mm: compaction: add /proc trigger for memory compaction
[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/slab.h>
19 #include <linux/hash.h>
20 #include <linux/sysfs.h>
21 #include <linux/dcache.h>
22 #include <linux/percpu.h>
23 #include <linux/ptrace.h>
24 #include <linux/vmstat.h>
25 #include <linux/vmalloc.h>
26 #include <linux/hardirq.h>
27 #include <linux/rculist.h>
28 #include <linux/uaccess.h>
29 #include <linux/syscalls.h>
30 #include <linux/anon_inodes.h>
31 #include <linux/kernel_stat.h>
32 #include <linux/perf_event.h>
33 #include <linux/ftrace_event.h>
34 #include <linux/hw_breakpoint.h>
35
36 #include <asm/irq_regs.h>
37
38 /*
39  * Each CPU has a list of per CPU events:
40  */
41 static DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
42
43 int perf_max_events __read_mostly = 1;
44 static int perf_reserved_percpu __read_mostly;
45 static int perf_overcommit __read_mostly = 1;
46
47 static atomic_t nr_events __read_mostly;
48 static atomic_t nr_mmap_events __read_mostly;
49 static atomic_t nr_comm_events __read_mostly;
50 static atomic_t nr_task_events __read_mostly;
51
52 /*
53  * perf event paranoia level:
54  *  -1 - not paranoid at all
55  *   0 - disallow raw tracepoint access for unpriv
56  *   1 - disallow cpu events for unpriv
57  *   2 - disallow kernel profiling for unpriv
58  */
59 int sysctl_perf_event_paranoid __read_mostly = 1;
60
61 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
62
63 /*
64  * max perf event sample rate
65  */
66 int sysctl_perf_event_sample_rate __read_mostly = 100000;
67
68 static atomic64_t perf_event_id;
69
70 /*
71  * Lock for (sysadmin-configurable) event reservations:
72  */
73 static DEFINE_SPINLOCK(perf_resource_lock);
74
75 /*
76  * Architecture provided APIs - weak aliases:
77  */
78 extern __weak const struct pmu *hw_perf_event_init(struct perf_event *event)
79 {
80         return NULL;
81 }
82
83 void __weak hw_perf_disable(void)               { barrier(); }
84 void __weak hw_perf_enable(void)                { barrier(); }
85
86 void __weak perf_event_print_debug(void)        { }
87
88 static DEFINE_PER_CPU(int, perf_disable_count);
89
90 void perf_disable(void)
91 {
92         if (!__get_cpu_var(perf_disable_count)++)
93                 hw_perf_disable();
94 }
95
96 void perf_enable(void)
97 {
98         if (!--__get_cpu_var(perf_disable_count))
99                 hw_perf_enable();
100 }
101
102 static void get_ctx(struct perf_event_context *ctx)
103 {
104         WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
105 }
106
107 static void free_ctx(struct rcu_head *head)
108 {
109         struct perf_event_context *ctx;
110
111         ctx = container_of(head, struct perf_event_context, rcu_head);
112         kfree(ctx);
113 }
114
115 static void put_ctx(struct perf_event_context *ctx)
116 {
117         if (atomic_dec_and_test(&ctx->refcount)) {
118                 if (ctx->parent_ctx)
119                         put_ctx(ctx->parent_ctx);
120                 if (ctx->task)
121                         put_task_struct(ctx->task);
122                 call_rcu(&ctx->rcu_head, free_ctx);
123         }
124 }
125
126 static void unclone_ctx(struct perf_event_context *ctx)
127 {
128         if (ctx->parent_ctx) {
129                 put_ctx(ctx->parent_ctx);
130                 ctx->parent_ctx = NULL;
131         }
132 }
133
134 /*
135  * If we inherit events we want to return the parent event id
136  * to userspace.
137  */
138 static u64 primary_event_id(struct perf_event *event)
139 {
140         u64 id = event->id;
141
142         if (event->parent)
143                 id = event->parent->id;
144
145         return id;
146 }
147
148 /*
149  * Get the perf_event_context for a task and lock it.
150  * This has to cope with with the fact that until it is locked,
151  * the context could get moved to another task.
152  */
153 static struct perf_event_context *
154 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
155 {
156         struct perf_event_context *ctx;
157
158         rcu_read_lock();
159  retry:
160         ctx = rcu_dereference(task->perf_event_ctxp);
161         if (ctx) {
162                 /*
163                  * If this context is a clone of another, it might
164                  * get swapped for another underneath us by
165                  * perf_event_task_sched_out, though the
166                  * rcu_read_lock() protects us from any context
167                  * getting freed.  Lock the context and check if it
168                  * got swapped before we could get the lock, and retry
169                  * if so.  If we locked the right context, then it
170                  * can't get swapped on us any more.
171                  */
172                 raw_spin_lock_irqsave(&ctx->lock, *flags);
173                 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
174                         raw_spin_unlock_irqrestore(&ctx->lock, *flags);
175                         goto retry;
176                 }
177
178                 if (!atomic_inc_not_zero(&ctx->refcount)) {
179                         raw_spin_unlock_irqrestore(&ctx->lock, *flags);
180                         ctx = NULL;
181                 }
182         }
183         rcu_read_unlock();
184         return ctx;
185 }
186
187 /*
188  * Get the context for a task and increment its pin_count so it
189  * can't get swapped to another task.  This also increments its
190  * reference count so that the context can't get freed.
191  */
192 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
193 {
194         struct perf_event_context *ctx;
195         unsigned long flags;
196
197         ctx = perf_lock_task_context(task, &flags);
198         if (ctx) {
199                 ++ctx->pin_count;
200                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
201         }
202         return ctx;
203 }
204
205 static void perf_unpin_context(struct perf_event_context *ctx)
206 {
207         unsigned long flags;
208
209         raw_spin_lock_irqsave(&ctx->lock, flags);
210         --ctx->pin_count;
211         raw_spin_unlock_irqrestore(&ctx->lock, flags);
212         put_ctx(ctx);
213 }
214
215 static inline u64 perf_clock(void)
216 {
217         return cpu_clock(raw_smp_processor_id());
218 }
219
220 /*
221  * Update the record of the current time in a context.
222  */
223 static void update_context_time(struct perf_event_context *ctx)
224 {
225         u64 now = perf_clock();
226
227         ctx->time += now - ctx->timestamp;
228         ctx->timestamp = now;
229 }
230
231 /*
232  * Update the total_time_enabled and total_time_running fields for a event.
233  */
234 static void update_event_times(struct perf_event *event)
235 {
236         struct perf_event_context *ctx = event->ctx;
237         u64 run_end;
238
239         if (event->state < PERF_EVENT_STATE_INACTIVE ||
240             event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
241                 return;
242
243         if (ctx->is_active)
244                 run_end = ctx->time;
245         else
246                 run_end = event->tstamp_stopped;
247
248         event->total_time_enabled = run_end - event->tstamp_enabled;
249
250         if (event->state == PERF_EVENT_STATE_INACTIVE)
251                 run_end = event->tstamp_stopped;
252         else
253                 run_end = ctx->time;
254
255         event->total_time_running = run_end - event->tstamp_running;
256 }
257
258 /*
259  * Update total_time_enabled and total_time_running for all events in a group.
260  */
261 static void update_group_times(struct perf_event *leader)
262 {
263         struct perf_event *event;
264
265         update_event_times(leader);
266         list_for_each_entry(event, &leader->sibling_list, group_entry)
267                 update_event_times(event);
268 }
269
270 static struct list_head *
271 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
272 {
273         if (event->attr.pinned)
274                 return &ctx->pinned_groups;
275         else
276                 return &ctx->flexible_groups;
277 }
278
279 /*
280  * Add a event from the lists for its context.
281  * Must be called with ctx->mutex and ctx->lock held.
282  */
283 static void
284 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
285 {
286         struct perf_event *group_leader = event->group_leader;
287
288         /*
289          * Depending on whether it is a standalone or sibling event,
290          * add it straight to the context's event list, or to the group
291          * leader's sibling list:
292          */
293         if (group_leader == event) {
294                 struct list_head *list;
295
296                 if (is_software_event(event))
297                         event->group_flags |= PERF_GROUP_SOFTWARE;
298
299                 list = ctx_group_list(event, ctx);
300                 list_add_tail(&event->group_entry, list);
301         } else {
302                 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
303                     !is_software_event(event))
304                         group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
305
306                 list_add_tail(&event->group_entry, &group_leader->sibling_list);
307                 group_leader->nr_siblings++;
308         }
309
310         list_add_rcu(&event->event_entry, &ctx->event_list);
311         ctx->nr_events++;
312         if (event->attr.inherit_stat)
313                 ctx->nr_stat++;
314 }
315
316 /*
317  * Remove a event from the lists for its context.
318  * Must be called with ctx->mutex and ctx->lock held.
319  */
320 static void
321 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
322 {
323         if (list_empty(&event->group_entry))
324                 return;
325         ctx->nr_events--;
326         if (event->attr.inherit_stat)
327                 ctx->nr_stat--;
328
329         list_del_init(&event->group_entry);
330         list_del_rcu(&event->event_entry);
331
332         if (event->group_leader != event)
333                 event->group_leader->nr_siblings--;
334
335         update_group_times(event);
336
337         /*
338          * If event was in error state, then keep it
339          * that way, otherwise bogus counts will be
340          * returned on read(). The only way to get out
341          * of error state is by explicit re-enabling
342          * of the event
343          */
344         if (event->state > PERF_EVENT_STATE_OFF)
345                 event->state = PERF_EVENT_STATE_OFF;
346 }
347
348 static void
349 perf_destroy_group(struct perf_event *event, struct perf_event_context *ctx)
350 {
351         struct perf_event *sibling, *tmp;
352
353         /*
354          * If this was a group event with sibling events then
355          * upgrade the siblings to singleton events by adding them
356          * to the context list directly:
357          */
358         list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
359                 struct list_head *list;
360
361                 list = ctx_group_list(event, ctx);
362                 list_move_tail(&sibling->group_entry, list);
363                 sibling->group_leader = sibling;
364
365                 /* Inherit group flags from the previous leader */
366                 sibling->group_flags = event->group_flags;
367         }
368 }
369
370 static void
371 event_sched_out(struct perf_event *event,
372                   struct perf_cpu_context *cpuctx,
373                   struct perf_event_context *ctx)
374 {
375         if (event->state != PERF_EVENT_STATE_ACTIVE)
376                 return;
377
378         event->state = PERF_EVENT_STATE_INACTIVE;
379         if (event->pending_disable) {
380                 event->pending_disable = 0;
381                 event->state = PERF_EVENT_STATE_OFF;
382         }
383         event->tstamp_stopped = ctx->time;
384         event->pmu->disable(event);
385         event->oncpu = -1;
386
387         if (!is_software_event(event))
388                 cpuctx->active_oncpu--;
389         ctx->nr_active--;
390         if (event->attr.exclusive || !cpuctx->active_oncpu)
391                 cpuctx->exclusive = 0;
392 }
393
394 static void
395 group_sched_out(struct perf_event *group_event,
396                 struct perf_cpu_context *cpuctx,
397                 struct perf_event_context *ctx)
398 {
399         struct perf_event *event;
400
401         if (group_event->state != PERF_EVENT_STATE_ACTIVE)
402                 return;
403
404         event_sched_out(group_event, cpuctx, ctx);
405
406         /*
407          * Schedule out siblings (if any):
408          */
409         list_for_each_entry(event, &group_event->sibling_list, group_entry)
410                 event_sched_out(event, cpuctx, ctx);
411
412         if (group_event->attr.exclusive)
413                 cpuctx->exclusive = 0;
414 }
415
416 /*
417  * Cross CPU call to remove a performance event
418  *
419  * We disable the event on the hardware level first. After that we
420  * remove it from the context list.
421  */
422 static void __perf_event_remove_from_context(void *info)
423 {
424         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
425         struct perf_event *event = info;
426         struct perf_event_context *ctx = event->ctx;
427
428         /*
429          * If this is a task context, we need to check whether it is
430          * the current task context of this cpu. If not it has been
431          * scheduled out before the smp call arrived.
432          */
433         if (ctx->task && cpuctx->task_ctx != ctx)
434                 return;
435
436         raw_spin_lock(&ctx->lock);
437         /*
438          * Protect the list operation against NMI by disabling the
439          * events on a global level.
440          */
441         perf_disable();
442
443         event_sched_out(event, cpuctx, ctx);
444
445         list_del_event(event, ctx);
446
447         if (!ctx->task) {
448                 /*
449                  * Allow more per task events with respect to the
450                  * reservation:
451                  */
452                 cpuctx->max_pertask =
453                         min(perf_max_events - ctx->nr_events,
454                             perf_max_events - perf_reserved_percpu);
455         }
456
457         perf_enable();
458         raw_spin_unlock(&ctx->lock);
459 }
460
461
462 /*
463  * Remove the event from a task's (or a CPU's) list of events.
464  *
465  * Must be called with ctx->mutex held.
466  *
467  * CPU events are removed with a smp call. For task events we only
468  * call when the task is on a CPU.
469  *
470  * If event->ctx is a cloned context, callers must make sure that
471  * every task struct that event->ctx->task could possibly point to
472  * remains valid.  This is OK when called from perf_release since
473  * that only calls us on the top-level context, which can't be a clone.
474  * When called from perf_event_exit_task, it's OK because the
475  * context has been detached from its task.
476  */
477 static void perf_event_remove_from_context(struct perf_event *event)
478 {
479         struct perf_event_context *ctx = event->ctx;
480         struct task_struct *task = ctx->task;
481
482         if (!task) {
483                 /*
484                  * Per cpu events are removed via an smp call and
485                  * the removal is always successful.
486                  */
487                 smp_call_function_single(event->cpu,
488                                          __perf_event_remove_from_context,
489                                          event, 1);
490                 return;
491         }
492
493 retry:
494         task_oncpu_function_call(task, __perf_event_remove_from_context,
495                                  event);
496
497         raw_spin_lock_irq(&ctx->lock);
498         /*
499          * If the context is active we need to retry the smp call.
500          */
501         if (ctx->nr_active && !list_empty(&event->group_entry)) {
502                 raw_spin_unlock_irq(&ctx->lock);
503                 goto retry;
504         }
505
506         /*
507          * The lock prevents that this context is scheduled in so we
508          * can remove the event safely, if the call above did not
509          * succeed.
510          */
511         if (!list_empty(&event->group_entry))
512                 list_del_event(event, ctx);
513         raw_spin_unlock_irq(&ctx->lock);
514 }
515
516 /*
517  * Cross CPU call to disable a performance event
518  */
519 static void __perf_event_disable(void *info)
520 {
521         struct perf_event *event = info;
522         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
523         struct perf_event_context *ctx = event->ctx;
524
525         /*
526          * If this is a per-task event, need to check whether this
527          * event's task is the current task on this cpu.
528          */
529         if (ctx->task && cpuctx->task_ctx != ctx)
530                 return;
531
532         raw_spin_lock(&ctx->lock);
533
534         /*
535          * If the event is on, turn it off.
536          * If it is in error state, leave it in error state.
537          */
538         if (event->state >= PERF_EVENT_STATE_INACTIVE) {
539                 update_context_time(ctx);
540                 update_group_times(event);
541                 if (event == event->group_leader)
542                         group_sched_out(event, cpuctx, ctx);
543                 else
544                         event_sched_out(event, cpuctx, ctx);
545                 event->state = PERF_EVENT_STATE_OFF;
546         }
547
548         raw_spin_unlock(&ctx->lock);
549 }
550
551 /*
552  * Disable a event.
553  *
554  * If event->ctx is a cloned context, callers must make sure that
555  * every task struct that event->ctx->task could possibly point to
556  * remains valid.  This condition is satisifed when called through
557  * perf_event_for_each_child or perf_event_for_each because they
558  * hold the top-level event's child_mutex, so any descendant that
559  * goes to exit will block in sync_child_event.
560  * When called from perf_pending_event it's OK because event->ctx
561  * is the current context on this CPU and preemption is disabled,
562  * hence we can't get into perf_event_task_sched_out for this context.
563  */
564 void perf_event_disable(struct perf_event *event)
565 {
566         struct perf_event_context *ctx = event->ctx;
567         struct task_struct *task = ctx->task;
568
569         if (!task) {
570                 /*
571                  * Disable the event on the cpu that it's on
572                  */
573                 smp_call_function_single(event->cpu, __perf_event_disable,
574                                          event, 1);
575                 return;
576         }
577
578  retry:
579         task_oncpu_function_call(task, __perf_event_disable, event);
580
581         raw_spin_lock_irq(&ctx->lock);
582         /*
583          * If the event is still active, we need to retry the cross-call.
584          */
585         if (event->state == PERF_EVENT_STATE_ACTIVE) {
586                 raw_spin_unlock_irq(&ctx->lock);
587                 goto retry;
588         }
589
590         /*
591          * Since we have the lock this context can't be scheduled
592          * in, so we can change the state safely.
593          */
594         if (event->state == PERF_EVENT_STATE_INACTIVE) {
595                 update_group_times(event);
596                 event->state = PERF_EVENT_STATE_OFF;
597         }
598
599         raw_spin_unlock_irq(&ctx->lock);
600 }
601
602 static int
603 event_sched_in(struct perf_event *event,
604                  struct perf_cpu_context *cpuctx,
605                  struct perf_event_context *ctx)
606 {
607         if (event->state <= PERF_EVENT_STATE_OFF)
608                 return 0;
609
610         event->state = PERF_EVENT_STATE_ACTIVE;
611         event->oncpu = smp_processor_id();
612         /*
613          * The new state must be visible before we turn it on in the hardware:
614          */
615         smp_wmb();
616
617         if (event->pmu->enable(event)) {
618                 event->state = PERF_EVENT_STATE_INACTIVE;
619                 event->oncpu = -1;
620                 return -EAGAIN;
621         }
622
623         event->tstamp_running += ctx->time - event->tstamp_stopped;
624
625         if (!is_software_event(event))
626                 cpuctx->active_oncpu++;
627         ctx->nr_active++;
628
629         if (event->attr.exclusive)
630                 cpuctx->exclusive = 1;
631
632         return 0;
633 }
634
635 static int
636 group_sched_in(struct perf_event *group_event,
637                struct perf_cpu_context *cpuctx,
638                struct perf_event_context *ctx)
639 {
640         struct perf_event *event, *partial_group = NULL;
641         const struct pmu *pmu = group_event->pmu;
642         bool txn = false;
643         int ret;
644
645         if (group_event->state == PERF_EVENT_STATE_OFF)
646                 return 0;
647
648         /* Check if group transaction availabe */
649         if (pmu->start_txn)
650                 txn = true;
651
652         if (txn)
653                 pmu->start_txn(pmu);
654
655         if (event_sched_in(group_event, cpuctx, ctx))
656                 return -EAGAIN;
657
658         /*
659          * Schedule in siblings as one group (if any):
660          */
661         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
662                 if (event_sched_in(event, cpuctx, ctx)) {
663                         partial_group = event;
664                         goto group_error;
665                 }
666         }
667
668         if (!txn)
669                 return 0;
670
671         ret = pmu->commit_txn(pmu);
672         if (!ret) {
673                 pmu->cancel_txn(pmu);
674                 return 0;
675         }
676
677 group_error:
678         if (txn)
679                 pmu->cancel_txn(pmu);
680
681         /*
682          * Groups can be scheduled in as one unit only, so undo any
683          * partial group before returning:
684          */
685         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
686                 if (event == partial_group)
687                         break;
688                 event_sched_out(event, cpuctx, ctx);
689         }
690         event_sched_out(group_event, cpuctx, ctx);
691
692         return -EAGAIN;
693 }
694
695 /*
696  * Work out whether we can put this event group on the CPU now.
697  */
698 static int group_can_go_on(struct perf_event *event,
699                            struct perf_cpu_context *cpuctx,
700                            int can_add_hw)
701 {
702         /*
703          * Groups consisting entirely of software events can always go on.
704          */
705         if (event->group_flags & PERF_GROUP_SOFTWARE)
706                 return 1;
707         /*
708          * If an exclusive group is already on, no other hardware
709          * events can go on.
710          */
711         if (cpuctx->exclusive)
712                 return 0;
713         /*
714          * If this group is exclusive and there are already
715          * events on the CPU, it can't go on.
716          */
717         if (event->attr.exclusive && cpuctx->active_oncpu)
718                 return 0;
719         /*
720          * Otherwise, try to add it if all previous groups were able
721          * to go on.
722          */
723         return can_add_hw;
724 }
725
726 static void add_event_to_ctx(struct perf_event *event,
727                                struct perf_event_context *ctx)
728 {
729         list_add_event(event, ctx);
730         event->tstamp_enabled = ctx->time;
731         event->tstamp_running = ctx->time;
732         event->tstamp_stopped = ctx->time;
733 }
734
735 /*
736  * Cross CPU call to install and enable a performance event
737  *
738  * Must be called with ctx->mutex held
739  */
740 static void __perf_install_in_context(void *info)
741 {
742         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
743         struct perf_event *event = info;
744         struct perf_event_context *ctx = event->ctx;
745         struct perf_event *leader = event->group_leader;
746         int err;
747
748         /*
749          * If this is a task context, we need to check whether it is
750          * the current task context of this cpu. If not it has been
751          * scheduled out before the smp call arrived.
752          * Or possibly this is the right context but it isn't
753          * on this cpu because it had no events.
754          */
755         if (ctx->task && cpuctx->task_ctx != ctx) {
756                 if (cpuctx->task_ctx || ctx->task != current)
757                         return;
758                 cpuctx->task_ctx = ctx;
759         }
760
761         raw_spin_lock(&ctx->lock);
762         ctx->is_active = 1;
763         update_context_time(ctx);
764
765         /*
766          * Protect the list operation against NMI by disabling the
767          * events on a global level. NOP for non NMI based events.
768          */
769         perf_disable();
770
771         add_event_to_ctx(event, ctx);
772
773         if (event->cpu != -1 && event->cpu != smp_processor_id())
774                 goto unlock;
775
776         /*
777          * Don't put the event on if it is disabled or if
778          * it is in a group and the group isn't on.
779          */
780         if (event->state != PERF_EVENT_STATE_INACTIVE ||
781             (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
782                 goto unlock;
783
784         /*
785          * An exclusive event can't go on if there are already active
786          * hardware events, and no hardware event can go on if there
787          * is already an exclusive event on.
788          */
789         if (!group_can_go_on(event, cpuctx, 1))
790                 err = -EEXIST;
791         else
792                 err = event_sched_in(event, cpuctx, ctx);
793
794         if (err) {
795                 /*
796                  * This event couldn't go on.  If it is in a group
797                  * then we have to pull the whole group off.
798                  * If the event group is pinned then put it in error state.
799                  */
800                 if (leader != event)
801                         group_sched_out(leader, cpuctx, ctx);
802                 if (leader->attr.pinned) {
803                         update_group_times(leader);
804                         leader->state = PERF_EVENT_STATE_ERROR;
805                 }
806         }
807
808         if (!err && !ctx->task && cpuctx->max_pertask)
809                 cpuctx->max_pertask--;
810
811  unlock:
812         perf_enable();
813
814         raw_spin_unlock(&ctx->lock);
815 }
816
817 /*
818  * Attach a performance event to a context
819  *
820  * First we add the event to the list with the hardware enable bit
821  * in event->hw_config cleared.
822  *
823  * If the event is attached to a task which is on a CPU we use a smp
824  * call to enable it in the task context. The task might have been
825  * scheduled away, but we check this in the smp call again.
826  *
827  * Must be called with ctx->mutex held.
828  */
829 static void
830 perf_install_in_context(struct perf_event_context *ctx,
831                         struct perf_event *event,
832                         int cpu)
833 {
834         struct task_struct *task = ctx->task;
835
836         if (!task) {
837                 /*
838                  * Per cpu events are installed via an smp call and
839                  * the install is always successful.
840                  */
841                 smp_call_function_single(cpu, __perf_install_in_context,
842                                          event, 1);
843                 return;
844         }
845
846 retry:
847         task_oncpu_function_call(task, __perf_install_in_context,
848                                  event);
849
850         raw_spin_lock_irq(&ctx->lock);
851         /*
852          * we need to retry the smp call.
853          */
854         if (ctx->is_active && list_empty(&event->group_entry)) {
855                 raw_spin_unlock_irq(&ctx->lock);
856                 goto retry;
857         }
858
859         /*
860          * The lock prevents that this context is scheduled in so we
861          * can add the event safely, if it the call above did not
862          * succeed.
863          */
864         if (list_empty(&event->group_entry))
865                 add_event_to_ctx(event, ctx);
866         raw_spin_unlock_irq(&ctx->lock);
867 }
868
869 /*
870  * Put a event into inactive state and update time fields.
871  * Enabling the leader of a group effectively enables all
872  * the group members that aren't explicitly disabled, so we
873  * have to update their ->tstamp_enabled also.
874  * Note: this works for group members as well as group leaders
875  * since the non-leader members' sibling_lists will be empty.
876  */
877 static void __perf_event_mark_enabled(struct perf_event *event,
878                                         struct perf_event_context *ctx)
879 {
880         struct perf_event *sub;
881
882         event->state = PERF_EVENT_STATE_INACTIVE;
883         event->tstamp_enabled = ctx->time - event->total_time_enabled;
884         list_for_each_entry(sub, &event->sibling_list, group_entry)
885                 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
886                         sub->tstamp_enabled =
887                                 ctx->time - sub->total_time_enabled;
888 }
889
890 /*
891  * Cross CPU call to enable a performance event
892  */
893 static void __perf_event_enable(void *info)
894 {
895         struct perf_event *event = info;
896         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
897         struct perf_event_context *ctx = event->ctx;
898         struct perf_event *leader = event->group_leader;
899         int err;
900
901         /*
902          * If this is a per-task event, need to check whether this
903          * event's task is the current task on this cpu.
904          */
905         if (ctx->task && cpuctx->task_ctx != ctx) {
906                 if (cpuctx->task_ctx || ctx->task != current)
907                         return;
908                 cpuctx->task_ctx = ctx;
909         }
910
911         raw_spin_lock(&ctx->lock);
912         ctx->is_active = 1;
913         update_context_time(ctx);
914
915         if (event->state >= PERF_EVENT_STATE_INACTIVE)
916                 goto unlock;
917         __perf_event_mark_enabled(event, ctx);
918
919         if (event->cpu != -1 && event->cpu != smp_processor_id())
920                 goto unlock;
921
922         /*
923          * If the event is in a group and isn't the group leader,
924          * then don't put it on unless the group is on.
925          */
926         if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
927                 goto unlock;
928
929         if (!group_can_go_on(event, cpuctx, 1)) {
930                 err = -EEXIST;
931         } else {
932                 perf_disable();
933                 if (event == leader)
934                         err = group_sched_in(event, cpuctx, ctx);
935                 else
936                         err = event_sched_in(event, cpuctx, ctx);
937                 perf_enable();
938         }
939
940         if (err) {
941                 /*
942                  * If this event can't go on and it's part of a
943                  * group, then the whole group has to come off.
944                  */
945                 if (leader != event)
946                         group_sched_out(leader, cpuctx, ctx);
947                 if (leader->attr.pinned) {
948                         update_group_times(leader);
949                         leader->state = PERF_EVENT_STATE_ERROR;
950                 }
951         }
952
953  unlock:
954         raw_spin_unlock(&ctx->lock);
955 }
956
957 /*
958  * Enable a event.
959  *
960  * If event->ctx is a cloned context, callers must make sure that
961  * every task struct that event->ctx->task could possibly point to
962  * remains valid.  This condition is satisfied when called through
963  * perf_event_for_each_child or perf_event_for_each as described
964  * for perf_event_disable.
965  */
966 void perf_event_enable(struct perf_event *event)
967 {
968         struct perf_event_context *ctx = event->ctx;
969         struct task_struct *task = ctx->task;
970
971         if (!task) {
972                 /*
973                  * Enable the event on the cpu that it's on
974                  */
975                 smp_call_function_single(event->cpu, __perf_event_enable,
976                                          event, 1);
977                 return;
978         }
979
980         raw_spin_lock_irq(&ctx->lock);
981         if (event->state >= PERF_EVENT_STATE_INACTIVE)
982                 goto out;
983
984         /*
985          * If the event is in error state, clear that first.
986          * That way, if we see the event in error state below, we
987          * know that it has gone back into error state, as distinct
988          * from the task having been scheduled away before the
989          * cross-call arrived.
990          */
991         if (event->state == PERF_EVENT_STATE_ERROR)
992                 event->state = PERF_EVENT_STATE_OFF;
993
994  retry:
995         raw_spin_unlock_irq(&ctx->lock);
996         task_oncpu_function_call(task, __perf_event_enable, event);
997
998         raw_spin_lock_irq(&ctx->lock);
999
1000         /*
1001          * If the context is active and the event is still off,
1002          * we need to retry the cross-call.
1003          */
1004         if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1005                 goto retry;
1006
1007         /*
1008          * Since we have the lock this context can't be scheduled
1009          * in, so we can change the state safely.
1010          */
1011         if (event->state == PERF_EVENT_STATE_OFF)
1012                 __perf_event_mark_enabled(event, ctx);
1013
1014  out:
1015         raw_spin_unlock_irq(&ctx->lock);
1016 }
1017
1018 static int perf_event_refresh(struct perf_event *event, int refresh)
1019 {
1020         /*
1021          * not supported on inherited events
1022          */
1023         if (event->attr.inherit)
1024                 return -EINVAL;
1025
1026         atomic_add(refresh, &event->event_limit);
1027         perf_event_enable(event);
1028
1029         return 0;
1030 }
1031
1032 enum event_type_t {
1033         EVENT_FLEXIBLE = 0x1,
1034         EVENT_PINNED = 0x2,
1035         EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1036 };
1037
1038 static void ctx_sched_out(struct perf_event_context *ctx,
1039                           struct perf_cpu_context *cpuctx,
1040                           enum event_type_t event_type)
1041 {
1042         struct perf_event *event;
1043
1044         raw_spin_lock(&ctx->lock);
1045         ctx->is_active = 0;
1046         if (likely(!ctx->nr_events))
1047                 goto out;
1048         update_context_time(ctx);
1049
1050         perf_disable();
1051         if (!ctx->nr_active)
1052                 goto out_enable;
1053
1054         if (event_type & EVENT_PINNED)
1055                 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1056                         group_sched_out(event, cpuctx, ctx);
1057
1058         if (event_type & EVENT_FLEXIBLE)
1059                 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1060                         group_sched_out(event, cpuctx, ctx);
1061
1062  out_enable:
1063         perf_enable();
1064  out:
1065         raw_spin_unlock(&ctx->lock);
1066 }
1067
1068 /*
1069  * Test whether two contexts are equivalent, i.e. whether they
1070  * have both been cloned from the same version of the same context
1071  * and they both have the same number of enabled events.
1072  * If the number of enabled events is the same, then the set
1073  * of enabled events should be the same, because these are both
1074  * inherited contexts, therefore we can't access individual events
1075  * in them directly with an fd; we can only enable/disable all
1076  * events via prctl, or enable/disable all events in a family
1077  * via ioctl, which will have the same effect on both contexts.
1078  */
1079 static int context_equiv(struct perf_event_context *ctx1,
1080                          struct perf_event_context *ctx2)
1081 {
1082         return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1083                 && ctx1->parent_gen == ctx2->parent_gen
1084                 && !ctx1->pin_count && !ctx2->pin_count;
1085 }
1086
1087 static void __perf_event_sync_stat(struct perf_event *event,
1088                                      struct perf_event *next_event)
1089 {
1090         u64 value;
1091
1092         if (!event->attr.inherit_stat)
1093                 return;
1094
1095         /*
1096          * Update the event value, we cannot use perf_event_read()
1097          * because we're in the middle of a context switch and have IRQs
1098          * disabled, which upsets smp_call_function_single(), however
1099          * we know the event must be on the current CPU, therefore we
1100          * don't need to use it.
1101          */
1102         switch (event->state) {
1103         case PERF_EVENT_STATE_ACTIVE:
1104                 event->pmu->read(event);
1105                 /* fall-through */
1106
1107         case PERF_EVENT_STATE_INACTIVE:
1108                 update_event_times(event);
1109                 break;
1110
1111         default:
1112                 break;
1113         }
1114
1115         /*
1116          * In order to keep per-task stats reliable we need to flip the event
1117          * values when we flip the contexts.
1118          */
1119         value = atomic64_read(&next_event->count);
1120         value = atomic64_xchg(&event->count, value);
1121         atomic64_set(&next_event->count, value);
1122
1123         swap(event->total_time_enabled, next_event->total_time_enabled);
1124         swap(event->total_time_running, next_event->total_time_running);
1125
1126         /*
1127          * Since we swizzled the values, update the user visible data too.
1128          */
1129         perf_event_update_userpage(event);
1130         perf_event_update_userpage(next_event);
1131 }
1132
1133 #define list_next_entry(pos, member) \
1134         list_entry(pos->member.next, typeof(*pos), member)
1135
1136 static void perf_event_sync_stat(struct perf_event_context *ctx,
1137                                    struct perf_event_context *next_ctx)
1138 {
1139         struct perf_event *event, *next_event;
1140
1141         if (!ctx->nr_stat)
1142                 return;
1143
1144         update_context_time(ctx);
1145
1146         event = list_first_entry(&ctx->event_list,
1147                                    struct perf_event, event_entry);
1148
1149         next_event = list_first_entry(&next_ctx->event_list,
1150                                         struct perf_event, event_entry);
1151
1152         while (&event->event_entry != &ctx->event_list &&
1153                &next_event->event_entry != &next_ctx->event_list) {
1154
1155                 __perf_event_sync_stat(event, next_event);
1156
1157                 event = list_next_entry(event, event_entry);
1158                 next_event = list_next_entry(next_event, event_entry);
1159         }
1160 }
1161
1162 /*
1163  * Called from scheduler to remove the events of the current task,
1164  * with interrupts disabled.
1165  *
1166  * We stop each event and update the event value in event->count.
1167  *
1168  * This does not protect us against NMI, but disable()
1169  * sets the disabled bit in the control field of event _before_
1170  * accessing the event control register. If a NMI hits, then it will
1171  * not restart the event.
1172  */
1173 void perf_event_task_sched_out(struct task_struct *task,
1174                                  struct task_struct *next)
1175 {
1176         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1177         struct perf_event_context *ctx = task->perf_event_ctxp;
1178         struct perf_event_context *next_ctx;
1179         struct perf_event_context *parent;
1180         int do_switch = 1;
1181
1182         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0);
1183
1184         if (likely(!ctx || !cpuctx->task_ctx))
1185                 return;
1186
1187         rcu_read_lock();
1188         parent = rcu_dereference(ctx->parent_ctx);
1189         next_ctx = next->perf_event_ctxp;
1190         if (parent && next_ctx &&
1191             rcu_dereference(next_ctx->parent_ctx) == parent) {
1192                 /*
1193                  * Looks like the two contexts are clones, so we might be
1194                  * able to optimize the context switch.  We lock both
1195                  * contexts and check that they are clones under the
1196                  * lock (including re-checking that neither has been
1197                  * uncloned in the meantime).  It doesn't matter which
1198                  * order we take the locks because no other cpu could
1199                  * be trying to lock both of these tasks.
1200                  */
1201                 raw_spin_lock(&ctx->lock);
1202                 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1203                 if (context_equiv(ctx, next_ctx)) {
1204                         /*
1205                          * XXX do we need a memory barrier of sorts
1206                          * wrt to rcu_dereference() of perf_event_ctxp
1207                          */
1208                         task->perf_event_ctxp = next_ctx;
1209                         next->perf_event_ctxp = ctx;
1210                         ctx->task = next;
1211                         next_ctx->task = task;
1212                         do_switch = 0;
1213
1214                         perf_event_sync_stat(ctx, next_ctx);
1215                 }
1216                 raw_spin_unlock(&next_ctx->lock);
1217                 raw_spin_unlock(&ctx->lock);
1218         }
1219         rcu_read_unlock();
1220
1221         if (do_switch) {
1222                 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1223                 cpuctx->task_ctx = NULL;
1224         }
1225 }
1226
1227 static void task_ctx_sched_out(struct perf_event_context *ctx,
1228                                enum event_type_t event_type)
1229 {
1230         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1231
1232         if (!cpuctx->task_ctx)
1233                 return;
1234
1235         if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1236                 return;
1237
1238         ctx_sched_out(ctx, cpuctx, event_type);
1239         cpuctx->task_ctx = NULL;
1240 }
1241
1242 /*
1243  * Called with IRQs disabled
1244  */
1245 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1246 {
1247         task_ctx_sched_out(ctx, EVENT_ALL);
1248 }
1249
1250 /*
1251  * Called with IRQs disabled
1252  */
1253 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1254                               enum event_type_t event_type)
1255 {
1256         ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1257 }
1258
1259 static void
1260 ctx_pinned_sched_in(struct perf_event_context *ctx,
1261                     struct perf_cpu_context *cpuctx)
1262 {
1263         struct perf_event *event;
1264
1265         list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1266                 if (event->state <= PERF_EVENT_STATE_OFF)
1267                         continue;
1268                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1269                         continue;
1270
1271                 if (group_can_go_on(event, cpuctx, 1))
1272                         group_sched_in(event, cpuctx, ctx);
1273
1274                 /*
1275                  * If this pinned group hasn't been scheduled,
1276                  * put it in error state.
1277                  */
1278                 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1279                         update_group_times(event);
1280                         event->state = PERF_EVENT_STATE_ERROR;
1281                 }
1282         }
1283 }
1284
1285 static void
1286 ctx_flexible_sched_in(struct perf_event_context *ctx,
1287                       struct perf_cpu_context *cpuctx)
1288 {
1289         struct perf_event *event;
1290         int can_add_hw = 1;
1291
1292         list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1293                 /* Ignore events in OFF or ERROR state */
1294                 if (event->state <= PERF_EVENT_STATE_OFF)
1295                         continue;
1296                 /*
1297                  * Listen to the 'cpu' scheduling filter constraint
1298                  * of events:
1299                  */
1300                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1301                         continue;
1302
1303                 if (group_can_go_on(event, cpuctx, can_add_hw))
1304                         if (group_sched_in(event, cpuctx, ctx))
1305                                 can_add_hw = 0;
1306         }
1307 }
1308
1309 static void
1310 ctx_sched_in(struct perf_event_context *ctx,
1311              struct perf_cpu_context *cpuctx,
1312              enum event_type_t event_type)
1313 {
1314         raw_spin_lock(&ctx->lock);
1315         ctx->is_active = 1;
1316         if (likely(!ctx->nr_events))
1317                 goto out;
1318
1319         ctx->timestamp = perf_clock();
1320
1321         perf_disable();
1322
1323         /*
1324          * First go through the list and put on any pinned groups
1325          * in order to give them the best chance of going on.
1326          */
1327         if (event_type & EVENT_PINNED)
1328                 ctx_pinned_sched_in(ctx, cpuctx);
1329
1330         /* Then walk through the lower prio flexible groups */
1331         if (event_type & EVENT_FLEXIBLE)
1332                 ctx_flexible_sched_in(ctx, cpuctx);
1333
1334         perf_enable();
1335  out:
1336         raw_spin_unlock(&ctx->lock);
1337 }
1338
1339 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1340                              enum event_type_t event_type)
1341 {
1342         struct perf_event_context *ctx = &cpuctx->ctx;
1343
1344         ctx_sched_in(ctx, cpuctx, event_type);
1345 }
1346
1347 static void task_ctx_sched_in(struct task_struct *task,
1348                               enum event_type_t event_type)
1349 {
1350         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1351         struct perf_event_context *ctx = task->perf_event_ctxp;
1352
1353         if (likely(!ctx))
1354                 return;
1355         if (cpuctx->task_ctx == ctx)
1356                 return;
1357         ctx_sched_in(ctx, cpuctx, event_type);
1358         cpuctx->task_ctx = ctx;
1359 }
1360 /*
1361  * Called from scheduler to add the events of the current task
1362  * with interrupts disabled.
1363  *
1364  * We restore the event value and then enable it.
1365  *
1366  * This does not protect us against NMI, but enable()
1367  * sets the enabled bit in the control field of event _before_
1368  * accessing the event control register. If a NMI hits, then it will
1369  * keep the event running.
1370  */
1371 void perf_event_task_sched_in(struct task_struct *task)
1372 {
1373         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1374         struct perf_event_context *ctx = task->perf_event_ctxp;
1375
1376         if (likely(!ctx))
1377                 return;
1378
1379         if (cpuctx->task_ctx == ctx)
1380                 return;
1381
1382         perf_disable();
1383
1384         /*
1385          * We want to keep the following priority order:
1386          * cpu pinned (that don't need to move), task pinned,
1387          * cpu flexible, task flexible.
1388          */
1389         cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1390
1391         ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1392         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1393         ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1394
1395         cpuctx->task_ctx = ctx;
1396
1397         perf_enable();
1398 }
1399
1400 #define MAX_INTERRUPTS (~0ULL)
1401
1402 static void perf_log_throttle(struct perf_event *event, int enable);
1403
1404 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1405 {
1406         u64 frequency = event->attr.sample_freq;
1407         u64 sec = NSEC_PER_SEC;
1408         u64 divisor, dividend;
1409
1410         int count_fls, nsec_fls, frequency_fls, sec_fls;
1411
1412         count_fls = fls64(count);
1413         nsec_fls = fls64(nsec);
1414         frequency_fls = fls64(frequency);
1415         sec_fls = 30;
1416
1417         /*
1418          * We got @count in @nsec, with a target of sample_freq HZ
1419          * the target period becomes:
1420          *
1421          *             @count * 10^9
1422          * period = -------------------
1423          *          @nsec * sample_freq
1424          *
1425          */
1426
1427         /*
1428          * Reduce accuracy by one bit such that @a and @b converge
1429          * to a similar magnitude.
1430          */
1431 #define REDUCE_FLS(a, b)                \
1432 do {                                    \
1433         if (a##_fls > b##_fls) {        \
1434                 a >>= 1;                \
1435                 a##_fls--;              \
1436         } else {                        \
1437                 b >>= 1;                \
1438                 b##_fls--;              \
1439         }                               \
1440 } while (0)
1441
1442         /*
1443          * Reduce accuracy until either term fits in a u64, then proceed with
1444          * the other, so that finally we can do a u64/u64 division.
1445          */
1446         while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1447                 REDUCE_FLS(nsec, frequency);
1448                 REDUCE_FLS(sec, count);
1449         }
1450
1451         if (count_fls + sec_fls > 64) {
1452                 divisor = nsec * frequency;
1453
1454                 while (count_fls + sec_fls > 64) {
1455                         REDUCE_FLS(count, sec);
1456                         divisor >>= 1;
1457                 }
1458
1459                 dividend = count * sec;
1460         } else {
1461                 dividend = count * sec;
1462
1463                 while (nsec_fls + frequency_fls > 64) {
1464                         REDUCE_FLS(nsec, frequency);
1465                         dividend >>= 1;
1466                 }
1467
1468                 divisor = nsec * frequency;
1469         }
1470
1471         return div64_u64(dividend, divisor);
1472 }
1473
1474 static void perf_event_stop(struct perf_event *event)
1475 {
1476         if (!event->pmu->stop)
1477                 return event->pmu->disable(event);
1478
1479         return event->pmu->stop(event);
1480 }
1481
1482 static int perf_event_start(struct perf_event *event)
1483 {
1484         if (!event->pmu->start)
1485                 return event->pmu->enable(event);
1486
1487         return event->pmu->start(event);
1488 }
1489
1490 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1491 {
1492         struct hw_perf_event *hwc = &event->hw;
1493         u64 period, sample_period;
1494         s64 delta;
1495
1496         period = perf_calculate_period(event, nsec, count);
1497
1498         delta = (s64)(period - hwc->sample_period);
1499         delta = (delta + 7) / 8; /* low pass filter */
1500
1501         sample_period = hwc->sample_period + delta;
1502
1503         if (!sample_period)
1504                 sample_period = 1;
1505
1506         hwc->sample_period = sample_period;
1507
1508         if (atomic64_read(&hwc->period_left) > 8*sample_period) {
1509                 perf_disable();
1510                 perf_event_stop(event);
1511                 atomic64_set(&hwc->period_left, 0);
1512                 perf_event_start(event);
1513                 perf_enable();
1514         }
1515 }
1516
1517 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1518 {
1519         struct perf_event *event;
1520         struct hw_perf_event *hwc;
1521         u64 interrupts, now;
1522         s64 delta;
1523
1524         raw_spin_lock(&ctx->lock);
1525         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1526                 if (event->state != PERF_EVENT_STATE_ACTIVE)
1527                         continue;
1528
1529                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1530                         continue;
1531
1532                 hwc = &event->hw;
1533
1534                 interrupts = hwc->interrupts;
1535                 hwc->interrupts = 0;
1536
1537                 /*
1538                  * unthrottle events on the tick
1539                  */
1540                 if (interrupts == MAX_INTERRUPTS) {
1541                         perf_log_throttle(event, 1);
1542                         perf_disable();
1543                         event->pmu->unthrottle(event);
1544                         perf_enable();
1545                 }
1546
1547                 if (!event->attr.freq || !event->attr.sample_freq)
1548                         continue;
1549
1550                 perf_disable();
1551                 event->pmu->read(event);
1552                 now = atomic64_read(&event->count);
1553                 delta = now - hwc->freq_count_stamp;
1554                 hwc->freq_count_stamp = now;
1555
1556                 if (delta > 0)
1557                         perf_adjust_period(event, TICK_NSEC, delta);
1558                 perf_enable();
1559         }
1560         raw_spin_unlock(&ctx->lock);
1561 }
1562
1563 /*
1564  * Round-robin a context's events:
1565  */
1566 static void rotate_ctx(struct perf_event_context *ctx)
1567 {
1568         raw_spin_lock(&ctx->lock);
1569
1570         /* Rotate the first entry last of non-pinned groups */
1571         list_rotate_left(&ctx->flexible_groups);
1572
1573         raw_spin_unlock(&ctx->lock);
1574 }
1575
1576 void perf_event_task_tick(struct task_struct *curr)
1577 {
1578         struct perf_cpu_context *cpuctx;
1579         struct perf_event_context *ctx;
1580         int rotate = 0;
1581
1582         if (!atomic_read(&nr_events))
1583                 return;
1584
1585         cpuctx = &__get_cpu_var(perf_cpu_context);
1586         if (cpuctx->ctx.nr_events &&
1587             cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1588                 rotate = 1;
1589
1590         ctx = curr->perf_event_ctxp;
1591         if (ctx && ctx->nr_events && ctx->nr_events != ctx->nr_active)
1592                 rotate = 1;
1593
1594         perf_ctx_adjust_freq(&cpuctx->ctx);
1595         if (ctx)
1596                 perf_ctx_adjust_freq(ctx);
1597
1598         if (!rotate)
1599                 return;
1600
1601         perf_disable();
1602         cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1603         if (ctx)
1604                 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1605
1606         rotate_ctx(&cpuctx->ctx);
1607         if (ctx)
1608                 rotate_ctx(ctx);
1609
1610         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1611         if (ctx)
1612                 task_ctx_sched_in(curr, EVENT_FLEXIBLE);
1613         perf_enable();
1614 }
1615
1616 static int event_enable_on_exec(struct perf_event *event,
1617                                 struct perf_event_context *ctx)
1618 {
1619         if (!event->attr.enable_on_exec)
1620                 return 0;
1621
1622         event->attr.enable_on_exec = 0;
1623         if (event->state >= PERF_EVENT_STATE_INACTIVE)
1624                 return 0;
1625
1626         __perf_event_mark_enabled(event, ctx);
1627
1628         return 1;
1629 }
1630
1631 /*
1632  * Enable all of a task's events that have been marked enable-on-exec.
1633  * This expects task == current.
1634  */
1635 static void perf_event_enable_on_exec(struct task_struct *task)
1636 {
1637         struct perf_event_context *ctx;
1638         struct perf_event *event;
1639         unsigned long flags;
1640         int enabled = 0;
1641         int ret;
1642
1643         local_irq_save(flags);
1644         ctx = task->perf_event_ctxp;
1645         if (!ctx || !ctx->nr_events)
1646                 goto out;
1647
1648         __perf_event_task_sched_out(ctx);
1649
1650         raw_spin_lock(&ctx->lock);
1651
1652         list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1653                 ret = event_enable_on_exec(event, ctx);
1654                 if (ret)
1655                         enabled = 1;
1656         }
1657
1658         list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1659                 ret = event_enable_on_exec(event, ctx);
1660                 if (ret)
1661                         enabled = 1;
1662         }
1663
1664         /*
1665          * Unclone this context if we enabled any event.
1666          */
1667         if (enabled)
1668                 unclone_ctx(ctx);
1669
1670         raw_spin_unlock(&ctx->lock);
1671
1672         perf_event_task_sched_in(task);
1673  out:
1674         local_irq_restore(flags);
1675 }
1676
1677 /*
1678  * Cross CPU call to read the hardware event
1679  */
1680 static void __perf_event_read(void *info)
1681 {
1682         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1683         struct perf_event *event = info;
1684         struct perf_event_context *ctx = event->ctx;
1685
1686         /*
1687          * If this is a task context, we need to check whether it is
1688          * the current task context of this cpu.  If not it has been
1689          * scheduled out before the smp call arrived.  In that case
1690          * event->count would have been updated to a recent sample
1691          * when the event was scheduled out.
1692          */
1693         if (ctx->task && cpuctx->task_ctx != ctx)
1694                 return;
1695
1696         raw_spin_lock(&ctx->lock);
1697         update_context_time(ctx);
1698         update_event_times(event);
1699         raw_spin_unlock(&ctx->lock);
1700
1701         event->pmu->read(event);
1702 }
1703
1704 static u64 perf_event_read(struct perf_event *event)
1705 {
1706         /*
1707          * If event is enabled and currently active on a CPU, update the
1708          * value in the event structure:
1709          */
1710         if (event->state == PERF_EVENT_STATE_ACTIVE) {
1711                 smp_call_function_single(event->oncpu,
1712                                          __perf_event_read, event, 1);
1713         } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1714                 struct perf_event_context *ctx = event->ctx;
1715                 unsigned long flags;
1716
1717                 raw_spin_lock_irqsave(&ctx->lock, flags);
1718                 update_context_time(ctx);
1719                 update_event_times(event);
1720                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1721         }
1722
1723         return atomic64_read(&event->count);
1724 }
1725
1726 /*
1727  * Initialize the perf_event context in a task_struct:
1728  */
1729 static void
1730 __perf_event_init_context(struct perf_event_context *ctx,
1731                             struct task_struct *task)
1732 {
1733         raw_spin_lock_init(&ctx->lock);
1734         mutex_init(&ctx->mutex);
1735         INIT_LIST_HEAD(&ctx->pinned_groups);
1736         INIT_LIST_HEAD(&ctx->flexible_groups);
1737         INIT_LIST_HEAD(&ctx->event_list);
1738         atomic_set(&ctx->refcount, 1);
1739         ctx->task = task;
1740 }
1741
1742 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1743 {
1744         struct perf_event_context *ctx;
1745         struct perf_cpu_context *cpuctx;
1746         struct task_struct *task;
1747         unsigned long flags;
1748         int err;
1749
1750         if (pid == -1 && cpu != -1) {
1751                 /* Must be root to operate on a CPU event: */
1752                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1753                         return ERR_PTR(-EACCES);
1754
1755                 if (cpu < 0 || cpu >= nr_cpumask_bits)
1756                         return ERR_PTR(-EINVAL);
1757
1758                 /*
1759                  * We could be clever and allow to attach a event to an
1760                  * offline CPU and activate it when the CPU comes up, but
1761                  * that's for later.
1762                  */
1763                 if (!cpu_online(cpu))
1764                         return ERR_PTR(-ENODEV);
1765
1766                 cpuctx = &per_cpu(perf_cpu_context, cpu);
1767                 ctx = &cpuctx->ctx;
1768                 get_ctx(ctx);
1769
1770                 return ctx;
1771         }
1772
1773         rcu_read_lock();
1774         if (!pid)
1775                 task = current;
1776         else
1777                 task = find_task_by_vpid(pid);
1778         if (task)
1779                 get_task_struct(task);
1780         rcu_read_unlock();
1781
1782         if (!task)
1783                 return ERR_PTR(-ESRCH);
1784
1785         /*
1786          * Can't attach events to a dying task.
1787          */
1788         err = -ESRCH;
1789         if (task->flags & PF_EXITING)
1790                 goto errout;
1791
1792         /* Reuse ptrace permission checks for now. */
1793         err = -EACCES;
1794         if (!ptrace_may_access(task, PTRACE_MODE_READ))
1795                 goto errout;
1796
1797  retry:
1798         ctx = perf_lock_task_context(task, &flags);
1799         if (ctx) {
1800                 unclone_ctx(ctx);
1801                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1802         }
1803
1804         if (!ctx) {
1805                 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1806                 err = -ENOMEM;
1807                 if (!ctx)
1808                         goto errout;
1809                 __perf_event_init_context(ctx, task);
1810                 get_ctx(ctx);
1811                 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1812                         /*
1813                          * We raced with some other task; use
1814                          * the context they set.
1815                          */
1816                         kfree(ctx);
1817                         goto retry;
1818                 }
1819                 get_task_struct(task);
1820         }
1821
1822         put_task_struct(task);
1823         return ctx;
1824
1825  errout:
1826         put_task_struct(task);
1827         return ERR_PTR(err);
1828 }
1829
1830 static void perf_event_free_filter(struct perf_event *event);
1831
1832 static void free_event_rcu(struct rcu_head *head)
1833 {
1834         struct perf_event *event;
1835
1836         event = container_of(head, struct perf_event, rcu_head);
1837         if (event->ns)
1838                 put_pid_ns(event->ns);
1839         perf_event_free_filter(event);
1840         kfree(event);
1841 }
1842
1843 static void perf_pending_sync(struct perf_event *event);
1844
1845 static void free_event(struct perf_event *event)
1846 {
1847         perf_pending_sync(event);
1848
1849         if (!event->parent) {
1850                 atomic_dec(&nr_events);
1851                 if (event->attr.mmap)
1852                         atomic_dec(&nr_mmap_events);
1853                 if (event->attr.comm)
1854                         atomic_dec(&nr_comm_events);
1855                 if (event->attr.task)
1856                         atomic_dec(&nr_task_events);
1857         }
1858
1859         if (event->output) {
1860                 fput(event->output->filp);
1861                 event->output = NULL;
1862         }
1863
1864         if (event->destroy)
1865                 event->destroy(event);
1866
1867         put_ctx(event->ctx);
1868         call_rcu(&event->rcu_head, free_event_rcu);
1869 }
1870
1871 int perf_event_release_kernel(struct perf_event *event)
1872 {
1873         struct perf_event_context *ctx = event->ctx;
1874
1875         /*
1876          * Remove from the PMU, can't get re-enabled since we got
1877          * here because the last ref went.
1878          */
1879         perf_event_disable(event);
1880
1881         WARN_ON_ONCE(ctx->parent_ctx);
1882         /*
1883          * There are two ways this annotation is useful:
1884          *
1885          *  1) there is a lock recursion from perf_event_exit_task
1886          *     see the comment there.
1887          *
1888          *  2) there is a lock-inversion with mmap_sem through
1889          *     perf_event_read_group(), which takes faults while
1890          *     holding ctx->mutex, however this is called after
1891          *     the last filedesc died, so there is no possibility
1892          *     to trigger the AB-BA case.
1893          */
1894         mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
1895         raw_spin_lock_irq(&ctx->lock);
1896         list_del_event(event, ctx);
1897         perf_destroy_group(event, ctx);
1898         raw_spin_unlock_irq(&ctx->lock);
1899         mutex_unlock(&ctx->mutex);
1900
1901         mutex_lock(&event->owner->perf_event_mutex);
1902         list_del_init(&event->owner_entry);
1903         mutex_unlock(&event->owner->perf_event_mutex);
1904         put_task_struct(event->owner);
1905
1906         free_event(event);
1907
1908         return 0;
1909 }
1910 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1911
1912 /*
1913  * Called when the last reference to the file is gone.
1914  */
1915 static int perf_release(struct inode *inode, struct file *file)
1916 {
1917         struct perf_event *event = file->private_data;
1918
1919         file->private_data = NULL;
1920
1921         return perf_event_release_kernel(event);
1922 }
1923
1924 static int perf_event_read_size(struct perf_event *event)
1925 {
1926         int entry = sizeof(u64); /* value */
1927         int size = 0;
1928         int nr = 1;
1929
1930         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1931                 size += sizeof(u64);
1932
1933         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1934                 size += sizeof(u64);
1935
1936         if (event->attr.read_format & PERF_FORMAT_ID)
1937                 entry += sizeof(u64);
1938
1939         if (event->attr.read_format & PERF_FORMAT_GROUP) {
1940                 nr += event->group_leader->nr_siblings;
1941                 size += sizeof(u64);
1942         }
1943
1944         size += entry * nr;
1945
1946         return size;
1947 }
1948
1949 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1950 {
1951         struct perf_event *child;
1952         u64 total = 0;
1953
1954         *enabled = 0;
1955         *running = 0;
1956
1957         mutex_lock(&event->child_mutex);
1958         total += perf_event_read(event);
1959         *enabled += event->total_time_enabled +
1960                         atomic64_read(&event->child_total_time_enabled);
1961         *running += event->total_time_running +
1962                         atomic64_read(&event->child_total_time_running);
1963
1964         list_for_each_entry(child, &event->child_list, child_list) {
1965                 total += perf_event_read(child);
1966                 *enabled += child->total_time_enabled;
1967                 *running += child->total_time_running;
1968         }
1969         mutex_unlock(&event->child_mutex);
1970
1971         return total;
1972 }
1973 EXPORT_SYMBOL_GPL(perf_event_read_value);
1974
1975 static int perf_event_read_group(struct perf_event *event,
1976                                    u64 read_format, char __user *buf)
1977 {
1978         struct perf_event *leader = event->group_leader, *sub;
1979         int n = 0, size = 0, ret = -EFAULT;
1980         struct perf_event_context *ctx = leader->ctx;
1981         u64 values[5];
1982         u64 count, enabled, running;
1983
1984         mutex_lock(&ctx->mutex);
1985         count = perf_event_read_value(leader, &enabled, &running);
1986
1987         values[n++] = 1 + leader->nr_siblings;
1988         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1989                 values[n++] = enabled;
1990         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1991                 values[n++] = running;
1992         values[n++] = count;
1993         if (read_format & PERF_FORMAT_ID)
1994                 values[n++] = primary_event_id(leader);
1995
1996         size = n * sizeof(u64);
1997
1998         if (copy_to_user(buf, values, size))
1999                 goto unlock;
2000
2001         ret = size;
2002
2003         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2004                 n = 0;
2005
2006                 values[n++] = perf_event_read_value(sub, &enabled, &running);
2007                 if (read_format & PERF_FORMAT_ID)
2008                         values[n++] = primary_event_id(sub);
2009
2010                 size = n * sizeof(u64);
2011
2012                 if (copy_to_user(buf + ret, values, size)) {
2013                         ret = -EFAULT;
2014                         goto unlock;
2015                 }
2016
2017                 ret += size;
2018         }
2019 unlock:
2020         mutex_unlock(&ctx->mutex);
2021
2022         return ret;
2023 }
2024
2025 static int perf_event_read_one(struct perf_event *event,
2026                                  u64 read_format, char __user *buf)
2027 {
2028         u64 enabled, running;
2029         u64 values[4];
2030         int n = 0;
2031
2032         values[n++] = perf_event_read_value(event, &enabled, &running);
2033         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2034                 values[n++] = enabled;
2035         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2036                 values[n++] = running;
2037         if (read_format & PERF_FORMAT_ID)
2038                 values[n++] = primary_event_id(event);
2039
2040         if (copy_to_user(buf, values, n * sizeof(u64)))
2041                 return -EFAULT;
2042
2043         return n * sizeof(u64);
2044 }
2045
2046 /*
2047  * Read the performance event - simple non blocking version for now
2048  */
2049 static ssize_t
2050 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2051 {
2052         u64 read_format = event->attr.read_format;
2053         int ret;
2054
2055         /*
2056          * Return end-of-file for a read on a event that is in
2057          * error state (i.e. because it was pinned but it couldn't be
2058          * scheduled on to the CPU at some point).
2059          */
2060         if (event->state == PERF_EVENT_STATE_ERROR)
2061                 return 0;
2062
2063         if (count < perf_event_read_size(event))
2064                 return -ENOSPC;
2065
2066         WARN_ON_ONCE(event->ctx->parent_ctx);
2067         if (read_format & PERF_FORMAT_GROUP)
2068                 ret = perf_event_read_group(event, read_format, buf);
2069         else
2070                 ret = perf_event_read_one(event, read_format, buf);
2071
2072         return ret;
2073 }
2074
2075 static ssize_t
2076 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2077 {
2078         struct perf_event *event = file->private_data;
2079
2080         return perf_read_hw(event, buf, count);
2081 }
2082
2083 static unsigned int perf_poll(struct file *file, poll_table *wait)
2084 {
2085         struct perf_event *event = file->private_data;
2086         struct perf_mmap_data *data;
2087         unsigned int events = POLL_HUP;
2088
2089         rcu_read_lock();
2090         data = rcu_dereference(event->data);
2091         if (data)
2092                 events = atomic_xchg(&data->poll, 0);
2093         rcu_read_unlock();
2094
2095         poll_wait(file, &event->waitq, wait);
2096
2097         return events;
2098 }
2099
2100 static void perf_event_reset(struct perf_event *event)
2101 {
2102         (void)perf_event_read(event);
2103         atomic64_set(&event->count, 0);
2104         perf_event_update_userpage(event);
2105 }
2106
2107 /*
2108  * Holding the top-level event's child_mutex means that any
2109  * descendant process that has inherited this event will block
2110  * in sync_child_event if it goes to exit, thus satisfying the
2111  * task existence requirements of perf_event_enable/disable.
2112  */
2113 static void perf_event_for_each_child(struct perf_event *event,
2114                                         void (*func)(struct perf_event *))
2115 {
2116         struct perf_event *child;
2117
2118         WARN_ON_ONCE(event->ctx->parent_ctx);
2119         mutex_lock(&event->child_mutex);
2120         func(event);
2121         list_for_each_entry(child, &event->child_list, child_list)
2122                 func(child);
2123         mutex_unlock(&event->child_mutex);
2124 }
2125
2126 static void perf_event_for_each(struct perf_event *event,
2127                                   void (*func)(struct perf_event *))
2128 {
2129         struct perf_event_context *ctx = event->ctx;
2130         struct perf_event *sibling;
2131
2132         WARN_ON_ONCE(ctx->parent_ctx);
2133         mutex_lock(&ctx->mutex);
2134         event = event->group_leader;
2135
2136         perf_event_for_each_child(event, func);
2137         func(event);
2138         list_for_each_entry(sibling, &event->sibling_list, group_entry)
2139                 perf_event_for_each_child(event, func);
2140         mutex_unlock(&ctx->mutex);
2141 }
2142
2143 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2144 {
2145         struct perf_event_context *ctx = event->ctx;
2146         unsigned long size;
2147         int ret = 0;
2148         u64 value;
2149
2150         if (!event->attr.sample_period)
2151                 return -EINVAL;
2152
2153         size = copy_from_user(&value, arg, sizeof(value));
2154         if (size != sizeof(value))
2155                 return -EFAULT;
2156
2157         if (!value)
2158                 return -EINVAL;
2159
2160         raw_spin_lock_irq(&ctx->lock);
2161         if (event->attr.freq) {
2162                 if (value > sysctl_perf_event_sample_rate) {
2163                         ret = -EINVAL;
2164                         goto unlock;
2165                 }
2166
2167                 event->attr.sample_freq = value;
2168         } else {
2169                 event->attr.sample_period = value;
2170                 event->hw.sample_period = value;
2171         }
2172 unlock:
2173         raw_spin_unlock_irq(&ctx->lock);
2174
2175         return ret;
2176 }
2177
2178 static int perf_event_set_output(struct perf_event *event, int output_fd);
2179 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2180
2181 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2182 {
2183         struct perf_event *event = file->private_data;
2184         void (*func)(struct perf_event *);
2185         u32 flags = arg;
2186
2187         switch (cmd) {
2188         case PERF_EVENT_IOC_ENABLE:
2189                 func = perf_event_enable;
2190                 break;
2191         case PERF_EVENT_IOC_DISABLE:
2192                 func = perf_event_disable;
2193                 break;
2194         case PERF_EVENT_IOC_RESET:
2195                 func = perf_event_reset;
2196                 break;
2197
2198         case PERF_EVENT_IOC_REFRESH:
2199                 return perf_event_refresh(event, arg);
2200
2201         case PERF_EVENT_IOC_PERIOD:
2202                 return perf_event_period(event, (u64 __user *)arg);
2203
2204         case PERF_EVENT_IOC_SET_OUTPUT:
2205                 return perf_event_set_output(event, arg);
2206
2207         case PERF_EVENT_IOC_SET_FILTER:
2208                 return perf_event_set_filter(event, (void __user *)arg);
2209
2210         default:
2211                 return -ENOTTY;
2212         }
2213
2214         if (flags & PERF_IOC_FLAG_GROUP)
2215                 perf_event_for_each(event, func);
2216         else
2217                 perf_event_for_each_child(event, func);
2218
2219         return 0;
2220 }
2221
2222 int perf_event_task_enable(void)
2223 {
2224         struct perf_event *event;
2225
2226         mutex_lock(&current->perf_event_mutex);
2227         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2228                 perf_event_for_each_child(event, perf_event_enable);
2229         mutex_unlock(&current->perf_event_mutex);
2230
2231         return 0;
2232 }
2233
2234 int perf_event_task_disable(void)
2235 {
2236         struct perf_event *event;
2237
2238         mutex_lock(&current->perf_event_mutex);
2239         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2240                 perf_event_for_each_child(event, perf_event_disable);
2241         mutex_unlock(&current->perf_event_mutex);
2242
2243         return 0;
2244 }
2245
2246 #ifndef PERF_EVENT_INDEX_OFFSET
2247 # define PERF_EVENT_INDEX_OFFSET 0
2248 #endif
2249
2250 static int perf_event_index(struct perf_event *event)
2251 {
2252         if (event->state != PERF_EVENT_STATE_ACTIVE)
2253                 return 0;
2254
2255         return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2256 }
2257
2258 /*
2259  * Callers need to ensure there can be no nesting of this function, otherwise
2260  * the seqlock logic goes bad. We can not serialize this because the arch
2261  * code calls this from NMI context.
2262  */
2263 void perf_event_update_userpage(struct perf_event *event)
2264 {
2265         struct perf_event_mmap_page *userpg;
2266         struct perf_mmap_data *data;
2267
2268         rcu_read_lock();
2269         data = rcu_dereference(event->data);
2270         if (!data)
2271                 goto unlock;
2272
2273         userpg = data->user_page;
2274
2275         /*
2276          * Disable preemption so as to not let the corresponding user-space
2277          * spin too long if we get preempted.
2278          */
2279         preempt_disable();
2280         ++userpg->lock;
2281         barrier();
2282         userpg->index = perf_event_index(event);
2283         userpg->offset = atomic64_read(&event->count);
2284         if (event->state == PERF_EVENT_STATE_ACTIVE)
2285                 userpg->offset -= atomic64_read(&event->hw.prev_count);
2286
2287         userpg->time_enabled = event->total_time_enabled +
2288                         atomic64_read(&event->child_total_time_enabled);
2289
2290         userpg->time_running = event->total_time_running +
2291                         atomic64_read(&event->child_total_time_running);
2292
2293         barrier();
2294         ++userpg->lock;
2295         preempt_enable();
2296 unlock:
2297         rcu_read_unlock();
2298 }
2299
2300 static unsigned long perf_data_size(struct perf_mmap_data *data)
2301 {
2302         return data->nr_pages << (PAGE_SHIFT + data->data_order);
2303 }
2304
2305 #ifndef CONFIG_PERF_USE_VMALLOC
2306
2307 /*
2308  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2309  */
2310
2311 static struct page *
2312 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2313 {
2314         if (pgoff > data->nr_pages)
2315                 return NULL;
2316
2317         if (pgoff == 0)
2318                 return virt_to_page(data->user_page);
2319
2320         return virt_to_page(data->data_pages[pgoff - 1]);
2321 }
2322
2323 static struct perf_mmap_data *
2324 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2325 {
2326         struct perf_mmap_data *data;
2327         unsigned long size;
2328         int i;
2329
2330         WARN_ON(atomic_read(&event->mmap_count));
2331
2332         size = sizeof(struct perf_mmap_data);
2333         size += nr_pages * sizeof(void *);
2334
2335         data = kzalloc(size, GFP_KERNEL);
2336         if (!data)
2337                 goto fail;
2338
2339         data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2340         if (!data->user_page)
2341                 goto fail_user_page;
2342
2343         for (i = 0; i < nr_pages; i++) {
2344                 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2345                 if (!data->data_pages[i])
2346                         goto fail_data_pages;
2347         }
2348
2349         data->data_order = 0;
2350         data->nr_pages = nr_pages;
2351
2352         return data;
2353
2354 fail_data_pages:
2355         for (i--; i >= 0; i--)
2356                 free_page((unsigned long)data->data_pages[i]);
2357
2358         free_page((unsigned long)data->user_page);
2359
2360 fail_user_page:
2361         kfree(data);
2362
2363 fail:
2364         return NULL;
2365 }
2366
2367 static void perf_mmap_free_page(unsigned long addr)
2368 {
2369         struct page *page = virt_to_page((void *)addr);
2370
2371         page->mapping = NULL;
2372         __free_page(page);
2373 }
2374
2375 static void perf_mmap_data_free(struct perf_mmap_data *data)
2376 {
2377         int i;
2378
2379         perf_mmap_free_page((unsigned long)data->user_page);
2380         for (i = 0; i < data->nr_pages; i++)
2381                 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2382         kfree(data);
2383 }
2384
2385 #else
2386
2387 /*
2388  * Back perf_mmap() with vmalloc memory.
2389  *
2390  * Required for architectures that have d-cache aliasing issues.
2391  */
2392
2393 static struct page *
2394 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2395 {
2396         if (pgoff > (1UL << data->data_order))
2397                 return NULL;
2398
2399         return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2400 }
2401
2402 static void perf_mmap_unmark_page(void *addr)
2403 {
2404         struct page *page = vmalloc_to_page(addr);
2405
2406         page->mapping = NULL;
2407 }
2408
2409 static void perf_mmap_data_free_work(struct work_struct *work)
2410 {
2411         struct perf_mmap_data *data;
2412         void *base;
2413         int i, nr;
2414
2415         data = container_of(work, struct perf_mmap_data, work);
2416         nr = 1 << data->data_order;
2417
2418         base = data->user_page;
2419         for (i = 0; i < nr + 1; i++)
2420                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2421
2422         vfree(base);
2423         kfree(data);
2424 }
2425
2426 static void perf_mmap_data_free(struct perf_mmap_data *data)
2427 {
2428         schedule_work(&data->work);
2429 }
2430
2431 static struct perf_mmap_data *
2432 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2433 {
2434         struct perf_mmap_data *data;
2435         unsigned long size;
2436         void *all_buf;
2437
2438         WARN_ON(atomic_read(&event->mmap_count));
2439
2440         size = sizeof(struct perf_mmap_data);
2441         size += sizeof(void *);
2442
2443         data = kzalloc(size, GFP_KERNEL);
2444         if (!data)
2445                 goto fail;
2446
2447         INIT_WORK(&data->work, perf_mmap_data_free_work);
2448
2449         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2450         if (!all_buf)
2451                 goto fail_all_buf;
2452
2453         data->user_page = all_buf;
2454         data->data_pages[0] = all_buf + PAGE_SIZE;
2455         data->data_order = ilog2(nr_pages);
2456         data->nr_pages = 1;
2457
2458         return data;
2459
2460 fail_all_buf:
2461         kfree(data);
2462
2463 fail:
2464         return NULL;
2465 }
2466
2467 #endif
2468
2469 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2470 {
2471         struct perf_event *event = vma->vm_file->private_data;
2472         struct perf_mmap_data *data;
2473         int ret = VM_FAULT_SIGBUS;
2474
2475         if (vmf->flags & FAULT_FLAG_MKWRITE) {
2476                 if (vmf->pgoff == 0)
2477                         ret = 0;
2478                 return ret;
2479         }
2480
2481         rcu_read_lock();
2482         data = rcu_dereference(event->data);
2483         if (!data)
2484                 goto unlock;
2485
2486         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2487                 goto unlock;
2488
2489         vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2490         if (!vmf->page)
2491                 goto unlock;
2492
2493         get_page(vmf->page);
2494         vmf->page->mapping = vma->vm_file->f_mapping;
2495         vmf->page->index   = vmf->pgoff;
2496
2497         ret = 0;
2498 unlock:
2499         rcu_read_unlock();
2500
2501         return ret;
2502 }
2503
2504 static void
2505 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2506 {
2507         long max_size = perf_data_size(data);
2508
2509         atomic_set(&data->lock, -1);
2510
2511         if (event->attr.watermark) {
2512                 data->watermark = min_t(long, max_size,
2513                                         event->attr.wakeup_watermark);
2514         }
2515
2516         if (!data->watermark)
2517                 data->watermark = max_size / 2;
2518
2519
2520         rcu_assign_pointer(event->data, data);
2521 }
2522
2523 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2524 {
2525         struct perf_mmap_data *data;
2526
2527         data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2528         perf_mmap_data_free(data);
2529 }
2530
2531 static void perf_mmap_data_release(struct perf_event *event)
2532 {
2533         struct perf_mmap_data *data = event->data;
2534
2535         WARN_ON(atomic_read(&event->mmap_count));
2536
2537         rcu_assign_pointer(event->data, NULL);
2538         call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2539 }
2540
2541 static void perf_mmap_open(struct vm_area_struct *vma)
2542 {
2543         struct perf_event *event = vma->vm_file->private_data;
2544
2545         atomic_inc(&event->mmap_count);
2546 }
2547
2548 static void perf_mmap_close(struct vm_area_struct *vma)
2549 {
2550         struct perf_event *event = vma->vm_file->private_data;
2551
2552         WARN_ON_ONCE(event->ctx->parent_ctx);
2553         if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2554                 unsigned long size = perf_data_size(event->data);
2555                 struct user_struct *user = current_user();
2556
2557                 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2558                 vma->vm_mm->locked_vm -= event->data->nr_locked;
2559                 perf_mmap_data_release(event);
2560                 mutex_unlock(&event->mmap_mutex);
2561         }
2562 }
2563
2564 static const struct vm_operations_struct perf_mmap_vmops = {
2565         .open           = perf_mmap_open,
2566         .close          = perf_mmap_close,
2567         .fault          = perf_mmap_fault,
2568         .page_mkwrite   = perf_mmap_fault,
2569 };
2570
2571 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2572 {
2573         struct perf_event *event = file->private_data;
2574         unsigned long user_locked, user_lock_limit;
2575         struct user_struct *user = current_user();
2576         unsigned long locked, lock_limit;
2577         struct perf_mmap_data *data;
2578         unsigned long vma_size;
2579         unsigned long nr_pages;
2580         long user_extra, extra;
2581         int ret = 0;
2582
2583         if (!(vma->vm_flags & VM_SHARED))
2584                 return -EINVAL;
2585
2586         vma_size = vma->vm_end - vma->vm_start;
2587         nr_pages = (vma_size / PAGE_SIZE) - 1;
2588
2589         /*
2590          * If we have data pages ensure they're a power-of-two number, so we
2591          * can do bitmasks instead of modulo.
2592          */
2593         if (nr_pages != 0 && !is_power_of_2(nr_pages))
2594                 return -EINVAL;
2595
2596         if (vma_size != PAGE_SIZE * (1 + nr_pages))
2597                 return -EINVAL;
2598
2599         if (vma->vm_pgoff != 0)
2600                 return -EINVAL;
2601
2602         WARN_ON_ONCE(event->ctx->parent_ctx);
2603         mutex_lock(&event->mmap_mutex);
2604         if (event->output) {
2605                 ret = -EINVAL;
2606                 goto unlock;
2607         }
2608
2609         if (atomic_inc_not_zero(&event->mmap_count)) {
2610                 if (nr_pages != event->data->nr_pages)
2611                         ret = -EINVAL;
2612                 goto unlock;
2613         }
2614
2615         user_extra = nr_pages + 1;
2616         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2617
2618         /*
2619          * Increase the limit linearly with more CPUs:
2620          */
2621         user_lock_limit *= num_online_cpus();
2622
2623         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2624
2625         extra = 0;
2626         if (user_locked > user_lock_limit)
2627                 extra = user_locked - user_lock_limit;
2628
2629         lock_limit = rlimit(RLIMIT_MEMLOCK);
2630         lock_limit >>= PAGE_SHIFT;
2631         locked = vma->vm_mm->locked_vm + extra;
2632
2633         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2634                 !capable(CAP_IPC_LOCK)) {
2635                 ret = -EPERM;
2636                 goto unlock;
2637         }
2638
2639         WARN_ON(event->data);
2640
2641         data = perf_mmap_data_alloc(event, nr_pages);
2642         ret = -ENOMEM;
2643         if (!data)
2644                 goto unlock;
2645
2646         ret = 0;
2647         perf_mmap_data_init(event, data);
2648
2649         atomic_set(&event->mmap_count, 1);
2650         atomic_long_add(user_extra, &user->locked_vm);
2651         vma->vm_mm->locked_vm += extra;
2652         event->data->nr_locked = extra;
2653         if (vma->vm_flags & VM_WRITE)
2654                 event->data->writable = 1;
2655
2656 unlock:
2657         mutex_unlock(&event->mmap_mutex);
2658
2659         vma->vm_flags |= VM_RESERVED;
2660         vma->vm_ops = &perf_mmap_vmops;
2661
2662         return ret;
2663 }
2664
2665 static int perf_fasync(int fd, struct file *filp, int on)
2666 {
2667         struct inode *inode = filp->f_path.dentry->d_inode;
2668         struct perf_event *event = filp->private_data;
2669         int retval;
2670
2671         mutex_lock(&inode->i_mutex);
2672         retval = fasync_helper(fd, filp, on, &event->fasync);
2673         mutex_unlock(&inode->i_mutex);
2674
2675         if (retval < 0)
2676                 return retval;
2677
2678         return 0;
2679 }
2680
2681 static const struct file_operations perf_fops = {
2682         .llseek                 = no_llseek,
2683         .release                = perf_release,
2684         .read                   = perf_read,
2685         .poll                   = perf_poll,
2686         .unlocked_ioctl         = perf_ioctl,
2687         .compat_ioctl           = perf_ioctl,
2688         .mmap                   = perf_mmap,
2689         .fasync                 = perf_fasync,
2690 };
2691
2692 /*
2693  * Perf event wakeup
2694  *
2695  * If there's data, ensure we set the poll() state and publish everything
2696  * to user-space before waking everybody up.
2697  */
2698
2699 void perf_event_wakeup(struct perf_event *event)
2700 {
2701         wake_up_all(&event->waitq);
2702
2703         if (event->pending_kill) {
2704                 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2705                 event->pending_kill = 0;
2706         }
2707 }
2708
2709 /*
2710  * Pending wakeups
2711  *
2712  * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2713  *
2714  * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2715  * single linked list and use cmpxchg() to add entries lockless.
2716  */
2717
2718 static void perf_pending_event(struct perf_pending_entry *entry)
2719 {
2720         struct perf_event *event = container_of(entry,
2721                         struct perf_event, pending);
2722
2723         if (event->pending_disable) {
2724                 event->pending_disable = 0;
2725                 __perf_event_disable(event);
2726         }
2727
2728         if (event->pending_wakeup) {
2729                 event->pending_wakeup = 0;
2730                 perf_event_wakeup(event);
2731         }
2732 }
2733
2734 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2735
2736 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2737         PENDING_TAIL,
2738 };
2739
2740 static void perf_pending_queue(struct perf_pending_entry *entry,
2741                                void (*func)(struct perf_pending_entry *))
2742 {
2743         struct perf_pending_entry **head;
2744
2745         if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2746                 return;
2747
2748         entry->func = func;
2749
2750         head = &get_cpu_var(perf_pending_head);
2751
2752         do {
2753                 entry->next = *head;
2754         } while (cmpxchg(head, entry->next, entry) != entry->next);
2755
2756         set_perf_event_pending();
2757
2758         put_cpu_var(perf_pending_head);
2759 }
2760
2761 static int __perf_pending_run(void)
2762 {
2763         struct perf_pending_entry *list;
2764         int nr = 0;
2765
2766         list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2767         while (list != PENDING_TAIL) {
2768                 void (*func)(struct perf_pending_entry *);
2769                 struct perf_pending_entry *entry = list;
2770
2771                 list = list->next;
2772
2773                 func = entry->func;
2774                 entry->next = NULL;
2775                 /*
2776                  * Ensure we observe the unqueue before we issue the wakeup,
2777                  * so that we won't be waiting forever.
2778                  * -- see perf_not_pending().
2779                  */
2780                 smp_wmb();
2781
2782                 func(entry);
2783                 nr++;
2784         }
2785
2786         return nr;
2787 }
2788
2789 static inline int perf_not_pending(struct perf_event *event)
2790 {
2791         /*
2792          * If we flush on whatever cpu we run, there is a chance we don't
2793          * need to wait.
2794          */
2795         get_cpu();
2796         __perf_pending_run();
2797         put_cpu();
2798
2799         /*
2800          * Ensure we see the proper queue state before going to sleep
2801          * so that we do not miss the wakeup. -- see perf_pending_handle()
2802          */
2803         smp_rmb();
2804         return event->pending.next == NULL;
2805 }
2806
2807 static void perf_pending_sync(struct perf_event *event)
2808 {
2809         wait_event(event->waitq, perf_not_pending(event));
2810 }
2811
2812 void perf_event_do_pending(void)
2813 {
2814         __perf_pending_run();
2815 }
2816
2817 /*
2818  * Callchain support -- arch specific
2819  */
2820
2821 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2822 {
2823         return NULL;
2824 }
2825
2826 __weak
2827 void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip, int skip)
2828 {
2829 }
2830
2831
2832 /*
2833  * We assume there is only KVM supporting the callbacks.
2834  * Later on, we might change it to a list if there is
2835  * another virtualization implementation supporting the callbacks.
2836  */
2837 struct perf_guest_info_callbacks *perf_guest_cbs;
2838
2839 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2840 {
2841         perf_guest_cbs = cbs;
2842         return 0;
2843 }
2844 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
2845
2846 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2847 {
2848         perf_guest_cbs = NULL;
2849         return 0;
2850 }
2851 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
2852
2853 /*
2854  * Output
2855  */
2856 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2857                               unsigned long offset, unsigned long head)
2858 {
2859         unsigned long mask;
2860
2861         if (!data->writable)
2862                 return true;
2863
2864         mask = perf_data_size(data) - 1;
2865
2866         offset = (offset - tail) & mask;
2867         head   = (head   - tail) & mask;
2868
2869         if ((int)(head - offset) < 0)
2870                 return false;
2871
2872         return true;
2873 }
2874
2875 static void perf_output_wakeup(struct perf_output_handle *handle)
2876 {
2877         atomic_set(&handle->data->poll, POLL_IN);
2878
2879         if (handle->nmi) {
2880                 handle->event->pending_wakeup = 1;
2881                 perf_pending_queue(&handle->event->pending,
2882                                    perf_pending_event);
2883         } else
2884                 perf_event_wakeup(handle->event);
2885 }
2886
2887 /*
2888  * Curious locking construct.
2889  *
2890  * We need to ensure a later event_id doesn't publish a head when a former
2891  * event_id isn't done writing. However since we need to deal with NMIs we
2892  * cannot fully serialize things.
2893  *
2894  * What we do is serialize between CPUs so we only have to deal with NMI
2895  * nesting on a single CPU.
2896  *
2897  * We only publish the head (and generate a wakeup) when the outer-most
2898  * event_id completes.
2899  */
2900 static void perf_output_lock(struct perf_output_handle *handle)
2901 {
2902         struct perf_mmap_data *data = handle->data;
2903         int cur, cpu = get_cpu();
2904
2905         handle->locked = 0;
2906
2907         for (;;) {
2908                 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2909                 if (cur == -1) {
2910                         handle->locked = 1;
2911                         break;
2912                 }
2913                 if (cur == cpu)
2914                         break;
2915
2916                 cpu_relax();
2917         }
2918 }
2919
2920 static void perf_output_unlock(struct perf_output_handle *handle)
2921 {
2922         struct perf_mmap_data *data = handle->data;
2923         unsigned long head;
2924         int cpu;
2925
2926         data->done_head = data->head;
2927
2928         if (!handle->locked)
2929                 goto out;
2930
2931 again:
2932         /*
2933          * The xchg implies a full barrier that ensures all writes are done
2934          * before we publish the new head, matched by a rmb() in userspace when
2935          * reading this position.
2936          */
2937         while ((head = atomic_long_xchg(&data->done_head, 0)))
2938                 data->user_page->data_head = head;
2939
2940         /*
2941          * NMI can happen here, which means we can miss a done_head update.
2942          */
2943
2944         cpu = atomic_xchg(&data->lock, -1);
2945         WARN_ON_ONCE(cpu != smp_processor_id());
2946
2947         /*
2948          * Therefore we have to validate we did not indeed do so.
2949          */
2950         if (unlikely(atomic_long_read(&data->done_head))) {
2951                 /*
2952                  * Since we had it locked, we can lock it again.
2953                  */
2954                 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2955                         cpu_relax();
2956
2957                 goto again;
2958         }
2959
2960         if (atomic_xchg(&data->wakeup, 0))
2961                 perf_output_wakeup(handle);
2962 out:
2963         put_cpu();
2964 }
2965
2966 void perf_output_copy(struct perf_output_handle *handle,
2967                       const void *buf, unsigned int len)
2968 {
2969         unsigned int pages_mask;
2970         unsigned long offset;
2971         unsigned int size;
2972         void **pages;
2973
2974         offset          = handle->offset;
2975         pages_mask      = handle->data->nr_pages - 1;
2976         pages           = handle->data->data_pages;
2977
2978         do {
2979                 unsigned long page_offset;
2980                 unsigned long page_size;
2981                 int nr;
2982
2983                 nr          = (offset >> PAGE_SHIFT) & pages_mask;
2984                 page_size   = 1UL << (handle->data->data_order + PAGE_SHIFT);
2985                 page_offset = offset & (page_size - 1);
2986                 size        = min_t(unsigned int, page_size - page_offset, len);
2987
2988                 memcpy(pages[nr] + page_offset, buf, size);
2989
2990                 len         -= size;
2991                 buf         += size;
2992                 offset      += size;
2993         } while (len);
2994
2995         handle->offset = offset;
2996
2997         /*
2998          * Check we didn't copy past our reservation window, taking the
2999          * possible unsigned int wrap into account.
3000          */
3001         WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
3002 }
3003
3004 int perf_output_begin(struct perf_output_handle *handle,
3005                       struct perf_event *event, unsigned int size,
3006                       int nmi, int sample)
3007 {
3008         struct perf_event *output_event;
3009         struct perf_mmap_data *data;
3010         unsigned long tail, offset, head;
3011         int have_lost;
3012         struct {
3013                 struct perf_event_header header;
3014                 u64                      id;
3015                 u64                      lost;
3016         } lost_event;
3017
3018         rcu_read_lock();
3019         /*
3020          * For inherited events we send all the output towards the parent.
3021          */
3022         if (event->parent)
3023                 event = event->parent;
3024
3025         output_event = rcu_dereference(event->output);
3026         if (output_event)
3027                 event = output_event;
3028
3029         data = rcu_dereference(event->data);
3030         if (!data)
3031                 goto out;
3032
3033         handle->data    = data;
3034         handle->event   = event;
3035         handle->nmi     = nmi;
3036         handle->sample  = sample;
3037
3038         if (!data->nr_pages)
3039                 goto fail;
3040
3041         have_lost = atomic_read(&data->lost);
3042         if (have_lost)
3043                 size += sizeof(lost_event);
3044
3045         perf_output_lock(handle);
3046
3047         do {
3048                 /*
3049                  * Userspace could choose to issue a mb() before updating the
3050                  * tail pointer. So that all reads will be completed before the
3051                  * write is issued.
3052                  */
3053                 tail = ACCESS_ONCE(data->user_page->data_tail);
3054                 smp_rmb();
3055                 offset = head = atomic_long_read(&data->head);
3056                 head += size;
3057                 if (unlikely(!perf_output_space(data, tail, offset, head)))
3058                         goto fail;
3059         } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
3060
3061         handle->offset  = offset;
3062         handle->head    = head;
3063
3064         if (head - tail > data->watermark)
3065                 atomic_set(&data->wakeup, 1);
3066
3067         if (have_lost) {
3068                 lost_event.header.type = PERF_RECORD_LOST;
3069                 lost_event.header.misc = 0;
3070                 lost_event.header.size = sizeof(lost_event);
3071                 lost_event.id          = event->id;
3072                 lost_event.lost        = atomic_xchg(&data->lost, 0);
3073
3074                 perf_output_put(handle, lost_event);
3075         }
3076
3077         return 0;
3078
3079 fail:
3080         atomic_inc(&data->lost);
3081         perf_output_unlock(handle);
3082 out:
3083         rcu_read_unlock();
3084
3085         return -ENOSPC;
3086 }
3087
3088 void perf_output_end(struct perf_output_handle *handle)
3089 {
3090         struct perf_event *event = handle->event;
3091         struct perf_mmap_data *data = handle->data;
3092
3093         int wakeup_events = event->attr.wakeup_events;
3094
3095         if (handle->sample && wakeup_events) {
3096                 int events = atomic_inc_return(&data->events);
3097                 if (events >= wakeup_events) {
3098                         atomic_sub(wakeup_events, &data->events);
3099                         atomic_set(&data->wakeup, 1);
3100                 }
3101         }
3102
3103         perf_output_unlock(handle);
3104         rcu_read_unlock();
3105 }
3106
3107 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3108 {
3109         /*
3110          * only top level events have the pid namespace they were created in
3111          */
3112         if (event->parent)
3113                 event = event->parent;
3114
3115         return task_tgid_nr_ns(p, event->ns);
3116 }
3117
3118 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3119 {
3120         /*
3121          * only top level events have the pid namespace they were created in
3122          */
3123         if (event->parent)
3124                 event = event->parent;
3125
3126         return task_pid_nr_ns(p, event->ns);
3127 }
3128
3129 static void perf_output_read_one(struct perf_output_handle *handle,
3130                                  struct perf_event *event)
3131 {
3132         u64 read_format = event->attr.read_format;
3133         u64 values[4];
3134         int n = 0;
3135
3136         values[n++] = atomic64_read(&event->count);
3137         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3138                 values[n++] = event->total_time_enabled +
3139                         atomic64_read(&event->child_total_time_enabled);
3140         }
3141         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3142                 values[n++] = event->total_time_running +
3143                         atomic64_read(&event->child_total_time_running);
3144         }
3145         if (read_format & PERF_FORMAT_ID)
3146                 values[n++] = primary_event_id(event);
3147
3148         perf_output_copy(handle, values, n * sizeof(u64));
3149 }
3150
3151 /*
3152  * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3153  */
3154 static void perf_output_read_group(struct perf_output_handle *handle,
3155                             struct perf_event *event)
3156 {
3157         struct perf_event *leader = event->group_leader, *sub;
3158         u64 read_format = event->attr.read_format;
3159         u64 values[5];
3160         int n = 0;
3161
3162         values[n++] = 1 + leader->nr_siblings;
3163
3164         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3165                 values[n++] = leader->total_time_enabled;
3166
3167         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3168                 values[n++] = leader->total_time_running;
3169
3170         if (leader != event)
3171                 leader->pmu->read(leader);
3172
3173         values[n++] = atomic64_read(&leader->count);
3174         if (read_format & PERF_FORMAT_ID)
3175                 values[n++] = primary_event_id(leader);
3176
3177         perf_output_copy(handle, values, n * sizeof(u64));
3178
3179         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3180                 n = 0;
3181
3182                 if (sub != event)
3183                         sub->pmu->read(sub);
3184
3185                 values[n++] = atomic64_read(&sub->count);
3186                 if (read_format & PERF_FORMAT_ID)
3187                         values[n++] = primary_event_id(sub);
3188
3189                 perf_output_copy(handle, values, n * sizeof(u64));
3190         }
3191 }
3192
3193 static void perf_output_read(struct perf_output_handle *handle,
3194                              struct perf_event *event)
3195 {
3196         if (event->attr.read_format & PERF_FORMAT_GROUP)
3197                 perf_output_read_group(handle, event);
3198         else
3199                 perf_output_read_one(handle, event);
3200 }
3201
3202 void perf_output_sample(struct perf_output_handle *handle,
3203                         struct perf_event_header *header,
3204                         struct perf_sample_data *data,
3205                         struct perf_event *event)
3206 {
3207         u64 sample_type = data->type;
3208
3209         perf_output_put(handle, *header);
3210
3211         if (sample_type & PERF_SAMPLE_IP)
3212                 perf_output_put(handle, data->ip);
3213
3214         if (sample_type & PERF_SAMPLE_TID)
3215                 perf_output_put(handle, data->tid_entry);
3216
3217         if (sample_type & PERF_SAMPLE_TIME)
3218                 perf_output_put(handle, data->time);
3219
3220         if (sample_type & PERF_SAMPLE_ADDR)
3221                 perf_output_put(handle, data->addr);
3222
3223         if (sample_type & PERF_SAMPLE_ID)
3224                 perf_output_put(handle, data->id);
3225
3226         if (sample_type & PERF_SAMPLE_STREAM_ID)
3227                 perf_output_put(handle, data->stream_id);
3228
3229         if (sample_type & PERF_SAMPLE_CPU)
3230                 perf_output_put(handle, data->cpu_entry);
3231
3232         if (sample_type & PERF_SAMPLE_PERIOD)
3233                 perf_output_put(handle, data->period);
3234
3235         if (sample_type & PERF_SAMPLE_READ)
3236                 perf_output_read(handle, event);
3237
3238         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3239                 if (data->callchain) {
3240                         int size = 1;
3241
3242                         if (data->callchain)
3243                                 size += data->callchain->nr;
3244
3245                         size *= sizeof(u64);
3246
3247                         perf_output_copy(handle, data->callchain, size);
3248                 } else {
3249                         u64 nr = 0;
3250                         perf_output_put(handle, nr);
3251                 }
3252         }
3253
3254         if (sample_type & PERF_SAMPLE_RAW) {
3255                 if (data->raw) {
3256                         perf_output_put(handle, data->raw->size);
3257                         perf_output_copy(handle, data->raw->data,
3258                                          data->raw->size);
3259                 } else {
3260                         struct {
3261                                 u32     size;
3262                                 u32     data;
3263                         } raw = {
3264                                 .size = sizeof(u32),
3265                                 .data = 0,
3266                         };
3267                         perf_output_put(handle, raw);
3268                 }
3269         }
3270 }
3271
3272 void perf_prepare_sample(struct perf_event_header *header,
3273                          struct perf_sample_data *data,
3274                          struct perf_event *event,
3275                          struct pt_regs *regs)
3276 {
3277         u64 sample_type = event->attr.sample_type;
3278
3279         data->type = sample_type;
3280
3281         header->type = PERF_RECORD_SAMPLE;
3282         header->size = sizeof(*header);
3283
3284         header->misc = 0;
3285         header->misc |= perf_misc_flags(regs);
3286
3287         if (sample_type & PERF_SAMPLE_IP) {
3288                 data->ip = perf_instruction_pointer(regs);
3289
3290                 header->size += sizeof(data->ip);
3291         }
3292
3293         if (sample_type & PERF_SAMPLE_TID) {
3294                 /* namespace issues */
3295                 data->tid_entry.pid = perf_event_pid(event, current);
3296                 data->tid_entry.tid = perf_event_tid(event, current);
3297
3298                 header->size += sizeof(data->tid_entry);
3299         }
3300
3301         if (sample_type & PERF_SAMPLE_TIME) {
3302                 data->time = perf_clock();
3303
3304                 header->size += sizeof(data->time);
3305         }
3306
3307         if (sample_type & PERF_SAMPLE_ADDR)
3308                 header->size += sizeof(data->addr);
3309
3310         if (sample_type & PERF_SAMPLE_ID) {
3311                 data->id = primary_event_id(event);
3312
3313                 header->size += sizeof(data->id);
3314         }
3315
3316         if (sample_type & PERF_SAMPLE_STREAM_ID) {
3317                 data->stream_id = event->id;
3318
3319                 header->size += sizeof(data->stream_id);
3320         }
3321
3322         if (sample_type & PERF_SAMPLE_CPU) {
3323                 data->cpu_entry.cpu             = raw_smp_processor_id();
3324                 data->cpu_entry.reserved        = 0;
3325
3326                 header->size += sizeof(data->cpu_entry);
3327         }
3328
3329         if (sample_type & PERF_SAMPLE_PERIOD)
3330                 header->size += sizeof(data->period);
3331
3332         if (sample_type & PERF_SAMPLE_READ)
3333                 header->size += perf_event_read_size(event);
3334
3335         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3336                 int size = 1;
3337
3338                 data->callchain = perf_callchain(regs);
3339
3340                 if (data->callchain)
3341                         size += data->callchain->nr;
3342
3343                 header->size += size * sizeof(u64);
3344         }
3345
3346         if (sample_type & PERF_SAMPLE_RAW) {
3347                 int size = sizeof(u32);
3348
3349                 if (data->raw)
3350                         size += data->raw->size;
3351                 else
3352                         size += sizeof(u32);
3353
3354                 WARN_ON_ONCE(size & (sizeof(u64)-1));
3355                 header->size += size;
3356         }
3357 }
3358
3359 static void perf_event_output(struct perf_event *event, int nmi,
3360                                 struct perf_sample_data *data,
3361                                 struct pt_regs *regs)
3362 {
3363         struct perf_output_handle handle;
3364         struct perf_event_header header;
3365
3366         perf_prepare_sample(&header, data, event, regs);
3367
3368         if (perf_output_begin(&handle, event, header.size, nmi, 1))
3369                 return;
3370
3371         perf_output_sample(&handle, &header, data, event);
3372
3373         perf_output_end(&handle);
3374 }
3375
3376 /*
3377  * read event_id
3378  */
3379
3380 struct perf_read_event {
3381         struct perf_event_header        header;
3382
3383         u32                             pid;
3384         u32                             tid;
3385 };
3386
3387 static void
3388 perf_event_read_event(struct perf_event *event,
3389                         struct task_struct *task)
3390 {
3391         struct perf_output_handle handle;
3392         struct perf_read_event read_event = {
3393                 .header = {
3394                         .type = PERF_RECORD_READ,
3395                         .misc = 0,
3396                         .size = sizeof(read_event) + perf_event_read_size(event),
3397                 },
3398                 .pid = perf_event_pid(event, task),
3399                 .tid = perf_event_tid(event, task),
3400         };
3401         int ret;
3402
3403         ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3404         if (ret)
3405                 return;
3406
3407         perf_output_put(&handle, read_event);
3408         perf_output_read(&handle, event);
3409
3410         perf_output_end(&handle);
3411 }
3412
3413 /*
3414  * task tracking -- fork/exit
3415  *
3416  * enabled by: attr.comm | attr.mmap | attr.task
3417  */
3418
3419 struct perf_task_event {
3420         struct task_struct              *task;
3421         struct perf_event_context       *task_ctx;
3422
3423         struct {
3424                 struct perf_event_header        header;
3425
3426                 u32                             pid;
3427                 u32                             ppid;
3428                 u32                             tid;
3429                 u32                             ptid;
3430                 u64                             time;
3431         } event_id;
3432 };
3433
3434 static void perf_event_task_output(struct perf_event *event,
3435                                      struct perf_task_event *task_event)
3436 {
3437         struct perf_output_handle handle;
3438         struct task_struct *task = task_event->task;
3439         unsigned long flags;
3440         int size, ret;
3441
3442         /*
3443          * If this CPU attempts to acquire an rq lock held by a CPU spinning
3444          * in perf_output_lock() from interrupt context, it's game over.
3445          */
3446         local_irq_save(flags);
3447
3448         size  = task_event->event_id.header.size;
3449         ret = perf_output_begin(&handle, event, size, 0, 0);
3450
3451         if (ret) {
3452                 local_irq_restore(flags);
3453                 return;
3454         }
3455
3456         task_event->event_id.pid = perf_event_pid(event, task);
3457         task_event->event_id.ppid = perf_event_pid(event, current);
3458
3459         task_event->event_id.tid = perf_event_tid(event, task);
3460         task_event->event_id.ptid = perf_event_tid(event, current);
3461
3462         perf_output_put(&handle, task_event->event_id);
3463
3464         perf_output_end(&handle);
3465         local_irq_restore(flags);
3466 }
3467
3468 static int perf_event_task_match(struct perf_event *event)
3469 {
3470         if (event->state < PERF_EVENT_STATE_INACTIVE)
3471                 return 0;
3472
3473         if (event->cpu != -1 && event->cpu != smp_processor_id())
3474                 return 0;
3475
3476         if (event->attr.comm || event->attr.mmap || event->attr.task)
3477                 return 1;
3478
3479         return 0;
3480 }
3481
3482 static void perf_event_task_ctx(struct perf_event_context *ctx,
3483                                   struct perf_task_event *task_event)
3484 {
3485         struct perf_event *event;
3486
3487         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3488                 if (perf_event_task_match(event))
3489                         perf_event_task_output(event, task_event);
3490         }
3491 }
3492
3493 static void perf_event_task_event(struct perf_task_event *task_event)
3494 {
3495         struct perf_cpu_context *cpuctx;
3496         struct perf_event_context *ctx = task_event->task_ctx;
3497
3498         rcu_read_lock();
3499         cpuctx = &get_cpu_var(perf_cpu_context);
3500         perf_event_task_ctx(&cpuctx->ctx, task_event);
3501         if (!ctx)
3502                 ctx = rcu_dereference(current->perf_event_ctxp);
3503         if (ctx)
3504                 perf_event_task_ctx(ctx, task_event);
3505         put_cpu_var(perf_cpu_context);
3506         rcu_read_unlock();
3507 }
3508
3509 static void perf_event_task(struct task_struct *task,
3510                               struct perf_event_context *task_ctx,
3511                               int new)
3512 {
3513         struct perf_task_event task_event;
3514
3515         if (!atomic_read(&nr_comm_events) &&
3516             !atomic_read(&nr_mmap_events) &&
3517             !atomic_read(&nr_task_events))
3518                 return;
3519
3520         task_event = (struct perf_task_event){
3521                 .task     = task,
3522                 .task_ctx = task_ctx,
3523                 .event_id    = {
3524                         .header = {
3525                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3526                                 .misc = 0,
3527                                 .size = sizeof(task_event.event_id),
3528                         },
3529                         /* .pid  */
3530                         /* .ppid */
3531                         /* .tid  */
3532                         /* .ptid */
3533                         .time = perf_clock(),
3534                 },
3535         };
3536
3537         perf_event_task_event(&task_event);
3538 }
3539
3540 void perf_event_fork(struct task_struct *task)
3541 {
3542         perf_event_task(task, NULL, 1);
3543 }
3544
3545 /*
3546  * comm tracking
3547  */
3548
3549 struct perf_comm_event {
3550         struct task_struct      *task;
3551         char                    *comm;
3552         int                     comm_size;
3553
3554         struct {
3555                 struct perf_event_header        header;
3556
3557                 u32                             pid;
3558                 u32                             tid;
3559         } event_id;
3560 };
3561
3562 static void perf_event_comm_output(struct perf_event *event,
3563                                      struct perf_comm_event *comm_event)
3564 {
3565         struct perf_output_handle handle;
3566         int size = comm_event->event_id.header.size;
3567         int ret = perf_output_begin(&handle, event, size, 0, 0);
3568
3569         if (ret)
3570                 return;
3571
3572         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3573         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3574
3575         perf_output_put(&handle, comm_event->event_id);
3576         perf_output_copy(&handle, comm_event->comm,
3577                                    comm_event->comm_size);
3578         perf_output_end(&handle);
3579 }
3580
3581 static int perf_event_comm_match(struct perf_event *event)
3582 {
3583         if (event->state < PERF_EVENT_STATE_INACTIVE)
3584                 return 0;
3585
3586         if (event->cpu != -1 && event->cpu != smp_processor_id())
3587                 return 0;
3588
3589         if (event->attr.comm)
3590                 return 1;
3591
3592         return 0;
3593 }
3594
3595 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3596                                   struct perf_comm_event *comm_event)
3597 {
3598         struct perf_event *event;
3599
3600         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3601                 if (perf_event_comm_match(event))
3602                         perf_event_comm_output(event, comm_event);
3603         }
3604 }
3605
3606 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3607 {
3608         struct perf_cpu_context *cpuctx;
3609         struct perf_event_context *ctx;
3610         unsigned int size;
3611         char comm[TASK_COMM_LEN];
3612
3613         memset(comm, 0, sizeof(comm));
3614         strlcpy(comm, comm_event->task->comm, sizeof(comm));
3615         size = ALIGN(strlen(comm)+1, sizeof(u64));
3616
3617         comm_event->comm = comm;
3618         comm_event->comm_size = size;
3619
3620         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3621
3622         rcu_read_lock();
3623         cpuctx = &get_cpu_var(perf_cpu_context);
3624         perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3625         ctx = rcu_dereference(current->perf_event_ctxp);
3626         if (ctx)
3627                 perf_event_comm_ctx(ctx, comm_event);
3628         put_cpu_var(perf_cpu_context);
3629         rcu_read_unlock();
3630 }
3631
3632 void perf_event_comm(struct task_struct *task)
3633 {
3634         struct perf_comm_event comm_event;
3635
3636         if (task->perf_event_ctxp)
3637                 perf_event_enable_on_exec(task);
3638
3639         if (!atomic_read(&nr_comm_events))
3640                 return;
3641
3642         comm_event = (struct perf_comm_event){
3643                 .task   = task,
3644                 /* .comm      */
3645                 /* .comm_size */
3646                 .event_id  = {
3647                         .header = {
3648                                 .type = PERF_RECORD_COMM,
3649                                 .misc = 0,
3650                                 /* .size */
3651                         },
3652                         /* .pid */
3653                         /* .tid */
3654                 },
3655         };
3656
3657         perf_event_comm_event(&comm_event);
3658 }
3659
3660 /*
3661  * mmap tracking
3662  */
3663
3664 struct perf_mmap_event {
3665         struct vm_area_struct   *vma;
3666
3667         const char              *file_name;
3668         int                     file_size;
3669
3670         struct {
3671                 struct perf_event_header        header;
3672
3673                 u32                             pid;
3674                 u32                             tid;
3675                 u64                             start;
3676                 u64                             len;
3677                 u64                             pgoff;
3678         } event_id;
3679 };
3680
3681 static void perf_event_mmap_output(struct perf_event *event,
3682                                      struct perf_mmap_event *mmap_event)
3683 {
3684         struct perf_output_handle handle;
3685         int size = mmap_event->event_id.header.size;
3686         int ret = perf_output_begin(&handle, event, size, 0, 0);
3687
3688         if (ret)
3689                 return;
3690
3691         mmap_event->event_id.pid = perf_event_pid(event, current);
3692         mmap_event->event_id.tid = perf_event_tid(event, current);
3693
3694         perf_output_put(&handle, mmap_event->event_id);
3695         perf_output_copy(&handle, mmap_event->file_name,
3696                                    mmap_event->file_size);
3697         perf_output_end(&handle);
3698 }
3699
3700 static int perf_event_mmap_match(struct perf_event *event,
3701                                    struct perf_mmap_event *mmap_event)
3702 {
3703         if (event->state < PERF_EVENT_STATE_INACTIVE)
3704                 return 0;
3705
3706         if (event->cpu != -1 && event->cpu != smp_processor_id())
3707                 return 0;
3708
3709         if (event->attr.mmap)
3710                 return 1;
3711
3712         return 0;
3713 }
3714
3715 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3716                                   struct perf_mmap_event *mmap_event)
3717 {
3718         struct perf_event *event;
3719
3720         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3721                 if (perf_event_mmap_match(event, mmap_event))
3722                         perf_event_mmap_output(event, mmap_event);
3723         }
3724 }
3725
3726 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3727 {
3728         struct perf_cpu_context *cpuctx;
3729         struct perf_event_context *ctx;
3730         struct vm_area_struct *vma = mmap_event->vma;
3731         struct file *file = vma->vm_file;
3732         unsigned int size;
3733         char tmp[16];
3734         char *buf = NULL;
3735         const char *name;
3736
3737         memset(tmp, 0, sizeof(tmp));
3738
3739         if (file) {
3740                 /*
3741                  * d_path works from the end of the buffer backwards, so we
3742                  * need to add enough zero bytes after the string to handle
3743                  * the 64bit alignment we do later.
3744                  */
3745                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3746                 if (!buf) {
3747                         name = strncpy(tmp, "//enomem", sizeof(tmp));
3748                         goto got_name;
3749                 }
3750                 name = d_path(&file->f_path, buf, PATH_MAX);
3751                 if (IS_ERR(name)) {
3752                         name = strncpy(tmp, "//toolong", sizeof(tmp));
3753                         goto got_name;
3754                 }
3755         } else {
3756                 if (arch_vma_name(mmap_event->vma)) {
3757                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3758                                        sizeof(tmp));
3759                         goto got_name;
3760                 }
3761
3762                 if (!vma->vm_mm) {
3763                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
3764                         goto got_name;
3765                 }
3766
3767                 name = strncpy(tmp, "//anon", sizeof(tmp));
3768                 goto got_name;
3769         }
3770
3771 got_name:
3772         size = ALIGN(strlen(name)+1, sizeof(u64));
3773
3774         mmap_event->file_name = name;
3775         mmap_event->file_size = size;
3776
3777         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3778
3779         rcu_read_lock();
3780         cpuctx = &get_cpu_var(perf_cpu_context);
3781         perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3782         ctx = rcu_dereference(current->perf_event_ctxp);
3783         if (ctx)
3784                 perf_event_mmap_ctx(ctx, mmap_event);
3785         put_cpu_var(perf_cpu_context);
3786         rcu_read_unlock();
3787
3788         kfree(buf);
3789 }
3790
3791 void __perf_event_mmap(struct vm_area_struct *vma)
3792 {
3793         struct perf_mmap_event mmap_event;
3794
3795         if (!atomic_read(&nr_mmap_events))
3796                 return;
3797
3798         mmap_event = (struct perf_mmap_event){
3799                 .vma    = vma,
3800                 /* .file_name */
3801                 /* .file_size */
3802                 .event_id  = {
3803                         .header = {
3804                                 .type = PERF_RECORD_MMAP,
3805                                 .misc = PERF_RECORD_MISC_USER,
3806                                 /* .size */
3807                         },
3808                         /* .pid */
3809                         /* .tid */
3810                         .start  = vma->vm_start,
3811                         .len    = vma->vm_end - vma->vm_start,
3812                         .pgoff  = (u64)vma->vm_pgoff << PAGE_SHIFT,
3813                 },
3814         };
3815
3816         perf_event_mmap_event(&mmap_event);
3817 }
3818
3819 /*
3820  * IRQ throttle logging
3821  */
3822
3823 static void perf_log_throttle(struct perf_event *event, int enable)
3824 {
3825         struct perf_output_handle handle;
3826         int ret;
3827
3828         struct {
3829                 struct perf_event_header        header;
3830                 u64                             time;
3831                 u64                             id;
3832                 u64                             stream_id;
3833         } throttle_event = {
3834                 .header = {
3835                         .type = PERF_RECORD_THROTTLE,
3836                         .misc = 0,
3837                         .size = sizeof(throttle_event),
3838                 },
3839                 .time           = perf_clock(),
3840                 .id             = primary_event_id(event),
3841                 .stream_id      = event->id,
3842         };
3843
3844         if (enable)
3845                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3846
3847         ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3848         if (ret)
3849                 return;
3850
3851         perf_output_put(&handle, throttle_event);
3852         perf_output_end(&handle);
3853 }
3854
3855 /*
3856  * Generic event overflow handling, sampling.
3857  */
3858
3859 static int __perf_event_overflow(struct perf_event *event, int nmi,
3860                                    int throttle, struct perf_sample_data *data,
3861                                    struct pt_regs *regs)
3862 {
3863         int events = atomic_read(&event->event_limit);
3864         struct hw_perf_event *hwc = &event->hw;
3865         int ret = 0;
3866
3867         throttle = (throttle && event->pmu->unthrottle != NULL);
3868
3869         if (!throttle) {
3870                 hwc->interrupts++;
3871         } else {
3872                 if (hwc->interrupts != MAX_INTERRUPTS) {
3873                         hwc->interrupts++;
3874                         if (HZ * hwc->interrupts >
3875                                         (u64)sysctl_perf_event_sample_rate) {
3876                                 hwc->interrupts = MAX_INTERRUPTS;
3877                                 perf_log_throttle(event, 0);
3878                                 ret = 1;
3879                         }
3880                 } else {
3881                         /*
3882                          * Keep re-disabling events even though on the previous
3883                          * pass we disabled it - just in case we raced with a
3884                          * sched-in and the event got enabled again:
3885                          */
3886                         ret = 1;
3887                 }
3888         }
3889
3890         if (event->attr.freq) {
3891                 u64 now = perf_clock();
3892                 s64 delta = now - hwc->freq_time_stamp;
3893
3894                 hwc->freq_time_stamp = now;
3895
3896                 if (delta > 0 && delta < 2*TICK_NSEC)
3897                         perf_adjust_period(event, delta, hwc->last_period);
3898         }
3899
3900         /*
3901          * XXX event_limit might not quite work as expected on inherited
3902          * events
3903          */
3904
3905         event->pending_kill = POLL_IN;
3906         if (events && atomic_dec_and_test(&event->event_limit)) {
3907                 ret = 1;
3908                 event->pending_kill = POLL_HUP;
3909                 if (nmi) {
3910                         event->pending_disable = 1;
3911                         perf_pending_queue(&event->pending,
3912                                            perf_pending_event);
3913                 } else
3914                         perf_event_disable(event);
3915         }
3916
3917         if (event->overflow_handler)
3918                 event->overflow_handler(event, nmi, data, regs);
3919         else
3920                 perf_event_output(event, nmi, data, regs);
3921
3922         return ret;
3923 }
3924
3925 int perf_event_overflow(struct perf_event *event, int nmi,
3926                           struct perf_sample_data *data,
3927                           struct pt_regs *regs)
3928 {
3929         return __perf_event_overflow(event, nmi, 1, data, regs);
3930 }
3931
3932 /*
3933  * Generic software event infrastructure
3934  */
3935
3936 /*
3937  * We directly increment event->count and keep a second value in
3938  * event->hw.period_left to count intervals. This period event
3939  * is kept in the range [-sample_period, 0] so that we can use the
3940  * sign as trigger.
3941  */
3942
3943 static u64 perf_swevent_set_period(struct perf_event *event)
3944 {
3945         struct hw_perf_event *hwc = &event->hw;
3946         u64 period = hwc->last_period;
3947         u64 nr, offset;
3948         s64 old, val;
3949
3950         hwc->last_period = hwc->sample_period;
3951
3952 again:
3953         old = val = atomic64_read(&hwc->period_left);
3954         if (val < 0)
3955                 return 0;
3956
3957         nr = div64_u64(period + val, period);
3958         offset = nr * period;
3959         val -= offset;
3960         if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3961                 goto again;
3962
3963         return nr;
3964 }
3965
3966 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3967                                     int nmi, struct perf_sample_data *data,
3968                                     struct pt_regs *regs)
3969 {
3970         struct hw_perf_event *hwc = &event->hw;
3971         int throttle = 0;
3972
3973         data->period = event->hw.last_period;
3974         if (!overflow)
3975                 overflow = perf_swevent_set_period(event);
3976
3977         if (hwc->interrupts == MAX_INTERRUPTS)
3978                 return;
3979
3980         for (; overflow; overflow--) {
3981                 if (__perf_event_overflow(event, nmi, throttle,
3982                                             data, regs)) {
3983                         /*
3984                          * We inhibit the overflow from happening when
3985                          * hwc->interrupts == MAX_INTERRUPTS.
3986                          */
3987                         break;
3988                 }
3989                 throttle = 1;
3990         }
3991 }
3992
3993 static void perf_swevent_unthrottle(struct perf_event *event)
3994 {
3995         /*
3996          * Nothing to do, we already reset hwc->interrupts.
3997          */
3998 }
3999
4000 static void perf_swevent_add(struct perf_event *event, u64 nr,
4001                                int nmi, struct perf_sample_data *data,
4002                                struct pt_regs *regs)
4003 {
4004         struct hw_perf_event *hwc = &event->hw;
4005
4006         atomic64_add(nr, &event->count);
4007
4008         if (!regs)
4009                 return;
4010
4011         if (!hwc->sample_period)
4012                 return;
4013
4014         if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4015                 return perf_swevent_overflow(event, 1, nmi, data, regs);
4016
4017         if (atomic64_add_negative(nr, &hwc->period_left))
4018                 return;
4019
4020         perf_swevent_overflow(event, 0, nmi, data, regs);
4021 }
4022
4023 static int perf_tp_event_match(struct perf_event *event,
4024                                 struct perf_sample_data *data);
4025
4026 static int perf_exclude_event(struct perf_event *event,
4027                               struct pt_regs *regs)
4028 {
4029         if (regs) {
4030                 if (event->attr.exclude_user && user_mode(regs))
4031                         return 1;
4032
4033                 if (event->attr.exclude_kernel && !user_mode(regs))
4034                         return 1;
4035         }
4036
4037         return 0;
4038 }
4039
4040 static int perf_swevent_match(struct perf_event *event,
4041                                 enum perf_type_id type,
4042                                 u32 event_id,
4043                                 struct perf_sample_data *data,
4044                                 struct pt_regs *regs)
4045 {
4046         if (event->attr.type != type)
4047                 return 0;
4048
4049         if (event->attr.config != event_id)
4050                 return 0;
4051
4052         if (perf_exclude_event(event, regs))
4053                 return 0;
4054
4055         if (event->attr.type == PERF_TYPE_TRACEPOINT &&
4056             !perf_tp_event_match(event, data))
4057                 return 0;
4058
4059         return 1;
4060 }
4061
4062 static inline u64 swevent_hash(u64 type, u32 event_id)
4063 {
4064         u64 val = event_id | (type << 32);
4065
4066         return hash_64(val, SWEVENT_HLIST_BITS);
4067 }
4068
4069 static struct hlist_head *
4070 find_swevent_head(struct perf_cpu_context *ctx, u64 type, u32 event_id)
4071 {
4072         u64 hash;
4073         struct swevent_hlist *hlist;
4074
4075         hash = swevent_hash(type, event_id);
4076
4077         hlist = rcu_dereference(ctx->swevent_hlist);
4078         if (!hlist)
4079                 return NULL;
4080
4081         return &hlist->heads[hash];
4082 }
4083
4084 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4085                                     u64 nr, int nmi,
4086                                     struct perf_sample_data *data,
4087                                     struct pt_regs *regs)
4088 {
4089         struct perf_cpu_context *cpuctx;
4090         struct perf_event *event;
4091         struct hlist_node *node;
4092         struct hlist_head *head;
4093
4094         cpuctx = &__get_cpu_var(perf_cpu_context);
4095
4096         rcu_read_lock();
4097
4098         head = find_swevent_head(cpuctx, type, event_id);
4099
4100         if (!head)
4101                 goto end;
4102
4103         hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4104                 if (perf_swevent_match(event, type, event_id, data, regs))
4105                         perf_swevent_add(event, nr, nmi, data, regs);
4106         }
4107 end:
4108         rcu_read_unlock();
4109 }
4110
4111 int perf_swevent_get_recursion_context(void)
4112 {
4113         struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
4114         int rctx;
4115
4116         if (in_nmi())
4117                 rctx = 3;
4118         else if (in_irq())
4119                 rctx = 2;
4120         else if (in_softirq())
4121                 rctx = 1;
4122         else
4123                 rctx = 0;
4124
4125         if (cpuctx->recursion[rctx]) {
4126                 put_cpu_var(perf_cpu_context);
4127                 return -1;
4128         }
4129
4130         cpuctx->recursion[rctx]++;
4131         barrier();
4132
4133         return rctx;
4134 }
4135 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4136
4137 void perf_swevent_put_recursion_context(int rctx)
4138 {
4139         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4140         barrier();
4141         cpuctx->recursion[rctx]--;
4142         put_cpu_var(perf_cpu_context);
4143 }
4144 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
4145
4146
4147 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4148                             struct pt_regs *regs, u64 addr)
4149 {
4150         struct perf_sample_data data;
4151         int rctx;
4152
4153         rctx = perf_swevent_get_recursion_context();
4154         if (rctx < 0)
4155                 return;
4156
4157         perf_sample_data_init(&data, addr);
4158
4159         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4160
4161         perf_swevent_put_recursion_context(rctx);
4162 }
4163
4164 static void perf_swevent_read(struct perf_event *event)
4165 {
4166 }
4167
4168 static int perf_swevent_enable(struct perf_event *event)
4169 {
4170         struct hw_perf_event *hwc = &event->hw;
4171         struct perf_cpu_context *cpuctx;
4172         struct hlist_head *head;
4173
4174         cpuctx = &__get_cpu_var(perf_cpu_context);
4175
4176         if (hwc->sample_period) {
4177                 hwc->last_period = hwc->sample_period;
4178                 perf_swevent_set_period(event);
4179         }
4180
4181         head = find_swevent_head(cpuctx, event->attr.type, event->attr.config);
4182         if (WARN_ON_ONCE(!head))
4183                 return -EINVAL;
4184
4185         hlist_add_head_rcu(&event->hlist_entry, head);
4186
4187         return 0;
4188 }
4189
4190 static void perf_swevent_disable(struct perf_event *event)
4191 {
4192         hlist_del_rcu(&event->hlist_entry);
4193 }
4194
4195 static const struct pmu perf_ops_generic = {
4196         .enable         = perf_swevent_enable,
4197         .disable        = perf_swevent_disable,
4198         .read           = perf_swevent_read,
4199         .unthrottle     = perf_swevent_unthrottle,
4200 };
4201
4202 /*
4203  * hrtimer based swevent callback
4204  */
4205
4206 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4207 {
4208         enum hrtimer_restart ret = HRTIMER_RESTART;
4209         struct perf_sample_data data;
4210         struct pt_regs *regs;
4211         struct perf_event *event;
4212         u64 period;
4213
4214         event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4215         event->pmu->read(event);
4216
4217         perf_sample_data_init(&data, 0);
4218         data.period = event->hw.last_period;
4219         regs = get_irq_regs();
4220
4221         if (regs && !perf_exclude_event(event, regs)) {
4222                 if (!(event->attr.exclude_idle && current->pid == 0))
4223                         if (perf_event_overflow(event, 0, &data, regs))
4224                                 ret = HRTIMER_NORESTART;
4225         }
4226
4227         period = max_t(u64, 10000, event->hw.sample_period);
4228         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4229
4230         return ret;
4231 }
4232
4233 static void perf_swevent_start_hrtimer(struct perf_event *event)
4234 {
4235         struct hw_perf_event *hwc = &event->hw;
4236
4237         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4238         hwc->hrtimer.function = perf_swevent_hrtimer;
4239         if (hwc->sample_period) {
4240                 u64 period;
4241
4242                 if (hwc->remaining) {
4243                         if (hwc->remaining < 0)
4244                                 period = 10000;
4245                         else
4246                                 period = hwc->remaining;
4247                         hwc->remaining = 0;
4248                 } else {
4249                         period = max_t(u64, 10000, hwc->sample_period);
4250                 }
4251                 __hrtimer_start_range_ns(&hwc->hrtimer,
4252                                 ns_to_ktime(period), 0,
4253                                 HRTIMER_MODE_REL, 0);
4254         }
4255 }
4256
4257 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4258 {
4259         struct hw_perf_event *hwc = &event->hw;
4260
4261         if (hwc->sample_period) {
4262                 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4263                 hwc->remaining = ktime_to_ns(remaining);
4264
4265                 hrtimer_cancel(&hwc->hrtimer);
4266         }
4267 }
4268
4269 /*
4270  * Software event: cpu wall time clock
4271  */
4272
4273 static void cpu_clock_perf_event_update(struct perf_event *event)
4274 {
4275         int cpu = raw_smp_processor_id();
4276         s64 prev;
4277         u64 now;
4278
4279         now = cpu_clock(cpu);
4280         prev = atomic64_xchg(&event->hw.prev_count, now);
4281         atomic64_add(now - prev, &event->count);
4282 }
4283
4284 static int cpu_clock_perf_event_enable(struct perf_event *event)
4285 {
4286         struct hw_perf_event *hwc = &event->hw;
4287         int cpu = raw_smp_processor_id();
4288
4289         atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4290         perf_swevent_start_hrtimer(event);
4291
4292         return 0;
4293 }
4294
4295 static void cpu_clock_perf_event_disable(struct perf_event *event)
4296 {
4297         perf_swevent_cancel_hrtimer(event);
4298         cpu_clock_perf_event_update(event);
4299 }
4300
4301 static void cpu_clock_perf_event_read(struct perf_event *event)
4302 {
4303         cpu_clock_perf_event_update(event);
4304 }
4305
4306 static const struct pmu perf_ops_cpu_clock = {
4307         .enable         = cpu_clock_perf_event_enable,
4308         .disable        = cpu_clock_perf_event_disable,
4309         .read           = cpu_clock_perf_event_read,
4310 };
4311
4312 /*
4313  * Software event: task time clock
4314  */
4315
4316 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4317 {
4318         u64 prev;
4319         s64 delta;
4320
4321         prev = atomic64_xchg(&event->hw.prev_count, now);
4322         delta = now - prev;
4323         atomic64_add(delta, &event->count);
4324 }
4325
4326 static int task_clock_perf_event_enable(struct perf_event *event)
4327 {
4328         struct hw_perf_event *hwc = &event->hw;
4329         u64 now;
4330
4331         now = event->ctx->time;
4332
4333         atomic64_set(&hwc->prev_count, now);
4334
4335         perf_swevent_start_hrtimer(event);
4336
4337         return 0;
4338 }
4339
4340 static void task_clock_perf_event_disable(struct perf_event *event)
4341 {
4342         perf_swevent_cancel_hrtimer(event);
4343         task_clock_perf_event_update(event, event->ctx->time);
4344
4345 }
4346
4347 static void task_clock_perf_event_read(struct perf_event *event)
4348 {
4349         u64 time;
4350
4351         if (!in_nmi()) {
4352                 update_context_time(event->ctx);
4353                 time = event->ctx->time;
4354         } else {
4355                 u64 now = perf_clock();
4356                 u64 delta = now - event->ctx->timestamp;
4357                 time = event->ctx->time + delta;
4358         }
4359
4360         task_clock_perf_event_update(event, time);
4361 }
4362
4363 static const struct pmu perf_ops_task_clock = {
4364         .enable         = task_clock_perf_event_enable,
4365         .disable        = task_clock_perf_event_disable,
4366         .read           = task_clock_perf_event_read,
4367 };
4368
4369 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4370 {
4371         struct swevent_hlist *hlist;
4372
4373         hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4374         kfree(hlist);
4375 }
4376
4377 static void swevent_hlist_release(struct perf_cpu_context *cpuctx)
4378 {
4379         struct swevent_hlist *hlist;
4380
4381         if (!cpuctx->swevent_hlist)
4382                 return;
4383
4384         hlist = cpuctx->swevent_hlist;
4385         rcu_assign_pointer(cpuctx->swevent_hlist, NULL);
4386         call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4387 }
4388
4389 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4390 {
4391         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4392
4393         mutex_lock(&cpuctx->hlist_mutex);
4394
4395         if (!--cpuctx->hlist_refcount)
4396                 swevent_hlist_release(cpuctx);
4397
4398         mutex_unlock(&cpuctx->hlist_mutex);
4399 }
4400
4401 static void swevent_hlist_put(struct perf_event *event)
4402 {
4403         int cpu;
4404
4405         if (event->cpu != -1) {
4406                 swevent_hlist_put_cpu(event, event->cpu);
4407                 return;
4408         }
4409
4410         for_each_possible_cpu(cpu)
4411                 swevent_hlist_put_cpu(event, cpu);
4412 }
4413
4414 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4415 {
4416         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4417         int err = 0;
4418
4419         mutex_lock(&cpuctx->hlist_mutex);
4420
4421         if (!cpuctx->swevent_hlist && cpu_online(cpu)) {
4422                 struct swevent_hlist *hlist;
4423
4424                 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4425                 if (!hlist) {
4426                         err = -ENOMEM;
4427                         goto exit;
4428                 }
4429                 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
4430         }
4431         cpuctx->hlist_refcount++;
4432  exit:
4433         mutex_unlock(&cpuctx->hlist_mutex);
4434
4435         return err;
4436 }
4437
4438 static int swevent_hlist_get(struct perf_event *event)
4439 {
4440         int err;
4441         int cpu, failed_cpu;
4442
4443         if (event->cpu != -1)
4444                 return swevent_hlist_get_cpu(event, event->cpu);
4445
4446         get_online_cpus();
4447         for_each_possible_cpu(cpu) {
4448                 err = swevent_hlist_get_cpu(event, cpu);
4449                 if (err) {
4450                         failed_cpu = cpu;
4451                         goto fail;
4452                 }
4453         }
4454         put_online_cpus();
4455
4456         return 0;
4457  fail:
4458         for_each_possible_cpu(cpu) {
4459                 if (cpu == failed_cpu)
4460                         break;
4461                 swevent_hlist_put_cpu(event, cpu);
4462         }
4463
4464         put_online_cpus();
4465         return err;
4466 }
4467
4468 #ifdef CONFIG_EVENT_TRACING
4469
4470 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4471                    int entry_size, struct pt_regs *regs)
4472 {
4473         struct perf_sample_data data;
4474         struct perf_raw_record raw = {
4475                 .size = entry_size,
4476                 .data = record,
4477         };
4478
4479         perf_sample_data_init(&data, addr);
4480         data.raw = &raw;
4481
4482         /* Trace events already protected against recursion */
4483         do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4484                          &data, regs);
4485 }
4486 EXPORT_SYMBOL_GPL(perf_tp_event);
4487
4488 static int perf_tp_event_match(struct perf_event *event,
4489                                 struct perf_sample_data *data)
4490 {
4491         void *record = data->raw->data;
4492
4493         if (likely(!event->filter) || filter_match_preds(event->filter, record))
4494                 return 1;
4495         return 0;
4496 }
4497
4498 static void tp_perf_event_destroy(struct perf_event *event)
4499 {
4500         perf_trace_disable(event->attr.config);
4501         swevent_hlist_put(event);
4502 }
4503
4504 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4505 {
4506         int err;
4507
4508         /*
4509          * Raw tracepoint data is a severe data leak, only allow root to
4510          * have these.
4511          */
4512         if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4513                         perf_paranoid_tracepoint_raw() &&
4514                         !capable(CAP_SYS_ADMIN))
4515                 return ERR_PTR(-EPERM);
4516
4517         if (perf_trace_enable(event->attr.config))
4518                 return NULL;
4519
4520         event->destroy = tp_perf_event_destroy;
4521         err = swevent_hlist_get(event);
4522         if (err) {
4523                 perf_trace_disable(event->attr.config);
4524                 return ERR_PTR(err);
4525         }
4526
4527         return &perf_ops_generic;
4528 }
4529
4530 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4531 {
4532         char *filter_str;
4533         int ret;
4534
4535         if (event->attr.type != PERF_TYPE_TRACEPOINT)
4536                 return -EINVAL;
4537
4538         filter_str = strndup_user(arg, PAGE_SIZE);
4539         if (IS_ERR(filter_str))
4540                 return PTR_ERR(filter_str);
4541
4542         ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4543
4544         kfree(filter_str);
4545         return ret;
4546 }
4547
4548 static void perf_event_free_filter(struct perf_event *event)
4549 {
4550         ftrace_profile_free_filter(event);
4551 }
4552
4553 #else
4554
4555 static int perf_tp_event_match(struct perf_event *event,
4556                                 struct perf_sample_data *data)
4557 {
4558         return 1;
4559 }
4560
4561 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4562 {
4563         return NULL;
4564 }
4565
4566 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4567 {
4568         return -ENOENT;
4569 }
4570
4571 static void perf_event_free_filter(struct perf_event *event)
4572 {
4573 }
4574
4575 #endif /* CONFIG_EVENT_TRACING */
4576
4577 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4578 static void bp_perf_event_destroy(struct perf_event *event)
4579 {
4580         release_bp_slot(event);
4581 }
4582
4583 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4584 {
4585         int err;
4586
4587         err = register_perf_hw_breakpoint(bp);
4588         if (err)
4589                 return ERR_PTR(err);
4590
4591         bp->destroy = bp_perf_event_destroy;
4592
4593         return &perf_ops_bp;
4594 }
4595
4596 void perf_bp_event(struct perf_event *bp, void *data)
4597 {
4598         struct perf_sample_data sample;
4599         struct pt_regs *regs = data;
4600
4601         perf_sample_data_init(&sample, bp->attr.bp_addr);
4602
4603         if (!perf_exclude_event(bp, regs))
4604                 perf_swevent_add(bp, 1, 1, &sample, regs);
4605 }
4606 #else
4607 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4608 {
4609         return NULL;
4610 }
4611
4612 void perf_bp_event(struct perf_event *bp, void *regs)
4613 {
4614 }
4615 #endif
4616
4617 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4618
4619 static void sw_perf_event_destroy(struct perf_event *event)
4620 {
4621         u64 event_id = event->attr.config;
4622
4623         WARN_ON(event->parent);
4624
4625         atomic_dec(&perf_swevent_enabled[event_id]);
4626         swevent_hlist_put(event);
4627 }
4628
4629 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4630 {
4631         const struct pmu *pmu = NULL;
4632         u64 event_id = event->attr.config;
4633
4634         /*
4635          * Software events (currently) can't in general distinguish
4636          * between user, kernel and hypervisor events.
4637          * However, context switches and cpu migrations are considered
4638          * to be kernel events, and page faults are never hypervisor
4639          * events.
4640          */
4641         switch (event_id) {
4642         case PERF_COUNT_SW_CPU_CLOCK:
4643                 pmu = &perf_ops_cpu_clock;
4644
4645                 break;
4646         case PERF_COUNT_SW_TASK_CLOCK:
4647                 /*
4648                  * If the user instantiates this as a per-cpu event,
4649                  * use the cpu_clock event instead.
4650                  */
4651                 if (event->ctx->task)
4652                         pmu = &perf_ops_task_clock;
4653                 else
4654                         pmu = &perf_ops_cpu_clock;
4655
4656                 break;
4657         case PERF_COUNT_SW_PAGE_FAULTS:
4658         case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4659         case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4660         case PERF_COUNT_SW_CONTEXT_SWITCHES:
4661         case PERF_COUNT_SW_CPU_MIGRATIONS:
4662         case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4663         case PERF_COUNT_SW_EMULATION_FAULTS:
4664                 if (!event->parent) {
4665                         int err;
4666
4667                         err = swevent_hlist_get(event);
4668                         if (err)
4669                                 return ERR_PTR(err);
4670
4671                         atomic_inc(&perf_swevent_enabled[event_id]);
4672                         event->destroy = sw_perf_event_destroy;
4673                 }
4674                 pmu = &perf_ops_generic;
4675                 break;
4676         }
4677
4678         return pmu;
4679 }
4680
4681 /*
4682  * Allocate and initialize a event structure
4683  */
4684 static struct perf_event *
4685 perf_event_alloc(struct perf_event_attr *attr,
4686                    int cpu,
4687                    struct perf_event_context *ctx,
4688                    struct perf_event *group_leader,
4689                    struct perf_event *parent_event,
4690                    perf_overflow_handler_t overflow_handler,
4691                    gfp_t gfpflags)
4692 {
4693         const struct pmu *pmu;
4694         struct perf_event *event;
4695         struct hw_perf_event *hwc;
4696         long err;
4697
4698         event = kzalloc(sizeof(*event), gfpflags);
4699         if (!event)
4700                 return ERR_PTR(-ENOMEM);
4701
4702         /*
4703          * Single events are their own group leaders, with an
4704          * empty sibling list:
4705          */
4706         if (!group_leader)
4707                 group_leader = event;
4708
4709         mutex_init(&event->child_mutex);
4710         INIT_LIST_HEAD(&event->child_list);
4711
4712         INIT_LIST_HEAD(&event->group_entry);
4713         INIT_LIST_HEAD(&event->event_entry);
4714         INIT_LIST_HEAD(&event->sibling_list);
4715         init_waitqueue_head(&event->waitq);
4716
4717         mutex_init(&event->mmap_mutex);
4718
4719         event->cpu              = cpu;
4720         event->attr             = *attr;
4721         event->group_leader     = group_leader;
4722         event->pmu              = NULL;
4723         event->ctx              = ctx;
4724         event->oncpu            = -1;
4725
4726         event->parent           = parent_event;
4727
4728         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
4729         event->id               = atomic64_inc_return(&perf_event_id);
4730
4731         event->state            = PERF_EVENT_STATE_INACTIVE;
4732
4733         if (!overflow_handler && parent_event)
4734                 overflow_handler = parent_event->overflow_handler;
4735         
4736         event->overflow_handler = overflow_handler;
4737
4738         if (attr->disabled)
4739                 event->state = PERF_EVENT_STATE_OFF;
4740
4741         pmu = NULL;
4742
4743         hwc = &event->hw;
4744         hwc->sample_period = attr->sample_period;
4745         if (attr->freq && attr->sample_freq)
4746                 hwc->sample_period = 1;
4747         hwc->last_period = hwc->sample_period;
4748
4749         atomic64_set(&hwc->period_left, hwc->sample_period);
4750
4751         /*
4752          * we currently do not support PERF_FORMAT_GROUP on inherited events
4753          */
4754         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4755                 goto done;
4756
4757         switch (attr->type) {
4758         case PERF_TYPE_RAW:
4759         case PERF_TYPE_HARDWARE:
4760         case PERF_TYPE_HW_CACHE:
4761                 pmu = hw_perf_event_init(event);
4762                 break;
4763
4764         case PERF_TYPE_SOFTWARE:
4765                 pmu = sw_perf_event_init(event);
4766                 break;
4767
4768         case PERF_TYPE_TRACEPOINT:
4769                 pmu = tp_perf_event_init(event);
4770                 break;
4771
4772         case PERF_TYPE_BREAKPOINT:
4773                 pmu = bp_perf_event_init(event);
4774                 break;
4775
4776
4777         default:
4778                 break;
4779         }
4780 done:
4781         err = 0;
4782         if (!pmu)
4783                 err = -EINVAL;
4784         else if (IS_ERR(pmu))
4785                 err = PTR_ERR(pmu);
4786
4787         if (err) {
4788                 if (event->ns)
4789                         put_pid_ns(event->ns);
4790                 kfree(event);
4791                 return ERR_PTR(err);
4792         }
4793
4794         event->pmu = pmu;
4795
4796         if (!event->parent) {
4797                 atomic_inc(&nr_events);
4798                 if (event->attr.mmap)
4799                         atomic_inc(&nr_mmap_events);
4800                 if (event->attr.comm)
4801                         atomic_inc(&nr_comm_events);
4802                 if (event->attr.task)
4803                         atomic_inc(&nr_task_events);
4804         }
4805
4806         return event;
4807 }
4808
4809 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4810                           struct perf_event_attr *attr)
4811 {
4812         u32 size;
4813         int ret;
4814
4815         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4816                 return -EFAULT;
4817
4818         /*
4819          * zero the full structure, so that a short copy will be nice.
4820          */
4821         memset(attr, 0, sizeof(*attr));
4822
4823         ret = get_user(size, &uattr->size);
4824         if (ret)
4825                 return ret;
4826
4827         if (size > PAGE_SIZE)   /* silly large */
4828                 goto err_size;
4829
4830         if (!size)              /* abi compat */
4831                 size = PERF_ATTR_SIZE_VER0;
4832
4833         if (size < PERF_ATTR_SIZE_VER0)
4834                 goto err_size;
4835
4836         /*
4837          * If we're handed a bigger struct than we know of,
4838          * ensure all the unknown bits are 0 - i.e. new
4839          * user-space does not rely on any kernel feature
4840          * extensions we dont know about yet.
4841          */
4842         if (size > sizeof(*attr)) {
4843                 unsigned char __user *addr;
4844                 unsigned char __user *end;
4845                 unsigned char val;
4846
4847                 addr = (void __user *)uattr + sizeof(*attr);
4848                 end  = (void __user *)uattr + size;
4849
4850                 for (; addr < end; addr++) {
4851                         ret = get_user(val, addr);
4852                         if (ret)
4853                                 return ret;
4854                         if (val)
4855                                 goto err_size;
4856                 }
4857                 size = sizeof(*attr);
4858         }
4859
4860         ret = copy_from_user(attr, uattr, size);
4861         if (ret)
4862                 return -EFAULT;
4863
4864         /*
4865          * If the type exists, the corresponding creation will verify
4866          * the attr->config.
4867          */
4868         if (attr->type >= PERF_TYPE_MAX)
4869                 return -EINVAL;
4870
4871         if (attr->__reserved_1)
4872                 return -EINVAL;
4873
4874         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4875                 return -EINVAL;
4876
4877         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4878                 return -EINVAL;
4879
4880 out:
4881         return ret;
4882
4883 err_size:
4884         put_user(sizeof(*attr), &uattr->size);
4885         ret = -E2BIG;
4886         goto out;
4887 }
4888
4889 static int perf_event_set_output(struct perf_event *event, int output_fd)
4890 {
4891         struct perf_event *output_event = NULL;
4892         struct file *output_file = NULL;
4893         struct perf_event *old_output;
4894         int fput_needed = 0;
4895         int ret = -EINVAL;
4896
4897         if (!output_fd)
4898                 goto set;
4899
4900         output_file = fget_light(output_fd, &fput_needed);
4901         if (!output_file)
4902                 return -EBADF;
4903
4904         if (output_file->f_op != &perf_fops)
4905                 goto out;
4906
4907         output_event = output_file->private_data;
4908
4909         /* Don't chain output fds */
4910         if (output_event->output)
4911                 goto out;
4912
4913         /* Don't set an output fd when we already have an output channel */
4914         if (event->data)
4915                 goto out;
4916
4917         atomic_long_inc(&output_file->f_count);
4918
4919 set:
4920         mutex_lock(&event->mmap_mutex);
4921         old_output = event->output;
4922         rcu_assign_pointer(event->output, output_event);
4923         mutex_unlock(&event->mmap_mutex);
4924
4925         if (old_output) {
4926                 /*
4927                  * we need to make sure no existing perf_output_*()
4928                  * is still referencing this event.
4929                  */
4930                 synchronize_rcu();
4931                 fput(old_output->filp);
4932         }
4933
4934         ret = 0;
4935 out:
4936         fput_light(output_file, fput_needed);
4937         return ret;
4938 }
4939
4940 /**
4941  * sys_perf_event_open - open a performance event, associate it to a task/cpu
4942  *
4943  * @attr_uptr:  event_id type attributes for monitoring/sampling
4944  * @pid:                target pid
4945  * @cpu:                target cpu
4946  * @group_fd:           group leader event fd
4947  */
4948 SYSCALL_DEFINE5(perf_event_open,
4949                 struct perf_event_attr __user *, attr_uptr,
4950                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4951 {
4952         struct perf_event *event, *group_leader;
4953         struct perf_event_attr attr;
4954         struct perf_event_context *ctx;
4955         struct file *event_file = NULL;
4956         struct file *group_file = NULL;
4957         int fput_needed = 0;
4958         int fput_needed2 = 0;
4959         int err;
4960
4961         /* for future expandability... */
4962         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4963                 return -EINVAL;
4964
4965         err = perf_copy_attr(attr_uptr, &attr);
4966         if (err)
4967                 return err;
4968
4969         if (!attr.exclude_kernel) {
4970                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4971                         return -EACCES;
4972         }
4973
4974         if (attr.freq) {
4975                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4976                         return -EINVAL;
4977         }
4978
4979         /*
4980          * Get the target context (task or percpu):
4981          */
4982         ctx = find_get_context(pid, cpu);
4983         if (IS_ERR(ctx))
4984                 return PTR_ERR(ctx);
4985
4986         /*
4987          * Look up the group leader (we will attach this event to it):
4988          */
4989         group_leader = NULL;
4990         if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4991                 err = -EINVAL;
4992                 group_file = fget_light(group_fd, &fput_needed);
4993                 if (!group_file)
4994                         goto err_put_context;
4995                 if (group_file->f_op != &perf_fops)
4996                         goto err_put_context;
4997
4998                 group_leader = group_file->private_data;
4999                 /*
5000                  * Do not allow a recursive hierarchy (this new sibling
5001                  * becoming part of another group-sibling):
5002                  */
5003                 if (group_leader->group_leader != group_leader)
5004                         goto err_put_context;
5005                 /*
5006                  * Do not allow to attach to a group in a different
5007                  * task or CPU context:
5008                  */
5009                 if (group_leader->ctx != ctx)
5010                         goto err_put_context;
5011                 /*
5012                  * Only a group leader can be exclusive or pinned
5013                  */
5014                 if (attr.exclusive || attr.pinned)
5015                         goto err_put_context;
5016         }
5017
5018         event = perf_event_alloc(&attr, cpu, ctx, group_leader,
5019                                      NULL, NULL, GFP_KERNEL);
5020         err = PTR_ERR(event);
5021         if (IS_ERR(event))
5022                 goto err_put_context;
5023
5024         err = anon_inode_getfd("[perf_event]", &perf_fops, event, O_RDWR);
5025         if (err < 0)
5026                 goto err_free_put_context;
5027
5028         event_file = fget_light(err, &fput_needed2);
5029         if (!event_file)
5030                 goto err_free_put_context;
5031
5032         if (flags & PERF_FLAG_FD_OUTPUT) {
5033                 err = perf_event_set_output(event, group_fd);
5034                 if (err)
5035                         goto err_fput_free_put_context;
5036         }
5037
5038         event->filp = event_file;
5039         WARN_ON_ONCE(ctx->parent_ctx);
5040         mutex_lock(&ctx->mutex);
5041         perf_install_in_context(ctx, event, cpu);
5042         ++ctx->generation;
5043         mutex_unlock(&ctx->mutex);
5044
5045         event->owner = current;
5046         get_task_struct(current);
5047         mutex_lock(&current->perf_event_mutex);
5048         list_add_tail(&event->owner_entry, &current->perf_event_list);
5049         mutex_unlock(&current->perf_event_mutex);
5050
5051 err_fput_free_put_context:
5052         fput_light(event_file, fput_needed2);
5053
5054 err_free_put_context:
5055         if (err < 0)
5056                 free_event(event);
5057
5058 err_put_context:
5059         if (err < 0)
5060                 put_ctx(ctx);
5061
5062         fput_light(group_file, fput_needed);
5063
5064         return err;
5065 }
5066
5067 /**
5068  * perf_event_create_kernel_counter
5069  *
5070  * @attr: attributes of the counter to create
5071  * @cpu: cpu in which the counter is bound
5072  * @pid: task to profile
5073  */
5074 struct perf_event *
5075 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
5076                                  pid_t pid,
5077                                  perf_overflow_handler_t overflow_handler)
5078 {
5079         struct perf_event *event;
5080         struct perf_event_context *ctx;
5081         int err;
5082
5083         /*
5084          * Get the target context (task or percpu):
5085          */
5086
5087         ctx = find_get_context(pid, cpu);
5088         if (IS_ERR(ctx)) {
5089                 err = PTR_ERR(ctx);
5090                 goto err_exit;
5091         }
5092
5093         event = perf_event_alloc(attr, cpu, ctx, NULL,
5094                                  NULL, overflow_handler, GFP_KERNEL);
5095         if (IS_ERR(event)) {
5096                 err = PTR_ERR(event);
5097                 goto err_put_context;
5098         }
5099
5100         event->filp = NULL;
5101         WARN_ON_ONCE(ctx->parent_ctx);
5102         mutex_lock(&ctx->mutex);
5103         perf_install_in_context(ctx, event, cpu);
5104         ++ctx->generation;
5105         mutex_unlock(&ctx->mutex);
5106
5107         event->owner = current;
5108         get_task_struct(current);
5109         mutex_lock(&current->perf_event_mutex);
5110         list_add_tail(&event->owner_entry, &current->perf_event_list);
5111         mutex_unlock(&current->perf_event_mutex);
5112
5113         return event;
5114
5115  err_put_context:
5116         put_ctx(ctx);
5117  err_exit:
5118         return ERR_PTR(err);
5119 }
5120 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
5121
5122 /*
5123  * inherit a event from parent task to child task:
5124  */
5125 static struct perf_event *
5126 inherit_event(struct perf_event *parent_event,
5127               struct task_struct *parent,
5128               struct perf_event_context *parent_ctx,
5129               struct task_struct *child,
5130               struct perf_event *group_leader,
5131               struct perf_event_context *child_ctx)
5132 {
5133         struct perf_event *child_event;
5134
5135         /*
5136          * Instead of creating recursive hierarchies of events,
5137          * we link inherited events back to the original parent,
5138          * which has a filp for sure, which we use as the reference
5139          * count:
5140          */
5141         if (parent_event->parent)
5142                 parent_event = parent_event->parent;
5143
5144         child_event = perf_event_alloc(&parent_event->attr,
5145                                            parent_event->cpu, child_ctx,
5146                                            group_leader, parent_event,
5147                                            NULL, GFP_KERNEL);
5148         if (IS_ERR(child_event))
5149                 return child_event;
5150         get_ctx(child_ctx);
5151
5152         /*
5153          * Make the child state follow the state of the parent event,
5154          * not its attr.disabled bit.  We hold the parent's mutex,
5155          * so we won't race with perf_event_{en, dis}able_family.
5156          */
5157         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
5158                 child_event->state = PERF_EVENT_STATE_INACTIVE;
5159         else
5160                 child_event->state = PERF_EVENT_STATE_OFF;
5161
5162         if (parent_event->attr.freq) {
5163                 u64 sample_period = parent_event->hw.sample_period;
5164                 struct hw_perf_event *hwc = &child_event->hw;
5165
5166                 hwc->sample_period = sample_period;
5167                 hwc->last_period   = sample_period;
5168
5169                 atomic64_set(&hwc->period_left, sample_period);
5170         }
5171
5172         child_event->overflow_handler = parent_event->overflow_handler;
5173
5174         /*
5175          * Link it up in the child's context:
5176          */
5177         add_event_to_ctx(child_event, child_ctx);
5178
5179         /*
5180          * Get a reference to the parent filp - we will fput it
5181          * when the child event exits. This is safe to do because
5182          * we are in the parent and we know that the filp still
5183          * exists and has a nonzero count:
5184          */
5185         atomic_long_inc(&parent_event->filp->f_count);
5186
5187         /*
5188          * Link this into the parent event's child list
5189          */
5190         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5191         mutex_lock(&parent_event->child_mutex);
5192         list_add_tail(&child_event->child_list, &parent_event->child_list);
5193         mutex_unlock(&parent_event->child_mutex);
5194
5195         return child_event;
5196 }
5197
5198 static int inherit_group(struct perf_event *parent_event,
5199               struct task_struct *parent,
5200               struct perf_event_context *parent_ctx,
5201               struct task_struct *child,
5202               struct perf_event_context *child_ctx)
5203 {
5204         struct perf_event *leader;
5205         struct perf_event *sub;
5206         struct perf_event *child_ctr;
5207
5208         leader = inherit_event(parent_event, parent, parent_ctx,
5209                                  child, NULL, child_ctx);
5210         if (IS_ERR(leader))
5211                 return PTR_ERR(leader);
5212         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5213                 child_ctr = inherit_event(sub, parent, parent_ctx,
5214                                             child, leader, child_ctx);
5215                 if (IS_ERR(child_ctr))
5216                         return PTR_ERR(child_ctr);
5217         }
5218         return 0;
5219 }
5220
5221 static void sync_child_event(struct perf_event *child_event,
5222                                struct task_struct *child)
5223 {
5224         struct perf_event *parent_event = child_event->parent;
5225         u64 child_val;
5226
5227         if (child_event->attr.inherit_stat)
5228                 perf_event_read_event(child_event, child);
5229
5230         child_val = atomic64_read(&child_event->count);
5231
5232         /*
5233          * Add back the child's count to the parent's count:
5234          */
5235         atomic64_add(child_val, &parent_event->count);
5236         atomic64_add(child_event->total_time_enabled,
5237                      &parent_event->child_total_time_enabled);
5238         atomic64_add(child_event->total_time_running,
5239                      &parent_event->child_total_time_running);
5240
5241         /*
5242          * Remove this event from the parent's list
5243          */
5244         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5245         mutex_lock(&parent_event->child_mutex);
5246         list_del_init(&child_event->child_list);
5247         mutex_unlock(&parent_event->child_mutex);
5248
5249         /*
5250          * Release the parent event, if this was the last
5251          * reference to it.
5252          */
5253         fput(parent_event->filp);
5254 }
5255
5256 static void
5257 __perf_event_exit_task(struct perf_event *child_event,
5258                          struct perf_event_context *child_ctx,
5259                          struct task_struct *child)
5260 {
5261         struct perf_event *parent_event;
5262
5263         perf_event_remove_from_context(child_event);
5264
5265         parent_event = child_event->parent;
5266         /*
5267          * It can happen that parent exits first, and has events
5268          * that are still around due to the child reference. These
5269          * events need to be zapped - but otherwise linger.
5270          */
5271         if (parent_event) {
5272                 sync_child_event(child_event, child);
5273                 free_event(child_event);
5274         }
5275 }
5276
5277 /*
5278  * When a child task exits, feed back event values to parent events.
5279  */
5280 void perf_event_exit_task(struct task_struct *child)
5281 {
5282         struct perf_event *child_event, *tmp;
5283         struct perf_event_context *child_ctx;
5284         unsigned long flags;
5285
5286         if (likely(!child->perf_event_ctxp)) {
5287                 perf_event_task(child, NULL, 0);
5288                 return;
5289         }
5290
5291         local_irq_save(flags);
5292         /*
5293          * We can't reschedule here because interrupts are disabled,
5294          * and either child is current or it is a task that can't be
5295          * scheduled, so we are now safe from rescheduling changing
5296          * our context.
5297          */
5298         child_ctx = child->perf_event_ctxp;
5299         __perf_event_task_sched_out(child_ctx);
5300
5301         /*
5302          * Take the context lock here so that if find_get_context is
5303          * reading child->perf_event_ctxp, we wait until it has
5304          * incremented the context's refcount before we do put_ctx below.
5305          */
5306         raw_spin_lock(&child_ctx->lock);
5307         child->perf_event_ctxp = NULL;
5308         /*
5309          * If this context is a clone; unclone it so it can't get
5310          * swapped to another process while we're removing all
5311          * the events from it.
5312          */
5313         unclone_ctx(child_ctx);
5314         update_context_time(child_ctx);
5315         raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5316
5317         /*
5318          * Report the task dead after unscheduling the events so that we
5319          * won't get any samples after PERF_RECORD_EXIT. We can however still
5320          * get a few PERF_RECORD_READ events.
5321          */
5322         perf_event_task(child, child_ctx, 0);
5323
5324         /*
5325          * We can recurse on the same lock type through:
5326          *
5327          *   __perf_event_exit_task()
5328          *     sync_child_event()
5329          *       fput(parent_event->filp)
5330          *         perf_release()
5331          *           mutex_lock(&ctx->mutex)
5332          *
5333          * But since its the parent context it won't be the same instance.
5334          */
5335         mutex_lock(&child_ctx->mutex);
5336
5337 again:
5338         list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5339                                  group_entry)
5340                 __perf_event_exit_task(child_event, child_ctx, child);
5341
5342         list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5343                                  group_entry)
5344                 __perf_event_exit_task(child_event, child_ctx, child);
5345
5346         /*
5347          * If the last event was a group event, it will have appended all
5348          * its siblings to the list, but we obtained 'tmp' before that which
5349          * will still point to the list head terminating the iteration.
5350          */
5351         if (!list_empty(&child_ctx->pinned_groups) ||
5352             !list_empty(&child_ctx->flexible_groups))
5353                 goto again;
5354
5355         mutex_unlock(&child_ctx->mutex);
5356
5357         put_ctx(child_ctx);
5358 }
5359
5360 static void perf_free_event(struct perf_event *event,
5361                             struct perf_event_context *ctx)
5362 {
5363         struct perf_event *parent = event->parent;
5364
5365         if (WARN_ON_ONCE(!parent))
5366                 return;
5367
5368         mutex_lock(&parent->child_mutex);
5369         list_del_init(&event->child_list);
5370         mutex_unlock(&parent->child_mutex);
5371
5372         fput(parent->filp);
5373
5374         list_del_event(event, ctx);
5375         free_event(event);
5376 }
5377
5378 /*
5379  * free an unexposed, unused context as created by inheritance by
5380  * init_task below, used by fork() in case of fail.
5381  */
5382 void perf_event_free_task(struct task_struct *task)
5383 {
5384         struct perf_event_context *ctx = task->perf_event_ctxp;
5385         struct perf_event *event, *tmp;
5386
5387         if (!ctx)
5388                 return;
5389
5390         mutex_lock(&ctx->mutex);
5391 again:
5392         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5393                 perf_free_event(event, ctx);
5394
5395         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5396                                  group_entry)
5397                 perf_free_event(event, ctx);
5398
5399         if (!list_empty(&ctx->pinned_groups) ||
5400             !list_empty(&ctx->flexible_groups))
5401                 goto again;
5402
5403         mutex_unlock(&ctx->mutex);
5404
5405         put_ctx(ctx);
5406 }
5407
5408 static int
5409 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5410                    struct perf_event_context *parent_ctx,
5411                    struct task_struct *child,
5412                    int *inherited_all)
5413 {
5414         int ret;
5415         struct perf_event_context *child_ctx = child->perf_event_ctxp;
5416
5417         if (!event->attr.inherit) {
5418                 *inherited_all = 0;
5419                 return 0;
5420         }
5421
5422         if (!child_ctx) {
5423                 /*
5424                  * This is executed from the parent task context, so
5425                  * inherit events that have been marked for cloning.
5426                  * First allocate and initialize a context for the
5427                  * child.
5428                  */
5429
5430                 child_ctx = kzalloc(sizeof(struct perf_event_context),
5431                                     GFP_KERNEL);
5432                 if (!child_ctx)
5433                         return -ENOMEM;
5434
5435                 __perf_event_init_context(child_ctx, child);
5436                 child->perf_event_ctxp = child_ctx;
5437                 get_task_struct(child);
5438         }
5439
5440         ret = inherit_group(event, parent, parent_ctx,
5441                             child, child_ctx);
5442
5443         if (ret)
5444                 *inherited_all = 0;
5445
5446         return ret;
5447 }
5448
5449
5450 /*
5451  * Initialize the perf_event context in task_struct
5452  */
5453 int perf_event_init_task(struct task_struct *child)
5454 {
5455         struct perf_event_context *child_ctx, *parent_ctx;
5456         struct perf_event_context *cloned_ctx;
5457         struct perf_event *event;
5458         struct task_struct *parent = current;
5459         int inherited_all = 1;
5460         int ret = 0;
5461
5462         child->perf_event_ctxp = NULL;
5463
5464         mutex_init(&child->perf_event_mutex);
5465         INIT_LIST_HEAD(&child->perf_event_list);
5466
5467         if (likely(!parent->perf_event_ctxp))
5468                 return 0;
5469
5470         /*
5471          * If the parent's context is a clone, pin it so it won't get
5472          * swapped under us.
5473          */
5474         parent_ctx = perf_pin_task_context(parent);
5475
5476         /*
5477          * No need to check if parent_ctx != NULL here; since we saw
5478          * it non-NULL earlier, the only reason for it to become NULL
5479          * is if we exit, and since we're currently in the middle of
5480          * a fork we can't be exiting at the same time.
5481          */
5482
5483         /*
5484          * Lock the parent list. No need to lock the child - not PID
5485          * hashed yet and not running, so nobody can access it.
5486          */
5487         mutex_lock(&parent_ctx->mutex);
5488
5489         /*
5490          * We dont have to disable NMIs - we are only looking at
5491          * the list, not manipulating it:
5492          */
5493         list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5494                 ret = inherit_task_group(event, parent, parent_ctx, child,
5495                                          &inherited_all);
5496                 if (ret)
5497                         break;
5498         }
5499
5500         list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5501                 ret = inherit_task_group(event, parent, parent_ctx, child,
5502                                          &inherited_all);
5503                 if (ret)
5504                         break;
5505         }
5506
5507         child_ctx = child->perf_event_ctxp;
5508
5509         if (child_ctx && inherited_all) {
5510                 /*
5511                  * Mark the child context as a clone of the parent
5512                  * context, or of whatever the parent is a clone of.
5513                  * Note that if the parent is a clone, it could get
5514                  * uncloned at any point, but that doesn't matter
5515                  * because the list of events and the generation
5516                  * count can't have changed since we took the mutex.
5517                  */
5518                 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5519                 if (cloned_ctx) {
5520                         child_ctx->parent_ctx = cloned_ctx;
5521                         child_ctx->parent_gen = parent_ctx->parent_gen;
5522                 } else {
5523                         child_ctx->parent_ctx = parent_ctx;
5524                         child_ctx->parent_gen = parent_ctx->generation;
5525                 }
5526                 get_ctx(child_ctx->parent_ctx);
5527         }
5528
5529         mutex_unlock(&parent_ctx->mutex);
5530
5531         perf_unpin_context(parent_ctx);
5532
5533         return ret;
5534 }
5535
5536 static void __init perf_event_init_all_cpus(void)
5537 {
5538         int cpu;
5539         struct perf_cpu_context *cpuctx;
5540
5541         for_each_possible_cpu(cpu) {
5542                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5543                 mutex_init(&cpuctx->hlist_mutex);
5544                 __perf_event_init_context(&cpuctx->ctx, NULL);
5545         }
5546 }
5547
5548 static void __cpuinit perf_event_init_cpu(int cpu)
5549 {
5550         struct perf_cpu_context *cpuctx;
5551
5552         cpuctx = &per_cpu(perf_cpu_context, cpu);
5553
5554         spin_lock(&perf_resource_lock);
5555         cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5556         spin_unlock(&perf_resource_lock);
5557
5558         mutex_lock(&cpuctx->hlist_mutex);
5559         if (cpuctx->hlist_refcount > 0) {
5560                 struct swevent_hlist *hlist;
5561
5562                 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5563                 WARN_ON_ONCE(!hlist);
5564                 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
5565         }
5566         mutex_unlock(&cpuctx->hlist_mutex);
5567 }
5568
5569 #ifdef CONFIG_HOTPLUG_CPU
5570 static void __perf_event_exit_cpu(void *info)
5571 {
5572         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5573         struct perf_event_context *ctx = &cpuctx->ctx;
5574         struct perf_event *event, *tmp;
5575
5576         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5577                 __perf_event_remove_from_context(event);
5578         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5579                 __perf_event_remove_from_context(event);
5580 }
5581 static void perf_event_exit_cpu(int cpu)
5582 {
5583         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5584         struct perf_event_context *ctx = &cpuctx->ctx;
5585
5586         mutex_lock(&cpuctx->hlist_mutex);
5587         swevent_hlist_release(cpuctx);
5588         mutex_unlock(&cpuctx->hlist_mutex);
5589
5590         mutex_lock(&ctx->mutex);
5591         smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5592         mutex_unlock(&ctx->mutex);
5593 }
5594 #else
5595 static inline void perf_event_exit_cpu(int cpu) { }
5596 #endif
5597
5598 static int __cpuinit
5599 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5600 {
5601         unsigned int cpu = (long)hcpu;
5602
5603         switch (action) {
5604
5605         case CPU_UP_PREPARE:
5606         case CPU_UP_PREPARE_FROZEN:
5607                 perf_event_init_cpu(cpu);
5608                 break;
5609
5610         case CPU_DOWN_PREPARE:
5611         case CPU_DOWN_PREPARE_FROZEN:
5612                 perf_event_exit_cpu(cpu);
5613                 break;
5614
5615         default:
5616                 break;
5617         }
5618
5619         return NOTIFY_OK;
5620 }
5621
5622 /*
5623  * This has to have a higher priority than migration_notifier in sched.c.
5624  */
5625 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5626         .notifier_call          = perf_cpu_notify,
5627         .priority               = 20,
5628 };
5629
5630 void __init perf_event_init(void)
5631 {
5632         perf_event_init_all_cpus();
5633         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5634                         (void *)(long)smp_processor_id());
5635         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5636                         (void *)(long)smp_processor_id());
5637         register_cpu_notifier(&perf_cpu_nb);
5638 }
5639
5640 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class,
5641                                         struct sysdev_class_attribute *attr,
5642                                         char *buf)
5643 {
5644         return sprintf(buf, "%d\n", perf_reserved_percpu);
5645 }
5646
5647 static ssize_t
5648 perf_set_reserve_percpu(struct sysdev_class *class,
5649                         struct sysdev_class_attribute *attr,
5650                         const char *buf,
5651                         size_t count)
5652 {
5653         struct perf_cpu_context *cpuctx;
5654         unsigned long val;
5655         int err, cpu, mpt;
5656
5657         err = strict_strtoul(buf, 10, &val);
5658         if (err)
5659                 return err;
5660         if (val > perf_max_events)
5661                 return -EINVAL;
5662
5663         spin_lock(&perf_resource_lock);
5664         perf_reserved_percpu = val;
5665         for_each_online_cpu(cpu) {
5666                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5667                 raw_spin_lock_irq(&cpuctx->ctx.lock);
5668                 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5669                           perf_max_events - perf_reserved_percpu);
5670                 cpuctx->max_pertask = mpt;
5671                 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5672         }
5673         spin_unlock(&perf_resource_lock);
5674
5675         return count;
5676 }
5677
5678 static ssize_t perf_show_overcommit(struct sysdev_class *class,
5679                                     struct sysdev_class_attribute *attr,
5680                                     char *buf)
5681 {
5682         return sprintf(buf, "%d\n", perf_overcommit);
5683 }
5684
5685 static ssize_t
5686 perf_set_overcommit(struct sysdev_class *class,
5687                     struct sysdev_class_attribute *attr,
5688                     const char *buf, size_t count)
5689 {
5690         unsigned long val;
5691         int err;
5692
5693         err = strict_strtoul(buf, 10, &val);
5694         if (err)
5695                 return err;
5696         if (val > 1)
5697                 return -EINVAL;
5698
5699         spin_lock(&perf_resource_lock);
5700         perf_overcommit = val;
5701         spin_unlock(&perf_resource_lock);
5702
5703         return count;
5704 }
5705
5706 static SYSDEV_CLASS_ATTR(
5707                                 reserve_percpu,
5708                                 0644,
5709                                 perf_show_reserve_percpu,
5710                                 perf_set_reserve_percpu
5711                         );
5712
5713 static SYSDEV_CLASS_ATTR(
5714                                 overcommit,
5715                                 0644,
5716                                 perf_show_overcommit,
5717                                 perf_set_overcommit
5718                         );
5719
5720 static struct attribute *perfclass_attrs[] = {
5721         &attr_reserve_percpu.attr,
5722         &attr_overcommit.attr,
5723         NULL
5724 };
5725
5726 static struct attribute_group perfclass_attr_group = {
5727         .attrs                  = perfclass_attrs,
5728         .name                   = "perf_events",
5729 };
5730
5731 static int __init perf_event_sysfs_init(void)
5732 {
5733         return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5734                                   &perfclass_attr_group);
5735 }
5736 device_initcall(perf_event_sysfs_init);