perf: Optimize perf_swevent_ctx_event()
[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         event = list_first_entry(&ctx->event_list,
1124                                    struct perf_event, event_entry);
1125
1126         next_event = list_first_entry(&next_ctx->event_list,
1127                                         struct perf_event, event_entry);
1128
1129         while (&event->event_entry != &ctx->event_list &&
1130                &next_event->event_entry != &next_ctx->event_list) {
1131
1132                 __perf_event_sync_stat(event, next_event);
1133
1134                 event = list_next_entry(event, event_entry);
1135                 next_event = list_next_entry(next_event, event_entry);
1136         }
1137 }
1138
1139 /*
1140  * Called from scheduler to remove the events of the current task,
1141  * with interrupts disabled.
1142  *
1143  * We stop each event and update the event value in event->count.
1144  *
1145  * This does not protect us against NMI, but disable()
1146  * sets the disabled bit in the control field of event _before_
1147  * accessing the event control register. If a NMI hits, then it will
1148  * not restart the event.
1149  */
1150 void perf_event_task_sched_out(struct task_struct *task,
1151                                  struct task_struct *next, int cpu)
1152 {
1153         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1154         struct perf_event_context *ctx = task->perf_event_ctxp;
1155         struct perf_event_context *next_ctx;
1156         struct perf_event_context *parent;
1157         struct pt_regs *regs;
1158         int do_switch = 1;
1159
1160         regs = task_pt_regs(task);
1161         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1162
1163         if (likely(!ctx || !cpuctx->task_ctx))
1164                 return;
1165
1166         update_context_time(ctx);
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_entry(struct perf_event *event,
1788                                    u64 read_format, char __user *buf)
1789 {
1790         int n = 0, count = 0;
1791         u64 values[2];
1792
1793         values[n++] = perf_event_read_value(event);
1794         if (read_format & PERF_FORMAT_ID)
1795                 values[n++] = primary_event_id(event);
1796
1797         count = n * sizeof(u64);
1798
1799         if (copy_to_user(buf, values, count))
1800                 return -EFAULT;
1801
1802         return count;
1803 }
1804
1805 static int perf_event_read_group(struct perf_event *event,
1806                                    u64 read_format, char __user *buf)
1807 {
1808         struct perf_event *leader = event->group_leader, *sub;
1809         int n = 0, size = 0, err = -EFAULT;
1810         u64 values[3];
1811
1812         values[n++] = 1 + leader->nr_siblings;
1813         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1814                 values[n++] = leader->total_time_enabled +
1815                         atomic64_read(&leader->child_total_time_enabled);
1816         }
1817         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1818                 values[n++] = leader->total_time_running +
1819                         atomic64_read(&leader->child_total_time_running);
1820         }
1821
1822         size = n * sizeof(u64);
1823
1824         if (copy_to_user(buf, values, size))
1825                 return -EFAULT;
1826
1827         err = perf_event_read_entry(leader, read_format, buf + size);
1828         if (err < 0)
1829                 return err;
1830
1831         size += err;
1832
1833         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1834                 err = perf_event_read_entry(sub, read_format,
1835                                 buf + size);
1836                 if (err < 0)
1837                         return err;
1838
1839                 size += err;
1840         }
1841
1842         return size;
1843 }
1844
1845 static int perf_event_read_one(struct perf_event *event,
1846                                  u64 read_format, char __user *buf)
1847 {
1848         u64 values[4];
1849         int n = 0;
1850
1851         values[n++] = perf_event_read_value(event);
1852         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1853                 values[n++] = event->total_time_enabled +
1854                         atomic64_read(&event->child_total_time_enabled);
1855         }
1856         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1857                 values[n++] = event->total_time_running +
1858                         atomic64_read(&event->child_total_time_running);
1859         }
1860         if (read_format & PERF_FORMAT_ID)
1861                 values[n++] = primary_event_id(event);
1862
1863         if (copy_to_user(buf, values, n * sizeof(u64)))
1864                 return -EFAULT;
1865
1866         return n * sizeof(u64);
1867 }
1868
1869 /*
1870  * Read the performance event - simple non blocking version for now
1871  */
1872 static ssize_t
1873 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
1874 {
1875         u64 read_format = event->attr.read_format;
1876         int ret;
1877
1878         /*
1879          * Return end-of-file for a read on a event that is in
1880          * error state (i.e. because it was pinned but it couldn't be
1881          * scheduled on to the CPU at some point).
1882          */
1883         if (event->state == PERF_EVENT_STATE_ERROR)
1884                 return 0;
1885
1886         if (count < perf_event_read_size(event))
1887                 return -ENOSPC;
1888
1889         WARN_ON_ONCE(event->ctx->parent_ctx);
1890         mutex_lock(&event->child_mutex);
1891         if (read_format & PERF_FORMAT_GROUP)
1892                 ret = perf_event_read_group(event, read_format, buf);
1893         else
1894                 ret = perf_event_read_one(event, read_format, buf);
1895         mutex_unlock(&event->child_mutex);
1896
1897         return ret;
1898 }
1899
1900 static ssize_t
1901 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1902 {
1903         struct perf_event *event = file->private_data;
1904
1905         return perf_read_hw(event, buf, count);
1906 }
1907
1908 static unsigned int perf_poll(struct file *file, poll_table *wait)
1909 {
1910         struct perf_event *event = file->private_data;
1911         struct perf_mmap_data *data;
1912         unsigned int events = POLL_HUP;
1913
1914         rcu_read_lock();
1915         data = rcu_dereference(event->data);
1916         if (data)
1917                 events = atomic_xchg(&data->poll, 0);
1918         rcu_read_unlock();
1919
1920         poll_wait(file, &event->waitq, wait);
1921
1922         return events;
1923 }
1924
1925 static void perf_event_reset(struct perf_event *event)
1926 {
1927         (void)perf_event_read(event);
1928         atomic64_set(&event->count, 0);
1929         perf_event_update_userpage(event);
1930 }
1931
1932 /*
1933  * Holding the top-level event's child_mutex means that any
1934  * descendant process that has inherited this event will block
1935  * in sync_child_event if it goes to exit, thus satisfying the
1936  * task existence requirements of perf_event_enable/disable.
1937  */
1938 static void perf_event_for_each_child(struct perf_event *event,
1939                                         void (*func)(struct perf_event *))
1940 {
1941         struct perf_event *child;
1942
1943         WARN_ON_ONCE(event->ctx->parent_ctx);
1944         mutex_lock(&event->child_mutex);
1945         func(event);
1946         list_for_each_entry(child, &event->child_list, child_list)
1947                 func(child);
1948         mutex_unlock(&event->child_mutex);
1949 }
1950
1951 static void perf_event_for_each(struct perf_event *event,
1952                                   void (*func)(struct perf_event *))
1953 {
1954         struct perf_event_context *ctx = event->ctx;
1955         struct perf_event *sibling;
1956
1957         WARN_ON_ONCE(ctx->parent_ctx);
1958         mutex_lock(&ctx->mutex);
1959         event = event->group_leader;
1960
1961         perf_event_for_each_child(event, func);
1962         func(event);
1963         list_for_each_entry(sibling, &event->sibling_list, group_entry)
1964                 perf_event_for_each_child(event, func);
1965         mutex_unlock(&ctx->mutex);
1966 }
1967
1968 static int perf_event_period(struct perf_event *event, u64 __user *arg)
1969 {
1970         struct perf_event_context *ctx = event->ctx;
1971         unsigned long size;
1972         int ret = 0;
1973         u64 value;
1974
1975         if (!event->attr.sample_period)
1976                 return -EINVAL;
1977
1978         size = copy_from_user(&value, arg, sizeof(value));
1979         if (size != sizeof(value))
1980                 return -EFAULT;
1981
1982         if (!value)
1983                 return -EINVAL;
1984
1985         spin_lock_irq(&ctx->lock);
1986         if (event->attr.freq) {
1987                 if (value > sysctl_perf_event_sample_rate) {
1988                         ret = -EINVAL;
1989                         goto unlock;
1990                 }
1991
1992                 event->attr.sample_freq = value;
1993         } else {
1994                 event->attr.sample_period = value;
1995                 event->hw.sample_period = value;
1996         }
1997 unlock:
1998         spin_unlock_irq(&ctx->lock);
1999
2000         return ret;
2001 }
2002
2003 static int perf_event_set_output(struct perf_event *event, int output_fd);
2004 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2005
2006 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2007 {
2008         struct perf_event *event = file->private_data;
2009         void (*func)(struct perf_event *);
2010         u32 flags = arg;
2011
2012         switch (cmd) {
2013         case PERF_EVENT_IOC_ENABLE:
2014                 func = perf_event_enable;
2015                 break;
2016         case PERF_EVENT_IOC_DISABLE:
2017                 func = perf_event_disable;
2018                 break;
2019         case PERF_EVENT_IOC_RESET:
2020                 func = perf_event_reset;
2021                 break;
2022
2023         case PERF_EVENT_IOC_REFRESH:
2024                 return perf_event_refresh(event, arg);
2025
2026         case PERF_EVENT_IOC_PERIOD:
2027                 return perf_event_period(event, (u64 __user *)arg);
2028
2029         case PERF_EVENT_IOC_SET_OUTPUT:
2030                 return perf_event_set_output(event, arg);
2031
2032         case PERF_EVENT_IOC_SET_FILTER:
2033                 return perf_event_set_filter(event, (void __user *)arg);
2034
2035         default:
2036                 return -ENOTTY;
2037         }
2038
2039         if (flags & PERF_IOC_FLAG_GROUP)
2040                 perf_event_for_each(event, func);
2041         else
2042                 perf_event_for_each_child(event, func);
2043
2044         return 0;
2045 }
2046
2047 int perf_event_task_enable(void)
2048 {
2049         struct perf_event *event;
2050
2051         mutex_lock(&current->perf_event_mutex);
2052         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2053                 perf_event_for_each_child(event, perf_event_enable);
2054         mutex_unlock(&current->perf_event_mutex);
2055
2056         return 0;
2057 }
2058
2059 int perf_event_task_disable(void)
2060 {
2061         struct perf_event *event;
2062
2063         mutex_lock(&current->perf_event_mutex);
2064         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2065                 perf_event_for_each_child(event, perf_event_disable);
2066         mutex_unlock(&current->perf_event_mutex);
2067
2068         return 0;
2069 }
2070
2071 #ifndef PERF_EVENT_INDEX_OFFSET
2072 # define PERF_EVENT_INDEX_OFFSET 0
2073 #endif
2074
2075 static int perf_event_index(struct perf_event *event)
2076 {
2077         if (event->state != PERF_EVENT_STATE_ACTIVE)
2078                 return 0;
2079
2080         return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2081 }
2082
2083 /*
2084  * Callers need to ensure there can be no nesting of this function, otherwise
2085  * the seqlock logic goes bad. We can not serialize this because the arch
2086  * code calls this from NMI context.
2087  */
2088 void perf_event_update_userpage(struct perf_event *event)
2089 {
2090         struct perf_event_mmap_page *userpg;
2091         struct perf_mmap_data *data;
2092
2093         rcu_read_lock();
2094         data = rcu_dereference(event->data);
2095         if (!data)
2096                 goto unlock;
2097
2098         userpg = data->user_page;
2099
2100         /*
2101          * Disable preemption so as to not let the corresponding user-space
2102          * spin too long if we get preempted.
2103          */
2104         preempt_disable();
2105         ++userpg->lock;
2106         barrier();
2107         userpg->index = perf_event_index(event);
2108         userpg->offset = atomic64_read(&event->count);
2109         if (event->state == PERF_EVENT_STATE_ACTIVE)
2110                 userpg->offset -= atomic64_read(&event->hw.prev_count);
2111
2112         userpg->time_enabled = event->total_time_enabled +
2113                         atomic64_read(&event->child_total_time_enabled);
2114
2115         userpg->time_running = event->total_time_running +
2116                         atomic64_read(&event->child_total_time_running);
2117
2118         barrier();
2119         ++userpg->lock;
2120         preempt_enable();
2121 unlock:
2122         rcu_read_unlock();
2123 }
2124
2125 static unsigned long perf_data_size(struct perf_mmap_data *data)
2126 {
2127         return data->nr_pages << (PAGE_SHIFT + data->data_order);
2128 }
2129
2130 #ifndef CONFIG_PERF_USE_VMALLOC
2131
2132 /*
2133  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2134  */
2135
2136 static struct page *
2137 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2138 {
2139         if (pgoff > data->nr_pages)
2140                 return NULL;
2141
2142         if (pgoff == 0)
2143                 return virt_to_page(data->user_page);
2144
2145         return virt_to_page(data->data_pages[pgoff - 1]);
2146 }
2147
2148 static struct perf_mmap_data *
2149 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2150 {
2151         struct perf_mmap_data *data;
2152         unsigned long size;
2153         int i;
2154
2155         WARN_ON(atomic_read(&event->mmap_count));
2156
2157         size = sizeof(struct perf_mmap_data);
2158         size += nr_pages * sizeof(void *);
2159
2160         data = kzalloc(size, GFP_KERNEL);
2161         if (!data)
2162                 goto fail;
2163
2164         data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2165         if (!data->user_page)
2166                 goto fail_user_page;
2167
2168         for (i = 0; i < nr_pages; i++) {
2169                 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2170                 if (!data->data_pages[i])
2171                         goto fail_data_pages;
2172         }
2173
2174         data->data_order = 0;
2175         data->nr_pages = nr_pages;
2176
2177         return data;
2178
2179 fail_data_pages:
2180         for (i--; i >= 0; i--)
2181                 free_page((unsigned long)data->data_pages[i]);
2182
2183         free_page((unsigned long)data->user_page);
2184
2185 fail_user_page:
2186         kfree(data);
2187
2188 fail:
2189         return NULL;
2190 }
2191
2192 static void perf_mmap_free_page(unsigned long addr)
2193 {
2194         struct page *page = virt_to_page((void *)addr);
2195
2196         page->mapping = NULL;
2197         __free_page(page);
2198 }
2199
2200 static void perf_mmap_data_free(struct perf_mmap_data *data)
2201 {
2202         int i;
2203
2204         perf_mmap_free_page((unsigned long)data->user_page);
2205         for (i = 0; i < data->nr_pages; i++)
2206                 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2207 }
2208
2209 #else
2210
2211 /*
2212  * Back perf_mmap() with vmalloc memory.
2213  *
2214  * Required for architectures that have d-cache aliasing issues.
2215  */
2216
2217 static struct page *
2218 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2219 {
2220         if (pgoff > (1UL << data->data_order))
2221                 return NULL;
2222
2223         return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2224 }
2225
2226 static void perf_mmap_unmark_page(void *addr)
2227 {
2228         struct page *page = vmalloc_to_page(addr);
2229
2230         page->mapping = NULL;
2231 }
2232
2233 static void perf_mmap_data_free_work(struct work_struct *work)
2234 {
2235         struct perf_mmap_data *data;
2236         void *base;
2237         int i, nr;
2238
2239         data = container_of(work, struct perf_mmap_data, work);
2240         nr = 1 << data->data_order;
2241
2242         base = data->user_page;
2243         for (i = 0; i < nr + 1; i++)
2244                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2245
2246         vfree(base);
2247 }
2248
2249 static void perf_mmap_data_free(struct perf_mmap_data *data)
2250 {
2251         schedule_work(&data->work);
2252 }
2253
2254 static struct perf_mmap_data *
2255 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2256 {
2257         struct perf_mmap_data *data;
2258         unsigned long size;
2259         void *all_buf;
2260
2261         WARN_ON(atomic_read(&event->mmap_count));
2262
2263         size = sizeof(struct perf_mmap_data);
2264         size += sizeof(void *);
2265
2266         data = kzalloc(size, GFP_KERNEL);
2267         if (!data)
2268                 goto fail;
2269
2270         INIT_WORK(&data->work, perf_mmap_data_free_work);
2271
2272         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2273         if (!all_buf)
2274                 goto fail_all_buf;
2275
2276         data->user_page = all_buf;
2277         data->data_pages[0] = all_buf + PAGE_SIZE;
2278         data->data_order = ilog2(nr_pages);
2279         data->nr_pages = 1;
2280
2281         return data;
2282
2283 fail_all_buf:
2284         kfree(data);
2285
2286 fail:
2287         return NULL;
2288 }
2289
2290 #endif
2291
2292 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2293 {
2294         struct perf_event *event = vma->vm_file->private_data;
2295         struct perf_mmap_data *data;
2296         int ret = VM_FAULT_SIGBUS;
2297
2298         if (vmf->flags & FAULT_FLAG_MKWRITE) {
2299                 if (vmf->pgoff == 0)
2300                         ret = 0;
2301                 return ret;
2302         }
2303
2304         rcu_read_lock();
2305         data = rcu_dereference(event->data);
2306         if (!data)
2307                 goto unlock;
2308
2309         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2310                 goto unlock;
2311
2312         vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2313         if (!vmf->page)
2314                 goto unlock;
2315
2316         get_page(vmf->page);
2317         vmf->page->mapping = vma->vm_file->f_mapping;
2318         vmf->page->index   = vmf->pgoff;
2319
2320         ret = 0;
2321 unlock:
2322         rcu_read_unlock();
2323
2324         return ret;
2325 }
2326
2327 static void
2328 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2329 {
2330         long max_size = perf_data_size(data);
2331
2332         atomic_set(&data->lock, -1);
2333
2334         if (event->attr.watermark) {
2335                 data->watermark = min_t(long, max_size,
2336                                         event->attr.wakeup_watermark);
2337         }
2338
2339         if (!data->watermark)
2340                 data->watermark = max_t(long, PAGE_SIZE, max_size / 2);
2341
2342
2343         rcu_assign_pointer(event->data, data);
2344 }
2345
2346 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2347 {
2348         struct perf_mmap_data *data;
2349
2350         data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2351         perf_mmap_data_free(data);
2352         kfree(data);
2353 }
2354
2355 static void perf_mmap_data_release(struct perf_event *event)
2356 {
2357         struct perf_mmap_data *data = event->data;
2358
2359         WARN_ON(atomic_read(&event->mmap_count));
2360
2361         rcu_assign_pointer(event->data, NULL);
2362         call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2363 }
2364
2365 static void perf_mmap_open(struct vm_area_struct *vma)
2366 {
2367         struct perf_event *event = vma->vm_file->private_data;
2368
2369         atomic_inc(&event->mmap_count);
2370 }
2371
2372 static void perf_mmap_close(struct vm_area_struct *vma)
2373 {
2374         struct perf_event *event = vma->vm_file->private_data;
2375
2376         WARN_ON_ONCE(event->ctx->parent_ctx);
2377         if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2378                 unsigned long size = perf_data_size(event->data);
2379                 struct user_struct *user = current_user();
2380
2381                 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2382                 vma->vm_mm->locked_vm -= event->data->nr_locked;
2383                 perf_mmap_data_release(event);
2384                 mutex_unlock(&event->mmap_mutex);
2385         }
2386 }
2387
2388 static const struct vm_operations_struct perf_mmap_vmops = {
2389         .open           = perf_mmap_open,
2390         .close          = perf_mmap_close,
2391         .fault          = perf_mmap_fault,
2392         .page_mkwrite   = perf_mmap_fault,
2393 };
2394
2395 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2396 {
2397         struct perf_event *event = file->private_data;
2398         unsigned long user_locked, user_lock_limit;
2399         struct user_struct *user = current_user();
2400         unsigned long locked, lock_limit;
2401         struct perf_mmap_data *data;
2402         unsigned long vma_size;
2403         unsigned long nr_pages;
2404         long user_extra, extra;
2405         int ret = 0;
2406
2407         if (!(vma->vm_flags & VM_SHARED))
2408                 return -EINVAL;
2409
2410         vma_size = vma->vm_end - vma->vm_start;
2411         nr_pages = (vma_size / PAGE_SIZE) - 1;
2412
2413         /*
2414          * If we have data pages ensure they're a power-of-two number, so we
2415          * can do bitmasks instead of modulo.
2416          */
2417         if (nr_pages != 0 && !is_power_of_2(nr_pages))
2418                 return -EINVAL;
2419
2420         if (vma_size != PAGE_SIZE * (1 + nr_pages))
2421                 return -EINVAL;
2422
2423         if (vma->vm_pgoff != 0)
2424                 return -EINVAL;
2425
2426         WARN_ON_ONCE(event->ctx->parent_ctx);
2427         mutex_lock(&event->mmap_mutex);
2428         if (event->output) {
2429                 ret = -EINVAL;
2430                 goto unlock;
2431         }
2432
2433         if (atomic_inc_not_zero(&event->mmap_count)) {
2434                 if (nr_pages != event->data->nr_pages)
2435                         ret = -EINVAL;
2436                 goto unlock;
2437         }
2438
2439         user_extra = nr_pages + 1;
2440         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2441
2442         /*
2443          * Increase the limit linearly with more CPUs:
2444          */
2445         user_lock_limit *= num_online_cpus();
2446
2447         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2448
2449         extra = 0;
2450         if (user_locked > user_lock_limit)
2451                 extra = user_locked - user_lock_limit;
2452
2453         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2454         lock_limit >>= PAGE_SHIFT;
2455         locked = vma->vm_mm->locked_vm + extra;
2456
2457         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2458                 !capable(CAP_IPC_LOCK)) {
2459                 ret = -EPERM;
2460                 goto unlock;
2461         }
2462
2463         WARN_ON(event->data);
2464
2465         data = perf_mmap_data_alloc(event, nr_pages);
2466         ret = -ENOMEM;
2467         if (!data)
2468                 goto unlock;
2469
2470         ret = 0;
2471         perf_mmap_data_init(event, data);
2472
2473         atomic_set(&event->mmap_count, 1);
2474         atomic_long_add(user_extra, &user->locked_vm);
2475         vma->vm_mm->locked_vm += extra;
2476         event->data->nr_locked = extra;
2477         if (vma->vm_flags & VM_WRITE)
2478                 event->data->writable = 1;
2479
2480 unlock:
2481         mutex_unlock(&event->mmap_mutex);
2482
2483         vma->vm_flags |= VM_RESERVED;
2484         vma->vm_ops = &perf_mmap_vmops;
2485
2486         return ret;
2487 }
2488
2489 static int perf_fasync(int fd, struct file *filp, int on)
2490 {
2491         struct inode *inode = filp->f_path.dentry->d_inode;
2492         struct perf_event *event = filp->private_data;
2493         int retval;
2494
2495         mutex_lock(&inode->i_mutex);
2496         retval = fasync_helper(fd, filp, on, &event->fasync);
2497         mutex_unlock(&inode->i_mutex);
2498
2499         if (retval < 0)
2500                 return retval;
2501
2502         return 0;
2503 }
2504
2505 static const struct file_operations perf_fops = {
2506         .release                = perf_release,
2507         .read                   = perf_read,
2508         .poll                   = perf_poll,
2509         .unlocked_ioctl         = perf_ioctl,
2510         .compat_ioctl           = perf_ioctl,
2511         .mmap                   = perf_mmap,
2512         .fasync                 = perf_fasync,
2513 };
2514
2515 /*
2516  * Perf event wakeup
2517  *
2518  * If there's data, ensure we set the poll() state and publish everything
2519  * to user-space before waking everybody up.
2520  */
2521
2522 void perf_event_wakeup(struct perf_event *event)
2523 {
2524         wake_up_all(&event->waitq);
2525
2526         if (event->pending_kill) {
2527                 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2528                 event->pending_kill = 0;
2529         }
2530 }
2531
2532 /*
2533  * Pending wakeups
2534  *
2535  * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2536  *
2537  * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2538  * single linked list and use cmpxchg() to add entries lockless.
2539  */
2540
2541 static void perf_pending_event(struct perf_pending_entry *entry)
2542 {
2543         struct perf_event *event = container_of(entry,
2544                         struct perf_event, pending);
2545
2546         if (event->pending_disable) {
2547                 event->pending_disable = 0;
2548                 __perf_event_disable(event);
2549         }
2550
2551         if (event->pending_wakeup) {
2552                 event->pending_wakeup = 0;
2553                 perf_event_wakeup(event);
2554         }
2555 }
2556
2557 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2558
2559 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2560         PENDING_TAIL,
2561 };
2562
2563 static void perf_pending_queue(struct perf_pending_entry *entry,
2564                                void (*func)(struct perf_pending_entry *))
2565 {
2566         struct perf_pending_entry **head;
2567
2568         if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2569                 return;
2570
2571         entry->func = func;
2572
2573         head = &get_cpu_var(perf_pending_head);
2574
2575         do {
2576                 entry->next = *head;
2577         } while (cmpxchg(head, entry->next, entry) != entry->next);
2578
2579         set_perf_event_pending();
2580
2581         put_cpu_var(perf_pending_head);
2582 }
2583
2584 static int __perf_pending_run(void)
2585 {
2586         struct perf_pending_entry *list;
2587         int nr = 0;
2588
2589         list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2590         while (list != PENDING_TAIL) {
2591                 void (*func)(struct perf_pending_entry *);
2592                 struct perf_pending_entry *entry = list;
2593
2594                 list = list->next;
2595
2596                 func = entry->func;
2597                 entry->next = NULL;
2598                 /*
2599                  * Ensure we observe the unqueue before we issue the wakeup,
2600                  * so that we won't be waiting forever.
2601                  * -- see perf_not_pending().
2602                  */
2603                 smp_wmb();
2604
2605                 func(entry);
2606                 nr++;
2607         }
2608
2609         return nr;
2610 }
2611
2612 static inline int perf_not_pending(struct perf_event *event)
2613 {
2614         /*
2615          * If we flush on whatever cpu we run, there is a chance we don't
2616          * need to wait.
2617          */
2618         get_cpu();
2619         __perf_pending_run();
2620         put_cpu();
2621
2622         /*
2623          * Ensure we see the proper queue state before going to sleep
2624          * so that we do not miss the wakeup. -- see perf_pending_handle()
2625          */
2626         smp_rmb();
2627         return event->pending.next == NULL;
2628 }
2629
2630 static void perf_pending_sync(struct perf_event *event)
2631 {
2632         wait_event(event->waitq, perf_not_pending(event));
2633 }
2634
2635 void perf_event_do_pending(void)
2636 {
2637         __perf_pending_run();
2638 }
2639
2640 /*
2641  * Callchain support -- arch specific
2642  */
2643
2644 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2645 {
2646         return NULL;
2647 }
2648
2649 /*
2650  * Output
2651  */
2652 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2653                               unsigned long offset, unsigned long head)
2654 {
2655         unsigned long mask;
2656
2657         if (!data->writable)
2658                 return true;
2659
2660         mask = perf_data_size(data) - 1;
2661
2662         offset = (offset - tail) & mask;
2663         head   = (head   - tail) & mask;
2664
2665         if ((int)(head - offset) < 0)
2666                 return false;
2667
2668         return true;
2669 }
2670
2671 static void perf_output_wakeup(struct perf_output_handle *handle)
2672 {
2673         atomic_set(&handle->data->poll, POLL_IN);
2674
2675         if (handle->nmi) {
2676                 handle->event->pending_wakeup = 1;
2677                 perf_pending_queue(&handle->event->pending,
2678                                    perf_pending_event);
2679         } else
2680                 perf_event_wakeup(handle->event);
2681 }
2682
2683 /*
2684  * Curious locking construct.
2685  *
2686  * We need to ensure a later event_id doesn't publish a head when a former
2687  * event_id isn't done writing. However since we need to deal with NMIs we
2688  * cannot fully serialize things.
2689  *
2690  * What we do is serialize between CPUs so we only have to deal with NMI
2691  * nesting on a single CPU.
2692  *
2693  * We only publish the head (and generate a wakeup) when the outer-most
2694  * event_id completes.
2695  */
2696 static void perf_output_lock(struct perf_output_handle *handle)
2697 {
2698         struct perf_mmap_data *data = handle->data;
2699         int cur, cpu = get_cpu();
2700
2701         handle->locked = 0;
2702
2703         for (;;) {
2704                 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2705                 if (cur == -1) {
2706                         handle->locked = 1;
2707                         break;
2708                 }
2709                 if (cur == cpu)
2710                         break;
2711
2712                 cpu_relax();
2713         }
2714 }
2715
2716 static void perf_output_unlock(struct perf_output_handle *handle)
2717 {
2718         struct perf_mmap_data *data = handle->data;
2719         unsigned long head;
2720         int cpu;
2721
2722         data->done_head = data->head;
2723
2724         if (!handle->locked)
2725                 goto out;
2726
2727 again:
2728         /*
2729          * The xchg implies a full barrier that ensures all writes are done
2730          * before we publish the new head, matched by a rmb() in userspace when
2731          * reading this position.
2732          */
2733         while ((head = atomic_long_xchg(&data->done_head, 0)))
2734                 data->user_page->data_head = head;
2735
2736         /*
2737          * NMI can happen here, which means we can miss a done_head update.
2738          */
2739
2740         cpu = atomic_xchg(&data->lock, -1);
2741         WARN_ON_ONCE(cpu != smp_processor_id());
2742
2743         /*
2744          * Therefore we have to validate we did not indeed do so.
2745          */
2746         if (unlikely(atomic_long_read(&data->done_head))) {
2747                 /*
2748                  * Since we had it locked, we can lock it again.
2749                  */
2750                 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2751                         cpu_relax();
2752
2753                 goto again;
2754         }
2755
2756         if (atomic_xchg(&data->wakeup, 0))
2757                 perf_output_wakeup(handle);
2758 out:
2759         put_cpu();
2760 }
2761
2762 void perf_output_copy(struct perf_output_handle *handle,
2763                       const void *buf, unsigned int len)
2764 {
2765         unsigned int pages_mask;
2766         unsigned long offset;
2767         unsigned int size;
2768         void **pages;
2769
2770         offset          = handle->offset;
2771         pages_mask      = handle->data->nr_pages - 1;
2772         pages           = handle->data->data_pages;
2773
2774         do {
2775                 unsigned long page_offset;
2776                 unsigned long page_size;
2777                 int nr;
2778
2779                 nr          = (offset >> PAGE_SHIFT) & pages_mask;
2780                 page_size   = 1UL << (handle->data->data_order + PAGE_SHIFT);
2781                 page_offset = offset & (page_size - 1);
2782                 size        = min_t(unsigned int, page_size - page_offset, len);
2783
2784                 memcpy(pages[nr] + page_offset, buf, size);
2785
2786                 len         -= size;
2787                 buf         += size;
2788                 offset      += size;
2789         } while (len);
2790
2791         handle->offset = offset;
2792
2793         /*
2794          * Check we didn't copy past our reservation window, taking the
2795          * possible unsigned int wrap into account.
2796          */
2797         WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2798 }
2799
2800 int perf_output_begin(struct perf_output_handle *handle,
2801                       struct perf_event *event, unsigned int size,
2802                       int nmi, int sample)
2803 {
2804         struct perf_event *output_event;
2805         struct perf_mmap_data *data;
2806         unsigned long tail, offset, head;
2807         int have_lost;
2808         struct {
2809                 struct perf_event_header header;
2810                 u64                      id;
2811                 u64                      lost;
2812         } lost_event;
2813
2814         rcu_read_lock();
2815         /*
2816          * For inherited events we send all the output towards the parent.
2817          */
2818         if (event->parent)
2819                 event = event->parent;
2820
2821         output_event = rcu_dereference(event->output);
2822         if (output_event)
2823                 event = output_event;
2824
2825         data = rcu_dereference(event->data);
2826         if (!data)
2827                 goto out;
2828
2829         handle->data    = data;
2830         handle->event   = event;
2831         handle->nmi     = nmi;
2832         handle->sample  = sample;
2833
2834         if (!data->nr_pages)
2835                 goto fail;
2836
2837         have_lost = atomic_read(&data->lost);
2838         if (have_lost)
2839                 size += sizeof(lost_event);
2840
2841         perf_output_lock(handle);
2842
2843         do {
2844                 /*
2845                  * Userspace could choose to issue a mb() before updating the
2846                  * tail pointer. So that all reads will be completed before the
2847                  * write is issued.
2848                  */
2849                 tail = ACCESS_ONCE(data->user_page->data_tail);
2850                 smp_rmb();
2851                 offset = head = atomic_long_read(&data->head);
2852                 head += size;
2853                 if (unlikely(!perf_output_space(data, tail, offset, head)))
2854                         goto fail;
2855         } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2856
2857         handle->offset  = offset;
2858         handle->head    = head;
2859
2860         if (head - tail > data->watermark)
2861                 atomic_set(&data->wakeup, 1);
2862
2863         if (have_lost) {
2864                 lost_event.header.type = PERF_RECORD_LOST;
2865                 lost_event.header.misc = 0;
2866                 lost_event.header.size = sizeof(lost_event);
2867                 lost_event.id          = event->id;
2868                 lost_event.lost        = atomic_xchg(&data->lost, 0);
2869
2870                 perf_output_put(handle, lost_event);
2871         }
2872
2873         return 0;
2874
2875 fail:
2876         atomic_inc(&data->lost);
2877         perf_output_unlock(handle);
2878 out:
2879         rcu_read_unlock();
2880
2881         return -ENOSPC;
2882 }
2883
2884 void perf_output_end(struct perf_output_handle *handle)
2885 {
2886         struct perf_event *event = handle->event;
2887         struct perf_mmap_data *data = handle->data;
2888
2889         int wakeup_events = event->attr.wakeup_events;
2890
2891         if (handle->sample && wakeup_events) {
2892                 int events = atomic_inc_return(&data->events);
2893                 if (events >= wakeup_events) {
2894                         atomic_sub(wakeup_events, &data->events);
2895                         atomic_set(&data->wakeup, 1);
2896                 }
2897         }
2898
2899         perf_output_unlock(handle);
2900         rcu_read_unlock();
2901 }
2902
2903 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
2904 {
2905         /*
2906          * only top level events have the pid namespace they were created in
2907          */
2908         if (event->parent)
2909                 event = event->parent;
2910
2911         return task_tgid_nr_ns(p, event->ns);
2912 }
2913
2914 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
2915 {
2916         /*
2917          * only top level events have the pid namespace they were created in
2918          */
2919         if (event->parent)
2920                 event = event->parent;
2921
2922         return task_pid_nr_ns(p, event->ns);
2923 }
2924
2925 static void perf_output_read_one(struct perf_output_handle *handle,
2926                                  struct perf_event *event)
2927 {
2928         u64 read_format = event->attr.read_format;
2929         u64 values[4];
2930         int n = 0;
2931
2932         values[n++] = atomic64_read(&event->count);
2933         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2934                 values[n++] = event->total_time_enabled +
2935                         atomic64_read(&event->child_total_time_enabled);
2936         }
2937         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2938                 values[n++] = event->total_time_running +
2939                         atomic64_read(&event->child_total_time_running);
2940         }
2941         if (read_format & PERF_FORMAT_ID)
2942                 values[n++] = primary_event_id(event);
2943
2944         perf_output_copy(handle, values, n * sizeof(u64));
2945 }
2946
2947 /*
2948  * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
2949  */
2950 static void perf_output_read_group(struct perf_output_handle *handle,
2951                             struct perf_event *event)
2952 {
2953         struct perf_event *leader = event->group_leader, *sub;
2954         u64 read_format = event->attr.read_format;
2955         u64 values[5];
2956         int n = 0;
2957
2958         values[n++] = 1 + leader->nr_siblings;
2959
2960         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2961                 values[n++] = leader->total_time_enabled;
2962
2963         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2964                 values[n++] = leader->total_time_running;
2965
2966         if (leader != event)
2967                 leader->pmu->read(leader);
2968
2969         values[n++] = atomic64_read(&leader->count);
2970         if (read_format & PERF_FORMAT_ID)
2971                 values[n++] = primary_event_id(leader);
2972
2973         perf_output_copy(handle, values, n * sizeof(u64));
2974
2975         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2976                 n = 0;
2977
2978                 if (sub != event)
2979                         sub->pmu->read(sub);
2980
2981                 values[n++] = atomic64_read(&sub->count);
2982                 if (read_format & PERF_FORMAT_ID)
2983                         values[n++] = primary_event_id(sub);
2984
2985                 perf_output_copy(handle, values, n * sizeof(u64));
2986         }
2987 }
2988
2989 static void perf_output_read(struct perf_output_handle *handle,
2990                              struct perf_event *event)
2991 {
2992         if (event->attr.read_format & PERF_FORMAT_GROUP)
2993                 perf_output_read_group(handle, event);
2994         else
2995                 perf_output_read_one(handle, event);
2996 }
2997
2998 void perf_output_sample(struct perf_output_handle *handle,
2999                         struct perf_event_header *header,
3000                         struct perf_sample_data *data,
3001                         struct perf_event *event)
3002 {
3003         u64 sample_type = data->type;
3004
3005         perf_output_put(handle, *header);
3006
3007         if (sample_type & PERF_SAMPLE_IP)
3008                 perf_output_put(handle, data->ip);
3009
3010         if (sample_type & PERF_SAMPLE_TID)
3011                 perf_output_put(handle, data->tid_entry);
3012
3013         if (sample_type & PERF_SAMPLE_TIME)
3014                 perf_output_put(handle, data->time);
3015
3016         if (sample_type & PERF_SAMPLE_ADDR)
3017                 perf_output_put(handle, data->addr);
3018
3019         if (sample_type & PERF_SAMPLE_ID)
3020                 perf_output_put(handle, data->id);
3021
3022         if (sample_type & PERF_SAMPLE_STREAM_ID)
3023                 perf_output_put(handle, data->stream_id);
3024
3025         if (sample_type & PERF_SAMPLE_CPU)
3026                 perf_output_put(handle, data->cpu_entry);
3027
3028         if (sample_type & PERF_SAMPLE_PERIOD)
3029                 perf_output_put(handle, data->period);
3030
3031         if (sample_type & PERF_SAMPLE_READ)
3032                 perf_output_read(handle, event);
3033
3034         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3035                 if (data->callchain) {
3036                         int size = 1;
3037
3038                         if (data->callchain)
3039                                 size += data->callchain->nr;
3040
3041                         size *= sizeof(u64);
3042
3043                         perf_output_copy(handle, data->callchain, size);
3044                 } else {
3045                         u64 nr = 0;
3046                         perf_output_put(handle, nr);
3047                 }
3048         }
3049
3050         if (sample_type & PERF_SAMPLE_RAW) {
3051                 if (data->raw) {
3052                         perf_output_put(handle, data->raw->size);
3053                         perf_output_copy(handle, data->raw->data,
3054                                          data->raw->size);
3055                 } else {
3056                         struct {
3057                                 u32     size;
3058                                 u32     data;
3059                         } raw = {
3060                                 .size = sizeof(u32),
3061                                 .data = 0,
3062                         };
3063                         perf_output_put(handle, raw);
3064                 }
3065         }
3066 }
3067
3068 void perf_prepare_sample(struct perf_event_header *header,
3069                          struct perf_sample_data *data,
3070                          struct perf_event *event,
3071                          struct pt_regs *regs)
3072 {
3073         u64 sample_type = event->attr.sample_type;
3074
3075         data->type = sample_type;
3076
3077         header->type = PERF_RECORD_SAMPLE;
3078         header->size = sizeof(*header);
3079
3080         header->misc = 0;
3081         header->misc |= perf_misc_flags(regs);
3082
3083         if (sample_type & PERF_SAMPLE_IP) {
3084                 data->ip = perf_instruction_pointer(regs);
3085
3086                 header->size += sizeof(data->ip);
3087         }
3088
3089         if (sample_type & PERF_SAMPLE_TID) {
3090                 /* namespace issues */
3091                 data->tid_entry.pid = perf_event_pid(event, current);
3092                 data->tid_entry.tid = perf_event_tid(event, current);
3093
3094                 header->size += sizeof(data->tid_entry);
3095         }
3096
3097         if (sample_type & PERF_SAMPLE_TIME) {
3098                 data->time = perf_clock();
3099
3100                 header->size += sizeof(data->time);
3101         }
3102
3103         if (sample_type & PERF_SAMPLE_ADDR)
3104                 header->size += sizeof(data->addr);
3105
3106         if (sample_type & PERF_SAMPLE_ID) {
3107                 data->id = primary_event_id(event);
3108
3109                 header->size += sizeof(data->id);
3110         }
3111
3112         if (sample_type & PERF_SAMPLE_STREAM_ID) {
3113                 data->stream_id = event->id;
3114
3115                 header->size += sizeof(data->stream_id);
3116         }
3117
3118         if (sample_type & PERF_SAMPLE_CPU) {
3119                 data->cpu_entry.cpu             = raw_smp_processor_id();
3120                 data->cpu_entry.reserved        = 0;
3121
3122                 header->size += sizeof(data->cpu_entry);
3123         }
3124
3125         if (sample_type & PERF_SAMPLE_PERIOD)
3126                 header->size += sizeof(data->period);
3127
3128         if (sample_type & PERF_SAMPLE_READ)
3129                 header->size += perf_event_read_size(event);
3130
3131         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3132                 int size = 1;
3133
3134                 data->callchain = perf_callchain(regs);
3135
3136                 if (data->callchain)
3137                         size += data->callchain->nr;
3138
3139                 header->size += size * sizeof(u64);
3140         }
3141
3142         if (sample_type & PERF_SAMPLE_RAW) {
3143                 int size = sizeof(u32);
3144
3145                 if (data->raw)
3146                         size += data->raw->size;
3147                 else
3148                         size += sizeof(u32);
3149
3150                 WARN_ON_ONCE(size & (sizeof(u64)-1));
3151                 header->size += size;
3152         }
3153 }
3154
3155 static void perf_event_output(struct perf_event *event, int nmi,
3156                                 struct perf_sample_data *data,
3157                                 struct pt_regs *regs)
3158 {
3159         struct perf_output_handle handle;
3160         struct perf_event_header header;
3161
3162         perf_prepare_sample(&header, data, event, regs);
3163
3164         if (perf_output_begin(&handle, event, header.size, nmi, 1))
3165                 return;
3166
3167         perf_output_sample(&handle, &header, data, event);
3168
3169         perf_output_end(&handle);
3170 }
3171
3172 /*
3173  * read event_id
3174  */
3175
3176 struct perf_read_event {
3177         struct perf_event_header        header;
3178
3179         u32                             pid;
3180         u32                             tid;
3181 };
3182
3183 static void
3184 perf_event_read_event(struct perf_event *event,
3185                         struct task_struct *task)
3186 {
3187         struct perf_output_handle handle;
3188         struct perf_read_event read_event = {
3189                 .header = {
3190                         .type = PERF_RECORD_READ,
3191                         .misc = 0,
3192                         .size = sizeof(read_event) + perf_event_read_size(event),
3193                 },
3194                 .pid = perf_event_pid(event, task),
3195                 .tid = perf_event_tid(event, task),
3196         };
3197         int ret;
3198
3199         ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3200         if (ret)
3201                 return;
3202
3203         perf_output_put(&handle, read_event);
3204         perf_output_read(&handle, event);
3205
3206         perf_output_end(&handle);
3207 }
3208
3209 /*
3210  * task tracking -- fork/exit
3211  *
3212  * enabled by: attr.comm | attr.mmap | attr.task
3213  */
3214
3215 struct perf_task_event {
3216         struct task_struct              *task;
3217         struct perf_event_context       *task_ctx;
3218
3219         struct {
3220                 struct perf_event_header        header;
3221
3222                 u32                             pid;
3223                 u32                             ppid;
3224                 u32                             tid;
3225                 u32                             ptid;
3226                 u64                             time;
3227         } event_id;
3228 };
3229
3230 static void perf_event_task_output(struct perf_event *event,
3231                                      struct perf_task_event *task_event)
3232 {
3233         struct perf_output_handle handle;
3234         int size;
3235         struct task_struct *task = task_event->task;
3236         int ret;
3237
3238         size  = task_event->event_id.header.size;
3239         ret = perf_output_begin(&handle, event, size, 0, 0);
3240
3241         if (ret)
3242                 return;
3243
3244         task_event->event_id.pid = perf_event_pid(event, task);
3245         task_event->event_id.ppid = perf_event_pid(event, current);
3246
3247         task_event->event_id.tid = perf_event_tid(event, task);
3248         task_event->event_id.ptid = perf_event_tid(event, current);
3249
3250         task_event->event_id.time = perf_clock();
3251
3252         perf_output_put(&handle, task_event->event_id);
3253
3254         perf_output_end(&handle);
3255 }
3256
3257 static int perf_event_task_match(struct perf_event *event)
3258 {
3259         if (event->attr.comm || event->attr.mmap || event->attr.task)
3260                 return 1;
3261
3262         return 0;
3263 }
3264
3265 static void perf_event_task_ctx(struct perf_event_context *ctx,
3266                                   struct perf_task_event *task_event)
3267 {
3268         struct perf_event *event;
3269
3270         if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3271                 return;
3272
3273         rcu_read_lock();
3274         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3275                 if (perf_event_task_match(event))
3276                         perf_event_task_output(event, task_event);
3277         }
3278         rcu_read_unlock();
3279 }
3280
3281 static void perf_event_task_event(struct perf_task_event *task_event)
3282 {
3283         struct perf_cpu_context *cpuctx;
3284         struct perf_event_context *ctx = task_event->task_ctx;
3285
3286         cpuctx = &get_cpu_var(perf_cpu_context);
3287         perf_event_task_ctx(&cpuctx->ctx, task_event);
3288         put_cpu_var(perf_cpu_context);
3289
3290         rcu_read_lock();
3291         if (!ctx)
3292                 ctx = rcu_dereference(task_event->task->perf_event_ctxp);
3293         if (ctx)
3294                 perf_event_task_ctx(ctx, task_event);
3295         rcu_read_unlock();
3296 }
3297
3298 static void perf_event_task(struct task_struct *task,
3299                               struct perf_event_context *task_ctx,
3300                               int new)
3301 {
3302         struct perf_task_event task_event;
3303
3304         if (!atomic_read(&nr_comm_events) &&
3305             !atomic_read(&nr_mmap_events) &&
3306             !atomic_read(&nr_task_events))
3307                 return;
3308
3309         task_event = (struct perf_task_event){
3310                 .task     = task,
3311                 .task_ctx = task_ctx,
3312                 .event_id    = {
3313                         .header = {
3314                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3315                                 .misc = 0,
3316                                 .size = sizeof(task_event.event_id),
3317                         },
3318                         /* .pid  */
3319                         /* .ppid */
3320                         /* .tid  */
3321                         /* .ptid */
3322                 },
3323         };
3324
3325         perf_event_task_event(&task_event);
3326 }
3327
3328 void perf_event_fork(struct task_struct *task)
3329 {
3330         perf_event_task(task, NULL, 1);
3331 }
3332
3333 /*
3334  * comm tracking
3335  */
3336
3337 struct perf_comm_event {
3338         struct task_struct      *task;
3339         char                    *comm;
3340         int                     comm_size;
3341
3342         struct {
3343                 struct perf_event_header        header;
3344
3345                 u32                             pid;
3346                 u32                             tid;
3347         } event_id;
3348 };
3349
3350 static void perf_event_comm_output(struct perf_event *event,
3351                                      struct perf_comm_event *comm_event)
3352 {
3353         struct perf_output_handle handle;
3354         int size = comm_event->event_id.header.size;
3355         int ret = perf_output_begin(&handle, event, size, 0, 0);
3356
3357         if (ret)
3358                 return;
3359
3360         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3361         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3362
3363         perf_output_put(&handle, comm_event->event_id);
3364         perf_output_copy(&handle, comm_event->comm,
3365                                    comm_event->comm_size);
3366         perf_output_end(&handle);
3367 }
3368
3369 static int perf_event_comm_match(struct perf_event *event)
3370 {
3371         if (event->attr.comm)
3372                 return 1;
3373
3374         return 0;
3375 }
3376
3377 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3378                                   struct perf_comm_event *comm_event)
3379 {
3380         struct perf_event *event;
3381
3382         if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3383                 return;
3384
3385         rcu_read_lock();
3386         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3387                 if (perf_event_comm_match(event))
3388                         perf_event_comm_output(event, comm_event);
3389         }
3390         rcu_read_unlock();
3391 }
3392
3393 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3394 {
3395         struct perf_cpu_context *cpuctx;
3396         struct perf_event_context *ctx;
3397         unsigned int size;
3398         char comm[TASK_COMM_LEN];
3399
3400         memset(comm, 0, sizeof(comm));
3401         strncpy(comm, comm_event->task->comm, sizeof(comm));
3402         size = ALIGN(strlen(comm)+1, sizeof(u64));
3403
3404         comm_event->comm = comm;
3405         comm_event->comm_size = size;
3406
3407         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3408
3409         cpuctx = &get_cpu_var(perf_cpu_context);
3410         perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3411         put_cpu_var(perf_cpu_context);
3412
3413         rcu_read_lock();
3414         /*
3415          * doesn't really matter which of the child contexts the
3416          * events ends up in.
3417          */
3418         ctx = rcu_dereference(current->perf_event_ctxp);
3419         if (ctx)
3420                 perf_event_comm_ctx(ctx, comm_event);
3421         rcu_read_unlock();
3422 }
3423
3424 void perf_event_comm(struct task_struct *task)
3425 {
3426         struct perf_comm_event comm_event;
3427
3428         if (task->perf_event_ctxp)
3429                 perf_event_enable_on_exec(task);
3430
3431         if (!atomic_read(&nr_comm_events))
3432                 return;
3433
3434         comm_event = (struct perf_comm_event){
3435                 .task   = task,
3436                 /* .comm      */
3437                 /* .comm_size */
3438                 .event_id  = {
3439                         .header = {
3440                                 .type = PERF_RECORD_COMM,
3441                                 .misc = 0,
3442                                 /* .size */
3443                         },
3444                         /* .pid */
3445                         /* .tid */
3446                 },
3447         };
3448
3449         perf_event_comm_event(&comm_event);
3450 }
3451
3452 /*
3453  * mmap tracking
3454  */
3455
3456 struct perf_mmap_event {
3457         struct vm_area_struct   *vma;
3458
3459         const char              *file_name;
3460         int                     file_size;
3461
3462         struct {
3463                 struct perf_event_header        header;
3464
3465                 u32                             pid;
3466                 u32                             tid;
3467                 u64                             start;
3468                 u64                             len;
3469                 u64                             pgoff;
3470         } event_id;
3471 };
3472
3473 static void perf_event_mmap_output(struct perf_event *event,
3474                                      struct perf_mmap_event *mmap_event)
3475 {
3476         struct perf_output_handle handle;
3477         int size = mmap_event->event_id.header.size;
3478         int ret = perf_output_begin(&handle, event, size, 0, 0);
3479
3480         if (ret)
3481                 return;
3482
3483         mmap_event->event_id.pid = perf_event_pid(event, current);
3484         mmap_event->event_id.tid = perf_event_tid(event, current);
3485
3486         perf_output_put(&handle, mmap_event->event_id);
3487         perf_output_copy(&handle, mmap_event->file_name,
3488                                    mmap_event->file_size);
3489         perf_output_end(&handle);
3490 }
3491
3492 static int perf_event_mmap_match(struct perf_event *event,
3493                                    struct perf_mmap_event *mmap_event)
3494 {
3495         if (event->attr.mmap)
3496                 return 1;
3497
3498         return 0;
3499 }
3500
3501 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3502                                   struct perf_mmap_event *mmap_event)
3503 {
3504         struct perf_event *event;
3505
3506         if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3507                 return;
3508
3509         rcu_read_lock();
3510         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3511                 if (perf_event_mmap_match(event, mmap_event))
3512                         perf_event_mmap_output(event, mmap_event);
3513         }
3514         rcu_read_unlock();
3515 }
3516
3517 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3518 {
3519         struct perf_cpu_context *cpuctx;
3520         struct perf_event_context *ctx;
3521         struct vm_area_struct *vma = mmap_event->vma;
3522         struct file *file = vma->vm_file;
3523         unsigned int size;
3524         char tmp[16];
3525         char *buf = NULL;
3526         const char *name;
3527
3528         memset(tmp, 0, sizeof(tmp));
3529
3530         if (file) {
3531                 /*
3532                  * d_path works from the end of the buffer backwards, so we
3533                  * need to add enough zero bytes after the string to handle
3534                  * the 64bit alignment we do later.
3535                  */
3536                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3537                 if (!buf) {
3538                         name = strncpy(tmp, "//enomem", sizeof(tmp));
3539                         goto got_name;
3540                 }
3541                 name = d_path(&file->f_path, buf, PATH_MAX);
3542                 if (IS_ERR(name)) {
3543                         name = strncpy(tmp, "//toolong", sizeof(tmp));
3544                         goto got_name;
3545                 }
3546         } else {
3547                 if (arch_vma_name(mmap_event->vma)) {
3548                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3549                                        sizeof(tmp));
3550                         goto got_name;
3551                 }
3552
3553                 if (!vma->vm_mm) {
3554                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
3555                         goto got_name;
3556                 }
3557
3558                 name = strncpy(tmp, "//anon", sizeof(tmp));
3559                 goto got_name;
3560         }
3561
3562 got_name:
3563         size = ALIGN(strlen(name)+1, sizeof(u64));
3564
3565         mmap_event->file_name = name;
3566         mmap_event->file_size = size;
3567
3568         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3569
3570         cpuctx = &get_cpu_var(perf_cpu_context);
3571         perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3572         put_cpu_var(perf_cpu_context);
3573
3574         rcu_read_lock();
3575         /*
3576          * doesn't really matter which of the child contexts the
3577          * events ends up in.
3578          */
3579         ctx = rcu_dereference(current->perf_event_ctxp);
3580         if (ctx)
3581                 perf_event_mmap_ctx(ctx, mmap_event);
3582         rcu_read_unlock();
3583
3584         kfree(buf);
3585 }
3586
3587 void __perf_event_mmap(struct vm_area_struct *vma)
3588 {
3589         struct perf_mmap_event mmap_event;
3590
3591         if (!atomic_read(&nr_mmap_events))
3592                 return;
3593
3594         mmap_event = (struct perf_mmap_event){
3595                 .vma    = vma,
3596                 /* .file_name */
3597                 /* .file_size */
3598                 .event_id  = {
3599                         .header = {
3600                                 .type = PERF_RECORD_MMAP,
3601                                 .misc = 0,
3602                                 /* .size */
3603                         },
3604                         /* .pid */
3605                         /* .tid */
3606                         .start  = vma->vm_start,
3607                         .len    = vma->vm_end - vma->vm_start,
3608                         .pgoff  = vma->vm_pgoff,
3609                 },
3610         };
3611
3612         perf_event_mmap_event(&mmap_event);
3613 }
3614
3615 /*
3616  * IRQ throttle logging
3617  */
3618
3619 static void perf_log_throttle(struct perf_event *event, int enable)
3620 {
3621         struct perf_output_handle handle;
3622         int ret;
3623
3624         struct {
3625                 struct perf_event_header        header;
3626                 u64                             time;
3627                 u64                             id;
3628                 u64                             stream_id;
3629         } throttle_event = {
3630                 .header = {
3631                         .type = PERF_RECORD_THROTTLE,
3632                         .misc = 0,
3633                         .size = sizeof(throttle_event),
3634                 },
3635                 .time           = perf_clock(),
3636                 .id             = primary_event_id(event),
3637                 .stream_id      = event->id,
3638         };
3639
3640         if (enable)
3641                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3642
3643         ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3644         if (ret)
3645                 return;
3646
3647         perf_output_put(&handle, throttle_event);
3648         perf_output_end(&handle);
3649 }
3650
3651 /*
3652  * Generic event overflow handling, sampling.
3653  */
3654
3655 static int __perf_event_overflow(struct perf_event *event, int nmi,
3656                                    int throttle, struct perf_sample_data *data,
3657                                    struct pt_regs *regs)
3658 {
3659         int events = atomic_read(&event->event_limit);
3660         struct hw_perf_event *hwc = &event->hw;
3661         int ret = 0;
3662
3663         throttle = (throttle && event->pmu->unthrottle != NULL);
3664
3665         if (!throttle) {
3666                 hwc->interrupts++;
3667         } else {
3668                 if (hwc->interrupts != MAX_INTERRUPTS) {
3669                         hwc->interrupts++;
3670                         if (HZ * hwc->interrupts >
3671                                         (u64)sysctl_perf_event_sample_rate) {
3672                                 hwc->interrupts = MAX_INTERRUPTS;
3673                                 perf_log_throttle(event, 0);
3674                                 ret = 1;
3675                         }
3676                 } else {
3677                         /*
3678                          * Keep re-disabling events even though on the previous
3679                          * pass we disabled it - just in case we raced with a
3680                          * sched-in and the event got enabled again:
3681                          */
3682                         ret = 1;
3683                 }
3684         }
3685
3686         if (event->attr.freq) {
3687                 u64 now = perf_clock();
3688                 s64 delta = now - hwc->freq_stamp;
3689
3690                 hwc->freq_stamp = now;
3691
3692                 if (delta > 0 && delta < TICK_NSEC)
3693                         perf_adjust_period(event, NSEC_PER_SEC / (int)delta);
3694         }
3695
3696         /*
3697          * XXX event_limit might not quite work as expected on inherited
3698          * events
3699          */
3700
3701         event->pending_kill = POLL_IN;
3702         if (events && atomic_dec_and_test(&event->event_limit)) {
3703                 ret = 1;
3704                 event->pending_kill = POLL_HUP;
3705                 if (nmi) {
3706                         event->pending_disable = 1;
3707                         perf_pending_queue(&event->pending,
3708                                            perf_pending_event);
3709                 } else
3710                         perf_event_disable(event);
3711         }
3712
3713         if (event->overflow_handler)
3714                 event->overflow_handler(event, nmi, data, regs);
3715         else
3716                 perf_event_output(event, nmi, data, regs);
3717
3718         return ret;
3719 }
3720
3721 int perf_event_overflow(struct perf_event *event, int nmi,
3722                           struct perf_sample_data *data,
3723                           struct pt_regs *regs)
3724 {
3725         return __perf_event_overflow(event, nmi, 1, data, regs);
3726 }
3727
3728 /*
3729  * Generic software event infrastructure
3730  */
3731
3732 /*
3733  * We directly increment event->count and keep a second value in
3734  * event->hw.period_left to count intervals. This period event
3735  * is kept in the range [-sample_period, 0] so that we can use the
3736  * sign as trigger.
3737  */
3738
3739 static u64 perf_swevent_set_period(struct perf_event *event)
3740 {
3741         struct hw_perf_event *hwc = &event->hw;
3742         u64 period = hwc->last_period;
3743         u64 nr, offset;
3744         s64 old, val;
3745
3746         hwc->last_period = hwc->sample_period;
3747
3748 again:
3749         old = val = atomic64_read(&hwc->period_left);
3750         if (val < 0)
3751                 return 0;
3752
3753         nr = div64_u64(period + val, period);
3754         offset = nr * period;
3755         val -= offset;
3756         if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3757                 goto again;
3758
3759         return nr;
3760 }
3761
3762 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3763                                     int nmi, struct perf_sample_data *data,
3764                                     struct pt_regs *regs)
3765 {
3766         struct hw_perf_event *hwc = &event->hw;
3767         int throttle = 0;
3768
3769         data->period = event->hw.last_period;
3770         if (!overflow)
3771                 overflow = perf_swevent_set_period(event);
3772
3773         if (hwc->interrupts == MAX_INTERRUPTS)
3774                 return;
3775
3776         for (; overflow; overflow--) {
3777                 if (__perf_event_overflow(event, nmi, throttle,
3778                                             data, regs)) {
3779                         /*
3780                          * We inhibit the overflow from happening when
3781                          * hwc->interrupts == MAX_INTERRUPTS.
3782                          */
3783                         break;
3784                 }
3785                 throttle = 1;
3786         }
3787 }
3788
3789 static void perf_swevent_unthrottle(struct perf_event *event)
3790 {
3791         /*
3792          * Nothing to do, we already reset hwc->interrupts.
3793          */
3794 }
3795
3796 static void perf_swevent_add(struct perf_event *event, u64 nr,
3797                                int nmi, struct perf_sample_data *data,
3798                                struct pt_regs *regs)
3799 {
3800         struct hw_perf_event *hwc = &event->hw;
3801
3802         atomic64_add(nr, &event->count);
3803
3804         if (!regs)
3805                 return;
3806
3807         if (!hwc->sample_period)
3808                 return;
3809
3810         if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3811                 return perf_swevent_overflow(event, 1, nmi, data, regs);
3812
3813         if (atomic64_add_negative(nr, &hwc->period_left))
3814                 return;
3815
3816         perf_swevent_overflow(event, 0, nmi, data, regs);
3817 }
3818
3819 static int perf_swevent_is_counting(struct perf_event *event)
3820 {
3821         /*
3822          * The event is active, we're good!
3823          */
3824         if (event->state == PERF_EVENT_STATE_ACTIVE)
3825                 return 1;
3826
3827         /*
3828          * The event is off/error, not counting.
3829          */
3830         if (event->state != PERF_EVENT_STATE_INACTIVE)
3831                 return 0;
3832
3833         /*
3834          * The event is inactive, if the context is active
3835          * we're part of a group that didn't make it on the 'pmu',
3836          * not counting.
3837          */
3838         if (event->ctx->is_active)
3839                 return 0;
3840
3841         /*
3842          * We're inactive and the context is too, this means the
3843          * task is scheduled out, we're counting events that happen
3844          * to us, like migration events.
3845          */
3846         return 1;
3847 }
3848
3849 static int perf_tp_event_match(struct perf_event *event,
3850                                 struct perf_sample_data *data);
3851
3852 static int perf_swevent_match(struct perf_event *event,
3853                                 enum perf_type_id type,
3854                                 u32 event_id,
3855                                 struct perf_sample_data *data,
3856                                 struct pt_regs *regs)
3857 {
3858         if (!perf_swevent_is_counting(event))
3859                 return 0;
3860
3861         if (event->attr.type != type)
3862                 return 0;
3863         if (event->attr.config != event_id)
3864                 return 0;
3865
3866         if (regs) {
3867                 if (event->attr.exclude_user && user_mode(regs))
3868                         return 0;
3869
3870                 if (event->attr.exclude_kernel && !user_mode(regs))
3871                         return 0;
3872         }
3873
3874         if (event->attr.type == PERF_TYPE_TRACEPOINT &&
3875             !perf_tp_event_match(event, data))
3876                 return 0;
3877
3878         return 1;
3879 }
3880
3881 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
3882                                      enum perf_type_id type,
3883                                      u32 event_id, u64 nr, int nmi,
3884                                      struct perf_sample_data *data,
3885                                      struct pt_regs *regs)
3886 {
3887         struct perf_event *event;
3888
3889         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3890                 if (perf_swevent_match(event, type, event_id, data, regs))
3891                         perf_swevent_add(event, nr, nmi, data, regs);
3892         }
3893 }
3894
3895 static int *perf_swevent_recursion_context(struct perf_cpu_context *cpuctx)
3896 {
3897         if (in_nmi())
3898                 return &cpuctx->recursion[3];
3899
3900         if (in_irq())
3901                 return &cpuctx->recursion[2];
3902
3903         if (in_softirq())
3904                 return &cpuctx->recursion[1];
3905
3906         return &cpuctx->recursion[0];
3907 }
3908
3909 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
3910                                     u64 nr, int nmi,
3911                                     struct perf_sample_data *data,
3912                                     struct pt_regs *regs)
3913 {
3914         struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3915         int *recursion = perf_swevent_recursion_context(cpuctx);
3916         struct perf_event_context *ctx;
3917
3918         if (*recursion)
3919                 goto out;
3920
3921         (*recursion)++;
3922         barrier();
3923
3924         rcu_read_lock();
3925         perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
3926                                  nr, nmi, data, regs);
3927         /*
3928          * doesn't really matter which of the child contexts the
3929          * events ends up in.
3930          */
3931         ctx = rcu_dereference(current->perf_event_ctxp);
3932         if (ctx)
3933                 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
3934         rcu_read_unlock();
3935
3936         barrier();
3937         (*recursion)--;
3938
3939 out:
3940         put_cpu_var(perf_cpu_context);
3941 }
3942
3943 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
3944                             struct pt_regs *regs, u64 addr)
3945 {
3946         struct perf_sample_data data = {
3947                 .addr = addr,
3948         };
3949
3950         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi,
3951                                 &data, regs);
3952 }
3953
3954 static void perf_swevent_read(struct perf_event *event)
3955 {
3956 }
3957
3958 static int perf_swevent_enable(struct perf_event *event)
3959 {
3960         struct hw_perf_event *hwc = &event->hw;
3961
3962         if (hwc->sample_period) {
3963                 hwc->last_period = hwc->sample_period;
3964                 perf_swevent_set_period(event);
3965         }
3966         return 0;
3967 }
3968
3969 static void perf_swevent_disable(struct perf_event *event)
3970 {
3971 }
3972
3973 static const struct pmu perf_ops_generic = {
3974         .enable         = perf_swevent_enable,
3975         .disable        = perf_swevent_disable,
3976         .read           = perf_swevent_read,
3977         .unthrottle     = perf_swevent_unthrottle,
3978 };
3979
3980 /*
3981  * hrtimer based swevent callback
3982  */
3983
3984 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
3985 {
3986         enum hrtimer_restart ret = HRTIMER_RESTART;
3987         struct perf_sample_data data;
3988         struct pt_regs *regs;
3989         struct perf_event *event;
3990         u64 period;
3991
3992         event   = container_of(hrtimer, struct perf_event, hw.hrtimer);
3993         event->pmu->read(event);
3994
3995         data.addr = 0;
3996         regs = get_irq_regs();
3997         /*
3998          * In case we exclude kernel IPs or are somehow not in interrupt
3999          * context, provide the next best thing, the user IP.
4000          */
4001         if ((event->attr.exclude_kernel || !regs) &&
4002                         !event->attr.exclude_user)
4003                 regs = task_pt_regs(current);
4004
4005         if (regs) {
4006                 if (!(event->attr.exclude_idle && current->pid == 0))
4007                         if (perf_event_overflow(event, 0, &data, regs))
4008                                 ret = HRTIMER_NORESTART;
4009         }
4010
4011         period = max_t(u64, 10000, event->hw.sample_period);
4012         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4013
4014         return ret;
4015 }
4016
4017 static void perf_swevent_start_hrtimer(struct perf_event *event)
4018 {
4019         struct hw_perf_event *hwc = &event->hw;
4020
4021         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4022         hwc->hrtimer.function = perf_swevent_hrtimer;
4023         if (hwc->sample_period) {
4024                 u64 period;
4025
4026                 if (hwc->remaining) {
4027                         if (hwc->remaining < 0)
4028                                 period = 10000;
4029                         else
4030                                 period = hwc->remaining;
4031                         hwc->remaining = 0;
4032                 } else {
4033                         period = max_t(u64, 10000, hwc->sample_period);
4034                 }
4035                 __hrtimer_start_range_ns(&hwc->hrtimer,
4036                                 ns_to_ktime(period), 0,
4037                                 HRTIMER_MODE_REL, 0);
4038         }
4039 }
4040
4041 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4042 {
4043         struct hw_perf_event *hwc = &event->hw;
4044
4045         if (hwc->sample_period) {
4046                 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4047                 hwc->remaining = ktime_to_ns(remaining);
4048
4049                 hrtimer_cancel(&hwc->hrtimer);
4050         }
4051 }
4052
4053 /*
4054  * Software event: cpu wall time clock
4055  */
4056
4057 static void cpu_clock_perf_event_update(struct perf_event *event)
4058 {
4059         int cpu = raw_smp_processor_id();
4060         s64 prev;
4061         u64 now;
4062
4063         now = cpu_clock(cpu);
4064         prev = atomic64_read(&event->hw.prev_count);
4065         atomic64_set(&event->hw.prev_count, now);
4066         atomic64_add(now - prev, &event->count);
4067 }
4068
4069 static int cpu_clock_perf_event_enable(struct perf_event *event)
4070 {
4071         struct hw_perf_event *hwc = &event->hw;
4072         int cpu = raw_smp_processor_id();
4073
4074         atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4075         perf_swevent_start_hrtimer(event);
4076
4077         return 0;
4078 }
4079
4080 static void cpu_clock_perf_event_disable(struct perf_event *event)
4081 {
4082         perf_swevent_cancel_hrtimer(event);
4083         cpu_clock_perf_event_update(event);
4084 }
4085
4086 static void cpu_clock_perf_event_read(struct perf_event *event)
4087 {
4088         cpu_clock_perf_event_update(event);
4089 }
4090
4091 static const struct pmu perf_ops_cpu_clock = {
4092         .enable         = cpu_clock_perf_event_enable,
4093         .disable        = cpu_clock_perf_event_disable,
4094         .read           = cpu_clock_perf_event_read,
4095 };
4096
4097 /*
4098  * Software event: task time clock
4099  */
4100
4101 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4102 {
4103         u64 prev;
4104         s64 delta;
4105
4106         prev = atomic64_xchg(&event->hw.prev_count, now);
4107         delta = now - prev;
4108         atomic64_add(delta, &event->count);
4109 }
4110
4111 static int task_clock_perf_event_enable(struct perf_event *event)
4112 {
4113         struct hw_perf_event *hwc = &event->hw;
4114         u64 now;
4115
4116         now = event->ctx->time;
4117
4118         atomic64_set(&hwc->prev_count, now);
4119
4120         perf_swevent_start_hrtimer(event);
4121
4122         return 0;
4123 }
4124
4125 static void task_clock_perf_event_disable(struct perf_event *event)
4126 {
4127         perf_swevent_cancel_hrtimer(event);
4128         task_clock_perf_event_update(event, event->ctx->time);
4129
4130 }
4131
4132 static void task_clock_perf_event_read(struct perf_event *event)
4133 {
4134         u64 time;
4135
4136         if (!in_nmi()) {
4137                 update_context_time(event->ctx);
4138                 time = event->ctx->time;
4139         } else {
4140                 u64 now = perf_clock();
4141                 u64 delta = now - event->ctx->timestamp;
4142                 time = event->ctx->time + delta;
4143         }
4144
4145         task_clock_perf_event_update(event, time);
4146 }
4147
4148 static const struct pmu perf_ops_task_clock = {
4149         .enable         = task_clock_perf_event_enable,
4150         .disable        = task_clock_perf_event_disable,
4151         .read           = task_clock_perf_event_read,
4152 };
4153
4154 #ifdef CONFIG_EVENT_PROFILE
4155
4156 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4157                           int entry_size)
4158 {
4159         struct perf_raw_record raw = {
4160                 .size = entry_size,
4161                 .data = record,
4162         };
4163
4164         struct perf_sample_data data = {
4165                 .addr = addr,
4166                 .raw = &raw,
4167         };
4168
4169         struct pt_regs *regs = get_irq_regs();
4170
4171         if (!regs)
4172                 regs = task_pt_regs(current);
4173
4174         do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4175                                 &data, regs);
4176 }
4177 EXPORT_SYMBOL_GPL(perf_tp_event);
4178
4179 static int perf_tp_event_match(struct perf_event *event,
4180                                 struct perf_sample_data *data)
4181 {
4182         void *record = data->raw->data;
4183
4184         if (likely(!event->filter) || filter_match_preds(event->filter, record))
4185                 return 1;
4186         return 0;
4187 }
4188
4189 static void tp_perf_event_destroy(struct perf_event *event)
4190 {
4191         ftrace_profile_disable(event->attr.config);
4192 }
4193
4194 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4195 {
4196         /*
4197          * Raw tracepoint data is a severe data leak, only allow root to
4198          * have these.
4199          */
4200         if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4201                         perf_paranoid_tracepoint_raw() &&
4202                         !capable(CAP_SYS_ADMIN))
4203                 return ERR_PTR(-EPERM);
4204
4205         if (ftrace_profile_enable(event->attr.config))
4206                 return NULL;
4207
4208         event->destroy = tp_perf_event_destroy;
4209
4210         return &perf_ops_generic;
4211 }
4212
4213 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4214 {
4215         char *filter_str;
4216         int ret;
4217
4218         if (event->attr.type != PERF_TYPE_TRACEPOINT)
4219                 return -EINVAL;
4220
4221         filter_str = strndup_user(arg, PAGE_SIZE);
4222         if (IS_ERR(filter_str))
4223                 return PTR_ERR(filter_str);
4224
4225         ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4226
4227         kfree(filter_str);
4228         return ret;
4229 }
4230
4231 static void perf_event_free_filter(struct perf_event *event)
4232 {
4233         ftrace_profile_free_filter(event);
4234 }
4235
4236 #else
4237
4238 static int perf_tp_event_match(struct perf_event *event,
4239                                 struct perf_sample_data *data)
4240 {
4241         return 1;
4242 }
4243
4244 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4245 {
4246         return NULL;
4247 }
4248
4249 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4250 {
4251         return -ENOENT;
4252 }
4253
4254 static void perf_event_free_filter(struct perf_event *event)
4255 {
4256 }
4257
4258 #endif /* CONFIG_EVENT_PROFILE */
4259
4260 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4261 static void bp_perf_event_destroy(struct perf_event *event)
4262 {
4263         release_bp_slot(event);
4264 }
4265
4266 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4267 {
4268         int err;
4269         /*
4270          * The breakpoint is already filled if we haven't created the counter
4271          * through perf syscall
4272          * FIXME: manage to get trigerred to NULL if it comes from syscalls
4273          */
4274         if (!bp->callback)
4275                 err = register_perf_hw_breakpoint(bp);
4276         else
4277                 err = __register_perf_hw_breakpoint(bp);
4278         if (err)
4279                 return ERR_PTR(err);
4280
4281         bp->destroy = bp_perf_event_destroy;
4282
4283         return &perf_ops_bp;
4284 }
4285
4286 void perf_bp_event(struct perf_event *bp, void *regs)
4287 {
4288         /* TODO */
4289 }
4290 #else
4291 static void bp_perf_event_destroy(struct perf_event *event)
4292 {
4293 }
4294
4295 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4296 {
4297         return NULL;
4298 }
4299
4300 void perf_bp_event(struct perf_event *bp, void *regs)
4301 {
4302 }
4303 #endif
4304
4305 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4306
4307 static void sw_perf_event_destroy(struct perf_event *event)
4308 {
4309         u64 event_id = event->attr.config;
4310
4311         WARN_ON(event->parent);
4312
4313         atomic_dec(&perf_swevent_enabled[event_id]);
4314 }
4315
4316 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4317 {
4318         const struct pmu *pmu = NULL;
4319         u64 event_id = event->attr.config;
4320
4321         /*
4322          * Software events (currently) can't in general distinguish
4323          * between user, kernel and hypervisor events.
4324          * However, context switches and cpu migrations are considered
4325          * to be kernel events, and page faults are never hypervisor
4326          * events.
4327          */
4328         switch (event_id) {
4329         case PERF_COUNT_SW_CPU_CLOCK:
4330                 pmu = &perf_ops_cpu_clock;
4331
4332                 break;
4333         case PERF_COUNT_SW_TASK_CLOCK:
4334                 /*
4335                  * If the user instantiates this as a per-cpu event,
4336                  * use the cpu_clock event instead.
4337                  */
4338                 if (event->ctx->task)
4339                         pmu = &perf_ops_task_clock;
4340                 else
4341                         pmu = &perf_ops_cpu_clock;
4342
4343                 break;
4344         case PERF_COUNT_SW_PAGE_FAULTS:
4345         case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4346         case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4347         case PERF_COUNT_SW_CONTEXT_SWITCHES:
4348         case PERF_COUNT_SW_CPU_MIGRATIONS:
4349         case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4350         case PERF_COUNT_SW_EMULATION_FAULTS:
4351                 if (!event->parent) {
4352                         atomic_inc(&perf_swevent_enabled[event_id]);
4353                         event->destroy = sw_perf_event_destroy;
4354                 }
4355                 pmu = &perf_ops_generic;
4356                 break;
4357         }
4358
4359         return pmu;
4360 }
4361
4362 /*
4363  * Allocate and initialize a event structure
4364  */
4365 static struct perf_event *
4366 perf_event_alloc(struct perf_event_attr *attr,
4367                    int cpu,
4368                    struct perf_event_context *ctx,
4369                    struct perf_event *group_leader,
4370                    struct perf_event *parent_event,
4371                    perf_callback_t callback,
4372                    gfp_t gfpflags)
4373 {
4374         const struct pmu *pmu;
4375         struct perf_event *event;
4376         struct hw_perf_event *hwc;
4377         long err;
4378
4379         event = kzalloc(sizeof(*event), gfpflags);
4380         if (!event)
4381                 return ERR_PTR(-ENOMEM);
4382
4383         /*
4384          * Single events are their own group leaders, with an
4385          * empty sibling list:
4386          */
4387         if (!group_leader)
4388                 group_leader = event;
4389
4390         mutex_init(&event->child_mutex);
4391         INIT_LIST_HEAD(&event->child_list);
4392
4393         INIT_LIST_HEAD(&event->group_entry);
4394         INIT_LIST_HEAD(&event->event_entry);
4395         INIT_LIST_HEAD(&event->sibling_list);
4396         init_waitqueue_head(&event->waitq);
4397
4398         mutex_init(&event->mmap_mutex);
4399
4400         event->cpu              = cpu;
4401         event->attr             = *attr;
4402         event->group_leader     = group_leader;
4403         event->pmu              = NULL;
4404         event->ctx              = ctx;
4405         event->oncpu            = -1;
4406
4407         event->parent           = parent_event;
4408
4409         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
4410         event->id               = atomic64_inc_return(&perf_event_id);
4411
4412         event->state            = PERF_EVENT_STATE_INACTIVE;
4413
4414         if (!callback && parent_event)
4415                 callback = parent_event->callback;
4416         
4417         event->callback = callback;
4418
4419         if (attr->disabled)
4420                 event->state = PERF_EVENT_STATE_OFF;
4421
4422         pmu = NULL;
4423
4424         hwc = &event->hw;
4425         hwc->sample_period = attr->sample_period;
4426         if (attr->freq && attr->sample_freq)
4427                 hwc->sample_period = 1;
4428         hwc->last_period = hwc->sample_period;
4429
4430         atomic64_set(&hwc->period_left, hwc->sample_period);
4431
4432         /*
4433          * we currently do not support PERF_FORMAT_GROUP on inherited events
4434          */
4435         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4436                 goto done;
4437
4438         switch (attr->type) {
4439         case PERF_TYPE_RAW:
4440         case PERF_TYPE_HARDWARE:
4441         case PERF_TYPE_HW_CACHE:
4442                 pmu = hw_perf_event_init(event);
4443                 break;
4444
4445         case PERF_TYPE_SOFTWARE:
4446                 pmu = sw_perf_event_init(event);
4447                 break;
4448
4449         case PERF_TYPE_TRACEPOINT:
4450                 pmu = tp_perf_event_init(event);
4451                 break;
4452
4453         case PERF_TYPE_BREAKPOINT:
4454                 pmu = bp_perf_event_init(event);
4455                 break;
4456
4457
4458         default:
4459                 break;
4460         }
4461 done:
4462         err = 0;
4463         if (!pmu)
4464                 err = -EINVAL;
4465         else if (IS_ERR(pmu))
4466                 err = PTR_ERR(pmu);
4467
4468         if (err) {
4469                 if (event->ns)
4470                         put_pid_ns(event->ns);
4471                 kfree(event);
4472                 return ERR_PTR(err);
4473         }
4474
4475         event->pmu = pmu;
4476
4477         if (!event->parent) {
4478                 atomic_inc(&nr_events);
4479                 if (event->attr.mmap)
4480                         atomic_inc(&nr_mmap_events);
4481                 if (event->attr.comm)
4482                         atomic_inc(&nr_comm_events);
4483                 if (event->attr.task)
4484                         atomic_inc(&nr_task_events);
4485         }
4486
4487         return event;
4488 }
4489
4490 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4491                           struct perf_event_attr *attr)
4492 {
4493         u32 size;
4494         int ret;
4495
4496         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4497                 return -EFAULT;
4498
4499         /*
4500          * zero the full structure, so that a short copy will be nice.
4501          */
4502         memset(attr, 0, sizeof(*attr));
4503
4504         ret = get_user(size, &uattr->size);
4505         if (ret)
4506                 return ret;
4507
4508         if (size > PAGE_SIZE)   /* silly large */
4509                 goto err_size;
4510
4511         if (!size)              /* abi compat */
4512                 size = PERF_ATTR_SIZE_VER0;
4513
4514         if (size < PERF_ATTR_SIZE_VER0)
4515                 goto err_size;
4516
4517         /*
4518          * If we're handed a bigger struct than we know of,
4519          * ensure all the unknown bits are 0 - i.e. new
4520          * user-space does not rely on any kernel feature
4521          * extensions we dont know about yet.
4522          */
4523         if (size > sizeof(*attr)) {
4524                 unsigned char __user *addr;
4525                 unsigned char __user *end;
4526                 unsigned char val;
4527
4528                 addr = (void __user *)uattr + sizeof(*attr);
4529                 end  = (void __user *)uattr + size;
4530
4531                 for (; addr < end; addr++) {
4532                         ret = get_user(val, addr);
4533                         if (ret)
4534                                 return ret;
4535                         if (val)
4536                                 goto err_size;
4537                 }
4538                 size = sizeof(*attr);
4539         }
4540
4541         ret = copy_from_user(attr, uattr, size);
4542         if (ret)
4543                 return -EFAULT;
4544
4545         /*
4546          * If the type exists, the corresponding creation will verify
4547          * the attr->config.
4548          */
4549         if (attr->type >= PERF_TYPE_MAX)
4550                 return -EINVAL;
4551
4552         if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3)
4553                 return -EINVAL;
4554
4555         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4556                 return -EINVAL;
4557
4558         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4559                 return -EINVAL;
4560
4561 out:
4562         return ret;
4563
4564 err_size:
4565         put_user(sizeof(*attr), &uattr->size);
4566         ret = -E2BIG;
4567         goto out;
4568 }
4569
4570 static int perf_event_set_output(struct perf_event *event, int output_fd)
4571 {
4572         struct perf_event *output_event = NULL;
4573         struct file *output_file = NULL;
4574         struct perf_event *old_output;
4575         int fput_needed = 0;
4576         int ret = -EINVAL;
4577
4578         if (!output_fd)
4579                 goto set;
4580
4581         output_file = fget_light(output_fd, &fput_needed);
4582         if (!output_file)
4583                 return -EBADF;
4584
4585         if (output_file->f_op != &perf_fops)
4586                 goto out;
4587
4588         output_event = output_file->private_data;
4589
4590         /* Don't chain output fds */
4591         if (output_event->output)
4592                 goto out;
4593
4594         /* Don't set an output fd when we already have an output channel */
4595         if (event->data)
4596                 goto out;
4597
4598         atomic_long_inc(&output_file->f_count);
4599
4600 set:
4601         mutex_lock(&event->mmap_mutex);
4602         old_output = event->output;
4603         rcu_assign_pointer(event->output, output_event);
4604         mutex_unlock(&event->mmap_mutex);
4605
4606         if (old_output) {
4607                 /*
4608                  * we need to make sure no existing perf_output_*()
4609                  * is still referencing this event.
4610                  */
4611                 synchronize_rcu();
4612                 fput(old_output->filp);
4613         }
4614
4615         ret = 0;
4616 out:
4617         fput_light(output_file, fput_needed);
4618         return ret;
4619 }
4620
4621 /**
4622  * sys_perf_event_open - open a performance event, associate it to a task/cpu
4623  *
4624  * @attr_uptr:  event_id type attributes for monitoring/sampling
4625  * @pid:                target pid
4626  * @cpu:                target cpu
4627  * @group_fd:           group leader event fd
4628  */
4629 SYSCALL_DEFINE5(perf_event_open,
4630                 struct perf_event_attr __user *, attr_uptr,
4631                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4632 {
4633         struct perf_event *event, *group_leader;
4634         struct perf_event_attr attr;
4635         struct perf_event_context *ctx;
4636         struct file *event_file = NULL;
4637         struct file *group_file = NULL;
4638         int fput_needed = 0;
4639         int fput_needed2 = 0;
4640         int err;
4641
4642         /* for future expandability... */
4643         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4644                 return -EINVAL;
4645
4646         err = perf_copy_attr(attr_uptr, &attr);
4647         if (err)
4648                 return err;
4649
4650         if (!attr.exclude_kernel) {
4651                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4652                         return -EACCES;
4653         }
4654
4655         if (attr.freq) {
4656                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4657                         return -EINVAL;
4658         }
4659
4660         /*
4661          * Get the target context (task or percpu):
4662          */
4663         ctx = find_get_context(pid, cpu);
4664         if (IS_ERR(ctx))
4665                 return PTR_ERR(ctx);
4666
4667         /*
4668          * Look up the group leader (we will attach this event to it):
4669          */
4670         group_leader = NULL;
4671         if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4672                 err = -EINVAL;
4673                 group_file = fget_light(group_fd, &fput_needed);
4674                 if (!group_file)
4675                         goto err_put_context;
4676                 if (group_file->f_op != &perf_fops)
4677                         goto err_put_context;
4678
4679                 group_leader = group_file->private_data;
4680                 /*
4681                  * Do not allow a recursive hierarchy (this new sibling
4682                  * becoming part of another group-sibling):
4683                  */
4684                 if (group_leader->group_leader != group_leader)
4685                         goto err_put_context;
4686                 /*
4687                  * Do not allow to attach to a group in a different
4688                  * task or CPU context:
4689                  */
4690                 if (group_leader->ctx != ctx)
4691                         goto err_put_context;
4692                 /*
4693                  * Only a group leader can be exclusive or pinned
4694                  */
4695                 if (attr.exclusive || attr.pinned)
4696                         goto err_put_context;
4697         }
4698
4699         event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4700                                      NULL, NULL, GFP_KERNEL);
4701         err = PTR_ERR(event);
4702         if (IS_ERR(event))
4703                 goto err_put_context;
4704
4705         err = anon_inode_getfd("[perf_event]", &perf_fops, event, 0);
4706         if (err < 0)
4707                 goto err_free_put_context;
4708
4709         event_file = fget_light(err, &fput_needed2);
4710         if (!event_file)
4711                 goto err_free_put_context;
4712
4713         if (flags & PERF_FLAG_FD_OUTPUT) {
4714                 err = perf_event_set_output(event, group_fd);
4715                 if (err)
4716                         goto err_fput_free_put_context;
4717         }
4718
4719         event->filp = event_file;
4720         WARN_ON_ONCE(ctx->parent_ctx);
4721         mutex_lock(&ctx->mutex);
4722         perf_install_in_context(ctx, event, cpu);
4723         ++ctx->generation;
4724         mutex_unlock(&ctx->mutex);
4725
4726         event->owner = current;
4727         get_task_struct(current);
4728         mutex_lock(&current->perf_event_mutex);
4729         list_add_tail(&event->owner_entry, &current->perf_event_list);
4730         mutex_unlock(&current->perf_event_mutex);
4731
4732 err_fput_free_put_context:
4733         fput_light(event_file, fput_needed2);
4734
4735 err_free_put_context:
4736         if (err < 0)
4737                 kfree(event);
4738
4739 err_put_context:
4740         if (err < 0)
4741                 put_ctx(ctx);
4742
4743         fput_light(group_file, fput_needed);
4744
4745         return err;
4746 }
4747
4748 /**
4749  * perf_event_create_kernel_counter
4750  *
4751  * @attr: attributes of the counter to create
4752  * @cpu: cpu in which the counter is bound
4753  * @pid: task to profile
4754  */
4755 struct perf_event *
4756 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4757                                  pid_t pid, perf_callback_t callback)
4758 {
4759         struct perf_event *event;
4760         struct perf_event_context *ctx;
4761         int err;
4762
4763         /*
4764          * Get the target context (task or percpu):
4765          */
4766
4767         ctx = find_get_context(pid, cpu);
4768         if (IS_ERR(ctx))
4769                 return NULL;
4770
4771         event = perf_event_alloc(attr, cpu, ctx, NULL,
4772                                      NULL, callback, GFP_KERNEL);
4773         err = PTR_ERR(event);
4774         if (IS_ERR(event))
4775                 goto err_put_context;
4776
4777         event->filp = NULL;
4778         WARN_ON_ONCE(ctx->parent_ctx);
4779         mutex_lock(&ctx->mutex);
4780         perf_install_in_context(ctx, event, cpu);
4781         ++ctx->generation;
4782         mutex_unlock(&ctx->mutex);
4783
4784         event->owner = current;
4785         get_task_struct(current);
4786         mutex_lock(&current->perf_event_mutex);
4787         list_add_tail(&event->owner_entry, &current->perf_event_list);
4788         mutex_unlock(&current->perf_event_mutex);
4789
4790         return event;
4791
4792 err_put_context:
4793         if (err < 0)
4794                 put_ctx(ctx);
4795
4796         return NULL;
4797 }
4798 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4799
4800 /*
4801  * inherit a event from parent task to child task:
4802  */
4803 static struct perf_event *
4804 inherit_event(struct perf_event *parent_event,
4805               struct task_struct *parent,
4806               struct perf_event_context *parent_ctx,
4807               struct task_struct *child,
4808               struct perf_event *group_leader,
4809               struct perf_event_context *child_ctx)
4810 {
4811         struct perf_event *child_event;
4812
4813         /*
4814          * Instead of creating recursive hierarchies of events,
4815          * we link inherited events back to the original parent,
4816          * which has a filp for sure, which we use as the reference
4817          * count:
4818          */
4819         if (parent_event->parent)
4820                 parent_event = parent_event->parent;
4821
4822         child_event = perf_event_alloc(&parent_event->attr,
4823                                            parent_event->cpu, child_ctx,
4824                                            group_leader, parent_event,
4825                                            NULL, GFP_KERNEL);
4826         if (IS_ERR(child_event))
4827                 return child_event;
4828         get_ctx(child_ctx);
4829
4830         /*
4831          * Make the child state follow the state of the parent event,
4832          * not its attr.disabled bit.  We hold the parent's mutex,
4833          * so we won't race with perf_event_{en, dis}able_family.
4834          */
4835         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
4836                 child_event->state = PERF_EVENT_STATE_INACTIVE;
4837         else
4838                 child_event->state = PERF_EVENT_STATE_OFF;
4839
4840         if (parent_event->attr.freq)
4841                 child_event->hw.sample_period = parent_event->hw.sample_period;
4842
4843         child_event->overflow_handler = parent_event->overflow_handler;
4844
4845         /*
4846          * Link it up in the child's context:
4847          */
4848         add_event_to_ctx(child_event, child_ctx);
4849
4850         /*
4851          * Get a reference to the parent filp - we will fput it
4852          * when the child event exits. This is safe to do because
4853          * we are in the parent and we know that the filp still
4854          * exists and has a nonzero count:
4855          */
4856         atomic_long_inc(&parent_event->filp->f_count);
4857
4858         /*
4859          * Link this into the parent event's child list
4860          */
4861         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4862         mutex_lock(&parent_event->child_mutex);
4863         list_add_tail(&child_event->child_list, &parent_event->child_list);
4864         mutex_unlock(&parent_event->child_mutex);
4865
4866         return child_event;
4867 }
4868
4869 static int inherit_group(struct perf_event *parent_event,
4870               struct task_struct *parent,
4871               struct perf_event_context *parent_ctx,
4872               struct task_struct *child,
4873               struct perf_event_context *child_ctx)
4874 {
4875         struct perf_event *leader;
4876         struct perf_event *sub;
4877         struct perf_event *child_ctr;
4878
4879         leader = inherit_event(parent_event, parent, parent_ctx,
4880                                  child, NULL, child_ctx);
4881         if (IS_ERR(leader))
4882                 return PTR_ERR(leader);
4883         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
4884                 child_ctr = inherit_event(sub, parent, parent_ctx,
4885                                             child, leader, child_ctx);
4886                 if (IS_ERR(child_ctr))
4887                         return PTR_ERR(child_ctr);
4888         }
4889         return 0;
4890 }
4891
4892 static void sync_child_event(struct perf_event *child_event,
4893                                struct task_struct *child)
4894 {
4895         struct perf_event *parent_event = child_event->parent;
4896         u64 child_val;
4897
4898         if (child_event->attr.inherit_stat)
4899                 perf_event_read_event(child_event, child);
4900
4901         child_val = atomic64_read(&child_event->count);
4902
4903         /*
4904          * Add back the child's count to the parent's count:
4905          */
4906         atomic64_add(child_val, &parent_event->count);
4907         atomic64_add(child_event->total_time_enabled,
4908                      &parent_event->child_total_time_enabled);
4909         atomic64_add(child_event->total_time_running,
4910                      &parent_event->child_total_time_running);
4911
4912         /*
4913          * Remove this event from the parent's list
4914          */
4915         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4916         mutex_lock(&parent_event->child_mutex);
4917         list_del_init(&child_event->child_list);
4918         mutex_unlock(&parent_event->child_mutex);
4919
4920         /*
4921          * Release the parent event, if this was the last
4922          * reference to it.
4923          */
4924         fput(parent_event->filp);
4925 }
4926
4927 static void
4928 __perf_event_exit_task(struct perf_event *child_event,
4929                          struct perf_event_context *child_ctx,
4930                          struct task_struct *child)
4931 {
4932         struct perf_event *parent_event;
4933
4934         update_event_times(child_event);
4935         perf_event_remove_from_context(child_event);
4936
4937         parent_event = child_event->parent;
4938         /*
4939          * It can happen that parent exits first, and has events
4940          * that are still around due to the child reference. These
4941          * events need to be zapped - but otherwise linger.
4942          */
4943         if (parent_event) {
4944                 sync_child_event(child_event, child);
4945                 free_event(child_event);
4946         }
4947 }
4948
4949 /*
4950  * When a child task exits, feed back event values to parent events.
4951  */
4952 void perf_event_exit_task(struct task_struct *child)
4953 {
4954         struct perf_event *child_event, *tmp;
4955         struct perf_event_context *child_ctx;
4956         unsigned long flags;
4957
4958         if (likely(!child->perf_event_ctxp)) {
4959                 perf_event_task(child, NULL, 0);
4960                 return;
4961         }
4962
4963         local_irq_save(flags);
4964         /*
4965          * We can't reschedule here because interrupts are disabled,
4966          * and either child is current or it is a task that can't be
4967          * scheduled, so we are now safe from rescheduling changing
4968          * our context.
4969          */
4970         child_ctx = child->perf_event_ctxp;
4971         __perf_event_task_sched_out(child_ctx);
4972
4973         /*
4974          * Take the context lock here so that if find_get_context is
4975          * reading child->perf_event_ctxp, we wait until it has
4976          * incremented the context's refcount before we do put_ctx below.
4977          */
4978         spin_lock(&child_ctx->lock);
4979         child->perf_event_ctxp = NULL;
4980         /*
4981          * If this context is a clone; unclone it so it can't get
4982          * swapped to another process while we're removing all
4983          * the events from it.
4984          */
4985         unclone_ctx(child_ctx);
4986         spin_unlock_irqrestore(&child_ctx->lock, flags);
4987
4988         /*
4989          * Report the task dead after unscheduling the events so that we
4990          * won't get any samples after PERF_RECORD_EXIT. We can however still
4991          * get a few PERF_RECORD_READ events.
4992          */
4993         perf_event_task(child, child_ctx, 0);
4994
4995         /*
4996          * We can recurse on the same lock type through:
4997          *
4998          *   __perf_event_exit_task()
4999          *     sync_child_event()
5000          *       fput(parent_event->filp)
5001          *         perf_release()
5002          *           mutex_lock(&ctx->mutex)
5003          *
5004          * But since its the parent context it won't be the same instance.
5005          */
5006         mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
5007
5008 again:
5009         list_for_each_entry_safe(child_event, tmp, &child_ctx->group_list,
5010                                  group_entry)
5011                 __perf_event_exit_task(child_event, child_ctx, child);
5012
5013         /*
5014          * If the last event was a group event, it will have appended all
5015          * its siblings to the list, but we obtained 'tmp' before that which
5016          * will still point to the list head terminating the iteration.
5017          */
5018         if (!list_empty(&child_ctx->group_list))
5019                 goto again;
5020
5021         mutex_unlock(&child_ctx->mutex);
5022
5023         put_ctx(child_ctx);
5024 }
5025
5026 /*
5027  * free an unexposed, unused context as created by inheritance by
5028  * init_task below, used by fork() in case of fail.
5029  */
5030 void perf_event_free_task(struct task_struct *task)
5031 {
5032         struct perf_event_context *ctx = task->perf_event_ctxp;
5033         struct perf_event *event, *tmp;
5034
5035         if (!ctx)
5036                 return;
5037
5038         mutex_lock(&ctx->mutex);
5039 again:
5040         list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry) {
5041                 struct perf_event *parent = event->parent;
5042
5043                 if (WARN_ON_ONCE(!parent))
5044                         continue;
5045
5046                 mutex_lock(&parent->child_mutex);
5047                 list_del_init(&event->child_list);
5048                 mutex_unlock(&parent->child_mutex);
5049
5050                 fput(parent->filp);
5051
5052                 list_del_event(event, ctx);
5053                 free_event(event);
5054         }
5055
5056         if (!list_empty(&ctx->group_list))
5057                 goto again;
5058
5059         mutex_unlock(&ctx->mutex);
5060
5061         put_ctx(ctx);
5062 }
5063
5064 /*
5065  * Initialize the perf_event context in task_struct
5066  */
5067 int perf_event_init_task(struct task_struct *child)
5068 {
5069         struct perf_event_context *child_ctx, *parent_ctx;
5070         struct perf_event_context *cloned_ctx;
5071         struct perf_event *event;
5072         struct task_struct *parent = current;
5073         int inherited_all = 1;
5074         int ret = 0;
5075
5076         child->perf_event_ctxp = NULL;
5077
5078         mutex_init(&child->perf_event_mutex);
5079         INIT_LIST_HEAD(&child->perf_event_list);
5080
5081         if (likely(!parent->perf_event_ctxp))
5082                 return 0;
5083
5084         /*
5085          * This is executed from the parent task context, so inherit
5086          * events that have been marked for cloning.
5087          * First allocate and initialize a context for the child.
5088          */
5089
5090         child_ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
5091         if (!child_ctx)
5092                 return -ENOMEM;
5093
5094         __perf_event_init_context(child_ctx, child);
5095         child->perf_event_ctxp = child_ctx;
5096         get_task_struct(child);
5097
5098         /*
5099          * If the parent's context is a clone, pin it so it won't get
5100          * swapped under us.
5101          */
5102         parent_ctx = perf_pin_task_context(parent);
5103
5104         /*
5105          * No need to check if parent_ctx != NULL here; since we saw
5106          * it non-NULL earlier, the only reason for it to become NULL
5107          * is if we exit, and since we're currently in the middle of
5108          * a fork we can't be exiting at the same time.
5109          */
5110
5111         /*
5112          * Lock the parent list. No need to lock the child - not PID
5113          * hashed yet and not running, so nobody can access it.
5114          */
5115         mutex_lock(&parent_ctx->mutex);
5116
5117         /*
5118          * We dont have to disable NMIs - we are only looking at
5119          * the list, not manipulating it:
5120          */
5121         list_for_each_entry(event, &parent_ctx->group_list, group_entry) {
5122
5123                 if (!event->attr.inherit) {
5124                         inherited_all = 0;
5125                         continue;
5126                 }
5127
5128                 ret = inherit_group(event, parent, parent_ctx,
5129                                              child, child_ctx);
5130                 if (ret) {
5131                         inherited_all = 0;
5132                         break;
5133                 }
5134         }
5135
5136         if (inherited_all) {
5137                 /*
5138                  * Mark the child context as a clone of the parent
5139                  * context, or of whatever the parent is a clone of.
5140                  * Note that if the parent is a clone, it could get
5141                  * uncloned at any point, but that doesn't matter
5142                  * because the list of events and the generation
5143                  * count can't have changed since we took the mutex.
5144                  */
5145                 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5146                 if (cloned_ctx) {
5147                         child_ctx->parent_ctx = cloned_ctx;
5148                         child_ctx->parent_gen = parent_ctx->parent_gen;
5149                 } else {
5150                         child_ctx->parent_ctx = parent_ctx;
5151                         child_ctx->parent_gen = parent_ctx->generation;
5152                 }
5153                 get_ctx(child_ctx->parent_ctx);
5154         }
5155
5156         mutex_unlock(&parent_ctx->mutex);
5157
5158         perf_unpin_context(parent_ctx);
5159
5160         return ret;
5161 }
5162
5163 static void __cpuinit perf_event_init_cpu(int cpu)
5164 {
5165         struct perf_cpu_context *cpuctx;
5166
5167         cpuctx = &per_cpu(perf_cpu_context, cpu);
5168         __perf_event_init_context(&cpuctx->ctx, NULL);
5169
5170         spin_lock(&perf_resource_lock);
5171         cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5172         spin_unlock(&perf_resource_lock);
5173
5174         hw_perf_event_setup(cpu);
5175 }
5176
5177 #ifdef CONFIG_HOTPLUG_CPU
5178 static void __perf_event_exit_cpu(void *info)
5179 {
5180         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5181         struct perf_event_context *ctx = &cpuctx->ctx;
5182         struct perf_event *event, *tmp;
5183
5184         list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry)
5185                 __perf_event_remove_from_context(event);
5186 }
5187 static void perf_event_exit_cpu(int cpu)
5188 {
5189         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5190         struct perf_event_context *ctx = &cpuctx->ctx;
5191
5192         mutex_lock(&ctx->mutex);
5193         smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5194         mutex_unlock(&ctx->mutex);
5195 }
5196 #else
5197 static inline void perf_event_exit_cpu(int cpu) { }
5198 #endif
5199
5200 static int __cpuinit
5201 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5202 {
5203         unsigned int cpu = (long)hcpu;
5204
5205         switch (action) {
5206
5207         case CPU_UP_PREPARE:
5208         case CPU_UP_PREPARE_FROZEN:
5209                 perf_event_init_cpu(cpu);
5210                 break;
5211
5212         case CPU_ONLINE:
5213         case CPU_ONLINE_FROZEN:
5214                 hw_perf_event_setup_online(cpu);
5215                 break;
5216
5217         case CPU_DOWN_PREPARE:
5218         case CPU_DOWN_PREPARE_FROZEN:
5219                 perf_event_exit_cpu(cpu);
5220                 break;
5221
5222         default:
5223                 break;
5224         }
5225
5226         return NOTIFY_OK;
5227 }
5228
5229 /*
5230  * This has to have a higher priority than migration_notifier in sched.c.
5231  */
5232 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5233         .notifier_call          = perf_cpu_notify,
5234         .priority               = 20,
5235 };
5236
5237 void __init perf_event_init(void)
5238 {
5239         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5240                         (void *)(long)smp_processor_id());
5241         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5242                         (void *)(long)smp_processor_id());
5243         register_cpu_notifier(&perf_cpu_nb);
5244 }
5245
5246 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
5247 {
5248         return sprintf(buf, "%d\n", perf_reserved_percpu);
5249 }
5250
5251 static ssize_t
5252 perf_set_reserve_percpu(struct sysdev_class *class,
5253                         const char *buf,
5254                         size_t count)
5255 {
5256         struct perf_cpu_context *cpuctx;
5257         unsigned long val;
5258         int err, cpu, mpt;
5259
5260         err = strict_strtoul(buf, 10, &val);
5261         if (err)
5262                 return err;
5263         if (val > perf_max_events)
5264                 return -EINVAL;
5265
5266         spin_lock(&perf_resource_lock);
5267         perf_reserved_percpu = val;
5268         for_each_online_cpu(cpu) {
5269                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5270                 spin_lock_irq(&cpuctx->ctx.lock);
5271                 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5272                           perf_max_events - perf_reserved_percpu);
5273                 cpuctx->max_pertask = mpt;
5274                 spin_unlock_irq(&cpuctx->ctx.lock);
5275         }
5276         spin_unlock(&perf_resource_lock);
5277
5278         return count;
5279 }
5280
5281 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
5282 {
5283         return sprintf(buf, "%d\n", perf_overcommit);
5284 }
5285
5286 static ssize_t
5287 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
5288 {
5289         unsigned long val;
5290         int err;
5291
5292         err = strict_strtoul(buf, 10, &val);
5293         if (err)
5294                 return err;
5295         if (val > 1)
5296                 return -EINVAL;
5297
5298         spin_lock(&perf_resource_lock);
5299         perf_overcommit = val;
5300         spin_unlock(&perf_resource_lock);
5301
5302         return count;
5303 }
5304
5305 static SYSDEV_CLASS_ATTR(
5306                                 reserve_percpu,
5307                                 0644,
5308                                 perf_show_reserve_percpu,
5309                                 perf_set_reserve_percpu
5310                         );
5311
5312 static SYSDEV_CLASS_ATTR(
5313                                 overcommit,
5314                                 0644,
5315                                 perf_show_overcommit,
5316                                 perf_set_overcommit
5317                         );
5318
5319 static struct attribute *perfclass_attrs[] = {
5320         &attr_reserve_percpu.attr,
5321         &attr_overcommit.attr,
5322         NULL
5323 };
5324
5325 static struct attribute_group perfclass_attr_group = {
5326         .attrs                  = perfclass_attrs,
5327         .name                   = "perf_events",
5328 };
5329
5330 static int __init perf_event_sysfs_init(void)
5331 {
5332         return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5333                                   &perfclass_attr_group);
5334 }
5335 device_initcall(perf_event_sysfs_init);