpu)
106 {
107         return 0;
108 }
109
110 void __weak perf_event_print_debug(void)        { }
111
112 static DEFINE_PER_CPU(int, perf_disable_count);
113
114 void __perf_disable(void)
115 {
116         __get_cpu_var(perf_disable_count)++;
117 }
118
119 bool __perf_enable(void)
120 {
121         return !--__get_cpu_var(perf_disable_count);
122 }
123
124 void perf_disable(void)
125 {
126         __perf_disable();
127         hw_perf_disable();
128 }
129
130 void perf_enable(void)
131 {
132         if (__perf_enable())
133                 hw_perf_enable();
134 }
135
136 static void get_ctx(struct perf_event_context *ctx)
137 {
138         WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
139 }
140
141 static void free_ctx(struct rcu_head *head)
142 {
143         struct perf_event_context *ctx;
144
145         ctx = container_of(head, struct perf_event_context, rcu_head);
146         kfree(ctx);
147 }
148
149 static void put_ctx(struct perf_event_context *ctx)
150 {
151         if (atomic_dec_and_test(&ctx->refcount)) {
152                 if (ctx->parent_ctx)
153                         put_ctx(ctx->parent_ctx);
154                 if (ctx->task)
155                         put_task_struct(ctx->task);
156                 call_rcu(&ctx->rcu_head, free_ctx);
157         }
158 }
159
160 static void unclone_ctx(struct perf_event_context *ctx)
161 {
162         if (ctx->parent_ctx) {
163                 put_ctx(ctx->parent_ctx);
164                 ctx->parent_ctx = NULL;
165         }
166 }
167
168 /*
169  * If we inherit events we want to return the parent event id
170  * to userspace.
171  */
172 static u64 primary_event_id(struct perf_event *event)
173 {
174         u64 id = event->id;
175
176         if (event->parent)
177                 id = event->parent->id;
178
179         return id;
180 }
181
182 /*
183  * Get the perf_event_context for a task and lock it.
184  * This has to cope with with the fact that until it is locked,
185  * the context could get moved to another task.
186  */
187 static struct perf_event_context *
188 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
189 {
190         struct perf_event_context *ctx;
191
192         rcu_read_lock();
193  retry:
194         ctx = rcu_dereference(task->perf_event_ctxp);
195         if (ctx) {
196                 /*
197                  * If this context is a clone of another, it might
198                  * get swapped for another underneath us by
199                  * perf_event_task_sched_out, though the
200                  * rcu_read_lock() protects us from any context
201                  * getting freed.  Lock the context and check if it
202                  * got swapped before we could get the lock, and retry
203                  * if so.  If we locked the right context, then it
204                  * can't get swapped on us any more.
205                  */
206                 spin_lock_irqsave(&ctx->lock, *flags);
207                 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
208                         spin_unlock_irqrestore(&ctx->lock, *flags);
209                         goto retry;
210                 }
211
212                 if (!atomic_inc_not_zero(&ctx->refcount)) {
213                         spin_unlock_irqrestore(&ctx->lock, *flags);
214                         ctx = NULL;
215                 }
216         }
217         rcu_read_unlock();
218         return ctx;
219 }
220
221 /*
222  * Get the context for a task and increment its pin_count so it
223  * can't get swapped to another task.  This also increments its
224  * reference count so that the context can't get freed.
225  */
226 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
227 {
228         struct perf_event_context *ctx;
229         unsigned long flags;
230
231         ctx = perf_lock_task_context(task, &flags);
232         if (ctx) {
233                 ++ctx->pin_count;
234                 spin_unlock_irqrestore(&ctx->lock, flags);
235         }
236         return ctx;
237 }
238
239 static void perf_unpin_context(struct perf_event_context *ctx)
240 {
241         unsigned long flags;
242
243         spin_lock_irqsave(&ctx->lock, flags);
244         --ctx->pin_count;
245         spin_unlock_irqrestore(&ctx->lock, flags);
246         put_ctx(ctx);
247 }
248
249 static inline u64 perf_clock(void)
250 {
251         return cpu_clock(smp_processor_id());
252 }
253
254 /*
255  * Update the record of the current time in a context.
256  */
257 static void update_context_time(struct perf_event_context *ctx)
258 {
259         u64 now = perf_clock();
260
261         ctx->time += now - ctx->timestamp;
262         ctx->timestamp = now;
263 }
264
265 /*
266  * Update the total_time_enabled and total_time_running fields for a event.
267  */
268 static void update_event_times(struct perf_event *event)
269 {
270         struct perf_event_context *ctx = event->ctx;
271         u64 run_end;
272
273         if (event->state < PERF_EVENT_STATE_INACTIVE ||
274             event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
275                 return;
276
277         if (ctx->is_active)
278                 run_end = ctx->time;
279         else
280                 run_end = event->tstamp_stopped;
281
282         event->total_time_enabled = run_end - event->tstamp_enabled;
283
284         if (event->state == PERF_EVENT_STATE_INACTIVE)
285                 run_end = event->tstamp_stopped;
286         else
287                 run_end = ctx->time;
288
289         event->total_time_running = run_end - event->tstamp_running;
290 }
291
292 /*
293  * Add a event from the lists for its context.
294  * Must be called with ctx->mutex and ctx->lock held.
295  */
296 static void
297 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
298 {
299         struct perf_event *group_leader = event->group_leader;
300
301         /*
302          * Depending on whether it is a standalone or sibling event,
303          * add it straight to the context's event list, or to the group
304          * leader's sibling list:
305          */
306         if (group_leader == event)
307                 list_add_tail(&event->group_entry, &ctx->group_list);
308         else {
309                 list_add_tail(&event->group_entry, &group_leader->sibling_list);
310                 group_leader->nr_siblings++;
311         }
312
313         list_add_rcu(&event->event_entry, &ctx->event_list);
314         ctx->nr_events++;
315         if (event->attr.inherit_stat)
316                 ctx->nr_stat++;
317 }
318
319 /*
320  * Remove a event from the lists for its context.
321  * Must be called with ctx->mutex and ctx->lock held.
322  */
323 static void
324 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
325 {
326         struct perf_event *sibling, *tmp;
327
328         if (list_empty(&event->group_entry))
329                 return;
330         ctx->nr_events--;
331         if (event->attr.inherit_stat)
332                 ctx->nr_stat--;
333
334         list_del_init(&event->group_entry);
335         list_del_rcu(&event->event_entry);
336
337         if (event->group_leader != event)
338                 event->group_leader->nr_siblings--;
339
340         update_event_times(event);
341         event->state = PERF_EVENT_STATE_OFF;
342
343         /*
344          * If this was a group event with sibling events then
345          * upgrade the siblings to singleton events by adding them
346          * to the context list directly:
347          */
348         list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
349
350                 list_move_tail(&sibling->group_entry, &ctx->group_list);
351                 sibling->group_leader = sibling;
352         }
353 }
354
355 static void
356 event_sched_out(struct perf_event *event,
357                   struct perf_cpu_context *cpuctx,
358                   struct perf_event_context *ctx)
359 {
360         if (event->state != PERF_EVENT_STATE_ACTIVE)
361                 return;
362
363         event->state = PERF_EVENT_STATE_INACTIVE;
364         if (event->pending_disable) {
365                 event->pending_disable = 0;
366                 event->state = PERF_EVENT_STATE_OFF;
367         }
368         event->tstamp_stopped = ctx->time;
369         event->pmu->disable(event);
370         event->oncpu = -1;
371
372         if (!is_software_event(event))
373                 cpuctx->active_oncpu--;
374         ctx->nr_active--;
375         if (event->attr.exclusive || !cpuctx->active_oncpu)
376                 cpuctx->exclusive = 0;
377 }
378
379 static void
380 group_sched_out(struct perf_event *group_event,
381                 struct perf_cpu_context *cpuctx,
382                 struct perf_event_context *ctx)
383 {
384         struct perf_event *event;
385
386         if (group_event->state != PERF_EVENT_STATE_ACTIVE)
387                 return;
388
389         event_sched_out(group_event, cpuctx, ctx);
390
391         /*
392          * Schedule out siblings (if any):
393          */
394         list_for_each_entry(event, &group_event->sibling_list, group_entry)
395                 event_sched_out(event, cpuctx, ctx);
396
397         if (group_event->attr.exclusive)
398                 cpuctx->exclusive = 0;
399 }
400
401 /*
402  * Cross CPU call to remove a performance event
403  *
404  * We disable the event on the hardware level first. After that we
405  * remove it from the context list.
406  */
407 static void __perf_event_remove_from_context(void *info)
408 {
409         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
410         struct perf_event *event = info;
411         struct perf_event_context *ctx = event->ctx;
412
413         /*
414          * If this is a task context, we need to check whether it is
415          * the current task context of this cpu. If not it has been
416          * scheduled out before the smp call arrived.
417          */
418         if (ctx->task && cpuctx->task_ctx != ctx)
419                 return;
420
421         spin_lock(&ctx->lock);
422         /*
423          * Protect the list operation against NMI by disabling the
424          * events on a global level.
425          */
426         perf_disable();
427
428         event_sched_out(event, cpuctx, ctx);
429
430         list_del_event(event, ctx);
431
432         if (!ctx->task) {
433                 /*
434                  * Allow more per task events with respect to the
435                  * reservation:
436                  */
437                 cpuctx->max_pertask =
438                         min(perf_max_events - ctx->nr_events,
439                             perf_max_events - perf_reserved_percpu);
440         }
441
442         perf_enable();
443         spin_unlock(&ctx->lock);
444 }
445
446
447 /*
448  * Remove the event from a task's (or a CPU's) list of events.
449  *
450  * Must be called with ctx->mutex held.
451  *
452  * CPU events are removed with a smp call. For task events we only
453  * call when the task is on a CPU.
454  *
455  * If event->ctx is a cloned context, callers must make sure that
456  * every task struct that event->ctx->task could possibly point to
457  * remains valid.  This is OK when called from perf_release since
458  * that only calls us on the top-level context, which can't be a clone.
459  * When called from perf_event_exit_task, it's OK because the
460  * context has been detached from its task.
461  */
462 static void perf_event_remove_from_context(struct perf_event *event)
463 {
464         struct perf_event_context *ctx = event->ctx;
465         struct task_struct *task = ctx->task;
466
467         if (!task) {
468                 /*
469                  * Per cpu events are removed via an smp call and
470                  * the removal is always sucessful.
471                  */
472                 smp_call_function_single(event->cpu,
473                                          __perf_event_remove_from_context,
474                                          event, 1);
475                 return;
476         }
477
478 retry:
479         task_oncpu_function_call(task, __perf_event_remove_from_context,
480                                  event);
481
482         spin_lock_irq(&ctx->lock);
483         /*
484          * If the context is active we need to retry the smp call.
485          */
486         if (ctx->nr_active && !list_empty(&event->group_entry)) {
487                 spin_unlock_irq(&ctx->lock);
488                 goto retry;
489         }
490
491         /*
492          * The lock prevents that this context is scheduled in so we
493          * can remove the event safely, if the call above did not
494          * succeed.
495          */
496         if (!list_empty(&event->group_entry))
497                 list_del_event(event, ctx);
498         spin_unlock_irq(&ctx->lock);
499 }
500
501 /*
502  * Update total_time_enabled and total_time_running for all events in a group.
503  */
504 static void update_group_times(struct perf_event *leader)
505 {
506         struct perf_event *event;
507
508         update_event_times(leader);
509         list_for_each_entry(event, &leader->sibling_list, group_entry)
510                 update_event_times(event);
511 }
512
513 /*
514  * Cross CPU call to disable a performance event
515  */
516 static void __perf_event_disable(void *info)
517 {
518         struct perf_event *event = info;
519         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
520         struct perf_event_context *ctx = event->ctx;
521
522         /*
523          * If this is a per-task event, need to check whether this
524          * event's task is the current task on this cpu.
525          */
526         if (ctx->task && cpuctx->task_ctx != ctx)
527                 return;
528
529         spin_lock(&ctx->lock);
530
531         /*
532          * If the event is on, turn it off.
533          * If it is in error state, leave it in error state.
534          */
535         if (event->state >= PERF_EVENT_STATE_INACTIVE) {
536                 update_context_time(ctx);
537                 update_group_times(event);
538                 if (event == event->group_leader)
539                         group_sched_out(event, cpuctx, ctx);
540                 else
541                         event_sched_out(event, cpuctx, ctx);
542                 event->state = PERF_EVENT_STATE_OFF;
543         }
544
545         spin_unlock(&ctx->lock);
546 }
547
548 /*
549  * Disable a event.
550  *
551  * If event->ctx is a cloned context, callers must make sure that
552  * every task struct that event->ctx->task could possibly point to
553  * remains valid.  This condition is satisifed when called through
554  * perf_event_for_each_child or perf_event_for_each because they
555  * hold the top-level event's child_mutex, so any descendant that
556  * goes to exit will block in sync_child_event.
557  * When called from perf_pending_event it's OK because event->ctx
558  * is the current context on this CPU and preemption is disabled,
559  * hence we can't get into perf_event_task_sched_out for this context.
560  */
561 static void perf_event_disable(struct perf_event *event)
562 {
563         struct perf_event_context *ctx = event->ctx;
564         struct task_struct *task = ctx->task;
565
566         if (!task) {
567                 /*
568                  * Disable the event on the cpu that it's on
569                  */
570                 smp_call_function_single(event->cpu, __perf_event_disable,
571                                          event, 1);
572                 return;
573         }
574
575  retry:
576         task_oncpu_function_call(task, __perf_event_disable, event);
577
578         spin_lock_irq(&ctx->lock);
579         /*
580          * If the event is still active, we need to retry the cross-call.
581          */
582         if (event->state == PERF_EVENT_STATE_ACTIVE) {
583                 spin_unlock_irq(&ctx->lock);
584                 goto retry;
585         }
586
587         /*
588          * Since we have the lock this context can't be scheduled
589          * in, so we can change the state safely.
590          */
591         if (event->state == PERF_EVENT_STATE_INACTIVE) {
592                 update_group_times(event);
593                 event->state = PERF_EVENT_STATE_OFF;
594         }
595
596         spin_unlock_irq(&ctx->lock);
597 }
598
599 static int
600 event_sched_in(struct perf_event *event,
601                  struct perf_cpu_context *cpuctx,
602                  struct perf_event_context *ctx,
603                  int cpu)
604 {
605         if (event->state <= PERF_EVENT_STATE_OFF)
606                 return 0;
607
608         event->state = PERF_EVENT_STATE_ACTIVE;
609         event->oncpu = cpu;     /* TODO: put 'cpu' into cpuctx->cpu */
610         /*
611          * The new state must be visible before we turn it on in the hardware:
612          */
613         smp_wmb();
614
615         if (event->pmu->enable(event)) {
616                 event->state = PERF_EVENT_STATE_INACTIVE;
617                 event->oncpu = -1;
618                 return -EAGAIN;
619         }
620
621         event->tstamp_running += ctx->time - event->tstamp_stopped;
622
623         if (!is_software_event(event))
624                 cpuctx->active_oncpu++;
625         ctx->nr_active++;
626
627         if (event->attr.exclusive)
628                 cpuctx->exclusive = 1;
629
630         return 0;
631 }
632
633 static int
634 group_sched_in(struct perf_event *group_event,
635                struct perf_cpu_context *cpuctx,
636                struct perf_event_context *ctx,
637                int cpu)
638 {
639         struct perf_event *event, *partial_group;
640         int ret;
641
642         if (group_event->state == PERF_EVENT_STATE_OFF)
643                 return 0;
644
645         ret = hw_perf_group_sched_in(group_event, cpuctx, ctx, cpu);
646         if (ret)
647                 return ret < 0 ? ret : 0;
648
649         if (event_sched_in(group_event, cpuctx, ctx, cpu))
650                 return -EAGAIN;
651
652         /*
653          * Schedule in siblings as one group (if any):
654          */
655         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
656                 if (event_sched_in(event, cpuctx, ctx, cpu)) {
657                         partial_group = event;
658                         goto group_error;
659                 }
660         }
661
662         return 0;
663
664 group_error:
665         /*
666          * Groups can be scheduled in as one unit only, so undo any
667          * partial group before returning:
668          */
669         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
670                 if (event == partial_group)
671                         break;
672                 event_sched_out(event, cpuctx, ctx);
673         }
674         event_sched_out(group_event, cpuctx, ctx);
675
676         return -EAGAIN;
677 }
678
679 /*
680  * Return 1 for a group consisting entirely of software events,
681  * 0 if the group contains any hardware events.
682  */
683 static int is_software_only_group(struct perf_event *leader)
684 {
685         struct perf_event *event;
686
687         if (!is_software_event(leader))
688                 return 0;
689
690         list_for_each_entry(event, &leader->sibling_list, group_entry)
691                 if (!is_software_event(event))
692                         return 0;
693
694         return 1;
695 }
696
697 /*
698  * Work out whether we can put this event group on the CPU now.
699  */
700 static int group_can_go_on(struct perf_event *event,
701                            struct perf_cpu_context *cpuctx,
702                            int can_add_hw)
703 {
704         /*
705          * Groups consisting entirely of software events can always go on.
706          */
707         if (is_software_only_group(event))
708                 return 1;
709         /*
710          * If an exclusive group is already on, no other hardware
711          * events can go on.
712          */
713         if (cpuctx->exclusive)
714                 return 0;
715         /*
716          * If this group is exclusive and there are already
717          * events on the CPU, it can't go on.
718          */
719         if (event->attr.exclusive && cpuctx->active_oncpu)
720                 return 0;
721         /*
722          * Otherwise, try to add it if all previous groups were able
723          * to go on.
724          */
725         return can_add_hw;
726 }
727
728 static void add_event_to_ctx(struct perf_event *event,
729                                struct perf_event_context *ctx)
730 {
731         list_add_event(event, ctx);
732         event->tstamp_enabled = ctx->time;
733         event->tstamp_running = ctx->time;
734         event->tstamp_stopped = ctx->time;
735 }
736
737 /*
738  * Cross CPU call to install and enable a performance event
739  *
740  * Must be called with ctx->mutex held
741  */
742 static void __perf_install_in_context(void *info)
743 {
744         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
745         struct perf_event *event = info;
746         struct perf_event_context *ctx = event->ctx;
747         struct perf_event *leader = event->group_leader;
748         int cpu = smp_processor_id();
749         int err;
750
751         /*
752          * If this is a task context, we need to check whether it is
753          * the current task context of this cpu. If not it has been
754          * scheduled out before the smp call arrived.
755          * Or possibly this is the right context but it isn't
756          * on this cpu because it had no events.
757          */
758         if (ctx->task && cpuctx->task_ctx != ctx) {
759                 if (cpuctx->task_ctx || ctx->task != current)
760                         return;
761                 cpuctx->task_ctx = ctx;
762         }
763
764         spin_lock(&ctx->lock);
765         ctx->is_active = 1;
766         update_context_time(ctx);
767
768         /*
769          * Protect the list operation against NMI by disabling the
770          * events on a global level. NOP for non NMI based events.
771          */
772         perf_disable();
773
774         add_event_to_ctx(event, ctx);
775
776         /*
777          * Don't put the event on if it is disabled or if
778          * it is in a group and the group isn't on.
779          */
780         if (event->state != PERF_EVENT_STATE_INACTIVE ||
781             (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
782                 goto unlock;
783
784         /*
785          * An exclusive event can't go on if there are already active
786          * hardware events, and no hardware event can go on if there
787          * is already an exclusive event on.
788          */
789         if (!group_can_go_on(event, cpuctx, 1))
790                 err = -EEXIST;
791         else
792                 err = event_sched_in(event, cpuctx, ctx, cpu);
793
794         if (err) {
795                 /*
796                  * This event couldn't go on.  If it is in a group
797                  * then we have to pull the whole group off.
798                  * If the event group is pinned then put it in error state.
799                  */
800                 if (leader != event)
801                         group_sched_out(leader, cpuctx, ctx);
802                 if (leader->attr.pinned) {
803                         update_group_times(leader);
804                         leader->state = PERF_EVENT_STATE_ERROR;
805                 }
806         }
807
808         if (!err && !ctx->task && cpuctx->max_pertask)
809                 cpuctx->max_pertask--;
810
811  unlock:
812         perf_enable();
813
814         spin_unlock(&ctx->lock);
815 }
816
817 /*
818  * Attach a performance event to a context
819  *
820  * First we add the event to the list with the hardware enable bit
821  * in event->hw_config cleared.
822  *
823  * If the event is attached to a task which is on a CPU we use a smp
824  * call to enable it in the task context. The task might have been
825  * scheduled away, but we check this in the smp call again.
826  *
827  * Must be called with ctx->mutex held.
828  */
829 static void
830 perf_install_in_context(struct perf_event_context *ctx,
831                         struct perf_event *event,
832                         int cpu)
833 {
834         struct task_struct *task = ctx->task;
835
836         if (!task) {
837                 /*
838                  * Per cpu events are installed via an smp call and
839                  * the install is always sucessful.
840                  */
841                 smp_call_function_single(cpu, __perf_install_in_context,
842                                          event, 1);
843                 return;
844         }
845
846 retry:
847         task_oncpu_function_call(task, __perf_install_in_context,
848                                  event);
849
850         spin_lock_irq(&ctx->lock);
851         /*
852          * we need to retry the smp call.
853          */
854         if (ctx->is_active && list_empty(&event->group_entry)) {
855                 spin_unlock_irq(&ctx->lock);
856                 goto retry;
857         }
858
859         /*
860          * The lock prevents that this context is scheduled in so we
861          * can add the event safely, if it the call above did not
862          * succeed.
863          */
864         if (list_empty(&event->group_entry))
865                 add_event_to_ctx(event, ctx);
866         spin_unlock_irq(&ctx->lock);
867 }
868
869 /*
870  * Put a event into inactive state and update time fields.
871  * Enabling the leader of a group effectively enables all
872  * the group members that aren't explicitly disabled, so we
873  * have to update their ->tstamp_enabled also.
874  * Note: this works for group members as well as group leaders
875  * since the non-leader members' sibling_lists will be empty.
876  */
877 static void __perf_event_mark_enabled(struct perf_event *event,
878                                         struct perf_event_context *ctx)
879 {
880         struct perf_event *sub;
881
882         event->state = PERF_EVENT_STATE_INACTIVE;
883         event->tstamp_enabled = ctx->time - event->total_time_enabled;
884         list_for_each_entry(sub, &event->sibling_list, group_entry)
885                 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
886                         sub->tstamp_enabled =
887                                 ctx->time - sub->total_time_enabled;
888 }
889
890 /*
891  * Cross CPU call to enable a performance event
892  */
893 static void __perf_event_enable(void *info)
894 {
895         struct perf_event *event = info;
896         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
897         struct perf_event_context *ctx = event->ctx;
898         struct perf_event *leader = event->group_leader;
899         int err;
900
901         /*
902          * If this is a per-task event, need to check whether this
903          * event's task is the current task on this cpu.
904          */
905         if (ctx->task && cpuctx->task_ctx != ctx) {
906                 if (cpuctx->task_ctx || ctx->task != current)
907                         return;
908                 cpuctx->task_ctx = ctx;
909         }
910
911         spin_lock(&ctx->lock);
912         ctx->is_active = 1;
913         update_context_time(ctx);
914
915         if (event->state >= PERF_EVENT_STATE_INACTIVE)
916                 goto unlock;
917         __perf_event_mark_enabled(event, ctx);
918
919         /*
920          * If the event is in a group and isn't the group leader,
921          * then don't put it on unless the group is on.
922          */
923         if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
924                 goto unlock;
925
926         if (!group_can_go_on(event, cpuctx, 1)) {
927                 err = -EEXIST;
928         } else {
929                 perf_disable();
930                 if (event == leader)
931                         err = group_sched_in(event, cpuctx, ctx,
932                                              smp_processor_id());
933                 else
934                         err = event_sched_in(event, cpuctx, ctx,
935                                                smp_processor_id());
936                 perf_enable();
937         }
938
939         if (err) {
940                 /*
941                  * If this event can't go on and it's part of a
942                  * group, then the whole group has to come off.
943                  */
944                 if (leader != event)
945                         group_sched_out(leader, cpuctx, ctx);
946                 if (leader->attr.pinned) {
947                         update_group_times(leader);
948                         leader->state = PERF_EVENT_STATE_ERROR;
949                 }
950         }
951
952  unlock:
953         spin_unlock(&ctx->lock);
954 }
955
956 /*
957  * Enable a event.
958  *
959  * If event->ctx is a cloned context, callers must make sure that
960  * every task struct that event->ctx->task could possibly point to
961  * remains valid.  This condition is satisfied when called through
962  * perf_event_for_each_child or perf_event_for_each as described
963  * for perf_event_disable.
964  */
965 static void perf_event_enable(struct perf_event *event)
966 {
967         struct perf_event_context *ctx = event->ctx;
968         struct task_struct *task = ctx->task;
969
970         if (!task) {
971                 /*
972                  * Enable the event on the cpu that it's on
973                  */
974                 smp_call_function_single(event->cpu, __perf_event_enable,
975                                          event, 1);
976                 return;
977         }
978
979         spin_lock_irq(&ctx->lock);
980         if (event->state >= PERF_EVENT_STATE_INACTIVE)
981                 goto out;
982
983         /*
984          * If the event is in error state, clear that first.
985          * That way, if we see the event in error state below, we
986          * know that it has gone back into error state, as distinct
987          * from the task having been scheduled away before the
988          * cross-call arrived.
989          */
990         if (event->state == PERF_EVENT_STATE_ERROR)
991                 event->state = PERF_EVENT_STATE_OFF;
992
993  retry:
994         spin_unlock_irq(&ctx->lock);
995         task_oncpu_function_call(task, __perf_event_enable, event);
996
997         spin_lock_irq(&ctx->lock);
998
999         /*
1000          * If the context is active and the event is still off,
1001          * we need to retry the cross-call.
1002          */
1003         if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1004                 goto retry;
1005
1006         /*
1007          * Since we have the lock this context can't be scheduled
1008          * in, so we can change the state safely.
1009          */
1010         if (event->state == PERF_EVENT_STATE_OFF)
1011                 __perf_event_mark_enabled(event, ctx);
1012
1013  out:
1014         spin_unlock_irq(&ctx->lock);
1015 }
1016
1017 static int perf_event_refresh(struct perf_event *event, int refresh)
1018 {
1019         /*
1020          * not supported on inherited events
1021          */
1022         if (event->attr.inherit)
1023                 return -EINVAL;
1024
1025         atomic_add(refresh, &event->event_limit);
1026         perf_event_enable(event);
1027
1028         return 0;
1029 }
1030
1031 void __perf_event_sched_out(struct perf_event_context *ctx,
1032                               struct perf_cpu_context *cpuctx)
1033 {
1034         struct perf_event *event;
1035
1036         spin_lock(&ctx->lock);
1037         ctx->is_active = 0;
1038         if (likely(!ctx->nr_events))
1039                 goto out;
1040         update_context_time(ctx);
1041
1042         perf_disable();
1043         if (ctx->nr_active) {
1044                 list_for_each_entry(event, &ctx->group_list, group_entry)
1045                         group_sched_out(event, cpuctx, ctx);
1046         }
1047         perf_enable();
1048  out:
1049         spin_unlock(&ctx->lock);
1050 }
1051
1052 /*
1053  * Test whether two contexts are equivalent, i.e. whether they
1054  * have both been cloned from the same version of the same context
1055  * and they both have the same number of enabled events.
1056  * If the number of enabled events is the same, then the set
1057  * of enabled events should be the same, because these are both
1058  * inherited contexts, therefore we can't access individual events
1059  * in them directly with an fd; we can only enable/disable all
1060  * events via prctl, or enable/disable all events in a family
1061  * via ioctl, which will have the same effect on both contexts.
1062  */
1063 static int context_equiv(struct perf_event_context *ctx1,
1064                          struct perf_event_context *ctx2)
1065 {
1066         return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1067                 && ctx1->parent_gen == ctx2->parent_gen
1068                 && !ctx1->pin_count && !ctx2->pin_count;
1069 }
1070
1071 static void __perf_event_sync_stat(struct perf_event *event,
1072                                      struct perf_event *next_event)
1073 {
1074         u64 value;
1075
1076         if (!event->attr.inherit_stat)
1077                 return;
1078
1079         /*
1080          * Update the event value, we cannot use perf_event_read()
1081          * because we're in the middle of a context switch and have IRQs
1082          * disabled, which upsets smp_call_function_single(), however
1083          * we know the event must be on the current CPU, therefore we
1084          * don't need to use it.
1085          */
1086         switch (event->state) {
1087         case PERF_EVENT_STATE_ACTIVE:
1088                 event->pmu->read(event);
1089                 /* fall-through */
1090
1091         case PERF_EVENT_STATE_INACTIVE:
1092                 update_event_times(event);
1093                 break;
1094
1095         default:
1096                 break;
1097         }
1098
1099         /*
1100          * In order to keep per-task stats reliable we need to flip the event
1101          * values when we flip the contexts.
1102          */
1103         value = atomic64_read(&next_event->count);
1104         value = atomic64_xchg(&event->count, value);
1105         atomic64_set(&next_event->count, value);
1106
1107         swap(event->total_time_enabled, next_event->total_time_enabled);
1108         swap(event->total_time_running, next_event->total_time_running);
1109
1110         /*
1111          * Since we swizzled the values, update the user visible data too.
1112          */
1113         perf_event_update_userpage(event);
1114         perf_event_update_userpage(next_event);
1115 }
1116
1117 #define list_next_entry(pos, member) \
1118         list_entry(pos->member.next, typeof(*pos), member)
1119
1120 static void perf_event_sync_stat(struct perf_event_context *ctx,
1121                                    struct perf_event_context *next_ctx)
1122 {
1123         struct perf_event *event, *next_event;
1124
1125         if (!ctx->nr_stat)
1126                 return;
1127
1128         update_context_time(ctx);
1129
1130         event = list_first_entry(&ctx->event_list,
1131                                    struct perf_event, event_entry);
1132
1133         next_event = list_first_entry(&next_ctx->event_list,
1134                                         struct perf_event, event_entry);
1135
1136         while (&event->event_entry != &ctx->event_list &&
1137                &next_event->event_entry != &next_ctx->event_list) {
1138
1139                 __perf_event_sync_stat(event, next_event);
1140
1141                 event = list_next_entry(event, event_entry);
1142                 next_event = list_next_entry(next_event, event_entry);
1143         }
1144 }
1145
1146 /*
1147  * Called from scheduler to remove the events of the current task,
1148  * with interrupts disabled.
1149  *
1150  * We stop each event and update the event value in event->count.
1151  *
1152  * This does not protect us against NMI, but disable()
1153  * sets the disabled bit in the control field of event _before_
1154  * accessing the event control register. If a NMI hits, then it will
1155  * not restart the event.
1156  */
1157 void perf_event_task_sched_out(struct task_struct *task,
1158                                  struct task_struct *next, int cpu)
1159 {
1160         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1161         struct perf_event_context *ctx = task->perf_event_ctxp;
1162         struct perf_event_context *next_ctx;
1163         struct perf_event_context *parent;
1164         struct pt_regs *regs;
1165         int do_switch = 1;
1166
1167         regs = task_pt_regs(task);
1168         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1169
1170         if (likely(!ctx || !cpuctx->task_ctx))
1171                 return;
1172
1173         rcu_read_lock();
1174         parent = rcu_dereference(ctx->parent_ctx);
1175         next_ctx = next->perf_event_ctxp;
1176         if (parent && next_ctx &&
1177             rcu_dereference(next_ctx->parent_ctx) == parent) {
1178                 /*
1179                  * Looks like the two contexts are clones, so we might be
1180                  * able to optimize the context switch.  We lock both
1181                  * contexts and check that they are clones under the
1182                  * lock (including re-checking that neither has been
1183                  * uncloned in the meantime).  It doesn't matter which
1184                  * order we take the locks because no other cpu could
1185                  * be trying to lock both of these tasks.
1186                  */
1187                 spin_lock(&ctx->lock);
1188                 spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1189                 if (context_equiv(ctx, next_ctx)) {
1190                         /*
1191                          * XXX do we need a memory barrier of sorts
1192                          * wrt to rcu_dereference() of perf_event_ctxp
1193                          */
1194                         task->perf_event_ctxp = next_ctx;
1195                         next->perf_event_ctxp = ctx;
1196                         ctx->task = next;
1197                         next_ctx->task = task;
1198                         do_switch = 0;
1199
1200                         perf_event_sync_stat(ctx, next_ctx);
1201                 }
1202                 spin_unlock(&next_ctx->lock);
1203                 spin_unlock(&ctx->lock);
1204         }
1205         rcu_read_unlock();
1206
1207         if (do_switch) {
1208                 __perf_event_sched_out(ctx, cpuctx);
1209                 cpuctx->task_ctx = NULL;
1210         }
1211 }
1212
1213 /*
1214  * Called with IRQs disabled
1215  */
1216 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1217 {
1218         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1219
1220         if (!cpuctx->task_ctx)
1221                 return;
1222
1223         if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1224                 return;
1225
1226         __perf_event_sched_out(ctx, cpuctx);
1227         cpuctx->task_ctx = NULL;
1228 }
1229
1230 /*
1231  * Called with IRQs disabled
1232  */
1233 static void perf_event_cpu_sched_out(struct perf_cpu_context *cpuctx)
1234 {
1235         __perf_event_sched_out(&cpuctx->ctx, cpuctx);
1236 }
1237
1238 static void
1239 __perf_event_sched_in(struct perf_event_context *ctx,
1240                         struct perf_cpu_context *cpuctx, int cpu)
1241 {
1242         struct perf_event *event;
1243         int can_add_hw = 1;
1244
1245         spin_lock(&ctx->lock);
1246         ctx->is_active = 1;
1247         if (likely(!ctx->nr_events))
1248                 goto out;
1249
1250         ctx->timestamp = perf_clock();
1251
1252         perf_disable();
1253
1254         /*
1255          * First go through the list and put on any pinned groups
1256          * in order to give them the best chance of going on.
1257          */
1258         list_for_each_entry(event, &ctx->group_list, group_entry) {
1259                 if (event->state <= PERF_EVENT_STATE_OFF ||
1260                     !event->attr.pinned)
1261                         continue;
1262                 if (event->cpu != -1 && event->cpu != cpu)
1263                         continue;
1264
1265                 if (group_can_go_on(event, cpuctx, 1))
1266                         group_sched_in(event, cpuctx, ctx, cpu);
1267
1268                 /*
1269                  * If this pinned group hasn't been scheduled,
1270                  * put it in error state.
1271                  */
1272                 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1273                         update_group_times(event);
1274                         event->state = PERF_EVENT_STATE_ERROR;
1275                 }
1276         }
1277
1278         list_for_each_entry(event, &ctx->group_list, group_entry) {
1279                 /*
1280                  * Ignore events in OFF or ERROR state, and
1281                  * ignore pinned events since we did them already.
1282                  */
1283                 if (event->state <= PERF_EVENT_STATE_OFF ||
1284                     event->attr.pinned)
1285                         continue;
1286
1287                 /*
1288                  * Listen to the 'cpu' scheduling filter constraint
1289                  * of events:
1290                  */
1291                 if (event->cpu != -1 && event->cpu != cpu)
1292                         continue;
1293
1294                 if (group_can_go_on(event, cpuctx, can_add_hw))
1295                         if (group_sched_in(event, cpuctx, ctx, cpu))
1296                                 can_add_hw = 0;
1297         }
1298         perf_enable();
1299  out:
1300         spin_unlock(&ctx->lock);
1301 }
1302
1303 /*
1304  * Called from scheduler to add the events of the current task
1305  * with interrupts disabled.
1306  *
1307  * We restore the event value and then enable it.
1308  *
1309  * This does not protect us against NMI, but enable()
1310  * sets the enabled bit in the control field of event _before_
1311  * accessing the event control register. If a NMI hits, then it will
1312  * keep the event running.
1313  */
1314 void perf_event_task_sched_in(struct task_struct *task, int cpu)
1315 {
1316         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1317         struct perf_event_context *ctx = task->perf_event_ctxp;
1318
1319         if (likely(!ctx))
1320                 return;
1321         if (cpuctx->task_ctx == ctx)
1322                 return;
1323         __perf_event_sched_in(ctx, cpuctx, cpu);
1324         cpuctx->task_ctx = ctx;
1325 }
1326
1327 static void perf_event_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
1328 {
1329         struct perf_event_context *ctx = &cpuctx->ctx;
1330
1331         __perf_event_sched_in(ctx, cpuctx, cpu);
1332 }
1333
1334 #define MAX_INTERRUPTS (~0ULL)
1335
1336 static void perf_log_throttle(struct perf_event *event, int enable);
1337
1338 static void perf_adjust_period(struct perf_event *event, u64 events)
1339 {
1340         struct hw_perf_event *hwc = &event->hw;
1341         u64 period, sample_period;
1342         s64 delta;
1343
1344         events *= hwc->sample_period;
1345         period = div64_u64(events, event->attr.sample_freq);
1346
1347         delta = (s64)(period - hwc->sample_period);
1348         delta = (delta + 7) / 8; /* low pass filter */
1349
1350         sample_period = hwc->sample_period + delta;
1351
1352         if (!sample_period)
1353                 sample_period = 1;
1354
1355         hwc->sample_period = sample_period;
1356 }
1357
1358 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1359 {
1360         struct perf_event *event;
1361         struct hw_perf_event *hwc;
1362         u64 interrupts, freq;
1363
1364         spin_lock(&ctx->lock);
1365         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1366                 if (event->state != PERF_EVENT_STATE_ACTIVE)
1367                         continue;
1368
1369                 hwc = &event->hw;
1370
1371                 interrupts = hwc->interrupts;
1372                 hwc->interrupts = 0;
1373
1374                 /*
1375                  * unthrottle events on the tick
1376                  */
1377                 if (interrupts == MAX_INTERRUPTS) {
1378                         perf_log_throttle(event, 1);
1379                         event->pmu->unthrottle(event);
1380                         interrupts = 2*sysctl_perf_event_sample_rate/HZ;
1381                 }
1382
1383                 if (!event->attr.freq || !event->attr.sample_freq)
1384                         continue;
1385
1386                 /*
1387                  * if the specified freq < HZ then we need to skip ticks
1388                  */
1389                 if (event->attr.sample_freq < HZ) {
1390                         freq = event->attr.sample_freq;
1391
1392                         hwc->freq_count += freq;
1393                         hwc->freq_interrupts += interrupts;
1394
1395                         if (hwc->freq_count < HZ)
1396                                 continue;
1397
1398                         interrupts = hwc->freq_interrupts;
1399                         hwc->freq_interrupts = 0;
1400                         hwc->freq_count -= HZ;
1401                 } else
1402                         freq = HZ;
1403
1404                 perf_adjust_period(event, freq * interrupts);
1405
1406                 /*
1407                  * In order to avoid being stalled by an (accidental) huge
1408                  * sample period, force reset the sample period if we didn't
1409                  * get any events in this freq period.
1410                  */
1411                 if (!interrupts) {
1412                         perf_disable();
1413                         event->pmu->disable(event);
1414                         atomic64_set(&hwc->period_left, 0);
1415                         event->pmu->enable(event);
1416                         perf_enable();
1417                 }
1418         }
1419         spin_unlock(&ctx->lock);
1420 }
1421
1422 /*
1423  * Round-robin a context's events:
1424  */
1425 static void rotate_ctx(struct perf_event_context *ctx)
1426 {
1427         struct perf_event *event;
1428
1429         if (!ctx->nr_events)
1430                 return;
1431
1432         spin_lock(&ctx->lock);
1433         /*
1434          * Rotate the first entry last (works just fine for group events too):
1435          */
1436         perf_disable();
1437         list_for_each_entry(event, &ctx->group_list, group_entry) {
1438                 list_move_tail(&event->group_entry, &ctx->group_list);
1439                 break;
1440         }
1441         perf_enable();
1442
1443         spin_unlock(&ctx->lock);
1444 }
1445
1446 void perf_event_task_tick(struct task_struct *curr, int cpu)
1447 {
1448         struct perf_cpu_context *cpuctx;
1449         struct perf_event_context *ctx;
1450
1451         if (!atomic_read(&nr_events))
1452                 return;
1453
1454         cpuctx = &per_cpu(perf_cpu_context, cpu);
1455         ctx = curr->perf_event_ctxp;
1456
1457         perf_ctx_adjust_freq(&cpuctx->ctx);
1458         if (ctx)
1459                 perf_ctx_adjust_freq(ctx);
1460
1461         perf_event_cpu_sched_out(cpuctx);
1462         if (ctx)
1463                 __perf_event_task_sched_out(ctx);
1464
1465         rotate_ctx(&cpuctx->ctx);
1466         if (ctx)
1467                 rotate_ctx(ctx);
1468
1469         perf_event_cpu_sched_in(cpuctx, cpu);
1470         if (ctx)
1471                 perf_event_task_sched_in(curr, cpu);
1472 }
1473
1474 /*
1475  * Enable all of a task's events that have been marked enable-on-exec.
1476  * This expects task == current.
1477  */
1478 static void perf_event_enable_on_exec(struct task_struct *task)
1479 {
1480         struct perf_event_context *ctx;
1481         struct perf_event *event;
1482         unsigned long flags;
1483         int enabled = 0;
1484
1485         local_irq_save(flags);
1486         ctx = task->perf_event_ctxp;
1487         if (!ctx || !ctx->nr_events)
1488                 goto out;
1489
1490         __perf_event_task_sched_out(ctx);
1491
1492         spin_lock(&ctx->lock);
1493
1494         list_for_each_entry(event, &ctx->group_list, group_entry) {
1495                 if (!event->attr.enable_on_exec)
1496                         continue;
1497                 event->attr.enable_on_exec = 0;
1498                 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1499                         continue;
1500                 __perf_event_mark_enabled(event, ctx);
1501                 enabled = 1;
1502         }
1503
1504         /*
1505          * Unclone this context if we enabled any event.
1506          */
1507         if (enabled)
1508                 unclone_ctx(ctx);
1509
1510         spin_unlock(&ctx->lock);
1511
1512         perf_event_task_sched_in(task, smp_processor_id());
1513  out:
1514         local_irq_restore(flags);
1515 }
1516
1517 /*
1518  * Cross CPU call to read the hardware event
1519  */
1520 static void __perf_event_read(void *info)
1521 {
1522         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1523         struct perf_event *event = info;
1524         struct perf_event_context *ctx = event->ctx;
1525
1526         /*
1527          * If this is a task context, we need to check whether it is
1528          * the current task context of this cpu.  If not it has been
1529          * scheduled out before the smp call arrived.  In that case
1530          * event->count would have been updated to a recent sample
1531          * when the event was scheduled out.
1532          */
1533         if (ctx->task && cpuctx->task_ctx != ctx)
1534                 return;
1535
1536         spin_lock(&ctx->lock);
1537         update_context_time(ctx);
1538         update_event_times(event);
1539         spin_unlock(&ctx->lock);
1540
1541         event->pmu->read(event);
1542 }
1543
1544 static u64 perf_event_read(struct perf_event *event)
1545 {
1546         /*
1547          * If event is enabled and currently active on a CPU, update the
1548          * value in the event structure:
1549          */
1550         if (event->state == PERF_EVENT_STATE_ACTIVE) {
1551                 smp_call_function_single(event->oncpu,
1552                                          __perf_event_read, event, 1);
1553         } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1554                 struct perf_event_context *ctx = event->ctx;
1555                 unsigned long flags;
1556
1557                 spin_lock_irqsave(&ctx->lock, flags);
1558                 update_context_time(ctx);
1559                 update_event_times(event);
1560                 spin_unlock_irqrestore(&ctx->lock, flags);
1561         }
1562
1563         return atomic64_read(&event->count);
1564 }
1565
1566 /*
1567  * Initialize the perf_event context in a task_struct:
1568  */
1569 static void
1570 __perf_event_init_context(struct perf_event_context *ctx,
1571                             struct task_struct *task)
1572 {
1573         memset(ctx, 0, sizeof(*ctx));
1574         spin_lock_init(&ctx->lock);
1575         mutex_init(&ctx->mutex);
1576         INIT_LIST_HEAD(&ctx->group_list);
1577         INIT_LIST_HEAD(&ctx->event_list);
1578         atomic_set(&ctx->refcount, 1);
1579         ctx->task = task;
1580 }
1581
1582 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1583 {
1584         struct perf_event_context *ctx;
1585         struct perf_cpu_context *cpuctx;
1586         struct task_struct *task;
1587         unsigned long flags;
1588         int err;
1589
1590         /*
1591          * If cpu is not a wildcard then this is a percpu event:
1592          */
1593         if (cpu != -1) {
1594                 /* Must be root to operate on a CPU event: */
1595                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1596                         return ERR_PTR(-EACCES);
1597
1598                 if (cpu < 0 || cpu > num_possible_cpus())
1599                         return ERR_PTR(-EINVAL);
1600
1601                 /*
1602                  * We could be clever and allow to attach a event to an
1603                  * offline CPU and activate it when the CPU comes up, but
1604                  * that's for later.
1605                  */
1606                 if (!cpu_isset(cpu, cpu_online_map))
1607                         return ERR_PTR(-ENODEV);
1608
1609                 cpuctx = &per_cpu(perf_cpu_context, cpu);
1610                 ctx = &cpuctx->ctx;
1611                 get_ctx(ctx);
1612
1613                 return ctx;
1614         }
1615
1616         rcu_read_lock();
1617         if (!pid)
1618                 task = current;
1619         else
1620                 task = find_task_by_vpid(pid);
1621         if (task)
1622                 get_task_struct(task);
1623         rcu_read_unlock();
1624
1625         if (!task)
1626                 return ERR_PTR(-ESRCH);
1627
1628         /*
1629          * Can't attach events to a dying task.
1630          */
1631         err = -ESRCH;
1632         if (task->flags & PF_EXITING)
1633                 goto errout;
1634
1635         /* Reuse ptrace permission checks for now. */
1636         err = -EACCES;
1637         if (!ptrace_may_access(task, PTRACE_MODE_READ))
1638                 goto errout;
1639
1640  retry:
1641         ctx = perf_lock_task_context(task, &flags);
1642         if (ctx) {
1643                 unclone_ctx(ctx);
1644                 spin_unlock_irqrestore(&ctx->lock, flags);
1645         }
1646
1647         if (!ctx) {
1648                 ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1649                 err = -ENOMEM;
1650                 if (!ctx)
1651                         goto errout;
1652                 __perf_event_init_context(ctx, task);
1653                 get_ctx(ctx);
1654                 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1655                         /*
1656                          * We raced with some other task; use
1657                          * the context they set.
1658                          */
1659                         kfree(ctx);
1660                         goto retry;
1661                 }
1662                 get_task_struct(task);
1663         }
1664
1665         put_task_struct(task);
1666         return ctx;
1667
1668  errout:
1669         put_task_struct(task);
1670         return ERR_PTR(err);
1671 }
1672
1673 static void perf_event_free_filter(struct perf_event *event);
1674
1675 static void free_event_rcu(struct rcu_head *head)
1676 {
1677         struct perf_event *event;
1678
1679         event = container_of(head, struct perf_event, rcu_head);
1680         if (event->ns)
1681                 put_pid_ns(event->ns);
1682         perf_event_free_filter(event);
1683         kfree(event);
1684 }
1685
1686 static void perf_pending_sync(struct perf_event *event);
1687
1688 static void free_event(struct perf_event *event)
1689 {
1690         perf_pending_sync(event);
1691
1692         if (!event->parent) {
1693                 atomic_dec(&nr_events);
1694                 if (event->attr.mmap)
1695                         atomic_dec(&nr_mmap_events);
1696                 if (event->attr.comm)
1697                         atomic_dec(&nr_comm_events);
1698                 if (event->attr.task)
1699                         atomic_dec(&nr_task_events);
1700         }
1701
1702         if (event->output) {
1703                 fput(event->output->filp);
1704                 event->output = NULL;
1705         }
1706
1707         if (event->destroy)
1708                 event->destroy(event);
1709
1710         put_ctx(event->ctx);
1711         call_rcu(&event->rcu_head, free_event_rcu);
1712 }
1713
1714 int perf_event_release_kernel(struct perf_event *event)
1715 {
1716         struct perf_event_context *ctx = event->ctx;
1717
1718         WARN_ON_ONCE(ctx->parent_ctx);
1719         mutex_lock(&ctx->mutex);
1720         perf_event_remove_from_context(event);
1721         mutex_unlock(&ctx->mutex);
1722
1723         mutex_lock(&event->owner->perf_event_mutex);
1724         list_del_init(&event->owner_entry);
1725         mutex_unlock(&event->owner->perf_event_mutex);
1726         put_task_struct(event->owner);
1727
1728         free_event(event);
1729
1730         return 0;
1731 }
1732 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1733
1734 /*
1735  * Called when the last reference to the file is gone.
1736  */
1737 static int perf_release(struct inode *inode, struct file *file)
1738 {
1739         struct perf_event *event = file->private_data;
1740
1741         file->private_data = NULL;
1742
1743         return perf_event_release_kernel(event);
1744 }
1745
1746 static int perf_event_read_size(struct perf_event *event)
1747 {
1748         int entry = sizeof(u64); /* value */
1749         int size = 0;
1750         int nr = 1;
1751
1752         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1753                 size += sizeof(u64);
1754
1755         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1756                 size += sizeof(u64);
1757
1758         if (event->attr.read_format & PERF_FORMAT_ID)
1759                 entry += sizeof(u64);
1760
1761         if (event->attr.read_format & PERF_FORMAT_GROUP) {
1762                 nr += event->group_leader->nr_siblings;
1763                 size += sizeof(u64);
1764         }
1765
1766         size += entry * nr;
1767
1768         return size;
1769 }
1770
1771 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1772 {
1773         struct perf_event *child;
1774         u64 total = 0;
1775
1776         *enabled = 0;
1777         *running = 0;
1778
1779         mutex_lock(&event->child_mutex);
1780         total += perf_event_read(event);
1781         *enabled += event->total_time_enabled +
1782                         atomic64_read(&event->child_total_time_enabled);
1783         *running += event->total_time_running +
1784                         atomic64_read(&event->child_total_time_running);
1785
1786         list_for_each_entry(child, &event->child_list, child_list) {
1787                 total += perf_event_read(child);
1788                 *enabled += child->total_time_enabled;
1789                 *running += child->total_time_running;
1790         }
1791         mutex_unlock(&event->child_mutex);
1792
1793         return total;
1794 }
1795 EXPORT_SYMBOL_GPL(perf_event_read_value);
1796
1797 static int perf_event_read_group(struct perf_event *event,
1798                                    u64 read_format, char __user *buf)
1799 {
1800         struct perf_event *leader = event->group_leader, *sub;
1801         int n = 0, size = 0, ret = -EFAULT;
1802         struct perf_event_context *ctx = leader->ctx;
1803         u64 values[5];
1804         u64 count, enabled, running;
1805
1806         mutex_lock(&ctx->mutex);
1807         count = perf_event_read_value(leader, &enabled, &running);
1808
1809         values[n++] = 1 + leader->nr_siblings;
1810         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1811                 values[n++] = enabled;
1812         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1813                 values[n++] = running;
1814         values[n++] = count;
1815         if (read_format & PERF_FORMAT_ID)
1816                 values[n++] = primary_event_id(leader);
1817
1818         size = n * sizeof(u64);
1819
1820         if (copy_to_user(buf, values, size))
1821                 goto unlock;
1822
1823         ret = size;
1824
1825         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1826                 n = 0;
1827
1828                 values[n++] = perf_event_read_value(sub, &enabled, &running);
1829                 if (read_format & PERF_FORMAT_ID)
1830                         values[n++] = primary_event_id(sub);
1831
1832                 size = n * sizeof(u64);
1833
1834                 if (copy_to_user(buf + size, values, size)) {
1835                         ret = -EFAULT;
1836                         goto unlock;
1837                 }
1838
1839                 ret += size;
1840         }
1841 unlock:
1842         mutex_unlock(&ctx->mutex);
1843
1844         return ret;
1845 }
1846
1847 static int perf_event_read_one(struct perf_event *event,
1848                                  u64 read_format, char __user *buf)
1849 {
1850         u64 enabled, running;
1851         u64 values[4];
1852         int n = 0;
1853
1854         values[n++] = perf_event_read_value(event, &enabled, &running);
1855         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1856                 values[n++] = enabled;
1857         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1858                 values[n++] = running;
1859         if (read_format & PERF_FORMAT_ID)
1860                 values[n++] = primary_event_id(event);
1861
1862         if (copy_to_user(buf, values, n * sizeof(u64)))
1863                 return -EFAULT;
1864
1865         return n * sizeof(u64);
1866 }
1867
1868 /*
1869  * Read the performance event - simple non blocking version for now
1870  */
1871 static ssize_t
1872 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
1873 {
1874         u64 read_format = event->attr.read_format;
1875         int ret;
1876
1877         /*
1878          * Return end-of-file for a read on a event that is in
1879          * error state (i.e. because it was pinned but it couldn't be
1880          * scheduled on to the CPU at some point).
1881          */
1882         if (event->state == PERF_EVENT_STATE_ERROR)
1883                 return 0;
1884
1885         if (count < perf_event_read_size(event))
1886                 return -ENOSPC;
1887
1888         WARN_ON_ONCE(event->ctx->parent_ctx);
1889         if (read_format & PERF_FORMAT_GROUP)
1890                 ret = perf_event_read_group(event, read_format, buf);
1891         else
1892                 ret = perf_event_read_one(event, read_format, buf);
1893
1894         return ret;
1895 }
1896
1897 static ssize_t
1898 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1899 {
1900         struct perf_event *event = file->private_data;
1901
1902         return perf_read_hw(event, buf, count);
1903 }
1904
1905 static unsigned int perf_poll(struct file *file, poll_table *wait)
1906 {
1907         struct perf_event *event = file->private_data;
1908         struct perf_mmap_data *data;
1909         unsigned int events = POLL_HUP;
1910
1911         rcu_read_lock();
1912         data = rcu_dereference(event->data);
1913         if (data)
1914                 events = atomic_xchg(&data->poll, 0);
1915         rcu_read_unlock();
1916
1917         poll_wait(file, &event->waitq, wait);
1918
1919         return events;
1920 }
1921
1922 static void perf_event_reset(struct perf_event *event)
1923 {
1924         (void)perf_event_read(event);
1925         atomic64_set(&event->count, 0);
1926         perf_event_update_userpage(event);
1927 }
1928
1929 /*
1930  * Holding the top-level event's child_mutex means that any
1931  * descendant process that has inherited this event will block
1932  * in sync_child_event if it goes to exit, thus satisfying the
1933  * task existence requirements of perf_event_enable/disable.
1934  */
1935 static void perf_event_for_each_child(struct perf_event *event,
1936                                         void (*func)(struct perf_event *))
1937 {
1938         struct perf_event *child;
1939
1940         WARN_ON_ONCE(event->ctx->parent_ctx);
1941         mutex_lock(&event->child_mutex);
1942         func(event);
1943         list_for_each_entry(child, &event->child_list, child_list)
1944                 func(child);
1945         mutex_unlock(&event->child_mutex);
1946 }
1947
1948 static void perf_event_for_each(struct perf_event *event,
1949                                   void (*func)(struct perf_event *))
1950 {
1951         struct perf_event_context *ctx = event->ctx;
1952         struct perf_event *sibling;
1953
1954         WARN_ON_ONCE(ctx->parent_ctx);
1955         mutex_lock(&ctx->mutex);
1956         event = event->group_leader;
1957
1958         perf_event_for_each_child(event, func);
1959         func(event);
1960         list_for_each_entry(sibling, &event->sibling_list, group_entry)
1961                 perf_event_for_each_child(event, func);
1962         mutex_unlock(&ctx->mutex);
1963 }
1964
1965 static int perf_event_period(struct perf_event *event, u64 __user *arg)
1966 {
1967         struct perf_event_context *ctx = event->ctx;
1968         unsigned long size;
1969         int ret = 0;
1970         u64 value;
1971
1972         if (!event->attr.sample_period)
1973                 return -EINVAL;
1974
1975         size = copy_from_user(&value, arg, sizeof(value));
1976         if (size != sizeof(value))
1977                 return -EFAULT;
1978
1979         if (!value)
1980                 return -EINVAL;
1981
1982         spin_lock_irq(&ctx->lock);
1983         if (event->attr.freq) {
1984                 if (value > sysctl_perf_event_sample_rate) {
1985                         ret = -EINVAL;
1986                         goto unlock;
1987                 }
1988
1989                 event->attr.sample_freq = value;
1990         } else {
1991                 event->attr.sample_period = value;
1992                 event->hw.sample_period = value;
1993         }
1994 unlock:
1995         spin_unlock_irq(&ctx->lock);
1996
1997         return ret;
1998 }
1999
2000 static int perf_event_set_output(struct perf_event *event, int output_fd);
2001 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2002
2003 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2004 {
2005         struct perf_event *event = file->private_data;
2006         void (*func)(struct perf_event *);
2007         u32 flags = arg;
2008
2009         switch (cmd) {
2010         case PERF_EVENT_IOC_ENABLE:
2011                 func = perf_event_enable;
2012                 break;
2013         case PERF_EVENT_IOC_DISABLE:
2014                 func = perf_event_disable;
2015                 break;
2016         case PERF_EVENT_IOC_RESET:
2017                 func = perf_event_reset;
2018                 break;
2019
2020         case PERF_EVENT_IOC_REFRESH:
2021                 return perf_event_refresh(event, arg);
2022
2023         case PERF_EVENT_IOC_PERIOD:
2024                 return perf_event_period(event, (u64 __user *)arg);
2025
2026         case PERF_EVENT_IOC_SET_OUTPUT:
2027                 return perf_event_set_output(event, arg);
2028
2029         case PERF_EVENT_IOC_SET_FILTER:
2030                 return perf_event_set_filter(event, (void __user *)arg);
2031
2032         default:
2033                 return -ENOTTY;
2034         }
2035
2036         if (flags & PERF_IOC_FLAG_GROUP)
2037                 perf_event_for_each(event, func);
2038         else
2039                 perf_event_for_each_child(event, func);
2040
2041         return 0;
2042 }
2043
2044 int perf_event_task_enable(void)
2045 {
2046         struct perf_event *event;
2047
2048         mutex_lock(&current->perf_event_mutex);
2049         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2050                 perf_event_for_each_child(event, perf_event_enable);
2051         mutex_unlock(&current->perf_event_mutex);
2052
2053         return 0;
2054 }
2055
2056 int perf_event_task_disable(void)
2057 {
2058         struct perf_event *event;
2059
2060         mutex_lock(&current->perf_event_mutex);
2061         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2062                 perf_event_for_each_child(event, perf_event_disable);
2063         mutex_unlock(&current->perf_event_mutex);
2064
2065         return 0;
2066 }
2067
2068 #ifndef PERF_EVENT_INDEX_OFFSET
2069 # define PERF_EVENT_INDEX_OFFSET 0
2070 #endif
2071
2072 static int perf_event_index(struct perf_event *event)
2073 {
2074         if (event->state != PERF_EVENT_STATE_ACTIVE)
2075                 return 0;
2076
2077         return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2078 }
2079
2080 /*
2081  * Callers need to ensure there can be no nesting of this function, otherwise
2082  * the seqlock logic goes bad. We can not serialize this because the arch
2083  * code calls this from NMI context.
2084  */
2085 void perf_event_update_userpage(struct perf_event *event)
2086 {
2087         struct perf_event_mmap_page *userpg;
2088         struct perf_mmap_data *data;
2089
2090         rcu_read_lock();
2091         data = rcu_dereference(event->data);
2092         if (!data)
2093                 goto unlock;
2094
2095         userpg = data->user_page;
2096
2097         /*
2098          * Disable preemption so as to not let the corresponding user-space
2099          * spin too long if we get preempted.
2100          */
2101         preempt_disable();
2102         ++userpg->lock;
2103         barrier();
2104         userpg->index = perf_event_index(event);
2105         userpg->offset = atomic64_read(&event->count);
2106         if (event->state == PERF_EVENT_STATE_ACTIVE)
2107                 userpg->offset -= atomic64_read(&event->hw.prev_count);
2108
2109         userpg->time_enabled = event->total_time_enabled +
2110                         atomic64_read(&event->child_total_time_enabled);
2111
2112         userpg->time_running = event->total_time_running +
2113                         atomic64_read(&event->child_total_time_running);
2114
2115         barrier();
2116         ++userpg->lock;
2117         preempt_enable();
2118 unlock:
2119         rcu_read_unlock();
2120 }
2121
2122 static unsigned long perf_data_size(struct perf_mmap_data *data)
2123 {
2124         return data->nr_pages << (PAGE_SHIFT + data->data_order);
2125 }
2126
2127 #ifndef CONFIG_PERF_USE_VMALLOC
2128
2129 /*
2130  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2131  */
2132
2133 static struct page *
2134 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2135 {
2136         if (pgoff > data->nr_pages)
2137                 return NULL;
2138
2139         if (pgoff == 0)
2140                 return virt_to_page(data->user_page);
2141
2142         return virt_to_page(data->data_pages[pgoff - 1]);
2143 }
2144
2145 static struct perf_mmap_data *
2146 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2147 {
2148         struct perf_mmap_data *data;
2149         unsigned long size;
2150         int i;
2151
2152         WARN_ON(atomic_read(&event->mmap_count));
2153
2154         size = sizeof(struct perf_mmap_data);
2155         size += nr_pages * sizeof(void *);
2156
2157         data = kzalloc(size, GFP_KERNEL);
2158         if (!data)
2159                 goto fail;
2160
2161         data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2162         if (!data->user_page)
2163                 goto fail_user_page;
2164
2165         for (i = 0; i < nr_pages; i++) {
2166                 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2167                 if (!data->data_pages[i])
2168                         goto fail_data_pages;
2169         }
2170
2171         data->data_order = 0;
2172         data->nr_pages = nr_pages;
2173
2174         return data;
2175
2176 fail_data_pages:
2177         for (i--; i >= 0; i--)
2178                 free_page((unsigned long)data->data_pages[i]);
2179
2180         free_page((unsigned long)data->user_page);
2181
2182 fail_user_page:
2183         kfree(data);
2184
2185 fail:
2186         return NULL;
2187 }
2188
2189 static void perf_mmap_free_page(unsigned long addr)
2190 {
2191         struct page *page = virt_to_page((void *)addr);
2192
2193         page->mapping = NULL;
2194         __free_page(page);
2195 }
2196
2197 static void perf_mmap_data_free(struct perf_mmap_data *data)
2198 {
2199         int i;
2200
2201         perf_mmap_free_page((unsigned long)data->user_page);
2202         for (i = 0; i < data->nr_pages; i++)
2203                 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2204 }
2205
2206 #else
2207
2208 /*
2209  * Back perf_mmap() with vmalloc memory.
2210  *
2211  * Required for architectures that have d-cache aliasing issues.
2212  */
2213
2214 static struct page *
2215 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2216 {
2217         if (pgoff > (1UL << data->data_order))
2218                 return NULL;
2219
2220         return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2221 }
2222
2223 static void perf_mmap_unmark_page(void *addr)
2224 {
2225         struct page *page = vmalloc_to_page(addr);
2226
2227         page->mapping = NULL;
2228 }
2229
2230 static void perf_mmap_data_free_work(struct work_struct *work)
2231 {
2232         struct perf_mmap_data *data;
2233         void *base;
2234         int i, nr;
2235
2236         data = container_of(work, struct perf_mmap_data, work);
2237         nr = 1 << data->data_order;
2238
2239         base = data->user_page;
2240         for (i = 0; i < nr + 1; i++)
2241                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2242
2243         vfree(base);
2244 }
2245
2246 static void perf_mmap_data_free(struct perf_mmap_data *data)
2247 {
2248         schedule_work(&data->work);
2249 }
2250
2251 static struct perf_mmap_data *
2252 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2253 {
2254         struct perf_mmap_data *data;
2255         unsigned long size;
2256         void *all_buf;
2257
2258         WARN_ON(atomic_read(&event->mmap_count));
2259
2260         size = sizeof(struct perf_mmap_data);
2261         size += sizeof(void *);
2262
2263         data = kzalloc(size, GFP_KERNEL);
2264         if (!data)
2265                 goto fail;
2266
2267         INIT_WORK(&data->work, perf_mmap_data_free_work);
2268
2269         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2270         if (!all_buf)
2271                 goto fail_all_buf;
2272
2273         data->user_page = all_buf;
2274         data->data_pages[0] = all_buf + PAGE_SIZE;
2275         data->data_order = ilog2(nr_pages);
2276         data->nr_pages = 1;
2277
2278         return data;
2279
2280 fail_all_buf:
2281         kfree(data);
2282
2283 fail:
2284         return NULL;
2285 }
2286
2287 #endif
2288
2289 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2290 {
2291         struct perf_event *event = vma->vm_file->private_data;
2292         struct perf_mmap_data *data;
2293         int ret = VM_FAULT_SIGBUS;
2294
2295         if (vmf->flags & FAULT_FLAG_MKWRITE) {
2296                 if (vmf->pgoff == 0)
2297                         ret = 0;
2298                 return ret;
2299         }
2300
2301         rcu_read_lock();
2302         data = rcu_dereference(event->data);
2303         if (!data)
2304                 goto unlock;
2305
2306         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2307                 goto unlock;
2308
2309         vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2310         if (!vmf->page)
2311                 goto unlock;
2312
2313         get_page(vmf->page);
2314         vmf->page->mapping = vma->vm_file->f_mapping;
2315         vmf->page->index   = vmf->pgoff;
2316
2317         ret = 0;
2318 unlock:
2319         rcu_read_unlock();
2320
2321         return ret;
2322 }
2323
2324 static void
2325 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2326 {
2327         long max_size = perf_data_size(data);
2328
2329         atomic_set(&data->lock, -1);
2330
2331         if (event->attr.watermark) {
2332                 data->watermark = min_t(long, max_size,
2333                                         event->attr.wakeup_watermark);
2334         }
2335
2336         if (!data->watermark)
2337                 data->watermark = max_size / 2;
2338
2339
2340         rcu_assign_pointer(event->data, data);
2341 }
2342
2343 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2344 {
2345         struct perf_mmap_data *data;
2346
2347         data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2348         perf_mmap_data_free(data);
2349         kfree(data);
2350 }
2351
2352 static void perf_mmap_data_release(struct perf_event *event)
2353 {
2354         struct perf_mmap_data *data = event->data;
2355
2356         WARN_ON(atomic_read(&event->mmap_count));
2357
2358         rcu_assign_pointer(event->data, NULL);
2359         call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2360 }
2361
2362 static void perf_mmap_open(struct vm_area_struct *vma)
2363 {
2364         struct perf_event *event = vma->vm_file->private_data;
2365
2366         atomic_inc(&event->mmap_count);
2367 }
2368
2369 static void perf_mmap_close(struct vm_area_struct *vma)
2370 {
2371         struct perf_event *event = vma->vm_file->private_data;
2372
2373         WARN_ON_ONCE(event->ctx->parent_ctx);
2374         if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2375                 unsigned long size = perf_data_size(event->data);
2376                 struct user_struct *user = current_user();
2377
2378                 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2379                 vma->vm_mm->locked_vm -= event->data->nr_locked;
2380                 perf_mmap_data_release(event);
2381                 mutex_unlock(&event->mmap_mutex);
2382         }
2383 }
2384
2385 static const struct vm_operations_struct perf_mmap_vmops = {
2386         .open           = perf_mmap_open,
2387         .close          = perf_mmap_close,
2388         .fault          = perf_mmap_fault,
2389         .page_mkwrite   = perf_mmap_fault,
2390 };
2391
2392 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2393 {
2394         struct perf_event *event = file->private_data;
2395         unsigned long user_locked, user_lock_limit;
2396         struct user_struct *user = current_user();
2397         unsigned long locked, lock_limit;
2398         struct perf_mmap_data *data;
2399         unsigned long vma_size;
2400         unsigned long nr_pages;
2401         long user_extra, extra;
2402         int ret = 0;
2403
2404         if (!(vma->vm_flags & VM_SHARED))
2405                 return -EINVAL;
2406
2407         vma_size = vma->vm_end - vma->vm_start;
2408         nr_pages = (vma_size / PAGE_SIZE) - 1;
2409
2410         /*
2411          * If we have data pages ensure they're a power-of-two number, so we
2412          * can do bitmasks instead of modulo.
2413          */
2414         if (nr_pages != 0 && !is_power_of_2(nr_pages))
2415                 return -EINVAL;
2416
2417         if (vma_size != PAGE_SIZE * (1 + nr_pages))
2418                 return -EINVAL;
2419
2420         if (vma->vm_pgoff != 0)
2421                 return -EINVAL;
2422
2423         WARN_ON_ONCE(event->ctx->parent_ctx);
2424         mutex_lock(&event->mmap_mutex);
2425         if (event->output) {
2426                 ret = -EINVAL;
2427                 goto unlock;
2428         }
2429
2430         if (atomic_inc_not_zero(&event->mmap_count)) {
2431                 if (nr_pages != event->data->nr_pages)
2432                         ret = -EINVAL;
2433                 goto unlock;
2434         }
2435
2436         user_extra = nr_pages + 1;
2437         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2438
2439         /*
2440          * Increase the limit linearly with more CPUs:
2441          */
2442         user_lock_limit *= num_online_cpus();
2443
2444         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2445
2446         extra = 0;
2447         if (user_locked > user_lock_limit)
2448                 extra = user_locked - user_lock_limit;
2449
2450         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2451         lock_limit >>= PAGE_SHIFT;
2452         locked = vma->vm_mm->locked_vm + extra;
2453
2454         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2455                 !capable(CAP_IPC_LOCK)) {
2456                 ret = -EPERM;
2457                 goto unlock;
2458         }
2459
2460         WARN_ON(event->data);
2461
2462         data = perf_mmap_data_alloc(event, nr_pages);
2463         ret = -ENOMEM;
2464         if (!data)
2465                 goto unlock;
2466
2467         ret = 0;
2468         perf_mmap_data_init(event, data);
2469
2470         atomic_set(&event->mmap_count, 1);
2471         atomic_long_add(user_extra, &user->locked_vm);
2472         vma->vm_mm->locked_vm += extra;
2473         event->data->nr_locked = extra;
2474         if (vma->vm_flags & VM_WRITE)
2475                 event->data->writable = 1;
2476
2477 unlock:
2478         mutex_unlock(&event->mmap_mutex);
2479
2480         vma->vm_flags |= VM_RESERVED;
2481         vma->vm_ops = &perf_mmap_vmops;
2482
2483         return ret;
2484 }
2485
2486 static int perf_fasync(int fd, struct file *filp, int on)
2487 {
2488         struct inode *inode = filp->f_path.dentry->d_inode;
2489         struct perf_event *event = filp->private_data;
2490         int retval;
2491
2492         mutex_lock(&inode->i_mutex);
2493         retval = fasync_helper(fd, filp, on, &event->fasync);
2494         mutex_unlock(&inode->i_mutex);
2495
2496         if (retval < 0)
2497                 return retval;
2498
2499         return 0;
2500 }
2501
2502 static const struct file_operations perf_fops = {
2503         .release                = perf_release,
2504         .read                   = perf_read,
2505         .poll                   = perf_poll,
2506         .unlocked_ioctl         = perf_ioctl,
2507         .compat_ioctl           = perf_ioctl,
2508         .mmap                   = perf_mmap,
2509         .fasync                 = perf_fasync,
2510 };
2511
2512 /*
2513  * Perf event wakeup
2514  *
2515  * If there's data, ensure we set the poll() state and publish everything
2516  * to user-space before waking everybody up.
2517  */
2518
2519 void perf_event_wakeup(struct perf_event *event)
2520 {
2521         wake_up_all(&event->waitq);
2522
2523         if (event->pending_kill) {
2524                 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2525                 event->pending_kill = 0;
2526         }
2527 }
2528
2529 /*
2530  * Pending wakeups
2531  *
2532  * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2533  *
2534  * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2535  * single linked list and use cmpxchg() to add entries lockless.
2536  */
2537
2538 static void perf_pending_event(struct perf_pending_entry *entry)
2539 {
2540         struct perf_event *event = container_of(entry,
2541                         struct perf_event, pending);
2542
2543         if (event->pending_disable) {
2544                 event->pending_disable = 0;
2545                 __perf_event_disable(event);
2546         }
2547
2548         if (event->pending_wakeup) {
2549                 event->pending_wakeup = 0;
2550                 perf_event_wakeup(event);
2551         }
2552 }
2553
2554 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2555
2556 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2557         PENDING_TAIL,
2558 };
2559
2560 static void perf_pending_queue(struct perf_pending_entry *entry,
2561                                void (*func)(struct perf_pending_entry *))
2562 {
2563         struct perf_pending_entry **head;
2564
2565         if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2566                 return;
2567
2568         entry->func = func;
2569
2570         head = &get_cpu_var(perf_pending_head);
2571
2572         do {
2573                 entry->next = *head;
2574         } while (cmpxchg(head, entry->next, entry) != entry->next);
2575
2576         set_perf_event_pending();
2577
2578         put_cpu_var(perf_pending_head);
2579 }
2580
2581 static int __perf_pending_run(void)
2582 {
2583         struct perf_pending_entry *list;
2584         int nr = 0;
2585
2586         list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2587         while (list != PENDING_TAIL) {
2588                 void (*func)(struct perf_pending_entry *);
2589                 struct perf_pending_entry *entry = list;
2590
2591                 list = list->next;
2592
2593                 func = entry->func;
2594                 entry->next = NULL;
2595                 /*
2596                  * Ensure we observe the unqueue before we issue the wakeup,
2597                  * so that we won't be waiting forever.
2598                  * -- see perf_not_pending().
2599                  */
2600                 smp_wmb();
2601
2602                 func(entry);
2603                 nr++;
2604         }
2605
2606         return nr;
2607 }
2608
2609 static inline int perf_not_pending(struct perf_event *event)
2610 {
2611         /*
2612          * If we flush on whatever cpu we run, there is a chance we don't
2613          * need to wait.
2614          */
2615         get_cpu();
2616         __perf_pending_run();
2617         put_cpu();
2618
2619         /*
2620          * Ensure we see the proper queue state before going to sleep
2621          * so that we do not miss the wakeup. -- see perf_pending_handle()
2622          */
2623         smp_rmb();
2624         return event->pending.next == NULL;
2625 }
2626
2627 static void perf_pending_sync(struct perf_event *event)
2628 {
2629         wait_event(event->waitq, perf_not_pending(event));
2630 }
2631
2632 void perf_event_do_pending(void)
2633 {
2634         __perf_pending_run();
2635 }
2636
2637 /*
2638  * Callchain support -- arch specific
2639  */
2640
2641 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2642 {
2643         return NULL;
2644 }
2645
2646 /*
2647  * Output
2648  */
2649 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2650                               unsigned long offset, unsigned long head)
2651 {
2652         unsigned long mask;
2653
2654         if (!data->writable)
2655                 return true;
2656
2657         mask = perf_data_size(data) - 1;
2658
2659         offset = (offset - tail) & mask;
2660         head   = (head   - tail) & mask;
2661
2662         if ((int)(head - offset) < 0)
2663                 return false;
2664
2665         return true;
2666 }
2667
2668 static void perf_output_wakeup(struct perf_output_handle *handle)
2669 {
2670         atomic_set(&handle->data->poll, POLL_IN);
2671
2672         if (handle->nmi) {
2673                 handle->event->pending_wakeup = 1;
2674                 perf_pending_queue(&handle->event->pending,
2675                                    perf_pending_event);
2676         } else
2677                 perf_event_wakeup(handle->event);
2678 }
2679
2680 /*
2681  * Curious locking construct.
2682  *
2683  * We need to ensure a later event_id doesn't publish a head when a former
2684  * event_id isn't done writing. However since we need to deal with NMIs we
2685  * cannot fully serialize things.
2686  *
2687  * What we do is serialize between CPUs so we only have to deal with NMI
2688  * nesting on a single CPU.
2689  *
2690  * We only publish the head (and generate a wakeup) when the outer-most
2691  * event_id completes.
2692  */
2693 static void perf_output_lock(struct perf_output_handle *handle)
2694 {
2695         struct perf_mmap_data *data = handle->data;
2696         int cur, cpu = get_cpu();
2697
2698         handle->locked = 0;
2699
2700         for (;;) {
2701                 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2702                 if (cur == -1) {
2703                         handle->locked = 1;
2704                         break;
2705                 }
2706                 if (cur == cpu)
2707                         break;
2708
2709                 cpu_relax();
2710         }
2711 }
2712
2713 static void perf_output_unlock(struct perf_output_handle *handle)
2714 {
2715         struct perf_mmap_data *data = handle->data;
2716         unsigned long head;
2717         int cpu;
2718
2719         data->done_head = data->head;
2720
2721         if (!handle->locked)
2722                 goto out;
2723
2724 again:
2725         /*
2726          * The xchg implies a full barrier that ensures all writes are done
2727          * before we publish the new head, matched by a rmb() in userspace when
2728          * reading this position.
2729          */
2730         while ((head = atomic_long_xchg(&data->done_head, 0)))
2731                 data->user_page->data_head = head;
2732
2733         /*
2734          * NMI can happen here, which means we can miss a done_head update.
2735          */
2736
2737         cpu = atomic_xchg(&data->lock, -1);
2738         WARN_ON_ONCE(cpu != smp_processor_id());
2739
2740         /*
2741          * Therefore we have to validate we did not indeed do so.
2742          */
2743         if (unlikely(atomic_long_read(&data->done_head))) {
2744                 /*
2745                  * Since we had it locked, we can lock it again.
2746                  */
2747                 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2748                         cpu_relax();
2749
2750                 goto again;
2751         }
2752
2753         if (atomic_xchg(&data->wakeup, 0))
2754                 perf_output_wakeup(handle);
2755 out:
2756         put_cpu();
2757 }
2758
2759 void perf_output_copy(struct perf_output_handle *handle,
2760                       const void *buf, unsigned int len)
2761 {
2762         unsigned int pages_mask;
2763         unsigned long offset;
2764         unsigned int size;
2765         void **pages;
2766
2767         offset          = handle->offset;
2768         pages_mask      = handle->data->nr_pages - 1;
2769         pages           = handle->data->data_pages;
2770
2771         do {
2772                 unsigned long page_offset;
2773                 unsigned long page_size;
2774                 int nr;
2775
2776                 nr          = (offset >> PAGE_SHIFT) & pages_mask;
2777                 page_size   = 1UL << (handle->data->data_order + PAGE_SHIFT);
2778                 page_offset = offset & (page_size - 1);
2779                 size        = min_t(unsigned int, page_size - page_offset, len);
2780
2781                 memcpy(pages[nr] + page_offset, buf, size);
2782
2783                 len         -= size;
2784                 buf         += size;
2785                 offset      += size;
2786         } while (len);
2787
2788         handle->offset = offset;
2789
2790         /*
2791          * Check we didn't copy past our reservation window, taking the
2792          * possible unsigned int wrap into account.
2793          */
2794         WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2795 }
2796
2797 int perf_output_begin(struct perf_output_handle *handle,
2798                       struct perf_event *event, unsigned int size,
2799                       int nmi, int sample)
2800 {
2801         struct perf_event *output_event;
2802         struct perf_mmap_data *data;
2803         unsigned long tail, offset, head;
2804         int have_lost;
2805         struct {
2806                 struct perf_event_header header;
2807                 u64                      id;
2808                 u64                      lost;
2809         } lost_event;
2810
2811         rcu_read_lock();
2812         /*
2813          * For inherited events we send all the output towards the parent.
2814          */
2815         if (event->parent)
2816                 event = event->parent;
2817
2818         output_event = rcu_dereference(event->output);
2819         if (output_event)
2820                 event = output_event;
2821
2822         data = rcu_dereference(event->data);
2823         if (!data)
2824                 goto out;
2825
2826         handle->data    = data;
2827         handle->event   = event;
2828         handle->nmi     = nmi;
2829         handle->sample  = sample;
2830
2831         if (!data->nr_pages)
2832                 goto fail;
2833
2834         have_lost = atomic_read(&data->lost);
2835         if (have_lost)
2836                 size += sizeof(lost_event);
2837
2838         perf_output_lock(handle);
2839
2840         do {
2841                 /*
2842                  * Userspace could choose to issue a mb() before updating the
2843                  * tail pointer. So that all reads will be completed before the
2844                  * write is issued.
2845                  */
2846                 tail = ACCESS_ONCE(data->user_page->data_tail);
2847                 smp_rmb();
2848                 offset = head = atomic_long_read(&data->head);
2849                 head += size;
2850                 if (unlikely(!perf_output_space(data, tail, offset, head)))
2851                         goto fail;
2852         } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2853
2854         handle->offset  = offset;
2855         handle->head    = head;
2856
2857         if (head - tail > data->watermark)
2858                 atomic_set(&data->wakeup, 1);
2859
2860         if (have_lost) {
2861                 lost_event.header.type = PERF_RECORD_LOST;
2862                 lost_event.header.misc = 0;
2863                 lost_event.header.size = sizeof(lost_event);
2864                 lost_event.id          = event->id;
2865                 lost_event.lost        = atomic_xchg(&data->lost, 0);
2866
2867                 perf_output_put(handle, lost_event);
2868         }
2869
2870         return 0;
2871
2872 fail:
2873         atomic_inc(&data->lost);
2874         perf_output_unlock(handle);
2875 out:
2876         rcu_read_unlock();
2877
2878         return -ENOSPC;
2879 }
2880
2881 void perf_output_end(struct perf_output_handle *handle)
2882 {
2883         struct perf_event *event = handle->event;
2884         struct perf_mmap_data *data = handle->data;
2885
2886         int wakeup_events = event->attr.wakeup_events;
2887
2888         if (handle->sample && wakeup_events) {
2889                 int events = atomic_inc_return(&data->events);
2890                 if (events >= wakeup_events) {
2891                         atomic_sub(wakeup_events, &data->events);
2892                         atomic_set(&data->wakeup, 1);
2893                 }
2894         }
2895
2896         perf_output_unlock(handle);
2897         rcu_read_unlock();
2898 }
2899
2900 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
2901 {
2902         /*
2903          * only top level events have the pid namespace they were created in
2904          */
2905         if (event->parent)
2906                 event = event->parent;
2907
2908         return task_tgid_nr_ns(p, event->ns);
2909 }
2910
2911 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
2912 {
2913         /*
2914          * only top level events have the pid namespace they were created in
2915          */
2916         if (event->parent)
2917                 event = event->parent;
2918
2919         return task_pid_nr_ns(p, event->ns);
2920 }
2921
2922 static void perf_output_read_one(struct perf_output_handle *handle,
2923                                  struct perf_event *event)
2924 {
2925         u64 read_format = event->attr.read_format;
2926         u64 values[4];
2927         int n = 0;
2928
2929         values[n++] = atomic64_read(&event->count);
2930         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2931                 values[n++] = event->total_time_enabled +
2932                         atomic64_read(&event->child_total_time_enabled);
2933         }
2934         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2935                 values[n++] = event->total_time_running +
2936                         atomic64_read(&event->child_total_time_running);
2937         }
2938         if (read_format & PERF_FORMAT_ID)
2939                 values[n++] = primary_event_id(event);
2940
2941         perf_output_copy(handle, values, n * sizeof(u64));
2942 }
2943
2944 /*
2945  * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
2946  */
2947 static void perf_output_read_group(struct perf_output_handle *handle,
2948                             struct perf_event *event)
2949 {
2950         struct perf_event *leader = event->group_leader, *sub;
2951         u64 read_format = event->attr.read_format;
2952         u64 values[5];
2953         int n = 0;
2954
2955         values[n++] = 1 + leader->nr_siblings;
2956
2957         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2958                 values[n++] = leader->total_time_enabled;
2959
2960         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2961                 values[n++] = leader->total_time_running;
2962
2963         if (leader != event)
2964                 leader->pmu->read(leader);
2965
2966         values[n++] = atomic64_read(&leader->count);
2967         if (read_format & PERF_FORMAT_ID)
2968                 values[n++] = primary_event_id(leader);
2969
2970         perf_output_copy(handle, values, n * sizeof(u64));
2971
2972         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2973                 n = 0;
2974
2975                 if (sub != event)
2976                         sub->pmu->read(sub);
2977
2978                 values[n++] = atomic64_read(&sub->count);
2979                 if (read_format & PERF_FORMAT_ID)
2980                         values[n++] = primary_event_id(sub);
2981
2982                 perf_output_copy(handle, values, n * sizeof(u64));
2983         }
2984 }
2985
2986 static void perf_output_read(struct perf_output_handle *handle,
2987                              struct perf_event *event)
2988 {
2989         if (event->attr.read_format & PERF_FORMAT_GROUP)
2990                 perf_output_read_group(handle, event);
2991         else
2992                 perf_output_read_one(handle, event);
2993 }
2994
2995 void perf_output_sample(struct perf_output_handle *handle,
2996                         struct perf_event_header *header,
2997                         struct perf_sample_data *data,
2998                         struct perf_event *event)
2999 {
3000         u64 sample_type = data->type;
3001
3002         perf_output_put(handle, *header);
3003
3004         if (sample_type & PERF_SAMPLE_IP)
3005                 perf_output_put(handle, data->ip);
3006
3007         if (sample_type & PERF_SAMPLE_TID)
3008                 perf_output_put(handle, data->tid_entry);
3009
3010         if (sample_type & PERF_SAMPLE_TIME)
3011                 perf_output_put(handle, data->time);
3012
3013         if (sample_type & PERF_SAMPLE_ADDR)
3014                 perf_output_put(handle, data->addr);
3015
3016         if (sample_type & PERF_SAMPLE_ID)
3017                 perf_output_put(handle, data->id);
3018
3019         if (sample_type & PERF_SAMPLE_STREAM_ID)
3020                 perf_output_put(handle, data->stream_id);
3021
3022         if (sample_type & PERF_SAMPLE_CPU)
3023                 perf_output_put(handle, data->cpu_entry);
3024
3025         if (sample_type & PERF_SAMPLE_PERIOD)
3026                 perf_output_put(handle, data->period);
3027
3028         if (sample_type & PERF_SAMPLE_READ)
3029                 perf_output_read(handle, event);
3030
3031         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3032                 if (data->callchain) {
3033                         int size = 1;
3034
3035                         if (data->callchain)
3036                                 size += data->callchain->nr;
3037
3038                         size *= sizeof(u64);
3039
3040                         perf_output_copy(handle, data->callchain, size);
3041                 } else {
3042                         u64 nr = 0;
3043                         perf_output_put(handle, nr);
3044                 }
3045         }
3046
3047         if (sample_type & PERF_SAMPLE_RAW) {
3048                 if (data->raw) {
3049                         perf_output_put(handle, data->raw->size);
3050                         perf_output_copy(handle, data->raw->data,
3051                                          data->raw->size);
3052                 } else {
3053                         struct {
3054                                 u32     size;
3055                                 u32     data;
3056                         } raw = {
3057                                 .size = sizeof(u32),
3058                                 .data = 0,
3059                         };
3060                         perf_output_put(handle, raw);
3061                 }
3062         }
3063 }
3064
3065 void perf_prepare_sample(struct perf_event_header *header,
3066                          struct perf_sample_data *data,
3067                          struct perf_event *event,
3068                          struct pt_regs *regs)
3069 {
3070         u64 sample_type = event->attr.sample_type;
3071
3072         data->type = sample_type;
3073
3074         header->type = PERF_RECORD_SAMPLE;
3075         header->size = sizeof(*header);
3076
3077         header->misc = 0;
3078         header->misc |= perf_misc_flags(regs);
3079
3080         if (sample_type & PERF_SAMPLE_IP) {
3081                 data->ip = perf_instruction_pointer(regs);
3082
3083                 header->size += sizeof(data->ip);
3084         }
3085
3086         if (sample_type & PERF_SAMPLE_TID) {
3087                 /* namespace issues */
3088                 data->tid_entry.pid = perf_event_pid(event, current);
3089                 data->tid_entry.tid = perf_event_tid(event, current);
3090
3091                 header->size += sizeof(data->tid_entry);
3092         }
3093
3094         if (sample_type & PERF_SAMPLE_TIME) {
3095                 data->time = perf_clock();
3096
3097                 header->size += sizeof(data->time);
3098         }
3099
3100         if (sample_type & PERF_SAMPLE_ADDR)
3101                 header->size += sizeof(data->addr);
3102
3103         if (sample_type & PERF_SAMPLE_ID) {
3104                 data->id = primary_event_id(event);
3105
3106                 header->size += sizeof(data->id);
3107         }
3108
3109         if (sample_type & PERF_SAMPLE_STREAM_ID) {
3110                 data->stream_id = event->id;
3111
3112                 header->size += sizeof(data->stream_id);
3113         }
3114
3115         if (sample_type & PERF_SAMPLE_CPU) {
3116                 data->cpu_entry.cpu             = raw_smp_processor_id();
3117                 data->cpu_entry.reserved        = 0;
3118
3119                 header->size += sizeof(data->cpu_entry);
3120         }
3121
3122         if (sample_type & PERF_SAMPLE_PERIOD)
3123                 header->size += sizeof(data->period);
3124
3125         if (sample_type & PERF_SAMPLE_READ)
3126                 header->size += perf_event_read_size(event);
3127
3128         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3129                 int size = 1;
3130
3131                 data->callchain = perf_callchain(regs);
3132
3133                 if (data->callchain)
3134                         size += data->callchain->nr;
3135
3136                 header->size += size * sizeof(u64);
3137         }
3138
3139         if (sample_type & PERF_SAMPLE_RAW) {
3140                 int size = sizeof(u32);
3141
3142                 if (data->raw)
3143                         size += data->raw->size;
3144                 else
3145                         size += sizeof(u32);
3146
3147                 WARN_ON_ONCE(size & (sizeof(u64)-1));
3148                 header->size += size;
3149         }
3150 }
3151
3152 static void perf_event_output(struct perf_event *event, int nmi,
3153                                 struct perf_sample_data *data,
3154                                 struct pt_regs *regs)
3155 {
3156         struct perf_output_handle handle;
3157         struct perf_event_header header;
3158
3159         perf_prepare_sample(&header, data, event, regs);
3160
3161         if (perf_output_begin(&handle, event, header.size, nmi, 1))
3162                 return;
3163
3164         perf_output_sample(&handle, &header, data, event);
3165
3166         perf_output_end(&handle);
3167 }
3168
3169 /*
3170  * read event_id
3171  */
3172
3173 struct perf_read_event {
3174         struct perf_event_header        header;
3175
3176         u32                             pid;
3177         u32                             tid;
3178 };
3179
3180 static void
3181 perf_event_read_event(struct perf_event *event,
3182                         struct task_struct *task)
3183 {
3184         struct perf_output_handle handle;
3185         struct perf_read_event read_event = {
3186                 .header = {
3187                         .type = PERF_RECORD_READ,
3188                         .misc = 0,
3189                         .size = sizeof(read_event) + perf_event_read_size(event),
3190                 },
3191                 .pid = perf_event_pid(event, task),
3192                 .tid = perf_event_tid(event, task),
3193         };
3194         int ret;
3195
3196         ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3197         if (ret)
3198                 return;
3199
3200         perf_output_put(&handle, read_event);
3201         perf_output_read(&handle, event);
3202
3203         perf_output_end(&handle);
3204 }
3205
3206 /*
3207  * task tracking -- fork/exit
3208  *
3209  * enabled by: attr.comm | attr.mmap | attr.task
3210  */
3211
3212 struct perf_task_event {
3213         struct task_struct              *task;
3214         struct perf_event_context       *task_ctx;
3215
3216         struct {
3217                 struct perf_event_header        header;
3218
3219                 u32                             pid;
3220                 u32                             ppid;
3221                 u32                             tid;
3222                 u32                             ptid;
3223                 u64                             time;
3224         } event_id;
3225 };
3226
3227 static void perf_event_task_output(struct perf_event *event,
3228                                      struct perf_task_event *task_event)
3229 {
3230         struct perf_output_handle handle;
3231         int size;
3232         struct task_struct *task = task_event->task;
3233         int ret;
3234
3235         size  = task_event->event_id.header.size;
3236         ret = perf_output_begin(&handle, event, size, 0, 0);
3237
3238         if (ret)
3239                 return;
3240
3241         task_event->event_id.pid = perf_event_pid(event, task);
3242         task_event->event_id.ppid = perf_event_pid(event, current);
3243
3244         task_event->event_id.tid = perf_event_tid(event, task);
3245         task_event->event_id.ptid = perf_event_tid(event, current);
3246
3247         task_event->event_id.time = perf_clock();
3248
3249         perf_output_put(&handle, task_event->event_id);
3250
3251         perf_output_end(&handle);
3252 }
3253
3254 static int perf_event_task_match(struct perf_event *event)
3255 {
3256         if (event->attr.comm || event->attr.mmap || event->attr.task)
3257                 return 1;
3258
3259         return 0;
3260 }
3261
3262 static void perf_event_task_ctx(struct perf_event_context *ctx,
3263                                   struct perf_task_event *task_event)
3264 {
3265         struct perf_event *event;
3266
3267         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3268                 if (perf_event_task_match(event))
3269                         perf_event_task_output(event, task_event);
3270         }
3271 }
3272
3273 static void perf_event_task_event(struct perf_task_event *task_event)
3274 {
3275         struct perf_cpu_context *cpuctx;
3276         struct perf_event_context *ctx = task_event->task_ctx;
3277
3278         rcu_read_lock();
3279         cpuctx = &get_cpu_var(perf_cpu_context);
3280         perf_event_task_ctx(&cpuctx->ctx, task_event);
3281         put_cpu_var(perf_cpu_context);
3282
3283         if (!ctx)
3284                 ctx = rcu_dereference(task_event->task->perf_event_ctxp);
3285         if (ctx)
3286                 perf_event_task_ctx(ctx, task_event);
3287         rcu_read_unlock();
3288 }
3289
3290 static void perf_event_task(struct task_struct *task,
3291                               struct perf_event_context *task_ctx,
3292                               int new)
3293 {
3294         struct perf_task_event task_event;
3295
3296         if (!atomic_read(&nr_comm_events) &&
3297             !atomic_read(&nr_mmap_events) &&
3298             !atomic_read(&nr_task_events))
3299                 return;
3300
3301         task_event = (struct perf_task_event){
3302                 .task     = task,
3303                 .task_ctx = task_ctx,
3304                 .event_id    = {
3305                         .header = {
3306                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3307                                 .misc = 0,
3308                                 .size = sizeof(task_event.event_id),
3309                         },
3310                         /* .pid  */
3311                         /* .ppid */
3312                         /* .tid  */
3313                         /* .ptid */
3314                 },
3315         };
3316
3317         perf_event_task_event(&task_event);
3318 }
3319
3320 void perf_event_fork(struct task_struct *task)
3321 {
3322         perf_event_task(task, NULL, 1);
3323 }
3324
3325 /*
3326  * comm tracking
3327  */
3328
3329 struct perf_comm_event {
3330         struct task_struct      *task;
3331         char                    *comm;
3332         int                     comm_size;
3333
3334         struct {
3335                 struct perf_event_header        header;
3336
3337                 u32                             pid;
3338                 u32                             tid;
3339         } event_id;
3340 };
3341
3342 static void perf_event_comm_output(struct perf_event *event,
3343                                      struct perf_comm_event *comm_event)
3344 {
3345         struct perf_output_handle handle;
3346         int size = comm_event->event_id.header.size;
3347         int ret = perf_output_begin(&handle, event, size, 0, 0);
3348
3349         if (ret)
3350                 return;
3351
3352         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3353         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3354
3355         perf_output_put(&handle, comm_event->event_id);
3356         perf_output_copy(&handle, comm_event->comm,
3357                                    comm_event->comm_size);
3358         perf_output_end(&handle);
3359 }
3360
3361 static int perf_event_comm_match(struct perf_event *event)
3362 {
3363         if (event->attr.comm)
3364                 return 1;
3365
3366         return 0;
3367 }
3368
3369 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3370                                   struct perf_comm_event *comm_event)
3371 {
3372         struct perf_event *event;
3373
3374         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3375                 if (perf_event_comm_match(event))
3376                         perf_event_comm_output(event, comm_event);
3377         }
3378 }
3379
3380 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3381 {
3382         struct perf_cpu_context *cpuctx;
3383         struct perf_event_context *ctx;
3384         unsigned int size;
3385         char comm[TASK_COMM_LEN];
3386
3387         memset(comm, 0, sizeof(comm));
3388         strlcpy(comm, comm_event->task->comm, sizeof(comm));
3389         size = ALIGN(strlen(comm)+1, sizeof(u64));
3390
3391         comm_event->comm = comm;
3392         comm_event->comm_size = size;
3393
3394         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3395
3396         rcu_read_lock();
3397         cpuctx = &get_cpu_var(perf_cpu_context);
3398         perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3399         put_cpu_var(perf_cpu_context);
3400
3401         /*
3402          * doesn't really matter which of the child contexts the
3403          * events ends up in.
3404          */
3405         ctx = rcu_dereference(current->perf_event_ctxp);
3406         if (ctx)
3407                 perf_event_comm_ctx(ctx, comm_event);
3408         rcu_read_unlock();
3409 }
3410
3411 void perf_event_comm(struct task_struct *task)
3412 {
3413         struct perf_comm_event comm_event;
3414
3415         if (task->perf_event_ctxp)
3416                 perf_event_enable_on_exec(task);
3417
3418         if (!atomic_read(&nr_comm_events))
3419                 return;
3420
3421         comm_event = (struct perf_comm_event){
3422                 .task   = task,
3423                 /* .comm      */
3424                 /* .comm_size */
3425                 .event_id  = {
3426                         .header = {
3427                                 .type = PERF_RECORD_COMM,
3428                                 .misc = 0,
3429                                 /* .size */
3430                         },
3431                         /* .pid */
3432                         /* .tid */
3433                 },
3434         };
3435
3436         perf_event_comm_event(&comm_event);
3437 }
3438
3439 /*
3440  * mmap tracking
3441  */
3442
3443 struct perf_mmap_event {
3444         struct vm_area_struct   *vma;
3445
3446         const char              *file_name;
3447         int                     file_size;
3448
3449         struct {
3450                 struct perf_event_header        header;
3451
3452                 u32                             pid;
3453                 u32                             tid;
3454                 u64                             start;
3455                 u64                             len;
3456                 u64                             pgoff;
3457         } event_id;
3458 };
3459
3460 static void perf_event_mmap_output(struct perf_event *event,
3461                                      struct perf_mmap_event *mmap_event)
3462 {
3463         struct perf_output_handle handle;
3464         int size = mmap_event->event_id.header.size;
3465         int ret = perf_output_begin(&handle, event, size, 0, 0);
3466
3467         if (ret)
3468                 return;
3469
3470         mmap_event->event_id.pid = perf_event_pid(event, current);
3471         mmap_event->event_id.tid = perf_event_tid(event, current);
3472
3473         perf_output_put(&handle, mmap_event->event_id);
3474         perf_output_copy(&handle, mmap_event->file_name,
3475                                    mmap_event->file_size);
3476         perf_output_end(&handle);
3477 }
3478
3479 static int perf_event_mmap_match(struct perf_event *event,
3480                                    struct perf_mmap_event *mmap_event)
3481 {
3482         if (event->attr.mmap)
3483                 return 1;
3484
3485         return 0;
3486 }
3487
3488 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3489                                   struct perf_mmap_event *mmap_event)
3490 {
3491         struct perf_event *event;
3492
3493         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3494                 if (perf_event_mmap_match(event, mmap_event))
3495                         perf_event_mmap_output(event, mmap_event);
3496         }
3497 }
3498
3499 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3500 {
3501         struct perf_cpu_context *cpuctx;
3502         struct perf_event_context *ctx;
3503         struct vm_area_struct *vma = mmap_event->vma;
3504         struct file *file = vma->vm_file;
3505         unsigned int size;
3506         char tmp[16];
3507         char *buf = NULL;
3508         const char *name;
3509
3510         memset(tmp, 0, sizeof(tmp));
3511
3512         if (file) {
3513                 /*
3514                  * d_path works from the end of the buffer backwards, so we
3515                  * need to add enough zero bytes after the string to handle
3516                  * the 64bit alignment we do later.
3517                  */
3518                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3519                 if (!buf) {
3520                         name = strncpy(tmp, "//enomem", sizeof(tmp));
3521                         goto got_name;
3522                 }
3523                 name = d_path(&file->f_path, buf, PATH_MAX);
3524                 if (IS_ERR(name)) {
3525                         name = strncpy(tmp, "//toolong", sizeof(tmp));
3526                         goto got_name;
3527                 }
3528         } else {
3529                 if (arch_vma_name(mmap_event->vma)) {
3530                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3531                                        sizeof(tmp));
3532                         goto got_name;
3533                 }
3534
3535                 if (!vma->vm_mm) {
3536                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
3537                         goto got_name;
3538                 }
3539
3540                 name = strncpy(tmp, "//anon", sizeof(tmp));
3541                 goto got_name;
3542         }
3543
3544 got_name:
3545         size = ALIGN(strlen(name)+1, sizeof(u64));
3546
3547         mmap_event->file_name = name;
3548         mmap_event->file_size = size;
3549
3550         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3551
3552         rcu_read_lock();
3553         cpuctx = &get_cpu_var(perf_cpu_context);
3554         perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3555         put_cpu_var(perf_cpu_context);
3556
3557         /*
3558          * doesn't really matter which of the child contexts the
3559          * events ends up in.
3560          */
3561         ctx = rcu_dereference(current->perf_event_ctxp);
3562         if (ctx)
3563                 perf_event_mmap_ctx(ctx, mmap_event);
3564         rcu_read_unlock();
3565
3566         kfree(buf);
3567 }
3568
3569 void __perf_event_mmap(struct vm_area_struct *vma)
3570 {
3571         struct perf_mmap_event mmap_event;
3572
3573         if (!atomic_read(&nr_mmap_events))
3574                 return;
3575
3576         mmap_event = (struct perf_mmap_event){
3577                 .vma    = vma,
3578                 /* .file_name */
3579                 /* .file_size */
3580                 .event_id  = {
3581                         .header = {
3582                                 .type = PERF_RECORD_MMAP,
3583                                 .misc = 0,
3584                                 /* .size */
3585                         },
3586                         /* .pid */
3587                         /* .tid */
3588                         .start  = vma->vm_start,
3589                         .len    = vma->vm_end - vma->vm_start,
3590                         .pgoff  = vma->vm_pgoff,
3591                 },
3592         };
3593
3594         perf_event_mmap_event(&mmap_event);
3595 }
3596
3597 /*
3598  * IRQ throttle logging
3599  */
3600
3601 static void perf_log_throttle(struct perf_event *event, int enable)
3602 {
3603         struct perf_output_handle handle;
3604         int ret;
3605
3606         struct {
3607                 struct perf_event_header        header;
3608                 u64                             time;
3609                 u64                             id;
3610                 u64                             stream_id;
3611         } throttle_event = {
3612                 .header = {
3613                         .type = PERF_RECORD_THROTTLE,
3614                         .misc = 0,
3615                         .size = sizeof(throttle_event),
3616                 },
3617                 .time           = perf_clock(),
3618                 .id             = primary_event_id(event),
3619                 .stream_id      = event->id,
3620         };
3621
3622         if (enable)
3623                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3624
3625         ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3626         if (ret)
3627                 return;
3628
3629         perf_output_put(&handle, throttle_event);
3630         perf_output_end(&handle);
3631 }
3632
3633 /*
3634  * Generic event overflow handling, sampling.
3635  */
3636
3637 static int __perf_event_overflow(struct perf_event *event, int nmi,
3638                                    int throttle, struct perf_sample_data *data,
3639                                    struct pt_regs *regs)
3640 {
3641         int events = atomic_read(&event->event_limit);
3642         struct hw_perf_event *hwc = &event->hw;
3643         int ret = 0;
3644
3645         throttle = (throttle && event->pmu->unthrottle != NULL);
3646
3647         if (!throttle) {
3648                 hwc->interrupts++;
3649         } else {
3650                 if (hwc->interrupts != MAX_INTERRUPTS) {
3651                         hwc->interrupts++;
3652                         if (HZ * hwc->interrupts >
3653                                         (u64)sysctl_perf_event_sample_rate) {
3654                                 hwc->interrupts = MAX_INTERRUPTS;
3655                                 perf_log_throttle(event, 0);
3656                                 ret = 1;
3657                         }
3658                 } else {
3659                         /*
3660                          * Keep re-disabling events even though on the previous
3661                          * pass we disabled it - just in case we raced with a
3662                          * sched-in and the event got enabled again:
3663                          */
3664                         ret = 1;
3665                 }
3666         }
3667
3668         if (event->attr.freq) {
3669                 u64 now = perf_clock();
3670                 s64 delta = now - hwc->freq_stamp;
3671
3672                 hwc->freq_stamp = now;
3673
3674                 if (delta > 0 && delta < TICK_NSEC)
3675                         perf_adjust_period(event, NSEC_PER_SEC / (int)delta);
3676         }
3677
3678         /*
3679          * XXX event_limit might not quite work as expected on inherited
3680          * events
3681          */
3682
3683         event->pending_kill = POLL_IN;
3684         if (events && atomic_dec_and_test(&event->event_limit)) {
3685                 ret = 1;
3686                 event->pending_kill = POLL_HUP;
3687                 if (nmi) {
3688                         event->pending_disable = 1;
3689                         perf_pending_queue(&event->pending,
3690                                            perf_pending_event);
3691                 } else
3692                         perf_event_disable(event);
3693         }
3694
3695         if (event->overflow_handler)
3696                 event->overflow_handler(event, nmi, data, regs);
3697         else
3698                 perf_event_output(event, nmi, data, regs);
3699
3700         return ret;
3701 }
3702
3703 int perf_event_overflow(struct perf_event *event, int nmi,
3704                           struct perf_sample_data *data,
3705                           struct pt_regs *regs)
3706 {
3707         return __perf_event_overflow(event, nmi, 1, data, regs);
3708 }
3709
3710 /*
3711  * Generic software event infrastructure
3712  */
3713
3714 /*
3715  * We directly increment event->count and keep a second value in
3716  * event->hw.period_left to count intervals. This period event
3717  * is kept in the range [-sample_period, 0] so that we can use the
3718  * sign as trigger.
3719  */
3720
3721 static u64 perf_swevent_set_period(struct perf_event *event)
3722 {
3723         struct hw_perf_event *hwc = &event->hw;
3724         u64 period = hwc->last_period;
3725         u64 nr, offset;
3726         s64 old, val;
3727
3728         hwc->last_period = hwc->sample_period;
3729
3730 again:
3731         old = val = atomic64_read(&hwc->period_left);
3732         if (val < 0)
3733                 return 0;
3734
3735         nr = div64_u64(period + val, period);
3736         offset = nr * period;
3737         val -= offset;
3738         if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3739                 goto again;
3740
3741         return nr;
3742 }
3743
3744 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3745                                     int nmi, struct perf_sample_data *data,
3746                                     struct pt_regs *regs)
3747 {
3748         struct hw_perf_event *hwc = &event->hw;
3749         int throttle = 0;
3750
3751         data->period = event->hw.last_period;
3752         if (!overflow)
3753                 overflow = perf_swevent_set_period(event);
3754
3755         if (hwc->interrupts == MAX_INTERRUPTS)
3756                 return;
3757
3758         for (; overflow; overflow--) {
3759                 if (__perf_event_overflow(event, nmi, throttle,
3760                                             data, regs)) {
3761                         /*
3762                          * We inhibit the overflow from happening when
3763                          * hwc->interrupts == MAX_INTERRUPTS.
3764                          */
3765                         break;
3766                 }
3767                 throttle = 1;
3768         }
3769 }
3770
3771 static void perf_swevent_unthrottle(struct perf_event *event)
3772 {
3773         /*
3774          * Nothing to do, we already reset hwc->interrupts.
3775          */
3776 }
3777
3778 static void perf_swevent_add(struct perf_event *event, u64 nr,
3779                                int nmi, struct perf_sample_data *data,
3780                                struct pt_regs *regs)
3781 {
3782         struct hw_perf_event *hwc = &event->hw;
3783
3784         atomic64_add(nr, &event->count);
3785
3786         if (!regs)
3787                 return;
3788
3789         if (!hwc->sample_period)
3790                 return;
3791
3792         if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3793                 return perf_swevent_overflow(event, 1, nmi, data, regs);
3794
3795         if (atomic64_add_negative(nr, &hwc->period_left))
3796                 return;
3797
3798         perf_swevent_overflow(event, 0, nmi, data, regs);
3799 }
3800
3801 static int perf_swevent_is_counting(struct perf_event *event)
3802 {
3803         /*
3804          * The event is active, we're good!
3805          */
3806         if (event->state == PERF_EVENT_STATE_ACTIVE)
3807                 return 1;
3808
3809         /*
3810          * The event is off/error, not counting.
3811          */
3812         if (event->state != PERF_EVENT_STATE_INACTIVE)
3813                 return 0;
3814
3815         /*
3816          * The event is inactive, if the context is active
3817          * we're part of a group that didn't make it on the 'pmu',
3818          * not counting.
3819          */
3820         if (event->ctx->is_active)
3821                 return 0;
3822
3823         /*
3824          * We're inactive and the context is too, this means the
3825          * task is scheduled out, we're counting events that happen
3826          * to us, like migration events.
3827          */
3828         return 1;
3829 }
3830
3831 static int perf_tp_event_match(struct perf_event *event,
3832                                 struct perf_sample_data *data);
3833
3834 static int perf_swevent_match(struct perf_event *event,
3835                                 enum perf_type_id type,
3836                                 u32 event_id,
3837                                 struct perf_sample_data *data,
3838                                 struct pt_regs *regs)
3839 {
3840         if (!perf_swevent_is_counting(event))
3841                 return 0;
3842
3843         if (event->attr.type != type)
3844                 return 0;
3845         if (event->attr.config != event_id)
3846                 return 0;
3847
3848         if (regs) {
3849                 if (event->attr.exclude_user && user_mode(regs))
3850                         return 0;
3851
3852                 if (event->attr.exclude_kernel && !user_mode(regs))
3853                         return 0;
3854         }
3855
3856         if (event->attr.type == PERF_TYPE_TRACEPOINT &&
3857             !perf_tp_event_match(event, data))
3858                 return 0;
3859
3860         return 1;
3861 }
3862
3863 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
3864                                      enum perf_type_id type,
3865                                      u32 event_id, u64 nr, int nmi,
3866                                      struct perf_sample_data *data,
3867                                      struct pt_regs *regs)
3868 {
3869         struct perf_event *event;
3870
3871         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3872                 if (perf_swevent_match(event, type, event_id, data, regs))
3873                         perf_swevent_add(event, nr, nmi, data, regs);
3874         }
3875 }
3876
3877 int perf_swevent_get_recursion_context(void)
3878 {
3879         struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3880         int rctx;
3881
3882         if (in_nmi())
3883                 rctx = 3;
3884         else if (in_irq())
3885                 rctx = 2;
3886         else if (in_softirq())
3887                 rctx = 1;
3888         else
3889                 rctx = 0;
3890
3891         if (cpuctx->recursion[rctx]) {
3892                 put_cpu_var(perf_cpu_context);
3893                 return -1;
3894         }
3895
3896         cpuctx->recursion[rctx]++;
3897         barrier();
3898
3899         return rctx;
3900 }
3901 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
3902
3903 void perf_swevent_put_recursion_context(int rctx)
3904 {
3905         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
3906         barrier();
3907         cpuctx->recursion[rctx]++;
3908         put_cpu_var(perf_cpu_context);
3909 }
3910 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
3911
3912 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
3913                                     u64 nr, int nmi,
3914                                     struct perf_sample_data *data,
3915                                     struct pt_regs *regs)
3916 {
3917         struct perf_cpu_context *cpuctx;
3918         struct perf_event_context *ctx;
3919
3920         cpuctx = &__get_cpu_var(perf_cpu_context);
3921         rcu_read_lock();
3922         perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
3923                                  nr, nmi, data, regs);
3924         /*
3925          * doesn't really matter which of the child contexts the
3926          * events ends up in.
3927          */
3928         ctx = rcu_dereference(current->perf_event_ctxp);
3929         if (ctx)
3930                 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
3931         rcu_read_unlock();
3932 }
3933
3934 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
3935                             struct pt_regs *regs, u64 addr)
3936 {
3937         struct perf_sample_data data;
3938         int rctx;
3939
3940         rctx = perf_swevent_get_recursion_context();
3941         if (rctx < 0)
3942                 return;
3943
3944         data.addr = addr;
3945         data.raw  = NULL;
3946
3947         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
3948
3949         perf_swevent_put_recursion_context(rctx);
3950 }
3951
3952 static void perf_swevent_read(struct perf_event *event)
3953 {
3954 }
3955
3956 static int perf_swevent_enable(struct perf_event *event)
3957 {
3958         struct hw_perf_event *hwc = &event->hw;
3959
3960         if (hwc->sample_period) {
3961                 hwc->last_period = hwc->sample_period;
3962                 perf_swevent_set_period(event);
3963         }
3964         return 0;
3965 }
3966
3967 static void perf_swevent_disable(struct perf_event *event)
3968 {
3969 }
3970
3971 static const struct pmu perf_ops_generic = {
3972         .enable         = perf_swevent_enable,
3973         .disable        = perf_swevent_disable,
3974         .read           = perf_swevent_read,
3975         .unthrottle     = perf_swevent_unthrottle,
3976 };
3977
3978 /*
3979  * hrtimer based swevent callback
3980  */
3981
3982 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
3983 {
3984         enum hrtimer_restart ret = HRTIMER_RESTART;
3985         struct perf_sample_data data;
3986         struct pt_regs *regs;
3987         struct perf_event *event;
3988         u64 period;
3989
3990         event   = container_of(hrtimer, struct perf_event, hw.hrtimer);
3991         event->pmu->read(event);
3992
3993         data.addr = 0;
3994         regs = get_irq_regs();
3995         /*
3996          * In case we exclude kernel IPs or are somehow not in interrupt
3997          * context, provide the next best thing, the user IP.
3998          */
3999         if ((event->attr.exclude_kernel || !regs) &&
4000                         !event->attr.exclude_user)
4001                 regs = task_pt_regs(current);
4002
4003         if (regs) {
4004                 if (!(event->attr.exclude_idle && current->pid == 0))
4005                         if (perf_event_overflow(event, 0, &data, regs))
4006                                 ret = HRTIMER_NORESTART;
4007         }
4008
4009         period = max_t(u64, 10000, event->hw.sample_period);
4010         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4011
4012         return ret;
4013 }
4014
4015 static void perf_swevent_start_hrtimer(struct perf_event *event)
4016 {
4017         struct hw_perf_event *hwc = &event->hw;
4018
4019         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4020         hwc->hrtimer.function = perf_swevent_hrtimer;
4021         if (hwc->sample_period) {
4022                 u64 period;
4023
4024                 if (hwc->remaining) {
4025                         if (hwc->remaining < 0)
4026                                 period = 10000;
4027                         else
4028                                 period = hwc->remaining;
4029                         hwc->remaining = 0;
4030                 } else {
4031                         period = max_t(u64, 10000, hwc->sample_period);
4032                 }
4033                 __hrtimer_start_range_ns(&hwc->hrtimer,
4034                                 ns_to_ktime(period), 0,
4035                                 HRTIMER_MODE_REL, 0);
4036         }
4037 }
4038
4039 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4040 {
4041         struct hw_perf_event *hwc = &event->hw;
4042
4043         if (hwc->sample_period) {
4044                 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4045                 hwc->remaining = ktime_to_ns(remaining);
4046
4047                 hrtimer_cancel(&hwc->hrtimer);
4048         }
4049 }
4050
4051 /*
4052  * Software event: cpu wall time clock
4053  */
4054
4055 static void cpu_clock_perf_event_update(struct perf_event *event)
4056 {
4057         int cpu = raw_smp_processor_id();
4058         s64 prev;
4059         u64 now;
4060
4061         now = cpu_clock(cpu);
4062         prev = atomic64_read(&event->hw.prev_count);
4063         atomic64_set(&event->hw.prev_count, now);
4064         atomic64_add(now - prev, &event->count);
4065 }
4066
4067 static int cpu_clock_perf_event_enable(struct perf_event *event)
4068 {
4069         struct hw_perf_event *hwc = &event->hw;
4070         int cpu = raw_smp_processor_id();
4071
4072         atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4073         perf_swevent_start_hrtimer(event);
4074
4075         return 0;
4076 }
4077
4078 static void cpu_clock_perf_event_disable(struct perf_event *event)
4079 {
4080         perf_swevent_cancel_hrtimer(event);
4081         cpu_clock_perf_event_update(event);
4082 }
4083
4084 static void cpu_clock_perf_event_read(struct perf_event *event)
4085 {
4086         cpu_clock_perf_event_update(event);
4087 }
4088
4089 static const struct pmu perf_ops_cpu_clock = {
4090         .enable         = cpu_clock_perf_event_enable,
4091         .disable        = cpu_clock_perf_event_disable,
4092         .read           = cpu_clock_perf_event_read,
4093 };
4094
4095 /*
4096  * Software event: task time clock
4097  */
4098
4099 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4100 {
4101         u64 prev;
4102         s64 delta;
4103
4104         prev = atomic64_xchg(&event->hw.prev_count, now);
4105         delta = now - prev;
4106         atomic64_add(delta, &event->count);
4107 }
4108
4109 static int task_clock_perf_event_enable(struct perf_event *event)
4110 {
4111         struct hw_perf_event *hwc = &event->hw;
4112         u64 now;
4113
4114         now = event->ctx->time;
4115
4116         atomic64_set(&hwc->prev_count, now);
4117
4118         perf_swevent_start_hrtimer(event);
4119
4120         return 0;
4121 }
4122
4123 static void task_clock_perf_event_disable(struct perf_event *event)
4124 {
4125         perf_swevent_cancel_hrtimer(event);
4126         task_clock_perf_event_update(event, event->ctx->time);
4127
4128 }
4129
4130 static void task_clock_perf_event_read(struct perf_event *event)
4131 {
4132         u64 time;
4133
4134         if (!in_nmi()) {
4135                 update_context_time(event->ctx);
4136                 time = event->ctx->time;
4137         } else {
4138                 u64 now = perf_clock();
4139                 u64 delta = now - event->ctx->timestamp;
4140                 time = event->ctx->time + delta;
4141         }
4142
4143         task_clock_perf_event_update(event, time);
4144 }
4145
4146 static const struct pmu perf_ops_task_clock = {
4147         .enable         = task_clock_perf_event_enable,
4148         .disable        = task_clock_perf_event_disable,
4149         .read           = task_clock_perf_event_read,
4150 };
4151
4152 #ifdef CONFIG_EVENT_PROFILE
4153
4154 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4155                           int entry_size)
4156 {
4157         struct perf_raw_record raw = {
4158                 .size = entry_size,
4159                 .data = record,
4160         };
4161
4162         struct perf_sample_data data = {
4163                 .addr = addr,
4164                 .raw = &raw,
4165         };
4166
4167         struct pt_regs *regs = get_irq_regs();
4168
4169         if (!regs)
4170                 regs = task_pt_regs(current);
4171
4172         /* Trace events already protected against recursion */
4173         do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4174                                 &data, regs);
4175 }
4176 EXPORT_SYMBOL_GPL(perf_tp_event);
4177
4178 static int perf_tp_event_match(struct perf_event *event,
4179                                 struct perf_sample_data *data)
4180 {
4181         void *record = data->raw->data;
4182
4183         if (likely(!event->filter) || filter_match_preds(event->filter, record))
4184                 return 1;
4185         return 0;
4186 }
4187
4188 static void tp_perf_event_destroy(struct perf_event *event)
4189 {
4190         ftrace_profile_disable(event->attr.config);
4191 }
4192
4193 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4194 {
4195         /*
4196          * Raw tracepoint data is a severe data leak, only allow root to
4197          * have these.
4198          */
4199         if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4200                         perf_paranoid_tracepoint_raw() &&
4201                         !capable(CAP_SYS_ADMIN))
4202                 return ERR_PTR(-EPERM);
4203
4204         if (ftrace_profile_enable(event->attr.config))
4205                 return NULL;
4206
4207         event->destroy = tp_perf_event_destroy;
4208
4209         return &perf_ops_generic;
4210 }
4211
4212 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4213 {
4214         char *filter_str;
4215         int ret;
4216
4217         if (event->attr.type != PERF_TYPE_TRACEPOINT)
4218                 return -EINVAL;
4219
4220         filter_str = strndup_user(arg, PAGE_SIZE);
4221         if (IS_ERR(filter_str))
4222                 return PTR_ERR(filter_str);
4223
4224         ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4225
4226         kfree(filter_str);
4227         return ret;
4228 }
4229
4230 static void perf_event_free_filter(struct perf_event *event)
4231 {
4232         ftrace_profile_free_filter(event);
4233 }
4234
4235 #else
4236
4237 static int perf_tp_event_match(struct perf_event *event,
4238                                 struct perf_sample_data *data)
4239 {
4240         return 1;
4241 }
4242
4243 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4244 {
4245         return NULL;
4246 }
4247
4248 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4249 {
4250         return -ENOENT;
4251 }
4252
4253 static void perf_event_free_filter(struct perf_event *event)
4254 {
4255 }
4256
4257 #endif /* CONFIG_EVENT_PROFILE */
4258
4259 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4260 static void bp_perf_event_destroy(struct perf_event *event)
4261 {
4262         release_bp_slot(event);
4263 }
4264
4265 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4266 {
4267         int err;
4268         /*
4269          * The breakpoint is already filled if we haven't created the counter
4270          * through perf syscall
4271          * FIXME: manage to get trigerred to NULL if it comes from syscalls
4272          */
4273         if (!bp->callback)
4274                 err = register_perf_hw_breakpoint(bp);
4275         else
4276                 err = __register_perf_hw_breakpoint(bp);
4277         if (err)
4278                 return ERR_PTR(err);
4279
4280         bp->destroy = bp_perf_event_destroy;
4281
4282         return &perf_ops_bp;
4283 }
4284
4285 void perf_bp_event(struct perf_event *bp, void *regs)
4286 {
4287         /* TODO */
4288 }
4289 #else
4290 static void bp_perf_event_destroy(struct perf_event *event)
4291 {
4292 }
4293
4294 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4295 {
4296         return NULL;
4297 }
4298
4299 void perf_bp_event(struct perf_event *bp, void *regs)
4300 {
4301 }
4302 #endif
4303
4304 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4305
4306 static void sw_perf_event_destroy(struct perf_event *event)
4307 {
4308         u64 event_id = event->attr.config;
4309
4310         WARN_ON(event->parent);
4311
4312         atomic_dec(&perf_swevent_enabled[event_id]);
4313 }
4314
4315 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4316 {
4317         const struct pmu *pmu = NULL;
4318         u64 event_id = event->attr.config;
4319
4320         /*
4321          * Software events (currently) can't in general distinguish
4322          * between user, kernel and hypervisor events.
4323          * However, context switches and cpu migrations are considered
4324          * to be kernel events, and page faults are never hypervisor
4325          * events.
4326          */
4327         switch (event_id) {
4328         case PERF_COUNT_SW_CPU_CLOCK:
4329                 pmu = &perf_ops_cpu_clock;
4330
4331                 break;
4332         case PERF_COUNT_SW_TASK_CLOCK:
4333                 /*
4334                  * If the user instantiates this as a per-cpu event,
4335                  * use the cpu_clock event instead.
4336                  */
4337                 if (event->ctx->task)
4338                         pmu = &perf_ops_task_clock;
4339                 else
4340                         pmu = &perf_ops_cpu_clock;
4341
4342                 break;
4343         case PERF_COUNT_SW_PAGE_FAULTS:
4344         case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4345         case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4346         case PERF_COUNT_SW_CONTEXT_SWITCHES:
4347         case PERF_COUNT_SW_CPU_MIGRATIONS:
4348         case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4349         case PERF_COUNT_SW_EMULATION_FAULTS:
4350                 if (!event->parent) {
4351                         atomic_inc(&perf_swevent_enabled[event_id]);
4352                         event->destroy = sw_perf_event_destroy;
4353                 }
4354                 pmu = &perf_ops_generic;
4355                 break;
4356         }
4357
4358         return pmu;
4359 }
4360
4361 /*
4362  * Allocate and initialize a event structure
4363  */
4364 static struct perf_event *
4365 perf_event_alloc(struct perf_event_attr *attr,
4366                    int cpu,
4367                    struct perf_event_context *ctx,
4368                    struct perf_event *group_leader,
4369                    struct perf_event *parent_event,
4370                    perf_callback_t callback,
4371                    gfp_t gfpflags)
4372 {
4373         const struct pmu *pmu;
4374         struct perf_event *event;
4375         struct hw_perf_event *hwc;
4376         long err;
4377
4378         event = kzalloc(sizeof(*event), gfpflags);
4379         if (!event)
4380                 return ERR_PTR(-ENOMEM);
4381
4382         /*
4383          * Single events are their own group leaders, with an
4384          * empty sibling list:
4385          */
4386         if (!group_leader)
4387                 group_leader = event;
4388
4389         mutex_init(&event->child_mutex);
4390         INIT_LIST_HEAD(&event->child_list);
4391
4392         INIT_LIST_HEAD(&event->group_entry);
4393         INIT_LIST_HEAD(&event->event_entry);
4394         INIT_LIST_HEAD(&event->sibling_list);
4395         init_waitqueue_head(&event->waitq);
4396
4397         mutex_init(&event->mmap_mutex);
4398
4399         event->cpu              = cpu;
4400         event->attr             = *attr;
4401         event->group_leader     = group_leader;
4402         event->pmu              = NULL;
4403         event->ctx              = ctx;
4404         event->oncpu            = -1;
4405
4406         event->parent           = parent_event;
4407
4408         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
4409         event->id               = atomic64_inc_return(&perf_event_id);
4410
4411         event->state            = PERF_EVENT_STATE_INACTIVE;
4412
4413         if (!callback && parent_event)
4414                 callback = parent_event->callback;
4415         
4416         event->callback = callback;
4417
4418         if (attr->disabled)
4419                 event->state = PERF_EVENT_STATE_OFF;
4420
4421         pmu = NULL;
4422
4423         hwc = &event->hw;
4424         hwc->sample_period = attr->sample_period;
4425         if (attr->freq && attr->sample_freq)
4426                 hwc->sample_period = 1;
4427         hwc->last_period = hwc->sample_period;
4428
4429         atomic64_set(&hwc->period_left, hwc->sample_period);
4430
4431         /*
4432          * we currently do not support PERF_FORMAT_GROUP on inherited events
4433          */
4434         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4435                 goto done;
4436
4437         switch (attr->type) {
4438         case PERF_TYPE_RAW:
4439         case PERF_TYPE_HARDWARE:
4440         case PERF_TYPE_HW_CACHE:
4441                 pmu = hw_perf_event_init(event);
4442                 break;
4443
4444         case PERF_TYPE_SOFTWARE:
4445                 pmu = sw_perf_event_init(event);
4446                 break;
4447
4448         case PERF_TYPE_TRACEPOINT:
4449                 pmu = tp_perf_event_init(event);
4450                 break;
4451
4452         case PERF_TYPE_BREAKPOINT:
4453                 pmu = bp_perf_event_init(event);
4454                 break;
4455
4456
4457         default:
4458                 break;
4459         }
4460 done:
4461         err = 0;
4462         if (!pmu)
4463                 err = -EINVAL;
4464         else if (IS_ERR(pmu))
4465                 err = PTR_ERR(pmu);
4466
4467         if (err) {
4468                 if (event->ns)
4469                         put_pid_ns(event->ns);
4470                 kfree(event);
4471                 return ERR_PTR(err);
4472         }
4473
4474         event->pmu = pmu;
4475
4476         if (!event->parent) {
4477                 atomic_inc(&nr_events);
4478                 if (event->attr.mmap)
4479                         atomic_inc(&nr_mmap_events);
4480                 if (event->attr.comm)
4481                         atomic_inc(&nr_comm_events);
4482                 if (event->attr.task)
4483                         atomic_inc(&nr_task_events);
4484         }
4485
4486         return event;
4487 }
4488
4489 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4490                           struct perf_event_attr *attr)
4491 {
4492         u32 size;
4493         int ret;
4494
4495         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4496                 return -EFAULT;
4497
4498         /*
4499          * zero the full structure, so that a short copy will be nice.
4500          */
4501         memset(attr, 0, sizeof(*attr));
4502
4503         ret = get_user(size, &uattr->size);
4504         if (ret)
4505                 return ret;
4506
4507         if (size > PAGE_SIZE)   /* silly large */
4508                 goto err_size;
4509
4510         if (!size)              /* abi compat */
4511                 size = PERF_ATTR_SIZE_VER0;
4512
4513         if (size < PERF_ATTR_SIZE_VER0)
4514                 goto err_size;
4515
4516         /*
4517          * If we're handed a bigger struct than we know of,
4518          * ensure all the unknown bits are 0 - i.e. new
4519          * user-space does not rely on any kernel feature
4520          * extensions we dont know about yet.
4521          */
4522         if (size > sizeof(*attr)) {
4523                 unsigned char __user *addr;
4524                 unsigned char __user *end;
4525                 unsigned char val;
4526
4527                 addr = (void __user *)uattr + sizeof(*attr);
4528                 end  = (void __user *)uattr + size;
4529
4530                 for (; addr < end; addr++) {
4531                         ret = get_user(val, addr);
4532                         if (ret)
4533                                 return ret;
4534                         if (val)
4535                                 goto err_size;
4536                 }
4537                 size = sizeof(*attr);
4538         }
4539
4540         ret = copy_from_user(attr, uattr, size);
4541         if (ret)
4542                 return -EFAULT;
4543
4544         /*
4545          * If the type exists, the corresponding creation will verify
4546          * the attr->config.
4547          */
4548         if (attr->type >= PERF_TYPE_MAX)
4549                 return -EINVAL;
4550
4551         if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3)
4552                 return -EINVAL;
4553
4554         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4555                 return -EINVAL;
4556
4557         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4558                 return -EINVAL;
4559
4560 out:
4561         return ret;
4562
4563 err_size:
4564         put_user(sizeof(*attr), &uattr->size);
4565         ret = -E2BIG;
4566         goto out;
4567 }
4568
4569 static int perf_event_set_output(struct perf_event *event, int output_fd)
4570 {
4571         struct perf_event *output_event = NULL;
4572         struct file *output_file = NULL;
4573         struct perf_event *old_output;
4574         int fput_needed = 0;
4575         int ret = -EINVAL;
4576
4577         if (!output_fd)
4578                 goto set;
4579
4580         output_file = fget_light(output_fd, &fput_needed);
4581         if (!output_file)
4582                 return -EBADF;
4583
4584         if (output_file->f_op != &perf_fops)
4585                 goto out;
4586
4587         output_event = output_file->private_data;
4588
4589         /* Don't chain output fds */
4590         if (output_event->output)
4591                 goto out;
4592
4593         /* Don't set an output fd when we already have an output channel */
4594         if (event->data)
4595                 goto out;
4596
4597         atomic_long_inc(&output_file->f_count);
4598
4599 set:
4600         mutex_lock(&event->mmap_mutex);
4601         old_output = event->output;
4602         rcu_assign_pointer(event->output, output_event);
4603         mutex_unlock(&event->mmap_mutex);
4604
4605         if (old_output) {
4606                 /*
4607                  * we need to make sure no existing perf_output_*()
4608                  * is still referencing this event.
4609                  */
4610                 synchronize_rcu();
4611                 fput(old_output->filp);
4612         }
4613
4614         ret = 0;
4615 out:
4616         fput_light(output_file, fput_needed);
4617         return ret;
4618 }
4619
4620 /**
4621  * sys_perf_event_open - open a performance event, associate it to a task/cpu
4622  *
4623  * @attr_uptr:  event_id type attributes for monitoring/sampling
4624  * @pid:                target pid
4625  * @cpu:                target cpu
4626  * @group_fd:           group leader event fd
4627  */
4628 SYSCALL_DEFINE5(perf_event_open,
4629                 struct perf_event_attr __user *, attr_uptr,
4630                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4631 {
4632         struct perf_event *event, *group_leader;
4633         struct perf_event_attr attr;
4634         struct perf_event_context *ctx;
4635         struct file *event_file = NULL;
4636         struct file *group_file = NULL;
4637         int fput_needed = 0;
4638         int fput_needed2 = 0;
4639         int err;
4640
4641         /* for future expandability... */
4642         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4643                 return -EINVAL;
4644
4645         err = perf_copy_attr(attr_uptr, &attr);
4646         if (err)
4647                 return err;
4648
4649         if (!attr.exclude_kernel) {
4650                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4651                         return -EACCES;
4652         }
4653
4654         if (attr.freq) {
4655                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4656                         return -EINVAL;
4657         }
4658
4659         /*
4660          * Get the target context (task or percpu):
4661          */
4662         ctx = find_get_context(pid, cpu);
4663         if (IS_ERR(ctx))
4664                 return PTR_ERR(ctx);
4665
4666         /*
4667          * Look up the group leader (we will attach this event to it):
4668          */
4669         group_leader = NULL;
4670         if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4671                 err = -EINVAL;
4672                 group_file = fget_light(group_fd, &fput_needed);
4673                 if (!group_file)
4674                         goto err_put_context;
4675                 if (group_file->f_op != &perf_fops)
4676                         goto err_put_context;
4677
4678                 group_leader = group_file->private_data;
4679                 /*
4680                  * Do not allow a recursive hierarchy (this new sibling
4681                  * becoming part of another group-sibling):
4682                  */
4683                 if (group_leader->group_leader != group_leader)
4684                         goto err_put_context;
4685                 /*
4686                  * Do not allow to attach to a group in a different
4687                  * task or CPU context:
4688                  */
4689                 if (group_leader->ctx != ctx)
4690                         goto err_put_context;
4691                 /*
4692                  * Only a group leader can be exclusive or pinned
4693                  */
4694                 if (attr.exclusive || attr.pinned)
4695                         goto err_put_context;
4696         }
4697
4698         event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4699                                      NULL, NULL, GFP_KERNEL);
4700         err = PTR_ERR(event);
4701         if (IS_ERR(event))
4702                 goto err_put_context;
4703
4704         err = anon_inode_getfd("[perf_event]", &perf_fops, event, 0);
4705         if (err < 0)
4706                 goto err_free_put_context;
4707
4708         event_file = fget_light(err, &fput_needed2);
4709         if (!event_file)
4710                 goto err_free_put_context;
4711
4712         if (flags & PERF_FLAG_FD_OUTPUT) {
4713                 err = perf_event_set_output(event, group_fd);
4714                 if (err)
4715                         goto err_fput_free_put_context;
4716         }
4717
4718         event->filp = event_file;
4719         WARN_ON_ONCE(ctx->parent_ctx);
4720         mutex_lock(&ctx->mutex);
4721         perf_install_in_context(ctx, event, cpu);
4722         ++ctx->generation;
4723         mutex_unlock(&ctx->mutex);
4724
4725         event->owner = current;
4726         get_task_struct(current);
4727         mutex_lock(&current->perf_event_mutex);
4728         list_add_tail(&event->owner_entry, &current->perf_event_list);
4729         mutex_unlock(&current->perf_event_mutex);
4730
4731 err_fput_free_put_context:
4732         fput_light(event_file, fput_needed2);
4733
4734 err_free_put_context:
4735         if (err < 0)
4736                 kfree(event);
4737
4738 err_put_context:
4739         if (err < 0)
4740                 put_ctx(ctx);
4741
4742         fput_light(group_file, fput_needed);
4743
4744         return err;
4745 }
4746
4747 /**
4748  * perf_event_create_kernel_counter
4749  *
4750  * @attr: attributes of the counter to create
4751  * @cpu: cpu in which the counter is bound
4752  * @pid: task to profile
4753  */
4754 struct perf_event *
4755 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4756                                  pid_t pid, perf_callback_t callback)
4757 {
4758         struct perf_event *event;
4759         struct perf_event_context *ctx;
4760         int err;
4761
4762         /*
4763          * Get the target context (task or percpu):
4764          */
4765
4766         ctx = find_get_context(pid, cpu);
4767         if (IS_ERR(ctx))
4768                 return NULL;
4769
4770         event = perf_event_alloc(attr, cpu, ctx, NULL,
4771                                      NULL, callback, GFP_KERNEL);
4772         err = PTR_ERR(event);
4773         if (IS_ERR(event))
4774                 goto err_put_context;
4775
4776         event->filp = NULL;
4777         WARN_ON_ONCE(ctx->parent_ctx);
4778         mutex_lock(&ctx->mutex);
4779         perf_install_in_context(ctx, event, cpu);
4780         ++ctx->generation;
4781         mutex_unlock(&ctx->mutex);
4782
4783         event->owner = current;
4784         get_task_struct(current);
4785         mutex_lock(&current->perf_event_mutex);
4786         list_add_tail(&event->owner_entry, &current->perf_event_list);
4787         mutex_unlock(&current->perf_event_mutex);
4788
4789         return event;
4790
4791 err_put_context:
4792         if (err < 0)
4793                 put_ctx(ctx);
4794
4795         return NULL;
4796 }
4797 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4798
4799 /*
4800  * inherit a event from parent task to child task:
4801  */
4802 static struct perf_event *
4803 inherit_event(struct perf_event *parent_event,
4804               struct task_struct *parent,
4805               struct perf_event_context *parent_ctx,
4806               struct task_struct *child,
4807               struct perf_event *group_leader,
4808               struct perf_event_context *child_ctx)
4809 {
4810         struct perf_event *child_event;
4811
4812         /*
4813          * Instead of creating recursive hierarchies of events,
4814          * we link inherited events back to the original parent,
4815          * which has a filp for sure, which we use as the reference
4816          * count:
4817          */
4818         if (parent_event->parent)
4819                 parent_event = parent_event->parent;
4820
4821         child_event = perf_event_alloc(&parent_event->attr,
4822                                            parent_event->cpu, child_ctx,
4823                                            group_leader, parent_event,
4824                                            NULL, GFP_KERNEL);
4825         if (IS_ERR(child_event))
4826                 return child_event;
4827         get_ctx(child_ctx);
4828
4829         /*
4830          * Make the child state follow the state of the parent event,
4831          * not its attr.disabled bit.  We hold the parent's mutex,
4832          * so we won't race with perf_event_{en, dis}able_family.
4833          */
4834         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
4835                 child_event->state = PERF_EVENT_STATE_INACTIVE;
4836         else
4837                 child_event->state = PERF_EVENT_STATE_OFF;
4838
4839         if (parent_event->attr.freq)
4840                 child_event->hw.sample_period = parent_event->hw.sample_period;
4841
4842         child_event->overflow_handler = parent_event->overflow_handler;
4843
4844         /*
4845          * Link it up in the child's context:
4846          */
4847         add_event_to_ctx(child_event, child_ctx);
4848
4849         /*
4850          * Get a reference to the parent filp - we will fput it
4851          * when the child event exits. This is safe to do because
4852          * we are in the parent and we know that the filp still
4853          * exists and has a nonzero count:
4854          */
4855         atomic_long_inc(&parent_event->filp->f_count);
4856
4857         /*
4858          * Link this into the parent event's child list
4859          */
4860         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4861         mutex_lock(&parent_event->child_mutex);
4862         list_add_tail(&child_event->child_list, &parent_event->child_list);
4863         mutex_unlock(&parent_event->child_mutex);
4864
4865         return child_event;
4866 }
4867
4868 static int inherit_group(struct perf_event *parent_event,
4869               struct task_struct *parent,
4870               struct perf_event_context *parent_ctx,
4871               struct task_struct *child,
4872               struct perf_event_context *child_ctx)
4873 {
4874         struct perf_event *leader;
4875         struct perf_event *sub;
4876         struct perf_event *child_ctr;
4877
4878         leader = inherit_event(parent_event, parent, parent_ctx,
4879                                  child, NULL, child_ctx);
4880         if (IS_ERR(leader))
4881                 return PTR_ERR(leader);
4882         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
4883                 child_ctr = inherit_event(sub, parent, parent_ctx,
4884                                             child, leader, child_ctx);
4885                 if (IS_ERR(child_ctr))
4886                         return PTR_ERR(child_ctr);
4887         }
4888         return 0;
4889 }
4890
4891 static void sync_child_event(struct perf_event *child_event,
4892                                struct task_struct *child)
4893 {
4894         struct perf_event *parent_event = child_event->parent;
4895         u64 child_val;
4896
4897         if (child_event->attr.inherit_stat)
4898                 perf_event_read_event(child_event, child);
4899
4900         child_val = atomic64_read(&child_event->count);
4901
4902         /*
4903          * Add back the child's count to the parent's count:
4904          */
4905         atomic64_add(child_val, &parent_event->count);
4906         atomic64_add(child_event->total_time_enabled,
4907                      &parent_event->child_total_time_enabled);
4908         atomic64_add(child_event->total_time_running,
4909                      &parent_event->child_total_time_running);
4910
4911         /*
4912          * Remove this event from the parent's list
4913          */
4914         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4915         mutex_lock(&parent_event->child_mutex);
4916         list_del_init(&child_event->child_list);
4917         mutex_unlock(&parent_event->child_mutex);
4918
4919         /*
4920          * Release the parent event, if this was the last
4921          * reference to it.
4922          */
4923         fput(parent_event->filp);
4924 }
4925
4926 static void
4927 __perf_event_exit_task(struct perf_event *child_event,
4928                          struct perf_event_context *child_ctx,
4929                          struct task_struct *child)
4930 {
4931         struct perf_event *parent_event;
4932
4933         perf_event_remove_from_context(child_event);
4934
4935         parent_event = child_event->parent;
4936         /*
4937          * It can happen that parent exits first, and has events
4938          * that are still around due to the child reference. These
4939          * events need to be zapped - but otherwise linger.
4940          */
4941         if (parent_event) {
4942                 sync_child_event(child_event, child);
4943                 free_event(child_event);
4944         }
4945 }
4946
4947 /*
4948  * When a child task exits, feed back event values to parent events.
4949  */
4950 void perf_event_exit_task(struct task_struct *child)
4951 {
4952         struct perf_event *child_event, *tmp;
4953         struct perf_event_context *child_ctx;
4954         unsigned long flags;
4955
4956         if (likely(!child->perf_event_ctxp)) {
4957                 perf_event_task(child, NULL, 0);
4958                 return;
4959         }
4960
4961         local_irq_save(flags);
4962         /*
4963          * We can't reschedule here because interrupts are disabled,
4964          * and either child is current or it is a task that can't be
4965          * scheduled, so we are now safe from rescheduling changing
4966          * our context.
4967          */
4968         child_ctx = child->perf_event_ctxp;
4969         __perf_event_task_sched_out(child_ctx);
4970
4971         /*
4972          * Take the context lock here so that if find_get_context is
4973          * reading child->perf_event_ctxp, we wait until it has
4974          * incremented the context's refcount before we do put_ctx below.
4975          */
4976         spin_lock(&child_ctx->lock);
4977         child->perf_event_ctxp = NULL;
4978         /*
4979          * If this context is a clone; unclone it so it can't get
4980          * swapped to another process while we're removing all
4981          * the events from it.
4982          */
4983         unclone_ctx(child_ctx);
4984         update_context_time(child_ctx);
4985         spin_unlock_irqrestore(&child_ctx->lock, flags);
4986
4987         /*
4988          * Report the task dead after unscheduling the events so that we
4989          * won't get any samples after PERF_RECORD_EXIT. We can however still
4990          * get a few PERF_RECORD_READ events.
4991          */
4992         perf_event_task(child, child_ctx, 0);
4993
4994         /*
4995          * We can recurse on the same lock type through:
4996          *
4997          *   __perf_event_exit_task()
4998          *     sync_child_event()
4999          *       fput(parent_event->filp)
5000          *         perf_release()
5001          *           mutex_lock(&ctx->mutex)
5002          *
5003          * But since its the parent context it won't be the same instance.
5004          */
5005         mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
5006
5007 again:
5008         list_for_each_entry_safe(child_event, tmp, &child_ctx->group_list,
5009                                  group_entry)
5010                 __perf_event_exit_task(child_event, child_ctx, child);
5011
5012         /*
5013          * If the last event was a group event, it will have appended all
5014          * its siblings to the list, but we obtained 'tmp' before that which
5015          * will still point to the list head terminating the iteration.
5016          */
5017         if (!list_empty(&child_ctx->group_list))
5018                 goto again;
5019
5020         mutex_unlock(&child_ctx->mutex);
5021
5022         put_ctx(child_ctx);
5023 }
5024
5025 /*
5026  * free an unexposed, unused context as created by inheritance by
5027  * init_task below, used by fork() in case of fail.
5028  */
5029 void perf_event_free_task(struct task_struct *task)
5030 {
5031         struct perf_event_context *ctx = task->perf_event_ctxp;
5032         struct perf_event *event, *tmp;
5033
5034         if (!ctx)
5035                 return;
5036
5037         mutex_lock(&ctx->mutex);
5038 again:
5039         list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry) {
5040                 struct perf_event *parent = event->parent;
5041
5042                 if (WARN_ON_ONCE(!parent))
5043                         continue;
5044
5045                 mutex_lock(&parent->child_mutex);
5046                 list_del_init(&event->child_list);
5047                 mutex_unlock(&parent->child_mutex);
5048
5049                 fput(parent->filp);
5050
5051                 list_del_event(event, ctx);
5052                 free_event(event);
5053         }
5054
5055         if (!list_empty(&ctx->group_list))
5056                 goto again;
5057
5058         mutex_unlock(&ctx->mutex);
5059
5060         put_ctx(ctx);
5061 }
5062
5063 /*
5064  * Initialize the perf_event context in task_struct
5065  */
5066 int perf_event_init_task(struct task_struct *child)
5067 {
5068         struct perf_event_context *child_ctx, *parent_ctx;
5069         struct perf_event_context *cloned_ctx;
5070         struct perf_event *event;
5071         struct task_struct *parent = current;
5072         int inherited_all = 1;
5073         int ret = 0;
5074
5075         child->perf_event_ctxp = NULL;
5076
5077         mutex_init(&child->perf_event_mutex);
5078         INIT_LIST_HEAD(&child->perf_event_list);
5079
5080         if (likely(!parent->perf_event_ctxp))
5081                 return 0;
5082
5083         /*
5084          * This is executed from the parent task context, so inherit
5085          * events that have been marked for cloning.
5086          * First allocate and initialize a context for the child.
5087          */
5088
5089         child_ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
5090         if (!child_ctx)
5091                 return -ENOMEM;
5092
5093         __perf_event_init_context(child_ctx, child);
5094         child->perf_event_ctxp = child_ctx;
5095         get_task_struct(child);
5096
5097         /*
5098          * If the parent's context is a clone, pin it so it won't get
5099          * swapped under us.
5100          */
5101         parent_ctx = perf_pin_task_context(parent);
5102
5103         /*
5104          * No need to check if parent_ctx != NULL here; since we saw
5105          * it non-NULL earlier, the only reason for it to become NULL
5106          * is if we exit, and since we're currently in the middle of
5107          * a fork we can't be exiting at the same time.
5108          */
5109
5110         /*
5111          * Lock the parent list. No need to lock the child - not PID
5112          * hashed yet and not running, so nobody can access it.
5113          */
5114         mutex_lock(&parent_ctx->mutex);
5115
5116         /*
5117          * We dont have to disable NMIs - we are only looking at
5118          * the list, not manipulating it:
5119          */
5120         list_for_each_entry(event, &parent_ctx->group_list, group_entry) {
5121
5122                 if (!event->attr.inherit) {
5123                         inherited_all = 0;
5124                         continue;
5125                 }
5126
5127                 ret = inherit_group(event, parent, parent_ctx,
5128                                              child, child_ctx);
5129                 if (ret) {
5130                         inherited_all = 0;
5131                         break;
5132                 }
5133         }
5134
5135         if (inherited_all) {
5136                 /*
5137                  * Mark the child context as a clone of the parent
5138                  * context, or of whatever the parent is a clone of.
5139                  * Note that if the parent is a clone, it could get
5140                  * uncloned at any point, but that doesn't matter
5141                  * because the list of events and the generation
5142                  * count can't have changed since we took the mutex.
5143                  */
5144                 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5145                 if (cloned_ctx) {
5146                         child_ctx->parent_ctx = cloned_ctx;
5147                         child_ctx->parent_gen = parent_ctx->parent_gen;
5148                 } else {
5149                         child_ctx->parent_ctx = parent_ctx;
5150                         child_ctx->parent_gen = parent_ctx->generation;
5151                 }
5152                 get_ctx(child_ctx->parent_ctx);
5153         }
5154
5155         mutex_unlock(&parent_ctx->mutex);
5156
5157         perf_unpin_context(parent_ctx);
5158
5159         return ret;
5160 }
5161
5162 static void __cpuinit perf_event_init_cpu(int cpu)
5163 {
5164         struct perf_cpu_context *cpuctx;
5165
5166         cpuctx = &per_cpu(perf_cpu_context, cpu);
5167         __perf_event_init_context(&cpuctx->ctx, NULL);
5168
5169         spin_lock(&perf_resource_lock);
5170         cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5171         spin_unlock(&perf_resource_lock);
5172
5173         hw_perf_event_setup(cpu);
5174 }
5175
5176 #ifdef CONFIG_HOTPLUG_CPU
5177 static void __perf_event_exit_cpu(void *info)
5178 {
5179         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5180         struct perf_event_context *ctx = &cpuctx->ctx;
5181         struct perf_event *event, *tmp;
5182
5183         list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry)
5184                 __perf_event_remove_from_context(event);
5185 }
5186 static void perf_event_exit_cpu(int cpu)
5187 {
5188         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5189         struct perf_event_context *ctx = &cpuctx->ctx;
5190
5191         mutex_lock(&ctx->mutex);
5192         smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5193         mutex_unlock(&ctx->mutex);
5194 }
5195 #else
5196 static inline void perf_event_exit_cpu(int cpu) { }
5197 #endif
5198
5199 static int __cpuinit
5200 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5201 {
5202         unsigned int cpu = (long)hcpu;
5203
5204         switch (action) {
5205
5206         case CPU_UP_PREPARE:
5207         case CPU_UP_PREPARE_FROZEN:
5208                 perf_event_init_cpu(cpu);
5209                 break;
5210
5211         case CPU_ONLINE:
5212         case CPU_ONLINE_FROZEN:
5213                 hw_perf_event_setup_online(cpu);
5214                 break;
5215
5216         case CPU_DOWN_PREPARE:
5217         case CPU_DOWN_PREPARE_FROZEN:
5218                 perf_event_exit_cpu(cpu);
5219                 break;
5220
5221         default:
5222                 break;
5223         }
5224
5225         return NOTIFY_OK;
5226 }
5227
5228 /*
5229  * This has to have a higher priority than migration_notifier in sched.c.
5230  */
5231 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5232         .notifier_call          = perf_cpu_notify,
5233         .priority               = 20,
5234 };
5235
5236 void __init perf_event_init(void)
5237 {
5238         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5239                         (void *)(long)smp_processor_id());
5240         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5241                         (void *)(long)smp_processor_id());
5242         register_cpu_notifier(&perf_cpu_nb);
5243 }
5244
5245 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
5246 {
5247         return sprintf(buf, "%d\n", perf_reserved_percpu);
5248 }
5249
5250 static ssize_t
5251 perf_set_reserve_percpu(struct sysdev_class *class,
5252                         const char *buf,
5253                         size_t count)
5254 {
5255         struct perf_cpu_context *cpuctx;
5256         unsigned long val;
5257         int err, cpu, mpt;
5258
5259         err = strict_strtoul(buf, 10, &val);
5260         if (err)
5261                 return err;
5262         if (val > perf_max_events)
5263                 return -EINVAL;
5264
5265         spin_lock(&perf_resource_lock);
5266         perf_reserved_percpu = val;
5267         for_each_online_cpu(cpu) {
5268                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5269                 spin_lock_irq(&cpuctx->ctx.lock);
5270                 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5271                           perf_max_events - perf_reserved_percpu);
5272                 cpuctx->max_pertask = mpt;
5273                 spin_unlock_irq(&cpuctx->ctx.lock);
5274         }
5275         spin_unlock(&perf_resource_lock);
5276
5277         return count;
5278 }
5279
5280 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
5281 {
5282         return sprintf(buf, "%d\n", perf_overcommit);
5283 }
5284
5285 static ssize_t
5286 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
5287 {
5288         unsigned long val;
5289         int err;
5290
5291         err = strict_strtoul(buf, 10, &val);
5292         if (err)
5293                 return err;
5294         if (val > 1)
5295                 return -EINVAL;
5296
5297         spin_lock(&perf_resource_lock);
5298         perf_overcommit = val;
5299         spin_unlock(&perf_resource_lock);
5300
5301         return count;
5302 }
5303
5304 static SYSDEV_CLASS_ATTR(
5305                                 reserve_percpu,
5306                                 0644,
5307                                 perf_show_reserve_percpu,
5308                                 perf_set_reserve_percpu
5309                         );
5310
5311 static SYSDEV_CLASS_ATTR(
5312                                 overcommit,
5313                                 0644,
5314                                 perf_show_overcommit,
5315                                 perf_set_overcommit
5316                         );
5317
5318 static struct attribute *perfclass_attrs[] = {
5319         &attr_reserve_percpu.attr,
5320         &attr_overcommit.attr,
5321         NULL
5322 };
5323
5324 static struct attribute_group perfclass_attr_group = {
5325         .attrs                  = perfclass_attrs,
5326         .name                   = "perf_events",
5327 };
5328
5329 static int __init perf_event_sysfs_init(void)
5330 {
5331         return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5332                                   &perfclass_attr_group);
5333 }
5334 device_initcall(perf_event_sysfs_init);