Merge branch 'perf/urgent' of git://git.kernel.org/pub/scm/linux/kernel/git/frederic...
[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                 return;
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         preempt_enable();
2958 }
2959
2960 void perf_output_copy(struct perf_output_handle *handle,
2961                       const void *buf, unsigned int len)
2962 {
2963         unsigned int pages_mask;
2964         unsigned long offset;
2965         unsigned int size;
2966         void **pages;
2967
2968         offset          = handle->offset;
2969         pages_mask      = handle->data->nr_pages - 1;
2970         pages           = handle->data->data_pages;
2971
2972         do {
2973                 unsigned long page_offset;
2974                 unsigned long page_size;
2975                 int nr;
2976
2977                 nr          = (offset >> PAGE_SHIFT) & pages_mask;
2978                 page_size   = 1UL << (handle->data->data_order + PAGE_SHIFT);
2979                 page_offset = offset & (page_size - 1);
2980                 size        = min_t(unsigned int, page_size - page_offset, len);
2981
2982                 memcpy(pages[nr] + page_offset, buf, size);
2983
2984                 len         -= size;
2985                 buf         += size;
2986                 offset      += size;
2987         } while (len);
2988
2989         handle->offset = offset;
2990
2991         /*
2992          * Check we didn't copy past our reservation window, taking the
2993          * possible unsigned int wrap into account.
2994          */
2995         WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2996 }
2997
2998 int perf_output_begin(struct perf_output_handle *handle,
2999                       struct perf_event *event, unsigned int size,
3000                       int nmi, int sample)
3001 {
3002         struct perf_event *output_event;
3003         struct perf_mmap_data *data;
3004         unsigned long tail, offset, head;
3005         int have_lost;
3006         struct {
3007                 struct perf_event_header header;
3008                 u64                      id;
3009                 u64                      lost;
3010         } lost_event;
3011
3012         rcu_read_lock();
3013         /*
3014          * For inherited events we send all the output towards the parent.
3015          */
3016         if (event->parent)
3017                 event = event->parent;
3018
3019         output_event = rcu_dereference(event->output);
3020         if (output_event)
3021                 event = output_event;
3022
3023         data = rcu_dereference(event->data);
3024         if (!data)
3025                 goto out;
3026
3027         handle->data    = data;
3028         handle->event   = event;
3029         handle->nmi     = nmi;
3030         handle->sample  = sample;
3031
3032         if (!data->nr_pages)
3033                 goto out;
3034
3035         have_lost = local_read(&data->lost);
3036         if (have_lost)
3037                 size += sizeof(lost_event);
3038
3039         perf_output_get_handle(handle);
3040
3041         do {
3042                 /*
3043                  * Userspace could choose to issue a mb() before updating the
3044                  * tail pointer. So that all reads will be completed before the
3045                  * write is issued.
3046                  */
3047                 tail = ACCESS_ONCE(data->user_page->data_tail);
3048                 smp_rmb();
3049                 offset = head = local_read(&data->head);
3050                 head += size;
3051                 if (unlikely(!perf_output_space(data, tail, offset, head)))
3052                         goto fail;
3053         } while (local_cmpxchg(&data->head, offset, head) != offset);
3054
3055         handle->offset  = offset;
3056         handle->head    = head;
3057
3058         if (head - tail > data->watermark)
3059                 local_inc(&data->wakeup);
3060
3061         if (have_lost) {
3062                 lost_event.header.type = PERF_RECORD_LOST;
3063                 lost_event.header.misc = 0;
3064                 lost_event.header.size = sizeof(lost_event);
3065                 lost_event.id          = event->id;
3066                 lost_event.lost        = local_xchg(&data->lost, 0);
3067
3068                 perf_output_put(handle, lost_event);
3069         }
3070
3071         return 0;
3072
3073 fail:
3074         local_inc(&data->lost);
3075         perf_output_put_handle(handle);
3076 out:
3077         rcu_read_unlock();
3078
3079         return -ENOSPC;
3080 }
3081
3082 void perf_output_end(struct perf_output_handle *handle)
3083 {
3084         struct perf_event *event = handle->event;
3085         struct perf_mmap_data *data = handle->data;
3086
3087         int wakeup_events = event->attr.wakeup_events;
3088
3089         if (handle->sample && wakeup_events) {
3090                 int events = local_inc_return(&data->events);
3091                 if (events >= wakeup_events) {
3092                         local_sub(wakeup_events, &data->events);
3093                         local_inc(&data->wakeup);
3094                 }
3095         }
3096
3097         perf_output_put_handle(handle);
3098         rcu_read_unlock();
3099 }
3100
3101 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3102 {
3103         /*
3104          * only top level events have the pid namespace they were created in
3105          */
3106         if (event->parent)
3107                 event = event->parent;
3108
3109         return task_tgid_nr_ns(p, event->ns);
3110 }
3111
3112 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3113 {
3114         /*
3115          * only top level events have the pid namespace they were created in
3116          */
3117         if (event->parent)
3118                 event = event->parent;
3119
3120         return task_pid_nr_ns(p, event->ns);
3121 }
3122
3123 static void perf_output_read_one(struct perf_output_handle *handle,
3124                                  struct perf_event *event)
3125 {
3126         u64 read_format = event->attr.read_format;
3127         u64 values[4];
3128         int n = 0;
3129
3130         values[n++] = atomic64_read(&event->count);
3131         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3132                 values[n++] = event->total_time_enabled +
3133                         atomic64_read(&event->child_total_time_enabled);
3134         }
3135         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3136                 values[n++] = event->total_time_running +
3137                         atomic64_read(&event->child_total_time_running);
3138         }
3139         if (read_format & PERF_FORMAT_ID)
3140                 values[n++] = primary_event_id(event);
3141
3142         perf_output_copy(handle, values, n * sizeof(u64));
3143 }
3144
3145 /*
3146  * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3147  */
3148 static void perf_output_read_group(struct perf_output_handle *handle,
3149                             struct perf_event *event)
3150 {
3151         struct perf_event *leader = event->group_leader, *sub;
3152         u64 read_format = event->attr.read_format;
3153         u64 values[5];
3154         int n = 0;
3155
3156         values[n++] = 1 + leader->nr_siblings;
3157
3158         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3159                 values[n++] = leader->total_time_enabled;
3160
3161         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3162                 values[n++] = leader->total_time_running;
3163
3164         if (leader != event)
3165                 leader->pmu->read(leader);
3166
3167         values[n++] = atomic64_read(&leader->count);
3168         if (read_format & PERF_FORMAT_ID)
3169                 values[n++] = primary_event_id(leader);
3170
3171         perf_output_copy(handle, values, n * sizeof(u64));
3172
3173         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3174                 n = 0;
3175
3176                 if (sub != event)
3177                         sub->pmu->read(sub);
3178
3179                 values[n++] = atomic64_read(&sub->count);
3180                 if (read_format & PERF_FORMAT_ID)
3181                         values[n++] = primary_event_id(sub);
3182
3183                 perf_output_copy(handle, values, n * sizeof(u64));
3184         }
3185 }
3186
3187 static void perf_output_read(struct perf_output_handle *handle,
3188                              struct perf_event *event)
3189 {
3190         if (event->attr.read_format & PERF_FORMAT_GROUP)
3191                 perf_output_read_group(handle, event);
3192         else
3193                 perf_output_read_one(handle, event);
3194 }
3195
3196 void perf_output_sample(struct perf_output_handle *handle,
3197                         struct perf_event_header *header,
3198                         struct perf_sample_data *data,
3199                         struct perf_event *event)
3200 {
3201         u64 sample_type = data->type;
3202
3203         perf_output_put(handle, *header);
3204
3205         if (sample_type & PERF_SAMPLE_IP)
3206                 perf_output_put(handle, data->ip);
3207
3208         if (sample_type & PERF_SAMPLE_TID)
3209                 perf_output_put(handle, data->tid_entry);
3210
3211         if (sample_type & PERF_SAMPLE_TIME)
3212                 perf_output_put(handle, data->time);
3213
3214         if (sample_type & PERF_SAMPLE_ADDR)
3215                 perf_output_put(handle, data->addr);
3216
3217         if (sample_type & PERF_SAMPLE_ID)
3218                 perf_output_put(handle, data->id);
3219
3220         if (sample_type & PERF_SAMPLE_STREAM_ID)
3221                 perf_output_put(handle, data->stream_id);
3222
3223         if (sample_type & PERF_SAMPLE_CPU)
3224                 perf_output_put(handle, data->cpu_entry);
3225
3226         if (sample_type & PERF_SAMPLE_PERIOD)
3227                 perf_output_put(handle, data->period);
3228
3229         if (sample_type & PERF_SAMPLE_READ)
3230                 perf_output_read(handle, event);
3231
3232         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3233                 if (data->callchain) {
3234                         int size = 1;
3235
3236                         if (data->callchain)
3237                                 size += data->callchain->nr;
3238
3239                         size *= sizeof(u64);
3240
3241                         perf_output_copy(handle, data->callchain, size);
3242                 } else {
3243                         u64 nr = 0;
3244                         perf_output_put(handle, nr);
3245                 }
3246         }
3247
3248         if (sample_type & PERF_SAMPLE_RAW) {
3249                 if (data->raw) {
3250                         perf_output_put(handle, data->raw->size);
3251                         perf_output_copy(handle, data->raw->data,
3252                                          data->raw->size);
3253                 } else {
3254                         struct {
3255                                 u32     size;
3256                                 u32     data;
3257                         } raw = {
3258                                 .size = sizeof(u32),
3259                                 .data = 0,
3260                         };
3261                         perf_output_put(handle, raw);
3262                 }
3263         }
3264 }
3265
3266 void perf_prepare_sample(struct perf_event_header *header,
3267                          struct perf_sample_data *data,
3268                          struct perf_event *event,
3269                          struct pt_regs *regs)
3270 {
3271         u64 sample_type = event->attr.sample_type;
3272
3273         data->type = sample_type;
3274
3275         header->type = PERF_RECORD_SAMPLE;
3276         header->size = sizeof(*header);
3277
3278         header->misc = 0;
3279         header->misc |= perf_misc_flags(regs);
3280
3281         if (sample_type & PERF_SAMPLE_IP) {
3282                 data->ip = perf_instruction_pointer(regs);
3283
3284                 header->size += sizeof(data->ip);
3285         }
3286
3287         if (sample_type & PERF_SAMPLE_TID) {
3288                 /* namespace issues */
3289                 data->tid_entry.pid = perf_event_pid(event, current);
3290                 data->tid_entry.tid = perf_event_tid(event, current);
3291
3292                 header->size += sizeof(data->tid_entry);
3293         }
3294
3295         if (sample_type & PERF_SAMPLE_TIME) {
3296                 data->time = perf_clock();
3297
3298                 header->size += sizeof(data->time);
3299         }
3300
3301         if (sample_type & PERF_SAMPLE_ADDR)
3302                 header->size += sizeof(data->addr);
3303
3304         if (sample_type & PERF_SAMPLE_ID) {
3305                 data->id = primary_event_id(event);
3306
3307                 header->size += sizeof(data->id);
3308         }
3309
3310         if (sample_type & PERF_SAMPLE_STREAM_ID) {
3311                 data->stream_id = event->id;
3312
3313                 header->size += sizeof(data->stream_id);
3314         }
3315
3316         if (sample_type & PERF_SAMPLE_CPU) {
3317                 data->cpu_entry.cpu             = raw_smp_processor_id();
3318                 data->cpu_entry.reserved        = 0;
3319
3320                 header->size += sizeof(data->cpu_entry);
3321         }
3322
3323         if (sample_type & PERF_SAMPLE_PERIOD)
3324                 header->size += sizeof(data->period);
3325
3326         if (sample_type & PERF_SAMPLE_READ)
3327                 header->size += perf_event_read_size(event);
3328
3329         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3330                 int size = 1;
3331
3332                 data->callchain = perf_callchain(regs);
3333
3334                 if (data->callchain)
3335                         size += data->callchain->nr;
3336
3337                 header->size += size * sizeof(u64);
3338         }
3339
3340         if (sample_type & PERF_SAMPLE_RAW) {
3341                 int size = sizeof(u32);
3342
3343                 if (data->raw)
3344                         size += data->raw->size;
3345                 else
3346                         size += sizeof(u32);
3347
3348                 WARN_ON_ONCE(size & (sizeof(u64)-1));
3349                 header->size += size;
3350         }
3351 }
3352
3353 static void perf_event_output(struct perf_event *event, int nmi,
3354                                 struct perf_sample_data *data,
3355                                 struct pt_regs *regs)
3356 {
3357         struct perf_output_handle handle;
3358         struct perf_event_header header;
3359
3360         perf_prepare_sample(&header, data, event, regs);
3361
3362         if (perf_output_begin(&handle, event, header.size, nmi, 1))
3363                 return;
3364
3365         perf_output_sample(&handle, &header, data, event);
3366
3367         perf_output_end(&handle);
3368 }
3369
3370 /*
3371  * read event_id
3372  */
3373
3374 struct perf_read_event {
3375         struct perf_event_header        header;
3376
3377         u32                             pid;
3378         u32                             tid;
3379 };
3380
3381 static void
3382 perf_event_read_event(struct perf_event *event,
3383                         struct task_struct *task)
3384 {
3385         struct perf_output_handle handle;
3386         struct perf_read_event read_event = {
3387                 .header = {
3388                         .type = PERF_RECORD_READ,
3389                         .misc = 0,
3390                         .size = sizeof(read_event) + perf_event_read_size(event),
3391                 },
3392                 .pid = perf_event_pid(event, task),
3393                 .tid = perf_event_tid(event, task),
3394         };
3395         int ret;
3396
3397         ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3398         if (ret)
3399                 return;
3400
3401         perf_output_put(&handle, read_event);
3402         perf_output_read(&handle, event);
3403
3404         perf_output_end(&handle);
3405 }
3406
3407 /*
3408  * task tracking -- fork/exit
3409  *
3410  * enabled by: attr.comm | attr.mmap | attr.task
3411  */
3412
3413 struct perf_task_event {
3414         struct task_struct              *task;
3415         struct perf_event_context       *task_ctx;
3416
3417         struct {
3418                 struct perf_event_header        header;
3419
3420                 u32                             pid;
3421                 u32                             ppid;
3422                 u32                             tid;
3423                 u32                             ptid;
3424                 u64                             time;
3425         } event_id;
3426 };
3427
3428 static void perf_event_task_output(struct perf_event *event,
3429                                      struct perf_task_event *task_event)
3430 {
3431         struct perf_output_handle handle;
3432         struct task_struct *task = task_event->task;
3433         int size, ret;
3434
3435         size  = task_event->event_id.header.size;
3436         ret = perf_output_begin(&handle, event, size, 0, 0);
3437
3438         if (ret)
3439                 return;
3440
3441         task_event->event_id.pid = perf_event_pid(event, task);
3442         task_event->event_id.ppid = perf_event_pid(event, current);
3443
3444         task_event->event_id.tid = perf_event_tid(event, task);
3445         task_event->event_id.ptid = perf_event_tid(event, current);
3446
3447         perf_output_put(&handle, task_event->event_id);
3448
3449         perf_output_end(&handle);
3450 }
3451
3452 static int perf_event_task_match(struct perf_event *event)
3453 {
3454         if (event->state < PERF_EVENT_STATE_INACTIVE)
3455                 return 0;
3456
3457         if (event->cpu != -1 && event->cpu != smp_processor_id())
3458                 return 0;
3459
3460         if (event->attr.comm || event->attr.mmap || event->attr.task)
3461                 return 1;
3462
3463         return 0;
3464 }
3465
3466 static void perf_event_task_ctx(struct perf_event_context *ctx,
3467                                   struct perf_task_event *task_event)
3468 {
3469         struct perf_event *event;
3470
3471         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3472                 if (perf_event_task_match(event))
3473                         perf_event_task_output(event, task_event);
3474         }
3475 }
3476
3477 static void perf_event_task_event(struct perf_task_event *task_event)
3478 {
3479         struct perf_cpu_context *cpuctx;
3480         struct perf_event_context *ctx = task_event->task_ctx;
3481
3482         rcu_read_lock();
3483         cpuctx = &get_cpu_var(perf_cpu_context);
3484         perf_event_task_ctx(&cpuctx->ctx, task_event);
3485         if (!ctx)
3486                 ctx = rcu_dereference(current->perf_event_ctxp);
3487         if (ctx)
3488                 perf_event_task_ctx(ctx, task_event);
3489         put_cpu_var(perf_cpu_context);
3490         rcu_read_unlock();
3491 }
3492
3493 static void perf_event_task(struct task_struct *task,
3494                               struct perf_event_context *task_ctx,
3495                               int new)
3496 {
3497         struct perf_task_event task_event;
3498
3499         if (!atomic_read(&nr_comm_events) &&
3500             !atomic_read(&nr_mmap_events) &&
3501             !atomic_read(&nr_task_events))
3502                 return;
3503
3504         task_event = (struct perf_task_event){
3505                 .task     = task,
3506                 .task_ctx = task_ctx,
3507                 .event_id    = {
3508                         .header = {
3509                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3510                                 .misc = 0,
3511                                 .size = sizeof(task_event.event_id),
3512                         },
3513                         /* .pid  */
3514                         /* .ppid */
3515                         /* .tid  */
3516                         /* .ptid */
3517                         .time = perf_clock(),
3518                 },
3519         };
3520
3521         perf_event_task_event(&task_event);
3522 }
3523
3524 void perf_event_fork(struct task_struct *task)
3525 {
3526         perf_event_task(task, NULL, 1);
3527 }
3528
3529 /*
3530  * comm tracking
3531  */
3532
3533 struct perf_comm_event {
3534         struct task_struct      *task;
3535         char                    *comm;
3536         int                     comm_size;
3537
3538         struct {
3539                 struct perf_event_header        header;
3540
3541                 u32                             pid;
3542                 u32                             tid;
3543         } event_id;
3544 };
3545
3546 static void perf_event_comm_output(struct perf_event *event,
3547                                      struct perf_comm_event *comm_event)
3548 {
3549         struct perf_output_handle handle;
3550         int size = comm_event->event_id.header.size;
3551         int ret = perf_output_begin(&handle, event, size, 0, 0);
3552
3553         if (ret)
3554                 return;
3555
3556         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3557         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3558
3559         perf_output_put(&handle, comm_event->event_id);
3560         perf_output_copy(&handle, comm_event->comm,
3561                                    comm_event->comm_size);
3562         perf_output_end(&handle);
3563 }
3564
3565 static int perf_event_comm_match(struct perf_event *event)
3566 {
3567         if (event->state < PERF_EVENT_STATE_INACTIVE)
3568                 return 0;
3569
3570         if (event->cpu != -1 && event->cpu != smp_processor_id())
3571                 return 0;
3572
3573         if (event->attr.comm)
3574                 return 1;
3575
3576         return 0;
3577 }
3578
3579 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3580                                   struct perf_comm_event *comm_event)
3581 {
3582         struct perf_event *event;
3583
3584         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3585                 if (perf_event_comm_match(event))
3586                         perf_event_comm_output(event, comm_event);
3587         }
3588 }
3589
3590 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3591 {
3592         struct perf_cpu_context *cpuctx;
3593         struct perf_event_context *ctx;
3594         unsigned int size;
3595         char comm[TASK_COMM_LEN];
3596
3597         memset(comm, 0, sizeof(comm));
3598         strlcpy(comm, comm_event->task->comm, sizeof(comm));
3599         size = ALIGN(strlen(comm)+1, sizeof(u64));
3600
3601         comm_event->comm = comm;
3602         comm_event->comm_size = size;
3603
3604         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3605
3606         rcu_read_lock();
3607         cpuctx = &get_cpu_var(perf_cpu_context);
3608         perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3609         ctx = rcu_dereference(current->perf_event_ctxp);
3610         if (ctx)
3611                 perf_event_comm_ctx(ctx, comm_event);
3612         put_cpu_var(perf_cpu_context);
3613         rcu_read_unlock();
3614 }
3615
3616 void perf_event_comm(struct task_struct *task)
3617 {
3618         struct perf_comm_event comm_event;
3619
3620         if (task->perf_event_ctxp)
3621                 perf_event_enable_on_exec(task);
3622
3623         if (!atomic_read(&nr_comm_events))
3624                 return;
3625
3626         comm_event = (struct perf_comm_event){
3627                 .task   = task,
3628                 /* .comm      */
3629                 /* .comm_size */
3630                 .event_id  = {
3631                         .header = {
3632                                 .type = PERF_RECORD_COMM,
3633                                 .misc = 0,
3634                                 /* .size */
3635                         },
3636                         /* .pid */
3637                         /* .tid */
3638                 },
3639         };
3640
3641         perf_event_comm_event(&comm_event);
3642 }
3643
3644 /*
3645  * mmap tracking
3646  */
3647
3648 struct perf_mmap_event {
3649         struct vm_area_struct   *vma;
3650
3651         const char              *file_name;
3652         int                     file_size;
3653
3654         struct {
3655                 struct perf_event_header        header;
3656
3657                 u32                             pid;
3658                 u32                             tid;
3659                 u64                             start;
3660                 u64                             len;
3661                 u64                             pgoff;
3662         } event_id;
3663 };
3664
3665 static void perf_event_mmap_output(struct perf_event *event,
3666                                      struct perf_mmap_event *mmap_event)
3667 {
3668         struct perf_output_handle handle;
3669         int size = mmap_event->event_id.header.size;
3670         int ret = perf_output_begin(&handle, event, size, 0, 0);
3671
3672         if (ret)
3673                 return;
3674
3675         mmap_event->event_id.pid = perf_event_pid(event, current);
3676         mmap_event->event_id.tid = perf_event_tid(event, current);
3677
3678         perf_output_put(&handle, mmap_event->event_id);
3679         perf_output_copy(&handle, mmap_event->file_name,
3680                                    mmap_event->file_size);
3681         perf_output_end(&handle);
3682 }
3683
3684 static int perf_event_mmap_match(struct perf_event *event,
3685                                    struct perf_mmap_event *mmap_event)
3686 {
3687         if (event->state < PERF_EVENT_STATE_INACTIVE)
3688                 return 0;
3689
3690         if (event->cpu != -1 && event->cpu != smp_processor_id())
3691                 return 0;
3692
3693         if (event->attr.mmap)
3694                 return 1;
3695
3696         return 0;
3697 }
3698
3699 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3700                                   struct perf_mmap_event *mmap_event)
3701 {
3702         struct perf_event *event;
3703
3704         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3705                 if (perf_event_mmap_match(event, mmap_event))
3706                         perf_event_mmap_output(event, mmap_event);
3707         }
3708 }
3709
3710 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3711 {
3712         struct perf_cpu_context *cpuctx;
3713         struct perf_event_context *ctx;
3714         struct vm_area_struct *vma = mmap_event->vma;
3715         struct file *file = vma->vm_file;
3716         unsigned int size;
3717         char tmp[16];
3718         char *buf = NULL;
3719         const char *name;
3720
3721         memset(tmp, 0, sizeof(tmp));
3722
3723         if (file) {
3724                 /*
3725                  * d_path works from the end of the buffer backwards, so we
3726                  * need to add enough zero bytes after the string to handle
3727                  * the 64bit alignment we do later.
3728                  */
3729                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3730                 if (!buf) {
3731                         name = strncpy(tmp, "//enomem", sizeof(tmp));
3732                         goto got_name;
3733                 }
3734                 name = d_path(&file->f_path, buf, PATH_MAX);
3735                 if (IS_ERR(name)) {
3736                         name = strncpy(tmp, "//toolong", sizeof(tmp));
3737                         goto got_name;
3738                 }
3739         } else {
3740                 if (arch_vma_name(mmap_event->vma)) {
3741                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3742                                        sizeof(tmp));
3743                         goto got_name;
3744                 }
3745
3746                 if (!vma->vm_mm) {
3747                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
3748                         goto got_name;
3749                 }
3750
3751                 name = strncpy(tmp, "//anon", sizeof(tmp));
3752                 goto got_name;
3753         }
3754
3755 got_name:
3756         size = ALIGN(strlen(name)+1, sizeof(u64));
3757
3758         mmap_event->file_name = name;
3759         mmap_event->file_size = size;
3760
3761         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3762
3763         rcu_read_lock();
3764         cpuctx = &get_cpu_var(perf_cpu_context);
3765         perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3766         ctx = rcu_dereference(current->perf_event_ctxp);
3767         if (ctx)
3768                 perf_event_mmap_ctx(ctx, mmap_event);
3769         put_cpu_var(perf_cpu_context);
3770         rcu_read_unlock();
3771
3772         kfree(buf);
3773 }
3774
3775 void __perf_event_mmap(struct vm_area_struct *vma)
3776 {
3777         struct perf_mmap_event mmap_event;
3778
3779         if (!atomic_read(&nr_mmap_events))
3780                 return;
3781
3782         mmap_event = (struct perf_mmap_event){
3783                 .vma    = vma,
3784                 /* .file_name */
3785                 /* .file_size */
3786                 .event_id  = {
3787                         .header = {
3788                                 .type = PERF_RECORD_MMAP,
3789                                 .misc = PERF_RECORD_MISC_USER,
3790                                 /* .size */
3791                         },
3792                         /* .pid */
3793                         /* .tid */
3794                         .start  = vma->vm_start,
3795                         .len    = vma->vm_end - vma->vm_start,
3796                         .pgoff  = (u64)vma->vm_pgoff << PAGE_SHIFT,
3797                 },
3798         };
3799
3800         perf_event_mmap_event(&mmap_event);
3801 }
3802
3803 /*
3804  * IRQ throttle logging
3805  */
3806
3807 static void perf_log_throttle(struct perf_event *event, int enable)
3808 {
3809         struct perf_output_handle handle;
3810         int ret;
3811
3812         struct {
3813                 struct perf_event_header        header;
3814                 u64                             time;
3815                 u64                             id;
3816                 u64                             stream_id;
3817         } throttle_event = {
3818                 .header = {
3819                         .type = PERF_RECORD_THROTTLE,
3820                         .misc = 0,
3821                         .size = sizeof(throttle_event),
3822                 },
3823                 .time           = perf_clock(),
3824                 .id             = primary_event_id(event),
3825                 .stream_id      = event->id,
3826         };
3827
3828         if (enable)
3829                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3830
3831         ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3832         if (ret)
3833                 return;
3834
3835         perf_output_put(&handle, throttle_event);
3836         perf_output_end(&handle);
3837 }
3838
3839 /*
3840  * Generic event overflow handling, sampling.
3841  */
3842
3843 static int __perf_event_overflow(struct perf_event *event, int nmi,
3844                                    int throttle, struct perf_sample_data *data,
3845                                    struct pt_regs *regs)
3846 {
3847         int events = atomic_read(&event->event_limit);
3848         struct hw_perf_event *hwc = &event->hw;
3849         int ret = 0;
3850
3851         throttle = (throttle && event->pmu->unthrottle != NULL);
3852
3853         if (!throttle) {
3854                 hwc->interrupts++;
3855         } else {
3856                 if (hwc->interrupts != MAX_INTERRUPTS) {
3857                         hwc->interrupts++;
3858                         if (HZ * hwc->interrupts >
3859                                         (u64)sysctl_perf_event_sample_rate) {
3860                                 hwc->interrupts = MAX_INTERRUPTS;
3861                                 perf_log_throttle(event, 0);
3862                                 ret = 1;
3863                         }
3864                 } else {
3865                         /*
3866                          * Keep re-disabling events even though on the previous
3867                          * pass we disabled it - just in case we raced with a
3868                          * sched-in and the event got enabled again:
3869                          */
3870                         ret = 1;
3871                 }
3872         }
3873
3874         if (event->attr.freq) {
3875                 u64 now = perf_clock();
3876                 s64 delta = now - hwc->freq_time_stamp;
3877
3878                 hwc->freq_time_stamp = now;
3879
3880                 if (delta > 0 && delta < 2*TICK_NSEC)
3881                         perf_adjust_period(event, delta, hwc->last_period);
3882         }
3883
3884         /*
3885          * XXX event_limit might not quite work as expected on inherited
3886          * events
3887          */
3888
3889         event->pending_kill = POLL_IN;
3890         if (events && atomic_dec_and_test(&event->event_limit)) {
3891                 ret = 1;
3892                 event->pending_kill = POLL_HUP;
3893                 if (nmi) {
3894                         event->pending_disable = 1;
3895                         perf_pending_queue(&event->pending,
3896                                            perf_pending_event);
3897                 } else
3898                         perf_event_disable(event);
3899         }
3900
3901         if (event->overflow_handler)
3902                 event->overflow_handler(event, nmi, data, regs);
3903         else
3904                 perf_event_output(event, nmi, data, regs);
3905
3906         return ret;
3907 }
3908
3909 int perf_event_overflow(struct perf_event *event, int nmi,
3910                           struct perf_sample_data *data,
3911                           struct pt_regs *regs)
3912 {
3913         return __perf_event_overflow(event, nmi, 1, data, regs);
3914 }
3915
3916 /*
3917  * Generic software event infrastructure
3918  */
3919
3920 /*
3921  * We directly increment event->count and keep a second value in
3922  * event->hw.period_left to count intervals. This period event
3923  * is kept in the range [-sample_period, 0] so that we can use the
3924  * sign as trigger.
3925  */
3926
3927 static u64 perf_swevent_set_period(struct perf_event *event)
3928 {
3929         struct hw_perf_event *hwc = &event->hw;
3930         u64 period = hwc->last_period;
3931         u64 nr, offset;
3932         s64 old, val;
3933
3934         hwc->last_period = hwc->sample_period;
3935
3936 again:
3937         old = val = atomic64_read(&hwc->period_left);
3938         if (val < 0)
3939                 return 0;
3940
3941         nr = div64_u64(period + val, period);
3942         offset = nr * period;
3943         val -= offset;
3944         if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3945                 goto again;
3946
3947         return nr;
3948 }
3949
3950 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3951                                     int nmi, struct perf_sample_data *data,
3952                                     struct pt_regs *regs)
3953 {
3954         struct hw_perf_event *hwc = &event->hw;
3955         int throttle = 0;
3956
3957         data->period = event->hw.last_period;
3958         if (!overflow)
3959                 overflow = perf_swevent_set_period(event);
3960
3961         if (hwc->interrupts == MAX_INTERRUPTS)
3962                 return;
3963
3964         for (; overflow; overflow--) {
3965                 if (__perf_event_overflow(event, nmi, throttle,
3966                                             data, regs)) {
3967                         /*
3968                          * We inhibit the overflow from happening when
3969                          * hwc->interrupts == MAX_INTERRUPTS.
3970                          */
3971                         break;
3972                 }
3973                 throttle = 1;
3974         }
3975 }
3976
3977 static void perf_swevent_unthrottle(struct perf_event *event)
3978 {
3979         /*
3980          * Nothing to do, we already reset hwc->interrupts.
3981          */
3982 }
3983
3984 static void perf_swevent_add(struct perf_event *event, u64 nr,
3985                                int nmi, struct perf_sample_data *data,
3986                                struct pt_regs *regs)
3987 {
3988         struct hw_perf_event *hwc = &event->hw;
3989
3990         atomic64_add(nr, &event->count);
3991
3992         if (!regs)
3993                 return;
3994
3995         if (!hwc->sample_period)
3996                 return;
3997
3998         if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3999                 return perf_swevent_overflow(event, 1, nmi, data, regs);
4000
4001         if (atomic64_add_negative(nr, &hwc->period_left))
4002                 return;
4003
4004         perf_swevent_overflow(event, 0, nmi, data, regs);
4005 }
4006
4007 static int perf_tp_event_match(struct perf_event *event,
4008                                 struct perf_sample_data *data);
4009
4010 static int perf_exclude_event(struct perf_event *event,
4011                               struct pt_regs *regs)
4012 {
4013         if (regs) {
4014                 if (event->attr.exclude_user && user_mode(regs))
4015                         return 1;
4016
4017                 if (event->attr.exclude_kernel && !user_mode(regs))
4018                         return 1;
4019         }
4020
4021         return 0;
4022 }
4023
4024 static int perf_swevent_match(struct perf_event *event,
4025                                 enum perf_type_id type,
4026                                 u32 event_id,
4027                                 struct perf_sample_data *data,
4028                                 struct pt_regs *regs)
4029 {
4030         if (event->attr.type != type)
4031                 return 0;
4032
4033         if (event->attr.config != event_id)
4034                 return 0;
4035
4036         if (perf_exclude_event(event, regs))
4037                 return 0;
4038
4039         if (event->attr.type == PERF_TYPE_TRACEPOINT &&
4040             !perf_tp_event_match(event, data))
4041                 return 0;
4042
4043         return 1;
4044 }
4045
4046 static inline u64 swevent_hash(u64 type, u32 event_id)
4047 {
4048         u64 val = event_id | (type << 32);
4049
4050         return hash_64(val, SWEVENT_HLIST_BITS);
4051 }
4052
4053 static inline struct hlist_head *
4054 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4055 {
4056         u64 hash = swevent_hash(type, event_id);
4057
4058         return &hlist->heads[hash];
4059 }
4060
4061 /* For the read side: events when they trigger */
4062 static inline struct hlist_head *
4063 find_swevent_head_rcu(struct perf_cpu_context *ctx, u64 type, u32 event_id)
4064 {
4065         struct swevent_hlist *hlist;
4066
4067         hlist = rcu_dereference(ctx->swevent_hlist);
4068         if (!hlist)
4069                 return NULL;
4070
4071         return __find_swevent_head(hlist, type, event_id);
4072 }
4073
4074 /* For the event head insertion and removal in the hlist */
4075 static inline struct hlist_head *
4076 find_swevent_head(struct perf_cpu_context *ctx, struct perf_event *event)
4077 {
4078         struct swevent_hlist *hlist;
4079         u32 event_id = event->attr.config;
4080         u64 type = event->attr.type;
4081
4082         /*
4083          * Event scheduling is always serialized against hlist allocation
4084          * and release. Which makes the protected version suitable here.
4085          * The context lock guarantees that.
4086          */
4087         hlist = rcu_dereference_protected(ctx->swevent_hlist,
4088                                           lockdep_is_held(&event->ctx->lock));
4089         if (!hlist)
4090                 return NULL;
4091
4092         return __find_swevent_head(hlist, type, event_id);
4093 }
4094
4095 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4096                                     u64 nr, int nmi,
4097                                     struct perf_sample_data *data,
4098                                     struct pt_regs *regs)
4099 {
4100         struct perf_cpu_context *cpuctx;
4101         struct perf_event *event;
4102         struct hlist_node *node;
4103         struct hlist_head *head;
4104
4105         cpuctx = &__get_cpu_var(perf_cpu_context);
4106
4107         rcu_read_lock();
4108
4109         head = find_swevent_head_rcu(cpuctx, type, event_id);
4110
4111         if (!head)
4112                 goto end;
4113
4114         hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4115                 if (perf_swevent_match(event, type, event_id, data, regs))
4116                         perf_swevent_add(event, nr, nmi, data, regs);
4117         }
4118 end:
4119         rcu_read_unlock();
4120 }
4121
4122 int perf_swevent_get_recursion_context(void)
4123 {
4124         struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
4125         int rctx;
4126
4127         if (in_nmi())
4128                 rctx = 3;
4129         else if (in_irq())
4130                 rctx = 2;
4131         else if (in_softirq())
4132                 rctx = 1;
4133         else
4134                 rctx = 0;
4135
4136         if (cpuctx->recursion[rctx]) {
4137                 put_cpu_var(perf_cpu_context);
4138                 return -1;
4139         }
4140
4141         cpuctx->recursion[rctx]++;
4142         barrier();
4143
4144         return rctx;
4145 }
4146 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4147
4148 void perf_swevent_put_recursion_context(int rctx)
4149 {
4150         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4151         barrier();
4152         cpuctx->recursion[rctx]--;
4153         put_cpu_var(perf_cpu_context);
4154 }
4155 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
4156
4157
4158 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4159                             struct pt_regs *regs, u64 addr)
4160 {
4161         struct perf_sample_data data;
4162         int rctx;
4163
4164         rctx = perf_swevent_get_recursion_context();
4165         if (rctx < 0)
4166                 return;
4167
4168         perf_sample_data_init(&data, addr);
4169
4170         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4171
4172         perf_swevent_put_recursion_context(rctx);
4173 }
4174
4175 static void perf_swevent_read(struct perf_event *event)
4176 {
4177 }
4178
4179 static int perf_swevent_enable(struct perf_event *event)
4180 {
4181         struct hw_perf_event *hwc = &event->hw;
4182         struct perf_cpu_context *cpuctx;
4183         struct hlist_head *head;
4184
4185         cpuctx = &__get_cpu_var(perf_cpu_context);
4186
4187         if (hwc->sample_period) {
4188                 hwc->last_period = hwc->sample_period;
4189                 perf_swevent_set_period(event);
4190         }
4191
4192         head = find_swevent_head(cpuctx, event);
4193         if (WARN_ON_ONCE(!head))
4194                 return -EINVAL;
4195
4196         hlist_add_head_rcu(&event->hlist_entry, head);
4197
4198         return 0;
4199 }
4200
4201 static void perf_swevent_disable(struct perf_event *event)
4202 {
4203         hlist_del_rcu(&event->hlist_entry);
4204 }
4205
4206 static const struct pmu perf_ops_generic = {
4207         .enable         = perf_swevent_enable,
4208         .disable        = perf_swevent_disable,
4209         .read           = perf_swevent_read,
4210         .unthrottle     = perf_swevent_unthrottle,
4211 };
4212
4213 /*
4214  * hrtimer based swevent callback
4215  */
4216
4217 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4218 {
4219         enum hrtimer_restart ret = HRTIMER_RESTART;
4220         struct perf_sample_data data;
4221         struct pt_regs *regs;
4222         struct perf_event *event;
4223         u64 period;
4224
4225         event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4226         event->pmu->read(event);
4227
4228         perf_sample_data_init(&data, 0);
4229         data.period = event->hw.last_period;
4230         regs = get_irq_regs();
4231
4232         if (regs && !perf_exclude_event(event, regs)) {
4233                 if (!(event->attr.exclude_idle && current->pid == 0))
4234                         if (perf_event_overflow(event, 0, &data, regs))
4235                                 ret = HRTIMER_NORESTART;
4236         }
4237
4238         period = max_t(u64, 10000, event->hw.sample_period);
4239         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4240
4241         return ret;
4242 }
4243
4244 static void perf_swevent_start_hrtimer(struct perf_event *event)
4245 {
4246         struct hw_perf_event *hwc = &event->hw;
4247
4248         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4249         hwc->hrtimer.function = perf_swevent_hrtimer;
4250         if (hwc->sample_period) {
4251                 u64 period;
4252
4253                 if (hwc->remaining) {
4254                         if (hwc->remaining < 0)
4255                                 period = 10000;
4256                         else
4257                                 period = hwc->remaining;
4258                         hwc->remaining = 0;
4259                 } else {
4260                         period = max_t(u64, 10000, hwc->sample_period);
4261                 }
4262                 __hrtimer_start_range_ns(&hwc->hrtimer,
4263                                 ns_to_ktime(period), 0,
4264                                 HRTIMER_MODE_REL, 0);
4265         }
4266 }
4267
4268 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4269 {
4270         struct hw_perf_event *hwc = &event->hw;
4271
4272         if (hwc->sample_period) {
4273                 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4274                 hwc->remaining = ktime_to_ns(remaining);
4275
4276                 hrtimer_cancel(&hwc->hrtimer);
4277         }
4278 }
4279
4280 /*
4281  * Software event: cpu wall time clock
4282  */
4283
4284 static void cpu_clock_perf_event_update(struct perf_event *event)
4285 {
4286         int cpu = raw_smp_processor_id();
4287         s64 prev;
4288         u64 now;
4289
4290         now = cpu_clock(cpu);
4291         prev = atomic64_xchg(&event->hw.prev_count, now);
4292         atomic64_add(now - prev, &event->count);
4293 }
4294
4295 static int cpu_clock_perf_event_enable(struct perf_event *event)
4296 {
4297         struct hw_perf_event *hwc = &event->hw;
4298         int cpu = raw_smp_processor_id();
4299
4300         atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4301         perf_swevent_start_hrtimer(event);
4302
4303         return 0;
4304 }
4305
4306 static void cpu_clock_perf_event_disable(struct perf_event *event)
4307 {
4308         perf_swevent_cancel_hrtimer(event);
4309         cpu_clock_perf_event_update(event);
4310 }
4311
4312 static void cpu_clock_perf_event_read(struct perf_event *event)
4313 {
4314         cpu_clock_perf_event_update(event);
4315 }
4316
4317 static const struct pmu perf_ops_cpu_clock = {
4318         .enable         = cpu_clock_perf_event_enable,
4319         .disable        = cpu_clock_perf_event_disable,
4320         .read           = cpu_clock_perf_event_read,
4321 };
4322
4323 /*
4324  * Software event: task time clock
4325  */
4326
4327 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4328 {
4329         u64 prev;
4330         s64 delta;
4331
4332         prev = atomic64_xchg(&event->hw.prev_count, now);
4333         delta = now - prev;
4334         atomic64_add(delta, &event->count);
4335 }
4336
4337 static int task_clock_perf_event_enable(struct perf_event *event)
4338 {
4339         struct hw_perf_event *hwc = &event->hw;
4340         u64 now;
4341
4342         now = event->ctx->time;
4343
4344         atomic64_set(&hwc->prev_count, now);
4345
4346         perf_swevent_start_hrtimer(event);
4347
4348         return 0;
4349 }
4350
4351 static void task_clock_perf_event_disable(struct perf_event *event)
4352 {
4353         perf_swevent_cancel_hrtimer(event);
4354         task_clock_perf_event_update(event, event->ctx->time);
4355
4356 }
4357
4358 static void task_clock_perf_event_read(struct perf_event *event)
4359 {
4360         u64 time;
4361
4362         if (!in_nmi()) {
4363                 update_context_time(event->ctx);
4364                 time = event->ctx->time;
4365         } else {
4366                 u64 now = perf_clock();
4367                 u64 delta = now - event->ctx->timestamp;
4368                 time = event->ctx->time + delta;
4369         }
4370
4371         task_clock_perf_event_update(event, time);
4372 }
4373
4374 static const struct pmu perf_ops_task_clock = {
4375         .enable         = task_clock_perf_event_enable,
4376         .disable        = task_clock_perf_event_disable,
4377         .read           = task_clock_perf_event_read,
4378 };
4379
4380 /* Deref the hlist from the update side */
4381 static inline struct swevent_hlist *
4382 swevent_hlist_deref(struct perf_cpu_context *cpuctx)
4383 {
4384         return rcu_dereference_protected(cpuctx->swevent_hlist,
4385                                          lockdep_is_held(&cpuctx->hlist_mutex));
4386 }
4387
4388 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4389 {
4390         struct swevent_hlist *hlist;
4391
4392         hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4393         kfree(hlist);
4394 }
4395
4396 static void swevent_hlist_release(struct perf_cpu_context *cpuctx)
4397 {
4398         struct swevent_hlist *hlist = swevent_hlist_deref(cpuctx);
4399
4400         if (!hlist)
4401                 return;
4402
4403         rcu_assign_pointer(cpuctx->swevent_hlist, NULL);
4404         call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4405 }
4406
4407 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4408 {
4409         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4410
4411         mutex_lock(&cpuctx->hlist_mutex);
4412
4413         if (!--cpuctx->hlist_refcount)
4414                 swevent_hlist_release(cpuctx);
4415
4416         mutex_unlock(&cpuctx->hlist_mutex);
4417 }
4418
4419 static void swevent_hlist_put(struct perf_event *event)
4420 {
4421         int cpu;
4422
4423         if (event->cpu != -1) {
4424                 swevent_hlist_put_cpu(event, event->cpu);
4425                 return;
4426         }
4427
4428         for_each_possible_cpu(cpu)
4429                 swevent_hlist_put_cpu(event, cpu);
4430 }
4431
4432 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4433 {
4434         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4435         int err = 0;
4436
4437         mutex_lock(&cpuctx->hlist_mutex);
4438
4439         if (!swevent_hlist_deref(cpuctx) && cpu_online(cpu)) {
4440                 struct swevent_hlist *hlist;
4441
4442                 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4443                 if (!hlist) {
4444                         err = -ENOMEM;
4445                         goto exit;
4446                 }
4447                 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
4448         }
4449         cpuctx->hlist_refcount++;
4450  exit:
4451         mutex_unlock(&cpuctx->hlist_mutex);
4452
4453         return err;
4454 }
4455
4456 static int swevent_hlist_get(struct perf_event *event)
4457 {
4458         int err;
4459         int cpu, failed_cpu;
4460
4461         if (event->cpu != -1)
4462                 return swevent_hlist_get_cpu(event, event->cpu);
4463
4464         get_online_cpus();
4465         for_each_possible_cpu(cpu) {
4466                 err = swevent_hlist_get_cpu(event, cpu);
4467                 if (err) {
4468                         failed_cpu = cpu;
4469                         goto fail;
4470                 }
4471         }
4472         put_online_cpus();
4473
4474         return 0;
4475  fail:
4476         for_each_possible_cpu(cpu) {
4477                 if (cpu == failed_cpu)
4478                         break;
4479                 swevent_hlist_put_cpu(event, cpu);
4480         }
4481
4482         put_online_cpus();
4483         return err;
4484 }
4485
4486 #ifdef CONFIG_EVENT_TRACING
4487
4488 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4489                    int entry_size, struct pt_regs *regs, void *event)
4490 {
4491         const int type = PERF_TYPE_TRACEPOINT;
4492         struct perf_sample_data data;
4493         struct perf_raw_record raw = {
4494                 .size = entry_size,
4495                 .data = record,
4496         };
4497
4498         perf_sample_data_init(&data, addr);
4499         data.raw = &raw;
4500
4501         if (!event) {
4502                 do_perf_sw_event(type, event_id, count, 1, &data, regs);
4503                 return;
4504         }
4505
4506         if (perf_swevent_match(event, type, event_id, &data, regs))
4507                 perf_swevent_add(event, count, 1, &data, regs);
4508 }
4509 EXPORT_SYMBOL_GPL(perf_tp_event);
4510
4511 static int perf_tp_event_match(struct perf_event *event,
4512                                 struct perf_sample_data *data)
4513 {
4514         void *record = data->raw->data;
4515
4516         if (likely(!event->filter) || filter_match_preds(event->filter, record))
4517                 return 1;
4518         return 0;
4519 }
4520
4521 static void tp_perf_event_destroy(struct perf_event *event)
4522 {
4523         perf_trace_disable(event->attr.config);
4524         swevent_hlist_put(event);
4525 }
4526
4527 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4528 {
4529         int err;
4530
4531         /*
4532          * Raw tracepoint data is a severe data leak, only allow root to
4533          * have these.
4534          */
4535         if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4536                         perf_paranoid_tracepoint_raw() &&
4537                         !capable(CAP_SYS_ADMIN))
4538                 return ERR_PTR(-EPERM);
4539
4540         if (perf_trace_enable(event->attr.config, event))
4541                 return NULL;
4542
4543         event->destroy = tp_perf_event_destroy;
4544         err = swevent_hlist_get(event);
4545         if (err) {
4546                 perf_trace_disable(event->attr.config);
4547                 return ERR_PTR(err);
4548         }
4549
4550         return &perf_ops_generic;
4551 }
4552
4553 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4554 {
4555         char *filter_str;
4556         int ret;
4557
4558         if (event->attr.type != PERF_TYPE_TRACEPOINT)
4559                 return -EINVAL;
4560
4561         filter_str = strndup_user(arg, PAGE_SIZE);
4562         if (IS_ERR(filter_str))
4563                 return PTR_ERR(filter_str);
4564
4565         ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4566
4567         kfree(filter_str);
4568         return ret;
4569 }
4570
4571 static void perf_event_free_filter(struct perf_event *event)
4572 {
4573         ftrace_profile_free_filter(event);
4574 }
4575
4576 #else
4577
4578 static int perf_tp_event_match(struct perf_event *event,
4579                                 struct perf_sample_data *data)
4580 {
4581         return 1;
4582 }
4583
4584 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4585 {
4586         return NULL;
4587 }
4588
4589 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4590 {
4591         return -ENOENT;
4592 }
4593
4594 static void perf_event_free_filter(struct perf_event *event)
4595 {
4596 }
4597
4598 #endif /* CONFIG_EVENT_TRACING */
4599
4600 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4601 static void bp_perf_event_destroy(struct perf_event *event)
4602 {
4603         release_bp_slot(event);
4604 }
4605
4606 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4607 {
4608         int err;
4609
4610         err = register_perf_hw_breakpoint(bp);
4611         if (err)
4612                 return ERR_PTR(err);
4613
4614         bp->destroy = bp_perf_event_destroy;
4615
4616         return &perf_ops_bp;
4617 }
4618
4619 void perf_bp_event(struct perf_event *bp, void *data)
4620 {
4621         struct perf_sample_data sample;
4622         struct pt_regs *regs = data;
4623
4624         perf_sample_data_init(&sample, bp->attr.bp_addr);
4625
4626         if (!perf_exclude_event(bp, regs))
4627                 perf_swevent_add(bp, 1, 1, &sample, regs);
4628 }
4629 #else
4630 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4631 {
4632         return NULL;
4633 }
4634
4635 void perf_bp_event(struct perf_event *bp, void *regs)
4636 {
4637 }
4638 #endif
4639
4640 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4641
4642 static void sw_perf_event_destroy(struct perf_event *event)
4643 {
4644         u64 event_id = event->attr.config;
4645
4646         WARN_ON(event->parent);
4647
4648         atomic_dec(&perf_swevent_enabled[event_id]);
4649         swevent_hlist_put(event);
4650 }
4651
4652 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4653 {
4654         const struct pmu *pmu = NULL;
4655         u64 event_id = event->attr.config;
4656
4657         /*
4658          * Software events (currently) can't in general distinguish
4659          * between user, kernel and hypervisor events.
4660          * However, context switches and cpu migrations are considered
4661          * to be kernel events, and page faults are never hypervisor
4662          * events.
4663          */
4664         switch (event_id) {
4665         case PERF_COUNT_SW_CPU_CLOCK:
4666                 pmu = &perf_ops_cpu_clock;
4667
4668                 break;
4669         case PERF_COUNT_SW_TASK_CLOCK:
4670                 /*
4671                  * If the user instantiates this as a per-cpu event,
4672                  * use the cpu_clock event instead.
4673                  */
4674                 if (event->ctx->task)
4675                         pmu = &perf_ops_task_clock;
4676                 else
4677                         pmu = &perf_ops_cpu_clock;
4678
4679                 break;
4680         case PERF_COUNT_SW_PAGE_FAULTS:
4681         case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4682         case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4683         case PERF_COUNT_SW_CONTEXT_SWITCHES:
4684         case PERF_COUNT_SW_CPU_MIGRATIONS:
4685         case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4686         case PERF_COUNT_SW_EMULATION_FAULTS:
4687                 if (!event->parent) {
4688                         int err;
4689
4690                         err = swevent_hlist_get(event);
4691                         if (err)
4692                                 return ERR_PTR(err);
4693
4694                         atomic_inc(&perf_swevent_enabled[event_id]);
4695                         event->destroy = sw_perf_event_destroy;
4696                 }
4697                 pmu = &perf_ops_generic;
4698                 break;
4699         }
4700
4701         return pmu;
4702 }
4703
4704 /*
4705  * Allocate and initialize a event structure
4706  */
4707 static struct perf_event *
4708 perf_event_alloc(struct perf_event_attr *attr,
4709                    int cpu,
4710                    struct perf_event_context *ctx,
4711                    struct perf_event *group_leader,
4712                    struct perf_event *parent_event,
4713                    perf_overflow_handler_t overflow_handler,
4714                    gfp_t gfpflags)
4715 {
4716         const struct pmu *pmu;
4717         struct perf_event *event;
4718         struct hw_perf_event *hwc;
4719         long err;
4720
4721         event = kzalloc(sizeof(*event), gfpflags);
4722         if (!event)
4723                 return ERR_PTR(-ENOMEM);
4724
4725         /*
4726          * Single events are their own group leaders, with an
4727          * empty sibling list:
4728          */
4729         if (!group_leader)
4730                 group_leader = event;
4731
4732         mutex_init(&event->child_mutex);
4733         INIT_LIST_HEAD(&event->child_list);
4734
4735         INIT_LIST_HEAD(&event->group_entry);
4736         INIT_LIST_HEAD(&event->event_entry);
4737         INIT_LIST_HEAD(&event->sibling_list);
4738         init_waitqueue_head(&event->waitq);
4739
4740         mutex_init(&event->mmap_mutex);
4741
4742         event->cpu              = cpu;
4743         event->attr             = *attr;
4744         event->group_leader     = group_leader;
4745         event->pmu              = NULL;
4746         event->ctx              = ctx;
4747         event->oncpu            = -1;
4748
4749         event->parent           = parent_event;
4750
4751         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
4752         event->id               = atomic64_inc_return(&perf_event_id);
4753
4754         event->state            = PERF_EVENT_STATE_INACTIVE;
4755
4756         if (!overflow_handler && parent_event)
4757                 overflow_handler = parent_event->overflow_handler;
4758         
4759         event->overflow_handler = overflow_handler;
4760
4761         if (attr->disabled)
4762                 event->state = PERF_EVENT_STATE_OFF;
4763
4764         pmu = NULL;
4765
4766         hwc = &event->hw;
4767         hwc->sample_period = attr->sample_period;
4768         if (attr->freq && attr->sample_freq)
4769                 hwc->sample_period = 1;
4770         hwc->last_period = hwc->sample_period;
4771
4772         atomic64_set(&hwc->period_left, hwc->sample_period);
4773
4774         /*
4775          * we currently do not support PERF_FORMAT_GROUP on inherited events
4776          */
4777         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4778                 goto done;
4779
4780         switch (attr->type) {
4781         case PERF_TYPE_RAW:
4782         case PERF_TYPE_HARDWARE:
4783         case PERF_TYPE_HW_CACHE:
4784                 pmu = hw_perf_event_init(event);
4785                 break;
4786
4787         case PERF_TYPE_SOFTWARE:
4788                 pmu = sw_perf_event_init(event);
4789                 break;
4790
4791         case PERF_TYPE_TRACEPOINT:
4792                 pmu = tp_perf_event_init(event);
4793                 break;
4794
4795         case PERF_TYPE_BREAKPOINT:
4796                 pmu = bp_perf_event_init(event);
4797                 break;
4798
4799
4800         default:
4801                 break;
4802         }
4803 done:
4804         err = 0;
4805         if (!pmu)
4806                 err = -EINVAL;
4807         else if (IS_ERR(pmu))
4808                 err = PTR_ERR(pmu);
4809
4810         if (err) {
4811                 if (event->ns)
4812                         put_pid_ns(event->ns);
4813                 kfree(event);
4814                 return ERR_PTR(err);
4815         }
4816
4817         event->pmu = pmu;
4818
4819         if (!event->parent) {
4820                 atomic_inc(&nr_events);
4821                 if (event->attr.mmap)
4822                         atomic_inc(&nr_mmap_events);
4823                 if (event->attr.comm)
4824                         atomic_inc(&nr_comm_events);
4825                 if (event->attr.task)
4826                         atomic_inc(&nr_task_events);
4827         }
4828
4829         return event;
4830 }
4831
4832 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4833                           struct perf_event_attr *attr)
4834 {
4835         u32 size;
4836         int ret;
4837
4838         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4839                 return -EFAULT;
4840
4841         /*
4842          * zero the full structure, so that a short copy will be nice.
4843          */
4844         memset(attr, 0, sizeof(*attr));
4845
4846         ret = get_user(size, &uattr->size);
4847         if (ret)
4848                 return ret;
4849
4850         if (size > PAGE_SIZE)   /* silly large */
4851                 goto err_size;
4852
4853         if (!size)              /* abi compat */
4854                 size = PERF_ATTR_SIZE_VER0;
4855
4856         if (size < PERF_ATTR_SIZE_VER0)
4857                 goto err_size;
4858
4859         /*
4860          * If we're handed a bigger struct than we know of,
4861          * ensure all the unknown bits are 0 - i.e. new
4862          * user-space does not rely on any kernel feature
4863          * extensions we dont know about yet.
4864          */
4865         if (size > sizeof(*attr)) {
4866                 unsigned char __user *addr;
4867                 unsigned char __user *end;
4868                 unsigned char val;
4869
4870                 addr = (void __user *)uattr + sizeof(*attr);
4871                 end  = (void __user *)uattr + size;
4872
4873                 for (; addr < end; addr++) {
4874                         ret = get_user(val, addr);
4875                         if (ret)
4876                                 return ret;
4877                         if (val)
4878                                 goto err_size;
4879                 }
4880                 size = sizeof(*attr);
4881         }
4882
4883         ret = copy_from_user(attr, uattr, size);
4884         if (ret)
4885                 return -EFAULT;
4886
4887         /*
4888          * If the type exists, the corresponding creation will verify
4889          * the attr->config.
4890          */
4891         if (attr->type >= PERF_TYPE_MAX)
4892                 return -EINVAL;
4893
4894         if (attr->__reserved_1)
4895                 return -EINVAL;
4896
4897         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4898                 return -EINVAL;
4899
4900         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4901                 return -EINVAL;
4902
4903 out:
4904         return ret;
4905
4906 err_size:
4907         put_user(sizeof(*attr), &uattr->size);
4908         ret = -E2BIG;
4909         goto out;
4910 }
4911
4912 static int perf_event_set_output(struct perf_event *event, int output_fd)
4913 {
4914         struct perf_event *output_event = NULL;
4915         struct file *output_file = NULL;
4916         struct perf_event *old_output;
4917         int fput_needed = 0;
4918         int ret = -EINVAL;
4919
4920         if (!output_fd)
4921                 goto set;
4922
4923         output_file = fget_light(output_fd, &fput_needed);
4924         if (!output_file)
4925                 return -EBADF;
4926
4927         if (output_file->f_op != &perf_fops)
4928                 goto out;
4929
4930         output_event = output_file->private_data;
4931
4932         /* Don't chain output fds */
4933         if (output_event->output)
4934                 goto out;
4935
4936         /* Don't set an output fd when we already have an output channel */
4937         if (event->data)
4938                 goto out;
4939
4940         atomic_long_inc(&output_file->f_count);
4941
4942 set:
4943         mutex_lock(&event->mmap_mutex);
4944         old_output = event->output;
4945         rcu_assign_pointer(event->output, output_event);
4946         mutex_unlock(&event->mmap_mutex);
4947
4948         if (old_output) {
4949                 /*
4950                  * we need to make sure no existing perf_output_*()
4951                  * is still referencing this event.
4952                  */
4953                 synchronize_rcu();
4954                 fput(old_output->filp);
4955         }
4956
4957         ret = 0;
4958 out:
4959         fput_light(output_file, fput_needed);
4960         return ret;
4961 }
4962
4963 /**
4964  * sys_perf_event_open - open a performance event, associate it to a task/cpu
4965  *
4966  * @attr_uptr:  event_id type attributes for monitoring/sampling
4967  * @pid:                target pid
4968  * @cpu:                target cpu
4969  * @group_fd:           group leader event fd
4970  */
4971 SYSCALL_DEFINE5(perf_event_open,
4972                 struct perf_event_attr __user *, attr_uptr,
4973                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4974 {
4975         struct perf_event *event, *group_leader;
4976         struct perf_event_attr attr;
4977         struct perf_event_context *ctx;
4978         struct file *event_file = NULL;
4979         struct file *group_file = NULL;
4980         int fput_needed = 0;
4981         int fput_needed2 = 0;
4982         int err;
4983
4984         /* for future expandability... */
4985         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4986                 return -EINVAL;
4987
4988         err = perf_copy_attr(attr_uptr, &attr);
4989         if (err)
4990                 return err;
4991
4992         if (!attr.exclude_kernel) {
4993                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4994                         return -EACCES;
4995         }
4996
4997         if (attr.freq) {
4998                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4999                         return -EINVAL;
5000         }
5001
5002         /*
5003          * Get the target context (task or percpu):
5004          */
5005         ctx = find_get_context(pid, cpu);
5006         if (IS_ERR(ctx))
5007                 return PTR_ERR(ctx);
5008
5009         /*
5010          * Look up the group leader (we will attach this event to it):
5011          */
5012         group_leader = NULL;
5013         if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
5014                 err = -EINVAL;
5015                 group_file = fget_light(group_fd, &fput_needed);
5016                 if (!group_file)
5017                         goto err_put_context;
5018                 if (group_file->f_op != &perf_fops)
5019                         goto err_put_context;
5020
5021                 group_leader = group_file->private_data;
5022                 /*
5023                  * Do not allow a recursive hierarchy (this new sibling
5024                  * becoming part of another group-sibling):
5025                  */
5026                 if (group_leader->group_leader != group_leader)
5027                         goto err_put_context;
5028                 /*
5029                  * Do not allow to attach to a group in a different
5030                  * task or CPU context:
5031                  */
5032                 if (group_leader->ctx != ctx)
5033                         goto err_put_context;
5034                 /*
5035                  * Only a group leader can be exclusive or pinned
5036                  */
5037                 if (attr.exclusive || attr.pinned)
5038                         goto err_put_context;
5039         }
5040
5041         event = perf_event_alloc(&attr, cpu, ctx, group_leader,
5042                                      NULL, NULL, GFP_KERNEL);
5043         err = PTR_ERR(event);
5044         if (IS_ERR(event))
5045                 goto err_put_context;
5046
5047         err = anon_inode_getfd("[perf_event]", &perf_fops, event, O_RDWR);
5048         if (err < 0)
5049                 goto err_free_put_context;
5050
5051         event_file = fget_light(err, &fput_needed2);
5052         if (!event_file)
5053                 goto err_free_put_context;
5054
5055         if (flags & PERF_FLAG_FD_OUTPUT) {
5056                 err = perf_event_set_output(event, group_fd);
5057                 if (err)
5058                         goto err_fput_free_put_context;
5059         }
5060
5061         event->filp = event_file;
5062         WARN_ON_ONCE(ctx->parent_ctx);
5063         mutex_lock(&ctx->mutex);
5064         perf_install_in_context(ctx, event, cpu);
5065         ++ctx->generation;
5066         mutex_unlock(&ctx->mutex);
5067
5068         event->owner = current;
5069         get_task_struct(current);
5070         mutex_lock(&current->perf_event_mutex);
5071         list_add_tail(&event->owner_entry, &current->perf_event_list);
5072         mutex_unlock(&current->perf_event_mutex);
5073
5074 err_fput_free_put_context:
5075         fput_light(event_file, fput_needed2);
5076
5077 err_free_put_context:
5078         if (err < 0)
5079                 free_event(event);
5080
5081 err_put_context:
5082         if (err < 0)
5083                 put_ctx(ctx);
5084
5085         fput_light(group_file, fput_needed);
5086
5087         return err;
5088 }
5089
5090 /**
5091  * perf_event_create_kernel_counter
5092  *
5093  * @attr: attributes of the counter to create
5094  * @cpu: cpu in which the counter is bound
5095  * @pid: task to profile
5096  */
5097 struct perf_event *
5098 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
5099                                  pid_t pid,
5100                                  perf_overflow_handler_t overflow_handler)
5101 {
5102         struct perf_event *event;
5103         struct perf_event_context *ctx;
5104         int err;
5105
5106         /*
5107          * Get the target context (task or percpu):
5108          */
5109
5110         ctx = find_get_context(pid, cpu);
5111         if (IS_ERR(ctx)) {
5112                 err = PTR_ERR(ctx);
5113                 goto err_exit;
5114         }
5115
5116         event = perf_event_alloc(attr, cpu, ctx, NULL,
5117                                  NULL, overflow_handler, GFP_KERNEL);
5118         if (IS_ERR(event)) {
5119                 err = PTR_ERR(event);
5120                 goto err_put_context;
5121         }
5122
5123         event->filp = NULL;
5124         WARN_ON_ONCE(ctx->parent_ctx);
5125         mutex_lock(&ctx->mutex);
5126         perf_install_in_context(ctx, event, cpu);
5127         ++ctx->generation;
5128         mutex_unlock(&ctx->mutex);
5129
5130         event->owner = current;
5131         get_task_struct(current);
5132         mutex_lock(&current->perf_event_mutex);
5133         list_add_tail(&event->owner_entry, &current->perf_event_list);
5134         mutex_unlock(&current->perf_event_mutex);
5135
5136         return event;
5137
5138  err_put_context:
5139         put_ctx(ctx);
5140  err_exit:
5141         return ERR_PTR(err);
5142 }
5143 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
5144
5145 /*
5146  * inherit a event from parent task to child task:
5147  */
5148 static struct perf_event *
5149 inherit_event(struct perf_event *parent_event,
5150               struct task_struct *parent,
5151               struct perf_event_context *parent_ctx,
5152               struct task_struct *child,
5153               struct perf_event *group_leader,
5154               struct perf_event_context *child_ctx)
5155 {
5156         struct perf_event *child_event;
5157
5158         /*
5159          * Instead of creating recursive hierarchies of events,
5160          * we link inherited events back to the original parent,
5161          * which has a filp for sure, which we use as the reference
5162          * count:
5163          */
5164         if (parent_event->parent)
5165                 parent_event = parent_event->parent;
5166
5167         child_event = perf_event_alloc(&parent_event->attr,
5168                                            parent_event->cpu, child_ctx,
5169                                            group_leader, parent_event,
5170                                            NULL, GFP_KERNEL);
5171         if (IS_ERR(child_event))
5172                 return child_event;
5173         get_ctx(child_ctx);
5174
5175         /*
5176          * Make the child state follow the state of the parent event,
5177          * not its attr.disabled bit.  We hold the parent's mutex,
5178          * so we won't race with perf_event_{en, dis}able_family.
5179          */
5180         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
5181                 child_event->state = PERF_EVENT_STATE_INACTIVE;
5182         else
5183                 child_event->state = PERF_EVENT_STATE_OFF;
5184
5185         if (parent_event->attr.freq) {
5186                 u64 sample_period = parent_event->hw.sample_period;
5187                 struct hw_perf_event *hwc = &child_event->hw;
5188
5189                 hwc->sample_period = sample_period;
5190                 hwc->last_period   = sample_period;
5191
5192                 atomic64_set(&hwc->period_left, sample_period);
5193         }
5194
5195         child_event->overflow_handler = parent_event->overflow_handler;
5196
5197         /*
5198          * Link it up in the child's context:
5199          */
5200         add_event_to_ctx(child_event, child_ctx);
5201
5202         /*
5203          * Get a reference to the parent filp - we will fput it
5204          * when the child event exits. This is safe to do because
5205          * we are in the parent and we know that the filp still
5206          * exists and has a nonzero count:
5207          */
5208         atomic_long_inc(&parent_event->filp->f_count);
5209
5210         /*
5211          * Link this into the parent event's child list
5212          */
5213         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5214         mutex_lock(&parent_event->child_mutex);
5215         list_add_tail(&child_event->child_list, &parent_event->child_list);
5216         mutex_unlock(&parent_event->child_mutex);
5217
5218         return child_event;
5219 }
5220
5221 static int inherit_group(struct perf_event *parent_event,
5222               struct task_struct *parent,
5223               struct perf_event_context *parent_ctx,
5224               struct task_struct *child,
5225               struct perf_event_context *child_ctx)
5226 {
5227         struct perf_event *leader;
5228         struct perf_event *sub;
5229         struct perf_event *child_ctr;
5230
5231         leader = inherit_event(parent_event, parent, parent_ctx,
5232                                  child, NULL, child_ctx);
5233         if (IS_ERR(leader))
5234                 return PTR_ERR(leader);
5235         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5236                 child_ctr = inherit_event(sub, parent, parent_ctx,
5237                                             child, leader, child_ctx);
5238                 if (IS_ERR(child_ctr))
5239                         return PTR_ERR(child_ctr);
5240         }
5241         return 0;
5242 }
5243
5244 static void sync_child_event(struct perf_event *child_event,
5245                                struct task_struct *child)
5246 {
5247         struct perf_event *parent_event = child_event->parent;
5248         u64 child_val;
5249
5250         if (child_event->attr.inherit_stat)
5251                 perf_event_read_event(child_event, child);
5252
5253         child_val = atomic64_read(&child_event->count);
5254
5255         /*
5256          * Add back the child's count to the parent's count:
5257          */
5258         atomic64_add(child_val, &parent_event->count);
5259         atomic64_add(child_event->total_time_enabled,
5260                      &parent_event->child_total_time_enabled);
5261         atomic64_add(child_event->total_time_running,
5262                      &parent_event->child_total_time_running);
5263
5264         /*
5265          * Remove this event from the parent's list
5266          */
5267         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5268         mutex_lock(&parent_event->child_mutex);
5269         list_del_init(&child_event->child_list);
5270         mutex_unlock(&parent_event->child_mutex);
5271
5272         /*
5273          * Release the parent event, if this was the last
5274          * reference to it.
5275          */
5276         fput(parent_event->filp);
5277 }
5278
5279 static void
5280 __perf_event_exit_task(struct perf_event *child_event,
5281                          struct perf_event_context *child_ctx,
5282                          struct task_struct *child)
5283 {
5284         struct perf_event *parent_event;
5285
5286         perf_event_remove_from_context(child_event);
5287
5288         parent_event = child_event->parent;
5289         /*
5290          * It can happen that parent exits first, and has events
5291          * that are still around due to the child reference. These
5292          * events need to be zapped - but otherwise linger.
5293          */
5294         if (parent_event) {
5295                 sync_child_event(child_event, child);
5296                 free_event(child_event);
5297         }
5298 }
5299
5300 /*
5301  * When a child task exits, feed back event values to parent events.
5302  */
5303 void perf_event_exit_task(struct task_struct *child)
5304 {
5305         struct perf_event *child_event, *tmp;
5306         struct perf_event_context *child_ctx;
5307         unsigned long flags;
5308
5309         if (likely(!child->perf_event_ctxp)) {
5310                 perf_event_task(child, NULL, 0);
5311                 return;
5312         }
5313
5314         local_irq_save(flags);
5315         /*
5316          * We can't reschedule here because interrupts are disabled,
5317          * and either child is current or it is a task that can't be
5318          * scheduled, so we are now safe from rescheduling changing
5319          * our context.
5320          */
5321         child_ctx = child->perf_event_ctxp;
5322         __perf_event_task_sched_out(child_ctx);
5323
5324         /*
5325          * Take the context lock here so that if find_get_context is
5326          * reading child->perf_event_ctxp, we wait until it has
5327          * incremented the context's refcount before we do put_ctx below.
5328          */
5329         raw_spin_lock(&child_ctx->lock);
5330         child->perf_event_ctxp = NULL;
5331         /*
5332          * If this context is a clone; unclone it so it can't get
5333          * swapped to another process while we're removing all
5334          * the events from it.
5335          */
5336         unclone_ctx(child_ctx);
5337         update_context_time(child_ctx);
5338         raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5339
5340         /*
5341          * Report the task dead after unscheduling the events so that we
5342          * won't get any samples after PERF_RECORD_EXIT. We can however still
5343          * get a few PERF_RECORD_READ events.
5344          */
5345         perf_event_task(child, child_ctx, 0);
5346
5347         /*
5348          * We can recurse on the same lock type through:
5349          *
5350          *   __perf_event_exit_task()
5351          *     sync_child_event()
5352          *       fput(parent_event->filp)
5353          *         perf_release()
5354          *           mutex_lock(&ctx->mutex)
5355          *
5356          * But since its the parent context it won't be the same instance.
5357          */
5358         mutex_lock(&child_ctx->mutex);
5359
5360 again:
5361         list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5362                                  group_entry)
5363                 __perf_event_exit_task(child_event, child_ctx, child);
5364
5365         list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5366                                  group_entry)
5367                 __perf_event_exit_task(child_event, child_ctx, child);
5368
5369         /*
5370          * If the last event was a group event, it will have appended all
5371          * its siblings to the list, but we obtained 'tmp' before that which
5372          * will still point to the list head terminating the iteration.
5373          */
5374         if (!list_empty(&child_ctx->pinned_groups) ||
5375             !list_empty(&child_ctx->flexible_groups))
5376                 goto again;
5377
5378         mutex_unlock(&child_ctx->mutex);
5379
5380         put_ctx(child_ctx);
5381 }
5382
5383 static void perf_free_event(struct perf_event *event,
5384                             struct perf_event_context *ctx)
5385 {
5386         struct perf_event *parent = event->parent;
5387
5388         if (WARN_ON_ONCE(!parent))
5389                 return;
5390
5391         mutex_lock(&parent->child_mutex);
5392         list_del_init(&event->child_list);
5393         mutex_unlock(&parent->child_mutex);
5394
5395         fput(parent->filp);
5396
5397         list_del_event(event, ctx);
5398         free_event(event);
5399 }
5400
5401 /*
5402  * free an unexposed, unused context as created by inheritance by
5403  * init_task below, used by fork() in case of fail.
5404  */
5405 void perf_event_free_task(struct task_struct *task)
5406 {
5407         struct perf_event_context *ctx = task->perf_event_ctxp;
5408         struct perf_event *event, *tmp;
5409
5410         if (!ctx)
5411                 return;
5412
5413         mutex_lock(&ctx->mutex);
5414 again:
5415         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5416                 perf_free_event(event, ctx);
5417
5418         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5419                                  group_entry)
5420                 perf_free_event(event, ctx);
5421
5422         if (!list_empty(&ctx->pinned_groups) ||
5423             !list_empty(&ctx->flexible_groups))
5424                 goto again;
5425
5426         mutex_unlock(&ctx->mutex);
5427
5428         put_ctx(ctx);
5429 }
5430
5431 static int
5432 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5433                    struct perf_event_context *parent_ctx,
5434                    struct task_struct *child,
5435                    int *inherited_all)
5436 {
5437         int ret;
5438         struct perf_event_context *child_ctx = child->perf_event_ctxp;
5439
5440         if (!event->attr.inherit) {
5441                 *inherited_all = 0;
5442                 return 0;
5443         }
5444
5445         if (!child_ctx) {
5446                 /*
5447                  * This is executed from the parent task context, so
5448                  * inherit events that have been marked for cloning.
5449                  * First allocate and initialize a context for the
5450                  * child.
5451                  */
5452
5453                 child_ctx = kzalloc(sizeof(struct perf_event_context),
5454                                     GFP_KERNEL);
5455                 if (!child_ctx)
5456                         return -ENOMEM;
5457
5458                 __perf_event_init_context(child_ctx, child);
5459                 child->perf_event_ctxp = child_ctx;
5460                 get_task_struct(child);
5461         }
5462
5463         ret = inherit_group(event, parent, parent_ctx,
5464                             child, child_ctx);
5465
5466         if (ret)
5467                 *inherited_all = 0;
5468
5469         return ret;
5470 }
5471
5472
5473 /*
5474  * Initialize the perf_event context in task_struct
5475  */
5476 int perf_event_init_task(struct task_struct *child)
5477 {
5478         struct perf_event_context *child_ctx, *parent_ctx;
5479         struct perf_event_context *cloned_ctx;
5480         struct perf_event *event;
5481         struct task_struct *parent = current;
5482         int inherited_all = 1;
5483         int ret = 0;
5484
5485         child->perf_event_ctxp = NULL;
5486
5487         mutex_init(&child->perf_event_mutex);
5488         INIT_LIST_HEAD(&child->perf_event_list);
5489
5490         if (likely(!parent->perf_event_ctxp))
5491                 return 0;
5492
5493         /*
5494          * If the parent's context is a clone, pin it so it won't get
5495          * swapped under us.
5496          */
5497         parent_ctx = perf_pin_task_context(parent);
5498
5499         /*
5500          * No need to check if parent_ctx != NULL here; since we saw
5501          * it non-NULL earlier, the only reason for it to become NULL
5502          * is if we exit, and since we're currently in the middle of
5503          * a fork we can't be exiting at the same time.
5504          */
5505
5506         /*
5507          * Lock the parent list. No need to lock the child - not PID
5508          * hashed yet and not running, so nobody can access it.
5509          */
5510         mutex_lock(&parent_ctx->mutex);
5511
5512         /*
5513          * We dont have to disable NMIs - we are only looking at
5514          * the list, not manipulating it:
5515          */
5516         list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5517                 ret = inherit_task_group(event, parent, parent_ctx, child,
5518                                          &inherited_all);
5519                 if (ret)
5520                         break;
5521         }
5522
5523         list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5524                 ret = inherit_task_group(event, parent, parent_ctx, child,
5525                                          &inherited_all);
5526                 if (ret)
5527                         break;
5528         }
5529
5530         child_ctx = child->perf_event_ctxp;
5531
5532         if (child_ctx && inherited_all) {
5533                 /*
5534                  * Mark the child context as a clone of the parent
5535                  * context, or of whatever the parent is a clone of.
5536                  * Note that if the parent is a clone, it could get
5537                  * uncloned at any point, but that doesn't matter
5538                  * because the list of events and the generation
5539                  * count can't have changed since we took the mutex.
5540                  */
5541                 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5542                 if (cloned_ctx) {
5543                         child_ctx->parent_ctx = cloned_ctx;
5544                         child_ctx->parent_gen = parent_ctx->parent_gen;
5545                 } else {
5546                         child_ctx->parent_ctx = parent_ctx;
5547                         child_ctx->parent_gen = parent_ctx->generation;
5548                 }
5549                 get_ctx(child_ctx->parent_ctx);
5550         }
5551
5552         mutex_unlock(&parent_ctx->mutex);
5553
5554         perf_unpin_context(parent_ctx);
5555
5556         return ret;
5557 }
5558
5559 static void __init perf_event_init_all_cpus(void)
5560 {
5561         int cpu;
5562         struct perf_cpu_context *cpuctx;
5563
5564         for_each_possible_cpu(cpu) {
5565                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5566                 mutex_init(&cpuctx->hlist_mutex);
5567                 __perf_event_init_context(&cpuctx->ctx, NULL);
5568         }
5569 }
5570
5571 static void __cpuinit perf_event_init_cpu(int cpu)
5572 {
5573         struct perf_cpu_context *cpuctx;
5574
5575         cpuctx = &per_cpu(perf_cpu_context, cpu);
5576
5577         spin_lock(&perf_resource_lock);
5578         cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5579         spin_unlock(&perf_resource_lock);
5580
5581         mutex_lock(&cpuctx->hlist_mutex);
5582         if (cpuctx->hlist_refcount > 0) {
5583                 struct swevent_hlist *hlist;
5584
5585                 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5586                 WARN_ON_ONCE(!hlist);
5587                 rcu_assign_pointer(cpuctx->swevent_hlist, hlist);
5588         }
5589         mutex_unlock(&cpuctx->hlist_mutex);
5590 }
5591
5592 #ifdef CONFIG_HOTPLUG_CPU
5593 static void __perf_event_exit_cpu(void *info)
5594 {
5595         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5596         struct perf_event_context *ctx = &cpuctx->ctx;
5597         struct perf_event *event, *tmp;
5598
5599         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5600                 __perf_event_remove_from_context(event);
5601         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5602                 __perf_event_remove_from_context(event);
5603 }
5604 static void perf_event_exit_cpu(int cpu)
5605 {
5606         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5607         struct perf_event_context *ctx = &cpuctx->ctx;
5608
5609         mutex_lock(&cpuctx->hlist_mutex);
5610         swevent_hlist_release(cpuctx);
5611         mutex_unlock(&cpuctx->hlist_mutex);
5612
5613         mutex_lock(&ctx->mutex);
5614         smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5615         mutex_unlock(&ctx->mutex);
5616 }
5617 #else
5618 static inline void perf_event_exit_cpu(int cpu) { }
5619 #endif
5620
5621 static int __cpuinit
5622 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5623 {
5624         unsigned int cpu = (long)hcpu;
5625
5626         switch (action) {
5627
5628         case CPU_UP_PREPARE:
5629         case CPU_UP_PREPARE_FROZEN:
5630                 perf_event_init_cpu(cpu);
5631                 break;
5632
5633         case CPU_DOWN_PREPARE:
5634         case CPU_DOWN_PREPARE_FROZEN:
5635                 perf_event_exit_cpu(cpu);
5636                 break;
5637
5638         default:
5639                 break;
5640         }
5641
5642         return NOTIFY_OK;
5643 }
5644
5645 /*
5646  * This has to have a higher priority than migration_notifier in sched.c.
5647  */
5648 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5649         .notifier_call          = perf_cpu_notify,
5650         .priority               = 20,
5651 };
5652
5653 void __init perf_event_init(void)
5654 {
5655         perf_event_init_all_cpus();
5656         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5657                         (void *)(long)smp_processor_id());
5658         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5659                         (void *)(long)smp_processor_id());
5660         register_cpu_notifier(&perf_cpu_nb);
5661 }
5662
5663 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class,
5664                                         struct sysdev_class_attribute *attr,
5665                                         char *buf)
5666 {
5667         return sprintf(buf, "%d\n", perf_reserved_percpu);
5668 }
5669
5670 static ssize_t
5671 perf_set_reserve_percpu(struct sysdev_class *class,
5672                         struct sysdev_class_attribute *attr,
5673                         const char *buf,
5674                         size_t count)
5675 {
5676         struct perf_cpu_context *cpuctx;
5677         unsigned long val;
5678         int err, cpu, mpt;
5679
5680         err = strict_strtoul(buf, 10, &val);
5681         if (err)
5682                 return err;
5683         if (val > perf_max_events)
5684                 return -EINVAL;
5685
5686         spin_lock(&perf_resource_lock);
5687         perf_reserved_percpu = val;
5688         for_each_online_cpu(cpu) {
5689                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5690                 raw_spin_lock_irq(&cpuctx->ctx.lock);
5691                 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5692                           perf_max_events - perf_reserved_percpu);
5693                 cpuctx->max_pertask = mpt;
5694                 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5695         }
5696         spin_unlock(&perf_resource_lock);
5697
5698         return count;
5699 }
5700
5701 static ssize_t perf_show_overcommit(struct sysdev_class *class,
5702                                     struct sysdev_class_attribute *attr,
5703                                     char *buf)
5704 {
5705         return sprintf(buf, "%d\n", perf_overcommit);
5706 }
5707
5708 static ssize_t
5709 perf_set_overcommit(struct sysdev_class *class,
5710                     struct sysdev_class_attribute *attr,
5711                     const char *buf, size_t count)
5712 {
5713         unsigned long val;
5714         int err;
5715
5716         err = strict_strtoul(buf, 10, &val);
5717         if (err)
5718                 return err;
5719         if (val > 1)
5720                 return -EINVAL;
5721
5722         spin_lock(&perf_resource_lock);
5723         perf_overcommit = val;
5724         spin_unlock(&perf_resource_lock);
5725
5726         return count;
5727 }
5728
5729 static SYSDEV_CLASS_ATTR(
5730                                 reserve_percpu,
5731                                 0644,
5732                                 perf_show_reserve_percpu,
5733                                 perf_set_reserve_percpu
5734                         );
5735
5736 static SYSDEV_CLASS_ATTR(
5737                                 overcommit,
5738                                 0644,
5739                                 perf_show_overcommit,
5740                                 perf_set_overcommit
5741                         );
5742
5743 static struct attribute *perfclass_attrs[] = {
5744         &attr_reserve_percpu.attr,
5745         &attr_overcommit.attr,
5746         NULL
5747 };
5748
5749 static struct attribute_group perfclass_attr_group = {
5750         .attrs                  = perfclass_attrs,
5751         .name                   = "perf_events",
5752 };
5753
5754 static int __init perf_event_sysfs_init(void)
5755 {
5756         return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5757                                   &perfclass_attr_group);
5758 }
5759 device_initcall(perf_event_sysfs_init);