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