e66f6c400d131c5d2fe72b56f5f501467a988a96
[linux-2.6.git] / kernel / perf_event.c
1 /*
2  * Performance events core code:
3  *
4  *  Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5  *  Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6  *  Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7  *  Copyright  ©  2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
8  *
9  * For licensing details see kernel-base/COPYING
10  */
11
12 #include <linux/fs.h>
13 #include <linux/mm.h>
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/file.h>
17 #include <linux/poll.h>
18 #include <linux/sysfs.h>
19 #include <linux/dcache.h>
20 #include <linux/percpu.h>
21 #include <linux/ptrace.h>
22 #include <linux/vmstat.h>
23 #include <linux/vmalloc.h>
24 #include <linux/hardirq.h>
25 #include <linux/rculist.h>
26 #include <linux/uaccess.h>
27 #include <linux/syscalls.h>
28 #include <linux/anon_inodes.h>
29 #include <linux/kernel_stat.h>
30 #include <linux/perf_event.h>
31 #include <linux/ftrace_event.h>
32 #include <linux/hw_breakpoint.h>
33
34 #include <asm/irq_regs.h>
35
36 /*
37  * Each CPU has a list of per CPU events:
38  */
39 DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
40
41 int perf_max_events __read_mostly = 1;
42 static int perf_reserved_percpu __read_mostly;
43 static int perf_overcommit __read_mostly = 1;
44
45 static atomic_t nr_events __read_mostly;
46 static atomic_t nr_mmap_events __read_mostly;
47 static atomic_t nr_comm_events __read_mostly;
48 static atomic_t nr_task_events __read_mostly;
49
50 /*
51  * perf event paranoia level:
52  *  -1 - not paranoid at all
53  *   0 - disallow raw tracepoint access for unpriv
54  *   1 - disallow cpu events for unpriv
55  *   2 - disallow kernel profiling for unpriv
56  */
57 int sysctl_perf_event_paranoid __read_mostly = 1;
58
59 static inline bool perf_paranoid_tracepoint_raw(void)
60 {
61         return sysctl_perf_event_paranoid > -1;
62 }
63
64 static inline bool perf_paranoid_cpu(void)
65 {
66         return sysctl_perf_event_paranoid > 0;
67 }
68
69 static inline bool perf_paranoid_kernel(void)
70 {
71         return sysctl_perf_event_paranoid > 1;
72 }
73
74 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
75
76 /*
77  * max perf event sample rate
78  */
79 int sysctl_perf_event_sample_rate __read_mostly = 100000;
80
81 static atomic64_t perf_event_id;
82
83 /*
84  * Lock for (sysadmin-configurable) event reservations:
85  */
86 static DEFINE_SPINLOCK(perf_resource_lock);
87
88 /*
89  * Architecture provided APIs - weak aliases:
90  */
91 extern __weak const struct pmu *hw_perf_event_init(struct perf_event *event)
92 {
93         return NULL;
94 }
95
96 void __weak hw_perf_disable(void)               { barrier(); }
97 void __weak hw_perf_enable(void)                { barrier(); }
98
99 void __weak hw_perf_event_setup(int cpu)        { barrier(); }
100 void __weak hw_perf_event_setup_online(int cpu) { barrier(); }
101
102 int __weak
103 hw_perf_group_sched_in(struct perf_event *group_leader,
104                struct perf_cpu_context *cpuctx,
105                struct perf_event_context *ctx, int cpu)
106 {
107         return 0;
108 }
109
110 void __weak perf_event_print_debug(void)        { }
111
112 static DEFINE_PER_CPU(int, perf_disable_count);
113
114 void __perf_disable(void)
115 {
116         __get_cpu_var(perf_disable_count)++;
117 }
118
119 bool __perf_enable(void)
120 {
121         return !--__get_cpu_var(perf_disable_count);
122 }
123
124 void perf_disable(void)
125 {
126         __perf_disable();
127         hw_perf_disable();
128 }
129
130 void perf_enable(void)
131 {
132         if (__perf_enable())
133                 hw_perf_enable();
134 }
135
136 static void get_ctx(struct perf_event_context *ctx)
137 {
138         WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
139 }
140
141 static void free_ctx(struct rcu_head *head)
142 {
143         struct perf_event_context *ctx;
144
145         ctx = container_of(head, struct perf_event_context, rcu_head);
146         kfree(ctx);
147 }
148
149 static void put_ctx(struct perf_event_context *ctx)
150 {
151         if (atomic_dec_and_test(&ctx->refcount)) {
152                 if (ctx->parent_ctx)
153                         put_ctx(ctx->parent_ctx);
154                 if (ctx->task)
155                         put_task_struct(ctx->task);
156                 call_rcu(&ctx->rcu_head, free_ctx);
157         }
158 }
159
160 static void unclone_ctx(struct perf_event_context *ctx)
161 {
162         if (ctx->parent_ctx) {
163                 put_ctx(ctx->parent_ctx);
164                 ctx->parent_ctx = NULL;
165         }
166 }
167
168 /*
169  * If we inherit events we want to return the parent event id
170  * to userspace.
171  */
172 static u64 primary_event_id(struct perf_event *event)
173 {
174         u64 id = event->id;
175
176         if (event->parent)
177                 id = event->parent->id;
178
179         return id;
180 }
181
182 /*
183  * Get the perf_event_context for a task and lock it.
184  * This has to cope with with the fact that until it is locked,
185  * the context could get moved to another task.
186  */
187 static struct perf_event_context *
188 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
189 {
190         struct perf_event_context *ctx;
191
192         rcu_read_lock();
193  retry:
194         ctx = rcu_dereference(task->perf_event_ctxp);
195         if (ctx) {
196                 /*
197                  * If this context is a clone of another, it might
198                  * get swapped for another underneath us by
199                  * perf_event_task_sched_out, though the
200                  * rcu_read_lock() protects us from any context
201                  * getting freed.  Lock the context and check if it
202                  * got swapped before we could get the lock, and retry
203                  * if so.  If we locked the right context, then it
204                  * can't get swapped on us any more.
205                  */
206                 spin_lock_irqsave(&ctx->lock, *flags);
207                 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
208                         spin_unlock_irqrestore(&ctx->lock, *flags);
209                         goto retry;
210                 }
211
212                 if (!atomic_inc_not_zero(&ctx->refcount)) {
213                         spin_unlock_irqrestore(&ctx->lock, *flags);
214                         ctx = NULL;
215                 }
216         }
217         rcu_read_unlock();
218         return ctx;
219 }
220
221 /*
222  * Get the context for a task and increment its pin_count so it
223  * can't get swapped to another task.  This also increments its
224  * reference count so that the context can't get freed.
225  */
226 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
227 {
228         struct perf_event_context *ctx;
229         unsigned long flags;
230
231         ctx = perf_lock_task_context(task, &flags);
232         if (ctx) {
233                 ++ctx->pin_count;
234                 spin_unlock_irqrestore(&ctx->lock, flags);
235         }
236         return ctx;
237 }
238
239 static void perf_unpin_context(struct perf_event_context *ctx)
240 {
241         unsigned long flags;
242
243         spin_lock_irqsave(&ctx->lock, flags);
244         --ctx->pin_count;
245         spin_unlock_irqrestore(&ctx->lock, flags);
246         put_ctx(ctx);
247 }
248
249 /*
250  * Add a event from the lists for its context.
251  * Must be called with ctx->mutex and ctx->lock held.
252  */
253 static void
254 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
255 {
256         struct perf_event *group_leader = event->group_leader;
257
258         /*
259          * Depending on whether it is a standalone or sibling event,
260          * add it straight to the context's event list, or to the group
261          * leader's sibling list:
262          */
263         if (group_leader == event)
264                 list_add_tail(&event->group_entry, &ctx->group_list);
265         else {
266                 list_add_tail(&event->group_entry, &group_leader->sibling_list);
267                 group_leader->nr_siblings++;
268         }
269
270         list_add_rcu(&event->event_entry, &ctx->event_list);
271         ctx->nr_events++;
272         if (event->attr.inherit_stat)
273                 ctx->nr_stat++;
274 }
275
276 /*
277  * Remove a event from the lists for its context.
278  * Must be called with ctx->mutex and ctx->lock held.
279  */
280 static void
281 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
282 {
283         struct perf_event *sibling, *tmp;
284
285         if (list_empty(&event->group_entry))
286                 return;
287         ctx->nr_events--;
288         if (event->attr.inherit_stat)
289                 ctx->nr_stat--;
290
291         list_del_init(&event->group_entry);
292         list_del_rcu(&event->event_entry);
293
294         if (event->group_leader != event)
295                 event->group_leader->nr_siblings--;
296
297         /*
298          * If this was a group event with sibling events then
299          * upgrade the siblings to singleton events by adding them
300          * to the context list directly:
301          */
302         list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
303
304                 list_move_tail(&sibling->group_entry, &ctx->group_list);
305                 sibling->group_leader = sibling;
306         }
307 }
308
309 static void
310 event_sched_out(struct perf_event *event,
311                   struct perf_cpu_context *cpuctx,
312                   struct perf_event_context *ctx)
313 {
314         if (event->state != PERF_EVENT_STATE_ACTIVE)
315                 return;
316
317         event->state = PERF_EVENT_STATE_INACTIVE;
318         if (event->pending_disable) {
319                 event->pending_disable = 0;
320                 event->state = PERF_EVENT_STATE_OFF;
321         }
322         event->tstamp_stopped = ctx->time;
323         event->pmu->disable(event);
324         event->oncpu = -1;
325
326         if (!is_software_event(event))
327                 cpuctx->active_oncpu--;
328         ctx->nr_active--;
329         if (event->attr.exclusive || !cpuctx->active_oncpu)
330                 cpuctx->exclusive = 0;
331 }
332
333 static void
334 group_sched_out(struct perf_event *group_event,
335                 struct perf_cpu_context *cpuctx,
336                 struct perf_event_context *ctx)
337 {
338         struct perf_event *event;
339
340         if (group_event->state != PERF_EVENT_STATE_ACTIVE)
341                 return;
342
343         event_sched_out(group_event, cpuctx, ctx);
344
345         /*
346          * Schedule out siblings (if any):
347          */
348         list_for_each_entry(event, &group_event->sibling_list, group_entry)
349                 event_sched_out(event, cpuctx, ctx);
350
351         if (group_event->attr.exclusive)
352                 cpuctx->exclusive = 0;
353 }
354
355 /*
356  * Cross CPU call to remove a performance event
357  *
358  * We disable the event on the hardware level first. After that we
359  * remove it from the context list.
360  */
361 static void __perf_event_remove_from_context(void *info)
362 {
363         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
364         struct perf_event *event = info;
365         struct perf_event_context *ctx = event->ctx;
366
367         /*
368          * If this is a task context, we need to check whether it is
369          * the current task context of this cpu. If not it has been
370          * scheduled out before the smp call arrived.
371          */
372         if (ctx->task && cpuctx->task_ctx != ctx)
373                 return;
374
375         spin_lock(&ctx->lock);
376         /*
377          * Protect the list operation against NMI by disabling the
378          * events on a global level.
379          */
380         perf_disable();
381
382         event_sched_out(event, cpuctx, ctx);
383
384         list_del_event(event, ctx);
385
386         if (!ctx->task) {
387                 /*
388                  * Allow more per task events with respect to the
389                  * reservation:
390                  */
391                 cpuctx->max_pertask =
392                         min(perf_max_events - ctx->nr_events,
393                             perf_max_events - perf_reserved_percpu);
394         }
395
396         perf_enable();
397         spin_unlock(&ctx->lock);
398 }
399
400
401 /*
402  * Remove the event from a task's (or a CPU's) list of events.
403  *
404  * Must be called with ctx->mutex held.
405  *
406  * CPU events are removed with a smp call. For task events we only
407  * call when the task is on a CPU.
408  *
409  * If event->ctx is a cloned context, callers must make sure that
410  * every task struct that event->ctx->task could possibly point to
411  * remains valid.  This is OK when called from perf_release since
412  * that only calls us on the top-level context, which can't be a clone.
413  * When called from perf_event_exit_task, it's OK because the
414  * context has been detached from its task.
415  */
416 static void perf_event_remove_from_context(struct perf_event *event)
417 {
418         struct perf_event_context *ctx = event->ctx;
419         struct task_struct *task = ctx->task;
420
421         if (!task) {
422                 /*
423                  * Per cpu events are removed via an smp call and
424                  * the removal is always sucessful.
425                  */
426                 smp_call_function_single(event->cpu,
427                                          __perf_event_remove_from_context,
428                                          event, 1);
429                 return;
430         }
431
432 retry:
433         task_oncpu_function_call(task, __perf_event_remove_from_context,
434                                  event);
435
436         spin_lock_irq(&ctx->lock);
437         /*
438          * If the context is active we need to retry the smp call.
439          */
440         if (ctx->nr_active && !list_empty(&event->group_entry)) {
441                 spin_unlock_irq(&ctx->lock);
442                 goto retry;
443         }
444
445         /*
446          * The lock prevents that this context is scheduled in so we
447          * can remove the event safely, if the call above did not
448          * succeed.
449          */
450         if (!list_empty(&event->group_entry)) {
451                 list_del_event(event, ctx);
452         }
453         spin_unlock_irq(&ctx->lock);
454 }
455
456 static inline u64 perf_clock(void)
457 {
458         return cpu_clock(smp_processor_id());
459 }
460
461 /*
462  * Update the record of the current time in a context.
463  */
464 static void update_context_time(struct perf_event_context *ctx)
465 {
466         u64 now = perf_clock();
467
468         ctx->time += now - ctx->timestamp;
469         ctx->timestamp = now;
470 }
471
472 /*
473  * Update the total_time_enabled and total_time_running fields for a event.
474  */
475 static void update_event_times(struct perf_event *event)
476 {
477         struct perf_event_context *ctx = event->ctx;
478         u64 run_end;
479
480         if (event->state < PERF_EVENT_STATE_INACTIVE ||
481             event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
482                 return;
483
484         event->total_time_enabled = ctx->time - event->tstamp_enabled;
485
486         if (event->state == PERF_EVENT_STATE_INACTIVE)
487                 run_end = event->tstamp_stopped;
488         else
489                 run_end = ctx->time;
490
491         event->total_time_running = run_end - event->tstamp_running;
492 }
493
494 /*
495  * Update total_time_enabled and total_time_running for all events in a group.
496  */
497 static void update_group_times(struct perf_event *leader)
498 {
499         struct perf_event *event;
500
501         update_event_times(leader);
502         list_for_each_entry(event, &leader->sibling_list, group_entry)
503                 update_event_times(event);
504 }
505
506 /*
507  * Cross CPU call to disable a performance event
508  */
509 static void __perf_event_disable(void *info)
510 {
511         struct perf_event *event = info;
512         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
513         struct perf_event_context *ctx = event->ctx;
514
515         /*
516          * If this is a per-task event, need to check whether this
517          * event's task is the current task on this cpu.
518          */
519         if (ctx->task && cpuctx->task_ctx != ctx)
520                 return;
521
522         spin_lock(&ctx->lock);
523
524         /*
525          * If the event is on, turn it off.
526          * If it is in error state, leave it in error state.
527          */
528         if (event->state >= PERF_EVENT_STATE_INACTIVE) {
529                 update_context_time(ctx);
530                 update_group_times(event);
531                 if (event == event->group_leader)
532                         group_sched_out(event, cpuctx, ctx);
533                 else
534                         event_sched_out(event, cpuctx, ctx);
535                 event->state = PERF_EVENT_STATE_OFF;
536         }
537
538         spin_unlock(&ctx->lock);
539 }
540
541 /*
542  * Disable a event.
543  *
544  * If event->ctx is a cloned context, callers must make sure that
545  * every task struct that event->ctx->task could possibly point to
546  * remains valid.  This condition is satisifed when called through
547  * perf_event_for_each_child or perf_event_for_each because they
548  * hold the top-level event's child_mutex, so any descendant that
549  * goes to exit will block in sync_child_event.
550  * When called from perf_pending_event it's OK because event->ctx
551  * is the current context on this CPU and preemption is disabled,
552  * hence we can't get into perf_event_task_sched_out for this context.
553  */
554 static void perf_event_disable(struct perf_event *event)
555 {
556         struct perf_event_context *ctx = event->ctx;
557         struct task_struct *task = ctx->task;
558
559         if (!task) {
560                 /*
561                  * Disable the event on the cpu that it's on
562                  */
563                 smp_call_function_single(event->cpu, __perf_event_disable,
564                                          event, 1);
565                 return;
566         }
567
568  retry:
569         task_oncpu_function_call(task, __perf_event_disable, event);
570
571         spin_lock_irq(&ctx->lock);
572         /*
573          * If the event is still active, we need to retry the cross-call.
574          */
575         if (event->state == PERF_EVENT_STATE_ACTIVE) {
576                 spin_unlock_irq(&ctx->lock);
577                 goto retry;
578         }
579
580         /*
581          * Since we have the lock this context can't be scheduled
582          * in, so we can change the state safely.
583          */
584         if (event->state == PERF_EVENT_STATE_INACTIVE) {
585                 update_group_times(event);
586                 event->state = PERF_EVENT_STATE_OFF;
587         }
588
589         spin_unlock_irq(&ctx->lock);
590 }
591
592 static int
593 event_sched_in(struct perf_event *event,
594                  struct perf_cpu_context *cpuctx,
595                  struct perf_event_context *ctx,
596                  int cpu)
597 {
598         if (event->state <= PERF_EVENT_STATE_OFF)
599                 return 0;
600
601         event->state = PERF_EVENT_STATE_ACTIVE;
602         event->oncpu = cpu;     /* TODO: put 'cpu' into cpuctx->cpu */
603         /*
604          * The new state must be visible before we turn it on in the hardware:
605          */
606         smp_wmb();
607
608         if (event->pmu->enable(event)) {
609                 event->state = PERF_EVENT_STATE_INACTIVE;
610                 event->oncpu = -1;
611                 return -EAGAIN;
612         }
613
614         event->tstamp_running += ctx->time - event->tstamp_stopped;
615
616         if (!is_software_event(event))
617                 cpuctx->active_oncpu++;
618         ctx->nr_active++;
619
620         if (event->attr.exclusive)
621                 cpuctx->exclusive = 1;
622
623         return 0;
624 }
625
626 static int
627 group_sched_in(struct perf_event *group_event,
628                struct perf_cpu_context *cpuctx,
629                struct perf_event_context *ctx,
630                int cpu)
631 {
632         struct perf_event *event, *partial_group;
633         int ret;
634
635         if (group_event->state == PERF_EVENT_STATE_OFF)
636                 return 0;
637
638         ret = hw_perf_group_sched_in(group_event, cpuctx, ctx, cpu);
639         if (ret)
640                 return ret < 0 ? ret : 0;
641
642         if (event_sched_in(group_event, cpuctx, ctx, cpu))
643                 return -EAGAIN;
644
645         /*
646          * Schedule in siblings as one group (if any):
647          */
648         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
649                 if (event_sched_in(event, cpuctx, ctx, cpu)) {
650                         partial_group = event;
651                         goto group_error;
652                 }
653         }
654
655         return 0;
656
657 group_error:
658         /*
659          * Groups can be scheduled in as one unit only, so undo any
660          * partial group before returning:
661          */
662         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
663                 if (event == partial_group)
664                         break;
665                 event_sched_out(event, cpuctx, ctx);
666         }
667         event_sched_out(group_event, cpuctx, ctx);
668
669         return -EAGAIN;
670 }
671
672 /*
673  * Return 1 for a group consisting entirely of software events,
674  * 0 if the group contains any hardware events.
675  */
676 static int is_software_only_group(struct perf_event *leader)
677 {
678         struct perf_event *event;
679
680         if (!is_software_event(leader))
681                 return 0;
682
683         list_for_each_entry(event, &leader->sibling_list, group_entry)
684                 if (!is_software_event(event))
685                         return 0;
686
687         return 1;
688 }
689
690 /*
691  * Work out whether we can put this event group on the CPU now.
692  */
693 static int group_can_go_on(struct perf_event *event,
694                            struct perf_cpu_context *cpuctx,
695                            int can_add_hw)
696 {
697         /*
698          * Groups consisting entirely of software events can always go on.
699          */
700         if (is_software_only_group(event))
701                 return 1;
702         /*
703          * If an exclusive group is already on, no other hardware
704          * events can go on.
705          */
706         if (cpuctx->exclusive)
707                 return 0;
708         /*
709          * If this group is exclusive and there are already
710          * events on the CPU, it can't go on.
711          */
712         if (event->attr.exclusive && cpuctx->active_oncpu)
713                 return 0;
714         /*
715          * Otherwise, try to add it if all previous groups were able
716          * to go on.
717          */
718         return can_add_hw;
719 }
720
721 static void add_event_to_ctx(struct perf_event *event,
722                                struct perf_event_context *ctx)
723 {
724         list_add_event(event, ctx);
725         event->tstamp_enabled = ctx->time;
726         event->tstamp_running = ctx->time;
727         event->tstamp_stopped = ctx->time;
728 }
729
730 /*
731  * Cross CPU call to install and enable a performance event
732  *
733  * Must be called with ctx->mutex held
734  */
735 static void __perf_install_in_context(void *info)
736 {
737         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
738         struct perf_event *event = info;
739         struct perf_event_context *ctx = event->ctx;
740         struct perf_event *leader = event->group_leader;
741         int cpu = smp_processor_id();
742         int err;
743
744         /*
745          * If this is a task context, we need to check whether it is
746          * the current task context of this cpu. If not it has been
747          * scheduled out before the smp call arrived.
748          * Or possibly this is the right context but it isn't
749          * on this cpu because it had no events.
750          */
751         if (ctx->task && cpuctx->task_ctx != ctx) {
752                 if (cpuctx->task_ctx || ctx->task != current)
753                         return;
754                 cpuctx->task_ctx = ctx;
755         }
756
757         spin_lock(&ctx->lock);
758         ctx->is_active = 1;
759         update_context_time(ctx);
760
761         /*
762          * Protect the list operation against NMI by disabling the
763          * events on a global level. NOP for non NMI based events.
764          */
765         perf_disable();
766
767         add_event_to_ctx(event, ctx);
768
769         /*
770          * Don't put the event on if it is disabled or if
771          * it is in a group and the group isn't on.
772          */
773         if (event->state != PERF_EVENT_STATE_INACTIVE ||
774             (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
775                 goto unlock;
776
777         /*
778          * An exclusive event can't go on if there are already active
779          * hardware events, and no hardware event can go on if there
780          * is already an exclusive event on.
781          */
782         if (!group_can_go_on(event, cpuctx, 1))
783                 err = -EEXIST;
784         else
785                 err = event_sched_in(event, cpuctx, ctx, cpu);
786
787         if (err) {
788                 /*
789                  * This event couldn't go on.  If it is in a group
790                  * then we have to pull the whole group off.
791                  * If the event group is pinned then put it in error state.
792                  */
793                 if (leader != event)
794                         group_sched_out(leader, cpuctx, ctx);
795                 if (leader->attr.pinned) {
796                         update_group_times(leader);
797                         leader->state = PERF_EVENT_STATE_ERROR;
798                 }
799         }
800
801         if (!err && !ctx->task && cpuctx->max_pertask)
802                 cpuctx->max_pertask--;
803
804  unlock:
805         perf_enable();
806
807         spin_unlock(&ctx->lock);
808 }
809
810 /*
811  * Attach a performance event to a context
812  *
813  * First we add the event to the list with the hardware enable bit
814  * in event->hw_config cleared.
815  *
816  * If the event is attached to a task which is on a CPU we use a smp
817  * call to enable it in the task context. The task might have been
818  * scheduled away, but we check this in the smp call again.
819  *
820  * Must be called with ctx->mutex held.
821  */
822 static void
823 perf_install_in_context(struct perf_event_context *ctx,
824                         struct perf_event *event,
825                         int cpu)
826 {
827         struct task_struct *task = ctx->task;
828
829         if (!task) {
830                 /*
831                  * Per cpu events are installed via an smp call and
832                  * the install is always sucessful.
833                  */
834                 smp_call_function_single(cpu, __perf_install_in_context,
835                                          event, 1);
836                 return;
837         }
838
839 retry:
840         task_oncpu_function_call(task, __perf_install_in_context,
841                                  event);
842
843         spin_lock_irq(&ctx->lock);
844         /*
845          * we need to retry the smp call.
846          */
847         if (ctx->is_active && list_empty(&event->group_entry)) {
848                 spin_unlock_irq(&ctx->lock);
849                 goto retry;
850         }
851
852         /*
853          * The lock prevents that this context is scheduled in so we
854          * can add the event safely, if it the call above did not
855          * succeed.
856          */
857         if (list_empty(&event->group_entry))
858                 add_event_to_ctx(event, ctx);
859         spin_unlock_irq(&ctx->lock);
860 }
861
862 /*
863  * Put a event into inactive state and update time fields.
864  * Enabling the leader of a group effectively enables all
865  * the group members that aren't explicitly disabled, so we
866  * have to update their ->tstamp_enabled also.
867  * Note: this works for group members as well as group leaders
868  * since the non-leader members' sibling_lists will be empty.
869  */
870 static void __perf_event_mark_enabled(struct perf_event *event,
871                                         struct perf_event_context *ctx)
872 {
873         struct perf_event *sub;
874
875         event->state = PERF_EVENT_STATE_INACTIVE;
876         event->tstamp_enabled = ctx->time - event->total_time_enabled;
877         list_for_each_entry(sub, &event->sibling_list, group_entry)
878                 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
879                         sub->tstamp_enabled =
880                                 ctx->time - sub->total_time_enabled;
881 }
882
883 /*
884  * Cross CPU call to enable a performance event
885  */
886 static void __perf_event_enable(void *info)
887 {
888         struct perf_event *event = info;
889         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
890         struct perf_event_context *ctx = event->ctx;
891         struct perf_event *leader = event->group_leader;
892         int err;
893
894         /*
895          * If this is a per-task event, need to check whether this
896          * event's task is the current task on this cpu.
897          */
898         if (ctx->task && cpuctx->task_ctx != ctx) {
899                 if (cpuctx->task_ctx || ctx->task != current)
900                         return;
901                 cpuctx->task_ctx = ctx;
902         }
903
904         spin_lock(&ctx->lock);
905         ctx->is_active = 1;
906         update_context_time(ctx);
907
908         if (event->state >= PERF_EVENT_STATE_INACTIVE)
909                 goto unlock;
910         __perf_event_mark_enabled(event, ctx);
911
912         /*
913          * If the event is in a group and isn't the group leader,
914          * then don't put it on unless the group is on.
915          */
916         if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
917                 goto unlock;
918
919         if (!group_can_go_on(event, cpuctx, 1)) {
920                 err = -EEXIST;
921         } else {
922                 perf_disable();
923                 if (event == leader)
924                         err = group_sched_in(event, cpuctx, ctx,
925                                              smp_processor_id());
926                 else
927                         err = event_sched_in(event, cpuctx, ctx,
928                                                smp_processor_id());
929                 perf_enable();
930         }
931
932         if (err) {
933                 /*
934                  * If this event can't go on and it's part of a
935                  * group, then the whole group has to come off.
936                  */
937                 if (leader != event)
938                         group_sched_out(leader, cpuctx, ctx);
939                 if (leader->attr.pinned) {
940                         update_group_times(leader);
941                         leader->state = PERF_EVENT_STATE_ERROR;
942                 }
943         }
944
945  unlock:
946         spin_unlock(&ctx->lock);
947 }
948
949 /*
950  * Enable a event.
951  *
952  * If event->ctx is a cloned context, callers must make sure that
953  * every task struct that event->ctx->task could possibly point to
954  * remains valid.  This condition is satisfied when called through
955  * perf_event_for_each_child or perf_event_for_each as described
956  * for perf_event_disable.
957  */
958 static void perf_event_enable(struct perf_event *event)
959 {
960         struct perf_event_context *ctx = event->ctx;
961         struct task_struct *task = ctx->task;
962
963         if (!task) {
964                 /*
965                  * Enable the event on the cpu that it's on
966                  */
967                 smp_call_function_single(event->cpu, __perf_event_enable,
968                                          event, 1);
969                 return;
970         }
971
972         spin_lock_irq(&ctx->lock);
973         if (event->state >= PERF_EVENT_STATE_INACTIVE)
974                 goto out;
975
976         /*
977          * If the event is in error state, clear that first.
978          * That way, if we see the event in error state below, we
979          * know that it has gone back into error state, as distinct
980          * from the task having been scheduled away before the
981          * cross-call arrived.
982          */
983         if (event->state == PERF_EVENT_STATE_ERROR)
984                 event->state = PERF_EVENT_STATE_OFF;
985
986  retry:
987         spin_unlock_irq(&ctx->lock);
988         task_oncpu_function_call(task, __perf_event_enable, event);
989
990         spin_lock_irq(&ctx->lock);
991
992         /*
993          * If the context is active and the event is still off,
994          * we need to retry the cross-call.
995          */
996         if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
997                 goto retry;
998
999         /*
1000          * Since we have the lock this context can't be scheduled
1001          * in, so we can change the state safely.
1002          */
1003         if (event->state == PERF_EVENT_STATE_OFF)
1004                 __perf_event_mark_enabled(event, ctx);
1005
1006  out:
1007         spin_unlock_irq(&ctx->lock);
1008 }
1009
1010 static int perf_event_refresh(struct perf_event *event, int refresh)
1011 {
1012         /*
1013          * not supported on inherited events
1014          */
1015         if (event->attr.inherit)
1016                 return -EINVAL;
1017
1018         atomic_add(refresh, &event->event_limit);
1019         perf_event_enable(event);
1020
1021         return 0;
1022 }
1023
1024 void __perf_event_sched_out(struct perf_event_context *ctx,
1025                               struct perf_cpu_context *cpuctx)
1026 {
1027         struct perf_event *event;
1028
1029         spin_lock(&ctx->lock);
1030         ctx->is_active = 0;
1031         if (likely(!ctx->nr_events))
1032                 goto out;
1033         update_context_time(ctx);
1034
1035         perf_disable();
1036         if (ctx->nr_active)
1037                 list_for_each_entry(event, &ctx->group_list, group_entry)
1038                         group_sched_out(event, cpuctx, ctx);
1039
1040         perf_enable();
1041  out:
1042         spin_unlock(&ctx->lock);
1043 }
1044
1045 /*
1046  * Test whether two contexts are equivalent, i.e. whether they
1047  * have both been cloned from the same version of the same context
1048  * and they both have the same number of enabled events.
1049  * If the number of enabled events is the same, then the set
1050  * of enabled events should be the same, because these are both
1051  * inherited contexts, therefore we can't access individual events
1052  * in them directly with an fd; we can only enable/disable all
1053  * events via prctl, or enable/disable all events in a family
1054  * via ioctl, which will have the same effect on both contexts.
1055  */
1056 static int context_equiv(struct perf_event_context *ctx1,
1057                          struct perf_event_context *ctx2)
1058 {
1059         return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1060                 && ctx1->parent_gen == ctx2->parent_gen
1061                 && !ctx1->pin_count && !ctx2->pin_count;
1062 }
1063
1064 static void __perf_event_read(void *event);
1065
1066 static void __perf_event_sync_stat(struct perf_event *event,
1067                                      struct perf_event *next_event)
1068 {
1069         u64 value;
1070
1071         if (!event->attr.inherit_stat)
1072                 return;
1073
1074         /*
1075          * Update the event value, we cannot use perf_event_read()
1076          * because we're in the middle of a context switch and have IRQs
1077          * disabled, which upsets smp_call_function_single(), however
1078          * we know the event must be on the current CPU, therefore we
1079          * don't need to use it.
1080          */
1081         switch (event->state) {
1082         case PERF_EVENT_STATE_ACTIVE:
1083                 __perf_event_read(event);
1084                 break;
1085
1086         case PERF_EVENT_STATE_INACTIVE:
1087                 update_event_times(event);
1088                 break;
1089
1090         default:
1091                 break;
1092         }
1093
1094         /*
1095          * In order to keep per-task stats reliable we need to flip the event
1096          * values when we flip the contexts.
1097          */
1098         value = atomic64_read(&next_event->count);
1099         value = atomic64_xchg(&event->count, value);
1100         atomic64_set(&next_event->count, value);
1101
1102         swap(event->total_time_enabled, next_event->total_time_enabled);
1103         swap(event->total_time_running, next_event->total_time_running);
1104
1105         /*
1106          * Since we swizzled the values, update the user visible data too.
1107          */
1108         perf_event_update_userpage(event);
1109         perf_event_update_userpage(next_event);
1110 }
1111
1112 #define list_next_entry(pos, member) \
1113         list_entry(pos->member.next, typeof(*pos), member)
1114
1115 static void perf_event_sync_stat(struct perf_event_context *ctx,
1116                                    struct perf_event_context *next_ctx)
1117 {
1118         struct perf_event *event, *next_event;
1119
1120         if (!ctx->nr_stat)
1121                 return;
1122
1123         update_context_time(ctx);
1124
1125         event = list_first_entry(&ctx->event_list,
1126                                    struct perf_event, event_entry);
1127
1128         next_event = list_first_entry(&next_ctx->event_list,
1129                                         struct perf_event, event_entry);
1130
1131         while (&event->event_entry != &ctx->event_list &&
1132                &next_event->event_entry != &next_ctx->event_list) {
1133
1134                 __perf_event_sync_stat(event, next_event);
1135
1136                 event = list_next_entry(event, event_entry);
1137                 next_event = list_next_entry(next_event, event_entry);
1138         }
1139 }
1140
1141 /*
1142  * Called from scheduler to remove the events of the current task,
1143  * with interrupts disabled.
1144  *
1145  * We stop each event and update the event value in event->count.
1146  *
1147  * This does not protect us against NMI, but disable()
1148  * sets the disabled bit in the control field of event _before_
1149  * accessing the event control register. If a NMI hits, then it will
1150  * not restart the event.
1151  */
1152 void perf_event_task_sched_out(struct task_struct *task,
1153                                  struct task_struct *next, int cpu)
1154 {
1155         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1156         struct perf_event_context *ctx = task->perf_event_ctxp;
1157         struct perf_event_context *next_ctx;
1158         struct perf_event_context *parent;
1159         struct pt_regs *regs;
1160         int do_switch = 1;
1161
1162         regs = task_pt_regs(task);
1163         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1164
1165         if (likely(!ctx || !cpuctx->task_ctx))
1166                 return;
1167
1168         rcu_read_lock();
1169         parent = rcu_dereference(ctx->parent_ctx);
1170         next_ctx = next->perf_event_ctxp;
1171         if (parent && next_ctx &&
1172             rcu_dereference(next_ctx->parent_ctx) == parent) {
1173                 /*
1174                  * Looks like the two contexts are clones, so we might be
1175                  * able to optimize the context switch.  We lock both
1176                  * contexts and check that they are clones under the
1177                  * lock (including re-checking that neither has been
1178                  * uncloned in the meantime).  It doesn't matter which
1179                  * order we take the locks because no other cpu could
1180                  * be trying to lock both of these tasks.
1181                  */
1182                 spin_lock(&ctx->lock);
1183                 spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1184                 if (context_equiv(ctx, next_ctx)) {
1185                         /*
1186                          * XXX do we need a memory barrier of sorts
1187                          * wrt to rcu_dereference() of perf_event_ctxp
1188                          */
1189                         task->perf_event_ctxp = next_ctx;
1190                         next->perf_event_ctxp = ctx;
1191                         ctx->task = next;
1192                         next_ctx->task = task;
1193                         do_switch = 0;
1194
1195                         perf_event_sync_stat(ctx, next_ctx);
1196                 }
1197                 spin_unlock(&next_ctx->lock);
1198                 spin_unlock(&ctx->lock);
1199         }
1200         rcu_read_unlock();
1201
1202         if (do_switch) {
1203                 __perf_event_sched_out(ctx, cpuctx);
1204                 cpuctx->task_ctx = NULL;
1205         }
1206 }
1207
1208 /*
1209  * Called with IRQs disabled
1210  */
1211 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1212 {
1213         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1214
1215         if (!cpuctx->task_ctx)
1216                 return;
1217
1218         if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1219                 return;
1220
1221         __perf_event_sched_out(ctx, cpuctx);
1222         cpuctx->task_ctx = NULL;
1223 }
1224
1225 /*
1226  * Called with IRQs disabled
1227  */
1228 static void perf_event_cpu_sched_out(struct perf_cpu_context *cpuctx)
1229 {
1230         __perf_event_sched_out(&cpuctx->ctx, cpuctx);
1231 }
1232
1233 static void
1234 __perf_event_sched_in(struct perf_event_context *ctx,
1235                         struct perf_cpu_context *cpuctx, int cpu)
1236 {
1237         struct perf_event *event;
1238         int can_add_hw = 1;
1239
1240         spin_lock(&ctx->lock);
1241         ctx->is_active = 1;
1242         if (likely(!ctx->nr_events))
1243                 goto out;
1244
1245         ctx->timestamp = perf_clock();
1246
1247         perf_disable();
1248
1249         /*
1250          * First go through the list and put on any pinned groups
1251          * in order to give them the best chance of going on.
1252          */
1253         list_for_each_entry(event, &ctx->group_list, group_entry) {
1254                 if (event->state <= PERF_EVENT_STATE_OFF ||
1255                     !event->attr.pinned)
1256                         continue;
1257                 if (event->cpu != -1 && event->cpu != cpu)
1258                         continue;
1259
1260                 if (group_can_go_on(event, cpuctx, 1))
1261                         group_sched_in(event, cpuctx, ctx, cpu);
1262
1263                 /*
1264                  * If this pinned group hasn't been scheduled,
1265                  * put it in error state.
1266                  */
1267                 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1268                         update_group_times(event);
1269                         event->state = PERF_EVENT_STATE_ERROR;
1270                 }
1271         }
1272
1273         list_for_each_entry(event, &ctx->group_list, group_entry) {
1274                 /*
1275                  * Ignore events in OFF or ERROR state, and
1276                  * ignore pinned events since we did them already.
1277                  */
1278                 if (event->state <= PERF_EVENT_STATE_OFF ||
1279                     event->attr.pinned)
1280                         continue;
1281
1282                 /*
1283                  * Listen to the 'cpu' scheduling filter constraint
1284                  * of events:
1285                  */
1286                 if (event->cpu != -1 && event->cpu != cpu)
1287                         continue;
1288
1289                 if (group_can_go_on(event, cpuctx, can_add_hw))
1290                         if (group_sched_in(event, cpuctx, ctx, cpu))
1291                                 can_add_hw = 0;
1292         }
1293         perf_enable();
1294  out:
1295         spin_unlock(&ctx->lock);
1296 }
1297
1298 /*
1299  * Called from scheduler to add the events of the current task
1300  * with interrupts disabled.
1301  *
1302  * We restore the event value and then enable it.
1303  *
1304  * This does not protect us against NMI, but enable()
1305  * sets the enabled bit in the control field of event _before_
1306  * accessing the event control register. If a NMI hits, then it will
1307  * keep the event running.
1308  */
1309 void perf_event_task_sched_in(struct task_struct *task, int cpu)
1310 {
1311         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1312         struct perf_event_context *ctx = task->perf_event_ctxp;
1313
1314         if (likely(!ctx))
1315                 return;
1316         if (cpuctx->task_ctx == ctx)
1317                 return;
1318         __perf_event_sched_in(ctx, cpuctx, cpu);
1319         cpuctx->task_ctx = ctx;
1320 }
1321
1322 static void perf_event_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
1323 {
1324         struct perf_event_context *ctx = &cpuctx->ctx;
1325
1326         __perf_event_sched_in(ctx, cpuctx, cpu);
1327 }
1328
1329 #define MAX_INTERRUPTS (~0ULL)
1330
1331 static void perf_log_throttle(struct perf_event *event, int enable);
1332
1333 static void perf_adjust_period(struct perf_event *event, u64 events)
1334 {
1335         struct hw_perf_event *hwc = &event->hw;
1336         u64 period, sample_period;
1337         s64 delta;
1338
1339         events *= hwc->sample_period;
1340         period = div64_u64(events, event->attr.sample_freq);
1341
1342         delta = (s64)(period - hwc->sample_period);
1343         delta = (delta + 7) / 8; /* low pass filter */
1344
1345         sample_period = hwc->sample_period + delta;
1346
1347         if (!sample_period)
1348                 sample_period = 1;
1349
1350         hwc->sample_period = sample_period;
1351 }
1352
1353 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1354 {
1355         struct perf_event *event;
1356         struct hw_perf_event *hwc;
1357         u64 interrupts, freq;
1358
1359         spin_lock(&ctx->lock);
1360         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1361                 if (event->state != PERF_EVENT_STATE_ACTIVE)
1362                         continue;
1363
1364                 hwc = &event->hw;
1365
1366                 interrupts = hwc->interrupts;
1367                 hwc->interrupts = 0;
1368
1369                 /*
1370                  * unthrottle events on the tick
1371                  */
1372                 if (interrupts == MAX_INTERRUPTS) {
1373                         perf_log_throttle(event, 1);
1374                         event->pmu->unthrottle(event);
1375                         interrupts = 2*sysctl_perf_event_sample_rate/HZ;
1376                 }
1377
1378                 if (!event->attr.freq || !event->attr.sample_freq)
1379                         continue;
1380
1381                 /*
1382                  * if the specified freq < HZ then we need to skip ticks
1383                  */
1384                 if (event->attr.sample_freq < HZ) {
1385                         freq = event->attr.sample_freq;
1386
1387                         hwc->freq_count += freq;
1388                         hwc->freq_interrupts += interrupts;
1389
1390                         if (hwc->freq_count < HZ)
1391                                 continue;
1392
1393                         interrupts = hwc->freq_interrupts;
1394                         hwc->freq_interrupts = 0;
1395                         hwc->freq_count -= HZ;
1396                 } else
1397                         freq = HZ;
1398
1399                 perf_adjust_period(event, freq * interrupts);
1400
1401                 /*
1402                  * In order to avoid being stalled by an (accidental) huge
1403                  * sample period, force reset the sample period if we didn't
1404                  * get any events in this freq period.
1405                  */
1406                 if (!interrupts) {
1407                         perf_disable();
1408                         event->pmu->disable(event);
1409                         atomic64_set(&hwc->period_left, 0);
1410                         event->pmu->enable(event);
1411                         perf_enable();
1412                 }
1413         }
1414         spin_unlock(&ctx->lock);
1415 }
1416
1417 /*
1418  * Round-robin a context's events:
1419  */
1420 static void rotate_ctx(struct perf_event_context *ctx)
1421 {
1422         struct perf_event *event;
1423
1424         if (!ctx->nr_events)
1425                 return;
1426
1427         spin_lock(&ctx->lock);
1428         /*
1429          * Rotate the first entry last (works just fine for group events too):
1430          */
1431         perf_disable();
1432         list_for_each_entry(event, &ctx->group_list, group_entry) {
1433                 list_move_tail(&event->group_entry, &ctx->group_list);
1434                 break;
1435         }
1436         perf_enable();
1437
1438         spin_unlock(&ctx->lock);
1439 }
1440
1441 void perf_event_task_tick(struct task_struct *curr, int cpu)
1442 {
1443         struct perf_cpu_context *cpuctx;
1444         struct perf_event_context *ctx;
1445
1446         if (!atomic_read(&nr_events))
1447                 return;
1448
1449         cpuctx = &per_cpu(perf_cpu_context, cpu);
1450         ctx = curr->perf_event_ctxp;
1451
1452         perf_ctx_adjust_freq(&cpuctx->ctx);
1453         if (ctx)
1454                 perf_ctx_adjust_freq(ctx);
1455
1456         perf_event_cpu_sched_out(cpuctx);
1457         if (ctx)
1458                 __perf_event_task_sched_out(ctx);
1459
1460         rotate_ctx(&cpuctx->ctx);
1461         if (ctx)
1462                 rotate_ctx(ctx);
1463
1464         perf_event_cpu_sched_in(cpuctx, cpu);
1465         if (ctx)
1466                 perf_event_task_sched_in(curr, cpu);
1467 }
1468
1469 /*
1470  * Enable all of a task's events that have been marked enable-on-exec.
1471  * This expects task == current.
1472  */
1473 static void perf_event_enable_on_exec(struct task_struct *task)
1474 {
1475         struct perf_event_context *ctx;
1476         struct perf_event *event;
1477         unsigned long flags;
1478         int enabled = 0;
1479
1480         local_irq_save(flags);
1481         ctx = task->perf_event_ctxp;
1482         if (!ctx || !ctx->nr_events)
1483                 goto out;
1484
1485         __perf_event_task_sched_out(ctx);
1486
1487         spin_lock(&ctx->lock);
1488
1489         list_for_each_entry(event, &ctx->group_list, group_entry) {
1490                 if (!event->attr.enable_on_exec)
1491                         continue;
1492                 event->attr.enable_on_exec = 0;
1493                 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1494                         continue;
1495                 __perf_event_mark_enabled(event, ctx);
1496                 enabled = 1;
1497         }
1498
1499         /*
1500          * Unclone this context if we enabled any event.
1501          */
1502         if (enabled)
1503                 unclone_ctx(ctx);
1504
1505         spin_unlock(&ctx->lock);
1506
1507         perf_event_task_sched_in(task, smp_processor_id());
1508  out:
1509         local_irq_restore(flags);
1510 }
1511
1512 /*
1513  * Cross CPU call to read the hardware event
1514  */
1515 static void __perf_event_read(void *info)
1516 {
1517         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1518         struct perf_event *event = info;
1519         struct perf_event_context *ctx = event->ctx;
1520
1521         /*
1522          * If this is a task context, we need to check whether it is
1523          * the current task context of this cpu.  If not it has been
1524          * scheduled out before the smp call arrived.  In that case
1525          * event->count would have been updated to a recent sample
1526          * when the event was scheduled out.
1527          */
1528         if (ctx->task && cpuctx->task_ctx != ctx)
1529                 return;
1530
1531         if (ctx->is_active)
1532                 update_context_time(ctx);
1533         event->pmu->read(event);
1534         update_event_times(event);
1535 }
1536
1537 static u64 perf_event_read(struct perf_event *event)
1538 {
1539         /*
1540          * If event is enabled and currently active on a CPU, update the
1541          * value in the event structure:
1542          */
1543         if (event->state == PERF_EVENT_STATE_ACTIVE) {
1544                 smp_call_function_single(event->oncpu,
1545                                          __perf_event_read, event, 1);
1546         } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1547                 update_event_times(event);
1548         }
1549
1550         return atomic64_read(&event->count);
1551 }
1552
1553 /*
1554  * Initialize the perf_event context in a task_struct:
1555  */
1556 static void
1557 __perf_event_init_context(struct perf_event_context *ctx,
1558                             struct task_struct *task)
1559 {
1560         memset(ctx, 0, sizeof(*ctx));
1561         spin_lock_init(&ctx->lock);
1562         mutex_init(&ctx->mutex);
1563         INIT_LIST_HEAD(&ctx->group_list);
1564         INIT_LIST_HEAD(&ctx->event_list);
1565         atomic_set(&ctx->refcount, 1);
1566         ctx->task = task;
1567 }
1568
1569 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1570 {
1571         struct perf_event_context *ctx;
1572         struct perf_cpu_context *cpuctx;
1573         struct task_struct *task;
1574         unsigned long flags;
1575         int err;
1576
1577         /*
1578          * If cpu is not a wildcard then this is a percpu event:
1579          */
1580         if (cpu != -1) {
1581                 /* Must be root to operate on a CPU event: */
1582                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1583                         return ERR_PTR(-EACCES);
1584
1585                 if (cpu < 0 || cpu > num_possible_cpus())
1586                         return ERR_PTR(-EINVAL);
1587
1588                 /*
1589                  * We could be clever and allow to attach a event to an
1590                  * offline CPU and activate it when the CPU comes up, but
1591                  * that's for later.
1592                  */
1593                 if (!cpu_isset(cpu, cpu_online_map))
1594                         return ERR_PTR(-ENODEV);
1595
1596                 cpuctx = &per_cpu(perf_cpu_context, cpu);
1597                 ctx = &cpuctx->ctx;
1598                 get_ctx(ctx);
1599
1600                 return ctx;
1601         }
1602
1603         rcu_read_lock();
1604         if (!pid)
1605                 task = current;
1606         else
1607                 task = find_task_by_vpid(pid);
1608         if (task)
1609                 get_task_struct(task);
1610         rcu_read_unlock();
1611
1612         if (!task)
1613                 return ERR_PTR(-ESRCH);
1614
1615         /*
1616          * Can't attach events to a dying task.
1617          */
1618         err = -ESRCH;
1619         if (task->flags & PF_EXITING)
1620                 goto errout;
1621
1622         /* Reuse ptrace permission checks for now. */
1623         err = -EACCES;
1624         if (!ptrace_may_access(task, PTRACE_MODE_READ))
1625                 goto errout;
1626
1627  retry:
1628         ctx = perf_lock_task_context(task, &flags);
1629         if (ctx) {
1630                 unclone_ctx(ctx);
1631                 spin_unlock_irqrestore(&ctx->lock, flags);
1632         }
1633
1634         if (!ctx) {
1635                 ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1636                 err = -ENOMEM;
1637                 if (!ctx)
1638                         goto errout;
1639                 __perf_event_init_context(ctx, task);
1640                 get_ctx(ctx);
1641                 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1642                         /*
1643                          * We raced with some other task; use
1644                          * the context they set.
1645                          */
1646                         kfree(ctx);
1647                         goto retry;
1648                 }
1649                 get_task_struct(task);
1650         }
1651
1652         put_task_struct(task);
1653         return ctx;
1654
1655  errout:
1656         put_task_struct(task);
1657         return ERR_PTR(err);
1658 }
1659
1660 static void perf_event_free_filter(struct perf_event *event);
1661
1662 static void free_event_rcu(struct rcu_head *head)
1663 {
1664         struct perf_event *event;
1665
1666         event = container_of(head, struct perf_event, rcu_head);
1667         if (event->ns)
1668                 put_pid_ns(event->ns);
1669         perf_event_free_filter(event);
1670         kfree(event);
1671 }
1672
1673 static void perf_pending_sync(struct perf_event *event);
1674
1675 static void free_event(struct perf_event *event)
1676 {
1677         perf_pending_sync(event);
1678
1679         if (!event->parent) {
1680                 atomic_dec(&nr_events);
1681                 if (event->attr.mmap)
1682                         atomic_dec(&nr_mmap_events);
1683                 if (event->attr.comm)
1684                         atomic_dec(&nr_comm_events);
1685                 if (event->attr.task)
1686                         atomic_dec(&nr_task_events);
1687         }
1688
1689         if (event->output) {
1690                 fput(event->output->filp);
1691                 event->output = NULL;
1692         }
1693
1694         if (event->destroy)
1695                 event->destroy(event);
1696
1697         put_ctx(event->ctx);
1698         call_rcu(&event->rcu_head, free_event_rcu);
1699 }
1700
1701 /*
1702  * Called when the last reference to the file is gone.
1703  */
1704 static int perf_release(struct inode *inode, struct file *file)
1705 {
1706         struct perf_event *event = file->private_data;
1707         struct perf_event_context *ctx = event->ctx;
1708
1709         file->private_data = NULL;
1710
1711         WARN_ON_ONCE(ctx->parent_ctx);
1712         mutex_lock(&ctx->mutex);
1713         perf_event_remove_from_context(event);
1714         mutex_unlock(&ctx->mutex);
1715
1716         mutex_lock(&event->owner->perf_event_mutex);
1717         list_del_init(&event->owner_entry);
1718         mutex_unlock(&event->owner->perf_event_mutex);
1719         put_task_struct(event->owner);
1720
1721         free_event(event);
1722
1723         return 0;
1724 }
1725
1726 int perf_event_release_kernel(struct perf_event *event)
1727 {
1728         struct perf_event_context *ctx = event->ctx;
1729
1730         WARN_ON_ONCE(ctx->parent_ctx);
1731         mutex_lock(&ctx->mutex);
1732         perf_event_remove_from_context(event);
1733         mutex_unlock(&ctx->mutex);
1734
1735         mutex_lock(&event->owner->perf_event_mutex);
1736         list_del_init(&event->owner_entry);
1737         mutex_unlock(&event->owner->perf_event_mutex);
1738         put_task_struct(event->owner);
1739
1740         free_event(event);
1741
1742         return 0;
1743 }
1744 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1745
1746 static int perf_event_read_size(struct perf_event *event)
1747 {
1748         int entry = sizeof(u64); /* value */
1749         int size = 0;
1750         int nr = 1;
1751
1752         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1753                 size += sizeof(u64);
1754
1755         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1756                 size += sizeof(u64);
1757
1758         if (event->attr.read_format & PERF_FORMAT_ID)
1759                 entry += sizeof(u64);
1760
1761         if (event->attr.read_format & PERF_FORMAT_GROUP) {
1762                 nr += event->group_leader->nr_siblings;
1763                 size += sizeof(u64);
1764         }
1765
1766         size += entry * nr;
1767
1768         return size;
1769 }
1770
1771 u64 perf_event_read_value(struct perf_event *event)
1772 {
1773         struct perf_event *child;
1774         u64 total = 0;
1775
1776         total += perf_event_read(event);
1777         list_for_each_entry(child, &event->child_list, child_list)
1778                 total += perf_event_read(child);
1779
1780         return total;
1781 }
1782 EXPORT_SYMBOL_GPL(perf_event_read_value);
1783
1784 static int perf_event_read_group(struct perf_event *event,
1785                                    u64 read_format, char __user *buf)
1786 {
1787         struct perf_event *leader = event->group_leader, *sub;
1788         int n = 0, size = 0, ret = 0;
1789         u64 values[5];
1790         u64 count;
1791
1792         count = perf_event_read_value(leader);
1793
1794         values[n++] = 1 + leader->nr_siblings;
1795         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1796                 values[n++] = leader->total_time_enabled +
1797                         atomic64_read(&leader->child_total_time_enabled);
1798         }
1799         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1800                 values[n++] = leader->total_time_running +
1801                         atomic64_read(&leader->child_total_time_running);
1802         }
1803         values[n++] = count;
1804         if (read_format & PERF_FORMAT_ID)
1805                 values[n++] = primary_event_id(leader);
1806
1807         size = n * sizeof(u64);
1808
1809         if (copy_to_user(buf, values, size))
1810                 return -EFAULT;
1811
1812         ret += size;
1813
1814         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1815                 n = 0;
1816
1817                 values[n++] = perf_event_read_value(sub);
1818                 if (read_format & PERF_FORMAT_ID)
1819                         values[n++] = primary_event_id(sub);
1820
1821                 size = n * sizeof(u64);
1822
1823                 if (copy_to_user(buf + size, values, size))
1824                         return -EFAULT;
1825
1826                 ret += size;
1827         }
1828
1829         return ret;
1830 }
1831
1832 static int perf_event_read_one(struct perf_event *event,
1833                                  u64 read_format, char __user *buf)
1834 {
1835         u64 values[4];
1836         int n = 0;
1837
1838         values[n++] = perf_event_read_value(event);
1839         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1840                 values[n++] = event->total_time_enabled +
1841                         atomic64_read(&event->child_total_time_enabled);
1842         }
1843         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1844                 values[n++] = event->total_time_running +
1845                         atomic64_read(&event->child_total_time_running);
1846         }
1847         if (read_format & PERF_FORMAT_ID)
1848                 values[n++] = primary_event_id(event);
1849
1850         if (copy_to_user(buf, values, n * sizeof(u64)))
1851                 return -EFAULT;
1852
1853         return n * sizeof(u64);
1854 }
1855
1856 /*
1857  * Read the performance event - simple non blocking version for now
1858  */
1859 static ssize_t
1860 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
1861 {
1862         u64 read_format = event->attr.read_format;
1863         int ret;
1864
1865         /*
1866          * Return end-of-file for a read on a event that is in
1867          * error state (i.e. because it was pinned but it couldn't be
1868          * scheduled on to the CPU at some point).
1869          */
1870         if (event->state == PERF_EVENT_STATE_ERROR)
1871                 return 0;
1872
1873         if (count < perf_event_read_size(event))
1874                 return -ENOSPC;
1875
1876         WARN_ON_ONCE(event->ctx->parent_ctx);
1877         mutex_lock(&event->child_mutex);
1878         if (read_format & PERF_FORMAT_GROUP)
1879                 ret = perf_event_read_group(event, read_format, buf);
1880         else
1881                 ret = perf_event_read_one(event, read_format, buf);
1882         mutex_unlock(&event->child_mutex);
1883
1884         return ret;
1885 }
1886
1887 static ssize_t
1888 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1889 {
1890         struct perf_event *event = file->private_data;
1891
1892         return perf_read_hw(event, buf, count);
1893 }
1894
1895 static unsigned int perf_poll(struct file *file, poll_table *wait)
1896 {
1897         struct perf_event *event = file->private_data;
1898         struct perf_mmap_data *data;
1899         unsigned int events = POLL_HUP;
1900
1901         rcu_read_lock();
1902         data = rcu_dereference(event->data);
1903         if (data)
1904                 events = atomic_xchg(&data->poll, 0);
1905         rcu_read_unlock();
1906
1907         poll_wait(file, &event->waitq, wait);
1908
1909         return events;
1910 }
1911
1912 static void perf_event_reset(struct perf_event *event)
1913 {
1914         (void)perf_event_read(event);
1915         atomic64_set(&event->count, 0);
1916         perf_event_update_userpage(event);
1917 }
1918
1919 /*
1920  * Holding the top-level event's child_mutex means that any
1921  * descendant process that has inherited this event will block
1922  * in sync_child_event if it goes to exit, thus satisfying the
1923  * task existence requirements of perf_event_enable/disable.
1924  */
1925 static void perf_event_for_each_child(struct perf_event *event,
1926                                         void (*func)(struct perf_event *))
1927 {
1928         struct perf_event *child;
1929
1930         WARN_ON_ONCE(event->ctx->parent_ctx);
1931         mutex_lock(&event->child_mutex);
1932         func(event);
1933         list_for_each_entry(child, &event->child_list, child_list)
1934                 func(child);
1935         mutex_unlock(&event->child_mutex);
1936 }
1937
1938 static void perf_event_for_each(struct perf_event *event,
1939                                   void (*func)(struct perf_event *))
1940 {
1941         struct perf_event_context *ctx = event->ctx;
1942         struct perf_event *sibling;
1943
1944         WARN_ON_ONCE(ctx->parent_ctx);
1945         mutex_lock(&ctx->mutex);
1946         event = event->group_leader;
1947
1948         perf_event_for_each_child(event, func);
1949         func(event);
1950         list_for_each_entry(sibling, &event->sibling_list, group_entry)
1951                 perf_event_for_each_child(event, func);
1952         mutex_unlock(&ctx->mutex);
1953 }
1954
1955 static int perf_event_period(struct perf_event *event, u64 __user *arg)
1956 {
1957         struct perf_event_context *ctx = event->ctx;
1958         unsigned long size;
1959         int ret = 0;
1960         u64 value;
1961
1962         if (!event->attr.sample_period)
1963                 return -EINVAL;
1964
1965         size = copy_from_user(&value, arg, sizeof(value));
1966         if (size != sizeof(value))
1967                 return -EFAULT;
1968
1969         if (!value)
1970                 return -EINVAL;
1971
1972         spin_lock_irq(&ctx->lock);
1973         if (event->attr.freq) {
1974                 if (value > sysctl_perf_event_sample_rate) {
1975                         ret = -EINVAL;
1976                         goto unlock;
1977                 }
1978
1979                 event->attr.sample_freq = value;
1980         } else {
1981                 event->attr.sample_period = value;
1982                 event->hw.sample_period = value;
1983         }
1984 unlock:
1985         spin_unlock_irq(&ctx->lock);
1986
1987         return ret;
1988 }
1989
1990 static int perf_event_set_output(struct perf_event *event, int output_fd);
1991 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
1992
1993 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1994 {
1995         struct perf_event *event = file->private_data;
1996         void (*func)(struct perf_event *);
1997         u32 flags = arg;
1998
1999         switch (cmd) {
2000         case PERF_EVENT_IOC_ENABLE:
2001                 func = perf_event_enable;
2002                 break;
2003         case PERF_EVENT_IOC_DISABLE:
2004                 func = perf_event_disable;
2005                 break;
2006         case PERF_EVENT_IOC_RESET:
2007                 func = perf_event_reset;
2008                 break;
2009
2010         case PERF_EVENT_IOC_REFRESH:
2011                 return perf_event_refresh(event, arg);
2012
2013         case PERF_EVENT_IOC_PERIOD:
2014                 return perf_event_period(event, (u64 __user *)arg);
2015
2016         case PERF_EVENT_IOC_SET_OUTPUT:
2017                 return perf_event_set_output(event, arg);
2018
2019         case PERF_EVENT_IOC_SET_FILTER:
2020                 return perf_event_set_filter(event, (void __user *)arg);
2021
2022         default:
2023                 return -ENOTTY;
2024         }
2025
2026         if (flags & PERF_IOC_FLAG_GROUP)
2027                 perf_event_for_each(event, func);
2028         else
2029                 perf_event_for_each_child(event, func);
2030
2031         return 0;
2032 }
2033
2034 int perf_event_task_enable(void)
2035 {
2036         struct perf_event *event;
2037
2038         mutex_lock(&current->perf_event_mutex);
2039         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2040                 perf_event_for_each_child(event, perf_event_enable);
2041         mutex_unlock(&current->perf_event_mutex);
2042
2043         return 0;
2044 }
2045
2046 int perf_event_task_disable(void)
2047 {
2048         struct perf_event *event;
2049
2050         mutex_lock(&current->perf_event_mutex);
2051         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2052                 perf_event_for_each_child(event, perf_event_disable);
2053         mutex_unlock(&current->perf_event_mutex);
2054
2055         return 0;
2056 }
2057
2058 #ifndef PERF_EVENT_INDEX_OFFSET
2059 # define PERF_EVENT_INDEX_OFFSET 0
2060 #endif
2061
2062 static int perf_event_index(struct perf_event *event)
2063 {
2064         if (event->state != PERF_EVENT_STATE_ACTIVE)
2065                 return 0;
2066
2067         return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2068 }
2069
2070 /*
2071  * Callers need to ensure there can be no nesting of this function, otherwise
2072  * the seqlock logic goes bad. We can not serialize this because the arch
2073  * code calls this from NMI context.
2074  */
2075 void perf_event_update_userpage(struct perf_event *event)
2076 {
2077         struct perf_event_mmap_page *userpg;
2078         struct perf_mmap_data *data;
2079
2080         rcu_read_lock();
2081         data = rcu_dereference(event->data);
2082         if (!data)
2083                 goto unlock;
2084
2085         userpg = data->user_page;
2086
2087         /*
2088          * Disable preemption so as to not let the corresponding user-space
2089          * spin too long if we get preempted.
2090          */
2091         preempt_disable();
2092         ++userpg->lock;
2093         barrier();
2094         userpg->index = perf_event_index(event);
2095         userpg->offset = atomic64_read(&event->count);
2096         if (event->state == PERF_EVENT_STATE_ACTIVE)
2097                 userpg->offset -= atomic64_read(&event->hw.prev_count);
2098
2099         userpg->time_enabled = event->total_time_enabled +
2100                         atomic64_read(&event->child_total_time_enabled);
2101
2102         userpg->time_running = event->total_time_running +
2103                         atomic64_read(&event->child_total_time_running);
2104
2105         barrier();
2106         ++userpg->lock;
2107         preempt_enable();
2108 unlock:
2109         rcu_read_unlock();
2110 }
2111
2112 static unsigned long perf_data_size(struct perf_mmap_data *data)
2113 {
2114         return data->nr_pages << (PAGE_SHIFT + data->data_order);
2115 }
2116
2117 #ifndef CONFIG_PERF_USE_VMALLOC
2118
2119 /*
2120  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2121  */
2122
2123 static struct page *
2124 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2125 {
2126         if (pgoff > data->nr_pages)
2127                 return NULL;
2128
2129         if (pgoff == 0)
2130                 return virt_to_page(data->user_page);
2131
2132         return virt_to_page(data->data_pages[pgoff - 1]);
2133 }
2134
2135 static struct perf_mmap_data *
2136 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2137 {
2138         struct perf_mmap_data *data;
2139         unsigned long size;
2140         int i;
2141
2142         WARN_ON(atomic_read(&event->mmap_count));
2143
2144         size = sizeof(struct perf_mmap_data);
2145         size += nr_pages * sizeof(void *);
2146
2147         data = kzalloc(size, GFP_KERNEL);
2148         if (!data)
2149                 goto fail;
2150
2151         data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2152         if (!data->user_page)
2153                 goto fail_user_page;
2154
2155         for (i = 0; i < nr_pages; i++) {
2156                 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2157                 if (!data->data_pages[i])
2158                         goto fail_data_pages;
2159         }
2160
2161         data->data_order = 0;
2162         data->nr_pages = nr_pages;
2163
2164         return data;
2165
2166 fail_data_pages:
2167         for (i--; i >= 0; i--)
2168                 free_page((unsigned long)data->data_pages[i]);
2169
2170         free_page((unsigned long)data->user_page);
2171
2172 fail_user_page:
2173         kfree(data);
2174
2175 fail:
2176         return NULL;
2177 }
2178
2179 static void perf_mmap_free_page(unsigned long addr)
2180 {
2181         struct page *page = virt_to_page((void *)addr);
2182
2183         page->mapping = NULL;
2184         __free_page(page);
2185 }
2186
2187 static void perf_mmap_data_free(struct perf_mmap_data *data)
2188 {
2189         int i;
2190
2191         perf_mmap_free_page((unsigned long)data->user_page);
2192         for (i = 0; i < data->nr_pages; i++)
2193                 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2194 }
2195
2196 #else
2197
2198 /*
2199  * Back perf_mmap() with vmalloc memory.
2200  *
2201  * Required for architectures that have d-cache aliasing issues.
2202  */
2203
2204 static struct page *
2205 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2206 {
2207         if (pgoff > (1UL << data->data_order))
2208                 return NULL;
2209
2210         return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2211 }
2212
2213 static void perf_mmap_unmark_page(void *addr)
2214 {
2215         struct page *page = vmalloc_to_page(addr);
2216
2217         page->mapping = NULL;
2218 }
2219
2220 static void perf_mmap_data_free_work(struct work_struct *work)
2221 {
2222         struct perf_mmap_data *data;
2223         void *base;
2224         int i, nr;
2225
2226         data = container_of(work, struct perf_mmap_data, work);
2227         nr = 1 << data->data_order;
2228
2229         base = data->user_page;
2230         for (i = 0; i < nr + 1; i++)
2231                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2232
2233         vfree(base);
2234 }
2235
2236 static void perf_mmap_data_free(struct perf_mmap_data *data)
2237 {
2238         schedule_work(&data->work);
2239 }
2240
2241 static struct perf_mmap_data *
2242 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2243 {
2244         struct perf_mmap_data *data;
2245         unsigned long size;
2246         void *all_buf;
2247
2248         WARN_ON(atomic_read(&event->mmap_count));
2249
2250         size = sizeof(struct perf_mmap_data);
2251         size += sizeof(void *);
2252
2253         data = kzalloc(size, GFP_KERNEL);
2254         if (!data)
2255                 goto fail;
2256
2257         INIT_WORK(&data->work, perf_mmap_data_free_work);
2258
2259         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2260         if (!all_buf)
2261                 goto fail_all_buf;
2262
2263         data->user_page = all_buf;
2264         data->data_pages[0] = all_buf + PAGE_SIZE;
2265         data->data_order = ilog2(nr_pages);
2266         data->nr_pages = 1;
2267
2268         return data;
2269
2270 fail_all_buf:
2271         kfree(data);
2272
2273 fail:
2274         return NULL;
2275 }
2276
2277 #endif
2278
2279 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2280 {
2281         struct perf_event *event = vma->vm_file->private_data;
2282         struct perf_mmap_data *data;
2283         int ret = VM_FAULT_SIGBUS;
2284
2285         if (vmf->flags & FAULT_FLAG_MKWRITE) {
2286                 if (vmf->pgoff == 0)
2287                         ret = 0;
2288                 return ret;
2289         }
2290
2291         rcu_read_lock();
2292         data = rcu_dereference(event->data);
2293         if (!data)
2294                 goto unlock;
2295
2296         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2297                 goto unlock;
2298
2299         vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2300         if (!vmf->page)
2301                 goto unlock;
2302
2303         get_page(vmf->page);
2304         vmf->page->mapping = vma->vm_file->f_mapping;
2305         vmf->page->index   = vmf->pgoff;
2306
2307         ret = 0;
2308 unlock:
2309         rcu_read_unlock();
2310
2311         return ret;
2312 }
2313
2314 static void
2315 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2316 {
2317         long max_size = perf_data_size(data);
2318
2319         atomic_set(&data->lock, -1);
2320
2321         if (event->attr.watermark) {
2322                 data->watermark = min_t(long, max_size,
2323                                         event->attr.wakeup_watermark);
2324         }
2325
2326         if (!data->watermark)
2327                 data->watermark = max_t(long, PAGE_SIZE, max_size / 2);
2328
2329
2330         rcu_assign_pointer(event->data, data);
2331 }
2332
2333 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2334 {
2335         struct perf_mmap_data *data;
2336
2337         data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2338         perf_mmap_data_free(data);
2339         kfree(data);
2340 }
2341
2342 static void perf_mmap_data_release(struct perf_event *event)
2343 {
2344         struct perf_mmap_data *data = event->data;
2345
2346         WARN_ON(atomic_read(&event->mmap_count));
2347
2348         rcu_assign_pointer(event->data, NULL);
2349         call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2350 }
2351
2352 static void perf_mmap_open(struct vm_area_struct *vma)
2353 {
2354         struct perf_event *event = vma->vm_file->private_data;
2355
2356         atomic_inc(&event->mmap_count);
2357 }
2358
2359 static void perf_mmap_close(struct vm_area_struct *vma)
2360 {
2361         struct perf_event *event = vma->vm_file->private_data;
2362
2363         WARN_ON_ONCE(event->ctx->parent_ctx);
2364         if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2365                 unsigned long size = perf_data_size(event->data);
2366                 struct user_struct *user = current_user();
2367
2368                 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2369                 vma->vm_mm->locked_vm -= event->data->nr_locked;
2370                 perf_mmap_data_release(event);
2371                 mutex_unlock(&event->mmap_mutex);
2372         }
2373 }
2374
2375 static const struct vm_operations_struct perf_mmap_vmops = {
2376         .open           = perf_mmap_open,
2377         .close          = perf_mmap_close,
2378         .fault          = perf_mmap_fault,
2379         .page_mkwrite   = perf_mmap_fault,
2380 };
2381
2382 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2383 {
2384         struct perf_event *event = file->private_data;
2385         unsigned long user_locked, user_lock_limit;
2386         struct user_struct *user = current_user();
2387         unsigned long locked, lock_limit;
2388         struct perf_mmap_data *data;
2389         unsigned long vma_size;
2390         unsigned long nr_pages;
2391         long user_extra, extra;
2392         int ret = 0;
2393
2394         if (!(vma->vm_flags & VM_SHARED))
2395                 return -EINVAL;
2396
2397         vma_size = vma->vm_end - vma->vm_start;
2398         nr_pages = (vma_size / PAGE_SIZE) - 1;
2399
2400         /*
2401          * If we have data pages ensure they're a power-of-two number, so we
2402          * can do bitmasks instead of modulo.
2403          */
2404         if (nr_pages != 0 && !is_power_of_2(nr_pages))
2405                 return -EINVAL;
2406
2407         if (vma_size != PAGE_SIZE * (1 + nr_pages))
2408                 return -EINVAL;
2409
2410         if (vma->vm_pgoff != 0)
2411                 return -EINVAL;
2412
2413         WARN_ON_ONCE(event->ctx->parent_ctx);
2414         mutex_lock(&event->mmap_mutex);
2415         if (event->output) {
2416                 ret = -EINVAL;
2417                 goto unlock;
2418         }
2419
2420         if (atomic_inc_not_zero(&event->mmap_count)) {
2421                 if (nr_pages != event->data->nr_pages)
2422                         ret = -EINVAL;
2423                 goto unlock;
2424         }
2425
2426         user_extra = nr_pages + 1;
2427         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2428
2429         /*
2430          * Increase the limit linearly with more CPUs:
2431          */
2432         user_lock_limit *= num_online_cpus();
2433
2434         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2435
2436         extra = 0;
2437         if (user_locked > user_lock_limit)
2438                 extra = user_locked - user_lock_limit;
2439
2440         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2441         lock_limit >>= PAGE_SHIFT;
2442         locked = vma->vm_mm->locked_vm + extra;
2443
2444         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2445                 !capable(CAP_IPC_LOCK)) {
2446                 ret = -EPERM;
2447                 goto unlock;
2448         }
2449
2450         WARN_ON(event->data);
2451
2452         data = perf_mmap_data_alloc(event, nr_pages);
2453         ret = -ENOMEM;
2454         if (!data)
2455                 goto unlock;
2456
2457         ret = 0;
2458         perf_mmap_data_init(event, data);
2459
2460         atomic_set(&event->mmap_count, 1);
2461         atomic_long_add(user_extra, &user->locked_vm);
2462         vma->vm_mm->locked_vm += extra;
2463         event->data->nr_locked = extra;
2464         if (vma->vm_flags & VM_WRITE)
2465                 event->data->writable = 1;
2466
2467 unlock:
2468         mutex_unlock(&event->mmap_mutex);
2469
2470         vma->vm_flags |= VM_RESERVED;
2471         vma->vm_ops = &perf_mmap_vmops;
2472
2473         return ret;
2474 }
2475
2476 static int perf_fasync(int fd, struct file *filp, int on)
2477 {
2478         struct inode *inode = filp->f_path.dentry->d_inode;
2479         struct perf_event *event = filp->private_data;
2480         int retval;
2481
2482         mutex_lock(&inode->i_mutex);
2483         retval = fasync_helper(fd, filp, on, &event->fasync);
2484         mutex_unlock(&inode->i_mutex);
2485
2486         if (retval < 0)
2487                 return retval;
2488
2489         return 0;
2490 }
2491
2492 static const struct file_operations perf_fops = {
2493         .release                = perf_release,
2494         .read                   = perf_read,
2495         .poll                   = perf_poll,
2496         .unlocked_ioctl         = perf_ioctl,
2497         .compat_ioctl           = perf_ioctl,
2498         .mmap                   = perf_mmap,
2499         .fasync                 = perf_fasync,
2500 };
2501
2502 /*
2503  * Perf event wakeup
2504  *
2505  * If there's data, ensure we set the poll() state and publish everything
2506  * to user-space before waking everybody up.
2507  */
2508
2509 void perf_event_wakeup(struct perf_event *event)
2510 {
2511         wake_up_all(&event->waitq);
2512
2513         if (event->pending_kill) {
2514                 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2515                 event->pending_kill = 0;
2516         }
2517 }
2518
2519 /*
2520  * Pending wakeups
2521  *
2522  * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2523  *
2524  * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2525  * single linked list and use cmpxchg() to add entries lockless.
2526  */
2527
2528 static void perf_pending_event(struct perf_pending_entry *entry)
2529 {
2530         struct perf_event *event = container_of(entry,
2531                         struct perf_event, pending);
2532
2533         if (event->pending_disable) {
2534                 event->pending_disable = 0;
2535                 __perf_event_disable(event);
2536         }
2537
2538         if (event->pending_wakeup) {
2539                 event->pending_wakeup = 0;
2540                 perf_event_wakeup(event);
2541         }
2542 }
2543
2544 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2545
2546 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2547         PENDING_TAIL,
2548 };
2549
2550 static void perf_pending_queue(struct perf_pending_entry *entry,
2551                                void (*func)(struct perf_pending_entry *))
2552 {
2553         struct perf_pending_entry **head;
2554
2555         if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2556                 return;
2557
2558         entry->func = func;
2559
2560         head = &get_cpu_var(perf_pending_head);
2561
2562         do {
2563                 entry->next = *head;
2564         } while (cmpxchg(head, entry->next, entry) != entry->next);
2565
2566         set_perf_event_pending();
2567
2568         put_cpu_var(perf_pending_head);
2569 }
2570
2571 static int __perf_pending_run(void)
2572 {
2573         struct perf_pending_entry *list;
2574         int nr = 0;
2575
2576         list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2577         while (list != PENDING_TAIL) {
2578                 void (*func)(struct perf_pending_entry *);
2579                 struct perf_pending_entry *entry = list;
2580
2581                 list = list->next;
2582
2583                 func = entry->func;
2584                 entry->next = NULL;
2585                 /*
2586                  * Ensure we observe the unqueue before we issue the wakeup,
2587                  * so that we won't be waiting forever.
2588                  * -- see perf_not_pending().
2589                  */
2590                 smp_wmb();
2591
2592                 func(entry);
2593                 nr++;
2594         }
2595
2596         return nr;
2597 }
2598
2599 static inline int perf_not_pending(struct perf_event *event)
2600 {
2601         /*
2602          * If we flush on whatever cpu we run, there is a chance we don't
2603          * need to wait.
2604          */
2605         get_cpu();
2606         __perf_pending_run();
2607         put_cpu();
2608
2609         /*
2610          * Ensure we see the proper queue state before going to sleep
2611          * so that we do not miss the wakeup. -- see perf_pending_handle()
2612          */
2613         smp_rmb();
2614         return event->pending.next == NULL;
2615 }
2616
2617 static void perf_pending_sync(struct perf_event *event)
2618 {
2619         wait_event(event->waitq, perf_not_pending(event));
2620 }
2621
2622 void perf_event_do_pending(void)
2623 {
2624         __perf_pending_run();
2625 }
2626
2627 /*
2628  * Callchain support -- arch specific
2629  */
2630
2631 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2632 {
2633         return NULL;
2634 }
2635
2636 /*
2637  * Output
2638  */
2639 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2640                               unsigned long offset, unsigned long head)
2641 {
2642         unsigned long mask;
2643
2644         if (!data->writable)
2645                 return true;
2646
2647         mask = perf_data_size(data) - 1;
2648
2649         offset = (offset - tail) & mask;
2650         head   = (head   - tail) & mask;
2651
2652         if ((int)(head - offset) < 0)
2653                 return false;
2654
2655         return true;
2656 }
2657
2658 static void perf_output_wakeup(struct perf_output_handle *handle)
2659 {
2660         atomic_set(&handle->data->poll, POLL_IN);
2661
2662         if (handle->nmi) {
2663                 handle->event->pending_wakeup = 1;
2664                 perf_pending_queue(&handle->event->pending,
2665                                    perf_pending_event);
2666         } else
2667                 perf_event_wakeup(handle->event);
2668 }
2669
2670 /*
2671  * Curious locking construct.
2672  *
2673  * We need to ensure a later event_id doesn't publish a head when a former
2674  * event_id isn't done writing. However since we need to deal with NMIs we
2675  * cannot fully serialize things.
2676  *
2677  * What we do is serialize between CPUs so we only have to deal with NMI
2678  * nesting on a single CPU.
2679  *
2680  * We only publish the head (and generate a wakeup) when the outer-most
2681  * event_id completes.
2682  */
2683 static void perf_output_lock(struct perf_output_handle *handle)
2684 {
2685         struct perf_mmap_data *data = handle->data;
2686         int cur, cpu = get_cpu();
2687
2688         handle->locked = 0;
2689
2690         for (;;) {
2691                 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2692                 if (cur == -1) {
2693                         handle->locked = 1;
2694                         break;
2695                 }
2696                 if (cur == cpu)
2697                         break;
2698
2699                 cpu_relax();
2700         }
2701 }
2702
2703 static void perf_output_unlock(struct perf_output_handle *handle)
2704 {
2705         struct perf_mmap_data *data = handle->data;
2706         unsigned long head;
2707         int cpu;
2708
2709         data->done_head = data->head;
2710
2711         if (!handle->locked)
2712                 goto out;
2713
2714 again:
2715         /*
2716          * The xchg implies a full barrier that ensures all writes are done
2717          * before we publish the new head, matched by a rmb() in userspace when
2718          * reading this position.
2719          */
2720         while ((head = atomic_long_xchg(&data->done_head, 0)))
2721                 data->user_page->data_head = head;
2722
2723         /*
2724          * NMI can happen here, which means we can miss a done_head update.
2725          */
2726
2727         cpu = atomic_xchg(&data->lock, -1);
2728         WARN_ON_ONCE(cpu != smp_processor_id());
2729
2730         /*
2731          * Therefore we have to validate we did not indeed do so.
2732          */
2733         if (unlikely(atomic_long_read(&data->done_head))) {
2734                 /*
2735                  * Since we had it locked, we can lock it again.
2736                  */
2737                 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2738                         cpu_relax();
2739
2740                 goto again;
2741         }
2742
2743         if (atomic_xchg(&data->wakeup, 0))
2744                 perf_output_wakeup(handle);
2745 out:
2746         put_cpu();
2747 }
2748
2749 void perf_output_copy(struct perf_output_handle *handle,
2750                       const void *buf, unsigned int len)
2751 {
2752         unsigned int pages_mask;
2753         unsigned long offset;
2754         unsigned int size;
2755         void **pages;
2756
2757         offset          = handle->offset;
2758         pages_mask      = handle->data->nr_pages - 1;
2759         pages           = handle->data->data_pages;
2760
2761         do {
2762                 unsigned long page_offset;
2763                 unsigned long page_size;
2764                 int nr;
2765
2766                 nr          = (offset >> PAGE_SHIFT) & pages_mask;
2767                 page_size   = 1UL << (handle->data->data_order + PAGE_SHIFT);
2768                 page_offset = offset & (page_size - 1);
2769                 size        = min_t(unsigned int, page_size - page_offset, len);
2770
2771                 memcpy(pages[nr] + page_offset, buf, size);
2772
2773                 len         -= size;
2774                 buf         += size;
2775                 offset      += size;
2776         } while (len);
2777
2778         handle->offset = offset;
2779
2780         /*
2781          * Check we didn't copy past our reservation window, taking the
2782          * possible unsigned int wrap into account.
2783          */
2784         WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2785 }
2786
2787 int perf_output_begin(struct perf_output_handle *handle,
2788                       struct perf_event *event, unsigned int size,
2789                       int nmi, int sample)
2790 {
2791         struct perf_event *output_event;
2792         struct perf_mmap_data *data;
2793         unsigned long tail, offset, head;
2794         int have_lost;
2795         struct {
2796                 struct perf_event_header header;
2797                 u64                      id;
2798                 u64                      lost;
2799         } lost_event;
2800
2801         rcu_read_lock();
2802         /*
2803          * For inherited events we send all the output towards the parent.
2804          */
2805         if (event->parent)
2806                 event = event->parent;
2807
2808         output_event = rcu_dereference(event->output);
2809         if (output_event)
2810                 event = output_event;
2811
2812         data = rcu_dereference(event->data);
2813         if (!data)
2814                 goto out;
2815
2816         handle->data    = data;
2817         handle->event   = event;
2818         handle->nmi     = nmi;
2819         handle->sample  = sample;
2820
2821         if (!data->nr_pages)
2822                 goto fail;
2823
2824         have_lost = atomic_read(&data->lost);
2825         if (have_lost)
2826                 size += sizeof(lost_event);
2827
2828         perf_output_lock(handle);
2829
2830         do {
2831                 /*
2832                  * Userspace could choose to issue a mb() before updating the
2833                  * tail pointer. So that all reads will be completed before the
2834                  * write is issued.
2835                  */
2836                 tail = ACCESS_ONCE(data->user_page->data_tail);
2837                 smp_rmb();
2838                 offset = head = atomic_long_read(&data->head);
2839                 head += size;
2840                 if (unlikely(!perf_output_space(data, tail, offset, head)))
2841                         goto fail;
2842         } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2843
2844         handle->offset  = offset;
2845         handle->head    = head;
2846
2847         if (head - tail > data->watermark)
2848                 atomic_set(&data->wakeup, 1);
2849
2850         if (have_lost) {
2851                 lost_event.header.type = PERF_RECORD_LOST;
2852                 lost_event.header.misc = 0;
2853                 lost_event.header.size = sizeof(lost_event);
2854                 lost_event.id          = event->id;
2855                 lost_event.lost        = atomic_xchg(&data->lost, 0);
2856
2857                 perf_output_put(handle, lost_event);
2858         }
2859
2860         return 0;
2861
2862 fail:
2863         atomic_inc(&data->lost);
2864         perf_output_unlock(handle);
2865 out:
2866         rcu_read_unlock();
2867
2868         return -ENOSPC;
2869 }
2870
2871 void perf_output_end(struct perf_output_handle *handle)
2872 {
2873         struct perf_event *event = handle->event;
2874         struct perf_mmap_data *data = handle->data;
2875
2876         int wakeup_events = event->attr.wakeup_events;
2877
2878         if (handle->sample && wakeup_events) {
2879                 int events = atomic_inc_return(&data->events);
2880                 if (events >= wakeup_events) {
2881                         atomic_sub(wakeup_events, &data->events);
2882                         atomic_set(&data->wakeup, 1);
2883                 }
2884         }
2885
2886         perf_output_unlock(handle);
2887         rcu_read_unlock();
2888 }
2889
2890 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
2891 {
2892         /*
2893          * only top level events have the pid namespace they were created in
2894          */
2895         if (event->parent)
2896                 event = event->parent;
2897
2898         return task_tgid_nr_ns(p, event->ns);
2899 }
2900
2901 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
2902 {
2903         /*
2904          * only top level events have the pid namespace they were created in
2905          */
2906         if (event->parent)
2907                 event = event->parent;
2908
2909         return task_pid_nr_ns(p, event->ns);
2910 }
2911
2912 static void perf_output_read_one(struct perf_output_handle *handle,
2913                                  struct perf_event *event)
2914 {
2915         u64 read_format = event->attr.read_format;
2916         u64 values[4];
2917         int n = 0;
2918
2919         values[n++] = atomic64_read(&event->count);
2920         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2921                 values[n++] = event->total_time_enabled +
2922                         atomic64_read(&event->child_total_time_enabled);
2923         }
2924         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2925                 values[n++] = event->total_time_running +
2926                         atomic64_read(&event->child_total_time_running);
2927         }
2928         if (read_format & PERF_FORMAT_ID)
2929                 values[n++] = primary_event_id(event);
2930
2931         perf_output_copy(handle, values, n * sizeof(u64));
2932 }
2933
2934 /*
2935  * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
2936  */
2937 static void perf_output_read_group(struct perf_output_handle *handle,
2938                             struct perf_event *event)
2939 {
2940         struct perf_event *leader = event->group_leader, *sub;
2941         u64 read_format = event->attr.read_format;
2942         u64 values[5];
2943         int n = 0;
2944
2945         values[n++] = 1 + leader->nr_siblings;
2946
2947         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2948                 values[n++] = leader->total_time_enabled;
2949
2950         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2951                 values[n++] = leader->total_time_running;
2952
2953         if (leader != event)
2954                 leader->pmu->read(leader);
2955
2956         values[n++] = atomic64_read(&leader->count);
2957         if (read_format & PERF_FORMAT_ID)
2958                 values[n++] = primary_event_id(leader);
2959
2960         perf_output_copy(handle, values, n * sizeof(u64));
2961
2962         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2963                 n = 0;
2964
2965                 if (sub != event)
2966                         sub->pmu->read(sub);
2967
2968                 values[n++] = atomic64_read(&sub->count);
2969                 if (read_format & PERF_FORMAT_ID)
2970                         values[n++] = primary_event_id(sub);
2971
2972                 perf_output_copy(handle, values, n * sizeof(u64));
2973         }
2974 }
2975
2976 static void perf_output_read(struct perf_output_handle *handle,
2977                              struct perf_event *event)
2978 {
2979         if (event->attr.read_format & PERF_FORMAT_GROUP)
2980                 perf_output_read_group(handle, event);
2981         else
2982                 perf_output_read_one(handle, event);
2983 }
2984
2985 void perf_output_sample(struct perf_output_handle *handle,
2986                         struct perf_event_header *header,
2987                         struct perf_sample_data *data,
2988                         struct perf_event *event)
2989 {
2990         u64 sample_type = data->type;
2991
2992         perf_output_put(handle, *header);
2993
2994         if (sample_type & PERF_SAMPLE_IP)
2995                 perf_output_put(handle, data->ip);
2996
2997         if (sample_type & PERF_SAMPLE_TID)
2998                 perf_output_put(handle, data->tid_entry);
2999
3000         if (sample_type & PERF_SAMPLE_TIME)
3001                 perf_output_put(handle, data->time);
3002
3003         if (sample_type & PERF_SAMPLE_ADDR)
3004                 perf_output_put(handle, data->addr);
3005
3006         if (sample_type & PERF_SAMPLE_ID)
3007                 perf_output_put(handle, data->id);
3008
3009         if (sample_type & PERF_SAMPLE_STREAM_ID)
3010                 perf_output_put(handle, data->stream_id);
3011
3012         if (sample_type & PERF_SAMPLE_CPU)
3013                 perf_output_put(handle, data->cpu_entry);
3014
3015         if (sample_type & PERF_SAMPLE_PERIOD)
3016                 perf_output_put(handle, data->period);
3017
3018         if (sample_type & PERF_SAMPLE_READ)
3019                 perf_output_read(handle, event);
3020
3021         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3022                 if (data->callchain) {
3023                         int size = 1;
3024
3025                         if (data->callchain)
3026                                 size += data->callchain->nr;
3027
3028                         size *= sizeof(u64);
3029
3030                         perf_output_copy(handle, data->callchain, size);
3031                 } else {
3032                         u64 nr = 0;
3033                         perf_output_put(handle, nr);
3034                 }
3035         }
3036
3037         if (sample_type & PERF_SAMPLE_RAW) {
3038                 if (data->raw) {
3039                         perf_output_put(handle, data->raw->size);
3040                         perf_output_copy(handle, data->raw->data,
3041                                          data->raw->size);
3042                 } else {
3043                         struct {
3044                                 u32     size;
3045                                 u32     data;
3046                         } raw = {
3047                                 .size = sizeof(u32),
3048                                 .data = 0,
3049                         };
3050                         perf_output_put(handle, raw);
3051                 }
3052         }
3053 }
3054
3055 void perf_prepare_sample(struct perf_event_header *header,
3056                          struct perf_sample_data *data,
3057                          struct perf_event *event,
3058                          struct pt_regs *regs)
3059 {
3060         u64 sample_type = event->attr.sample_type;
3061
3062         data->type = sample_type;
3063
3064         header->type = PERF_RECORD_SAMPLE;
3065         header->size = sizeof(*header);
3066
3067         header->misc = 0;
3068         header->misc |= perf_misc_flags(regs);
3069
3070         if (sample_type & PERF_SAMPLE_IP) {
3071                 data->ip = perf_instruction_pointer(regs);
3072
3073                 header->size += sizeof(data->ip);
3074         }
3075
3076         if (sample_type & PERF_SAMPLE_TID) {
3077                 /* namespace issues */
3078                 data->tid_entry.pid = perf_event_pid(event, current);
3079                 data->tid_entry.tid = perf_event_tid(event, current);
3080
3081                 header->size += sizeof(data->tid_entry);
3082         }
3083
3084         if (sample_type & PERF_SAMPLE_TIME) {
3085                 data->time = perf_clock();
3086
3087                 header->size += sizeof(data->time);
3088         }
3089
3090         if (sample_type & PERF_SAMPLE_ADDR)
3091                 header->size += sizeof(data->addr);
3092
3093         if (sample_type & PERF_SAMPLE_ID) {
3094                 data->id = primary_event_id(event);
3095
3096                 header->size += sizeof(data->id);
3097         }
3098
3099         if (sample_type & PERF_SAMPLE_STREAM_ID) {
3100                 data->stream_id = event->id;
3101
3102                 header->size += sizeof(data->stream_id);
3103         }
3104
3105         if (sample_type & PERF_SAMPLE_CPU) {
3106                 data->cpu_entry.cpu             = raw_smp_processor_id();
3107                 data->cpu_entry.reserved        = 0;
3108
3109                 header->size += sizeof(data->cpu_entry);
3110         }
3111
3112         if (sample_type & PERF_SAMPLE_PERIOD)
3113                 header->size += sizeof(data->period);
3114
3115         if (sample_type & PERF_SAMPLE_READ)
3116                 header->size += perf_event_read_size(event);
3117
3118         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3119                 int size = 1;
3120
3121                 data->callchain = perf_callchain(regs);
3122
3123                 if (data->callchain)
3124                         size += data->callchain->nr;
3125
3126                 header->size += size * sizeof(u64);
3127         }
3128
3129         if (sample_type & PERF_SAMPLE_RAW) {
3130                 int size = sizeof(u32);
3131
3132                 if (data->raw)
3133                         size += data->raw->size;
3134                 else
3135                         size += sizeof(u32);
3136
3137                 WARN_ON_ONCE(size & (sizeof(u64)-1));
3138                 header->size += size;
3139         }
3140 }
3141
3142 static void perf_event_output(struct perf_event *event, int nmi,
3143                                 struct perf_sample_data *data,
3144                                 struct pt_regs *regs)
3145 {
3146         struct perf_output_handle handle;
3147         struct perf_event_header header;
3148
3149         perf_prepare_sample(&header, data, event, regs);
3150
3151         if (perf_output_begin(&handle, event, header.size, nmi, 1))
3152                 return;
3153
3154         perf_output_sample(&handle, &header, data, event);
3155
3156         perf_output_end(&handle);
3157 }
3158
3159 /*
3160  * read event_id
3161  */
3162
3163 struct perf_read_event {
3164         struct perf_event_header        header;
3165
3166         u32                             pid;
3167         u32                             tid;
3168 };
3169
3170 static void
3171 perf_event_read_event(struct perf_event *event,
3172                         struct task_struct *task)
3173 {
3174         struct perf_output_handle handle;
3175         struct perf_read_event read_event = {
3176                 .header = {
3177                         .type = PERF_RECORD_READ,
3178                         .misc = 0,
3179                         .size = sizeof(read_event) + perf_event_read_size(event),
3180                 },
3181                 .pid = perf_event_pid(event, task),
3182                 .tid = perf_event_tid(event, task),
3183         };
3184         int ret;
3185
3186         ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3187         if (ret)
3188                 return;
3189
3190         perf_output_put(&handle, read_event);
3191         perf_output_read(&handle, event);
3192
3193         perf_output_end(&handle);
3194 }
3195
3196 /*
3197  * task tracking -- fork/exit
3198  *
3199  * enabled by: attr.comm | attr.mmap | attr.task
3200  */
3201
3202 struct perf_task_event {
3203         struct task_struct              *task;
3204         struct perf_event_context       *task_ctx;
3205
3206         struct {
3207                 struct perf_event_header        header;
3208
3209                 u32                             pid;
3210                 u32                             ppid;
3211                 u32                             tid;
3212                 u32                             ptid;
3213                 u64                             time;
3214         } event_id;
3215 };
3216
3217 static void perf_event_task_output(struct perf_event *event,
3218                                      struct perf_task_event *task_event)
3219 {
3220         struct perf_output_handle handle;
3221         int size;
3222         struct task_struct *task = task_event->task;
3223         int ret;
3224
3225         size  = task_event->event_id.header.size;
3226         ret = perf_output_begin(&handle, event, size, 0, 0);
3227
3228         if (ret)
3229                 return;
3230
3231         task_event->event_id.pid = perf_event_pid(event, task);
3232         task_event->event_id.ppid = perf_event_pid(event, current);
3233
3234         task_event->event_id.tid = perf_event_tid(event, task);
3235         task_event->event_id.ptid = perf_event_tid(event, current);
3236
3237         task_event->event_id.time = perf_clock();
3238
3239         perf_output_put(&handle, task_event->event_id);
3240
3241         perf_output_end(&handle);
3242 }
3243
3244 static int perf_event_task_match(struct perf_event *event)
3245 {
3246         if (event->attr.comm || event->attr.mmap || event->attr.task)
3247                 return 1;
3248
3249         return 0;
3250 }
3251
3252 static void perf_event_task_ctx(struct perf_event_context *ctx,
3253                                   struct perf_task_event *task_event)
3254 {
3255         struct perf_event *event;
3256
3257         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3258                 if (perf_event_task_match(event))
3259                         perf_event_task_output(event, task_event);
3260         }
3261 }
3262
3263 static void perf_event_task_event(struct perf_task_event *task_event)
3264 {
3265         struct perf_cpu_context *cpuctx;
3266         struct perf_event_context *ctx = task_event->task_ctx;
3267
3268         rcu_read_lock();
3269         cpuctx = &get_cpu_var(perf_cpu_context);
3270         perf_event_task_ctx(&cpuctx->ctx, task_event);
3271         put_cpu_var(perf_cpu_context);
3272
3273         if (!ctx)
3274                 ctx = rcu_dereference(task_event->task->perf_event_ctxp);
3275         if (ctx)
3276                 perf_event_task_ctx(ctx, task_event);
3277         rcu_read_unlock();
3278 }
3279
3280 static void perf_event_task(struct task_struct *task,
3281                               struct perf_event_context *task_ctx,
3282                               int new)
3283 {
3284         struct perf_task_event task_event;
3285
3286         if (!atomic_read(&nr_comm_events) &&
3287             !atomic_read(&nr_mmap_events) &&
3288             !atomic_read(&nr_task_events))
3289                 return;
3290
3291         task_event = (struct perf_task_event){
3292                 .task     = task,
3293                 .task_ctx = task_ctx,
3294                 .event_id    = {
3295                         .header = {
3296                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3297                                 .misc = 0,
3298                                 .size = sizeof(task_event.event_id),
3299                         },
3300                         /* .pid  */
3301                         /* .ppid */
3302                         /* .tid  */
3303                         /* .ptid */
3304                 },
3305         };
3306
3307         perf_event_task_event(&task_event);
3308 }
3309
3310 void perf_event_fork(struct task_struct *task)
3311 {
3312         perf_event_task(task, NULL, 1);
3313 }
3314
3315 /*
3316  * comm tracking
3317  */
3318
3319 struct perf_comm_event {
3320         struct task_struct      *task;
3321         char                    *comm;
3322         int                     comm_size;
3323
3324         struct {
3325                 struct perf_event_header        header;
3326
3327                 u32                             pid;
3328                 u32                             tid;
3329         } event_id;
3330 };
3331
3332 static void perf_event_comm_output(struct perf_event *event,
3333                                      struct perf_comm_event *comm_event)
3334 {
3335         struct perf_output_handle handle;
3336         int size = comm_event->event_id.header.size;
3337         int ret = perf_output_begin(&handle, event, size, 0, 0);
3338
3339         if (ret)
3340                 return;
3341
3342         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3343         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3344
3345         perf_output_put(&handle, comm_event->event_id);
3346         perf_output_copy(&handle, comm_event->comm,
3347                                    comm_event->comm_size);
3348         perf_output_end(&handle);
3349 }
3350
3351 static int perf_event_comm_match(struct perf_event *event)
3352 {
3353         if (event->attr.comm)
3354                 return 1;
3355
3356         return 0;
3357 }
3358
3359 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3360                                   struct perf_comm_event *comm_event)
3361 {
3362         struct perf_event *event;
3363
3364         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3365                 if (perf_event_comm_match(event))
3366                         perf_event_comm_output(event, comm_event);
3367         }
3368 }
3369
3370 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3371 {
3372         struct perf_cpu_context *cpuctx;
3373         struct perf_event_context *ctx;
3374         unsigned int size;
3375         char comm[TASK_COMM_LEN];
3376
3377         memset(comm, 0, sizeof(comm));
3378         strncpy(comm, comm_event->task->comm, sizeof(comm));
3379         size = ALIGN(strlen(comm)+1, sizeof(u64));
3380
3381         comm_event->comm = comm;
3382         comm_event->comm_size = size;
3383
3384         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3385
3386         rcu_read_lock();
3387         cpuctx = &get_cpu_var(perf_cpu_context);
3388         perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3389         put_cpu_var(perf_cpu_context);
3390
3391         /*
3392          * doesn't really matter which of the child contexts the
3393          * events ends up in.
3394          */
3395         ctx = rcu_dereference(current->perf_event_ctxp);
3396         if (ctx)
3397                 perf_event_comm_ctx(ctx, comm_event);
3398         rcu_read_unlock();
3399 }
3400
3401 void perf_event_comm(struct task_struct *task)
3402 {
3403         struct perf_comm_event comm_event;
3404
3405         if (task->perf_event_ctxp)
3406                 perf_event_enable_on_exec(task);
3407
3408         if (!atomic_read(&nr_comm_events))
3409                 return;
3410
3411         comm_event = (struct perf_comm_event){
3412                 .task   = task,
3413                 /* .comm      */
3414                 /* .comm_size */
3415                 .event_id  = {
3416                         .header = {
3417                                 .type = PERF_RECORD_COMM,
3418                                 .misc = 0,
3419                                 /* .size */
3420                         },
3421                         /* .pid */
3422                         /* .tid */
3423                 },
3424         };
3425
3426         perf_event_comm_event(&comm_event);
3427 }
3428
3429 /*
3430  * mmap tracking
3431  */
3432
3433 struct perf_mmap_event {
3434         struct vm_area_struct   *vma;
3435
3436         const char              *file_name;
3437         int                     file_size;
3438
3439         struct {
3440                 struct perf_event_header        header;
3441
3442                 u32                             pid;
3443                 u32                             tid;
3444                 u64                             start;
3445                 u64                             len;
3446                 u64                             pgoff;
3447         } event_id;
3448 };
3449
3450 static void perf_event_mmap_output(struct perf_event *event,
3451                                      struct perf_mmap_event *mmap_event)
3452 {
3453         struct perf_output_handle handle;
3454         int size = mmap_event->event_id.header.size;
3455         int ret = perf_output_begin(&handle, event, size, 0, 0);
3456
3457         if (ret)
3458                 return;
3459
3460         mmap_event->event_id.pid = perf_event_pid(event, current);
3461         mmap_event->event_id.tid = perf_event_tid(event, current);
3462
3463         perf_output_put(&handle, mmap_event->event_id);
3464         perf_output_copy(&handle, mmap_event->file_name,
3465                                    mmap_event->file_size);
3466         perf_output_end(&handle);
3467 }
3468
3469 static int perf_event_mmap_match(struct perf_event *event,
3470                                    struct perf_mmap_event *mmap_event)
3471 {
3472         if (event->attr.mmap)
3473                 return 1;
3474
3475         return 0;
3476 }
3477
3478 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3479                                   struct perf_mmap_event *mmap_event)
3480 {
3481         struct perf_event *event;
3482
3483         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3484                 if (perf_event_mmap_match(event, mmap_event))
3485                         perf_event_mmap_output(event, mmap_event);
3486         }
3487 }
3488
3489 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3490 {
3491         struct perf_cpu_context *cpuctx;
3492         struct perf_event_context *ctx;
3493         struct vm_area_struct *vma = mmap_event->vma;
3494         struct file *file = vma->vm_file;
3495         unsigned int size;
3496         char tmp[16];
3497         char *buf = NULL;
3498         const char *name;
3499
3500         memset(tmp, 0, sizeof(tmp));
3501
3502         if (file) {
3503                 /*
3504                  * d_path works from the end of the buffer backwards, so we
3505                  * need to add enough zero bytes after the string to handle
3506                  * the 64bit alignment we do later.
3507                  */
3508                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3509                 if (!buf) {
3510                         name = strncpy(tmp, "//enomem", sizeof(tmp));
3511                         goto got_name;
3512                 }
3513                 name = d_path(&file->f_path, buf, PATH_MAX);
3514                 if (IS_ERR(name)) {
3515                         name = strncpy(tmp, "//toolong", sizeof(tmp));
3516                         goto got_name;
3517                 }
3518         } else {
3519                 if (arch_vma_name(mmap_event->vma)) {
3520                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3521                                        sizeof(tmp));
3522                         goto got_name;
3523                 }
3524
3525                 if (!vma->vm_mm) {
3526                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
3527                         goto got_name;
3528                 }
3529
3530                 name = strncpy(tmp, "//anon", sizeof(tmp));
3531                 goto got_name;
3532         }
3533
3534 got_name:
3535         size = ALIGN(strlen(name)+1, sizeof(u64));
3536
3537         mmap_event->file_name = name;
3538         mmap_event->file_size = size;
3539
3540         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3541
3542         rcu_read_lock();
3543         cpuctx = &get_cpu_var(perf_cpu_context);
3544         perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3545         put_cpu_var(perf_cpu_context);
3546
3547         /*
3548          * doesn't really matter which of the child contexts the
3549          * events ends up in.
3550          */
3551         ctx = rcu_dereference(current->perf_event_ctxp);
3552         if (ctx)
3553                 perf_event_mmap_ctx(ctx, mmap_event);
3554         rcu_read_unlock();
3555
3556         kfree(buf);
3557 }
3558
3559 void __perf_event_mmap(struct vm_area_struct *vma)
3560 {
3561         struct perf_mmap_event mmap_event;
3562
3563         if (!atomic_read(&nr_mmap_events))
3564                 return;
3565
3566         mmap_event = (struct perf_mmap_event){
3567                 .vma    = vma,
3568                 /* .file_name */
3569                 /* .file_size */
3570                 .event_id  = {
3571                         .header = {
3572                                 .type = PERF_RECORD_MMAP,
3573                                 .misc = 0,
3574                                 /* .size */
3575                         },
3576                         /* .pid */
3577                         /* .tid */
3578                         .start  = vma->vm_start,
3579                         .len    = vma->vm_end - vma->vm_start,
3580                         .pgoff  = vma->vm_pgoff,
3581                 },
3582         };
3583
3584         perf_event_mmap_event(&mmap_event);
3585 }
3586
3587 /*
3588  * IRQ throttle logging
3589  */
3590
3591 static void perf_log_throttle(struct perf_event *event, int enable)
3592 {
3593         struct perf_output_handle handle;
3594         int ret;
3595
3596         struct {
3597                 struct perf_event_header        header;
3598                 u64                             time;
3599                 u64                             id;
3600                 u64                             stream_id;
3601         } throttle_event = {
3602                 .header = {
3603                         .type = PERF_RECORD_THROTTLE,
3604                         .misc = 0,
3605                         .size = sizeof(throttle_event),
3606                 },
3607                 .time           = perf_clock(),
3608                 .id             = primary_event_id(event),
3609                 .stream_id      = event->id,
3610         };
3611
3612         if (enable)
3613                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3614
3615         ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3616         if (ret)
3617                 return;
3618
3619         perf_output_put(&handle, throttle_event);
3620         perf_output_end(&handle);
3621 }
3622
3623 /*
3624  * Generic event overflow handling, sampling.
3625  */
3626
3627 static int __perf_event_overflow(struct perf_event *event, int nmi,
3628                                    int throttle, struct perf_sample_data *data,
3629                                    struct pt_regs *regs)
3630 {
3631         int events = atomic_read(&event->event_limit);
3632         struct hw_perf_event *hwc = &event->hw;
3633         int ret = 0;
3634
3635         throttle = (throttle && event->pmu->unthrottle != NULL);
3636
3637         if (!throttle) {
3638                 hwc->interrupts++;
3639         } else {
3640                 if (hwc->interrupts != MAX_INTERRUPTS) {
3641                         hwc->interrupts++;
3642                         if (HZ * hwc->interrupts >
3643                                         (u64)sysctl_perf_event_sample_rate) {
3644                                 hwc->interrupts = MAX_INTERRUPTS;
3645                                 perf_log_throttle(event, 0);
3646                                 ret = 1;
3647                         }
3648                 } else {
3649                         /*
3650                          * Keep re-disabling events even though on the previous
3651                          * pass we disabled it - just in case we raced with a
3652                          * sched-in and the event got enabled again:
3653                          */
3654                         ret = 1;
3655                 }
3656         }
3657
3658         if (event->attr.freq) {
3659                 u64 now = perf_clock();
3660                 s64 delta = now - hwc->freq_stamp;
3661
3662                 hwc->freq_stamp = now;
3663
3664                 if (delta > 0 && delta < TICK_NSEC)
3665                         perf_adjust_period(event, NSEC_PER_SEC / (int)delta);
3666         }
3667
3668         /*
3669          * XXX event_limit might not quite work as expected on inherited
3670          * events
3671          */
3672
3673         event->pending_kill = POLL_IN;
3674         if (events && atomic_dec_and_test(&event->event_limit)) {
3675                 ret = 1;
3676                 event->pending_kill = POLL_HUP;
3677                 if (nmi) {
3678                         event->pending_disable = 1;
3679                         perf_pending_queue(&event->pending,
3680                                            perf_pending_event);
3681                 } else
3682                         perf_event_disable(event);
3683         }
3684
3685         if (event->overflow_handler)
3686                 event->overflow_handler(event, nmi, data, regs);
3687         else
3688                 perf_event_output(event, nmi, data, regs);
3689
3690         return ret;
3691 }
3692
3693 int perf_event_overflow(struct perf_event *event, int nmi,
3694                           struct perf_sample_data *data,
3695                           struct pt_regs *regs)
3696 {
3697         return __perf_event_overflow(event, nmi, 1, data, regs);
3698 }
3699
3700 /*
3701  * Generic software event infrastructure
3702  */
3703
3704 /*
3705  * We directly increment event->count and keep a second value in
3706  * event->hw.period_left to count intervals. This period event
3707  * is kept in the range [-sample_period, 0] so that we can use the
3708  * sign as trigger.
3709  */
3710
3711 static u64 perf_swevent_set_period(struct perf_event *event)
3712 {
3713         struct hw_perf_event *hwc = &event->hw;
3714         u64 period = hwc->last_period;
3715         u64 nr, offset;
3716         s64 old, val;
3717
3718         hwc->last_period = hwc->sample_period;
3719
3720 again:
3721         old = val = atomic64_read(&hwc->period_left);
3722         if (val < 0)
3723                 return 0;
3724
3725         nr = div64_u64(period + val, period);
3726         offset = nr * period;
3727         val -= offset;
3728         if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3729                 goto again;
3730
3731         return nr;
3732 }
3733
3734 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3735                                     int nmi, struct perf_sample_data *data,
3736                                     struct pt_regs *regs)
3737 {
3738         struct hw_perf_event *hwc = &event->hw;
3739         int throttle = 0;
3740
3741         data->period = event->hw.last_period;
3742         if (!overflow)
3743                 overflow = perf_swevent_set_period(event);
3744
3745         if (hwc->interrupts == MAX_INTERRUPTS)
3746                 return;
3747
3748         for (; overflow; overflow--) {
3749                 if (__perf_event_overflow(event, nmi, throttle,
3750                                             data, regs)) {
3751                         /*
3752                          * We inhibit the overflow from happening when
3753                          * hwc->interrupts == MAX_INTERRUPTS.
3754                          */
3755                         break;
3756                 }
3757                 throttle = 1;
3758         }
3759 }
3760
3761 static void perf_swevent_unthrottle(struct perf_event *event)
3762 {
3763         /*
3764          * Nothing to do, we already reset hwc->interrupts.
3765          */
3766 }
3767
3768 static void perf_swevent_add(struct perf_event *event, u64 nr,
3769                                int nmi, struct perf_sample_data *data,
3770                                struct pt_regs *regs)
3771 {
3772         struct hw_perf_event *hwc = &event->hw;
3773
3774         atomic64_add(nr, &event->count);
3775
3776         if (!regs)
3777                 return;
3778
3779         if (!hwc->sample_period)
3780                 return;
3781
3782         if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3783                 return perf_swevent_overflow(event, 1, nmi, data, regs);
3784
3785         if (atomic64_add_negative(nr, &hwc->period_left))
3786                 return;
3787
3788         perf_swevent_overflow(event, 0, nmi, data, regs);
3789 }
3790
3791 static int perf_swevent_is_counting(struct perf_event *event)
3792 {
3793         /*
3794          * The event is active, we're good!
3795          */
3796         if (event->state == PERF_EVENT_STATE_ACTIVE)
3797                 return 1;
3798
3799         /*
3800          * The event is off/error, not counting.
3801          */
3802         if (event->state != PERF_EVENT_STATE_INACTIVE)
3803                 return 0;
3804
3805         /*
3806          * The event is inactive, if the context is active
3807          * we're part of a group that didn't make it on the 'pmu',
3808          * not counting.
3809          */
3810         if (event->ctx->is_active)
3811                 return 0;
3812
3813         /*
3814          * We're inactive and the context is too, this means the
3815          * task is scheduled out, we're counting events that happen
3816          * to us, like migration events.
3817          */
3818         return 1;
3819 }
3820
3821 static int perf_tp_event_match(struct perf_event *event,
3822                                 struct perf_sample_data *data);
3823
3824 static int perf_swevent_match(struct perf_event *event,
3825                                 enum perf_type_id type,
3826                                 u32 event_id,
3827                                 struct perf_sample_data *data,
3828                                 struct pt_regs *regs)
3829 {
3830         if (!perf_swevent_is_counting(event))
3831                 return 0;
3832
3833         if (event->attr.type != type)
3834                 return 0;
3835         if (event->attr.config != event_id)
3836                 return 0;
3837
3838         if (regs) {
3839                 if (event->attr.exclude_user && user_mode(regs))
3840                         return 0;
3841
3842                 if (event->attr.exclude_kernel && !user_mode(regs))
3843                         return 0;
3844         }
3845
3846         if (event->attr.type == PERF_TYPE_TRACEPOINT &&
3847             !perf_tp_event_match(event, data))
3848                 return 0;
3849
3850         return 1;
3851 }
3852
3853 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
3854                                      enum perf_type_id type,
3855                                      u32 event_id, u64 nr, int nmi,
3856                                      struct perf_sample_data *data,
3857                                      struct pt_regs *regs)
3858 {
3859         struct perf_event *event;
3860
3861         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3862                 if (perf_swevent_match(event, type, event_id, data, regs))
3863                         perf_swevent_add(event, nr, nmi, data, regs);
3864         }
3865 }
3866
3867 static int *perf_swevent_recursion_context(struct perf_cpu_context *cpuctx)
3868 {
3869         if (in_nmi())
3870                 return &cpuctx->recursion[3];
3871
3872         if (in_irq())
3873                 return &cpuctx->recursion[2];
3874
3875         if (in_softirq())
3876                 return &cpuctx->recursion[1];
3877
3878         return &cpuctx->recursion[0];
3879 }
3880
3881 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
3882                                     u64 nr, int nmi,
3883                                     struct perf_sample_data *data,
3884                                     struct pt_regs *regs)
3885 {
3886         struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3887         int *recursion = perf_swevent_recursion_context(cpuctx);
3888         struct perf_event_context *ctx;
3889
3890         if (*recursion)
3891                 goto out;
3892
3893         (*recursion)++;
3894         barrier();
3895
3896         rcu_read_lock();
3897         perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
3898                                  nr, nmi, data, regs);
3899         /*
3900          * doesn't really matter which of the child contexts the
3901          * events ends up in.
3902          */
3903         ctx = rcu_dereference(current->perf_event_ctxp);
3904         if (ctx)
3905                 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
3906         rcu_read_unlock();
3907
3908         barrier();
3909         (*recursion)--;
3910
3911 out:
3912         put_cpu_var(perf_cpu_context);
3913 }
3914
3915 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
3916                             struct pt_regs *regs, u64 addr)
3917 {
3918         struct perf_sample_data data = {
3919                 .addr = addr,
3920         };
3921
3922         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi,
3923                                 &data, regs);
3924 }
3925
3926 static void perf_swevent_read(struct perf_event *event)
3927 {
3928 }
3929
3930 static int perf_swevent_enable(struct perf_event *event)
3931 {
3932         struct hw_perf_event *hwc = &event->hw;
3933
3934         if (hwc->sample_period) {
3935                 hwc->last_period = hwc->sample_period;
3936                 perf_swevent_set_period(event);
3937         }
3938         return 0;
3939 }
3940
3941 static void perf_swevent_disable(struct perf_event *event)
3942 {
3943 }
3944
3945 static const struct pmu perf_ops_generic = {
3946         .enable         = perf_swevent_enable,
3947         .disable        = perf_swevent_disable,
3948         .read           = perf_swevent_read,
3949         .unthrottle     = perf_swevent_unthrottle,
3950 };
3951
3952 /*
3953  * hrtimer based swevent callback
3954  */
3955
3956 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
3957 {
3958         enum hrtimer_restart ret = HRTIMER_RESTART;
3959         struct perf_sample_data data;
3960         struct pt_regs *regs;
3961         struct perf_event *event;
3962         u64 period;
3963
3964         event   = container_of(hrtimer, struct perf_event, hw.hrtimer);
3965         event->pmu->read(event);
3966
3967         data.addr = 0;
3968         regs = get_irq_regs();
3969         /*
3970          * In case we exclude kernel IPs or are somehow not in interrupt
3971          * context, provide the next best thing, the user IP.
3972          */
3973         if ((event->attr.exclude_kernel || !regs) &&
3974                         !event->attr.exclude_user)
3975                 regs = task_pt_regs(current);
3976
3977         if (regs) {
3978                 if (!(event->attr.exclude_idle && current->pid == 0))
3979                         if (perf_event_overflow(event, 0, &data, regs))
3980                                 ret = HRTIMER_NORESTART;
3981         }
3982
3983         period = max_t(u64, 10000, event->hw.sample_period);
3984         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
3985
3986         return ret;
3987 }
3988
3989 static void perf_swevent_start_hrtimer(struct perf_event *event)
3990 {
3991         struct hw_perf_event *hwc = &event->hw;
3992
3993         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3994         hwc->hrtimer.function = perf_swevent_hrtimer;
3995         if (hwc->sample_period) {
3996                 u64 period;
3997
3998                 if (hwc->remaining) {
3999                         if (hwc->remaining < 0)
4000                                 period = 10000;
4001                         else
4002                                 period = hwc->remaining;
4003                         hwc->remaining = 0;
4004                 } else {
4005                         period = max_t(u64, 10000, hwc->sample_period);
4006                 }
4007                 __hrtimer_start_range_ns(&hwc->hrtimer,
4008                                 ns_to_ktime(period), 0,
4009                                 HRTIMER_MODE_REL, 0);
4010         }
4011 }
4012
4013 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4014 {
4015         struct hw_perf_event *hwc = &event->hw;
4016
4017         if (hwc->sample_period) {
4018                 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4019                 hwc->remaining = ktime_to_ns(remaining);
4020
4021                 hrtimer_cancel(&hwc->hrtimer);
4022         }
4023 }
4024
4025 /*
4026  * Software event: cpu wall time clock
4027  */
4028
4029 static void cpu_clock_perf_event_update(struct perf_event *event)
4030 {
4031         int cpu = raw_smp_processor_id();
4032         s64 prev;
4033         u64 now;
4034
4035         now = cpu_clock(cpu);
4036         prev = atomic64_read(&event->hw.prev_count);
4037         atomic64_set(&event->hw.prev_count, now);
4038         atomic64_add(now - prev, &event->count);
4039 }
4040
4041 static int cpu_clock_perf_event_enable(struct perf_event *event)
4042 {
4043         struct hw_perf_event *hwc = &event->hw;
4044         int cpu = raw_smp_processor_id();
4045
4046         atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4047         perf_swevent_start_hrtimer(event);
4048
4049         return 0;
4050 }
4051
4052 static void cpu_clock_perf_event_disable(struct perf_event *event)
4053 {
4054         perf_swevent_cancel_hrtimer(event);
4055         cpu_clock_perf_event_update(event);
4056 }
4057
4058 static void cpu_clock_perf_event_read(struct perf_event *event)
4059 {
4060         cpu_clock_perf_event_update(event);
4061 }
4062
4063 static const struct pmu perf_ops_cpu_clock = {
4064         .enable         = cpu_clock_perf_event_enable,
4065         .disable        = cpu_clock_perf_event_disable,
4066         .read           = cpu_clock_perf_event_read,
4067 };
4068
4069 /*
4070  * Software event: task time clock
4071  */
4072
4073 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4074 {
4075         u64 prev;
4076         s64 delta;
4077
4078         prev = atomic64_xchg(&event->hw.prev_count, now);
4079         delta = now - prev;
4080         atomic64_add(delta, &event->count);
4081 }
4082
4083 static int task_clock_perf_event_enable(struct perf_event *event)
4084 {
4085         struct hw_perf_event *hwc = &event->hw;
4086         u64 now;
4087
4088         now = event->ctx->time;
4089
4090         atomic64_set(&hwc->prev_count, now);
4091
4092         perf_swevent_start_hrtimer(event);
4093
4094         return 0;
4095 }
4096
4097 static void task_clock_perf_event_disable(struct perf_event *event)
4098 {
4099         perf_swevent_cancel_hrtimer(event);
4100         task_clock_perf_event_update(event, event->ctx->time);
4101
4102 }
4103
4104 static void task_clock_perf_event_read(struct perf_event *event)
4105 {
4106         u64 time;
4107
4108         if (!in_nmi()) {
4109                 update_context_time(event->ctx);
4110                 time = event->ctx->time;
4111         } else {
4112                 u64 now = perf_clock();
4113                 u64 delta = now - event->ctx->timestamp;
4114                 time = event->ctx->time + delta;
4115         }
4116
4117         task_clock_perf_event_update(event, time);
4118 }
4119
4120 static const struct pmu perf_ops_task_clock = {
4121         .enable         = task_clock_perf_event_enable,
4122         .disable        = task_clock_perf_event_disable,
4123         .read           = task_clock_perf_event_read,
4124 };
4125
4126 #ifdef CONFIG_EVENT_PROFILE
4127
4128 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4129                           int entry_size)
4130 {
4131         struct perf_raw_record raw = {
4132                 .size = entry_size,
4133                 .data = record,
4134         };
4135
4136         struct perf_sample_data data = {
4137                 .addr = addr,
4138                 .raw = &raw,
4139         };
4140
4141         struct pt_regs *regs = get_irq_regs();
4142
4143         if (!regs)
4144                 regs = task_pt_regs(current);
4145
4146         do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4147                                 &data, regs);
4148 }
4149 EXPORT_SYMBOL_GPL(perf_tp_event);
4150
4151 static int perf_tp_event_match(struct perf_event *event,
4152                                 struct perf_sample_data *data)
4153 {
4154         void *record = data->raw->data;
4155
4156         if (likely(!event->filter) || filter_match_preds(event->filter, record))
4157                 return 1;
4158         return 0;
4159 }
4160
4161 static void tp_perf_event_destroy(struct perf_event *event)
4162 {
4163         ftrace_profile_disable(event->attr.config);
4164 }
4165
4166 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4167 {
4168         /*
4169          * Raw tracepoint data is a severe data leak, only allow root to
4170          * have these.
4171          */
4172         if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4173                         perf_paranoid_tracepoint_raw() &&
4174                         !capable(CAP_SYS_ADMIN))
4175                 return ERR_PTR(-EPERM);
4176
4177         if (ftrace_profile_enable(event->attr.config))
4178                 return NULL;
4179
4180         event->destroy = tp_perf_event_destroy;
4181
4182         return &perf_ops_generic;
4183 }
4184
4185 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4186 {
4187         char *filter_str;
4188         int ret;
4189
4190         if (event->attr.type != PERF_TYPE_TRACEPOINT)
4191                 return -EINVAL;
4192
4193         filter_str = strndup_user(arg, PAGE_SIZE);
4194         if (IS_ERR(filter_str))
4195                 return PTR_ERR(filter_str);
4196
4197         ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4198
4199         kfree(filter_str);
4200         return ret;
4201 }
4202
4203 static void perf_event_free_filter(struct perf_event *event)
4204 {
4205         ftrace_profile_free_filter(event);
4206 }
4207
4208 #else
4209
4210 static int perf_tp_event_match(struct perf_event *event,
4211                                 struct perf_sample_data *data)
4212 {
4213         return 1;
4214 }
4215
4216 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4217 {
4218         return NULL;
4219 }
4220
4221 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4222 {
4223         return -ENOENT;
4224 }
4225
4226 static void perf_event_free_filter(struct perf_event *event)
4227 {
4228 }
4229
4230 #endif /* CONFIG_EVENT_PROFILE */
4231
4232 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4233 static void bp_perf_event_destroy(struct perf_event *event)
4234 {
4235         release_bp_slot(event);
4236 }
4237
4238 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4239 {
4240         int err;
4241         /*
4242          * The breakpoint is already filled if we haven't created the counter
4243          * through perf syscall
4244          * FIXME: manage to get trigerred to NULL if it comes from syscalls
4245          */
4246         if (!bp->callback)
4247                 err = register_perf_hw_breakpoint(bp);
4248         else
4249                 err = __register_perf_hw_breakpoint(bp);
4250         if (err)
4251                 return ERR_PTR(err);
4252
4253         bp->destroy = bp_perf_event_destroy;
4254
4255         return &perf_ops_bp;
4256 }
4257
4258 void perf_bp_event(struct perf_event *bp, void *regs)
4259 {
4260         /* TODO */
4261 }
4262 #else
4263 static void bp_perf_event_destroy(struct perf_event *event)
4264 {
4265 }
4266
4267 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4268 {
4269         return NULL;
4270 }
4271
4272 void perf_bp_event(struct perf_event *bp, void *regs)
4273 {
4274 }
4275 #endif
4276
4277 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4278
4279 static void sw_perf_event_destroy(struct perf_event *event)
4280 {
4281         u64 event_id = event->attr.config;
4282
4283         WARN_ON(event->parent);
4284
4285         atomic_dec(&perf_swevent_enabled[event_id]);
4286 }
4287
4288 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4289 {
4290         const struct pmu *pmu = NULL;
4291         u64 event_id = event->attr.config;
4292
4293         /*
4294          * Software events (currently) can't in general distinguish
4295          * between user, kernel and hypervisor events.
4296          * However, context switches and cpu migrations are considered
4297          * to be kernel events, and page faults are never hypervisor
4298          * events.
4299          */
4300         switch (event_id) {
4301         case PERF_COUNT_SW_CPU_CLOCK:
4302                 pmu = &perf_ops_cpu_clock;
4303
4304                 break;
4305         case PERF_COUNT_SW_TASK_CLOCK:
4306                 /*
4307                  * If the user instantiates this as a per-cpu event,
4308                  * use the cpu_clock event instead.
4309                  */
4310                 if (event->ctx->task)
4311                         pmu = &perf_ops_task_clock;
4312                 else
4313                         pmu = &perf_ops_cpu_clock;
4314
4315                 break;
4316         case PERF_COUNT_SW_PAGE_FAULTS:
4317         case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4318         case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4319         case PERF_COUNT_SW_CONTEXT_SWITCHES:
4320         case PERF_COUNT_SW_CPU_MIGRATIONS:
4321         case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4322         case PERF_COUNT_SW_EMULATION_FAULTS:
4323                 if (!event->parent) {
4324                         atomic_inc(&perf_swevent_enabled[event_id]);
4325                         event->destroy = sw_perf_event_destroy;
4326                 }
4327                 pmu = &perf_ops_generic;
4328                 break;
4329         }
4330
4331         return pmu;
4332 }
4333
4334 /*
4335  * Allocate and initialize a event structure
4336  */
4337 static struct perf_event *
4338 perf_event_alloc(struct perf_event_attr *attr,
4339                    int cpu,
4340                    struct perf_event_context *ctx,
4341                    struct perf_event *group_leader,
4342                    struct perf_event *parent_event,
4343                    perf_callback_t callback,
4344                    gfp_t gfpflags)
4345 {
4346         const struct pmu *pmu;
4347         struct perf_event *event;
4348         struct hw_perf_event *hwc;
4349         long err;
4350
4351         event = kzalloc(sizeof(*event), gfpflags);
4352         if (!event)
4353                 return ERR_PTR(-ENOMEM);
4354
4355         /*
4356          * Single events are their own group leaders, with an
4357          * empty sibling list:
4358          */
4359         if (!group_leader)
4360                 group_leader = event;
4361
4362         mutex_init(&event->child_mutex);
4363         INIT_LIST_HEAD(&event->child_list);
4364
4365         INIT_LIST_HEAD(&event->group_entry);
4366         INIT_LIST_HEAD(&event->event_entry);
4367         INIT_LIST_HEAD(&event->sibling_list);
4368         init_waitqueue_head(&event->waitq);
4369
4370         mutex_init(&event->mmap_mutex);
4371
4372         event->cpu              = cpu;
4373         event->attr             = *attr;
4374         event->group_leader     = group_leader;
4375         event->pmu              = NULL;
4376         event->ctx              = ctx;
4377         event->oncpu            = -1;
4378
4379         event->parent           = parent_event;
4380
4381         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
4382         event->id               = atomic64_inc_return(&perf_event_id);
4383
4384         event->state            = PERF_EVENT_STATE_INACTIVE;
4385
4386         if (!callback && parent_event)
4387                 callback = parent_event->callback;
4388         
4389         event->callback = callback;
4390
4391         if (attr->disabled)
4392                 event->state = PERF_EVENT_STATE_OFF;
4393
4394         pmu = NULL;
4395
4396         hwc = &event->hw;
4397         hwc->sample_period = attr->sample_period;
4398         if (attr->freq && attr->sample_freq)
4399                 hwc->sample_period = 1;
4400         hwc->last_period = hwc->sample_period;
4401
4402         atomic64_set(&hwc->period_left, hwc->sample_period);
4403
4404         /*
4405          * we currently do not support PERF_FORMAT_GROUP on inherited events
4406          */
4407         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4408                 goto done;
4409
4410         switch (attr->type) {
4411         case PERF_TYPE_RAW:
4412         case PERF_TYPE_HARDWARE:
4413         case PERF_TYPE_HW_CACHE:
4414                 pmu = hw_perf_event_init(event);
4415                 break;
4416
4417         case PERF_TYPE_SOFTWARE:
4418                 pmu = sw_perf_event_init(event);
4419                 break;
4420
4421         case PERF_TYPE_TRACEPOINT:
4422                 pmu = tp_perf_event_init(event);
4423                 break;
4424
4425         case PERF_TYPE_BREAKPOINT:
4426                 pmu = bp_perf_event_init(event);
4427                 break;
4428
4429
4430         default:
4431                 break;
4432         }
4433 done:
4434         err = 0;
4435         if (!pmu)
4436                 err = -EINVAL;
4437         else if (IS_ERR(pmu))
4438                 err = PTR_ERR(pmu);
4439
4440         if (err) {
4441                 if (event->ns)
4442                         put_pid_ns(event->ns);
4443                 kfree(event);
4444                 return ERR_PTR(err);
4445         }
4446
4447         event->pmu = pmu;
4448
4449         if (!event->parent) {
4450                 atomic_inc(&nr_events);
4451                 if (event->attr.mmap)
4452                         atomic_inc(&nr_mmap_events);
4453                 if (event->attr.comm)
4454                         atomic_inc(&nr_comm_events);
4455                 if (event->attr.task)
4456                         atomic_inc(&nr_task_events);
4457         }
4458
4459         return event;
4460 }
4461
4462 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4463                           struct perf_event_attr *attr)
4464 {
4465         u32 size;
4466         int ret;
4467
4468         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4469                 return -EFAULT;
4470
4471         /*
4472          * zero the full structure, so that a short copy will be nice.
4473          */
4474         memset(attr, 0, sizeof(*attr));
4475
4476         ret = get_user(size, &uattr->size);
4477         if (ret)
4478                 return ret;
4479
4480         if (size > PAGE_SIZE)   /* silly large */
4481                 goto err_size;
4482
4483         if (!size)              /* abi compat */
4484                 size = PERF_ATTR_SIZE_VER0;
4485
4486         if (size < PERF_ATTR_SIZE_VER0)
4487                 goto err_size;
4488
4489         /*
4490          * If we're handed a bigger struct than we know of,
4491          * ensure all the unknown bits are 0 - i.e. new
4492          * user-space does not rely on any kernel feature
4493          * extensions we dont know about yet.
4494          */
4495         if (size > sizeof(*attr)) {
4496                 unsigned char __user *addr;
4497                 unsigned char __user *end;
4498                 unsigned char val;
4499
4500                 addr = (void __user *)uattr + sizeof(*attr);
4501                 end  = (void __user *)uattr + size;
4502
4503                 for (; addr < end; addr++) {
4504                         ret = get_user(val, addr);
4505                         if (ret)
4506                                 return ret;
4507                         if (val)
4508                                 goto err_size;
4509                 }
4510                 size = sizeof(*attr);
4511         }
4512
4513         ret = copy_from_user(attr, uattr, size);
4514         if (ret)
4515                 return -EFAULT;
4516
4517         /*
4518          * If the type exists, the corresponding creation will verify
4519          * the attr->config.
4520          */
4521         if (attr->type >= PERF_TYPE_MAX)
4522                 return -EINVAL;
4523
4524         if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3)
4525                 return -EINVAL;
4526
4527         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4528                 return -EINVAL;
4529
4530         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4531                 return -EINVAL;
4532
4533 out:
4534         return ret;
4535
4536 err_size:
4537         put_user(sizeof(*attr), &uattr->size);
4538         ret = -E2BIG;
4539         goto out;
4540 }
4541
4542 static int perf_event_set_output(struct perf_event *event, int output_fd)
4543 {
4544         struct perf_event *output_event = NULL;
4545         struct file *output_file = NULL;
4546         struct perf_event *old_output;
4547         int fput_needed = 0;
4548         int ret = -EINVAL;
4549
4550         if (!output_fd)
4551                 goto set;
4552
4553         output_file = fget_light(output_fd, &fput_needed);
4554         if (!output_file)
4555                 return -EBADF;
4556
4557         if (output_file->f_op != &perf_fops)
4558                 goto out;
4559
4560         output_event = output_file->private_data;
4561
4562         /* Don't chain output fds */
4563         if (output_event->output)
4564                 goto out;
4565
4566         /* Don't set an output fd when we already have an output channel */
4567         if (event->data)
4568                 goto out;
4569
4570         atomic_long_inc(&output_file->f_count);
4571
4572 set:
4573         mutex_lock(&event->mmap_mutex);
4574         old_output = event->output;
4575         rcu_assign_pointer(event->output, output_event);
4576         mutex_unlock(&event->mmap_mutex);
4577
4578         if (old_output) {
4579                 /*
4580                  * we need to make sure no existing perf_output_*()
4581                  * is still referencing this event.
4582                  */
4583                 synchronize_rcu();
4584                 fput(old_output->filp);
4585         }
4586
4587         ret = 0;
4588 out:
4589         fput_light(output_file, fput_needed);
4590         return ret;
4591 }
4592
4593 /**
4594  * sys_perf_event_open - open a performance event, associate it to a task/cpu
4595  *
4596  * @attr_uptr:  event_id type attributes for monitoring/sampling
4597  * @pid:                target pid
4598  * @cpu:                target cpu
4599  * @group_fd:           group leader event fd
4600  */
4601 SYSCALL_DEFINE5(perf_event_open,
4602                 struct perf_event_attr __user *, attr_uptr,
4603                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4604 {
4605         struct perf_event *event, *group_leader;
4606         struct perf_event_attr attr;
4607         struct perf_event_context *ctx;
4608         struct file *event_file = NULL;
4609         struct file *group_file = NULL;
4610         int fput_needed = 0;
4611         int fput_needed2 = 0;
4612         int err;
4613
4614         /* for future expandability... */
4615         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4616                 return -EINVAL;
4617
4618         err = perf_copy_attr(attr_uptr, &attr);
4619         if (err)
4620                 return err;
4621
4622         if (!attr.exclude_kernel) {
4623                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4624                         return -EACCES;
4625         }
4626
4627         if (attr.freq) {
4628                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4629                         return -EINVAL;
4630         }
4631
4632         /*
4633          * Get the target context (task or percpu):
4634          */
4635         ctx = find_get_context(pid, cpu);
4636         if (IS_ERR(ctx))
4637                 return PTR_ERR(ctx);
4638
4639         /*
4640          * Look up the group leader (we will attach this event to it):
4641          */
4642         group_leader = NULL;
4643         if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4644                 err = -EINVAL;
4645                 group_file = fget_light(group_fd, &fput_needed);
4646                 if (!group_file)
4647                         goto err_put_context;
4648                 if (group_file->f_op != &perf_fops)
4649                         goto err_put_context;
4650
4651                 group_leader = group_file->private_data;
4652                 /*
4653                  * Do not allow a recursive hierarchy (this new sibling
4654                  * becoming part of another group-sibling):
4655                  */
4656                 if (group_leader->group_leader != group_leader)
4657                         goto err_put_context;
4658                 /*
4659                  * Do not allow to attach to a group in a different
4660                  * task or CPU context:
4661                  */
4662                 if (group_leader->ctx != ctx)
4663                         goto err_put_context;
4664                 /*
4665                  * Only a group leader can be exclusive or pinned
4666                  */
4667                 if (attr.exclusive || attr.pinned)
4668                         goto err_put_context;
4669         }
4670
4671         event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4672                                      NULL, NULL, GFP_KERNEL);
4673         err = PTR_ERR(event);
4674         if (IS_ERR(event))
4675                 goto err_put_context;
4676
4677         err = anon_inode_getfd("[perf_event]", &perf_fops, event, 0);
4678         if (err < 0)
4679                 goto err_free_put_context;
4680
4681         event_file = fget_light(err, &fput_needed2);
4682         if (!event_file)
4683                 goto err_free_put_context;
4684
4685         if (flags & PERF_FLAG_FD_OUTPUT) {
4686                 err = perf_event_set_output(event, group_fd);
4687                 if (err)
4688                         goto err_fput_free_put_context;
4689         }
4690
4691         event->filp = event_file;
4692         WARN_ON_ONCE(ctx->parent_ctx);
4693         mutex_lock(&ctx->mutex);
4694         perf_install_in_context(ctx, event, cpu);
4695         ++ctx->generation;
4696         mutex_unlock(&ctx->mutex);
4697
4698         event->owner = current;
4699         get_task_struct(current);
4700         mutex_lock(&current->perf_event_mutex);
4701         list_add_tail(&event->owner_entry, &current->perf_event_list);
4702         mutex_unlock(&current->perf_event_mutex);
4703
4704 err_fput_free_put_context:
4705         fput_light(event_file, fput_needed2);
4706
4707 err_free_put_context:
4708         if (err < 0)
4709                 kfree(event);
4710
4711 err_put_context:
4712         if (err < 0)
4713                 put_ctx(ctx);
4714
4715         fput_light(group_file, fput_needed);
4716
4717         return err;
4718 }
4719
4720 /**
4721  * perf_event_create_kernel_counter
4722  *
4723  * @attr: attributes of the counter to create
4724  * @cpu: cpu in which the counter is bound
4725  * @pid: task to profile
4726  */
4727 struct perf_event *
4728 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4729                                  pid_t pid, perf_callback_t callback)
4730 {
4731         struct perf_event *event;
4732         struct perf_event_context *ctx;
4733         int err;
4734
4735         /*
4736          * Get the target context (task or percpu):
4737          */
4738
4739         ctx = find_get_context(pid, cpu);
4740         if (IS_ERR(ctx))
4741                 return NULL;
4742
4743         event = perf_event_alloc(attr, cpu, ctx, NULL,
4744                                      NULL, callback, GFP_KERNEL);
4745         err = PTR_ERR(event);
4746         if (IS_ERR(event))
4747                 goto err_put_context;
4748
4749         event->filp = NULL;
4750         WARN_ON_ONCE(ctx->parent_ctx);
4751         mutex_lock(&ctx->mutex);
4752         perf_install_in_context(ctx, event, cpu);
4753         ++ctx->generation;
4754         mutex_unlock(&ctx->mutex);
4755
4756         event->owner = current;
4757         get_task_struct(current);
4758         mutex_lock(&current->perf_event_mutex);
4759         list_add_tail(&event->owner_entry, &current->perf_event_list);
4760         mutex_unlock(&current->perf_event_mutex);
4761
4762         return event;
4763
4764 err_put_context:
4765         if (err < 0)
4766                 put_ctx(ctx);
4767
4768         return NULL;
4769 }
4770 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4771
4772 /*
4773  * inherit a event from parent task to child task:
4774  */
4775 static struct perf_event *
4776 inherit_event(struct perf_event *parent_event,
4777               struct task_struct *parent,
4778               struct perf_event_context *parent_ctx,
4779               struct task_struct *child,
4780               struct perf_event *group_leader,
4781               struct perf_event_context *child_ctx)
4782 {
4783         struct perf_event *child_event;
4784
4785         /*
4786          * Instead of creating recursive hierarchies of events,
4787          * we link inherited events back to the original parent,
4788          * which has a filp for sure, which we use as the reference
4789          * count:
4790          */
4791         if (parent_event->parent)
4792                 parent_event = parent_event->parent;
4793
4794         child_event = perf_event_alloc(&parent_event->attr,
4795                                            parent_event->cpu, child_ctx,
4796                                            group_leader, parent_event,
4797                                            NULL, GFP_KERNEL);
4798         if (IS_ERR(child_event))
4799                 return child_event;
4800         get_ctx(child_ctx);
4801
4802         /*
4803          * Make the child state follow the state of the parent event,
4804          * not its attr.disabled bit.  We hold the parent's mutex,
4805          * so we won't race with perf_event_{en, dis}able_family.
4806          */
4807         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
4808                 child_event->state = PERF_EVENT_STATE_INACTIVE;
4809         else
4810                 child_event->state = PERF_EVENT_STATE_OFF;
4811
4812         if (parent_event->attr.freq)
4813                 child_event->hw.sample_period = parent_event->hw.sample_period;
4814
4815         child_event->overflow_handler = parent_event->overflow_handler;
4816
4817         /*
4818          * Link it up in the child's context:
4819          */
4820         add_event_to_ctx(child_event, child_ctx);
4821
4822         /*
4823          * Get a reference to the parent filp - we will fput it
4824          * when the child event exits. This is safe to do because
4825          * we are in the parent and we know that the filp still
4826          * exists and has a nonzero count:
4827          */
4828         atomic_long_inc(&parent_event->filp->f_count);
4829
4830         /*
4831          * Link this into the parent event's child list
4832          */
4833         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4834         mutex_lock(&parent_event->child_mutex);
4835         list_add_tail(&child_event->child_list, &parent_event->child_list);
4836         mutex_unlock(&parent_event->child_mutex);
4837
4838         return child_event;
4839 }
4840
4841 static int inherit_group(struct perf_event *parent_event,
4842               struct task_struct *parent,
4843               struct perf_event_context *parent_ctx,
4844               struct task_struct *child,
4845               struct perf_event_context *child_ctx)
4846 {
4847         struct perf_event *leader;
4848         struct perf_event *sub;
4849         struct perf_event *child_ctr;
4850
4851         leader = inherit_event(parent_event, parent, parent_ctx,
4852                                  child, NULL, child_ctx);
4853         if (IS_ERR(leader))
4854                 return PTR_ERR(leader);
4855         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
4856                 child_ctr = inherit_event(sub, parent, parent_ctx,
4857                                             child, leader, child_ctx);
4858                 if (IS_ERR(child_ctr))
4859                         return PTR_ERR(child_ctr);
4860         }
4861         return 0;
4862 }
4863
4864 static void sync_child_event(struct perf_event *child_event,
4865                                struct task_struct *child)
4866 {
4867         struct perf_event *parent_event = child_event->parent;
4868         u64 child_val;
4869
4870         if (child_event->attr.inherit_stat)
4871                 perf_event_read_event(child_event, child);
4872
4873         child_val = atomic64_read(&child_event->count);
4874
4875         /*
4876          * Add back the child's count to the parent's count:
4877          */
4878         atomic64_add(child_val, &parent_event->count);
4879         atomic64_add(child_event->total_time_enabled,
4880                      &parent_event->child_total_time_enabled);
4881         atomic64_add(child_event->total_time_running,
4882                      &parent_event->child_total_time_running);
4883
4884         /*
4885          * Remove this event from the parent's list
4886          */
4887         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4888         mutex_lock(&parent_event->child_mutex);
4889         list_del_init(&child_event->child_list);
4890         mutex_unlock(&parent_event->child_mutex);
4891
4892         /*
4893          * Release the parent event, if this was the last
4894          * reference to it.
4895          */
4896         fput(parent_event->filp);
4897 }
4898
4899 static void
4900 __perf_event_exit_task(struct perf_event *child_event,
4901                          struct perf_event_context *child_ctx,
4902                          struct task_struct *child)
4903 {
4904         struct perf_event *parent_event;
4905
4906         update_event_times(child_event);
4907         perf_event_remove_from_context(child_event);
4908
4909         parent_event = child_event->parent;
4910         /*
4911          * It can happen that parent exits first, and has events
4912          * that are still around due to the child reference. These
4913          * events need to be zapped - but otherwise linger.
4914          */
4915         if (parent_event) {
4916                 sync_child_event(child_event, child);
4917                 free_event(child_event);
4918         }
4919 }
4920
4921 /*
4922  * When a child task exits, feed back event values to parent events.
4923  */
4924 void perf_event_exit_task(struct task_struct *child)
4925 {
4926         struct perf_event *child_event, *tmp;
4927         struct perf_event_context *child_ctx;
4928         unsigned long flags;
4929
4930         if (likely(!child->perf_event_ctxp)) {
4931                 perf_event_task(child, NULL, 0);
4932                 return;
4933         }
4934
4935         local_irq_save(flags);
4936         /*
4937          * We can't reschedule here because interrupts are disabled,
4938          * and either child is current or it is a task that can't be
4939          * scheduled, so we are now safe from rescheduling changing
4940          * our context.
4941          */
4942         child_ctx = child->perf_event_ctxp;
4943         __perf_event_task_sched_out(child_ctx);
4944
4945         /*
4946          * Take the context lock here so that if find_get_context is
4947          * reading child->perf_event_ctxp, we wait until it has
4948          * incremented the context's refcount before we do put_ctx below.
4949          */
4950         spin_lock(&child_ctx->lock);
4951         child->perf_event_ctxp = NULL;
4952         /*
4953          * If this context is a clone; unclone it so it can't get
4954          * swapped to another process while we're removing all
4955          * the events from it.
4956          */
4957         unclone_ctx(child_ctx);
4958         spin_unlock_irqrestore(&child_ctx->lock, flags);
4959
4960         /*
4961          * Report the task dead after unscheduling the events so that we
4962          * won't get any samples after PERF_RECORD_EXIT. We can however still
4963          * get a few PERF_RECORD_READ events.
4964          */
4965         perf_event_task(child, child_ctx, 0);
4966
4967         /*
4968          * We can recurse on the same lock type through:
4969          *
4970          *   __perf_event_exit_task()
4971          *     sync_child_event()
4972          *       fput(parent_event->filp)
4973          *         perf_release()
4974          *           mutex_lock(&ctx->mutex)
4975          *
4976          * But since its the parent context it won't be the same instance.
4977          */
4978         mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
4979
4980 again:
4981         list_for_each_entry_safe(child_event, tmp, &child_ctx->group_list,
4982                                  group_entry)
4983                 __perf_event_exit_task(child_event, child_ctx, child);
4984
4985         /*
4986          * If the last event was a group event, it will have appended all
4987          * its siblings to the list, but we obtained 'tmp' before that which
4988          * will still point to the list head terminating the iteration.
4989          */
4990         if (!list_empty(&child_ctx->group_list))
4991                 goto again;
4992
4993         mutex_unlock(&child_ctx->mutex);
4994
4995         put_ctx(child_ctx);
4996 }
4997
4998 /*
4999  * free an unexposed, unused context as created by inheritance by
5000  * init_task below, used by fork() in case of fail.
5001  */
5002 void perf_event_free_task(struct task_struct *task)
5003 {
5004         struct perf_event_context *ctx = task->perf_event_ctxp;
5005         struct perf_event *event, *tmp;
5006
5007         if (!ctx)
5008                 return;
5009
5010         mutex_lock(&ctx->mutex);
5011 again:
5012         list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry) {
5013                 struct perf_event *parent = event->parent;
5014
5015                 if (WARN_ON_ONCE(!parent))
5016                         continue;
5017
5018                 mutex_lock(&parent->child_mutex);
5019                 list_del_init(&event->child_list);
5020                 mutex_unlock(&parent->child_mutex);
5021
5022                 fput(parent->filp);
5023
5024                 list_del_event(event, ctx);
5025                 free_event(event);
5026         }
5027
5028         if (!list_empty(&ctx->group_list))
5029                 goto again;
5030
5031         mutex_unlock(&ctx->mutex);
5032
5033         put_ctx(ctx);
5034 }
5035
5036 /*
5037  * Initialize the perf_event context in task_struct
5038  */
5039 int perf_event_init_task(struct task_struct *child)
5040 {
5041         struct perf_event_context *child_ctx, *parent_ctx;
5042         struct perf_event_context *cloned_ctx;
5043         struct perf_event *event;
5044         struct task_struct *parent = current;
5045         int inherited_all = 1;
5046         int ret = 0;
5047
5048         child->perf_event_ctxp = NULL;
5049
5050         mutex_init(&child->perf_event_mutex);
5051         INIT_LIST_HEAD(&child->perf_event_list);
5052
5053         if (likely(!parent->perf_event_ctxp))
5054                 return 0;
5055
5056         /*
5057          * This is executed from the parent task context, so inherit
5058          * events that have been marked for cloning.
5059          * First allocate and initialize a context for the child.
5060          */
5061
5062         child_ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
5063         if (!child_ctx)
5064                 return -ENOMEM;
5065
5066         __perf_event_init_context(child_ctx, child);
5067         child->perf_event_ctxp = child_ctx;
5068         get_task_struct(child);
5069
5070         /*
5071          * If the parent's context is a clone, pin it so it won't get
5072          * swapped under us.
5073          */
5074         parent_ctx = perf_pin_task_context(parent);
5075
5076         /*
5077          * No need to check if parent_ctx != NULL here; since we saw
5078          * it non-NULL earlier, the only reason for it to become NULL
5079          * is if we exit, and since we're currently in the middle of
5080          * a fork we can't be exiting at the same time.
5081          */
5082
5083         /*
5084          * Lock the parent list. No need to lock the child - not PID
5085          * hashed yet and not running, so nobody can access it.
5086          */
5087         mutex_lock(&parent_ctx->mutex);
5088
5089         /*
5090          * We dont have to disable NMIs - we are only looking at
5091          * the list, not manipulating it:
5092          */
5093         list_for_each_entry(event, &parent_ctx->group_list, group_entry) {
5094
5095                 if (!event->attr.inherit) {
5096                         inherited_all = 0;
5097                         continue;
5098                 }
5099
5100                 ret = inherit_group(event, parent, parent_ctx,
5101                                              child, child_ctx);
5102                 if (ret) {
5103                         inherited_all = 0;
5104                         break;
5105                 }
5106         }
5107
5108         if (inherited_all) {
5109                 /*
5110                  * Mark the child context as a clone of the parent
5111                  * context, or of whatever the parent is a clone of.
5112                  * Note that if the parent is a clone, it could get
5113                  * uncloned at any point, but that doesn't matter
5114                  * because the list of events and the generation
5115                  * count can't have changed since we took the mutex.
5116                  */
5117                 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5118                 if (cloned_ctx) {
5119                         child_ctx->parent_ctx = cloned_ctx;
5120                         child_ctx->parent_gen = parent_ctx->parent_gen;
5121                 } else {
5122                         child_ctx->parent_ctx = parent_ctx;
5123                         child_ctx->parent_gen = parent_ctx->generation;
5124                 }
5125                 get_ctx(child_ctx->parent_ctx);
5126         }
5127
5128         mutex_unlock(&parent_ctx->mutex);
5129
5130         perf_unpin_context(parent_ctx);
5131
5132         return ret;
5133 }
5134
5135 static void __cpuinit perf_event_init_cpu(int cpu)
5136 {
5137         struct perf_cpu_context *cpuctx;
5138
5139         cpuctx = &per_cpu(perf_cpu_context, cpu);
5140         __perf_event_init_context(&cpuctx->ctx, NULL);
5141
5142         spin_lock(&perf_resource_lock);
5143         cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5144         spin_unlock(&perf_resource_lock);
5145
5146         hw_perf_event_setup(cpu);
5147 }
5148
5149 #ifdef CONFIG_HOTPLUG_CPU
5150 static void __perf_event_exit_cpu(void *info)
5151 {
5152         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5153         struct perf_event_context *ctx = &cpuctx->ctx;
5154         struct perf_event *event, *tmp;
5155
5156         list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry)
5157                 __perf_event_remove_from_context(event);
5158 }
5159 static void perf_event_exit_cpu(int cpu)
5160 {
5161         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5162         struct perf_event_context *ctx = &cpuctx->ctx;
5163
5164         mutex_lock(&ctx->mutex);
5165         smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5166         mutex_unlock(&ctx->mutex);
5167 }
5168 #else
5169 static inline void perf_event_exit_cpu(int cpu) { }
5170 #endif
5171
5172 static int __cpuinit
5173 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5174 {
5175         unsigned int cpu = (long)hcpu;
5176
5177         switch (action) {
5178
5179         case CPU_UP_PREPARE:
5180         case CPU_UP_PREPARE_FROZEN:
5181                 perf_event_init_cpu(cpu);
5182                 break;
5183
5184         case CPU_ONLINE:
5185         case CPU_ONLINE_FROZEN:
5186                 hw_perf_event_setup_online(cpu);
5187                 break;
5188
5189         case CPU_DOWN_PREPARE:
5190         case CPU_DOWN_PREPARE_FROZEN:
5191                 perf_event_exit_cpu(cpu);
5192                 break;
5193
5194         default:
5195                 break;
5196         }
5197
5198         return NOTIFY_OK;
5199 }
5200
5201 /*
5202  * This has to have a higher priority than migration_notifier in sched.c.
5203  */
5204 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5205         .notifier_call          = perf_cpu_notify,
5206         .priority               = 20,
5207 };
5208
5209 void __init perf_event_init(void)
5210 {
5211         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5212                         (void *)(long)smp_processor_id());
5213         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5214                         (void *)(long)smp_processor_id());
5215         register_cpu_notifier(&perf_cpu_nb);
5216 }
5217
5218 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
5219 {
5220         return sprintf(buf, "%d\n", perf_reserved_percpu);
5221 }
5222
5223 static ssize_t
5224 perf_set_reserve_percpu(struct sysdev_class *class,
5225                         const char *buf,
5226                         size_t count)
5227 {
5228         struct perf_cpu_context *cpuctx;
5229         unsigned long val;
5230         int err, cpu, mpt;
5231
5232         err = strict_strtoul(buf, 10, &val);
5233         if (err)
5234                 return err;
5235         if (val > perf_max_events)
5236                 return -EINVAL;
5237
5238         spin_lock(&perf_resource_lock);
5239         perf_reserved_percpu = val;
5240         for_each_online_cpu(cpu) {
5241                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5242                 spin_lock_irq(&cpuctx->ctx.lock);
5243                 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5244                           perf_max_events - perf_reserved_percpu);
5245                 cpuctx->max_pertask = mpt;
5246                 spin_unlock_irq(&cpuctx->ctx.lock);
5247         }
5248         spin_unlock(&perf_resource_lock);
5249
5250         return count;
5251 }
5252
5253 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
5254 {
5255         return sprintf(buf, "%d\n", perf_overcommit);
5256 }
5257
5258 static ssize_t
5259 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
5260 {
5261         unsigned long val;
5262         int err;
5263
5264         err = strict_strtoul(buf, 10, &val);
5265         if (err)
5266                 return err;
5267         if (val > 1)
5268                 return -EINVAL;
5269
5270         spin_lock(&perf_resource_lock);
5271         perf_overcommit = val;
5272         spin_unlock(&perf_resource_lock);
5273
5274         return count;
5275 }
5276
5277 static SYSDEV_CLASS_ATTR(
5278                                 reserve_percpu,
5279                                 0644,
5280                                 perf_show_reserve_percpu,
5281                                 perf_set_reserve_percpu
5282                         );
5283
5284 static SYSDEV_CLASS_ATTR(
5285                                 overcommit,
5286                                 0644,
5287                                 perf_show_overcommit,
5288                                 perf_set_overcommit
5289                         );
5290
5291 static struct attribute *perfclass_attrs[] = {
5292         &attr_reserve_percpu.attr,
5293         &attr_overcommit.attr,
5294         NULL
5295 };
5296
5297 static struct attribute_group perfclass_attr_group = {
5298         .attrs                  = perfclass_attrs,
5299         .name                   = "perf_events",
5300 };
5301
5302 static int __init perf_event_sysfs_init(void)
5303 {
5304         return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5305                                   &perfclass_attr_group);
5306 }
5307 device_initcall(perf_event_sysfs_init);