perf, trace: Optimize tracepoints by using per-tracepoint-per-cpu hlist to track...
[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 void *perf_mmap_alloc_page(int cpu)
2324 {
2325         struct page *page;
2326         int node;
2327
2328         node = (cpu == -1) ? cpu : cpu_to_node(cpu);
2329         page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
2330         if (!page)
2331                 return NULL;
2332
2333         return page_address(page);
2334 }
2335
2336 static struct perf_mmap_data *
2337 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2338 {
2339         struct perf_mmap_data *data;
2340         unsigned long size;
2341         int i;
2342
2343         WARN_ON(atomic_read(&event->mmap_count));
2344
2345         size = sizeof(struct perf_mmap_data);
2346         size += nr_pages * sizeof(void *);
2347
2348         data = kzalloc(size, GFP_KERNEL);
2349         if (!data)
2350                 goto fail;
2351
2352         data->user_page = perf_mmap_alloc_page(event->cpu);
2353         if (!data->user_page)
2354                 goto fail_user_page;
2355
2356         for (i = 0; i < nr_pages; i++) {
2357                 data->data_pages[i] = perf_mmap_alloc_page(event->cpu);
2358                 if (!data->data_pages[i])
2359                         goto fail_data_pages;
2360         }
2361
2362         data->data_order = 0;
2363         data->nr_pages = nr_pages;
2364
2365         return data;
2366
2367 fail_data_pages:
2368         for (i--; i >= 0; i--)
2369                 free_page((unsigned long)data->data_pages[i]);
2370
2371         free_page((unsigned long)data->user_page);
2372
2373 fail_user_page:
2374         kfree(data);
2375
2376 fail:
2377         return NULL;
2378 }
2379
2380 static void perf_mmap_free_page(unsigned long addr)
2381 {
2382         struct page *page = virt_to_page((void *)addr);
2383
2384         page->mapping = NULL;
2385         __free_page(page);
2386 }
2387
2388 static void perf_mmap_data_free(struct perf_mmap_data *data)
2389 {
2390         int i;
2391
2392         perf_mmap_free_page((unsigned long)data->user_page);
2393         for (i = 0; i < data->nr_pages; i++)
2394                 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2395         kfree(data);
2396 }
2397
2398 #else
2399
2400 /*
2401  * Back perf_mmap() with vmalloc memory.
2402  *
2403  * Required for architectures that have d-cache aliasing issues.
2404  */
2405
2406 static struct page *
2407 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2408 {
2409         if (pgoff > (1UL << data->data_order))
2410                 return NULL;
2411
2412         return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2413 }
2414
2415 static void perf_mmap_unmark_page(void *addr)
2416 {
2417         struct page *page = vmalloc_to_page(addr);
2418
2419         page->mapping = NULL;
2420 }
2421
2422 static void perf_mmap_data_free_work(struct work_struct *work)
2423 {
2424         struct perf_mmap_data *data;
2425         void *base;
2426         int i, nr;
2427
2428         data = container_of(work, struct perf_mmap_data, work);
2429         nr = 1 << data->data_order;
2430
2431         base = data->user_page;
2432         for (i = 0; i < nr + 1; i++)
2433                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2434
2435         vfree(base);
2436         kfree(data);
2437 }
2438
2439 static void perf_mmap_data_free(struct perf_mmap_data *data)
2440 {
2441         schedule_work(&data->work);
2442 }
2443
2444 static struct perf_mmap_data *
2445 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2446 {
2447         struct perf_mmap_data *data;
2448         unsigned long size;
2449         void *all_buf;
2450
2451         WARN_ON(atomic_read(&event->mmap_count));
2452
2453         size = sizeof(struct perf_mmap_data);
2454         size += sizeof(void *);
2455
2456         data = kzalloc(size, GFP_KERNEL);
2457         if (!data)
2458                 goto fail;
2459
2460         INIT_WORK(&data->work, perf_mmap_data_free_work);
2461
2462         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2463         if (!all_buf)
2464                 goto fail_all_buf;
2465
2466         data->user_page = all_buf;
2467         data->data_pages[0] = all_buf + PAGE_SIZE;
2468         data->data_order = ilog2(nr_pages);
2469         data->nr_pages = 1;
2470
2471         return data;
2472
2473 fail_all_buf:
2474         kfree(data);
2475
2476 fail:
2477         return NULL;
2478 }
2479
2480 #endif
2481
2482 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2483 {
2484         struct perf_event *event = vma->vm_file->private_data;
2485         struct perf_mmap_data *data;
2486         int ret = VM_FAULT_SIGBUS;
2487
2488         if (vmf->flags & FAULT_FLAG_MKWRITE) {
2489                 if (vmf->pgoff == 0)
2490                         ret = 0;
2491                 return ret;
2492         }
2493
2494         rcu_read_lock();
2495         data = rcu_dereference(event->data);
2496         if (!data)
2497                 goto unlock;
2498
2499         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2500                 goto unlock;
2501
2502         vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2503         if (!vmf->page)
2504                 goto unlock;
2505
2506         get_page(vmf->page);
2507         vmf->page->mapping = vma->vm_file->f_mapping;
2508         vmf->page->index   = vmf->pgoff;
2509
2510         ret = 0;
2511 unlock:
2512         rcu_read_unlock();
2513
2514         return ret;
2515 }
2516
2517 static void
2518 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2519 {
2520         long max_size = perf_data_size(data);
2521
2522         if (event->attr.watermark) {
2523                 data->watermark = min_t(long, max_size,
2524                                         event->attr.wakeup_watermark);
2525         }
2526
2527         if (!data->watermark)
2528                 data->watermark = max_size / 2;
2529
2530
2531         rcu_assign_pointer(event->data, data);
2532 }
2533
2534 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2535 {
2536         struct perf_mmap_data *data;
2537
2538         data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2539         perf_mmap_data_free(data);
2540 }
2541
2542 static void perf_mmap_data_release(struct perf_event *event)
2543 {
2544         struct perf_mmap_data *data = event->data;
2545
2546         WARN_ON(atomic_read(&event->mmap_count));
2547
2548         rcu_assign_pointer(event->data, NULL);
2549         call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2550 }
2551
2552 static void perf_mmap_open(struct vm_area_struct *vma)
2553 {
2554         struct perf_event *event = vma->vm_file->private_data;
2555
2556         atomic_inc(&event->mmap_count);
2557 }
2558
2559 static void perf_mmap_close(struct vm_area_struct *vma)
2560 {
2561         struct perf_event *event = vma->vm_file->private_data;
2562
2563         WARN_ON_ONCE(event->ctx->parent_ctx);
2564         if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2565                 unsigned long size = perf_data_size(event->data);
2566                 struct user_struct *user = current_user();
2567
2568                 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2569                 vma->vm_mm->locked_vm -= event->data->nr_locked;
2570                 perf_mmap_data_release(event);
2571                 mutex_unlock(&event->mmap_mutex);
2572         }
2573 }
2574
2575 static const struct vm_operations_struct perf_mmap_vmops = {
2576         .open           = perf_mmap_open,
2577         .close          = perf_mmap_close,
2578         .fault          = perf_mmap_fault,
2579         .page_mkwrite   = perf_mmap_fault,
2580 };
2581
2582 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2583 {
2584         struct perf_event *event = file->private_data;
2585         unsigned long user_locked, user_lock_limit;
2586         struct user_struct *user = current_user();
2587         unsigned long locked, lock_limit;
2588         struct perf_mmap_data *data;
2589         unsigned long vma_size;
2590         unsigned long nr_pages;
2591         long user_extra, extra;
2592         int ret = 0;
2593
2594         /*
2595          * Don't allow mmap() of inherited per-task counters. This would
2596          * create a performance issue due to all children writing to the
2597          * same buffer.
2598          */
2599         if (event->cpu == -1 && event->attr.inherit)
2600                 return -EINVAL;
2601
2602         if (!(vma->vm_flags & VM_SHARED))
2603                 return -EINVAL;
2604
2605         vma_size = vma->vm_end - vma->vm_start;
2606         nr_pages = (vma_size / PAGE_SIZE) - 1;
2607
2608         /*
2609          * If we have data pages ensure they're a power-of-two number, so we
2610          * can do bitmasks instead of modulo.
2611          */
2612         if (nr_pages != 0 && !is_power_of_2(nr_pages))
2613                 return -EINVAL;
2614
2615         if (vma_size != PAGE_SIZE * (1 + nr_pages))
2616                 return -EINVAL;
2617
2618         if (vma->vm_pgoff != 0)
2619                 return -EINVAL;
2620
2621         WARN_ON_ONCE(event->ctx->parent_ctx);
2622         mutex_lock(&event->mmap_mutex);
2623         if (event->output) {
2624                 ret = -EINVAL;
2625                 goto unlock;
2626         }
2627
2628         if (atomic_inc_not_zero(&event->mmap_count)) {
2629                 if (nr_pages != event->data->nr_pages)
2630                         ret = -EINVAL;
2631                 goto unlock;
2632         }
2633
2634         user_extra = nr_pages + 1;
2635         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2636
2637         /*
2638          * Increase the limit linearly with more CPUs:
2639          */
2640         user_lock_limit *= num_online_cpus();
2641
2642         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2643
2644         extra = 0;
2645         if (user_locked > user_lock_limit)
2646                 extra = user_locked - user_lock_limit;
2647
2648         lock_limit = rlimit(RLIMIT_MEMLOCK);
2649         lock_limit >>= PAGE_SHIFT;
2650         locked = vma->vm_mm->locked_vm + extra;
2651
2652         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2653                 !capable(CAP_IPC_LOCK)) {
2654                 ret = -EPERM;
2655                 goto unlock;
2656         }
2657
2658         WARN_ON(event->data);
2659
2660         data = perf_mmap_data_alloc(event, nr_pages);
2661         ret = -ENOMEM;
2662         if (!data)
2663                 goto unlock;
2664
2665         ret = 0;
2666         perf_mmap_data_init(event, data);
2667
2668         atomic_set(&event->mmap_count, 1);
2669         atomic_long_add(user_extra, &user->locked_vm);
2670         vma->vm_mm->locked_vm += extra;
2671         event->data->nr_locked = extra;
2672         if (vma->vm_flags & VM_WRITE)
2673                 event->data->writable = 1;
2674
2675 unlock:
2676         mutex_unlock(&event->mmap_mutex);
2677
2678         vma->vm_flags |= VM_RESERVED;
2679         vma->vm_ops = &perf_mmap_vmops;
2680
2681         return ret;
2682 }
2683
2684 static int perf_fasync(int fd, struct file *filp, int on)
2685 {
2686         struct inode *inode = filp->f_path.dentry->d_inode;
2687         struct perf_event *event = filp->private_data;
2688         int retval;
2689
2690         mutex_lock(&inode->i_mutex);
2691         retval = fasync_helper(fd, filp, on, &event->fasync);
2692         mutex_unlock(&inode->i_mutex);
2693
2694         if (retval < 0)
2695                 return retval;
2696
2697         return 0;
2698 }
2699
2700 static const struct file_operations perf_fops = {
2701         .llseek                 = no_llseek,
2702         .release                = perf_release,
2703         .read                   = perf_read,
2704         .poll                   = perf_poll,
2705         .unlocked_ioctl         = perf_ioctl,
2706         .compat_ioctl           = perf_ioctl,
2707         .mmap                   = perf_mmap,
2708         .fasync                 = perf_fasync,
2709 };
2710
2711 /*
2712  * Perf event wakeup
2713  *
2714  * If there's data, ensure we set the poll() state and publish everything
2715  * to user-space before waking everybody up.
2716  */
2717
2718 void perf_event_wakeup(struct perf_event *event)
2719 {
2720         wake_up_all(&event->waitq);
2721
2722         if (event->pending_kill) {
2723                 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2724                 event->pending_kill = 0;
2725         }
2726 }
2727
2728 /*
2729  * Pending wakeups
2730  *
2731  * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2732  *
2733  * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2734  * single linked list and use cmpxchg() to add entries lockless.
2735  */
2736
2737 static void perf_pending_event(struct perf_pending_entry *entry)
2738 {
2739         struct perf_event *event = container_of(entry,
2740                         struct perf_event, pending);
2741
2742         if (event->pending_disable) {
2743                 event->pending_disable = 0;
2744                 __perf_event_disable(event);
2745         }
2746
2747         if (event->pending_wakeup) {
2748                 event->pending_wakeup = 0;
2749                 perf_event_wakeup(event);
2750         }
2751 }
2752
2753 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2754
2755 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2756         PENDING_TAIL,
2757 };
2758
2759 static void perf_pending_queue(struct perf_pending_entry *entry,
2760                                void (*func)(struct perf_pending_entry *))
2761 {
2762         struct perf_pending_entry **head;
2763
2764         if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2765                 return;
2766
2767         entry->func = func;
2768
2769         head = &get_cpu_var(perf_pending_head);
2770
2771         do {
2772                 entry->next = *head;
2773         } while (cmpxchg(head, entry->next, entry) != entry->next);
2774
2775         set_perf_event_pending();
2776
2777         put_cpu_var(perf_pending_head);
2778 }
2779
2780 static int __perf_pending_run(void)
2781 {
2782         struct perf_pending_entry *list;
2783         int nr = 0;
2784
2785         list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2786         while (list != PENDING_TAIL) {
2787                 void (*func)(struct perf_pending_entry *);
2788                 struct perf_pending_entry *entry = list;
2789
2790                 list = list->next;
2791
2792                 func = entry->func;
2793                 entry->next = NULL;
2794                 /*
2795                  * Ensure we observe the unqueue before we issue the wakeup,
2796                  * so that we won't be waiting forever.
2797                  * -- see perf_not_pending().
2798                  */
2799                 smp_wmb();
2800
2801                 func(entry);
2802                 nr++;
2803         }
2804
2805         return nr;
2806 }
2807
2808 static inline int perf_not_pending(struct perf_event *event)
2809 {
2810         /*
2811          * If we flush on whatever cpu we run, there is a chance we don't
2812          * need to wait.
2813          */
2814         get_cpu();
2815         __perf_pending_run();
2816         put_cpu();
2817
2818         /*
2819          * Ensure we see the proper queue state before going to sleep
2820          * so that we do not miss the wakeup. -- see perf_pending_handle()
2821          */
2822         smp_rmb();
2823         return event->pending.next == NULL;
2824 }
2825
2826 static void perf_pending_sync(struct perf_event *event)
2827 {
2828         wait_event(event->waitq, perf_not_pending(event));
2829 }
2830
2831 void perf_event_do_pending(void)
2832 {
2833         __perf_pending_run();
2834 }
2835
2836 /*
2837  * Callchain support -- arch specific
2838  */
2839
2840 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2841 {
2842         return NULL;
2843 }
2844
2845 __weak
2846 void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip, int skip)
2847 {
2848 }
2849
2850
2851 /*
2852  * We assume there is only KVM supporting the callbacks.
2853  * Later on, we might change it to a list if there is
2854  * another virtualization implementation supporting the callbacks.
2855  */
2856 struct perf_guest_info_callbacks *perf_guest_cbs;
2857
2858 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2859 {
2860         perf_guest_cbs = cbs;
2861         return 0;
2862 }
2863 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
2864
2865 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
2866 {
2867         perf_guest_cbs = NULL;
2868         return 0;
2869 }
2870 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
2871
2872 /*
2873  * Output
2874  */
2875 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2876                               unsigned long offset, unsigned long head)
2877 {
2878         unsigned long mask;
2879
2880         if (!data->writable)
2881                 return true;
2882
2883         mask = perf_data_size(data) - 1;
2884
2885         offset = (offset - tail) & mask;
2886         head   = (head   - tail) & mask;
2887
2888         if ((int)(head - offset) < 0)
2889                 return false;
2890
2891         return true;
2892 }
2893
2894 static void perf_output_wakeup(struct perf_output_handle *handle)
2895 {
2896         atomic_set(&handle->data->poll, POLL_IN);
2897
2898         if (handle->nmi) {
2899                 handle->event->pending_wakeup = 1;
2900                 perf_pending_queue(&handle->event->pending,
2901                                    perf_pending_event);
2902         } else
2903                 perf_event_wakeup(handle->event);
2904 }
2905
2906 /*
2907  * We need to ensure a later event_id doesn't publish a head when a former
2908  * event isn't done writing. However since we need to deal with NMIs we
2909  * cannot fully serialize things.
2910  *
2911  * We only publish the head (and generate a wakeup) when the outer-most
2912  * event completes.
2913  */
2914 static void perf_output_get_handle(struct perf_output_handle *handle)
2915 {
2916         struct perf_mmap_data *data = handle->data;
2917
2918         preempt_disable();
2919         local_inc(&data->nest);
2920         handle->wakeup = local_read(&data->wakeup);
2921 }
2922
2923 static void perf_output_put_handle(struct perf_output_handle *handle)
2924 {
2925         struct perf_mmap_data *data = handle->data;
2926         unsigned long head;
2927
2928 again:
2929         head = local_read(&data->head);
2930
2931         /*
2932          * IRQ/NMI can happen here, which means we can miss a head update.
2933          */
2934
2935         if (!local_dec_and_test(&data->nest))
2936                 goto out;
2937
2938         /*
2939          * Publish the known good head. Rely on the full barrier implied
2940          * by atomic_dec_and_test() order the data->head read and this
2941          * write.
2942          */
2943         data->user_page->data_head = head;
2944
2945         /*
2946          * Now check if we missed an update, rely on the (compiler)
2947          * barrier in atomic_dec_and_test() to re-read data->head.
2948          */
2949         if (unlikely(head != local_read(&data->head))) {
2950                 local_inc(&data->nest);
2951                 goto again;
2952         }
2953
2954         if (handle->wakeup != local_read(&data->wakeup))
2955                 perf_output_wakeup(handle);
2956
2957  out:
2958         preempt_enable();
2959 }
2960
2961 void perf_output_copy(struct perf_output_handle *handle,
2962                       const void *buf, unsigned int len)
2963 {
2964         unsigned int pages_mask;
2965         unsigned long offset;
2966         unsigned int size;
2967         void **pages;
2968
2969         offset          = handle->offset;
2970         pages_mask      = handle->data->nr_pages - 1;
2971         pages           = handle->data->data_pages;
2972
2973         do {
2974                 unsigned long page_offset;
2975                 unsigned long page_size;
2976                 int nr;
2977
2978                 nr          = (offset >> PAGE_SHIFT) & pages_mask;
2979                 page_size   = 1UL << (handle->data->data_order + PAGE_SHIFT);
2980                 page_offset = offset & (page_size - 1);
2981                 size        = min_t(unsigned int, page_size - page_offset, len);
2982
2983                 memcpy(pages[nr] + page_offset, buf, size);
2984
2985                 len         -= size;
2986                 buf         += size;
2987                 offset      += size;
2988         } while (len);
2989
2990         handle->offset = offset;
2991
2992         /*
2993          * Check we didn't copy past our reservation window, taking the
2994          * possible unsigned int wrap into account.
2995          */
2996         WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2997 }
2998
2999 int perf_output_begin(struct perf_output_handle *handle,
3000                       struct perf_event *event, unsigned int size,
3001                       int nmi, int sample)
3002 {
3003         struct perf_event *output_event;
3004         struct perf_mmap_data *data;
3005         unsigned long tail, offset, head;
3006         int have_lost;
3007         struct {
3008                 struct perf_event_header header;
3009                 u64                      id;
3010                 u64                      lost;
3011         } lost_event;
3012
3013         rcu_read_lock();
3014         /*
3015          * For inherited events we send all the output towards the parent.
3016          */
3017         if (event->parent)
3018                 event = event->parent;
3019
3020         output_event = rcu_dereference(event->output);
3021         if (output_event)
3022                 event = output_event;
3023
3024         data = rcu_dereference(event->data);
3025         if (!data)
3026                 goto out;
3027
3028         handle->data    = data;
3029         handle->event   = event;
3030         handle->nmi     = nmi;
3031         handle->sample  = sample;
3032
3033         if (!data->nr_pages)
3034                 goto out;
3035
3036         have_lost = local_read(&data->lost);
3037         if (have_lost)
3038                 size += sizeof(lost_event);
3039
3040         perf_output_get_handle(handle);
3041
3042         do {
3043                 /*
3044                  * Userspace could choose to issue a mb() before updating the
3045                  * tail pointer. So that all reads will be completed before the
3046                  * write is issued.
3047                  */
3048                 tail = ACCESS_ONCE(data->user_page->data_tail);
3049                 smp_rmb();
3050                 offset = head = local_read(&data->head);
3051                 head += size;
3052                 if (unlikely(!perf_output_space(data, tail, offset, head)))
3053                         goto fail;
3054         } while (local_cmpxchg(&data->head, offset, head) != offset);
3055
3056         handle->offset  = offset;
3057         handle->head    = head;
3058
3059         if (head - tail > data->watermark)
3060                 local_inc(&data->wakeup);
3061
3062         if (have_lost) {
3063                 lost_event.header.type = PERF_RECORD_LOST;
3064                 lost_event.header.misc = 0;
3065                 lost_event.header.size = sizeof(lost_event);
3066                 lost_event.id          = event->id;
3067                 lost_event.lost        = local_xchg(&data->lost, 0);
3068
3069                 perf_output_put(handle, lost_event);
3070         }
3071
3072         return 0;
3073
3074 fail:
3075         local_inc(&data->lost);
3076         perf_output_put_handle(handle);
3077 out:
3078         rcu_read_unlock();
3079
3080         return -ENOSPC;
3081 }
3082
3083 void perf_output_end(struct perf_output_handle *handle)
3084 {
3085         struct perf_event *event = handle->event;
3086         struct perf_mmap_data *data = handle->data;
3087
3088         int wakeup_events = event->attr.wakeup_events;
3089
3090         if (handle->sample && wakeup_events) {
3091                 int events = local_inc_return(&data->events);
3092                 if (events >= wakeup_events) {
3093                         local_sub(wakeup_events, &data->events);
3094                         local_inc(&data->wakeup);
3095                 }
3096         }
3097
3098         perf_output_put_handle(handle);
3099         rcu_read_unlock();
3100 }
3101
3102 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3103 {
3104         /*
3105          * only top level events have the pid namespace they were created in
3106          */
3107         if (event->parent)
3108                 event = event->parent;
3109
3110         return task_tgid_nr_ns(p, event->ns);
3111 }
3112
3113 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3114 {
3115         /*
3116          * only top level events have the pid namespace they were created in
3117          */
3118         if (event->parent)
3119                 event = event->parent;
3120
3121         return task_pid_nr_ns(p, event->ns);
3122 }
3123
3124 static void perf_output_read_one(struct perf_output_handle *handle,
3125                                  struct perf_event *event)
3126 {
3127         u64 read_format = event->attr.read_format;
3128         u64 values[4];
3129         int n = 0;
3130
3131         values[n++] = atomic64_read(&event->count);
3132         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3133                 values[n++] = event->total_time_enabled +
3134                         atomic64_read(&event->child_total_time_enabled);
3135         }
3136         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3137                 values[n++] = event->total_time_running +
3138                         atomic64_read(&event->child_total_time_running);
3139         }
3140         if (read_format & PERF_FORMAT_ID)
3141                 values[n++] = primary_event_id(event);
3142
3143         perf_output_copy(handle, values, n * sizeof(u64));
3144 }
3145
3146 /*
3147  * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3148  */
3149 static void perf_output_read_group(struct perf_output_handle *handle,
3150                             struct perf_event *event)
3151 {
3152         struct perf_event *leader = event->group_leader, *sub;
3153         u64 read_format = event->attr.read_format;
3154         u64 values[5];
3155         int n = 0;
3156
3157         values[n++] = 1 + leader->nr_siblings;
3158
3159         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3160                 values[n++] = leader->total_time_enabled;
3161
3162         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3163                 values[n++] = leader->total_time_running;
3164
3165         if (leader != event)
3166                 leader->pmu->read(leader);
3167
3168         values[n++] = atomic64_read(&leader->count);
3169         if (read_format & PERF_FORMAT_ID)
3170                 values[n++] = primary_event_id(leader);
3171
3172         perf_output_copy(handle, values, n * sizeof(u64));
3173
3174         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3175                 n = 0;
3176
3177                 if (sub != event)
3178                         sub->pmu->read(sub);
3179
3180                 values[n++] = atomic64_read(&sub->count);
3181                 if (read_format & PERF_FORMAT_ID)
3182                         values[n++] = primary_event_id(sub);
3183
3184                 perf_output_copy(handle, values, n * sizeof(u64));
3185         }
3186 }
3187
3188 static void perf_output_read(struct perf_output_handle *handle,
3189                              struct perf_event *event)
3190 {
3191         if (event->attr.read_format & PERF_FORMAT_GROUP)
3192                 perf_output_read_group(handle, event);
3193         else
3194                 perf_output_read_one(handle, event);
3195 }
3196
3197 void perf_output_sample(struct perf_output_handle *handle,
3198                         struct perf_event_header *header,
3199                         struct perf_sample_data *data,
3200                         struct perf_event *event)
3201 {
3202         u64 sample_type = data->type;
3203
3204         perf_output_put(handle, *header);
3205
3206         if (sample_type & PERF_SAMPLE_IP)
3207                 perf_output_put(handle, data->ip);
3208
3209         if (sample_type & PERF_SAMPLE_TID)
3210                 perf_output_put(handle, data->tid_entry);
3211
3212         if (sample_type & PERF_SAMPLE_TIME)
3213                 perf_output_put(handle, data->time);
3214
3215         if (sample_type & PERF_SAMPLE_ADDR)
3216                 perf_output_put(handle, data->addr);
3217
3218         if (sample_type & PERF_SAMPLE_ID)
3219                 perf_output_put(handle, data->id);
3220
3221         if (sample_type & PERF_SAMPLE_STREAM_ID)
3222                 perf_output_put(handle, data->stream_id);
3223
3224         if (sample_type & PERF_SAMPLE_CPU)
3225                 perf_output_put(handle, data->cpu_entry);
3226
3227         if (sample_type & PERF_SAMPLE_PERIOD)
3228                 perf_output_put(handle, data->period);
3229
3230         if (sample_type & PERF_SAMPLE_READ)
3231                 perf_output_read(handle, event);
3232
3233         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3234                 if (data->callchain) {
3235                         int size = 1;
3236
3237                         if (data->callchain)
3238                                 size += data->callchain->nr;
3239
3240                         size *= sizeof(u64);
3241
3242                         perf_output_copy(handle, data->callchain, size);
3243                 } else {
3244                         u64 nr = 0;
3245                         perf_output_put(handle, nr);
3246                 }
3247         }
3248
3249         if (sample_type & PERF_SAMPLE_RAW) {
3250                 if (data->raw) {
3251                         perf_output_put(handle, data->raw->size);
3252                         perf_output_copy(handle, data->raw->data,
3253                                          data->raw->size);
3254                 } else {
3255                         struct {
3256                                 u32     size;
3257                                 u32     data;
3258                         } raw = {
3259                                 .size = sizeof(u32),
3260                                 .data = 0,
3261                         };
3262                         perf_output_put(handle, raw);
3263                 }
3264         }
3265 }
3266
3267 void perf_prepare_sample(struct perf_event_header *header,
3268                          struct perf_sample_data *data,
3269                          struct perf_event *event,
3270                          struct pt_regs *regs)
3271 {
3272         u64 sample_type = event->attr.sample_type;
3273
3274         data->type = sample_type;
3275
3276         header->type = PERF_RECORD_SAMPLE;
3277         header->size = sizeof(*header);
3278
3279         header->misc = 0;
3280         header->misc |= perf_misc_flags(regs);
3281
3282         if (sample_type & PERF_SAMPLE_IP) {
3283                 data->ip = perf_instruction_pointer(regs);
3284
3285                 header->size += sizeof(data->ip);
3286         }
3287
3288         if (sample_type & PERF_SAMPLE_TID) {
3289                 /* namespace issues */
3290                 data->tid_entry.pid = perf_event_pid(event, current);
3291                 data->tid_entry.tid = perf_event_tid(event, current);
3292
3293                 header->size += sizeof(data->tid_entry);
3294         }
3295
3296         if (sample_type & PERF_SAMPLE_TIME) {
3297                 data->time = perf_clock();
3298
3299                 header->size += sizeof(data->time);
3300         }
3301
3302         if (sample_type & PERF_SAMPLE_ADDR)
3303                 header->size += sizeof(data->addr);
3304
3305         if (sample_type & PERF_SAMPLE_ID) {
3306                 data->id = primary_event_id(event);
3307
3308                 header->size += sizeof(data->id);
3309         }
3310
3311         if (sample_type & PERF_SAMPLE_STREAM_ID) {
3312                 data->stream_id = event->id;
3313
3314                 header->size += sizeof(data->stream_id);
3315         }
3316
3317         if (sample_type & PERF_SAMPLE_CPU) {
3318                 data->cpu_entry.cpu             = raw_smp_processor_id();
3319                 data->cpu_entry.reserved        = 0;
3320
3321                 header->size += sizeof(data->cpu_entry);
3322         }
3323
3324         if (sample_type & PERF_SAMPLE_PERIOD)
3325                 header->size += sizeof(data->period);
3326
3327         if (sample_type & PERF_SAMPLE_READ)
3328                 header->size += perf_event_read_size(event);
3329
3330         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3331                 int size = 1;
3332
3333                 data->callchain = perf_callchain(regs);
3334
3335                 if (data->callchain)
3336                         size += data->callchain->nr;
3337
3338                 header->size += size * sizeof(u64);
3339         }
3340
3341         if (sample_type & PERF_SAMPLE_RAW) {
3342                 int size = sizeof(u32);
3343
3344                 if (data->raw)
3345                         size += data->raw->size;
3346                 else
3347                         size += sizeof(u32);
3348
3349                 WARN_ON_ONCE(size & (sizeof(u64)-1));
3350                 header->size += size;
3351         }
3352 }
3353
3354 static void perf_event_output(struct perf_event *event, int nmi,
3355                                 struct perf_sample_data *data,
3356                                 struct pt_regs *regs)
3357 {
3358         struct perf_output_handle handle;
3359         struct perf_event_header header;
3360
3361         perf_prepare_sample(&header, data, event, regs);
3362
3363         if (perf_output_begin(&handle, event, header.size, nmi, 1))
3364                 return;
3365
3366         perf_output_sample(&handle, &header, data, event);
3367
3368         perf_output_end(&handle);
3369 }
3370
3371 /*
3372  * read event_id
3373  */
3374
3375 struct perf_read_event {
3376         struct perf_event_header        header;
3377
3378         u32                             pid;
3379         u32                             tid;
3380 };
3381
3382 static void
3383 perf_event_read_event(struct perf_event *event,
3384                         struct task_struct *task)
3385 {
3386         struct perf_output_handle handle;
3387         struct perf_read_event read_event = {
3388                 .header = {
3389                         .type = PERF_RECORD_READ,
3390                         .misc = 0,
3391                         .size = sizeof(read_event) + perf_event_read_size(event),
3392                 },
3393                 .pid = perf_event_pid(event, task),
3394                 .tid = perf_event_tid(event, task),
3395         };
3396         int ret;
3397
3398         ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3399         if (ret)
3400                 return;
3401
3402         perf_output_put(&handle, read_event);
3403         perf_output_read(&handle, event);
3404
3405         perf_output_end(&handle);
3406 }
3407
3408 /*
3409  * task tracking -- fork/exit
3410  *
3411  * enabled by: attr.comm | attr.mmap | attr.task
3412  */
3413
3414 struct perf_task_event {
3415         struct task_struct              *task;
3416         struct perf_event_context       *task_ctx;
3417
3418         struct {
3419                 struct perf_event_header        header;
3420
3421                 u32                             pid;
3422                 u32                             ppid;
3423                 u32                             tid;
3424                 u32                             ptid;
3425                 u64                             time;
3426         } event_id;
3427 };
3428
3429 static void perf_event_task_output(struct perf_event *event,
3430                                      struct perf_task_event *task_event)
3431 {
3432         struct perf_output_handle handle;
3433         struct task_struct *task = task_event->task;
3434         int size, ret;
3435
3436         size  = task_event->event_id.header.size;
3437         ret = perf_output_begin(&handle, event, size, 0, 0);
3438
3439         if (ret)
3440                 return;
3441
3442         task_event->event_id.pid = perf_event_pid(event, task);
3443         task_event->event_id.ppid = perf_event_pid(event, current);
3444
3445         task_event->event_id.tid = perf_event_tid(event, task);
3446         task_event->event_id.ptid = perf_event_tid(event, current);
3447
3448         perf_output_put(&handle, task_event->event_id);
3449
3450         perf_output_end(&handle);
3451 }
3452
3453 static int perf_event_task_match(struct perf_event *event)
3454 {
3455         if (event->state < PERF_EVENT_STATE_INACTIVE)
3456                 return 0;
3457
3458         if (event->cpu != -1 && event->cpu != smp_processor_id())
3459                 return 0;
3460
3461         if (event->attr.comm || event->attr.mmap || event->attr.task)
3462                 return 1;
3463
3464         return 0;
3465 }
3466
3467 static void perf_event_task_ctx(struct perf_event_context *ctx,
3468                                   struct perf_task_event *task_event)
3469 {
3470         struct perf_event *event;
3471
3472         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3473                 if (perf_event_task_match(event))
3474                         perf_event_task_output(event, task_event);
3475         }
3476 }
3477
3478 static void perf_event_task_event(struct perf_task_event *task_event)
3479 {
3480         struct perf_cpu_context *cpuctx;
3481         struct perf_event_context *ctx = task_event->task_ctx;
3482
3483         rcu_read_lock();
3484         cpuctx = &get_cpu_var(perf_cpu_context);
3485         perf_event_task_ctx(&cpuctx->ctx, task_event);
3486         if (!ctx)
3487                 ctx = rcu_dereference(current->perf_event_ctxp);
3488         if (ctx)
3489                 perf_event_task_ctx(ctx, task_event);
3490         put_cpu_var(perf_cpu_context);
3491         rcu_read_unlock();
3492 }
3493
3494 static void perf_event_task(struct task_struct *task,
3495                               struct perf_event_context *task_ctx,
3496                               int new)
3497 {
3498         struct perf_task_event task_event;
3499
3500         if (!atomic_read(&nr_comm_events) &&
3501             !atomic_read(&nr_mmap_events) &&
3502             !atomic_read(&nr_task_events))
3503                 return;
3504
3505         task_event = (struct perf_task_event){
3506                 .task     = task,
3507                 .task_ctx = task_ctx,
3508                 .event_id    = {
3509                         .header = {
3510                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3511                                 .misc = 0,
3512                                 .size = sizeof(task_event.event_id),
3513                         },
3514                         /* .pid  */
3515                         /* .ppid */
3516                         /* .tid  */
3517                         /* .ptid */
3518                         .time = perf_clock(),
3519                 },
3520         };
3521
3522         perf_event_task_event(&task_event);
3523 }
3524
3525 void perf_event_fork(struct task_struct *task)
3526 {
3527         perf_event_task(task, NULL, 1);
3528 }
3529
3530 /*
3531  * comm tracking
3532  */
3533
3534 struct perf_comm_event {
3535         struct task_struct      *task;
3536         char                    *comm;
3537         int                     comm_size;
3538
3539         struct {
3540                 struct perf_event_header        header;
3541
3542                 u32                             pid;
3543                 u32                             tid;
3544         } event_id;
3545 };
3546
3547 static void perf_event_comm_output(struct perf_event *event,
3548                                      struct perf_comm_event *comm_event)
3549 {
3550         struct perf_output_handle handle;
3551         int size = comm_event->event_id.header.size;
3552         int ret = perf_output_begin(&handle, event, size, 0, 0);
3553
3554         if (ret)
3555                 return;
3556
3557         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3558         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3559
3560         perf_output_put(&handle, comm_event->event_id);
3561         perf_output_copy(&handle, comm_event->comm,
3562                                    comm_event->comm_size);
3563         perf_output_end(&handle);
3564 }
3565
3566 static int perf_event_comm_match(struct perf_event *event)
3567 {
3568         if (event->state < PERF_EVENT_STATE_INACTIVE)
3569                 return 0;
3570
3571         if (event->cpu != -1 && event->cpu != smp_processor_id())
3572                 return 0;
3573
3574         if (event->attr.comm)
3575                 return 1;
3576
3577         return 0;
3578 }
3579
3580 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3581                                   struct perf_comm_event *comm_event)
3582 {
3583         struct perf_event *event;
3584
3585         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3586                 if (perf_event_comm_match(event))
3587                         perf_event_comm_output(event, comm_event);
3588         }
3589 }
3590
3591 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3592 {
3593         struct perf_cpu_context *cpuctx;
3594         struct perf_event_context *ctx;
3595         unsigned int size;
3596         char comm[TASK_COMM_LEN];
3597
3598         memset(comm, 0, sizeof(comm));
3599         strlcpy(comm, comm_event->task->comm, sizeof(comm));
3600         size = ALIGN(strlen(comm)+1, sizeof(u64));
3601
3602         comm_event->comm = comm;
3603         comm_event->comm_size = size;
3604
3605         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3606
3607         rcu_read_lock();
3608         cpuctx = &get_cpu_var(perf_cpu_context);
3609         perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3610         ctx = rcu_dereference(current->perf_event_ctxp);
3611         if (ctx)
3612                 perf_event_comm_ctx(ctx, comm_event);
3613         put_cpu_var(perf_cpu_context);
3614         rcu_read_unlock();
3615 }
3616
3617 void perf_event_comm(struct task_struct *task)
3618 {
3619         struct perf_comm_event comm_event;
3620
3621         if (task->perf_event_ctxp)
3622                 perf_event_enable_on_exec(task);
3623
3624         if (!atomic_read(&nr_comm_events))
3625                 return;
3626
3627         comm_event = (struct perf_comm_event){
3628                 .task   = task,
3629                 /* .comm      */
3630                 /* .comm_size */
3631                 .event_id  = {
3632                         .header = {
3633                                 .type = PERF_RECORD_COMM,
3634                                 .misc = 0,
3635                                 /* .size */
3636                         },
3637                         /* .pid */
3638                         /* .tid */
3639                 },
3640         };
3641
3642         perf_event_comm_event(&comm_event);
3643 }
3644
3645 /*
3646  * mmap tracking
3647  */
3648
3649 struct perf_mmap_event {
3650         struct vm_area_struct   *vma;
3651
3652         const char              *file_name;
3653         int                     file_size;
3654
3655         struct {
3656                 struct perf_event_header        header;
3657
3658                 u32                             pid;
3659                 u32                             tid;
3660                 u64                             start;
3661                 u64                             len;
3662                 u64                             pgoff;
3663         } event_id;
3664 };
3665
3666 static void perf_event_mmap_output(struct perf_event *event,
3667                                      struct perf_mmap_event *mmap_event)
3668 {
3669         struct perf_output_handle handle;
3670         int size = mmap_event->event_id.header.size;
3671         int ret = perf_output_begin(&handle, event, size, 0, 0);
3672
3673         if (ret)
3674                 return;
3675
3676         mmap_event->event_id.pid = perf_event_pid(event, current);
3677         mmap_event->event_id.tid = perf_event_tid(event, current);
3678
3679         perf_output_put(&handle, mmap_event->event_id);
3680         perf_output_copy(&handle, mmap_event->file_name,
3681                                    mmap_event->file_size);
3682         perf_output_end(&handle);
3683 }
3684
3685 static int perf_event_mmap_match(struct perf_event *event,
3686                                    struct perf_mmap_event *mmap_event)
3687 {
3688         if (event->state < PERF_EVENT_STATE_INACTIVE)
3689                 return 0;
3690
3691         if (event->cpu != -1 && event->cpu != smp_processor_id())
3692                 return 0;
3693
3694         if (event->attr.mmap)
3695                 return 1;
3696
3697         return 0;
3698 }
3699
3700 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3701                                   struct perf_mmap_event *mmap_event)
3702 {
3703         struct perf_event *event;
3704
3705         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3706                 if (perf_event_mmap_match(event, mmap_event))
3707                         perf_event_mmap_output(event, mmap_event);
3708         }
3709 }
3710
3711 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3712 {
3713         struct perf_cpu_context *cpuctx;
3714         struct perf_event_context *ctx;
3715         struct vm_area_struct *vma = mmap_event->vma;
3716         struct file *file = vma->vm_file;
3717         unsigned int size;
3718         char tmp[16];
3719         char *buf = NULL;
3720         const char *name;
3721
3722         memset(tmp, 0, sizeof(tmp));
3723
3724         if (file) {
3725                 /*
3726                  * d_path works from the end of the buffer backwards, so we
3727                  * need to add enough zero bytes after the string to handle
3728                  * the 64bit alignment we do later.
3729                  */
3730                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3731                 if (!buf) {
3732                         name = strncpy(tmp, "//enomem", sizeof(tmp));
3733                         goto got_name;
3734                 }
3735                 name = d_path(&file->f_path, buf, PATH_MAX);
3736                 if (IS_ERR(name)) {
3737                         name = strncpy(tmp, "//toolong", sizeof(tmp));
3738                         goto got_name;
3739                 }
3740         } else {
3741                 if (arch_vma_name(mmap_event->vma)) {
3742                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3743                                        sizeof(tmp));
3744                         goto got_name;
3745                 }
3746
3747                 if (!vma->vm_mm) {
3748                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
3749                         goto got_name;
3750                 }
3751
3752                 name = strncpy(tmp, "//anon", sizeof(tmp));
3753                 goto got_name;
3754         }
3755
3756 got_name:
3757         size = ALIGN(strlen(name)+1, sizeof(u64));
3758
3759         mmap_event->file_name = name;
3760         mmap_event->file_size = size;
3761
3762         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3763
3764         rcu_read_lock();
3765         cpuctx = &get_cpu_var(perf_cpu_context);
3766         perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3767         ctx = rcu_dereference(current->perf_event_ctxp);
3768         if (ctx)
3769                 perf_event_mmap_ctx(ctx, mmap_event);
3770         put_cpu_var(perf_cpu_context);
3771         rcu_read_unlock();
3772
3773         kfree(buf);
3774 }
3775
3776 void __perf_event_mmap(struct vm_area_struct *vma)
3777 {
3778         struct perf_mmap_event mmap_event;
3779
3780         if (!atomic_read(&nr_mmap_events))
3781                 return;
3782
3783         mmap_event = (struct perf_mmap_event){
3784                 .vma    = vma,
3785                 /* .file_name */
3786                 /* .file_size */
3787                 .event_id  = {
3788                         .header = {
3789                                 .type = PERF_RECORD_MMAP,
3790                                 .misc = PERF_RECORD_MISC_USER,
3791                                 /* .size */
3792                         },
3793                         /* .pid */
3794                         /* .tid */
3795                         .start  = vma->vm_start,
3796                         .len    = vma->vm_end - vma->vm_start,
3797                         .pgoff  = (u64)vma->vm_pgoff << PAGE_SHIFT,
3798                 },
3799         };
3800
3801         perf_event_mmap_event(&mmap_event);
3802 }
3803
3804 /*
3805  * IRQ throttle logging
3806  */
3807
3808 static void perf_log_throttle(struct perf_event *event, int enable)
3809 {
3810         struct perf_output_handle handle;
3811         int ret;
3812
3813         struct {
3814                 struct perf_event_header        header;
3815                 u64                             time;
3816                 u64                             id;
3817                 u64                             stream_id;
3818         } throttle_event = {
3819                 .header = {
3820                         .type = PERF_RECORD_THROTTLE,
3821                         .misc = 0,
3822                         .size = sizeof(throttle_event),
3823                 },
3824                 .time           = perf_clock(),
3825                 .id             = primary_event_id(event),
3826                 .stream_id      = event->id,
3827         };
3828
3829         if (enable)
3830                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3831
3832         ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3833         if (ret)
3834                 return;
3835
3836         perf_output_put(&handle, throttle_event);
3837         perf_output_end(&handle);
3838 }
3839
3840 /*
3841  * Generic event overflow handling, sampling.
3842  */
3843
3844 static int __perf_event_overflow(struct perf_event *event, int nmi,
3845                                    int throttle, struct perf_sample_data *data,
3846                                    struct pt_regs *regs)
3847 {
3848         int events = atomic_read(&event->event_limit);
3849         struct hw_perf_event *hwc = &event->hw;
3850         int ret = 0;
3851
3852         throttle = (throttle && event->pmu->unthrottle != NULL);
3853
3854         if (!throttle) {
3855                 hwc->interrupts++;
3856         } else {
3857                 if (hwc->interrupts != MAX_INTERRUPTS) {
3858                         hwc->interrupts++;
3859                         if (HZ * hwc->interrupts >
3860                                         (u64)sysctl_perf_event_sample_rate) {
3861                                 hwc->interrupts = MAX_INTERRUPTS;
3862                                 perf_log_throttle(event, 0);
3863                                 ret = 1;
3864                         }
3865                 } else {
3866                         /*
3867                          * Keep re-disabling events even though on the previous
3868                          * pass we disabled it - just in case we raced with a
3869                          * sched-in and the event got enabled again:
3870                          */
3871                         ret = 1;
3872                 }
3873         }
3874
3875         if (event->attr.freq) {
3876                 u64 now = perf_clock();
3877                 s64 delta = now - hwc->freq_time_stamp;
3878
3879                 hwc->freq_time_stamp = now;
3880
3881                 if (delta > 0 && delta < 2*TICK_NSEC)
3882                         perf_adjust_period(event, delta, hwc->last_period);
3883         }
3884
3885         /*
3886          * XXX event_limit might not quite work as expected on inherited
3887          * events
3888          */
3889
3890         event->pending_kill = POLL_IN;
3891         if (events && atomic_dec_and_test(&event->event_limit)) {
3892                 ret = 1;
3893                 event->pending_kill = POLL_HUP;
3894                 if (nmi) {
3895                         event->pending_disable = 1;
3896                         perf_pending_queue(&event->pending,
3897                                            perf_pending_event);
3898                 } else
3899                         perf_event_disable(event);
3900         }
3901
3902         if (event->overflow_handler)
3903                 event->overflow_handler(event, nmi, data, regs);
3904         else
3905                 perf_event_output(event, nmi, data, regs);
3906
3907         return ret;
3908 }
3909
3910 int perf_event_overflow(struct perf_event *event, int nmi,
3911                           struct perf_sample_data *data,
3912                           struct pt_regs *regs)
3913 {
3914         return __perf_event_overflow(event, nmi, 1, data, regs);
3915 }
3916
3917 /*
3918  * Generic software event infrastructure
3919  */
3920
3921 /*
3922  * We directly increment event->count and keep a second value in
3923  * event->hw.period_left to count intervals. This period event
3924  * is kept in the range [-sample_period, 0] so that we can use the
3925  * sign as trigger.
3926  */
3927
3928 static u64 perf_swevent_set_period(struct perf_event *event)
3929 {
3930         struct hw_perf_event *hwc = &event->hw;
3931         u64 period = hwc->last_period;
3932         u64 nr, offset;
3933         s64 old, val;
3934
3935         hwc->last_period = hwc->sample_period;
3936
3937 again:
3938         old = val = atomic64_read(&hwc->period_left);
3939         if (val < 0)
3940                 return 0;
3941
3942         nr = div64_u64(period + val, period);
3943         offset = nr * period;
3944         val -= offset;
3945         if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3946                 goto again;
3947
3948         return nr;
3949 }
3950
3951 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3952                                     int nmi, struct perf_sample_data *data,
3953                                     struct pt_regs *regs)
3954 {
3955         struct hw_perf_event *hwc = &event->hw;
3956         int throttle = 0;
3957
3958         data->period = event->hw.last_period;
3959         if (!overflow)
3960                 overflow = perf_swevent_set_period(event);
3961
3962         if (hwc->interrupts == MAX_INTERRUPTS)
3963                 return;
3964
3965         for (; overflow; overflow--) {
3966                 if (__perf_event_overflow(event, nmi, throttle,
3967                                             data, regs)) {
3968                         /*
3969                          * We inhibit the overflow from happening when
3970                          * hwc->interrupts == MAX_INTERRUPTS.
3971                          */
3972                         break;
3973                 }
3974                 throttle = 1;
3975         }
3976 }
3977
3978 static void perf_swevent_unthrottle(struct perf_event *event)
3979 {
3980         /*
3981          * Nothing to do, we already reset hwc->interrupts.
3982          */
3983 }
3984
3985 static void perf_swevent_add(struct perf_event *event, u64 nr,
3986                                int nmi, struct perf_sample_data *data,
3987                                struct pt_regs *regs)
3988 {
3989         struct hw_perf_event *hwc = &event->hw;
3990
3991         atomic64_add(nr, &event->count);
3992
3993         if (!regs)
3994                 return;
3995
3996         if (!hwc->sample_period)
3997                 return;
3998
3999         if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4000                 return perf_swevent_overflow(event, 1, nmi, data, regs);
4001
4002         if (atomic64_add_negative(nr, &hwc->period_left))
4003                 return;
4004
4005         perf_swevent_overflow(event, 0, nmi, data, regs);
4006 }
4007
4008 static int perf_exclude_event(struct perf_event *event,
4009                               struct pt_regs *regs)
4010 {
4011         if (regs) {
4012                 if (event->attr.exclude_user && user_mode(regs))
4013                         return 1;
4014
4015                 if (event->attr.exclude_kernel && !user_mode(regs))
4016                         return 1;
4017         }
4018
4019         return 0;
4020 }
4021
4022 static int perf_swevent_match(struct perf_event *event,
4023                                 enum perf_type_id type,
4024                                 u32 event_id,
4025                                 struct perf_sample_data *data,
4026                                 struct pt_regs *regs)
4027 {
4028         if (event->attr.type != type)
4029                 return 0;
4030
4031         if (event->attr.config != event_id)
4032                 return 0;
4033
4034         if (perf_exclude_event(event, regs))
4035                 return 0;
4036
4037         return 1;
4038 }
4039
4040 static inline u64 swevent_hash(u64 type, u32 event_id)
4041 {
4042         u64 val = event_id | (type << 32);
4043
4044         return hash_64(val, SWEVENT_HLIST_BITS);
4045 }
4046
4047 static inline struct hlist_head *
4048 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4049 {
4050         u64 hash = swevent_hash(type, event_id);
4051
4052         return &hlist->heads[hash];
4053 }
4054
4055 /* For the read side: events when they trigger */
4056 static inline struct hlist_head *
4057 find_swevent_head_rcu(struct perf_cpu_context *ctx, u64 type, u32 event_id)
4058 {
4059         struct swevent_hlist *hlist;
4060
4061         hlist = rcu_dereference(ctx->swevent_hlist);
4062         if (!hlist)
4063                 return NULL;
4064
4065         return __find_swevent_head(hlist, type, event_id);
4066 }
4067
4068 /* For the event head insertion and removal in the hlist */
4069 static inline struct hlist_head *
4070 find_swevent_head(struct perf_cpu_context *ctx, struct perf_event *event)
4071 {
4072         struct swevent_hlist *hlist;
4073         u32 event_id = event->attr.config;
4074         u64 type = event->attr.type;
4075
4076         /*
4077          * Event scheduling is always serialized against hlist allocation
4078          * and release. Which makes the protected version suitable here.
4079          * The context lock guarantees that.
4080          */
4081         hlist = rcu_dereference_protected(ctx->swevent_hlist,
4082                                           lockdep_is_held(&event->ctx->lock));
4083         if (!hlist)
4084                 return NULL;
4085
4086         return __find_swevent_head(hlist, type, event_id);
4087 }
4088
4089 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4090                                     u64 nr, int nmi,
4091                                     struct perf_sample_data *data,
4092                                     struct pt_regs *regs)
4093 {
4094         struct perf_cpu_context *cpuctx;
4095         struct perf_event *event;
4096         struct hlist_node *node;
4097         struct hlist_head *head;
4098
4099         cpuctx = &__get_cpu_var(perf_cpu_context);
4100
4101         rcu_read_lock();
4102
4103         head = find_swevent_head_rcu(cpuctx, type, event_id);
4104
4105         if (!head)
4106                 goto end;
4107
4108         hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4109                 if (perf_swevent_match(event, type, event_id, data, regs))
4110                         perf_swevent_add(event, nr, nmi, data, regs);
4111         }
4112 end:
4113         rcu_read_unlock();
4114 }
4115
4116 int perf_swevent_get_recursion_context(void)
4117 {
4118         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4119         int rctx;
4120
4121         if (in_nmi())
4122                 rctx = 3;
4123         else if (in_irq())
4124                 rctx = 2;
4125         else if (in_softirq())
4126                 rctx = 1;
4127         else
4128                 rctx = 0;
4129
4130         if (cpuctx->recursion[rctx])
4131                 return -1;
4132
4133         cpuctx->recursion[rctx]++;
4134         barrier();
4135
4136         return rctx;
4137 }
4138 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4139
4140 void perf_swevent_put_recursion_context(int rctx)
4141 {
4142         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4143         barrier();
4144         cpuctx->recursion[rctx]--;
4145 }
4146 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
4147
4148
4149 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4150                             struct pt_regs *regs, u64 addr)
4151 {
4152         struct perf_sample_data data;
4153         int rctx;
4154
4155         preempt_disable_notrace();
4156         rctx = perf_swevent_get_recursion_context();
4157         if (rctx < 0)
4158                 return;
4159
4160         perf_sample_data_init(&data, addr);
4161
4162         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4163
4164         perf_swevent_put_recursion_context(rctx);
4165         preempt_enable_notrace();
4166 }
4167
4168 static void perf_swevent_read(struct perf_event *event)
4169 {
4170 }
4171
4172 static int perf_swevent_enable(struct perf_event *event)
4173 {
4174         struct hw_perf_event *hwc = &event->hw;
4175         struct perf_cpu_context *cpuctx;
4176         struct hlist_head *head;
4177
4178         cpuctx = &__get_cpu_var(perf_cpu_context);
4179
4180         if (hwc->sample_period) {
4181                 hwc->last_period = hwc->sample_period;
4182                 perf_swevent_set_period(event);
4183         }
4184
4185         head = find_swevent_head(cpuctx, event);
4186         if (WARN_ON_ONCE(!head))
4187                 return -EINVAL;
4188
4189         hlist_add_head_rcu(&event->hlist_entry, head);
4190
4191         return 0;
4192 }
4193
4194 static void perf_swevent_disable(struct perf_event *event)
4195 {
4196         hlist_del_rcu(&event->hlist_entry);
4197 }
4198
4199 static const struct pmu perf_ops_generic = {
4200         .enable         = perf_swevent_enable,
4201         .disable        = perf_swevent_disable,
4202         .read           = perf_swevent_read,
4203         .unthrottle     = perf_swevent_unthrottle,
4204 };
4205
4206 /*
4207  * hrtimer based swevent callback
4208  */
4209
4210 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4211 {
4212         enum hrtimer_restart ret = HRTIMER_RESTART;
4213         struct perf_sample_data data;
4214         struct pt_regs *regs;
4215         struct perf_event *event;
4216         u64 period;
4217
4218         event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4219         event->pmu->read(event);
4220
4221         perf_sample_data_init(&data, 0);
4222         data.period = event->hw.last_period;
4223         regs = get_irq_regs();
4224
4225         if (regs && !perf_exclude_event(event, regs)) {
4226                 if (!(event->attr.exclude_idle && current->pid == 0))
4227                         if (perf_event_overflow(event, 0, &data, regs))
4228                                 ret = HRTIMER_NORESTART;
4229         }
4230
4231         period = max_t(u64, 10000, event->hw.sample_period);
4232         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4233
4234         return ret;
4235 }
4236
4237 static void perf_swevent_start_hrtimer(struct perf_event *event)
4238 {
4239         struct hw_perf_event *hwc = &event->hw;
4240
4241         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4242         hwc->hrtimer.function = perf_swevent_hrtimer;
4243         if (hwc->sample_period) {
4244                 u64 period;
4245
4246                 if (hwc->remaining) {
4247                         if (hwc->remaining < 0)
4248                                 period = 10000;
4249                         else
4250                                 period = hwc->remaining;
4251                         hwc->remaining = 0;
4252                 } else {
4253                         period = max_t(u64, 10000, hwc->sample_period);
4254                 }
4255                 __hrtimer_start_range_ns(&hwc->hrtimer,
4256                                 ns_to_ktime(period), 0,
4257                                 HRTIMER_MODE_REL, 0);
4258         }
4259 }
4260
4261 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4262 {
4263         struct hw_perf_event *hwc = &event->hw;
4264
4265         if (hwc->sample_period) {
4266                 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4267                 hwc->remaining = ktime_to_ns(remaining);
4268
4269                 hrtimer_cancel(&hwc->hrtimer);
4270         }
4271 }
4272
4273 /*
4274  * Software event: cpu wall time clock
4275  */
4276
4277 static void cpu_clock_perf_event_update(struct perf_event *event)
4278 {
4279         int cpu = raw_smp_processor_id();
4280         s64 prev;
4281         u64 now;
4282
4283         now = cpu_clock(cpu);
4284         prev = atomic64_xchg(&event->hw.prev_count, now);
4285         atomic64_add(now - prev, &event->count);
4286 }
4287
4288 static int cpu_clock_perf_event_enable(struct perf_event *event)
4289 {
4290         struct hw_perf_event *hwc = &event->hw;
4291         int cpu = raw_smp_processor_id();
4292
4293         atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4294         perf_swevent_start_hrtimer(event);
4295
4296         return 0;
4297 }
4298
4299 static void cpu_clock_perf_event_disable(struct perf_event *event)
4300 {
4301         perf_swevent_cancel_hrtimer(event);
4302         cpu_clock_perf_event_update(event);
4303 }
4304
4305 static void cpu_clock_perf_event_read(struct perf_event *event)
4306 {
4307         cpu_clock_perf_event_update(event);
4308 }
4309
4310 static const struct pmu perf_ops_cpu_clock = {
4311         .enable         = cpu_clock_perf_event_enable,
4312         .disable        = cpu_clock_perf_event_disable,
4313         .read           = cpu_clock_perf_event_read,
4314 };
4315
4316 /*
4317  * Software event: task time clock
4318  */
4319
4320 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4321 {
4322         u64 prev;
4323         s64 delta;
4324
4325         prev = atomic64_xchg(&event->hw.prev_count, now);
4326         delta = now - prev;
4327         atomic64_add(delta, &event->count);
4328 }
4329
4330 static int task_clock_perf_event_enable(struct perf_event *event)
4331 {
4332         struct hw_perf_event *hwc = &event->hw;
4333         u64 now;
4334
4335         now = event->ctx->time;
4336
4337         atomic64_set(&hwc->prev_count, now);
4338
4339         perf_swevent_start_hrtimer(event);
4340
4341         return 0;
4342 }
4343
4344 static void task_clock_perf_event_disable(struct perf_event *event)
4345 {
4346         perf_swevent_cancel_hrtimer(event);
4347         task_clock_perf_event_update(event, event->ctx->time);
4348
4349 }
4350
4351 static void task_clock_perf_event_read(struct perf_event *event)
4352 {
4353         u64 time;
4354
4355         if (!in_nmi()) {
4356                 update_context_time(event->ctx);
4357                 time = event->ctx->time;
4358         } else {
4359                 u64 now = perf_clock();
4360                 u64 delta = now - event->ctx->timestamp;
4361                 time = event->ctx->time + delta;
4362         }
4363
4364         task_clock_perf_event_update(event, time);
4365 }
4366
4367 static const struct pmu perf_ops_task_clock = {
4368         .enable         = task_clock_perf_event_enable,
4369         .disable        = task_clock_perf_event_disable,
4370         .read           = task_clock_perf_event_read,
4371 };
4372
4373 /* Deref the hlist from the update side */
4374 static inline struct swevent_hlist *
4375 swevent_hlist_deref(struct perf_cpu_context *cpuctx)
4376 {
4377         return rcu_dereference_protected(cpuctx->swevent_hlist,
4378                                          lockdep_is_held(&cpuctx->hlist_mutex));
4379 }
4380
4381 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4382 {
4383         struct swevent_hlist *hlist;
4384
4385         hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4386         kfree(hlist);
4387 }
4388
4389 static void swevent_hlist_release(struct perf_cpu_context *cpuctx)
4390 {
4391         struct swevent_hlist *hlist = swevent_hlist_deref(cpuctx);
4392
4393         if (!hlist)
4394                 return;
4395
4396         rcu_assign_pointer(cpuctx->swevent_hlist, NULL);
4397         call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4398 }
4399
4400 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4401 {
4402         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4403
4404         mutex_lock(&cpuctx->hlist_mutex);
4405
4406         if (!--cpuctx->hlist_refcount)
4407                 swevent_hlist_release(cpuctx);
4408
4409         mutex_unlock(&cpuctx->hlist_mutex);
4410 }
4411
4412 static void swevent_hlist_put(struct perf_event *event)
4413 {
4414         int cpu;
4415
4416         if (event->cpu != -1) {
4417                 swevent_hlist_put_cpu(event, event->cpu);
4418                 return;
4419         }
4420
4421         for_each_possible_cpu(cpu)
4422                 swevent_hlist_put_cpu(event, cpu);
4423 }
4424
4425 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4426 {
4427         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4428         int err = 0;
4429
4430         mutex_lock(&cpuctx->hlist_mutex);
4431
4432         if (!swevent_hlist_deref(cpuctx) && cpu_online(cpu)) {
4433                 struct swevent_hlist *hlist;
4434
4435                 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4436                 if (!hlist) {
4437                         err = -ENOMEM;
4438                         goto exit;
4439                 }
4440                 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
4441         }
4442         cpuctx->hlist_refcount++;
4443  exit:
4444         mutex_unlock(&cpuctx->hlist_mutex);
4445
4446         return err;
4447 }
4448
4449 static int swevent_hlist_get(struct perf_event *event)
4450 {
4451         int err;
4452         int cpu, failed_cpu;
4453
4454         if (event->cpu != -1)
4455                 return swevent_hlist_get_cpu(event, event->cpu);
4456
4457         get_online_cpus();
4458         for_each_possible_cpu(cpu) {
4459                 err = swevent_hlist_get_cpu(event, cpu);
4460                 if (err) {
4461                         failed_cpu = cpu;
4462                         goto fail;
4463                 }
4464         }
4465         put_online_cpus();
4466
4467         return 0;
4468  fail:
4469         for_each_possible_cpu(cpu) {
4470                 if (cpu == failed_cpu)
4471                         break;
4472                 swevent_hlist_put_cpu(event, cpu);
4473         }
4474
4475         put_online_cpus();
4476         return err;
4477 }
4478
4479 #ifdef CONFIG_EVENT_TRACING
4480
4481 static const struct pmu perf_ops_tracepoint = {
4482         .enable         = perf_trace_enable,
4483         .disable        = perf_trace_disable,
4484         .read           = perf_swevent_read,
4485         .unthrottle     = perf_swevent_unthrottle,
4486 };
4487
4488 static int perf_tp_filter_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 int perf_tp_event_match(struct perf_event *event,
4499                                 struct perf_sample_data *data,
4500                                 struct pt_regs *regs)
4501 {
4502         if (perf_exclude_event(event, regs))
4503                 return 0;
4504
4505         if (!perf_tp_filter_match(event, data))
4506                 return 0;
4507
4508         return 1;
4509 }
4510
4511 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
4512                    struct pt_regs *regs, struct hlist_head *head)
4513 {
4514         struct perf_sample_data data;
4515         struct perf_event *event;
4516         struct hlist_node *node;
4517
4518         struct perf_raw_record raw = {
4519                 .size = entry_size,
4520                 .data = record,
4521         };
4522
4523         perf_sample_data_init(&data, addr);
4524         data.raw = &raw;
4525
4526         rcu_read_lock();
4527         hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4528                 if (perf_tp_event_match(event, &data, regs))
4529                         perf_swevent_add(event, count, 1, &data, regs);
4530         }
4531         rcu_read_unlock();
4532 }
4533 EXPORT_SYMBOL_GPL(perf_tp_event);
4534
4535 static void tp_perf_event_destroy(struct perf_event *event)
4536 {
4537         perf_trace_destroy(event);
4538 }
4539
4540 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4541 {
4542         int err;
4543
4544         /*
4545          * Raw tracepoint data is a severe data leak, only allow root to
4546          * have these.
4547          */
4548         if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4549                         perf_paranoid_tracepoint_raw() &&
4550                         !capable(CAP_SYS_ADMIN))
4551                 return ERR_PTR(-EPERM);
4552
4553         err = perf_trace_init(event);
4554         if (err)
4555                 return NULL;
4556
4557         event->destroy = tp_perf_event_destroy;
4558
4559         return &perf_ops_tracepoint;
4560 }
4561
4562 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4563 {
4564         char *filter_str;
4565         int ret;
4566
4567         if (event->attr.type != PERF_TYPE_TRACEPOINT)
4568                 return -EINVAL;
4569
4570         filter_str = strndup_user(arg, PAGE_SIZE);
4571         if (IS_ERR(filter_str))
4572                 return PTR_ERR(filter_str);
4573
4574         ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4575
4576         kfree(filter_str);
4577         return ret;
4578 }
4579
4580 static void perf_event_free_filter(struct perf_event *event)
4581 {
4582         ftrace_profile_free_filter(event);
4583 }
4584
4585 #else
4586
4587 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4588 {
4589         return NULL;
4590 }
4591
4592 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4593 {
4594         return -ENOENT;
4595 }
4596
4597 static void perf_event_free_filter(struct perf_event *event)
4598 {
4599 }
4600
4601 #endif /* CONFIG_EVENT_TRACING */
4602
4603 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4604 static void bp_perf_event_destroy(struct perf_event *event)
4605 {
4606         release_bp_slot(event);
4607 }
4608
4609 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4610 {
4611         int err;
4612
4613         err = register_perf_hw_breakpoint(bp);
4614         if (err)
4615                 return ERR_PTR(err);
4616
4617         bp->destroy = bp_perf_event_destroy;
4618
4619         return &perf_ops_bp;
4620 }
4621
4622 void perf_bp_event(struct perf_event *bp, void *data)
4623 {
4624         struct perf_sample_data sample;
4625         struct pt_regs *regs = data;
4626
4627         perf_sample_data_init(&sample, bp->attr.bp_addr);
4628
4629         if (!perf_exclude_event(bp, regs))
4630                 perf_swevent_add(bp, 1, 1, &sample, regs);
4631 }
4632 #else
4633 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4634 {
4635         return NULL;
4636 }
4637
4638 void perf_bp_event(struct perf_event *bp, void *regs)
4639 {
4640 }
4641 #endif
4642
4643 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4644
4645 static void sw_perf_event_destroy(struct perf_event *event)
4646 {
4647         u64 event_id = event->attr.config;
4648
4649         WARN_ON(event->parent);
4650
4651         atomic_dec(&perf_swevent_enabled[event_id]);
4652         swevent_hlist_put(event);
4653 }
4654
4655 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4656 {
4657         const struct pmu *pmu = NULL;
4658         u64 event_id = event->attr.config;
4659
4660         /*
4661          * Software events (currently) can't in general distinguish
4662          * between user, kernel and hypervisor events.
4663          * However, context switches and cpu migrations are considered
4664          * to be kernel events, and page faults are never hypervisor
4665          * events.
4666          */
4667         switch (event_id) {
4668         case PERF_COUNT_SW_CPU_CLOCK:
4669                 pmu = &perf_ops_cpu_clock;
4670
4671                 break;
4672         case PERF_COUNT_SW_TASK_CLOCK:
4673                 /*
4674                  * If the user instantiates this as a per-cpu event,
4675                  * use the cpu_clock event instead.
4676                  */
4677                 if (event->ctx->task)
4678                         pmu = &perf_ops_task_clock;
4679                 else
4680                         pmu = &perf_ops_cpu_clock;
4681
4682                 break;
4683         case PERF_COUNT_SW_PAGE_FAULTS:
4684         case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4685         case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4686         case PERF_COUNT_SW_CONTEXT_SWITCHES:
4687         case PERF_COUNT_SW_CPU_MIGRATIONS:
4688         case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4689         case PERF_COUNT_SW_EMULATION_FAULTS:
4690                 if (!event->parent) {
4691                         int err;
4692
4693                         err = swevent_hlist_get(event);
4694                         if (err)
4695                                 return ERR_PTR(err);
4696
4697                         atomic_inc(&perf_swevent_enabled[event_id]);
4698                         event->destroy = sw_perf_event_destroy;
4699                 }
4700                 pmu = &perf_ops_generic;
4701                 break;
4702         }
4703
4704         return pmu;
4705 }
4706
4707 /*
4708  * Allocate and initialize a event structure
4709  */
4710 static struct perf_event *
4711 perf_event_alloc(struct perf_event_attr *attr,
4712                    int cpu,
4713                    struct perf_event_context *ctx,
4714                    struct perf_event *group_leader,
4715                    struct perf_event *parent_event,
4716                    perf_overflow_handler_t overflow_handler,
4717                    gfp_t gfpflags)
4718 {
4719         const struct pmu *pmu;
4720         struct perf_event *event;
4721         struct hw_perf_event *hwc;
4722         long err;
4723
4724         event = kzalloc(sizeof(*event), gfpflags);
4725         if (!event)
4726                 return ERR_PTR(-ENOMEM);
4727
4728         /*
4729          * Single events are their own group leaders, with an
4730          * empty sibling list:
4731          */
4732         if (!group_leader)
4733                 group_leader = event;
4734
4735         mutex_init(&event->child_mutex);
4736         INIT_LIST_HEAD(&event->child_list);
4737
4738         INIT_LIST_HEAD(&event->group_entry);
4739         INIT_LIST_HEAD(&event->event_entry);
4740         INIT_LIST_HEAD(&event->sibling_list);
4741         init_waitqueue_head(&event->waitq);
4742
4743         mutex_init(&event->mmap_mutex);
4744
4745         event->cpu              = cpu;
4746         event->attr             = *attr;
4747         event->group_leader     = group_leader;
4748         event->pmu              = NULL;
4749         event->ctx              = ctx;
4750         event->oncpu            = -1;
4751
4752         event->parent           = parent_event;
4753
4754         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
4755         event->id               = atomic64_inc_return(&perf_event_id);
4756
4757         event->state            = PERF_EVENT_STATE_INACTIVE;
4758
4759         if (!overflow_handler && parent_event)
4760                 overflow_handler = parent_event->overflow_handler;
4761         
4762         event->overflow_handler = overflow_handler;
4763
4764         if (attr->disabled)
4765                 event->state = PERF_EVENT_STATE_OFF;
4766
4767         pmu = NULL;
4768
4769         hwc = &event->hw;
4770         hwc->sample_period = attr->sample_period;
4771         if (attr->freq && attr->sample_freq)
4772                 hwc->sample_period = 1;
4773         hwc->last_period = hwc->sample_period;
4774
4775         atomic64_set(&hwc->period_left, hwc->sample_period);
4776
4777         /*
4778          * we currently do not support PERF_FORMAT_GROUP on inherited events
4779          */
4780         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4781                 goto done;
4782
4783         switch (attr->type) {
4784         case PERF_TYPE_RAW:
4785         case PERF_TYPE_HARDWARE:
4786         case PERF_TYPE_HW_CACHE:
4787                 pmu = hw_perf_event_init(event);
4788                 break;
4789
4790         case PERF_TYPE_SOFTWARE:
4791                 pmu = sw_perf_event_init(event);
4792                 break;
4793
4794         case PERF_TYPE_TRACEPOINT:
4795                 pmu = tp_perf_event_init(event);
4796                 break;
4797
4798         case PERF_TYPE_BREAKPOINT:
4799                 pmu = bp_perf_event_init(event);
4800                 break;
4801
4802
4803         default:
4804                 break;
4805         }
4806 done:
4807         err = 0;
4808         if (!pmu)
4809                 err = -EINVAL;
4810         else if (IS_ERR(pmu))
4811                 err = PTR_ERR(pmu);
4812
4813         if (err) {
4814                 if (event->ns)
4815                         put_pid_ns(event->ns);
4816                 kfree(event);
4817                 return ERR_PTR(err);
4818         }
4819
4820         event->pmu = pmu;
4821
4822         if (!event->parent) {
4823                 atomic_inc(&nr_events);
4824                 if (event->attr.mmap)
4825                         atomic_inc(&nr_mmap_events);
4826                 if (event->attr.comm)
4827                         atomic_inc(&nr_comm_events);
4828                 if (event->attr.task)
4829                         atomic_inc(&nr_task_events);
4830         }
4831
4832         return event;
4833 }
4834
4835 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4836                           struct perf_event_attr *attr)
4837 {
4838         u32 size;
4839         int ret;
4840
4841         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4842                 return -EFAULT;
4843
4844         /*
4845          * zero the full structure, so that a short copy will be nice.
4846          */
4847         memset(attr, 0, sizeof(*attr));
4848
4849         ret = get_user(size, &uattr->size);
4850         if (ret)
4851                 return ret;
4852
4853         if (size > PAGE_SIZE)   /* silly large */
4854                 goto err_size;
4855
4856         if (!size)              /* abi compat */
4857                 size = PERF_ATTR_SIZE_VER0;
4858
4859         if (size < PERF_ATTR_SIZE_VER0)
4860                 goto err_size;
4861
4862         /*
4863          * If we're handed a bigger struct than we know of,
4864          * ensure all the unknown bits are 0 - i.e. new
4865          * user-space does not rely on any kernel feature
4866          * extensions we dont know about yet.
4867          */
4868         if (size > sizeof(*attr)) {
4869                 unsigned char __user *addr;
4870                 unsigned char __user *end;
4871                 unsigned char val;
4872
4873                 addr = (void __user *)uattr + sizeof(*attr);
4874                 end  = (void __user *)uattr + size;
4875
4876                 for (; addr < end; addr++) {
4877                         ret = get_user(val, addr);
4878                         if (ret)
4879                                 return ret;
4880                         if (val)
4881                                 goto err_size;
4882                 }
4883                 size = sizeof(*attr);
4884         }
4885
4886         ret = copy_from_user(attr, uattr, size);
4887         if (ret)
4888                 return -EFAULT;
4889
4890         /*
4891          * If the type exists, the corresponding creation will verify
4892          * the attr->config.
4893          */
4894         if (attr->type >= PERF_TYPE_MAX)
4895                 return -EINVAL;
4896
4897         if (attr->__reserved_1)
4898                 return -EINVAL;
4899
4900         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4901                 return -EINVAL;
4902
4903         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4904                 return -EINVAL;
4905
4906 out:
4907         return ret;
4908
4909 err_size:
4910         put_user(sizeof(*attr), &uattr->size);
4911         ret = -E2BIG;
4912         goto out;
4913 }
4914
4915 static int perf_event_set_output(struct perf_event *event, int output_fd)
4916 {
4917         struct perf_event *output_event = NULL;
4918         struct file *output_file = NULL;
4919         struct perf_event *old_output;
4920         int fput_needed = 0;
4921         int ret = -EINVAL;
4922
4923         if (!output_fd)
4924                 goto set;
4925
4926         output_file = fget_light(output_fd, &fput_needed);
4927         if (!output_file)
4928                 return -EBADF;
4929
4930         if (output_file->f_op != &perf_fops)
4931                 goto out;
4932
4933         output_event = output_file->private_data;
4934
4935         /* Don't chain output fds */
4936         if (output_event->output)
4937                 goto out;
4938
4939         /* Don't set an output fd when we already have an output channel */
4940         if (event->data)
4941                 goto out;
4942
4943         atomic_long_inc(&output_file->f_count);
4944
4945 set:
4946         mutex_lock(&event->mmap_mutex);
4947         old_output = event->output;
4948         rcu_assign_pointer(event->output, output_event);
4949         mutex_unlock(&event->mmap_mutex);
4950
4951         if (old_output) {
4952                 /*
4953                  * we need to make sure no existing perf_output_*()
4954                  * is still referencing this event.
4955                  */
4956                 synchronize_rcu();
4957                 fput(old_output->filp);
4958         }
4959
4960         ret = 0;
4961 out:
4962         fput_light(output_file, fput_needed);
4963         return ret;
4964 }
4965
4966 /**
4967  * sys_perf_event_open - open a performance event, associate it to a task/cpu
4968  *
4969  * @attr_uptr:  event_id type attributes for monitoring/sampling
4970  * @pid:                target pid
4971  * @cpu:                target cpu
4972  * @group_fd:           group leader event fd
4973  */
4974 SYSCALL_DEFINE5(perf_event_open,
4975                 struct perf_event_attr __user *, attr_uptr,
4976                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4977 {
4978         struct perf_event *event, *group_leader;
4979         struct perf_event_attr attr;
4980         struct perf_event_context *ctx;
4981         struct file *event_file = NULL;
4982         struct file *group_file = NULL;
4983         int fput_needed = 0;
4984         int fput_needed2 = 0;
4985         int err;
4986
4987         /* for future expandability... */
4988         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4989                 return -EINVAL;
4990
4991         err = perf_copy_attr(attr_uptr, &attr);
4992         if (err)
4993                 return err;
4994
4995         if (!attr.exclude_kernel) {
4996                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4997                         return -EACCES;
4998         }
4999
5000         if (attr.freq) {
5001                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
5002                         return -EINVAL;
5003         }
5004
5005         /*
5006          * Get the target context (task or percpu):
5007          */
5008         ctx = find_get_context(pid, cpu);
5009         if (IS_ERR(ctx))
5010                 return PTR_ERR(ctx);
5011
5012         /*
5013          * Look up the group leader (we will attach this event to it):
5014          */
5015         group_leader = NULL;
5016         if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
5017                 err = -EINVAL;
5018                 group_file = fget_light(group_fd, &fput_needed);
5019                 if (!group_file)
5020                         goto err_put_context;
5021                 if (group_file->f_op != &perf_fops)
5022                         goto err_put_context;
5023
5024                 group_leader = group_file->private_data;
5025                 /*
5026                  * Do not allow a recursive hierarchy (this new sibling
5027                  * becoming part of another group-sibling):
5028                  */
5029                 if (group_leader->group_leader != group_leader)
5030                         goto err_put_context;
5031                 /*
5032                  * Do not allow to attach to a group in a different
5033                  * task or CPU context:
5034                  */
5035                 if (group_leader->ctx != ctx)
5036                         goto err_put_context;
5037                 /*
5038                  * Only a group leader can be exclusive or pinned
5039                  */
5040                 if (attr.exclusive || attr.pinned)
5041                         goto err_put_context;
5042         }
5043
5044         event = perf_event_alloc(&attr, cpu, ctx, group_leader,
5045                                      NULL, NULL, GFP_KERNEL);
5046         err = PTR_ERR(event);
5047         if (IS_ERR(event))
5048                 goto err_put_context;
5049
5050         err = anon_inode_getfd("[perf_event]", &perf_fops, event, O_RDWR);
5051         if (err < 0)
5052                 goto err_free_put_context;
5053
5054         event_file = fget_light(err, &fput_needed2);
5055         if (!event_file)
5056                 goto err_free_put_context;
5057
5058         if (flags & PERF_FLAG_FD_OUTPUT) {
5059                 err = perf_event_set_output(event, group_fd);
5060                 if (err)
5061                         goto err_fput_free_put_context;
5062         }
5063
5064         event->filp = event_file;
5065         WARN_ON_ONCE(ctx->parent_ctx);
5066         mutex_lock(&ctx->mutex);
5067         perf_install_in_context(ctx, event, cpu);
5068         ++ctx->generation;
5069         mutex_unlock(&ctx->mutex);
5070
5071         event->owner = current;
5072         get_task_struct(current);
5073         mutex_lock(&current->perf_event_mutex);
5074         list_add_tail(&event->owner_entry, &current->perf_event_list);
5075         mutex_unlock(&current->perf_event_mutex);
5076
5077 err_fput_free_put_context:
5078         fput_light(event_file, fput_needed2);
5079
5080 err_free_put_context:
5081         if (err < 0)
5082                 free_event(event);
5083
5084 err_put_context:
5085         if (err < 0)
5086                 put_ctx(ctx);
5087
5088         fput_light(group_file, fput_needed);
5089
5090         return err;
5091 }
5092
5093 /**
5094  * perf_event_create_kernel_counter
5095  *
5096  * @attr: attributes of the counter to create
5097  * @cpu: cpu in which the counter is bound
5098  * @pid: task to profile
5099  */
5100 struct perf_event *
5101 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
5102                                  pid_t pid,
5103                                  perf_overflow_handler_t overflow_handler)
5104 {
5105         struct perf_event *event;
5106         struct perf_event_context *ctx;
5107         int err;
5108
5109         /*
5110          * Get the target context (task or percpu):
5111          */
5112
5113         ctx = find_get_context(pid, cpu);
5114         if (IS_ERR(ctx)) {
5115                 err = PTR_ERR(ctx);
5116                 goto err_exit;
5117         }
5118
5119         event = perf_event_alloc(attr, cpu, ctx, NULL,
5120                                  NULL, overflow_handler, GFP_KERNEL);
5121         if (IS_ERR(event)) {
5122                 err = PTR_ERR(event);
5123                 goto err_put_context;
5124         }
5125
5126         event->filp = NULL;
5127         WARN_ON_ONCE(ctx->parent_ctx);
5128         mutex_lock(&ctx->mutex);
5129         perf_install_in_context(ctx, event, cpu);
5130         ++ctx->generation;
5131         mutex_unlock(&ctx->mutex);
5132
5133         event->owner = current;
5134         get_task_struct(current);
5135         mutex_lock(&current->perf_event_mutex);
5136         list_add_tail(&event->owner_entry, &current->perf_event_list);
5137         mutex_unlock(&current->perf_event_mutex);
5138
5139         return event;
5140
5141  err_put_context:
5142         put_ctx(ctx);
5143  err_exit:
5144         return ERR_PTR(err);
5145 }
5146 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
5147
5148 /*
5149  * inherit a event from parent task to child task:
5150  */
5151 static struct perf_event *
5152 inherit_event(struct perf_event *parent_event,
5153               struct task_struct *parent,
5154               struct perf_event_context *parent_ctx,
5155               struct task_struct *child,
5156               struct perf_event *group_leader,
5157               struct perf_event_context *child_ctx)
5158 {
5159         struct perf_event *child_event;
5160
5161         /*
5162          * Instead of creating recursive hierarchies of events,
5163          * we link inherited events back to the original parent,
5164          * which has a filp for sure, which we use as the reference
5165          * count:
5166          */
5167         if (parent_event->parent)
5168                 parent_event = parent_event->parent;
5169
5170         child_event = perf_event_alloc(&parent_event->attr,
5171                                            parent_event->cpu, child_ctx,
5172                                            group_leader, parent_event,
5173                                            NULL, GFP_KERNEL);
5174         if (IS_ERR(child_event))
5175                 return child_event;
5176         get_ctx(child_ctx);
5177
5178         /*
5179          * Make the child state follow the state of the parent event,
5180          * not its attr.disabled bit.  We hold the parent's mutex,
5181          * so we won't race with perf_event_{en, dis}able_family.
5182          */
5183         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
5184                 child_event->state = PERF_EVENT_STATE_INACTIVE;
5185         else
5186                 child_event->state = PERF_EVENT_STATE_OFF;
5187
5188         if (parent_event->attr.freq) {
5189                 u64 sample_period = parent_event->hw.sample_period;
5190                 struct hw_perf_event *hwc = &child_event->hw;
5191
5192                 hwc->sample_period = sample_period;
5193                 hwc->last_period   = sample_period;
5194
5195                 atomic64_set(&hwc->period_left, sample_period);
5196         }
5197
5198         child_event->overflow_handler = parent_event->overflow_handler;
5199
5200         /*
5201          * Link it up in the child's context:
5202          */
5203         add_event_to_ctx(child_event, child_ctx);
5204
5205         /*
5206          * Get a reference to the parent filp - we will fput it
5207          * when the child event exits. This is safe to do because
5208          * we are in the parent and we know that the filp still
5209          * exists and has a nonzero count:
5210          */
5211         atomic_long_inc(&parent_event->filp->f_count);
5212
5213         /*
5214          * Link this into the parent event's child list
5215          */
5216         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5217         mutex_lock(&parent_event->child_mutex);
5218         list_add_tail(&child_event->child_list, &parent_event->child_list);
5219         mutex_unlock(&parent_event->child_mutex);
5220
5221         return child_event;
5222 }
5223
5224 static int inherit_group(struct perf_event *parent_event,
5225               struct task_struct *parent,
5226               struct perf_event_context *parent_ctx,
5227               struct task_struct *child,
5228               struct perf_event_context *child_ctx)
5229 {
5230         struct perf_event *leader;
5231         struct perf_event *sub;
5232         struct perf_event *child_ctr;
5233
5234         leader = inherit_event(parent_event, parent, parent_ctx,
5235                                  child, NULL, child_ctx);
5236         if (IS_ERR(leader))
5237                 return PTR_ERR(leader);
5238         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5239                 child_ctr = inherit_event(sub, parent, parent_ctx,
5240                                             child, leader, child_ctx);
5241                 if (IS_ERR(child_ctr))
5242                         return PTR_ERR(child_ctr);
5243         }
5244         return 0;
5245 }
5246
5247 static void sync_child_event(struct perf_event *child_event,
5248                                struct task_struct *child)
5249 {
5250         struct perf_event *parent_event = child_event->parent;
5251         u64 child_val;
5252
5253         if (child_event->attr.inherit_stat)
5254                 perf_event_read_event(child_event, child);
5255
5256         child_val = atomic64_read(&child_event->count);
5257
5258         /*
5259          * Add back the child's count to the parent's count:
5260          */
5261         atomic64_add(child_val, &parent_event->count);
5262         atomic64_add(child_event->total_time_enabled,
5263                      &parent_event->child_total_time_enabled);
5264         atomic64_add(child_event->total_time_running,
5265                      &parent_event->child_total_time_running);
5266
5267         /*
5268          * Remove this event from the parent's list
5269          */
5270         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5271         mutex_lock(&parent_event->child_mutex);
5272         list_del_init(&child_event->child_list);
5273         mutex_unlock(&parent_event->child_mutex);
5274
5275         /*
5276          * Release the parent event, if this was the last
5277          * reference to it.
5278          */
5279         fput(parent_event->filp);
5280 }
5281
5282 static void
5283 __perf_event_exit_task(struct perf_event *child_event,
5284                          struct perf_event_context *child_ctx,
5285                          struct task_struct *child)
5286 {
5287         struct perf_event *parent_event;
5288
5289         perf_event_remove_from_context(child_event);
5290
5291         parent_event = child_event->parent;
5292         /*
5293          * It can happen that parent exits first, and has events
5294          * that are still around due to the child reference. These
5295          * events need to be zapped - but otherwise linger.
5296          */
5297         if (parent_event) {
5298                 sync_child_event(child_event, child);
5299                 free_event(child_event);
5300         }
5301 }
5302
5303 /*
5304  * When a child task exits, feed back event values to parent events.
5305  */
5306 void perf_event_exit_task(struct task_struct *child)
5307 {
5308         struct perf_event *child_event, *tmp;
5309         struct perf_event_context *child_ctx;
5310         unsigned long flags;
5311
5312         if (likely(!child->perf_event_ctxp)) {
5313                 perf_event_task(child, NULL, 0);
5314                 return;
5315         }
5316
5317         local_irq_save(flags);
5318         /*
5319          * We can't reschedule here because interrupts are disabled,
5320          * and either child is current or it is a task that can't be
5321          * scheduled, so we are now safe from rescheduling changing
5322          * our context.
5323          */
5324         child_ctx = child->perf_event_ctxp;
5325         __perf_event_task_sched_out(child_ctx);
5326
5327         /*
5328          * Take the context lock here so that if find_get_context is
5329          * reading child->perf_event_ctxp, we wait until it has
5330          * incremented the context's refcount before we do put_ctx below.
5331          */
5332         raw_spin_lock(&child_ctx->lock);
5333         child->perf_event_ctxp = NULL;
5334         /*
5335          * If this context is a clone; unclone it so it can't get
5336          * swapped to another process while we're removing all
5337          * the events from it.
5338          */
5339         unclone_ctx(child_ctx);
5340         update_context_time(child_ctx);
5341         raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5342
5343         /*
5344          * Report the task dead after unscheduling the events so that we
5345          * won't get any samples after PERF_RECORD_EXIT. We can however still
5346          * get a few PERF_RECORD_READ events.
5347          */
5348         perf_event_task(child, child_ctx, 0);
5349
5350         /*
5351          * We can recurse on the same lock type through:
5352          *
5353          *   __perf_event_exit_task()
5354          *     sync_child_event()
5355          *       fput(parent_event->filp)
5356          *         perf_release()
5357          *           mutex_lock(&ctx->mutex)
5358          *
5359          * But since its the parent context it won't be the same instance.
5360          */
5361         mutex_lock(&child_ctx->mutex);
5362
5363 again:
5364         list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5365                                  group_entry)
5366                 __perf_event_exit_task(child_event, child_ctx, child);
5367
5368         list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5369                                  group_entry)
5370                 __perf_event_exit_task(child_event, child_ctx, child);
5371
5372         /*
5373          * If the last event was a group event, it will have appended all
5374          * its siblings to the list, but we obtained 'tmp' before that which
5375          * will still point to the list head terminating the iteration.
5376          */
5377         if (!list_empty(&child_ctx->pinned_groups) ||
5378             !list_empty(&child_ctx->flexible_groups))
5379                 goto again;
5380
5381         mutex_unlock(&child_ctx->mutex);
5382
5383         put_ctx(child_ctx);
5384 }
5385
5386 static void perf_free_event(struct perf_event *event,
5387                             struct perf_event_context *ctx)
5388 {
5389         struct perf_event *parent = event->parent;
5390
5391         if (WARN_ON_ONCE(!parent))
5392                 return;
5393
5394         mutex_lock(&parent->child_mutex);
5395         list_del_init(&event->child_list);
5396         mutex_unlock(&parent->child_mutex);
5397
5398         fput(parent->filp);
5399
5400         list_del_event(event, ctx);
5401         free_event(event);
5402 }
5403
5404 /*
5405  * free an unexposed, unused context as created by inheritance by
5406  * init_task below, used by fork() in case of fail.
5407  */
5408 void perf_event_free_task(struct task_struct *task)
5409 {
5410         struct perf_event_context *ctx = task->perf_event_ctxp;
5411         struct perf_event *event, *tmp;
5412
5413         if (!ctx)
5414                 return;
5415
5416         mutex_lock(&ctx->mutex);
5417 again:
5418         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5419                 perf_free_event(event, ctx);
5420
5421         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5422                                  group_entry)
5423                 perf_free_event(event, ctx);
5424
5425         if (!list_empty(&ctx->pinned_groups) ||
5426             !list_empty(&ctx->flexible_groups))
5427                 goto again;
5428
5429         mutex_unlock(&ctx->mutex);
5430
5431         put_ctx(ctx);
5432 }
5433
5434 static int
5435 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5436                    struct perf_event_context *parent_ctx,
5437                    struct task_struct *child,
5438                    int *inherited_all)
5439 {
5440         int ret;
5441         struct perf_event_context *child_ctx = child->perf_event_ctxp;
5442
5443         if (!event->attr.inherit) {
5444                 *inherited_all = 0;
5445                 return 0;
5446         }
5447
5448         if (!child_ctx) {
5449                 /*
5450                  * This is executed from the parent task context, so
5451                  * inherit events that have been marked for cloning.
5452                  * First allocate and initialize a context for the
5453                  * child.
5454                  */
5455
5456                 child_ctx = kzalloc(sizeof(struct perf_event_context),
5457                                     GFP_KERNEL);
5458                 if (!child_ctx)
5459                         return -ENOMEM;
5460
5461                 __perf_event_init_context(child_ctx, child);
5462                 child->perf_event_ctxp = child_ctx;
5463                 get_task_struct(child);
5464         }
5465
5466         ret = inherit_group(event, parent, parent_ctx,
5467                             child, child_ctx);
5468
5469         if (ret)
5470                 *inherited_all = 0;
5471
5472         return ret;
5473 }
5474
5475
5476 /*
5477  * Initialize the perf_event context in task_struct
5478  */
5479 int perf_event_init_task(struct task_struct *child)
5480 {
5481         struct perf_event_context *child_ctx, *parent_ctx;
5482         struct perf_event_context *cloned_ctx;
5483         struct perf_event *event;
5484         struct task_struct *parent = current;
5485         int inherited_all = 1;
5486         int ret = 0;
5487
5488         child->perf_event_ctxp = NULL;
5489
5490         mutex_init(&child->perf_event_mutex);
5491         INIT_LIST_HEAD(&child->perf_event_list);
5492
5493         if (likely(!parent->perf_event_ctxp))
5494                 return 0;
5495
5496         /*
5497          * If the parent's context is a clone, pin it so it won't get
5498          * swapped under us.
5499          */
5500         parent_ctx = perf_pin_task_context(parent);
5501
5502         /*
5503          * No need to check if parent_ctx != NULL here; since we saw
5504          * it non-NULL earlier, the only reason for it to become NULL
5505          * is if we exit, and since we're currently in the middle of
5506          * a fork we can't be exiting at the same time.
5507          */
5508
5509         /*
5510          * Lock the parent list. No need to lock the child - not PID
5511          * hashed yet and not running, so nobody can access it.
5512          */
5513         mutex_lock(&parent_ctx->mutex);
5514
5515         /*
5516          * We dont have to disable NMIs - we are only looking at
5517          * the list, not manipulating it:
5518          */
5519         list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5520                 ret = inherit_task_group(event, parent, parent_ctx, child,
5521                                          &inherited_all);
5522                 if (ret)
5523                         break;
5524         }
5525
5526         list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5527                 ret = inherit_task_group(event, parent, parent_ctx, child,
5528                                          &inherited_all);
5529                 if (ret)
5530                         break;
5531         }
5532
5533         child_ctx = child->perf_event_ctxp;
5534
5535         if (child_ctx && inherited_all) {
5536                 /*
5537                  * Mark the child context as a clone of the parent
5538                  * context, or of whatever the parent is a clone of.
5539                  * Note that if the parent is a clone, it could get
5540                  * uncloned at any point, but that doesn't matter
5541                  * because the list of events and the generation
5542                  * count can't have changed since we took the mutex.
5543                  */
5544                 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5545                 if (cloned_ctx) {
5546                         child_ctx->parent_ctx = cloned_ctx;
5547                         child_ctx->parent_gen = parent_ctx->parent_gen;
5548                 } else {
5549                         child_ctx->parent_ctx = parent_ctx;
5550                         child_ctx->parent_gen = parent_ctx->generation;
5551                 }
5552                 get_ctx(child_ctx->parent_ctx);
5553         }
5554
5555         mutex_unlock(&parent_ctx->mutex);
5556
5557         perf_unpin_context(parent_ctx);
5558
5559         return ret;
5560 }
5561
5562 static void __init perf_event_init_all_cpus(void)
5563 {
5564         int cpu;
5565         struct perf_cpu_context *cpuctx;
5566
5567         for_each_possible_cpu(cpu) {
5568                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5569                 mutex_init(&cpuctx->hlist_mutex);
5570                 __perf_event_init_context(&cpuctx->ctx, NULL);
5571         }
5572 }
5573
5574 static void __cpuinit perf_event_init_cpu(int cpu)
5575 {
5576         struct perf_cpu_context *cpuctx;
5577
5578         cpuctx = &per_cpu(perf_cpu_context, cpu);
5579
5580         spin_lock(&perf_resource_lock);
5581         cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5582         spin_unlock(&perf_resource_lock);
5583
5584         mutex_lock(&cpuctx->hlist_mutex);
5585         if (cpuctx->hlist_refcount > 0) {
5586                 struct swevent_hlist *hlist;
5587
5588                 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5589                 WARN_ON_ONCE(!hlist);
5590                 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
5591         }
5592         mutex_unlock(&cpuctx->hlist_mutex);
5593 }
5594
5595 #ifdef CONFIG_HOTPLUG_CPU
5596 static void __perf_event_exit_cpu(void *info)
5597 {
5598         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5599         struct perf_event_context *ctx = &cpuctx->ctx;
5600         struct perf_event *event, *tmp;
5601
5602         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5603                 __perf_event_remove_from_context(event);
5604         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5605                 __perf_event_remove_from_context(event);
5606 }
5607 static void perf_event_exit_cpu(int cpu)
5608 {
5609         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5610         struct perf_event_context *ctx = &cpuctx->ctx;
5611
5612         mutex_lock(&cpuctx->hlist_mutex);
5613         swevent_hlist_release(cpuctx);
5614         mutex_unlock(&cpuctx->hlist_mutex);
5615
5616         mutex_lock(&ctx->mutex);
5617         smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5618         mutex_unlock(&ctx->mutex);
5619 }
5620 #else
5621 static inline void perf_event_exit_cpu(int cpu) { }
5622 #endif
5623
5624 static int __cpuinit
5625 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5626 {
5627         unsigned int cpu = (long)hcpu;
5628
5629         switch (action) {
5630
5631         case CPU_UP_PREPARE:
5632         case CPU_UP_PREPARE_FROZEN:
5633                 perf_event_init_cpu(cpu);
5634                 break;
5635
5636         case CPU_DOWN_PREPARE:
5637         case CPU_DOWN_PREPARE_FROZEN:
5638                 perf_event_exit_cpu(cpu);
5639                 break;
5640
5641         default:
5642                 break;
5643         }
5644
5645         return NOTIFY_OK;
5646 }
5647
5648 /*
5649  * This has to have a higher priority than migration_notifier in sched.c.
5650  */
5651 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5652         .notifier_call          = perf_cpu_notify,
5653         .priority               = 20,
5654 };
5655
5656 void __init perf_event_init(void)
5657 {
5658         perf_event_init_all_cpus();
5659         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5660                         (void *)(long)smp_processor_id());
5661         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5662                         (void *)(long)smp_processor_id());
5663         register_cpu_notifier(&perf_cpu_nb);
5664 }
5665
5666 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class,
5667                                         struct sysdev_class_attribute *attr,
5668                                         char *buf)
5669 {
5670         return sprintf(buf, "%d\n", perf_reserved_percpu);
5671 }
5672
5673 static ssize_t
5674 perf_set_reserve_percpu(struct sysdev_class *class,
5675                         struct sysdev_class_attribute *attr,
5676                         const char *buf,
5677                         size_t count)
5678 {
5679         struct perf_cpu_context *cpuctx;
5680         unsigned long val;
5681         int err, cpu, mpt;
5682
5683         err = strict_strtoul(buf, 10, &val);
5684         if (err)
5685                 return err;
5686         if (val > perf_max_events)
5687                 return -EINVAL;
5688
5689         spin_lock(&perf_resource_lock);
5690         perf_reserved_percpu = val;
5691         for_each_online_cpu(cpu) {
5692                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5693                 raw_spin_lock_irq(&cpuctx->ctx.lock);
5694                 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5695                           perf_max_events - perf_reserved_percpu);
5696                 cpuctx->max_pertask = mpt;
5697                 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5698         }
5699         spin_unlock(&perf_resource_lock);
5700
5701         return count;
5702 }
5703
5704 static ssize_t perf_show_overcommit(struct sysdev_class *class,
5705                                     struct sysdev_class_attribute *attr,
5706                                     char *buf)
5707 {
5708         return sprintf(buf, "%d\n", perf_overcommit);
5709 }
5710
5711 static ssize_t
5712 perf_set_overcommit(struct sysdev_class *class,
5713                     struct sysdev_class_attribute *attr,
5714                     const char *buf, size_t count)
5715 {
5716         unsigned long val;
5717         int err;
5718
5719         err = strict_strtoul(buf, 10, &val);
5720         if (err)
5721                 return err;
5722         if (val > 1)
5723                 return -EINVAL;
5724
5725         spin_lock(&perf_resource_lock);
5726         perf_overcommit = val;
5727         spin_unlock(&perf_resource_lock);
5728
5729         return count;
5730 }
5731
5732 static SYSDEV_CLASS_ATTR(
5733                                 reserve_percpu,
5734                                 0644,
5735                                 perf_show_reserve_percpu,
5736                                 perf_set_reserve_percpu
5737                         );
5738
5739 static SYSDEV_CLASS_ATTR(
5740                                 overcommit,
5741                                 0644,
5742                                 perf_show_overcommit,
5743                                 perf_set_overcommit
5744                         );
5745
5746 static struct attribute *perfclass_attrs[] = {
5747         &attr_reserve_percpu.attr,
5748         &attr_overcommit.attr,
5749         NULL
5750 };
5751
5752 static struct attribute_group perfclass_attr_group = {
5753         .attrs                  = perfclass_attrs,
5754         .name                   = "perf_events",
5755 };
5756
5757 static int __init perf_event_sysfs_init(void)
5758 {
5759         return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5760                                   &perfclass_attr_group);
5761 }
5762 device_initcall(perf_event_sysfs_init);