Merge branch 'for-davem' of ssh://master.kernel.org/pub/scm/linux/kernel/git/linville...
[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/idr.h>
17 #include <linux/file.h>
18 #include <linux/poll.h>
19 #include <linux/slab.h>
20 #include <linux/hash.h>
21 #include <linux/sysfs.h>
22 #include <linux/dcache.h>
23 #include <linux/percpu.h>
24 #include <linux/ptrace.h>
25 #include <linux/reboot.h>
26 #include <linux/vmstat.h>
27 #include <linux/device.h>
28 #include <linux/vmalloc.h>
29 #include <linux/hardirq.h>
30 #include <linux/rculist.h>
31 #include <linux/uaccess.h>
32 #include <linux/syscalls.h>
33 #include <linux/anon_inodes.h>
34 #include <linux/kernel_stat.h>
35 #include <linux/perf_event.h>
36 #include <linux/ftrace_event.h>
37 #include <linux/hw_breakpoint.h>
38
39 #include <asm/irq_regs.h>
40
41 enum event_type_t {
42         EVENT_FLEXIBLE = 0x1,
43         EVENT_PINNED = 0x2,
44         EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
45 };
46
47 atomic_t perf_task_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 static LIST_HEAD(pmus);
53 static DEFINE_MUTEX(pmus_lock);
54 static struct srcu_struct pmus_srcu;
55
56 /*
57  * perf event paranoia level:
58  *  -1 - not paranoid at all
59  *   0 - disallow raw tracepoint access for unpriv
60  *   1 - disallow cpu events for unpriv
61  *   2 - disallow kernel profiling for unpriv
62  */
63 int sysctl_perf_event_paranoid __read_mostly = 1;
64
65 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
66
67 /*
68  * max perf event sample rate
69  */
70 int sysctl_perf_event_sample_rate __read_mostly = 100000;
71
72 static atomic64_t perf_event_id;
73
74 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
75                               enum event_type_t event_type);
76
77 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
78                              enum event_type_t event_type);
79
80 void __weak perf_event_print_debug(void)        { }
81
82 extern __weak const char *perf_pmu_name(void)
83 {
84         return "pmu";
85 }
86
87 static inline u64 perf_clock(void)
88 {
89         return local_clock();
90 }
91
92 void perf_pmu_disable(struct pmu *pmu)
93 {
94         int *count = this_cpu_ptr(pmu->pmu_disable_count);
95         if (!(*count)++)
96                 pmu->pmu_disable(pmu);
97 }
98
99 void perf_pmu_enable(struct pmu *pmu)
100 {
101         int *count = this_cpu_ptr(pmu->pmu_disable_count);
102         if (!--(*count))
103                 pmu->pmu_enable(pmu);
104 }
105
106 static DEFINE_PER_CPU(struct list_head, rotation_list);
107
108 /*
109  * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
110  * because they're strictly cpu affine and rotate_start is called with IRQs
111  * disabled, while rotate_context is called from IRQ context.
112  */
113 static void perf_pmu_rotate_start(struct pmu *pmu)
114 {
115         struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
116         struct list_head *head = &__get_cpu_var(rotation_list);
117
118         WARN_ON(!irqs_disabled());
119
120         if (list_empty(&cpuctx->rotation_list))
121                 list_add(&cpuctx->rotation_list, head);
122 }
123
124 static void get_ctx(struct perf_event_context *ctx)
125 {
126         WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
127 }
128
129 static void free_ctx(struct rcu_head *head)
130 {
131         struct perf_event_context *ctx;
132
133         ctx = container_of(head, struct perf_event_context, rcu_head);
134         kfree(ctx);
135 }
136
137 static void put_ctx(struct perf_event_context *ctx)
138 {
139         if (atomic_dec_and_test(&ctx->refcount)) {
140                 if (ctx->parent_ctx)
141                         put_ctx(ctx->parent_ctx);
142                 if (ctx->task)
143                         put_task_struct(ctx->task);
144                 call_rcu(&ctx->rcu_head, free_ctx);
145         }
146 }
147
148 static void unclone_ctx(struct perf_event_context *ctx)
149 {
150         if (ctx->parent_ctx) {
151                 put_ctx(ctx->parent_ctx);
152                 ctx->parent_ctx = NULL;
153         }
154 }
155
156 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
157 {
158         /*
159          * only top level events have the pid namespace they were created in
160          */
161         if (event->parent)
162                 event = event->parent;
163
164         return task_tgid_nr_ns(p, event->ns);
165 }
166
167 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
168 {
169         /*
170          * only top level events have the pid namespace they were created in
171          */
172         if (event->parent)
173                 event = event->parent;
174
175         return task_pid_nr_ns(p, event->ns);
176 }
177
178 /*
179  * If we inherit events we want to return the parent event id
180  * to userspace.
181  */
182 static u64 primary_event_id(struct perf_event *event)
183 {
184         u64 id = event->id;
185
186         if (event->parent)
187                 id = event->parent->id;
188
189         return id;
190 }
191
192 /*
193  * Get the perf_event_context for a task and lock it.
194  * This has to cope with with the fact that until it is locked,
195  * the context could get moved to another task.
196  */
197 static struct perf_event_context *
198 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
199 {
200         struct perf_event_context *ctx;
201
202         rcu_read_lock();
203 retry:
204         ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
205         if (ctx) {
206                 /*
207                  * If this context is a clone of another, it might
208                  * get swapped for another underneath us by
209                  * perf_event_task_sched_out, though the
210                  * rcu_read_lock() protects us from any context
211                  * getting freed.  Lock the context and check if it
212                  * got swapped before we could get the lock, and retry
213                  * if so.  If we locked the right context, then it
214                  * can't get swapped on us any more.
215                  */
216                 raw_spin_lock_irqsave(&ctx->lock, *flags);
217                 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
218                         raw_spin_unlock_irqrestore(&ctx->lock, *flags);
219                         goto retry;
220                 }
221
222                 if (!atomic_inc_not_zero(&ctx->refcount)) {
223                         raw_spin_unlock_irqrestore(&ctx->lock, *flags);
224                         ctx = NULL;
225                 }
226         }
227         rcu_read_unlock();
228         return ctx;
229 }
230
231 /*
232  * Get the context for a task and increment its pin_count so it
233  * can't get swapped to another task.  This also increments its
234  * reference count so that the context can't get freed.
235  */
236 static struct perf_event_context *
237 perf_pin_task_context(struct task_struct *task, int ctxn)
238 {
239         struct perf_event_context *ctx;
240         unsigned long flags;
241
242         ctx = perf_lock_task_context(task, ctxn, &flags);
243         if (ctx) {
244                 ++ctx->pin_count;
245                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
246         }
247         return ctx;
248 }
249
250 static void perf_unpin_context(struct perf_event_context *ctx)
251 {
252         unsigned long flags;
253
254         raw_spin_lock_irqsave(&ctx->lock, flags);
255         --ctx->pin_count;
256         raw_spin_unlock_irqrestore(&ctx->lock, flags);
257         put_ctx(ctx);
258 }
259
260 /*
261  * Update the record of the current time in a context.
262  */
263 static void update_context_time(struct perf_event_context *ctx)
264 {
265         u64 now = perf_clock();
266
267         ctx->time += now - ctx->timestamp;
268         ctx->timestamp = now;
269 }
270
271 static u64 perf_event_time(struct perf_event *event)
272 {
273         struct perf_event_context *ctx = event->ctx;
274         return ctx ? ctx->time : 0;
275 }
276
277 /*
278  * Update the total_time_enabled and total_time_running fields for a event.
279  */
280 static void update_event_times(struct perf_event *event)
281 {
282         struct perf_event_context *ctx = event->ctx;
283         u64 run_end;
284
285         if (event->state < PERF_EVENT_STATE_INACTIVE ||
286             event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
287                 return;
288
289         if (ctx->is_active)
290                 run_end = perf_event_time(event);
291         else
292                 run_end = event->tstamp_stopped;
293
294         event->total_time_enabled = run_end - event->tstamp_enabled;
295
296         if (event->state == PERF_EVENT_STATE_INACTIVE)
297                 run_end = event->tstamp_stopped;
298         else
299                 run_end = perf_event_time(event);
300
301         event->total_time_running = run_end - event->tstamp_running;
302 }
303
304 /*
305  * Update total_time_enabled and total_time_running for all events in a group.
306  */
307 static void update_group_times(struct perf_event *leader)
308 {
309         struct perf_event *event;
310
311         update_event_times(leader);
312         list_for_each_entry(event, &leader->sibling_list, group_entry)
313                 update_event_times(event);
314 }
315
316 static struct list_head *
317 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
318 {
319         if (event->attr.pinned)
320                 return &ctx->pinned_groups;
321         else
322                 return &ctx->flexible_groups;
323 }
324
325 /*
326  * Add a event from the lists for its context.
327  * Must be called with ctx->mutex and ctx->lock held.
328  */
329 static void
330 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
331 {
332         WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
333         event->attach_state |= PERF_ATTACH_CONTEXT;
334
335         /*
336          * If we're a stand alone event or group leader, we go to the context
337          * list, group events are kept attached to the group so that
338          * perf_group_detach can, at all times, locate all siblings.
339          */
340         if (event->group_leader == event) {
341                 struct list_head *list;
342
343                 if (is_software_event(event))
344                         event->group_flags |= PERF_GROUP_SOFTWARE;
345
346                 list = ctx_group_list(event, ctx);
347                 list_add_tail(&event->group_entry, list);
348         }
349
350         list_add_rcu(&event->event_entry, &ctx->event_list);
351         if (!ctx->nr_events)
352                 perf_pmu_rotate_start(ctx->pmu);
353         ctx->nr_events++;
354         if (event->attr.inherit_stat)
355                 ctx->nr_stat++;
356 }
357
358 /*
359  * Called at perf_event creation and when events are attached/detached from a
360  * group.
361  */
362 static void perf_event__read_size(struct perf_event *event)
363 {
364         int entry = sizeof(u64); /* value */
365         int size = 0;
366         int nr = 1;
367
368         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
369                 size += sizeof(u64);
370
371         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
372                 size += sizeof(u64);
373
374         if (event->attr.read_format & PERF_FORMAT_ID)
375                 entry += sizeof(u64);
376
377         if (event->attr.read_format & PERF_FORMAT_GROUP) {
378                 nr += event->group_leader->nr_siblings;
379                 size += sizeof(u64);
380         }
381
382         size += entry * nr;
383         event->read_size = size;
384 }
385
386 static void perf_event__header_size(struct perf_event *event)
387 {
388         struct perf_sample_data *data;
389         u64 sample_type = event->attr.sample_type;
390         u16 size = 0;
391
392         perf_event__read_size(event);
393
394         if (sample_type & PERF_SAMPLE_IP)
395                 size += sizeof(data->ip);
396
397         if (sample_type & PERF_SAMPLE_ADDR)
398                 size += sizeof(data->addr);
399
400         if (sample_type & PERF_SAMPLE_PERIOD)
401                 size += sizeof(data->period);
402
403         if (sample_type & PERF_SAMPLE_READ)
404                 size += event->read_size;
405
406         event->header_size = size;
407 }
408
409 static void perf_event__id_header_size(struct perf_event *event)
410 {
411         struct perf_sample_data *data;
412         u64 sample_type = event->attr.sample_type;
413         u16 size = 0;
414
415         if (sample_type & PERF_SAMPLE_TID)
416                 size += sizeof(data->tid_entry);
417
418         if (sample_type & PERF_SAMPLE_TIME)
419                 size += sizeof(data->time);
420
421         if (sample_type & PERF_SAMPLE_ID)
422                 size += sizeof(data->id);
423
424         if (sample_type & PERF_SAMPLE_STREAM_ID)
425                 size += sizeof(data->stream_id);
426
427         if (sample_type & PERF_SAMPLE_CPU)
428                 size += sizeof(data->cpu_entry);
429
430         event->id_header_size = size;
431 }
432
433 static void perf_group_attach(struct perf_event *event)
434 {
435         struct perf_event *group_leader = event->group_leader, *pos;
436
437         /*
438          * We can have double attach due to group movement in perf_event_open.
439          */
440         if (event->attach_state & PERF_ATTACH_GROUP)
441                 return;
442
443         event->attach_state |= PERF_ATTACH_GROUP;
444
445         if (group_leader == event)
446                 return;
447
448         if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
449                         !is_software_event(event))
450                 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
451
452         list_add_tail(&event->group_entry, &group_leader->sibling_list);
453         group_leader->nr_siblings++;
454
455         perf_event__header_size(group_leader);
456
457         list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
458                 perf_event__header_size(pos);
459 }
460
461 /*
462  * Remove a event from the lists for its context.
463  * Must be called with ctx->mutex and ctx->lock held.
464  */
465 static void
466 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
467 {
468         /*
469          * We can have double detach due to exit/hot-unplug + close.
470          */
471         if (!(event->attach_state & PERF_ATTACH_CONTEXT))
472                 return;
473
474         event->attach_state &= ~PERF_ATTACH_CONTEXT;
475
476         ctx->nr_events--;
477         if (event->attr.inherit_stat)
478                 ctx->nr_stat--;
479
480         list_del_rcu(&event->event_entry);
481
482         if (event->group_leader == event)
483                 list_del_init(&event->group_entry);
484
485         update_group_times(event);
486
487         /*
488          * If event was in error state, then keep it
489          * that way, otherwise bogus counts will be
490          * returned on read(). The only way to get out
491          * of error state is by explicit re-enabling
492          * of the event
493          */
494         if (event->state > PERF_EVENT_STATE_OFF)
495                 event->state = PERF_EVENT_STATE_OFF;
496 }
497
498 static void perf_group_detach(struct perf_event *event)
499 {
500         struct perf_event *sibling, *tmp;
501         struct list_head *list = NULL;
502
503         /*
504          * We can have double detach due to exit/hot-unplug + close.
505          */
506         if (!(event->attach_state & PERF_ATTACH_GROUP))
507                 return;
508
509         event->attach_state &= ~PERF_ATTACH_GROUP;
510
511         /*
512          * If this is a sibling, remove it from its group.
513          */
514         if (event->group_leader != event) {
515                 list_del_init(&event->group_entry);
516                 event->group_leader->nr_siblings--;
517                 goto out;
518         }
519
520         if (!list_empty(&event->group_entry))
521                 list = &event->group_entry;
522
523         /*
524          * If this was a group event with sibling events then
525          * upgrade the siblings to singleton events by adding them
526          * to whatever list we are on.
527          */
528         list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
529                 if (list)
530                         list_move_tail(&sibling->group_entry, list);
531                 sibling->group_leader = sibling;
532
533                 /* Inherit group flags from the previous leader */
534                 sibling->group_flags = event->group_flags;
535         }
536
537 out:
538         perf_event__header_size(event->group_leader);
539
540         list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
541                 perf_event__header_size(tmp);
542 }
543
544 static inline int
545 event_filter_match(struct perf_event *event)
546 {
547         return event->cpu == -1 || event->cpu == smp_processor_id();
548 }
549
550 static void
551 event_sched_out(struct perf_event *event,
552                   struct perf_cpu_context *cpuctx,
553                   struct perf_event_context *ctx)
554 {
555         u64 tstamp = perf_event_time(event);
556         u64 delta;
557         /*
558          * An event which could not be activated because of
559          * filter mismatch still needs to have its timings
560          * maintained, otherwise bogus information is return
561          * via read() for time_enabled, time_running:
562          */
563         if (event->state == PERF_EVENT_STATE_INACTIVE
564             && !event_filter_match(event)) {
565                 delta = ctx->time - event->tstamp_stopped;
566                 event->tstamp_running += delta;
567                 event->tstamp_stopped = tstamp;
568         }
569
570         if (event->state != PERF_EVENT_STATE_ACTIVE)
571                 return;
572
573         event->state = PERF_EVENT_STATE_INACTIVE;
574         if (event->pending_disable) {
575                 event->pending_disable = 0;
576                 event->state = PERF_EVENT_STATE_OFF;
577         }
578         event->tstamp_stopped = tstamp;
579         event->pmu->del(event, 0);
580         event->oncpu = -1;
581
582         if (!is_software_event(event))
583                 cpuctx->active_oncpu--;
584         ctx->nr_active--;
585         if (event->attr.exclusive || !cpuctx->active_oncpu)
586                 cpuctx->exclusive = 0;
587 }
588
589 static void
590 group_sched_out(struct perf_event *group_event,
591                 struct perf_cpu_context *cpuctx,
592                 struct perf_event_context *ctx)
593 {
594         struct perf_event *event;
595         int state = group_event->state;
596
597         event_sched_out(group_event, cpuctx, ctx);
598
599         /*
600          * Schedule out siblings (if any):
601          */
602         list_for_each_entry(event, &group_event->sibling_list, group_entry)
603                 event_sched_out(event, cpuctx, ctx);
604
605         if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
606                 cpuctx->exclusive = 0;
607 }
608
609 static inline struct perf_cpu_context *
610 __get_cpu_context(struct perf_event_context *ctx)
611 {
612         return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
613 }
614
615 /*
616  * Cross CPU call to remove a performance event
617  *
618  * We disable the event on the hardware level first. After that we
619  * remove it from the context list.
620  */
621 static void __perf_event_remove_from_context(void *info)
622 {
623         struct perf_event *event = info;
624         struct perf_event_context *ctx = event->ctx;
625         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
626
627         /*
628          * If this is a task context, we need to check whether it is
629          * the current task context of this cpu. If not it has been
630          * scheduled out before the smp call arrived.
631          */
632         if (ctx->task && cpuctx->task_ctx != ctx)
633                 return;
634
635         raw_spin_lock(&ctx->lock);
636
637         event_sched_out(event, cpuctx, ctx);
638
639         list_del_event(event, ctx);
640
641         raw_spin_unlock(&ctx->lock);
642 }
643
644
645 /*
646  * Remove the event from a task's (or a CPU's) list of events.
647  *
648  * Must be called with ctx->mutex held.
649  *
650  * CPU events are removed with a smp call. For task events we only
651  * call when the task is on a CPU.
652  *
653  * If event->ctx is a cloned context, callers must make sure that
654  * every task struct that event->ctx->task could possibly point to
655  * remains valid.  This is OK when called from perf_release since
656  * that only calls us on the top-level context, which can't be a clone.
657  * When called from perf_event_exit_task, it's OK because the
658  * context has been detached from its task.
659  */
660 static void perf_event_remove_from_context(struct perf_event *event)
661 {
662         struct perf_event_context *ctx = event->ctx;
663         struct task_struct *task = ctx->task;
664
665         if (!task) {
666                 /*
667                  * Per cpu events are removed via an smp call and
668                  * the removal is always successful.
669                  */
670                 smp_call_function_single(event->cpu,
671                                          __perf_event_remove_from_context,
672                                          event, 1);
673                 return;
674         }
675
676 retry:
677         task_oncpu_function_call(task, __perf_event_remove_from_context,
678                                  event);
679
680         raw_spin_lock_irq(&ctx->lock);
681         /*
682          * If the context is active we need to retry the smp call.
683          */
684         if (ctx->nr_active && !list_empty(&event->group_entry)) {
685                 raw_spin_unlock_irq(&ctx->lock);
686                 goto retry;
687         }
688
689         /*
690          * The lock prevents that this context is scheduled in so we
691          * can remove the event safely, if the call above did not
692          * succeed.
693          */
694         if (!list_empty(&event->group_entry))
695                 list_del_event(event, ctx);
696         raw_spin_unlock_irq(&ctx->lock);
697 }
698
699 /*
700  * Cross CPU call to disable a performance event
701  */
702 static void __perf_event_disable(void *info)
703 {
704         struct perf_event *event = info;
705         struct perf_event_context *ctx = event->ctx;
706         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
707
708         /*
709          * If this is a per-task event, need to check whether this
710          * event's task is the current task on this cpu.
711          */
712         if (ctx->task && cpuctx->task_ctx != ctx)
713                 return;
714
715         raw_spin_lock(&ctx->lock);
716
717         /*
718          * If the event is on, turn it off.
719          * If it is in error state, leave it in error state.
720          */
721         if (event->state >= PERF_EVENT_STATE_INACTIVE) {
722                 update_context_time(ctx);
723                 update_group_times(event);
724                 if (event == event->group_leader)
725                         group_sched_out(event, cpuctx, ctx);
726                 else
727                         event_sched_out(event, cpuctx, ctx);
728                 event->state = PERF_EVENT_STATE_OFF;
729         }
730
731         raw_spin_unlock(&ctx->lock);
732 }
733
734 /*
735  * Disable a event.
736  *
737  * If event->ctx is a cloned context, callers must make sure that
738  * every task struct that event->ctx->task could possibly point to
739  * remains valid.  This condition is satisifed when called through
740  * perf_event_for_each_child or perf_event_for_each because they
741  * hold the top-level event's child_mutex, so any descendant that
742  * goes to exit will block in sync_child_event.
743  * When called from perf_pending_event it's OK because event->ctx
744  * is the current context on this CPU and preemption is disabled,
745  * hence we can't get into perf_event_task_sched_out for this context.
746  */
747 void perf_event_disable(struct perf_event *event)
748 {
749         struct perf_event_context *ctx = event->ctx;
750         struct task_struct *task = ctx->task;
751
752         if (!task) {
753                 /*
754                  * Disable the event on the cpu that it's on
755                  */
756                 smp_call_function_single(event->cpu, __perf_event_disable,
757                                          event, 1);
758                 return;
759         }
760
761 retry:
762         task_oncpu_function_call(task, __perf_event_disable, event);
763
764         raw_spin_lock_irq(&ctx->lock);
765         /*
766          * If the event is still active, we need to retry the cross-call.
767          */
768         if (event->state == PERF_EVENT_STATE_ACTIVE) {
769                 raw_spin_unlock_irq(&ctx->lock);
770                 goto retry;
771         }
772
773         /*
774          * Since we have the lock this context can't be scheduled
775          * in, so we can change the state safely.
776          */
777         if (event->state == PERF_EVENT_STATE_INACTIVE) {
778                 update_group_times(event);
779                 event->state = PERF_EVENT_STATE_OFF;
780         }
781
782         raw_spin_unlock_irq(&ctx->lock);
783 }
784
785 #define MAX_INTERRUPTS (~0ULL)
786
787 static void perf_log_throttle(struct perf_event *event, int enable);
788
789 static int
790 event_sched_in(struct perf_event *event,
791                  struct perf_cpu_context *cpuctx,
792                  struct perf_event_context *ctx)
793 {
794         u64 tstamp = perf_event_time(event);
795
796         if (event->state <= PERF_EVENT_STATE_OFF)
797                 return 0;
798
799         event->state = PERF_EVENT_STATE_ACTIVE;
800         event->oncpu = smp_processor_id();
801
802         /*
803          * Unthrottle events, since we scheduled we might have missed several
804          * ticks already, also for a heavily scheduling task there is little
805          * guarantee it'll get a tick in a timely manner.
806          */
807         if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
808                 perf_log_throttle(event, 1);
809                 event->hw.interrupts = 0;
810         }
811
812         /*
813          * The new state must be visible before we turn it on in the hardware:
814          */
815         smp_wmb();
816
817         if (event->pmu->add(event, PERF_EF_START)) {
818                 event->state = PERF_EVENT_STATE_INACTIVE;
819                 event->oncpu = -1;
820                 return -EAGAIN;
821         }
822
823         event->tstamp_running += tstamp - event->tstamp_stopped;
824
825         event->shadow_ctx_time = tstamp - ctx->timestamp;
826
827         if (!is_software_event(event))
828                 cpuctx->active_oncpu++;
829         ctx->nr_active++;
830
831         if (event->attr.exclusive)
832                 cpuctx->exclusive = 1;
833
834         return 0;
835 }
836
837 static int
838 group_sched_in(struct perf_event *group_event,
839                struct perf_cpu_context *cpuctx,
840                struct perf_event_context *ctx)
841 {
842         struct perf_event *event, *partial_group = NULL;
843         struct pmu *pmu = group_event->pmu;
844         u64 now = ctx->time;
845         bool simulate = false;
846
847         if (group_event->state == PERF_EVENT_STATE_OFF)
848                 return 0;
849
850         pmu->start_txn(pmu);
851
852         if (event_sched_in(group_event, cpuctx, ctx)) {
853                 pmu->cancel_txn(pmu);
854                 return -EAGAIN;
855         }
856
857         /*
858          * Schedule in siblings as one group (if any):
859          */
860         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
861                 if (event_sched_in(event, cpuctx, ctx)) {
862                         partial_group = event;
863                         goto group_error;
864                 }
865         }
866
867         if (!pmu->commit_txn(pmu))
868                 return 0;
869
870 group_error:
871         /*
872          * Groups can be scheduled in as one unit only, so undo any
873          * partial group before returning:
874          * The events up to the failed event are scheduled out normally,
875          * tstamp_stopped will be updated.
876          *
877          * The failed events and the remaining siblings need to have
878          * their timings updated as if they had gone thru event_sched_in()
879          * and event_sched_out(). This is required to get consistent timings
880          * across the group. This also takes care of the case where the group
881          * could never be scheduled by ensuring tstamp_stopped is set to mark
882          * the time the event was actually stopped, such that time delta
883          * calculation in update_event_times() is correct.
884          */
885         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
886                 if (event == partial_group)
887                         simulate = true;
888
889                 if (simulate) {
890                         event->tstamp_running += now - event->tstamp_stopped;
891                         event->tstamp_stopped = now;
892                 } else {
893                         event_sched_out(event, cpuctx, ctx);
894                 }
895         }
896         event_sched_out(group_event, cpuctx, ctx);
897
898         pmu->cancel_txn(pmu);
899
900         return -EAGAIN;
901 }
902
903 /*
904  * Work out whether we can put this event group on the CPU now.
905  */
906 static int group_can_go_on(struct perf_event *event,
907                            struct perf_cpu_context *cpuctx,
908                            int can_add_hw)
909 {
910         /*
911          * Groups consisting entirely of software events can always go on.
912          */
913         if (event->group_flags & PERF_GROUP_SOFTWARE)
914                 return 1;
915         /*
916          * If an exclusive group is already on, no other hardware
917          * events can go on.
918          */
919         if (cpuctx->exclusive)
920                 return 0;
921         /*
922          * If this group is exclusive and there are already
923          * events on the CPU, it can't go on.
924          */
925         if (event->attr.exclusive && cpuctx->active_oncpu)
926                 return 0;
927         /*
928          * Otherwise, try to add it if all previous groups were able
929          * to go on.
930          */
931         return can_add_hw;
932 }
933
934 static void add_event_to_ctx(struct perf_event *event,
935                                struct perf_event_context *ctx)
936 {
937         u64 tstamp = perf_event_time(event);
938
939         list_add_event(event, ctx);
940         perf_group_attach(event);
941         event->tstamp_enabled = tstamp;
942         event->tstamp_running = tstamp;
943         event->tstamp_stopped = tstamp;
944 }
945
946 /*
947  * Cross CPU call to install and enable a performance event
948  *
949  * Must be called with ctx->mutex held
950  */
951 static void __perf_install_in_context(void *info)
952 {
953         struct perf_event *event = info;
954         struct perf_event_context *ctx = event->ctx;
955         struct perf_event *leader = event->group_leader;
956         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
957         int err;
958
959         /*
960          * If this is a task context, we need to check whether it is
961          * the current task context of this cpu. If not it has been
962          * scheduled out before the smp call arrived.
963          * Or possibly this is the right context but it isn't
964          * on this cpu because it had no events.
965          */
966         if (ctx->task && cpuctx->task_ctx != ctx) {
967                 if (cpuctx->task_ctx || ctx->task != current)
968                         return;
969                 cpuctx->task_ctx = ctx;
970         }
971
972         raw_spin_lock(&ctx->lock);
973         ctx->is_active = 1;
974         update_context_time(ctx);
975
976         add_event_to_ctx(event, ctx);
977
978         if (!event_filter_match(event))
979                 goto unlock;
980
981         /*
982          * Don't put the event on if it is disabled or if
983          * it is in a group and the group isn't on.
984          */
985         if (event->state != PERF_EVENT_STATE_INACTIVE ||
986             (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
987                 goto unlock;
988
989         /*
990          * An exclusive event can't go on if there are already active
991          * hardware events, and no hardware event can go on if there
992          * is already an exclusive event on.
993          */
994         if (!group_can_go_on(event, cpuctx, 1))
995                 err = -EEXIST;
996         else
997                 err = event_sched_in(event, cpuctx, ctx);
998
999         if (err) {
1000                 /*
1001                  * This event couldn't go on.  If it is in a group
1002                  * then we have to pull the whole group off.
1003                  * If the event group is pinned then put it in error state.
1004                  */
1005                 if (leader != event)
1006                         group_sched_out(leader, cpuctx, ctx);
1007                 if (leader->attr.pinned) {
1008                         update_group_times(leader);
1009                         leader->state = PERF_EVENT_STATE_ERROR;
1010                 }
1011         }
1012
1013 unlock:
1014         raw_spin_unlock(&ctx->lock);
1015 }
1016
1017 /*
1018  * Attach a performance event to a context
1019  *
1020  * First we add the event to the list with the hardware enable bit
1021  * in event->hw_config cleared.
1022  *
1023  * If the event is attached to a task which is on a CPU we use a smp
1024  * call to enable it in the task context. The task might have been
1025  * scheduled away, but we check this in the smp call again.
1026  *
1027  * Must be called with ctx->mutex held.
1028  */
1029 static void
1030 perf_install_in_context(struct perf_event_context *ctx,
1031                         struct perf_event *event,
1032                         int cpu)
1033 {
1034         struct task_struct *task = ctx->task;
1035
1036         event->ctx = ctx;
1037
1038         if (!task) {
1039                 /*
1040                  * Per cpu events are installed via an smp call and
1041                  * the install is always successful.
1042                  */
1043                 smp_call_function_single(cpu, __perf_install_in_context,
1044                                          event, 1);
1045                 return;
1046         }
1047
1048 retry:
1049         task_oncpu_function_call(task, __perf_install_in_context,
1050                                  event);
1051
1052         raw_spin_lock_irq(&ctx->lock);
1053         /*
1054          * we need to retry the smp call.
1055          */
1056         if (ctx->is_active && list_empty(&event->group_entry)) {
1057                 raw_spin_unlock_irq(&ctx->lock);
1058                 goto retry;
1059         }
1060
1061         /*
1062          * The lock prevents that this context is scheduled in so we
1063          * can add the event safely, if it the call above did not
1064          * succeed.
1065          */
1066         if (list_empty(&event->group_entry))
1067                 add_event_to_ctx(event, ctx);
1068         raw_spin_unlock_irq(&ctx->lock);
1069 }
1070
1071 /*
1072  * Put a event into inactive state and update time fields.
1073  * Enabling the leader of a group effectively enables all
1074  * the group members that aren't explicitly disabled, so we
1075  * have to update their ->tstamp_enabled also.
1076  * Note: this works for group members as well as group leaders
1077  * since the non-leader members' sibling_lists will be empty.
1078  */
1079 static void __perf_event_mark_enabled(struct perf_event *event,
1080                                         struct perf_event_context *ctx)
1081 {
1082         struct perf_event *sub;
1083         u64 tstamp = perf_event_time(event);
1084
1085         event->state = PERF_EVENT_STATE_INACTIVE;
1086         event->tstamp_enabled = tstamp - event->total_time_enabled;
1087         list_for_each_entry(sub, &event->sibling_list, group_entry) {
1088                 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1089                         sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1090         }
1091 }
1092
1093 /*
1094  * Cross CPU call to enable a performance event
1095  */
1096 static void __perf_event_enable(void *info)
1097 {
1098         struct perf_event *event = info;
1099         struct perf_event_context *ctx = event->ctx;
1100         struct perf_event *leader = event->group_leader;
1101         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1102         int err;
1103
1104         /*
1105          * If this is a per-task event, need to check whether this
1106          * event's task is the current task on this cpu.
1107          */
1108         if (ctx->task && cpuctx->task_ctx != ctx) {
1109                 if (cpuctx->task_ctx || ctx->task != current)
1110                         return;
1111                 cpuctx->task_ctx = ctx;
1112         }
1113
1114         raw_spin_lock(&ctx->lock);
1115         ctx->is_active = 1;
1116         update_context_time(ctx);
1117
1118         if (event->state >= PERF_EVENT_STATE_INACTIVE)
1119                 goto unlock;
1120         __perf_event_mark_enabled(event, ctx);
1121
1122         if (!event_filter_match(event))
1123                 goto unlock;
1124
1125         /*
1126          * If the event is in a group and isn't the group leader,
1127          * then don't put it on unless the group is on.
1128          */
1129         if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
1130                 goto unlock;
1131
1132         if (!group_can_go_on(event, cpuctx, 1)) {
1133                 err = -EEXIST;
1134         } else {
1135                 if (event == leader)
1136                         err = group_sched_in(event, cpuctx, ctx);
1137                 else
1138                         err = event_sched_in(event, cpuctx, ctx);
1139         }
1140
1141         if (err) {
1142                 /*
1143                  * If this event can't go on and it's part of a
1144                  * group, then the whole group has to come off.
1145                  */
1146                 if (leader != event)
1147                         group_sched_out(leader, cpuctx, ctx);
1148                 if (leader->attr.pinned) {
1149                         update_group_times(leader);
1150                         leader->state = PERF_EVENT_STATE_ERROR;
1151                 }
1152         }
1153
1154 unlock:
1155         raw_spin_unlock(&ctx->lock);
1156 }
1157
1158 /*
1159  * Enable a event.
1160  *
1161  * If event->ctx is a cloned context, callers must make sure that
1162  * every task struct that event->ctx->task could possibly point to
1163  * remains valid.  This condition is satisfied when called through
1164  * perf_event_for_each_child or perf_event_for_each as described
1165  * for perf_event_disable.
1166  */
1167 void perf_event_enable(struct perf_event *event)
1168 {
1169         struct perf_event_context *ctx = event->ctx;
1170         struct task_struct *task = ctx->task;
1171
1172         if (!task) {
1173                 /*
1174                  * Enable the event on the cpu that it's on
1175                  */
1176                 smp_call_function_single(event->cpu, __perf_event_enable,
1177                                          event, 1);
1178                 return;
1179         }
1180
1181         raw_spin_lock_irq(&ctx->lock);
1182         if (event->state >= PERF_EVENT_STATE_INACTIVE)
1183                 goto out;
1184
1185         /*
1186          * If the event is in error state, clear that first.
1187          * That way, if we see the event in error state below, we
1188          * know that it has gone back into error state, as distinct
1189          * from the task having been scheduled away before the
1190          * cross-call arrived.
1191          */
1192         if (event->state == PERF_EVENT_STATE_ERROR)
1193                 event->state = PERF_EVENT_STATE_OFF;
1194
1195 retry:
1196         raw_spin_unlock_irq(&ctx->lock);
1197         task_oncpu_function_call(task, __perf_event_enable, event);
1198
1199         raw_spin_lock_irq(&ctx->lock);
1200
1201         /*
1202          * If the context is active and the event is still off,
1203          * we need to retry the cross-call.
1204          */
1205         if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1206                 goto retry;
1207
1208         /*
1209          * Since we have the lock this context can't be scheduled
1210          * in, so we can change the state safely.
1211          */
1212         if (event->state == PERF_EVENT_STATE_OFF)
1213                 __perf_event_mark_enabled(event, ctx);
1214
1215 out:
1216         raw_spin_unlock_irq(&ctx->lock);
1217 }
1218
1219 static int perf_event_refresh(struct perf_event *event, int refresh)
1220 {
1221         /*
1222          * not supported on inherited events
1223          */
1224         if (event->attr.inherit || !is_sampling_event(event))
1225                 return -EINVAL;
1226
1227         atomic_add(refresh, &event->event_limit);
1228         perf_event_enable(event);
1229
1230         return 0;
1231 }
1232
1233 static void ctx_sched_out(struct perf_event_context *ctx,
1234                           struct perf_cpu_context *cpuctx,
1235                           enum event_type_t event_type)
1236 {
1237         struct perf_event *event;
1238
1239         raw_spin_lock(&ctx->lock);
1240         perf_pmu_disable(ctx->pmu);
1241         ctx->is_active = 0;
1242         if (likely(!ctx->nr_events))
1243                 goto out;
1244         update_context_time(ctx);
1245
1246         if (!ctx->nr_active)
1247                 goto out;
1248
1249         if (event_type & EVENT_PINNED) {
1250                 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1251                         group_sched_out(event, cpuctx, ctx);
1252         }
1253
1254         if (event_type & EVENT_FLEXIBLE) {
1255                 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1256                         group_sched_out(event, cpuctx, ctx);
1257         }
1258 out:
1259         perf_pmu_enable(ctx->pmu);
1260         raw_spin_unlock(&ctx->lock);
1261 }
1262
1263 /*
1264  * Test whether two contexts are equivalent, i.e. whether they
1265  * have both been cloned from the same version of the same context
1266  * and they both have the same number of enabled events.
1267  * If the number of enabled events is the same, then the set
1268  * of enabled events should be the same, because these are both
1269  * inherited contexts, therefore we can't access individual events
1270  * in them directly with an fd; we can only enable/disable all
1271  * events via prctl, or enable/disable all events in a family
1272  * via ioctl, which will have the same effect on both contexts.
1273  */
1274 static int context_equiv(struct perf_event_context *ctx1,
1275                          struct perf_event_context *ctx2)
1276 {
1277         return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1278                 && ctx1->parent_gen == ctx2->parent_gen
1279                 && !ctx1->pin_count && !ctx2->pin_count;
1280 }
1281
1282 static void __perf_event_sync_stat(struct perf_event *event,
1283                                      struct perf_event *next_event)
1284 {
1285         u64 value;
1286
1287         if (!event->attr.inherit_stat)
1288                 return;
1289
1290         /*
1291          * Update the event value, we cannot use perf_event_read()
1292          * because we're in the middle of a context switch and have IRQs
1293          * disabled, which upsets smp_call_function_single(), however
1294          * we know the event must be on the current CPU, therefore we
1295          * don't need to use it.
1296          */
1297         switch (event->state) {
1298         case PERF_EVENT_STATE_ACTIVE:
1299                 event->pmu->read(event);
1300                 /* fall-through */
1301
1302         case PERF_EVENT_STATE_INACTIVE:
1303                 update_event_times(event);
1304                 break;
1305
1306         default:
1307                 break;
1308         }
1309
1310         /*
1311          * In order to keep per-task stats reliable we need to flip the event
1312          * values when we flip the contexts.
1313          */
1314         value = local64_read(&next_event->count);
1315         value = local64_xchg(&event->count, value);
1316         local64_set(&next_event->count, value);
1317
1318         swap(event->total_time_enabled, next_event->total_time_enabled);
1319         swap(event->total_time_running, next_event->total_time_running);
1320
1321         /*
1322          * Since we swizzled the values, update the user visible data too.
1323          */
1324         perf_event_update_userpage(event);
1325         perf_event_update_userpage(next_event);
1326 }
1327
1328 #define list_next_entry(pos, member) \
1329         list_entry(pos->member.next, typeof(*pos), member)
1330
1331 static void perf_event_sync_stat(struct perf_event_context *ctx,
1332                                    struct perf_event_context *next_ctx)
1333 {
1334         struct perf_event *event, *next_event;
1335
1336         if (!ctx->nr_stat)
1337                 return;
1338
1339         update_context_time(ctx);
1340
1341         event = list_first_entry(&ctx->event_list,
1342                                    struct perf_event, event_entry);
1343
1344         next_event = list_first_entry(&next_ctx->event_list,
1345                                         struct perf_event, event_entry);
1346
1347         while (&event->event_entry != &ctx->event_list &&
1348                &next_event->event_entry != &next_ctx->event_list) {
1349
1350                 __perf_event_sync_stat(event, next_event);
1351
1352                 event = list_next_entry(event, event_entry);
1353                 next_event = list_next_entry(next_event, event_entry);
1354         }
1355 }
1356
1357 void perf_event_context_sched_out(struct task_struct *task, int ctxn,
1358                                   struct task_struct *next)
1359 {
1360         struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
1361         struct perf_event_context *next_ctx;
1362         struct perf_event_context *parent;
1363         struct perf_cpu_context *cpuctx;
1364         int do_switch = 1;
1365
1366         if (likely(!ctx))
1367                 return;
1368
1369         cpuctx = __get_cpu_context(ctx);
1370         if (!cpuctx->task_ctx)
1371                 return;
1372
1373         rcu_read_lock();
1374         parent = rcu_dereference(ctx->parent_ctx);
1375         next_ctx = next->perf_event_ctxp[ctxn];
1376         if (parent && next_ctx &&
1377             rcu_dereference(next_ctx->parent_ctx) == parent) {
1378                 /*
1379                  * Looks like the two contexts are clones, so we might be
1380                  * able to optimize the context switch.  We lock both
1381                  * contexts and check that they are clones under the
1382                  * lock (including re-checking that neither has been
1383                  * uncloned in the meantime).  It doesn't matter which
1384                  * order we take the locks because no other cpu could
1385                  * be trying to lock both of these tasks.
1386                  */
1387                 raw_spin_lock(&ctx->lock);
1388                 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1389                 if (context_equiv(ctx, next_ctx)) {
1390                         /*
1391                          * XXX do we need a memory barrier of sorts
1392                          * wrt to rcu_dereference() of perf_event_ctxp
1393                          */
1394                         task->perf_event_ctxp[ctxn] = next_ctx;
1395                         next->perf_event_ctxp[ctxn] = ctx;
1396                         ctx->task = next;
1397                         next_ctx->task = task;
1398                         do_switch = 0;
1399
1400                         perf_event_sync_stat(ctx, next_ctx);
1401                 }
1402                 raw_spin_unlock(&next_ctx->lock);
1403                 raw_spin_unlock(&ctx->lock);
1404         }
1405         rcu_read_unlock();
1406
1407         if (do_switch) {
1408                 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1409                 cpuctx->task_ctx = NULL;
1410         }
1411 }
1412
1413 #define for_each_task_context_nr(ctxn)                                  \
1414         for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
1415
1416 /*
1417  * Called from scheduler to remove the events of the current task,
1418  * with interrupts disabled.
1419  *
1420  * We stop each event and update the event value in event->count.
1421  *
1422  * This does not protect us against NMI, but disable()
1423  * sets the disabled bit in the control field of event _before_
1424  * accessing the event control register. If a NMI hits, then it will
1425  * not restart the event.
1426  */
1427 void __perf_event_task_sched_out(struct task_struct *task,
1428                                  struct task_struct *next)
1429 {
1430         int ctxn;
1431
1432         for_each_task_context_nr(ctxn)
1433                 perf_event_context_sched_out(task, ctxn, next);
1434 }
1435
1436 static void task_ctx_sched_out(struct perf_event_context *ctx,
1437                                enum event_type_t event_type)
1438 {
1439         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1440
1441         if (!cpuctx->task_ctx)
1442                 return;
1443
1444         if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1445                 return;
1446
1447         ctx_sched_out(ctx, cpuctx, event_type);
1448         cpuctx->task_ctx = NULL;
1449 }
1450
1451 /*
1452  * Called with IRQs disabled
1453  */
1454 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1455                               enum event_type_t event_type)
1456 {
1457         ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1458 }
1459
1460 static void
1461 ctx_pinned_sched_in(struct perf_event_context *ctx,
1462                     struct perf_cpu_context *cpuctx)
1463 {
1464         struct perf_event *event;
1465
1466         list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1467                 if (event->state <= PERF_EVENT_STATE_OFF)
1468                         continue;
1469                 if (!event_filter_match(event))
1470                         continue;
1471
1472                 if (group_can_go_on(event, cpuctx, 1))
1473                         group_sched_in(event, cpuctx, ctx);
1474
1475                 /*
1476                  * If this pinned group hasn't been scheduled,
1477                  * put it in error state.
1478                  */
1479                 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1480                         update_group_times(event);
1481                         event->state = PERF_EVENT_STATE_ERROR;
1482                 }
1483         }
1484 }
1485
1486 static void
1487 ctx_flexible_sched_in(struct perf_event_context *ctx,
1488                       struct perf_cpu_context *cpuctx)
1489 {
1490         struct perf_event *event;
1491         int can_add_hw = 1;
1492
1493         list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1494                 /* Ignore events in OFF or ERROR state */
1495                 if (event->state <= PERF_EVENT_STATE_OFF)
1496                         continue;
1497                 /*
1498                  * Listen to the 'cpu' scheduling filter constraint
1499                  * of events:
1500                  */
1501                 if (!event_filter_match(event))
1502                         continue;
1503
1504                 if (group_can_go_on(event, cpuctx, can_add_hw)) {
1505                         if (group_sched_in(event, cpuctx, ctx))
1506                                 can_add_hw = 0;
1507                 }
1508         }
1509 }
1510
1511 static void
1512 ctx_sched_in(struct perf_event_context *ctx,
1513              struct perf_cpu_context *cpuctx,
1514              enum event_type_t event_type)
1515 {
1516         raw_spin_lock(&ctx->lock);
1517         ctx->is_active = 1;
1518         if (likely(!ctx->nr_events))
1519                 goto out;
1520
1521         ctx->timestamp = perf_clock();
1522
1523         /*
1524          * First go through the list and put on any pinned groups
1525          * in order to give them the best chance of going on.
1526          */
1527         if (event_type & EVENT_PINNED)
1528                 ctx_pinned_sched_in(ctx, cpuctx);
1529
1530         /* Then walk through the lower prio flexible groups */
1531         if (event_type & EVENT_FLEXIBLE)
1532                 ctx_flexible_sched_in(ctx, cpuctx);
1533
1534 out:
1535         raw_spin_unlock(&ctx->lock);
1536 }
1537
1538 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1539                              enum event_type_t event_type)
1540 {
1541         struct perf_event_context *ctx = &cpuctx->ctx;
1542
1543         ctx_sched_in(ctx, cpuctx, event_type);
1544 }
1545
1546 static void task_ctx_sched_in(struct perf_event_context *ctx,
1547                               enum event_type_t event_type)
1548 {
1549         struct perf_cpu_context *cpuctx;
1550
1551         cpuctx = __get_cpu_context(ctx);
1552         if (cpuctx->task_ctx == ctx)
1553                 return;
1554
1555         ctx_sched_in(ctx, cpuctx, event_type);
1556         cpuctx->task_ctx = ctx;
1557 }
1558
1559 void perf_event_context_sched_in(struct perf_event_context *ctx)
1560 {
1561         struct perf_cpu_context *cpuctx;
1562
1563         cpuctx = __get_cpu_context(ctx);
1564         if (cpuctx->task_ctx == ctx)
1565                 return;
1566
1567         perf_pmu_disable(ctx->pmu);
1568         /*
1569          * We want to keep the following priority order:
1570          * cpu pinned (that don't need to move), task pinned,
1571          * cpu flexible, task flexible.
1572          */
1573         cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1574
1575         ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1576         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1577         ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1578
1579         cpuctx->task_ctx = ctx;
1580
1581         /*
1582          * Since these rotations are per-cpu, we need to ensure the
1583          * cpu-context we got scheduled on is actually rotating.
1584          */
1585         perf_pmu_rotate_start(ctx->pmu);
1586         perf_pmu_enable(ctx->pmu);
1587 }
1588
1589 /*
1590  * Called from scheduler to add the events of the current task
1591  * with interrupts disabled.
1592  *
1593  * We restore the event value and then enable it.
1594  *
1595  * This does not protect us against NMI, but enable()
1596  * sets the enabled bit in the control field of event _before_
1597  * accessing the event control register. If a NMI hits, then it will
1598  * keep the event running.
1599  */
1600 void __perf_event_task_sched_in(struct task_struct *task)
1601 {
1602         struct perf_event_context *ctx;
1603         int ctxn;
1604
1605         for_each_task_context_nr(ctxn) {
1606                 ctx = task->perf_event_ctxp[ctxn];
1607                 if (likely(!ctx))
1608                         continue;
1609
1610                 perf_event_context_sched_in(ctx);
1611         }
1612 }
1613
1614 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1615 {
1616         u64 frequency = event->attr.sample_freq;
1617         u64 sec = NSEC_PER_SEC;
1618         u64 divisor, dividend;
1619
1620         int count_fls, nsec_fls, frequency_fls, sec_fls;
1621
1622         count_fls = fls64(count);
1623         nsec_fls = fls64(nsec);
1624         frequency_fls = fls64(frequency);
1625         sec_fls = 30;
1626
1627         /*
1628          * We got @count in @nsec, with a target of sample_freq HZ
1629          * the target period becomes:
1630          *
1631          *             @count * 10^9
1632          * period = -------------------
1633          *          @nsec * sample_freq
1634          *
1635          */
1636
1637         /*
1638          * Reduce accuracy by one bit such that @a and @b converge
1639          * to a similar magnitude.
1640          */
1641 #define REDUCE_FLS(a, b)                \
1642 do {                                    \
1643         if (a##_fls > b##_fls) {        \
1644                 a >>= 1;                \
1645                 a##_fls--;              \
1646         } else {                        \
1647                 b >>= 1;                \
1648                 b##_fls--;              \
1649         }                               \
1650 } while (0)
1651
1652         /*
1653          * Reduce accuracy until either term fits in a u64, then proceed with
1654          * the other, so that finally we can do a u64/u64 division.
1655          */
1656         while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1657                 REDUCE_FLS(nsec, frequency);
1658                 REDUCE_FLS(sec, count);
1659         }
1660
1661         if (count_fls + sec_fls > 64) {
1662                 divisor = nsec * frequency;
1663
1664                 while (count_fls + sec_fls > 64) {
1665                         REDUCE_FLS(count, sec);
1666                         divisor >>= 1;
1667                 }
1668
1669                 dividend = count * sec;
1670         } else {
1671                 dividend = count * sec;
1672
1673                 while (nsec_fls + frequency_fls > 64) {
1674                         REDUCE_FLS(nsec, frequency);
1675                         dividend >>= 1;
1676                 }
1677
1678                 divisor = nsec * frequency;
1679         }
1680
1681         if (!divisor)
1682                 return dividend;
1683
1684         return div64_u64(dividend, divisor);
1685 }
1686
1687 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1688 {
1689         struct hw_perf_event *hwc = &event->hw;
1690         s64 period, sample_period;
1691         s64 delta;
1692
1693         period = perf_calculate_period(event, nsec, count);
1694
1695         delta = (s64)(period - hwc->sample_period);
1696         delta = (delta + 7) / 8; /* low pass filter */
1697
1698         sample_period = hwc->sample_period + delta;
1699
1700         if (!sample_period)
1701                 sample_period = 1;
1702
1703         hwc->sample_period = sample_period;
1704
1705         if (local64_read(&hwc->period_left) > 8*sample_period) {
1706                 event->pmu->stop(event, PERF_EF_UPDATE);
1707                 local64_set(&hwc->period_left, 0);
1708                 event->pmu->start(event, PERF_EF_RELOAD);
1709         }
1710 }
1711
1712 static void perf_ctx_adjust_freq(struct perf_event_context *ctx, u64 period)
1713 {
1714         struct perf_event *event;
1715         struct hw_perf_event *hwc;
1716         u64 interrupts, now;
1717         s64 delta;
1718
1719         raw_spin_lock(&ctx->lock);
1720         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1721                 if (event->state != PERF_EVENT_STATE_ACTIVE)
1722                         continue;
1723
1724                 if (!event_filter_match(event))
1725                         continue;
1726
1727                 hwc = &event->hw;
1728
1729                 interrupts = hwc->interrupts;
1730                 hwc->interrupts = 0;
1731
1732                 /*
1733                  * unthrottle events on the tick
1734                  */
1735                 if (interrupts == MAX_INTERRUPTS) {
1736                         perf_log_throttle(event, 1);
1737                         event->pmu->start(event, 0);
1738                 }
1739
1740                 if (!event->attr.freq || !event->attr.sample_freq)
1741                         continue;
1742
1743                 event->pmu->read(event);
1744                 now = local64_read(&event->count);
1745                 delta = now - hwc->freq_count_stamp;
1746                 hwc->freq_count_stamp = now;
1747
1748                 if (delta > 0)
1749                         perf_adjust_period(event, period, delta);
1750         }
1751         raw_spin_unlock(&ctx->lock);
1752 }
1753
1754 /*
1755  * Round-robin a context's events:
1756  */
1757 static void rotate_ctx(struct perf_event_context *ctx)
1758 {
1759         raw_spin_lock(&ctx->lock);
1760
1761         /*
1762          * Rotate the first entry last of non-pinned groups. Rotation might be
1763          * disabled by the inheritance code.
1764          */
1765         if (!ctx->rotate_disable)
1766                 list_rotate_left(&ctx->flexible_groups);
1767
1768         raw_spin_unlock(&ctx->lock);
1769 }
1770
1771 /*
1772  * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
1773  * because they're strictly cpu affine and rotate_start is called with IRQs
1774  * disabled, while rotate_context is called from IRQ context.
1775  */
1776 static void perf_rotate_context(struct perf_cpu_context *cpuctx)
1777 {
1778         u64 interval = (u64)cpuctx->jiffies_interval * TICK_NSEC;
1779         struct perf_event_context *ctx = NULL;
1780         int rotate = 0, remove = 1;
1781
1782         if (cpuctx->ctx.nr_events) {
1783                 remove = 0;
1784                 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1785                         rotate = 1;
1786         }
1787
1788         ctx = cpuctx->task_ctx;
1789         if (ctx && ctx->nr_events) {
1790                 remove = 0;
1791                 if (ctx->nr_events != ctx->nr_active)
1792                         rotate = 1;
1793         }
1794
1795         perf_pmu_disable(cpuctx->ctx.pmu);
1796         perf_ctx_adjust_freq(&cpuctx->ctx, interval);
1797         if (ctx)
1798                 perf_ctx_adjust_freq(ctx, interval);
1799
1800         if (!rotate)
1801                 goto done;
1802
1803         cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1804         if (ctx)
1805                 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1806
1807         rotate_ctx(&cpuctx->ctx);
1808         if (ctx)
1809                 rotate_ctx(ctx);
1810
1811         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1812         if (ctx)
1813                 task_ctx_sched_in(ctx, EVENT_FLEXIBLE);
1814
1815 done:
1816         if (remove)
1817                 list_del_init(&cpuctx->rotation_list);
1818
1819         perf_pmu_enable(cpuctx->ctx.pmu);
1820 }
1821
1822 void perf_event_task_tick(void)
1823 {
1824         struct list_head *head = &__get_cpu_var(rotation_list);
1825         struct perf_cpu_context *cpuctx, *tmp;
1826
1827         WARN_ON(!irqs_disabled());
1828
1829         list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
1830                 if (cpuctx->jiffies_interval == 1 ||
1831                                 !(jiffies % cpuctx->jiffies_interval))
1832                         perf_rotate_context(cpuctx);
1833         }
1834 }
1835
1836 static int event_enable_on_exec(struct perf_event *event,
1837                                 struct perf_event_context *ctx)
1838 {
1839         if (!event->attr.enable_on_exec)
1840                 return 0;
1841
1842         event->attr.enable_on_exec = 0;
1843         if (event->state >= PERF_EVENT_STATE_INACTIVE)
1844                 return 0;
1845
1846         __perf_event_mark_enabled(event, ctx);
1847
1848         return 1;
1849 }
1850
1851 /*
1852  * Enable all of a task's events that have been marked enable-on-exec.
1853  * This expects task == current.
1854  */
1855 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
1856 {
1857         struct perf_event *event;
1858         unsigned long flags;
1859         int enabled = 0;
1860         int ret;
1861
1862         local_irq_save(flags);
1863         if (!ctx || !ctx->nr_events)
1864                 goto out;
1865
1866         task_ctx_sched_out(ctx, EVENT_ALL);
1867
1868         raw_spin_lock(&ctx->lock);
1869
1870         list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1871                 ret = event_enable_on_exec(event, ctx);
1872                 if (ret)
1873                         enabled = 1;
1874         }
1875
1876         list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1877                 ret = event_enable_on_exec(event, ctx);
1878                 if (ret)
1879                         enabled = 1;
1880         }
1881
1882         /*
1883          * Unclone this context if we enabled any event.
1884          */
1885         if (enabled)
1886                 unclone_ctx(ctx);
1887
1888         raw_spin_unlock(&ctx->lock);
1889
1890         perf_event_context_sched_in(ctx);
1891 out:
1892         local_irq_restore(flags);
1893 }
1894
1895 /*
1896  * Cross CPU call to read the hardware event
1897  */
1898 static void __perf_event_read(void *info)
1899 {
1900         struct perf_event *event = info;
1901         struct perf_event_context *ctx = event->ctx;
1902         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1903
1904         /*
1905          * If this is a task context, we need to check whether it is
1906          * the current task context of this cpu.  If not it has been
1907          * scheduled out before the smp call arrived.  In that case
1908          * event->count would have been updated to a recent sample
1909          * when the event was scheduled out.
1910          */
1911         if (ctx->task && cpuctx->task_ctx != ctx)
1912                 return;
1913
1914         raw_spin_lock(&ctx->lock);
1915         if (ctx->is_active)
1916                 update_context_time(ctx);
1917         update_event_times(event);
1918         if (event->state == PERF_EVENT_STATE_ACTIVE)
1919                 event->pmu->read(event);
1920         raw_spin_unlock(&ctx->lock);
1921 }
1922
1923 static inline u64 perf_event_count(struct perf_event *event)
1924 {
1925         return local64_read(&event->count) + atomic64_read(&event->child_count);
1926 }
1927
1928 static u64 perf_event_read(struct perf_event *event)
1929 {
1930         /*
1931          * If event is enabled and currently active on a CPU, update the
1932          * value in the event structure:
1933          */
1934         if (event->state == PERF_EVENT_STATE_ACTIVE) {
1935                 smp_call_function_single(event->oncpu,
1936                                          __perf_event_read, event, 1);
1937         } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1938                 struct perf_event_context *ctx = event->ctx;
1939                 unsigned long flags;
1940
1941                 raw_spin_lock_irqsave(&ctx->lock, flags);
1942                 /*
1943                  * may read while context is not active
1944                  * (e.g., thread is blocked), in that case
1945                  * we cannot update context time
1946                  */
1947                 if (ctx->is_active)
1948                         update_context_time(ctx);
1949                 update_event_times(event);
1950                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1951         }
1952
1953         return perf_event_count(event);
1954 }
1955
1956 /*
1957  * Callchain support
1958  */
1959
1960 struct callchain_cpus_entries {
1961         struct rcu_head                 rcu_head;
1962         struct perf_callchain_entry     *cpu_entries[0];
1963 };
1964
1965 static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
1966 static atomic_t nr_callchain_events;
1967 static DEFINE_MUTEX(callchain_mutex);
1968 struct callchain_cpus_entries *callchain_cpus_entries;
1969
1970
1971 __weak void perf_callchain_kernel(struct perf_callchain_entry *entry,
1972                                   struct pt_regs *regs)
1973 {
1974 }
1975
1976 __weak void perf_callchain_user(struct perf_callchain_entry *entry,
1977                                 struct pt_regs *regs)
1978 {
1979 }
1980
1981 static void release_callchain_buffers_rcu(struct rcu_head *head)
1982 {
1983         struct callchain_cpus_entries *entries;
1984         int cpu;
1985
1986         entries = container_of(head, struct callchain_cpus_entries, rcu_head);
1987
1988         for_each_possible_cpu(cpu)
1989                 kfree(entries->cpu_entries[cpu]);
1990
1991         kfree(entries);
1992 }
1993
1994 static void release_callchain_buffers(void)
1995 {
1996         struct callchain_cpus_entries *entries;
1997
1998         entries = callchain_cpus_entries;
1999         rcu_assign_pointer(callchain_cpus_entries, NULL);
2000         call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
2001 }
2002
2003 static int alloc_callchain_buffers(void)
2004 {
2005         int cpu;
2006         int size;
2007         struct callchain_cpus_entries *entries;
2008
2009         /*
2010          * We can't use the percpu allocation API for data that can be
2011          * accessed from NMI. Use a temporary manual per cpu allocation
2012          * until that gets sorted out.
2013          */
2014         size = offsetof(struct callchain_cpus_entries, cpu_entries[nr_cpu_ids]);
2015
2016         entries = kzalloc(size, GFP_KERNEL);
2017         if (!entries)
2018                 return -ENOMEM;
2019
2020         size = sizeof(struct perf_callchain_entry) * PERF_NR_CONTEXTS;
2021
2022         for_each_possible_cpu(cpu) {
2023                 entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
2024                                                          cpu_to_node(cpu));
2025                 if (!entries->cpu_entries[cpu])
2026                         goto fail;
2027         }
2028
2029         rcu_assign_pointer(callchain_cpus_entries, entries);
2030
2031         return 0;
2032
2033 fail:
2034         for_each_possible_cpu(cpu)
2035                 kfree(entries->cpu_entries[cpu]);
2036         kfree(entries);
2037
2038         return -ENOMEM;
2039 }
2040
2041 static int get_callchain_buffers(void)
2042 {
2043         int err = 0;
2044         int count;
2045
2046         mutex_lock(&callchain_mutex);
2047
2048         count = atomic_inc_return(&nr_callchain_events);
2049         if (WARN_ON_ONCE(count < 1)) {
2050                 err = -EINVAL;
2051                 goto exit;
2052         }
2053
2054         if (count > 1) {
2055                 /* If the allocation failed, give up */
2056                 if (!callchain_cpus_entries)
2057                         err = -ENOMEM;
2058                 goto exit;
2059         }
2060
2061         err = alloc_callchain_buffers();
2062         if (err)
2063                 release_callchain_buffers();
2064 exit:
2065         mutex_unlock(&callchain_mutex);
2066
2067         return err;
2068 }
2069
2070 static void put_callchain_buffers(void)
2071 {
2072         if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
2073                 release_callchain_buffers();
2074                 mutex_unlock(&callchain_mutex);
2075         }
2076 }
2077
2078 static int get_recursion_context(int *recursion)
2079 {
2080         int rctx;
2081
2082         if (in_nmi())
2083                 rctx = 3;
2084         else if (in_irq())
2085                 rctx = 2;
2086         else if (in_softirq())
2087                 rctx = 1;
2088         else
2089                 rctx = 0;
2090
2091         if (recursion[rctx])
2092                 return -1;
2093
2094         recursion[rctx]++;
2095         barrier();
2096
2097         return rctx;
2098 }
2099
2100 static inline void put_recursion_context(int *recursion, int rctx)
2101 {
2102         barrier();
2103         recursion[rctx]--;
2104 }
2105
2106 static struct perf_callchain_entry *get_callchain_entry(int *rctx)
2107 {
2108         int cpu;
2109         struct callchain_cpus_entries *entries;
2110
2111         *rctx = get_recursion_context(__get_cpu_var(callchain_recursion));
2112         if (*rctx == -1)
2113                 return NULL;
2114
2115         entries = rcu_dereference(callchain_cpus_entries);
2116         if (!entries)
2117                 return NULL;
2118
2119         cpu = smp_processor_id();
2120
2121         return &entries->cpu_entries[cpu][*rctx];
2122 }
2123
2124 static void
2125 put_callchain_entry(int rctx)
2126 {
2127         put_recursion_context(__get_cpu_var(callchain_recursion), rctx);
2128 }
2129
2130 static struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2131 {
2132         int rctx;
2133         struct perf_callchain_entry *entry;
2134
2135
2136         entry = get_callchain_entry(&rctx);
2137         if (rctx == -1)
2138                 return NULL;
2139
2140         if (!entry)
2141                 goto exit_put;
2142
2143         entry->nr = 0;
2144
2145         if (!user_mode(regs)) {
2146                 perf_callchain_store(entry, PERF_CONTEXT_KERNEL);
2147                 perf_callchain_kernel(entry, regs);
2148                 if (current->mm)
2149                         regs = task_pt_regs(current);
2150                 else
2151                         regs = NULL;
2152         }
2153
2154         if (regs) {
2155                 perf_callchain_store(entry, PERF_CONTEXT_USER);
2156                 perf_callchain_user(entry, regs);
2157         }
2158
2159 exit_put:
2160         put_callchain_entry(rctx);
2161
2162         return entry;
2163 }
2164
2165 /*
2166  * Initialize the perf_event context in a task_struct:
2167  */
2168 static void __perf_event_init_context(struct perf_event_context *ctx)
2169 {
2170         raw_spin_lock_init(&ctx->lock);
2171         mutex_init(&ctx->mutex);
2172         INIT_LIST_HEAD(&ctx->pinned_groups);
2173         INIT_LIST_HEAD(&ctx->flexible_groups);
2174         INIT_LIST_HEAD(&ctx->event_list);
2175         atomic_set(&ctx->refcount, 1);
2176 }
2177
2178 static struct perf_event_context *
2179 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2180 {
2181         struct perf_event_context *ctx;
2182
2183         ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2184         if (!ctx)
2185                 return NULL;
2186
2187         __perf_event_init_context(ctx);
2188         if (task) {
2189                 ctx->task = task;
2190                 get_task_struct(task);
2191         }
2192         ctx->pmu = pmu;
2193
2194         return ctx;
2195 }
2196
2197 static struct task_struct *
2198 find_lively_task_by_vpid(pid_t vpid)
2199 {
2200         struct task_struct *task;
2201         int err;
2202
2203         rcu_read_lock();
2204         if (!vpid)
2205                 task = current;
2206         else
2207                 task = find_task_by_vpid(vpid);
2208         if (task)
2209                 get_task_struct(task);
2210         rcu_read_unlock();
2211
2212         if (!task)
2213                 return ERR_PTR(-ESRCH);
2214
2215         /* Reuse ptrace permission checks for now. */
2216         err = -EACCES;
2217         if (!ptrace_may_access(task, PTRACE_MODE_READ))
2218                 goto errout;
2219
2220         return task;
2221 errout:
2222         put_task_struct(task);
2223         return ERR_PTR(err);
2224
2225 }
2226
2227 static struct perf_event_context *
2228 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
2229 {
2230         struct perf_event_context *ctx;
2231         struct perf_cpu_context *cpuctx;
2232         unsigned long flags;
2233         int ctxn, err;
2234
2235         if (!task) {
2236                 /* Must be root to operate on a CPU event: */
2237                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2238                         return ERR_PTR(-EACCES);
2239
2240                 /*
2241                  * We could be clever and allow to attach a event to an
2242                  * offline CPU and activate it when the CPU comes up, but
2243                  * that's for later.
2244                  */
2245                 if (!cpu_online(cpu))
2246                         return ERR_PTR(-ENODEV);
2247
2248                 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2249                 ctx = &cpuctx->ctx;
2250                 get_ctx(ctx);
2251
2252                 return ctx;
2253         }
2254
2255         err = -EINVAL;
2256         ctxn = pmu->task_ctx_nr;
2257         if (ctxn < 0)
2258                 goto errout;
2259
2260 retry:
2261         ctx = perf_lock_task_context(task, ctxn, &flags);
2262         if (ctx) {
2263                 unclone_ctx(ctx);
2264                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2265         }
2266
2267         if (!ctx) {
2268                 ctx = alloc_perf_context(pmu, task);
2269                 err = -ENOMEM;
2270                 if (!ctx)
2271                         goto errout;
2272
2273                 get_ctx(ctx);
2274
2275                 err = 0;
2276                 mutex_lock(&task->perf_event_mutex);
2277                 /*
2278                  * If it has already passed perf_event_exit_task().
2279                  * we must see PF_EXITING, it takes this mutex too.
2280                  */
2281                 if (task->flags & PF_EXITING)
2282                         err = -ESRCH;
2283                 else if (task->perf_event_ctxp[ctxn])
2284                         err = -EAGAIN;
2285                 else
2286                         rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
2287                 mutex_unlock(&task->perf_event_mutex);
2288
2289                 if (unlikely(err)) {
2290                         put_task_struct(task);
2291                         kfree(ctx);
2292
2293                         if (err == -EAGAIN)
2294                                 goto retry;
2295                         goto errout;
2296                 }
2297         }
2298
2299         return ctx;
2300
2301 errout:
2302         return ERR_PTR(err);
2303 }
2304
2305 static void perf_event_free_filter(struct perf_event *event);
2306
2307 static void free_event_rcu(struct rcu_head *head)
2308 {
2309         struct perf_event *event;
2310
2311         event = container_of(head, struct perf_event, rcu_head);
2312         if (event->ns)
2313                 put_pid_ns(event->ns);
2314         perf_event_free_filter(event);
2315         kfree(event);
2316 }
2317
2318 static void perf_buffer_put(struct perf_buffer *buffer);
2319
2320 static void free_event(struct perf_event *event)
2321 {
2322         irq_work_sync(&event->pending);
2323
2324         if (!event->parent) {
2325                 if (event->attach_state & PERF_ATTACH_TASK)
2326                         jump_label_dec(&perf_task_events);
2327                 if (event->attr.mmap || event->attr.mmap_data)
2328                         atomic_dec(&nr_mmap_events);
2329                 if (event->attr.comm)
2330                         atomic_dec(&nr_comm_events);
2331                 if (event->attr.task)
2332                         atomic_dec(&nr_task_events);
2333                 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2334                         put_callchain_buffers();
2335         }
2336
2337         if (event->buffer) {
2338                 perf_buffer_put(event->buffer);
2339                 event->buffer = NULL;
2340         }
2341
2342         if (event->destroy)
2343                 event->destroy(event);
2344
2345         if (event->ctx)
2346                 put_ctx(event->ctx);
2347
2348         call_rcu(&event->rcu_head, free_event_rcu);
2349 }
2350
2351 int perf_event_release_kernel(struct perf_event *event)
2352 {
2353         struct perf_event_context *ctx = event->ctx;
2354
2355         /*
2356          * Remove from the PMU, can't get re-enabled since we got
2357          * here because the last ref went.
2358          */
2359         perf_event_disable(event);
2360
2361         WARN_ON_ONCE(ctx->parent_ctx);
2362         /*
2363          * There are two ways this annotation is useful:
2364          *
2365          *  1) there is a lock recursion from perf_event_exit_task
2366          *     see the comment there.
2367          *
2368          *  2) there is a lock-inversion with mmap_sem through
2369          *     perf_event_read_group(), which takes faults while
2370          *     holding ctx->mutex, however this is called after
2371          *     the last filedesc died, so there is no possibility
2372          *     to trigger the AB-BA case.
2373          */
2374         mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2375         raw_spin_lock_irq(&ctx->lock);
2376         perf_group_detach(event);
2377         list_del_event(event, ctx);
2378         raw_spin_unlock_irq(&ctx->lock);
2379         mutex_unlock(&ctx->mutex);
2380
2381         free_event(event);
2382
2383         return 0;
2384 }
2385 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
2386
2387 /*
2388  * Called when the last reference to the file is gone.
2389  */
2390 static int perf_release(struct inode *inode, struct file *file)
2391 {
2392         struct perf_event *event = file->private_data;
2393         struct task_struct *owner;
2394
2395         file->private_data = NULL;
2396
2397         rcu_read_lock();
2398         owner = ACCESS_ONCE(event->owner);
2399         /*
2400          * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2401          * !owner it means the list deletion is complete and we can indeed
2402          * free this event, otherwise we need to serialize on
2403          * owner->perf_event_mutex.
2404          */
2405         smp_read_barrier_depends();
2406         if (owner) {
2407                 /*
2408                  * Since delayed_put_task_struct() also drops the last
2409                  * task reference we can safely take a new reference
2410                  * while holding the rcu_read_lock().
2411                  */
2412                 get_task_struct(owner);
2413         }
2414         rcu_read_unlock();
2415
2416         if (owner) {
2417                 mutex_lock(&owner->perf_event_mutex);
2418                 /*
2419                  * We have to re-check the event->owner field, if it is cleared
2420                  * we raced with perf_event_exit_task(), acquiring the mutex
2421                  * ensured they're done, and we can proceed with freeing the
2422                  * event.
2423                  */
2424                 if (event->owner)
2425                         list_del_init(&event->owner_entry);
2426                 mutex_unlock(&owner->perf_event_mutex);
2427                 put_task_struct(owner);
2428         }
2429
2430         return perf_event_release_kernel(event);
2431 }
2432
2433 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
2434 {
2435         struct perf_event *child;
2436         u64 total = 0;
2437
2438         *enabled = 0;
2439         *running = 0;
2440
2441         mutex_lock(&event->child_mutex);
2442         total += perf_event_read(event);
2443         *enabled += event->total_time_enabled +
2444                         atomic64_read(&event->child_total_time_enabled);
2445         *running += event->total_time_running +
2446                         atomic64_read(&event->child_total_time_running);
2447
2448         list_for_each_entry(child, &event->child_list, child_list) {
2449                 total += perf_event_read(child);
2450                 *enabled += child->total_time_enabled;
2451                 *running += child->total_time_running;
2452         }
2453         mutex_unlock(&event->child_mutex);
2454
2455         return total;
2456 }
2457 EXPORT_SYMBOL_GPL(perf_event_read_value);
2458
2459 static int perf_event_read_group(struct perf_event *event,
2460                                    u64 read_format, char __user *buf)
2461 {
2462         struct perf_event *leader = event->group_leader, *sub;
2463         int n = 0, size = 0, ret = -EFAULT;
2464         struct perf_event_context *ctx = leader->ctx;
2465         u64 values[5];
2466         u64 count, enabled, running;
2467
2468         mutex_lock(&ctx->mutex);
2469         count = perf_event_read_value(leader, &enabled, &running);
2470
2471         values[n++] = 1 + leader->nr_siblings;
2472         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2473                 values[n++] = enabled;
2474         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2475                 values[n++] = running;
2476         values[n++] = count;
2477         if (read_format & PERF_FORMAT_ID)
2478                 values[n++] = primary_event_id(leader);
2479
2480         size = n * sizeof(u64);
2481
2482         if (copy_to_user(buf, values, size))
2483                 goto unlock;
2484
2485         ret = size;
2486
2487         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2488                 n = 0;
2489
2490                 values[n++] = perf_event_read_value(sub, &enabled, &running);
2491                 if (read_format & PERF_FORMAT_ID)
2492                         values[n++] = primary_event_id(sub);
2493
2494                 size = n * sizeof(u64);
2495
2496                 if (copy_to_user(buf + ret, values, size)) {
2497                         ret = -EFAULT;
2498                         goto unlock;
2499                 }
2500
2501                 ret += size;
2502         }
2503 unlock:
2504         mutex_unlock(&ctx->mutex);
2505
2506         return ret;
2507 }
2508
2509 static int perf_event_read_one(struct perf_event *event,
2510                                  u64 read_format, char __user *buf)
2511 {
2512         u64 enabled, running;
2513         u64 values[4];
2514         int n = 0;
2515
2516         values[n++] = perf_event_read_value(event, &enabled, &running);
2517         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2518                 values[n++] = enabled;
2519         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2520                 values[n++] = running;
2521         if (read_format & PERF_FORMAT_ID)
2522                 values[n++] = primary_event_id(event);
2523
2524         if (copy_to_user(buf, values, n * sizeof(u64)))
2525                 return -EFAULT;
2526
2527         return n * sizeof(u64);
2528 }
2529
2530 /*
2531  * Read the performance event - simple non blocking version for now
2532  */
2533 static ssize_t
2534 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2535 {
2536         u64 read_format = event->attr.read_format;
2537         int ret;
2538
2539         /*
2540          * Return end-of-file for a read on a event that is in
2541          * error state (i.e. because it was pinned but it couldn't be
2542          * scheduled on to the CPU at some point).
2543          */
2544         if (event->state == PERF_EVENT_STATE_ERROR)
2545                 return 0;
2546
2547         if (count < event->read_size)
2548                 return -ENOSPC;
2549
2550         WARN_ON_ONCE(event->ctx->parent_ctx);
2551         if (read_format & PERF_FORMAT_GROUP)
2552                 ret = perf_event_read_group(event, read_format, buf);
2553         else
2554                 ret = perf_event_read_one(event, read_format, buf);
2555
2556         return ret;
2557 }
2558
2559 static ssize_t
2560 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2561 {
2562         struct perf_event *event = file->private_data;
2563
2564         return perf_read_hw(event, buf, count);
2565 }
2566
2567 static unsigned int perf_poll(struct file *file, poll_table *wait)
2568 {
2569         struct perf_event *event = file->private_data;
2570         struct perf_buffer *buffer;
2571         unsigned int events = POLL_HUP;
2572
2573         rcu_read_lock();
2574         buffer = rcu_dereference(event->buffer);
2575         if (buffer)
2576                 events = atomic_xchg(&buffer->poll, 0);
2577         rcu_read_unlock();
2578
2579         poll_wait(file, &event->waitq, wait);
2580
2581         return events;
2582 }
2583
2584 static void perf_event_reset(struct perf_event *event)
2585 {
2586         (void)perf_event_read(event);
2587         local64_set(&event->count, 0);
2588         perf_event_update_userpage(event);
2589 }
2590
2591 /*
2592  * Holding the top-level event's child_mutex means that any
2593  * descendant process that has inherited this event will block
2594  * in sync_child_event if it goes to exit, thus satisfying the
2595  * task existence requirements of perf_event_enable/disable.
2596  */
2597 static void perf_event_for_each_child(struct perf_event *event,
2598                                         void (*func)(struct perf_event *))
2599 {
2600         struct perf_event *child;
2601
2602         WARN_ON_ONCE(event->ctx->parent_ctx);
2603         mutex_lock(&event->child_mutex);
2604         func(event);
2605         list_for_each_entry(child, &event->child_list, child_list)
2606                 func(child);
2607         mutex_unlock(&event->child_mutex);
2608 }
2609
2610 static void perf_event_for_each(struct perf_event *event,
2611                                   void (*func)(struct perf_event *))
2612 {
2613         struct perf_event_context *ctx = event->ctx;
2614         struct perf_event *sibling;
2615
2616         WARN_ON_ONCE(ctx->parent_ctx);
2617         mutex_lock(&ctx->mutex);
2618         event = event->group_leader;
2619
2620         perf_event_for_each_child(event, func);
2621         func(event);
2622         list_for_each_entry(sibling, &event->sibling_list, group_entry)
2623                 perf_event_for_each_child(event, func);
2624         mutex_unlock(&ctx->mutex);
2625 }
2626
2627 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2628 {
2629         struct perf_event_context *ctx = event->ctx;
2630         int ret = 0;
2631         u64 value;
2632
2633         if (!is_sampling_event(event))
2634                 return -EINVAL;
2635
2636         if (copy_from_user(&value, arg, sizeof(value)))
2637                 return -EFAULT;
2638
2639         if (!value)
2640                 return -EINVAL;
2641
2642         raw_spin_lock_irq(&ctx->lock);
2643         if (event->attr.freq) {
2644                 if (value > sysctl_perf_event_sample_rate) {
2645                         ret = -EINVAL;
2646                         goto unlock;
2647                 }
2648
2649                 event->attr.sample_freq = value;
2650         } else {
2651                 event->attr.sample_period = value;
2652                 event->hw.sample_period = value;
2653         }
2654 unlock:
2655         raw_spin_unlock_irq(&ctx->lock);
2656
2657         return ret;
2658 }
2659
2660 static const struct file_operations perf_fops;
2661
2662 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
2663 {
2664         struct file *file;
2665
2666         file = fget_light(fd, fput_needed);
2667         if (!file)
2668                 return ERR_PTR(-EBADF);
2669
2670         if (file->f_op != &perf_fops) {
2671                 fput_light(file, *fput_needed);
2672                 *fput_needed = 0;
2673                 return ERR_PTR(-EBADF);
2674         }
2675
2676         return file->private_data;
2677 }
2678
2679 static int perf_event_set_output(struct perf_event *event,
2680                                  struct perf_event *output_event);
2681 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2682
2683 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2684 {
2685         struct perf_event *event = file->private_data;
2686         void (*func)(struct perf_event *);
2687         u32 flags = arg;
2688
2689         switch (cmd) {
2690         case PERF_EVENT_IOC_ENABLE:
2691                 func = perf_event_enable;
2692                 break;
2693         case PERF_EVENT_IOC_DISABLE:
2694                 func = perf_event_disable;
2695                 break;
2696         case PERF_EVENT_IOC_RESET:
2697                 func = perf_event_reset;
2698                 break;
2699
2700         case PERF_EVENT_IOC_REFRESH:
2701                 return perf_event_refresh(event, arg);
2702
2703         case PERF_EVENT_IOC_PERIOD:
2704                 return perf_event_period(event, (u64 __user *)arg);
2705
2706         case PERF_EVENT_IOC_SET_OUTPUT:
2707         {
2708                 struct perf_event *output_event = NULL;
2709                 int fput_needed = 0;
2710                 int ret;
2711
2712                 if (arg != -1) {
2713                         output_event = perf_fget_light(arg, &fput_needed);
2714                         if (IS_ERR(output_event))
2715                                 return PTR_ERR(output_event);
2716                 }
2717
2718                 ret = perf_event_set_output(event, output_event);
2719                 if (output_event)
2720                         fput_light(output_event->filp, fput_needed);
2721
2722                 return ret;
2723         }
2724
2725         case PERF_EVENT_IOC_SET_FILTER:
2726                 return perf_event_set_filter(event, (void __user *)arg);
2727
2728         default:
2729                 return -ENOTTY;
2730         }
2731
2732         if (flags & PERF_IOC_FLAG_GROUP)
2733                 perf_event_for_each(event, func);
2734         else
2735                 perf_event_for_each_child(event, func);
2736
2737         return 0;
2738 }
2739
2740 int perf_event_task_enable(void)
2741 {
2742         struct perf_event *event;
2743
2744         mutex_lock(&current->perf_event_mutex);
2745         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2746                 perf_event_for_each_child(event, perf_event_enable);
2747         mutex_unlock(&current->perf_event_mutex);
2748
2749         return 0;
2750 }
2751
2752 int perf_event_task_disable(void)
2753 {
2754         struct perf_event *event;
2755
2756         mutex_lock(&current->perf_event_mutex);
2757         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2758                 perf_event_for_each_child(event, perf_event_disable);
2759         mutex_unlock(&current->perf_event_mutex);
2760
2761         return 0;
2762 }
2763
2764 #ifndef PERF_EVENT_INDEX_OFFSET
2765 # define PERF_EVENT_INDEX_OFFSET 0
2766 #endif
2767
2768 static int perf_event_index(struct perf_event *event)
2769 {
2770         if (event->hw.state & PERF_HES_STOPPED)
2771                 return 0;
2772
2773         if (event->state != PERF_EVENT_STATE_ACTIVE)
2774                 return 0;
2775
2776         return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2777 }
2778
2779 /*
2780  * Callers need to ensure there can be no nesting of this function, otherwise
2781  * the seqlock logic goes bad. We can not serialize this because the arch
2782  * code calls this from NMI context.
2783  */
2784 void perf_event_update_userpage(struct perf_event *event)
2785 {
2786         struct perf_event_mmap_page *userpg;
2787         struct perf_buffer *buffer;
2788
2789         rcu_read_lock();
2790         buffer = rcu_dereference(event->buffer);
2791         if (!buffer)
2792                 goto unlock;
2793
2794         userpg = buffer->user_page;
2795
2796         /*
2797          * Disable preemption so as to not let the corresponding user-space
2798          * spin too long if we get preempted.
2799          */
2800         preempt_disable();
2801         ++userpg->lock;
2802         barrier();
2803         userpg->index = perf_event_index(event);
2804         userpg->offset = perf_event_count(event);
2805         if (event->state == PERF_EVENT_STATE_ACTIVE)
2806                 userpg->offset -= local64_read(&event->hw.prev_count);
2807
2808         userpg->time_enabled = event->total_time_enabled +
2809                         atomic64_read(&event->child_total_time_enabled);
2810
2811         userpg->time_running = event->total_time_running +
2812                         atomic64_read(&event->child_total_time_running);
2813
2814         barrier();
2815         ++userpg->lock;
2816         preempt_enable();
2817 unlock:
2818         rcu_read_unlock();
2819 }
2820
2821 static unsigned long perf_data_size(struct perf_buffer *buffer);
2822
2823 static void
2824 perf_buffer_init(struct perf_buffer *buffer, long watermark, int flags)
2825 {
2826         long max_size = perf_data_size(buffer);
2827
2828         if (watermark)
2829                 buffer->watermark = min(max_size, watermark);
2830
2831         if (!buffer->watermark)
2832                 buffer->watermark = max_size / 2;
2833
2834         if (flags & PERF_BUFFER_WRITABLE)
2835                 buffer->writable = 1;
2836
2837         atomic_set(&buffer->refcount, 1);
2838 }
2839
2840 #ifndef CONFIG_PERF_USE_VMALLOC
2841
2842 /*
2843  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2844  */
2845
2846 static struct page *
2847 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2848 {
2849         if (pgoff > buffer->nr_pages)
2850                 return NULL;
2851
2852         if (pgoff == 0)
2853                 return virt_to_page(buffer->user_page);
2854
2855         return virt_to_page(buffer->data_pages[pgoff - 1]);
2856 }
2857
2858 static void *perf_mmap_alloc_page(int cpu)
2859 {
2860         struct page *page;
2861         int node;
2862
2863         node = (cpu == -1) ? cpu : cpu_to_node(cpu);
2864         page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
2865         if (!page)
2866                 return NULL;
2867
2868         return page_address(page);
2869 }
2870
2871 static struct perf_buffer *
2872 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2873 {
2874         struct perf_buffer *buffer;
2875         unsigned long size;
2876         int i;
2877
2878         size = sizeof(struct perf_buffer);
2879         size += nr_pages * sizeof(void *);
2880
2881         buffer = kzalloc(size, GFP_KERNEL);
2882         if (!buffer)
2883                 goto fail;
2884
2885         buffer->user_page = perf_mmap_alloc_page(cpu);
2886         if (!buffer->user_page)
2887                 goto fail_user_page;
2888
2889         for (i = 0; i < nr_pages; i++) {
2890                 buffer->data_pages[i] = perf_mmap_alloc_page(cpu);
2891                 if (!buffer->data_pages[i])
2892                         goto fail_data_pages;
2893         }
2894
2895         buffer->nr_pages = nr_pages;
2896
2897         perf_buffer_init(buffer, watermark, flags);
2898
2899         return buffer;
2900
2901 fail_data_pages:
2902         for (i--; i >= 0; i--)
2903                 free_page((unsigned long)buffer->data_pages[i]);
2904
2905         free_page((unsigned long)buffer->user_page);
2906
2907 fail_user_page:
2908         kfree(buffer);
2909
2910 fail:
2911         return NULL;
2912 }
2913
2914 static void perf_mmap_free_page(unsigned long addr)
2915 {
2916         struct page *page = virt_to_page((void *)addr);
2917
2918         page->mapping = NULL;
2919         __free_page(page);
2920 }
2921
2922 static void perf_buffer_free(struct perf_buffer *buffer)
2923 {
2924         int i;
2925
2926         perf_mmap_free_page((unsigned long)buffer->user_page);
2927         for (i = 0; i < buffer->nr_pages; i++)
2928                 perf_mmap_free_page((unsigned long)buffer->data_pages[i]);
2929         kfree(buffer);
2930 }
2931
2932 static inline int page_order(struct perf_buffer *buffer)
2933 {
2934         return 0;
2935 }
2936
2937 #else
2938
2939 /*
2940  * Back perf_mmap() with vmalloc memory.
2941  *
2942  * Required for architectures that have d-cache aliasing issues.
2943  */
2944
2945 static inline int page_order(struct perf_buffer *buffer)
2946 {
2947         return buffer->page_order;
2948 }
2949
2950 static struct page *
2951 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2952 {
2953         if (pgoff > (1UL << page_order(buffer)))
2954                 return NULL;
2955
2956         return vmalloc_to_page((void *)buffer->user_page + pgoff * PAGE_SIZE);
2957 }
2958
2959 static void perf_mmap_unmark_page(void *addr)
2960 {
2961         struct page *page = vmalloc_to_page(addr);
2962
2963         page->mapping = NULL;
2964 }
2965
2966 static void perf_buffer_free_work(struct work_struct *work)
2967 {
2968         struct perf_buffer *buffer;
2969         void *base;
2970         int i, nr;
2971
2972         buffer = container_of(work, struct perf_buffer, work);
2973         nr = 1 << page_order(buffer);
2974
2975         base = buffer->user_page;
2976         for (i = 0; i < nr + 1; i++)
2977                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2978
2979         vfree(base);
2980         kfree(buffer);
2981 }
2982
2983 static void perf_buffer_free(struct perf_buffer *buffer)
2984 {
2985         schedule_work(&buffer->work);
2986 }
2987
2988 static struct perf_buffer *
2989 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2990 {
2991         struct perf_buffer *buffer;
2992         unsigned long size;
2993         void *all_buf;
2994
2995         size = sizeof(struct perf_buffer);
2996         size += sizeof(void *);
2997
2998         buffer = kzalloc(size, GFP_KERNEL);
2999         if (!buffer)
3000                 goto fail;
3001
3002         INIT_WORK(&buffer->work, perf_buffer_free_work);
3003
3004         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
3005         if (!all_buf)
3006                 goto fail_all_buf;
3007
3008         buffer->user_page = all_buf;
3009         buffer->data_pages[0] = all_buf + PAGE_SIZE;
3010         buffer->page_order = ilog2(nr_pages);
3011         buffer->nr_pages = 1;
3012
3013         perf_buffer_init(buffer, watermark, flags);
3014
3015         return buffer;
3016
3017 fail_all_buf:
3018         kfree(buffer);
3019
3020 fail:
3021         return NULL;
3022 }
3023
3024 #endif
3025
3026 static unsigned long perf_data_size(struct perf_buffer *buffer)
3027 {
3028         return buffer->nr_pages << (PAGE_SHIFT + page_order(buffer));
3029 }
3030
3031 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3032 {
3033         struct perf_event *event = vma->vm_file->private_data;
3034         struct perf_buffer *buffer;
3035         int ret = VM_FAULT_SIGBUS;
3036
3037         if (vmf->flags & FAULT_FLAG_MKWRITE) {
3038                 if (vmf->pgoff == 0)
3039                         ret = 0;
3040                 return ret;
3041         }
3042
3043         rcu_read_lock();
3044         buffer = rcu_dereference(event->buffer);
3045         if (!buffer)
3046                 goto unlock;
3047
3048         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3049                 goto unlock;
3050
3051         vmf->page = perf_mmap_to_page(buffer, vmf->pgoff);
3052         if (!vmf->page)
3053                 goto unlock;
3054
3055         get_page(vmf->page);
3056         vmf->page->mapping = vma->vm_file->f_mapping;
3057         vmf->page->index   = vmf->pgoff;
3058
3059         ret = 0;
3060 unlock:
3061         rcu_read_unlock();
3062
3063         return ret;
3064 }
3065
3066 static void perf_buffer_free_rcu(struct rcu_head *rcu_head)
3067 {
3068         struct perf_buffer *buffer;
3069
3070         buffer = container_of(rcu_head, struct perf_buffer, rcu_head);
3071         perf_buffer_free(buffer);
3072 }
3073
3074 static struct perf_buffer *perf_buffer_get(struct perf_event *event)
3075 {
3076         struct perf_buffer *buffer;
3077
3078         rcu_read_lock();
3079         buffer = rcu_dereference(event->buffer);
3080         if (buffer) {
3081                 if (!atomic_inc_not_zero(&buffer->refcount))
3082                         buffer = NULL;
3083         }
3084         rcu_read_unlock();
3085
3086         return buffer;
3087 }
3088
3089 static void perf_buffer_put(struct perf_buffer *buffer)
3090 {
3091         if (!atomic_dec_and_test(&buffer->refcount))
3092                 return;
3093
3094         call_rcu(&buffer->rcu_head, perf_buffer_free_rcu);
3095 }
3096
3097 static void perf_mmap_open(struct vm_area_struct *vma)
3098 {
3099         struct perf_event *event = vma->vm_file->private_data;
3100
3101         atomic_inc(&event->mmap_count);
3102 }
3103
3104 static void perf_mmap_close(struct vm_area_struct *vma)
3105 {
3106         struct perf_event *event = vma->vm_file->private_data;
3107
3108         if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
3109                 unsigned long size = perf_data_size(event->buffer);
3110                 struct user_struct *user = event->mmap_user;
3111                 struct perf_buffer *buffer = event->buffer;
3112
3113                 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
3114                 vma->vm_mm->locked_vm -= event->mmap_locked;
3115                 rcu_assign_pointer(event->buffer, NULL);
3116                 mutex_unlock(&event->mmap_mutex);
3117
3118                 perf_buffer_put(buffer);
3119                 free_uid(user);
3120         }
3121 }
3122
3123 static const struct vm_operations_struct perf_mmap_vmops = {
3124         .open           = perf_mmap_open,
3125         .close          = perf_mmap_close,
3126         .fault          = perf_mmap_fault,
3127         .page_mkwrite   = perf_mmap_fault,
3128 };
3129
3130 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3131 {
3132         struct perf_event *event = file->private_data;
3133         unsigned long user_locked, user_lock_limit;
3134         struct user_struct *user = current_user();
3135         unsigned long locked, lock_limit;
3136         struct perf_buffer *buffer;
3137         unsigned long vma_size;
3138         unsigned long nr_pages;
3139         long user_extra, extra;
3140         int ret = 0, flags = 0;
3141
3142         /*
3143          * Don't allow mmap() of inherited per-task counters. This would
3144          * create a performance issue due to all children writing to the
3145          * same buffer.
3146          */
3147         if (event->cpu == -1 && event->attr.inherit)
3148                 return -EINVAL;
3149
3150         if (!(vma->vm_flags & VM_SHARED))
3151                 return -EINVAL;
3152
3153         vma_size = vma->vm_end - vma->vm_start;
3154         nr_pages = (vma_size / PAGE_SIZE) - 1;
3155
3156         /*
3157          * If we have buffer pages ensure they're a power-of-two number, so we
3158          * can do bitmasks instead of modulo.
3159          */
3160         if (nr_pages != 0 && !is_power_of_2(nr_pages))
3161                 return -EINVAL;
3162
3163         if (vma_size != PAGE_SIZE * (1 + nr_pages))
3164                 return -EINVAL;
3165
3166         if (vma->vm_pgoff != 0)
3167                 return -EINVAL;
3168
3169         WARN_ON_ONCE(event->ctx->parent_ctx);
3170         mutex_lock(&event->mmap_mutex);
3171         if (event->buffer) {
3172                 if (event->buffer->nr_pages == nr_pages)
3173                         atomic_inc(&event->buffer->refcount);
3174                 else
3175                         ret = -EINVAL;
3176                 goto unlock;
3177         }
3178
3179         user_extra = nr_pages + 1;
3180         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3181
3182         /*
3183          * Increase the limit linearly with more CPUs:
3184          */
3185         user_lock_limit *= num_online_cpus();
3186
3187         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3188
3189         extra = 0;
3190         if (user_locked > user_lock_limit)
3191                 extra = user_locked - user_lock_limit;
3192
3193         lock_limit = rlimit(RLIMIT_MEMLOCK);
3194         lock_limit >>= PAGE_SHIFT;
3195         locked = vma->vm_mm->locked_vm + extra;
3196
3197         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3198                 !capable(CAP_IPC_LOCK)) {
3199                 ret = -EPERM;
3200                 goto unlock;
3201         }
3202
3203         WARN_ON(event->buffer);
3204
3205         if (vma->vm_flags & VM_WRITE)
3206                 flags |= PERF_BUFFER_WRITABLE;
3207
3208         buffer = perf_buffer_alloc(nr_pages, event->attr.wakeup_watermark,
3209                                    event->cpu, flags);
3210         if (!buffer) {
3211                 ret = -ENOMEM;
3212                 goto unlock;
3213         }
3214         rcu_assign_pointer(event->buffer, buffer);
3215
3216         atomic_long_add(user_extra, &user->locked_vm);
3217         event->mmap_locked = extra;
3218         event->mmap_user = get_current_user();
3219         vma->vm_mm->locked_vm += event->mmap_locked;
3220
3221 unlock:
3222         if (!ret)
3223                 atomic_inc(&event->mmap_count);
3224         mutex_unlock(&event->mmap_mutex);
3225
3226         vma->vm_flags |= VM_RESERVED;
3227         vma->vm_ops = &perf_mmap_vmops;
3228
3229         return ret;
3230 }
3231
3232 static int perf_fasync(int fd, struct file *filp, int on)
3233 {
3234         struct inode *inode = filp->f_path.dentry->d_inode;
3235         struct perf_event *event = filp->private_data;
3236         int retval;
3237
3238         mutex_lock(&inode->i_mutex);
3239         retval = fasync_helper(fd, filp, on, &event->fasync);
3240         mutex_unlock(&inode->i_mutex);
3241
3242         if (retval < 0)
3243                 return retval;
3244
3245         return 0;
3246 }
3247
3248 static const struct file_operations perf_fops = {
3249         .llseek                 = no_llseek,
3250         .release                = perf_release,
3251         .read                   = perf_read,
3252         .poll                   = perf_poll,
3253         .unlocked_ioctl         = perf_ioctl,
3254         .compat_ioctl           = perf_ioctl,
3255         .mmap                   = perf_mmap,
3256         .fasync                 = perf_fasync,
3257 };
3258
3259 /*
3260  * Perf event wakeup
3261  *
3262  * If there's data, ensure we set the poll() state and publish everything
3263  * to user-space before waking everybody up.
3264  */
3265
3266 void perf_event_wakeup(struct perf_event *event)
3267 {
3268         wake_up_all(&event->waitq);
3269
3270         if (event->pending_kill) {
3271                 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3272                 event->pending_kill = 0;
3273         }
3274 }
3275
3276 static void perf_pending_event(struct irq_work *entry)
3277 {
3278         struct perf_event *event = container_of(entry,
3279                         struct perf_event, pending);
3280
3281         if (event->pending_disable) {
3282                 event->pending_disable = 0;
3283                 __perf_event_disable(event);
3284         }
3285
3286         if (event->pending_wakeup) {
3287                 event->pending_wakeup = 0;
3288                 perf_event_wakeup(event);
3289         }
3290 }
3291
3292 /*
3293  * We assume there is only KVM supporting the callbacks.
3294  * Later on, we might change it to a list if there is
3295  * another virtualization implementation supporting the callbacks.
3296  */
3297 struct perf_guest_info_callbacks *perf_guest_cbs;
3298
3299 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3300 {
3301         perf_guest_cbs = cbs;
3302         return 0;
3303 }
3304 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3305
3306 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3307 {
3308         perf_guest_cbs = NULL;
3309         return 0;
3310 }
3311 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3312
3313 /*
3314  * Output
3315  */
3316 static bool perf_output_space(struct perf_buffer *buffer, unsigned long tail,
3317                               unsigned long offset, unsigned long head)
3318 {
3319         unsigned long mask;
3320
3321         if (!buffer->writable)
3322                 return true;
3323
3324         mask = perf_data_size(buffer) - 1;
3325
3326         offset = (offset - tail) & mask;
3327         head   = (head   - tail) & mask;
3328
3329         if ((int)(head - offset) < 0)
3330                 return false;
3331
3332         return true;
3333 }
3334
3335 static void perf_output_wakeup(struct perf_output_handle *handle)
3336 {
3337         atomic_set(&handle->buffer->poll, POLL_IN);
3338
3339         if (handle->nmi) {
3340                 handle->event->pending_wakeup = 1;
3341                 irq_work_queue(&handle->event->pending);
3342         } else
3343                 perf_event_wakeup(handle->event);
3344 }
3345
3346 /*
3347  * We need to ensure a later event_id doesn't publish a head when a former
3348  * event isn't done writing. However since we need to deal with NMIs we
3349  * cannot fully serialize things.
3350  *
3351  * We only publish the head (and generate a wakeup) when the outer-most
3352  * event completes.
3353  */
3354 static void perf_output_get_handle(struct perf_output_handle *handle)
3355 {
3356         struct perf_buffer *buffer = handle->buffer;
3357
3358         preempt_disable();
3359         local_inc(&buffer->nest);
3360         handle->wakeup = local_read(&buffer->wakeup);
3361 }
3362
3363 static void perf_output_put_handle(struct perf_output_handle *handle)
3364 {
3365         struct perf_buffer *buffer = handle->buffer;
3366         unsigned long head;
3367
3368 again:
3369         head = local_read(&buffer->head);
3370
3371         /*
3372          * IRQ/NMI can happen here, which means we can miss a head update.
3373          */
3374
3375         if (!local_dec_and_test(&buffer->nest))
3376                 goto out;
3377
3378         /*
3379          * Publish the known good head. Rely on the full barrier implied
3380          * by atomic_dec_and_test() order the buffer->head read and this
3381          * write.
3382          */
3383         buffer->user_page->data_head = head;
3384
3385         /*
3386          * Now check if we missed an update, rely on the (compiler)
3387          * barrier in atomic_dec_and_test() to re-read buffer->head.
3388          */
3389         if (unlikely(head != local_read(&buffer->head))) {
3390                 local_inc(&buffer->nest);
3391                 goto again;
3392         }
3393
3394         if (handle->wakeup != local_read(&buffer->wakeup))
3395                 perf_output_wakeup(handle);
3396
3397 out:
3398         preempt_enable();
3399 }
3400
3401 __always_inline void perf_output_copy(struct perf_output_handle *handle,
3402                       const void *buf, unsigned int len)
3403 {
3404         do {
3405                 unsigned long size = min_t(unsigned long, handle->size, len);
3406
3407                 memcpy(handle->addr, buf, size);
3408
3409                 len -= size;
3410                 handle->addr += size;
3411                 buf += size;
3412                 handle->size -= size;
3413                 if (!handle->size) {
3414                         struct perf_buffer *buffer = handle->buffer;
3415
3416                         handle->page++;
3417                         handle->page &= buffer->nr_pages - 1;
3418                         handle->addr = buffer->data_pages[handle->page];
3419                         handle->size = PAGE_SIZE << page_order(buffer);
3420                 }
3421         } while (len);
3422 }
3423
3424 static void __perf_event_header__init_id(struct perf_event_header *header,
3425                                          struct perf_sample_data *data,
3426                                          struct perf_event *event)
3427 {
3428         u64 sample_type = event->attr.sample_type;
3429
3430         data->type = sample_type;
3431         header->size += event->id_header_size;
3432
3433         if (sample_type & PERF_SAMPLE_TID) {
3434                 /* namespace issues */
3435                 data->tid_entry.pid = perf_event_pid(event, current);
3436                 data->tid_entry.tid = perf_event_tid(event, current);
3437         }
3438
3439         if (sample_type & PERF_SAMPLE_TIME)
3440                 data->time = perf_clock();
3441
3442         if (sample_type & PERF_SAMPLE_ID)
3443                 data->id = primary_event_id(event);
3444
3445         if (sample_type & PERF_SAMPLE_STREAM_ID)
3446                 data->stream_id = event->id;
3447
3448         if (sample_type & PERF_SAMPLE_CPU) {
3449                 data->cpu_entry.cpu      = raw_smp_processor_id();
3450                 data->cpu_entry.reserved = 0;
3451         }
3452 }
3453
3454 static void perf_event_header__init_id(struct perf_event_header *header,
3455                                        struct perf_sample_data *data,
3456                                        struct perf_event *event)
3457 {
3458         if (event->attr.sample_id_all)
3459                 __perf_event_header__init_id(header, data, event);
3460 }
3461
3462 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
3463                                            struct perf_sample_data *data)
3464 {
3465         u64 sample_type = data->type;
3466
3467         if (sample_type & PERF_SAMPLE_TID)
3468                 perf_output_put(handle, data->tid_entry);
3469
3470         if (sample_type & PERF_SAMPLE_TIME)
3471                 perf_output_put(handle, data->time);
3472
3473         if (sample_type & PERF_SAMPLE_ID)
3474                 perf_output_put(handle, data->id);
3475
3476         if (sample_type & PERF_SAMPLE_STREAM_ID)
3477                 perf_output_put(handle, data->stream_id);
3478
3479         if (sample_type & PERF_SAMPLE_CPU)
3480                 perf_output_put(handle, data->cpu_entry);
3481 }
3482
3483 static void perf_event__output_id_sample(struct perf_event *event,
3484                                          struct perf_output_handle *handle,
3485                                          struct perf_sample_data *sample)
3486 {
3487         if (event->attr.sample_id_all)
3488                 __perf_event__output_id_sample(handle, sample);
3489 }
3490
3491 int perf_output_begin(struct perf_output_handle *handle,
3492                       struct perf_event *event, unsigned int size,
3493                       int nmi, int sample)
3494 {
3495         struct perf_buffer *buffer;
3496         unsigned long tail, offset, head;
3497         int have_lost;
3498         struct perf_sample_data sample_data;
3499         struct {
3500                 struct perf_event_header header;
3501                 u64                      id;
3502                 u64                      lost;
3503         } lost_event;
3504
3505         rcu_read_lock();
3506         /*
3507          * For inherited events we send all the output towards the parent.
3508          */
3509         if (event->parent)
3510                 event = event->parent;
3511
3512         buffer = rcu_dereference(event->buffer);
3513         if (!buffer)
3514                 goto out;
3515
3516         handle->buffer  = buffer;
3517         handle->event   = event;
3518         handle->nmi     = nmi;
3519         handle->sample  = sample;
3520
3521         if (!buffer->nr_pages)
3522                 goto out;
3523
3524         have_lost = local_read(&buffer->lost);
3525         if (have_lost) {
3526                 lost_event.header.size = sizeof(lost_event);
3527                 perf_event_header__init_id(&lost_event.header, &sample_data,
3528                                            event);
3529                 size += lost_event.header.size;
3530         }
3531
3532         perf_output_get_handle(handle);
3533
3534         do {
3535                 /*
3536                  * Userspace could choose to issue a mb() before updating the
3537                  * tail pointer. So that all reads will be completed before the
3538                  * write is issued.
3539                  */
3540                 tail = ACCESS_ONCE(buffer->user_page->data_tail);
3541                 smp_rmb();
3542                 offset = head = local_read(&buffer->head);
3543                 head += size;
3544                 if (unlikely(!perf_output_space(buffer, tail, offset, head)))
3545                         goto fail;
3546         } while (local_cmpxchg(&buffer->head, offset, head) != offset);
3547
3548         if (head - local_read(&buffer->wakeup) > buffer->watermark)
3549                 local_add(buffer->watermark, &buffer->wakeup);
3550
3551         handle->page = offset >> (PAGE_SHIFT + page_order(buffer));
3552         handle->page &= buffer->nr_pages - 1;
3553         handle->size = offset & ((PAGE_SIZE << page_order(buffer)) - 1);
3554         handle->addr = buffer->data_pages[handle->page];
3555         handle->addr += handle->size;
3556         handle->size = (PAGE_SIZE << page_order(buffer)) - handle->size;
3557
3558         if (have_lost) {
3559                 lost_event.header.type = PERF_RECORD_LOST;
3560                 lost_event.header.misc = 0;
3561                 lost_event.id          = event->id;
3562                 lost_event.lost        = local_xchg(&buffer->lost, 0);
3563
3564                 perf_output_put(handle, lost_event);
3565                 perf_event__output_id_sample(event, handle, &sample_data);
3566         }
3567
3568         return 0;
3569
3570 fail:
3571         local_inc(&buffer->lost);
3572         perf_output_put_handle(handle);
3573 out:
3574         rcu_read_unlock();
3575
3576         return -ENOSPC;
3577 }
3578
3579 void perf_output_end(struct perf_output_handle *handle)
3580 {
3581         struct perf_event *event = handle->event;
3582         struct perf_buffer *buffer = handle->buffer;
3583
3584         int wakeup_events = event->attr.wakeup_events;
3585
3586         if (handle->sample && wakeup_events) {
3587                 int events = local_inc_return(&buffer->events);
3588                 if (events >= wakeup_events) {
3589                         local_sub(wakeup_events, &buffer->events);
3590                         local_inc(&buffer->wakeup);
3591                 }
3592         }
3593
3594         perf_output_put_handle(handle);
3595         rcu_read_unlock();
3596 }
3597
3598 static void perf_output_read_one(struct perf_output_handle *handle,
3599                                  struct perf_event *event,
3600                                  u64 enabled, u64 running)
3601 {
3602         u64 read_format = event->attr.read_format;
3603         u64 values[4];
3604         int n = 0;
3605
3606         values[n++] = perf_event_count(event);
3607         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3608                 values[n++] = enabled +
3609                         atomic64_read(&event->child_total_time_enabled);
3610         }
3611         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3612                 values[n++] = running +
3613                         atomic64_read(&event->child_total_time_running);
3614         }
3615         if (read_format & PERF_FORMAT_ID)
3616                 values[n++] = primary_event_id(event);
3617
3618         perf_output_copy(handle, values, n * sizeof(u64));
3619 }
3620
3621 /*
3622  * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3623  */
3624 static void perf_output_read_group(struct perf_output_handle *handle,
3625                             struct perf_event *event,
3626                             u64 enabled, u64 running)
3627 {
3628         struct perf_event *leader = event->group_leader, *sub;
3629         u64 read_format = event->attr.read_format;
3630         u64 values[5];
3631         int n = 0;
3632
3633         values[n++] = 1 + leader->nr_siblings;
3634
3635         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3636                 values[n++] = enabled;
3637
3638         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3639                 values[n++] = running;
3640
3641         if (leader != event)
3642                 leader->pmu->read(leader);
3643
3644         values[n++] = perf_event_count(leader);
3645         if (read_format & PERF_FORMAT_ID)
3646                 values[n++] = primary_event_id(leader);
3647
3648         perf_output_copy(handle, values, n * sizeof(u64));
3649
3650         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3651                 n = 0;
3652
3653                 if (sub != event)
3654                         sub->pmu->read(sub);
3655
3656                 values[n++] = perf_event_count(sub);
3657                 if (read_format & PERF_FORMAT_ID)
3658                         values[n++] = primary_event_id(sub);
3659
3660                 perf_output_copy(handle, values, n * sizeof(u64));
3661         }
3662 }
3663
3664 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
3665                                  PERF_FORMAT_TOTAL_TIME_RUNNING)
3666
3667 static void perf_output_read(struct perf_output_handle *handle,
3668                              struct perf_event *event)
3669 {
3670         u64 enabled = 0, running = 0, now, ctx_time;
3671         u64 read_format = event->attr.read_format;
3672
3673         /*
3674          * compute total_time_enabled, total_time_running
3675          * based on snapshot values taken when the event
3676          * was last scheduled in.
3677          *
3678          * we cannot simply called update_context_time()
3679          * because of locking issue as we are called in
3680          * NMI context
3681          */
3682         if (read_format & PERF_FORMAT_TOTAL_TIMES) {
3683                 now = perf_clock();
3684                 ctx_time = event->shadow_ctx_time + now;
3685                 enabled = ctx_time - event->tstamp_enabled;
3686                 running = ctx_time - event->tstamp_running;
3687         }
3688
3689         if (event->attr.read_format & PERF_FORMAT_GROUP)
3690                 perf_output_read_group(handle, event, enabled, running);
3691         else
3692                 perf_output_read_one(handle, event, enabled, running);
3693 }
3694
3695 void perf_output_sample(struct perf_output_handle *handle,
3696                         struct perf_event_header *header,
3697                         struct perf_sample_data *data,
3698                         struct perf_event *event)
3699 {
3700         u64 sample_type = data->type;
3701
3702         perf_output_put(handle, *header);
3703
3704         if (sample_type & PERF_SAMPLE_IP)
3705                 perf_output_put(handle, data->ip);
3706
3707         if (sample_type & PERF_SAMPLE_TID)
3708                 perf_output_put(handle, data->tid_entry);
3709
3710         if (sample_type & PERF_SAMPLE_TIME)
3711                 perf_output_put(handle, data->time);
3712
3713         if (sample_type & PERF_SAMPLE_ADDR)
3714                 perf_output_put(handle, data->addr);
3715
3716         if (sample_type & PERF_SAMPLE_ID)
3717                 perf_output_put(handle, data->id);
3718
3719         if (sample_type & PERF_SAMPLE_STREAM_ID)
3720                 perf_output_put(handle, data->stream_id);
3721
3722         if (sample_type & PERF_SAMPLE_CPU)
3723                 perf_output_put(handle, data->cpu_entry);
3724
3725         if (sample_type & PERF_SAMPLE_PERIOD)
3726                 perf_output_put(handle, data->period);
3727
3728         if (sample_type & PERF_SAMPLE_READ)
3729                 perf_output_read(handle, event);
3730
3731         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3732                 if (data->callchain) {
3733                         int size = 1;
3734
3735                         if (data->callchain)
3736                                 size += data->callchain->nr;
3737
3738                         size *= sizeof(u64);
3739
3740                         perf_output_copy(handle, data->callchain, size);
3741                 } else {
3742                         u64 nr = 0;
3743                         perf_output_put(handle, nr);
3744                 }
3745         }
3746
3747         if (sample_type & PERF_SAMPLE_RAW) {
3748                 if (data->raw) {
3749                         perf_output_put(handle, data->raw->size);
3750                         perf_output_copy(handle, data->raw->data,
3751                                          data->raw->size);
3752                 } else {
3753                         struct {
3754                                 u32     size;
3755                                 u32     data;
3756                         } raw = {
3757                                 .size = sizeof(u32),
3758                                 .data = 0,
3759                         };
3760                         perf_output_put(handle, raw);
3761                 }
3762         }
3763 }
3764
3765 void perf_prepare_sample(struct perf_event_header *header,
3766                          struct perf_sample_data *data,
3767                          struct perf_event *event,
3768                          struct pt_regs *regs)
3769 {
3770         u64 sample_type = event->attr.sample_type;
3771
3772         header->type = PERF_RECORD_SAMPLE;
3773         header->size = sizeof(*header) + event->header_size;
3774
3775         header->misc = 0;
3776         header->misc |= perf_misc_flags(regs);
3777
3778         __perf_event_header__init_id(header, data, event);
3779
3780         if (sample_type & PERF_SAMPLE_IP)
3781                 data->ip = perf_instruction_pointer(regs);
3782
3783         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3784                 int size = 1;
3785
3786                 data->callchain = perf_callchain(regs);
3787
3788                 if (data->callchain)
3789                         size += data->callchain->nr;
3790
3791                 header->size += size * sizeof(u64);
3792         }
3793
3794         if (sample_type & PERF_SAMPLE_RAW) {
3795                 int size = sizeof(u32);
3796
3797                 if (data->raw)
3798                         size += data->raw->size;
3799                 else
3800                         size += sizeof(u32);
3801
3802                 WARN_ON_ONCE(size & (sizeof(u64)-1));
3803                 header->size += size;
3804         }
3805 }
3806
3807 static void perf_event_output(struct perf_event *event, int nmi,
3808                                 struct perf_sample_data *data,
3809                                 struct pt_regs *regs)
3810 {
3811         struct perf_output_handle handle;
3812         struct perf_event_header header;
3813
3814         /* protect the callchain buffers */
3815         rcu_read_lock();
3816
3817         perf_prepare_sample(&header, data, event, regs);
3818
3819         if (perf_output_begin(&handle, event, header.size, nmi, 1))
3820                 goto exit;
3821
3822         perf_output_sample(&handle, &header, data, event);
3823
3824         perf_output_end(&handle);
3825
3826 exit:
3827         rcu_read_unlock();
3828 }
3829
3830 /*
3831  * read event_id
3832  */
3833
3834 struct perf_read_event {
3835         struct perf_event_header        header;
3836
3837         u32                             pid;
3838         u32                             tid;
3839 };
3840
3841 static void
3842 perf_event_read_event(struct perf_event *event,
3843                         struct task_struct *task)
3844 {
3845         struct perf_output_handle handle;
3846         struct perf_sample_data sample;
3847         struct perf_read_event read_event = {
3848                 .header = {
3849                         .type = PERF_RECORD_READ,
3850                         .misc = 0,
3851                         .size = sizeof(read_event) + event->read_size,
3852                 },
3853                 .pid = perf_event_pid(event, task),
3854                 .tid = perf_event_tid(event, task),
3855         };
3856         int ret;
3857
3858         perf_event_header__init_id(&read_event.header, &sample, event);
3859         ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3860         if (ret)
3861                 return;
3862
3863         perf_output_put(&handle, read_event);
3864         perf_output_read(&handle, event);
3865         perf_event__output_id_sample(event, &handle, &sample);
3866
3867         perf_output_end(&handle);
3868 }
3869
3870 /*
3871  * task tracking -- fork/exit
3872  *
3873  * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3874  */
3875
3876 struct perf_task_event {
3877         struct task_struct              *task;
3878         struct perf_event_context       *task_ctx;
3879
3880         struct {
3881                 struct perf_event_header        header;
3882
3883                 u32                             pid;
3884                 u32                             ppid;
3885                 u32                             tid;
3886                 u32                             ptid;
3887                 u64                             time;
3888         } event_id;
3889 };
3890
3891 static void perf_event_task_output(struct perf_event *event,
3892                                      struct perf_task_event *task_event)
3893 {
3894         struct perf_output_handle handle;
3895         struct perf_sample_data sample;
3896         struct task_struct *task = task_event->task;
3897         int ret, size = task_event->event_id.header.size;
3898
3899         perf_event_header__init_id(&task_event->event_id.header, &sample, event);
3900
3901         ret = perf_output_begin(&handle, event,
3902                                 task_event->event_id.header.size, 0, 0);
3903         if (ret)
3904                 goto out;
3905
3906         task_event->event_id.pid = perf_event_pid(event, task);
3907         task_event->event_id.ppid = perf_event_pid(event, current);
3908
3909         task_event->event_id.tid = perf_event_tid(event, task);
3910         task_event->event_id.ptid = perf_event_tid(event, current);
3911
3912         perf_output_put(&handle, task_event->event_id);
3913
3914         perf_event__output_id_sample(event, &handle, &sample);
3915
3916         perf_output_end(&handle);
3917 out:
3918         task_event->event_id.header.size = size;
3919 }
3920
3921 static int perf_event_task_match(struct perf_event *event)
3922 {
3923         if (event->state < PERF_EVENT_STATE_INACTIVE)
3924                 return 0;
3925
3926         if (!event_filter_match(event))
3927                 return 0;
3928
3929         if (event->attr.comm || event->attr.mmap ||
3930             event->attr.mmap_data || event->attr.task)
3931                 return 1;
3932
3933         return 0;
3934 }
3935
3936 static void perf_event_task_ctx(struct perf_event_context *ctx,
3937                                   struct perf_task_event *task_event)
3938 {
3939         struct perf_event *event;
3940
3941         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3942                 if (perf_event_task_match(event))
3943                         perf_event_task_output(event, task_event);
3944         }
3945 }
3946
3947 static void perf_event_task_event(struct perf_task_event *task_event)
3948 {
3949         struct perf_cpu_context *cpuctx;
3950         struct perf_event_context *ctx;
3951         struct pmu *pmu;
3952         int ctxn;
3953
3954         rcu_read_lock();
3955         list_for_each_entry_rcu(pmu, &pmus, entry) {
3956                 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
3957                 if (cpuctx->active_pmu != pmu)
3958                         goto next;
3959                 perf_event_task_ctx(&cpuctx->ctx, task_event);
3960
3961                 ctx = task_event->task_ctx;
3962                 if (!ctx) {
3963                         ctxn = pmu->task_ctx_nr;
3964                         if (ctxn < 0)
3965                                 goto next;
3966                         ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
3967                 }
3968                 if (ctx)
3969                         perf_event_task_ctx(ctx, task_event);
3970 next:
3971                 put_cpu_ptr(pmu->pmu_cpu_context);
3972         }
3973         rcu_read_unlock();
3974 }
3975
3976 static void perf_event_task(struct task_struct *task,
3977                               struct perf_event_context *task_ctx,
3978                               int new)
3979 {
3980         struct perf_task_event task_event;
3981
3982         if (!atomic_read(&nr_comm_events) &&
3983             !atomic_read(&nr_mmap_events) &&
3984             !atomic_read(&nr_task_events))
3985                 return;
3986
3987         task_event = (struct perf_task_event){
3988                 .task     = task,
3989                 .task_ctx = task_ctx,
3990                 .event_id    = {
3991                         .header = {
3992                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3993                                 .misc = 0,
3994                                 .size = sizeof(task_event.event_id),
3995                         },
3996                         /* .pid  */
3997                         /* .ppid */
3998                         /* .tid  */
3999                         /* .ptid */
4000                         .time = perf_clock(),
4001                 },
4002         };
4003
4004         perf_event_task_event(&task_event);
4005 }
4006
4007 void perf_event_fork(struct task_struct *task)
4008 {
4009         perf_event_task(task, NULL, 1);
4010 }
4011
4012 /*
4013  * comm tracking
4014  */
4015
4016 struct perf_comm_event {
4017         struct task_struct      *task;
4018         char                    *comm;
4019         int                     comm_size;
4020
4021         struct {
4022                 struct perf_event_header        header;
4023
4024                 u32                             pid;
4025                 u32                             tid;
4026         } event_id;
4027 };
4028
4029 static void perf_event_comm_output(struct perf_event *event,
4030                                      struct perf_comm_event *comm_event)
4031 {
4032         struct perf_output_handle handle;
4033         struct perf_sample_data sample;
4034         int size = comm_event->event_id.header.size;
4035         int ret;
4036
4037         perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4038         ret = perf_output_begin(&handle, event,
4039                                 comm_event->event_id.header.size, 0, 0);
4040
4041         if (ret)
4042                 goto out;
4043
4044         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4045         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4046
4047         perf_output_put(&handle, comm_event->event_id);
4048         perf_output_copy(&handle, comm_event->comm,
4049                                    comm_event->comm_size);
4050
4051         perf_event__output_id_sample(event, &handle, &sample);
4052
4053         perf_output_end(&handle);
4054 out:
4055         comm_event->event_id.header.size = size;
4056 }
4057
4058 static int perf_event_comm_match(struct perf_event *event)
4059 {
4060         if (event->state < PERF_EVENT_STATE_INACTIVE)
4061                 return 0;
4062
4063         if (!event_filter_match(event))
4064                 return 0;
4065
4066         if (event->attr.comm)
4067                 return 1;
4068
4069         return 0;
4070 }
4071
4072 static void perf_event_comm_ctx(struct perf_event_context *ctx,
4073                                   struct perf_comm_event *comm_event)
4074 {
4075         struct perf_event *event;
4076
4077         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4078                 if (perf_event_comm_match(event))
4079                         perf_event_comm_output(event, comm_event);
4080         }
4081 }
4082
4083 static void perf_event_comm_event(struct perf_comm_event *comm_event)
4084 {
4085         struct perf_cpu_context *cpuctx;
4086         struct perf_event_context *ctx;
4087         char comm[TASK_COMM_LEN];
4088         unsigned int size;
4089         struct pmu *pmu;
4090         int ctxn;
4091
4092         memset(comm, 0, sizeof(comm));
4093         strlcpy(comm, comm_event->task->comm, sizeof(comm));
4094         size = ALIGN(strlen(comm)+1, sizeof(u64));
4095
4096         comm_event->comm = comm;
4097         comm_event->comm_size = size;
4098
4099         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
4100         rcu_read_lock();
4101         list_for_each_entry_rcu(pmu, &pmus, entry) {
4102                 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4103                 if (cpuctx->active_pmu != pmu)
4104                         goto next;
4105                 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
4106
4107                 ctxn = pmu->task_ctx_nr;
4108                 if (ctxn < 0)
4109                         goto next;
4110
4111                 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4112                 if (ctx)
4113                         perf_event_comm_ctx(ctx, comm_event);
4114 next:
4115                 put_cpu_ptr(pmu->pmu_cpu_context);
4116         }
4117         rcu_read_unlock();
4118 }
4119
4120 void perf_event_comm(struct task_struct *task)
4121 {
4122         struct perf_comm_event comm_event;
4123         struct perf_event_context *ctx;
4124         int ctxn;
4125
4126         for_each_task_context_nr(ctxn) {
4127                 ctx = task->perf_event_ctxp[ctxn];
4128                 if (!ctx)
4129                         continue;
4130
4131                 perf_event_enable_on_exec(ctx);
4132         }
4133
4134         if (!atomic_read(&nr_comm_events))
4135                 return;
4136
4137         comm_event = (struct perf_comm_event){
4138                 .task   = task,
4139                 /* .comm      */
4140                 /* .comm_size */
4141                 .event_id  = {
4142                         .header = {
4143                                 .type = PERF_RECORD_COMM,
4144                                 .misc = 0,
4145                                 /* .size */
4146                         },
4147                         /* .pid */
4148                         /* .tid */
4149                 },
4150         };
4151
4152         perf_event_comm_event(&comm_event);
4153 }
4154
4155 /*
4156  * mmap tracking
4157  */
4158
4159 struct perf_mmap_event {
4160         struct vm_area_struct   *vma;
4161
4162         const char              *file_name;
4163         int                     file_size;
4164
4165         struct {
4166                 struct perf_event_header        header;
4167
4168                 u32                             pid;
4169                 u32                             tid;
4170                 u64                             start;
4171                 u64                             len;
4172                 u64                             pgoff;
4173         } event_id;
4174 };
4175
4176 static void perf_event_mmap_output(struct perf_event *event,
4177                                      struct perf_mmap_event *mmap_event)
4178 {
4179         struct perf_output_handle handle;
4180         struct perf_sample_data sample;
4181         int size = mmap_event->event_id.header.size;
4182         int ret;
4183
4184         perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
4185         ret = perf_output_begin(&handle, event,
4186                                 mmap_event->event_id.header.size, 0, 0);
4187         if (ret)
4188                 goto out;
4189
4190         mmap_event->event_id.pid = perf_event_pid(event, current);
4191         mmap_event->event_id.tid = perf_event_tid(event, current);
4192
4193         perf_output_put(&handle, mmap_event->event_id);
4194         perf_output_copy(&handle, mmap_event->file_name,
4195                                    mmap_event->file_size);
4196
4197         perf_event__output_id_sample(event, &handle, &sample);
4198
4199         perf_output_end(&handle);
4200 out:
4201         mmap_event->event_id.header.size = size;
4202 }
4203
4204 static int perf_event_mmap_match(struct perf_event *event,
4205                                    struct perf_mmap_event *mmap_event,
4206                                    int executable)
4207 {
4208         if (event->state < PERF_EVENT_STATE_INACTIVE)
4209                 return 0;
4210
4211         if (!event_filter_match(event))
4212                 return 0;
4213
4214         if ((!executable && event->attr.mmap_data) ||
4215             (executable && event->attr.mmap))
4216                 return 1;
4217
4218         return 0;
4219 }
4220
4221 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
4222                                   struct perf_mmap_event *mmap_event,
4223                                   int executable)
4224 {
4225         struct perf_event *event;
4226
4227         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4228                 if (perf_event_mmap_match(event, mmap_event, executable))
4229                         perf_event_mmap_output(event, mmap_event);
4230         }
4231 }
4232
4233 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4234 {
4235         struct perf_cpu_context *cpuctx;
4236         struct perf_event_context *ctx;
4237         struct vm_area_struct *vma = mmap_event->vma;
4238         struct file *file = vma->vm_file;
4239         unsigned int size;
4240         char tmp[16];
4241         char *buf = NULL;
4242         const char *name;
4243         struct pmu *pmu;
4244         int ctxn;
4245
4246         memset(tmp, 0, sizeof(tmp));
4247
4248         if (file) {
4249                 /*
4250                  * d_path works from the end of the buffer backwards, so we
4251                  * need to add enough zero bytes after the string to handle
4252                  * the 64bit alignment we do later.
4253                  */
4254                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4255                 if (!buf) {
4256                         name = strncpy(tmp, "//enomem", sizeof(tmp));
4257                         goto got_name;
4258                 }
4259                 name = d_path(&file->f_path, buf, PATH_MAX);
4260                 if (IS_ERR(name)) {
4261                         name = strncpy(tmp, "//toolong", sizeof(tmp));
4262                         goto got_name;
4263                 }
4264         } else {
4265                 if (arch_vma_name(mmap_event->vma)) {
4266                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4267                                        sizeof(tmp));
4268                         goto got_name;
4269                 }
4270
4271                 if (!vma->vm_mm) {
4272                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
4273                         goto got_name;
4274                 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4275                                 vma->vm_end >= vma->vm_mm->brk) {
4276                         name = strncpy(tmp, "[heap]", sizeof(tmp));
4277                         goto got_name;
4278                 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4279                                 vma->vm_end >= vma->vm_mm->start_stack) {
4280                         name = strncpy(tmp, "[stack]", sizeof(tmp));
4281                         goto got_name;
4282                 }
4283
4284                 name = strncpy(tmp, "//anon", sizeof(tmp));
4285                 goto got_name;
4286         }
4287
4288 got_name:
4289         size = ALIGN(strlen(name)+1, sizeof(u64));
4290
4291         mmap_event->file_name = name;
4292         mmap_event->file_size = size;
4293
4294         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4295
4296         rcu_read_lock();
4297         list_for_each_entry_rcu(pmu, &pmus, entry) {
4298                 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4299                 if (cpuctx->active_pmu != pmu)
4300                         goto next;
4301                 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
4302                                         vma->vm_flags & VM_EXEC);
4303
4304                 ctxn = pmu->task_ctx_nr;
4305                 if (ctxn < 0)
4306                         goto next;
4307
4308                 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4309                 if (ctx) {
4310                         perf_event_mmap_ctx(ctx, mmap_event,
4311                                         vma->vm_flags & VM_EXEC);
4312                 }
4313 next:
4314                 put_cpu_ptr(pmu->pmu_cpu_context);
4315         }
4316         rcu_read_unlock();
4317
4318         kfree(buf);
4319 }
4320
4321 void perf_event_mmap(struct vm_area_struct *vma)
4322 {
4323         struct perf_mmap_event mmap_event;
4324
4325         if (!atomic_read(&nr_mmap_events))
4326                 return;
4327
4328         mmap_event = (struct perf_mmap_event){
4329                 .vma    = vma,
4330                 /* .file_name */
4331                 /* .file_size */
4332                 .event_id  = {
4333                         .header = {
4334                                 .type = PERF_RECORD_MMAP,
4335                                 .misc = PERF_RECORD_MISC_USER,
4336                                 /* .size */
4337                         },
4338                         /* .pid */
4339                         /* .tid */
4340                         .start  = vma->vm_start,
4341                         .len    = vma->vm_end - vma->vm_start,
4342                         .pgoff  = (u64)vma->vm_pgoff << PAGE_SHIFT,
4343                 },
4344         };
4345
4346         perf_event_mmap_event(&mmap_event);
4347 }
4348
4349 /*
4350  * IRQ throttle logging
4351  */
4352
4353 static void perf_log_throttle(struct perf_event *event, int enable)
4354 {
4355         struct perf_output_handle handle;
4356         struct perf_sample_data sample;
4357         int ret;
4358
4359         struct {
4360                 struct perf_event_header        header;
4361                 u64                             time;
4362                 u64                             id;
4363                 u64                             stream_id;
4364         } throttle_event = {
4365                 .header = {
4366                         .type = PERF_RECORD_THROTTLE,
4367                         .misc = 0,
4368                         .size = sizeof(throttle_event),
4369                 },
4370                 .time           = perf_clock(),
4371                 .id             = primary_event_id(event),
4372                 .stream_id      = event->id,
4373         };
4374
4375         if (enable)
4376                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4377
4378         perf_event_header__init_id(&throttle_event.header, &sample, event);
4379
4380         ret = perf_output_begin(&handle, event,
4381                                 throttle_event.header.size, 1, 0);
4382         if (ret)
4383                 return;
4384
4385         perf_output_put(&handle, throttle_event);
4386         perf_event__output_id_sample(event, &handle, &sample);
4387         perf_output_end(&handle);
4388 }
4389
4390 /*
4391  * Generic event overflow handling, sampling.
4392  */
4393
4394 static int __perf_event_overflow(struct perf_event *event, int nmi,
4395                                    int throttle, struct perf_sample_data *data,
4396                                    struct pt_regs *regs)
4397 {
4398         int events = atomic_read(&event->event_limit);
4399         struct hw_perf_event *hwc = &event->hw;
4400         int ret = 0;
4401
4402         /*
4403          * Non-sampling counters might still use the PMI to fold short
4404          * hardware counters, ignore those.
4405          */
4406         if (unlikely(!is_sampling_event(event)))
4407                 return 0;
4408
4409         if (!throttle) {
4410                 hwc->interrupts++;
4411         } else {
4412                 if (hwc->interrupts != MAX_INTERRUPTS) {
4413                         hwc->interrupts++;
4414                         if (HZ * hwc->interrupts >
4415                                         (u64)sysctl_perf_event_sample_rate) {
4416                                 hwc->interrupts = MAX_INTERRUPTS;
4417                                 perf_log_throttle(event, 0);
4418                                 ret = 1;
4419                         }
4420                 } else {
4421                         /*
4422                          * Keep re-disabling events even though on the previous
4423                          * pass we disabled it - just in case we raced with a
4424                          * sched-in and the event got enabled again:
4425                          */
4426                         ret = 1;
4427                 }
4428         }
4429
4430         if (event->attr.freq) {
4431                 u64 now = perf_clock();
4432                 s64 delta = now - hwc->freq_time_stamp;
4433
4434                 hwc->freq_time_stamp = now;
4435
4436                 if (delta > 0 && delta < 2*TICK_NSEC)
4437                         perf_adjust_period(event, delta, hwc->last_period);
4438         }
4439
4440         /*
4441          * XXX event_limit might not quite work as expected on inherited
4442          * events
4443          */
4444
4445         event->pending_kill = POLL_IN;
4446         if (events && atomic_dec_and_test(&event->event_limit)) {
4447                 ret = 1;
4448                 event->pending_kill = POLL_HUP;
4449                 if (nmi) {
4450                         event->pending_disable = 1;
4451                         irq_work_queue(&event->pending);
4452                 } else
4453                         perf_event_disable(event);
4454         }
4455
4456         if (event->overflow_handler)
4457                 event->overflow_handler(event, nmi, data, regs);
4458         else
4459                 perf_event_output(event, nmi, data, regs);
4460
4461         return ret;
4462 }
4463
4464 int perf_event_overflow(struct perf_event *event, int nmi,
4465                           struct perf_sample_data *data,
4466                           struct pt_regs *regs)
4467 {
4468         return __perf_event_overflow(event, nmi, 1, data, regs);
4469 }
4470
4471 /*
4472  * Generic software event infrastructure
4473  */
4474
4475 struct swevent_htable {
4476         struct swevent_hlist            *swevent_hlist;
4477         struct mutex                    hlist_mutex;
4478         int                             hlist_refcount;
4479
4480         /* Recursion avoidance in each contexts */
4481         int                             recursion[PERF_NR_CONTEXTS];
4482 };
4483
4484 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
4485
4486 /*
4487  * We directly increment event->count and keep a second value in
4488  * event->hw.period_left to count intervals. This period event
4489  * is kept in the range [-sample_period, 0] so that we can use the
4490  * sign as trigger.
4491  */
4492
4493 static u64 perf_swevent_set_period(struct perf_event *event)
4494 {
4495         struct hw_perf_event *hwc = &event->hw;
4496         u64 period = hwc->last_period;
4497         u64 nr, offset;
4498         s64 old, val;
4499
4500         hwc->last_period = hwc->sample_period;
4501
4502 again:
4503         old = val = local64_read(&hwc->period_left);
4504         if (val < 0)
4505                 return 0;
4506
4507         nr = div64_u64(period + val, period);
4508         offset = nr * period;
4509         val -= offset;
4510         if (local64_cmpxchg(&hwc->period_left, old, val) != old)
4511                 goto again;
4512
4513         return nr;
4514 }
4515
4516 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4517                                     int nmi, struct perf_sample_data *data,
4518                                     struct pt_regs *regs)
4519 {
4520         struct hw_perf_event *hwc = &event->hw;
4521         int throttle = 0;
4522
4523         data->period = event->hw.last_period;
4524         if (!overflow)
4525                 overflow = perf_swevent_set_period(event);
4526
4527         if (hwc->interrupts == MAX_INTERRUPTS)
4528                 return;
4529
4530         for (; overflow; overflow--) {
4531                 if (__perf_event_overflow(event, nmi, throttle,
4532                                             data, regs)) {
4533                         /*
4534                          * We inhibit the overflow from happening when
4535                          * hwc->interrupts == MAX_INTERRUPTS.
4536                          */
4537                         break;
4538                 }
4539                 throttle = 1;
4540         }
4541 }
4542
4543 static void perf_swevent_event(struct perf_event *event, u64 nr,
4544                                int nmi, struct perf_sample_data *data,
4545                                struct pt_regs *regs)
4546 {
4547         struct hw_perf_event *hwc = &event->hw;
4548
4549         local64_add(nr, &event->count);
4550
4551         if (!regs)
4552                 return;
4553
4554         if (!is_sampling_event(event))
4555                 return;
4556
4557         if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4558                 return perf_swevent_overflow(event, 1, nmi, data, regs);
4559
4560         if (local64_add_negative(nr, &hwc->period_left))
4561                 return;
4562
4563         perf_swevent_overflow(event, 0, nmi, data, regs);
4564 }
4565
4566 static int perf_exclude_event(struct perf_event *event,
4567                               struct pt_regs *regs)
4568 {
4569         if (event->hw.state & PERF_HES_STOPPED)
4570                 return 0;
4571
4572         if (regs) {
4573                 if (event->attr.exclude_user && user_mode(regs))
4574                         return 1;
4575
4576                 if (event->attr.exclude_kernel && !user_mode(regs))
4577                         return 1;
4578         }
4579
4580         return 0;
4581 }
4582
4583 static int perf_swevent_match(struct perf_event *event,
4584                                 enum perf_type_id type,
4585                                 u32 event_id,
4586                                 struct perf_sample_data *data,
4587                                 struct pt_regs *regs)
4588 {
4589         if (event->attr.type != type)
4590                 return 0;
4591
4592         if (event->attr.config != event_id)
4593                 return 0;
4594
4595         if (perf_exclude_event(event, regs))
4596                 return 0;
4597
4598         return 1;
4599 }
4600
4601 static inline u64 swevent_hash(u64 type, u32 event_id)
4602 {
4603         u64 val = event_id | (type << 32);
4604
4605         return hash_64(val, SWEVENT_HLIST_BITS);
4606 }
4607
4608 static inline struct hlist_head *
4609 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4610 {
4611         u64 hash = swevent_hash(type, event_id);
4612
4613         return &hlist->heads[hash];
4614 }
4615
4616 /* For the read side: events when they trigger */
4617 static inline struct hlist_head *
4618 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
4619 {
4620         struct swevent_hlist *hlist;
4621
4622         hlist = rcu_dereference(swhash->swevent_hlist);
4623         if (!hlist)
4624                 return NULL;
4625
4626         return __find_swevent_head(hlist, type, event_id);
4627 }
4628
4629 /* For the event head insertion and removal in the hlist */
4630 static inline struct hlist_head *
4631 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
4632 {
4633         struct swevent_hlist *hlist;
4634         u32 event_id = event->attr.config;
4635         u64 type = event->attr.type;
4636
4637         /*
4638          * Event scheduling is always serialized against hlist allocation
4639          * and release. Which makes the protected version suitable here.
4640          * The context lock guarantees that.
4641          */
4642         hlist = rcu_dereference_protected(swhash->swevent_hlist,
4643                                           lockdep_is_held(&event->ctx->lock));
4644         if (!hlist)
4645                 return NULL;
4646
4647         return __find_swevent_head(hlist, type, event_id);
4648 }
4649
4650 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4651                                     u64 nr, int nmi,
4652                                     struct perf_sample_data *data,
4653                                     struct pt_regs *regs)
4654 {
4655         struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4656         struct perf_event *event;
4657         struct hlist_node *node;
4658         struct hlist_head *head;
4659
4660         rcu_read_lock();
4661         head = find_swevent_head_rcu(swhash, type, event_id);
4662         if (!head)
4663                 goto end;
4664
4665         hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4666                 if (perf_swevent_match(event, type, event_id, data, regs))
4667                         perf_swevent_event(event, nr, nmi, data, regs);
4668         }
4669 end:
4670         rcu_read_unlock();
4671 }
4672
4673 int perf_swevent_get_recursion_context(void)
4674 {
4675         struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4676
4677         return get_recursion_context(swhash->recursion);
4678 }
4679 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4680
4681 inline void perf_swevent_put_recursion_context(int rctx)
4682 {
4683         struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4684
4685         put_recursion_context(swhash->recursion, rctx);
4686 }
4687
4688 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4689                             struct pt_regs *regs, u64 addr)
4690 {
4691         struct perf_sample_data data;
4692         int rctx;
4693
4694         preempt_disable_notrace();
4695         rctx = perf_swevent_get_recursion_context();
4696         if (rctx < 0)
4697                 return;
4698
4699         perf_sample_data_init(&data, addr);
4700
4701         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4702
4703         perf_swevent_put_recursion_context(rctx);
4704         preempt_enable_notrace();
4705 }
4706
4707 static void perf_swevent_read(struct perf_event *event)
4708 {
4709 }
4710
4711 static int perf_swevent_add(struct perf_event *event, int flags)
4712 {
4713         struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4714         struct hw_perf_event *hwc = &event->hw;
4715         struct hlist_head *head;
4716
4717         if (is_sampling_event(event)) {
4718                 hwc->last_period = hwc->sample_period;
4719                 perf_swevent_set_period(event);
4720         }
4721
4722         hwc->state = !(flags & PERF_EF_START);
4723
4724         head = find_swevent_head(swhash, event);
4725         if (WARN_ON_ONCE(!head))
4726                 return -EINVAL;
4727
4728         hlist_add_head_rcu(&event->hlist_entry, head);
4729
4730         return 0;
4731 }
4732
4733 static void perf_swevent_del(struct perf_event *event, int flags)
4734 {
4735         hlist_del_rcu(&event->hlist_entry);
4736 }
4737
4738 static void perf_swevent_start(struct perf_event *event, int flags)
4739 {
4740         event->hw.state = 0;
4741 }
4742
4743 static void perf_swevent_stop(struct perf_event *event, int flags)
4744 {
4745         event->hw.state = PERF_HES_STOPPED;
4746 }
4747
4748 /* Deref the hlist from the update side */
4749 static inline struct swevent_hlist *
4750 swevent_hlist_deref(struct swevent_htable *swhash)
4751 {
4752         return rcu_dereference_protected(swhash->swevent_hlist,
4753                                          lockdep_is_held(&swhash->hlist_mutex));
4754 }
4755
4756 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4757 {
4758         struct swevent_hlist *hlist;
4759
4760         hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4761         kfree(hlist);
4762 }
4763
4764 static void swevent_hlist_release(struct swevent_htable *swhash)
4765 {
4766         struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
4767
4768         if (!hlist)
4769                 return;
4770
4771         rcu_assign_pointer(swhash->swevent_hlist, NULL);
4772         call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4773 }
4774
4775 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4776 {
4777         struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4778
4779         mutex_lock(&swhash->hlist_mutex);
4780
4781         if (!--swhash->hlist_refcount)
4782                 swevent_hlist_release(swhash);
4783
4784         mutex_unlock(&swhash->hlist_mutex);
4785 }
4786
4787 static void swevent_hlist_put(struct perf_event *event)
4788 {
4789         int cpu;
4790
4791         if (event->cpu != -1) {
4792                 swevent_hlist_put_cpu(event, event->cpu);
4793                 return;
4794         }
4795
4796         for_each_possible_cpu(cpu)
4797                 swevent_hlist_put_cpu(event, cpu);
4798 }
4799
4800 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4801 {
4802         struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4803         int err = 0;
4804
4805         mutex_lock(&swhash->hlist_mutex);
4806
4807         if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
4808                 struct swevent_hlist *hlist;
4809
4810                 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4811                 if (!hlist) {
4812                         err = -ENOMEM;
4813                         goto exit;
4814                 }
4815                 rcu_assign_pointer(swhash->swevent_hlist, hlist);
4816         }
4817         swhash->hlist_refcount++;
4818 exit:
4819         mutex_unlock(&swhash->hlist_mutex);
4820
4821         return err;
4822 }
4823
4824 static int swevent_hlist_get(struct perf_event *event)
4825 {
4826         int err;
4827         int cpu, failed_cpu;
4828
4829         if (event->cpu != -1)
4830                 return swevent_hlist_get_cpu(event, event->cpu);
4831
4832         get_online_cpus();
4833         for_each_possible_cpu(cpu) {
4834                 err = swevent_hlist_get_cpu(event, cpu);
4835                 if (err) {
4836                         failed_cpu = cpu;
4837                         goto fail;
4838                 }
4839         }
4840         put_online_cpus();
4841
4842         return 0;
4843 fail:
4844         for_each_possible_cpu(cpu) {
4845                 if (cpu == failed_cpu)
4846                         break;
4847                 swevent_hlist_put_cpu(event, cpu);
4848         }
4849
4850         put_online_cpus();
4851         return err;
4852 }
4853
4854 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4855
4856 static void sw_perf_event_destroy(struct perf_event *event)
4857 {
4858         u64 event_id = event->attr.config;
4859
4860         WARN_ON(event->parent);
4861
4862         jump_label_dec(&perf_swevent_enabled[event_id]);
4863         swevent_hlist_put(event);
4864 }
4865
4866 static int perf_swevent_init(struct perf_event *event)
4867 {
4868         int event_id = event->attr.config;
4869
4870         if (event->attr.type != PERF_TYPE_SOFTWARE)
4871                 return -ENOENT;
4872
4873         switch (event_id) {
4874         case PERF_COUNT_SW_CPU_CLOCK:
4875         case PERF_COUNT_SW_TASK_CLOCK:
4876                 return -ENOENT;
4877
4878         default:
4879                 break;
4880         }
4881
4882         if (event_id >= PERF_COUNT_SW_MAX)
4883                 return -ENOENT;
4884
4885         if (!event->parent) {
4886                 int err;
4887
4888                 err = swevent_hlist_get(event);
4889                 if (err)
4890                         return err;
4891
4892                 jump_label_inc(&perf_swevent_enabled[event_id]);
4893                 event->destroy = sw_perf_event_destroy;
4894         }
4895
4896         return 0;
4897 }
4898
4899 static struct pmu perf_swevent = {
4900         .task_ctx_nr    = perf_sw_context,
4901
4902         .event_init     = perf_swevent_init,
4903         .add            = perf_swevent_add,
4904         .del            = perf_swevent_del,
4905         .start          = perf_swevent_start,
4906         .stop           = perf_swevent_stop,
4907         .read           = perf_swevent_read,
4908 };
4909
4910 #ifdef CONFIG_EVENT_TRACING
4911
4912 static int perf_tp_filter_match(struct perf_event *event,
4913                                 struct perf_sample_data *data)
4914 {
4915         void *record = data->raw->data;
4916
4917         if (likely(!event->filter) || filter_match_preds(event->filter, record))
4918                 return 1;
4919         return 0;
4920 }
4921
4922 static int perf_tp_event_match(struct perf_event *event,
4923                                 struct perf_sample_data *data,
4924                                 struct pt_regs *regs)
4925 {
4926         /*
4927          * All tracepoints are from kernel-space.
4928          */
4929         if (event->attr.exclude_kernel)
4930                 return 0;
4931
4932         if (!perf_tp_filter_match(event, data))
4933                 return 0;
4934
4935         return 1;
4936 }
4937
4938 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
4939                    struct pt_regs *regs, struct hlist_head *head, int rctx)
4940 {
4941         struct perf_sample_data data;
4942         struct perf_event *event;
4943         struct hlist_node *node;
4944
4945         struct perf_raw_record raw = {
4946                 .size = entry_size,
4947                 .data = record,
4948         };
4949
4950         perf_sample_data_init(&data, addr);
4951         data.raw = &raw;
4952
4953         hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4954                 if (perf_tp_event_match(event, &data, regs))
4955                         perf_swevent_event(event, count, 1, &data, regs);
4956         }
4957
4958         perf_swevent_put_recursion_context(rctx);
4959 }
4960 EXPORT_SYMBOL_GPL(perf_tp_event);
4961
4962 static void tp_perf_event_destroy(struct perf_event *event)
4963 {
4964         perf_trace_destroy(event);
4965 }
4966
4967 static int perf_tp_event_init(struct perf_event *event)
4968 {
4969         int err;
4970
4971         if (event->attr.type != PERF_TYPE_TRACEPOINT)
4972                 return -ENOENT;
4973
4974         err = perf_trace_init(event);
4975         if (err)
4976                 return err;
4977
4978         event->destroy = tp_perf_event_destroy;
4979
4980         return 0;
4981 }
4982
4983 static struct pmu perf_tracepoint = {
4984         .task_ctx_nr    = perf_sw_context,
4985
4986         .event_init     = perf_tp_event_init,
4987         .add            = perf_trace_add,
4988         .del            = perf_trace_del,
4989         .start          = perf_swevent_start,
4990         .stop           = perf_swevent_stop,
4991         .read           = perf_swevent_read,
4992 };
4993
4994 static inline void perf_tp_register(void)
4995 {
4996         perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
4997 }
4998
4999 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5000 {
5001         char *filter_str;
5002         int ret;
5003
5004         if (event->attr.type != PERF_TYPE_TRACEPOINT)
5005                 return -EINVAL;
5006
5007         filter_str = strndup_user(arg, PAGE_SIZE);
5008         if (IS_ERR(filter_str))
5009                 return PTR_ERR(filter_str);
5010
5011         ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5012
5013         kfree(filter_str);
5014         return ret;
5015 }
5016
5017 static void perf_event_free_filter(struct perf_event *event)
5018 {
5019         ftrace_profile_free_filter(event);
5020 }
5021
5022 #else
5023
5024 static inline void perf_tp_register(void)
5025 {
5026 }
5027
5028 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5029 {
5030         return -ENOENT;
5031 }
5032
5033 static void perf_event_free_filter(struct perf_event *event)
5034 {
5035 }
5036
5037 #endif /* CONFIG_EVENT_TRACING */
5038
5039 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5040 void perf_bp_event(struct perf_event *bp, void *data)
5041 {
5042         struct perf_sample_data sample;
5043         struct pt_regs *regs = data;
5044
5045         perf_sample_data_init(&sample, bp->attr.bp_addr);
5046
5047         if (!bp->hw.state && !perf_exclude_event(bp, regs))
5048                 perf_swevent_event(bp, 1, 1, &sample, regs);
5049 }
5050 #endif
5051
5052 /*
5053  * hrtimer based swevent callback
5054  */
5055
5056 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
5057 {
5058         enum hrtimer_restart ret = HRTIMER_RESTART;
5059         struct perf_sample_data data;
5060         struct pt_regs *regs;
5061         struct perf_event *event;
5062         u64 period;
5063
5064         event = container_of(hrtimer, struct perf_event, hw.hrtimer);
5065         event->pmu->read(event);
5066
5067         perf_sample_data_init(&data, 0);
5068         data.period = event->hw.last_period;
5069         regs = get_irq_regs();
5070
5071         if (regs && !perf_exclude_event(event, regs)) {
5072                 if (!(event->attr.exclude_idle && current->pid == 0))
5073                         if (perf_event_overflow(event, 0, &data, regs))
5074                                 ret = HRTIMER_NORESTART;
5075         }
5076
5077         period = max_t(u64, 10000, event->hw.sample_period);
5078         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
5079
5080         return ret;
5081 }
5082
5083 static void perf_swevent_start_hrtimer(struct perf_event *event)
5084 {
5085         struct hw_perf_event *hwc = &event->hw;
5086         s64 period;
5087
5088         if (!is_sampling_event(event))
5089                 return;
5090
5091         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5092         hwc->hrtimer.function = perf_swevent_hrtimer;
5093
5094         period = local64_read(&hwc->period_left);
5095         if (period) {
5096                 if (period < 0)
5097                         period = 10000;
5098
5099                 local64_set(&hwc->period_left, 0);
5100         } else {
5101                 period = max_t(u64, 10000, hwc->sample_period);
5102         }
5103         __hrtimer_start_range_ns(&hwc->hrtimer,
5104                                 ns_to_ktime(period), 0,
5105                                 HRTIMER_MODE_REL_PINNED, 0);
5106 }
5107
5108 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
5109 {
5110         struct hw_perf_event *hwc = &event->hw;
5111
5112         if (is_sampling_event(event)) {
5113                 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
5114                 local64_set(&hwc->period_left, ktime_to_ns(remaining));
5115
5116                 hrtimer_cancel(&hwc->hrtimer);
5117         }
5118 }
5119
5120 /*
5121  * Software event: cpu wall time clock
5122  */
5123
5124 static void cpu_clock_event_update(struct perf_event *event)
5125 {
5126         s64 prev;
5127         u64 now;
5128
5129         now = local_clock();
5130         prev = local64_xchg(&event->hw.prev_count, now);
5131         local64_add(now - prev, &event->count);
5132 }
5133
5134 static void cpu_clock_event_start(struct perf_event *event, int flags)
5135 {
5136         local64_set(&event->hw.prev_count, local_clock());
5137         perf_swevent_start_hrtimer(event);
5138 }
5139
5140 static void cpu_clock_event_stop(struct perf_event *event, int flags)
5141 {
5142         perf_swevent_cancel_hrtimer(event);
5143         cpu_clock_event_update(event);
5144 }
5145
5146 static int cpu_clock_event_add(struct perf_event *event, int flags)
5147 {
5148         if (flags & PERF_EF_START)
5149                 cpu_clock_event_start(event, flags);
5150
5151         return 0;
5152 }
5153
5154 static void cpu_clock_event_del(struct perf_event *event, int flags)
5155 {
5156         cpu_clock_event_stop(event, flags);
5157 }
5158
5159 static void cpu_clock_event_read(struct perf_event *event)
5160 {
5161         cpu_clock_event_update(event);
5162 }
5163
5164 static int cpu_clock_event_init(struct perf_event *event)
5165 {
5166         if (event->attr.type != PERF_TYPE_SOFTWARE)
5167                 return -ENOENT;
5168
5169         if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
5170                 return -ENOENT;
5171
5172         return 0;
5173 }
5174
5175 static struct pmu perf_cpu_clock = {
5176         .task_ctx_nr    = perf_sw_context,
5177
5178         .event_init     = cpu_clock_event_init,
5179         .add            = cpu_clock_event_add,
5180         .del            = cpu_clock_event_del,
5181         .start          = cpu_clock_event_start,
5182         .stop           = cpu_clock_event_stop,
5183         .read           = cpu_clock_event_read,
5184 };
5185
5186 /*
5187  * Software event: task time clock
5188  */
5189
5190 static void task_clock_event_update(struct perf_event *event, u64 now)
5191 {
5192         u64 prev;
5193         s64 delta;
5194
5195         prev = local64_xchg(&event->hw.prev_count, now);
5196         delta = now - prev;
5197         local64_add(delta, &event->count);
5198 }
5199
5200 static void task_clock_event_start(struct perf_event *event, int flags)
5201 {
5202         local64_set(&event->hw.prev_count, event->ctx->time);
5203         perf_swevent_start_hrtimer(event);
5204 }
5205
5206 static void task_clock_event_stop(struct perf_event *event, int flags)
5207 {
5208         perf_swevent_cancel_hrtimer(event);
5209         task_clock_event_update(event, event->ctx->time);
5210 }
5211
5212 static int task_clock_event_add(struct perf_event *event, int flags)
5213 {
5214         if (flags & PERF_EF_START)
5215                 task_clock_event_start(event, flags);
5216
5217         return 0;
5218 }
5219
5220 static void task_clock_event_del(struct perf_event *event, int flags)
5221 {
5222         task_clock_event_stop(event, PERF_EF_UPDATE);
5223 }
5224
5225 static void task_clock_event_read(struct perf_event *event)
5226 {
5227         u64 time;
5228
5229         if (!in_nmi()) {
5230                 update_context_time(event->ctx);
5231                 time = event->ctx->time;
5232         } else {
5233                 u64 now = perf_clock();
5234                 u64 delta = now - event->ctx->timestamp;
5235                 time = event->ctx->time + delta;
5236         }
5237
5238         task_clock_event_update(event, time);
5239 }
5240
5241 static int task_clock_event_init(struct perf_event *event)
5242 {
5243         if (event->attr.type != PERF_TYPE_SOFTWARE)
5244                 return -ENOENT;
5245
5246         if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5247                 return -ENOENT;
5248
5249         return 0;
5250 }
5251
5252 static struct pmu perf_task_clock = {
5253         .task_ctx_nr    = perf_sw_context,
5254
5255         .event_init     = task_clock_event_init,
5256         .add            = task_clock_event_add,
5257         .del            = task_clock_event_del,
5258         .start          = task_clock_event_start,
5259         .stop           = task_clock_event_stop,
5260         .read           = task_clock_event_read,
5261 };
5262
5263 static void perf_pmu_nop_void(struct pmu *pmu)
5264 {
5265 }
5266
5267 static int perf_pmu_nop_int(struct pmu *pmu)
5268 {
5269         return 0;
5270 }
5271
5272 static void perf_pmu_start_txn(struct pmu *pmu)
5273 {
5274         perf_pmu_disable(pmu);
5275 }
5276
5277 static int perf_pmu_commit_txn(struct pmu *pmu)
5278 {
5279         perf_pmu_enable(pmu);
5280         return 0;
5281 }
5282
5283 static void perf_pmu_cancel_txn(struct pmu *pmu)
5284 {
5285         perf_pmu_enable(pmu);
5286 }
5287
5288 /*
5289  * Ensures all contexts with the same task_ctx_nr have the same
5290  * pmu_cpu_context too.
5291  */
5292 static void *find_pmu_context(int ctxn)
5293 {
5294         struct pmu *pmu;
5295
5296         if (ctxn < 0)
5297                 return NULL;
5298
5299         list_for_each_entry(pmu, &pmus, entry) {
5300                 if (pmu->task_ctx_nr == ctxn)
5301                         return pmu->pmu_cpu_context;
5302         }
5303
5304         return NULL;
5305 }
5306
5307 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
5308 {
5309         int cpu;
5310
5311         for_each_possible_cpu(cpu) {
5312                 struct perf_cpu_context *cpuctx;
5313
5314                 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5315
5316                 if (cpuctx->active_pmu == old_pmu)
5317                         cpuctx->active_pmu = pmu;
5318         }
5319 }
5320
5321 static void free_pmu_context(struct pmu *pmu)
5322 {
5323         struct pmu *i;
5324
5325         mutex_lock(&pmus_lock);
5326         /*
5327          * Like a real lame refcount.
5328          */
5329         list_for_each_entry(i, &pmus, entry) {
5330                 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
5331                         update_pmu_context(i, pmu);
5332                         goto out;
5333                 }
5334         }
5335
5336         free_percpu(pmu->pmu_cpu_context);
5337 out:
5338         mutex_unlock(&pmus_lock);
5339 }
5340 static struct idr pmu_idr;
5341
5342 static ssize_t
5343 type_show(struct device *dev, struct device_attribute *attr, char *page)
5344 {
5345         struct pmu *pmu = dev_get_drvdata(dev);
5346
5347         return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
5348 }
5349
5350 static struct device_attribute pmu_dev_attrs[] = {
5351        __ATTR_RO(type),
5352        __ATTR_NULL,
5353 };
5354
5355 static int pmu_bus_running;
5356 static struct bus_type pmu_bus = {
5357         .name           = "event_source",
5358         .dev_attrs      = pmu_dev_attrs,
5359 };
5360
5361 static void pmu_dev_release(struct device *dev)
5362 {
5363         kfree(dev);
5364 }
5365
5366 static int pmu_dev_alloc(struct pmu *pmu)
5367 {
5368         int ret = -ENOMEM;
5369
5370         pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
5371         if (!pmu->dev)
5372                 goto out;
5373
5374         device_initialize(pmu->dev);
5375         ret = dev_set_name(pmu->dev, "%s", pmu->name);
5376         if (ret)
5377                 goto free_dev;
5378
5379         dev_set_drvdata(pmu->dev, pmu);
5380         pmu->dev->bus = &pmu_bus;
5381         pmu->dev->release = pmu_dev_release;
5382         ret = device_add(pmu->dev);
5383         if (ret)
5384                 goto free_dev;
5385
5386 out:
5387         return ret;
5388
5389 free_dev:
5390         put_device(pmu->dev);
5391         goto out;
5392 }
5393
5394 static struct lock_class_key cpuctx_mutex;
5395
5396 int perf_pmu_register(struct pmu *pmu, char *name, int type)
5397 {
5398         int cpu, ret;
5399
5400         mutex_lock(&pmus_lock);
5401         ret = -ENOMEM;
5402         pmu->pmu_disable_count = alloc_percpu(int);
5403         if (!pmu->pmu_disable_count)
5404                 goto unlock;
5405
5406         pmu->type = -1;
5407         if (!name)
5408                 goto skip_type;
5409         pmu->name = name;
5410
5411         if (type < 0) {
5412                 int err = idr_pre_get(&pmu_idr, GFP_KERNEL);
5413                 if (!err)
5414                         goto free_pdc;
5415
5416                 err = idr_get_new_above(&pmu_idr, pmu, PERF_TYPE_MAX, &type);
5417                 if (err) {
5418                         ret = err;
5419                         goto free_pdc;
5420                 }
5421         }
5422         pmu->type = type;
5423
5424         if (pmu_bus_running) {
5425                 ret = pmu_dev_alloc(pmu);
5426                 if (ret)
5427                         goto free_idr;
5428         }
5429
5430 skip_type:
5431         pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
5432         if (pmu->pmu_cpu_context)
5433                 goto got_cpu_context;
5434
5435         pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
5436         if (!pmu->pmu_cpu_context)
5437                 goto free_dev;
5438
5439         for_each_possible_cpu(cpu) {
5440                 struct perf_cpu_context *cpuctx;
5441
5442                 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5443                 __perf_event_init_context(&cpuctx->ctx);
5444                 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
5445                 cpuctx->ctx.type = cpu_context;
5446                 cpuctx->ctx.pmu = pmu;
5447                 cpuctx->jiffies_interval = 1;
5448                 INIT_LIST_HEAD(&cpuctx->rotation_list);
5449                 cpuctx->active_pmu = pmu;
5450         }
5451
5452 got_cpu_context:
5453         if (!pmu->start_txn) {
5454                 if (pmu->pmu_enable) {
5455                         /*
5456                          * If we have pmu_enable/pmu_disable calls, install
5457                          * transaction stubs that use that to try and batch
5458                          * hardware accesses.
5459                          */
5460                         pmu->start_txn  = perf_pmu_start_txn;
5461                         pmu->commit_txn = perf_pmu_commit_txn;
5462                         pmu->cancel_txn = perf_pmu_cancel_txn;
5463                 } else {
5464                         pmu->start_txn  = perf_pmu_nop_void;
5465                         pmu->commit_txn = perf_pmu_nop_int;
5466                         pmu->cancel_txn = perf_pmu_nop_void;
5467                 }
5468         }
5469
5470         if (!pmu->pmu_enable) {
5471                 pmu->pmu_enable  = perf_pmu_nop_void;
5472                 pmu->pmu_disable = perf_pmu_nop_void;
5473         }
5474
5475         list_add_rcu(&pmu->entry, &pmus);
5476         ret = 0;
5477 unlock:
5478         mutex_unlock(&pmus_lock);
5479
5480         return ret;
5481
5482 free_dev:
5483         device_del(pmu->dev);
5484         put_device(pmu->dev);
5485
5486 free_idr:
5487         if (pmu->type >= PERF_TYPE_MAX)
5488                 idr_remove(&pmu_idr, pmu->type);
5489
5490 free_pdc:
5491         free_percpu(pmu->pmu_disable_count);
5492         goto unlock;
5493 }
5494
5495 void perf_pmu_unregister(struct pmu *pmu)
5496 {
5497         mutex_lock(&pmus_lock);
5498         list_del_rcu(&pmu->entry);
5499         mutex_unlock(&pmus_lock);
5500
5501         /*
5502          * We dereference the pmu list under both SRCU and regular RCU, so
5503          * synchronize against both of those.
5504          */
5505         synchronize_srcu(&pmus_srcu);
5506         synchronize_rcu();
5507
5508         free_percpu(pmu->pmu_disable_count);
5509         if (pmu->type >= PERF_TYPE_MAX)
5510                 idr_remove(&pmu_idr, pmu->type);
5511         device_del(pmu->dev);
5512         put_device(pmu->dev);
5513         free_pmu_context(pmu);
5514 }
5515
5516 struct pmu *perf_init_event(struct perf_event *event)
5517 {
5518         struct pmu *pmu = NULL;
5519         int idx;
5520
5521         idx = srcu_read_lock(&pmus_srcu);
5522
5523         rcu_read_lock();
5524         pmu = idr_find(&pmu_idr, event->attr.type);
5525         rcu_read_unlock();
5526         if (pmu)
5527                 goto unlock;
5528
5529         list_for_each_entry_rcu(pmu, &pmus, entry) {
5530                 int ret = pmu->event_init(event);
5531                 if (!ret)
5532                         goto unlock;
5533
5534                 if (ret != -ENOENT) {
5535                         pmu = ERR_PTR(ret);
5536                         goto unlock;
5537                 }
5538         }
5539         pmu = ERR_PTR(-ENOENT);
5540 unlock:
5541         srcu_read_unlock(&pmus_srcu, idx);
5542
5543         return pmu;
5544 }
5545
5546 /*
5547  * Allocate and initialize a event structure
5548  */
5549 static struct perf_event *
5550 perf_event_alloc(struct perf_event_attr *attr, int cpu,
5551                  struct task_struct *task,
5552                  struct perf_event *group_leader,
5553                  struct perf_event *parent_event,
5554                  perf_overflow_handler_t overflow_handler)
5555 {
5556         struct pmu *pmu;
5557         struct perf_event *event;
5558         struct hw_perf_event *hwc;
5559         long err;
5560
5561         if ((unsigned)cpu >= nr_cpu_ids) {
5562                 if (!task || cpu != -1)
5563                         return ERR_PTR(-EINVAL);
5564         }
5565
5566         event = kzalloc(sizeof(*event), GFP_KERNEL);
5567         if (!event)
5568                 return ERR_PTR(-ENOMEM);
5569
5570         /*
5571          * Single events are their own group leaders, with an
5572          * empty sibling list:
5573          */
5574         if (!group_leader)
5575                 group_leader = event;
5576
5577         mutex_init(&event->child_mutex);
5578         INIT_LIST_HEAD(&event->child_list);
5579
5580         INIT_LIST_HEAD(&event->group_entry);
5581         INIT_LIST_HEAD(&event->event_entry);
5582         INIT_LIST_HEAD(&event->sibling_list);
5583         init_waitqueue_head(&event->waitq);
5584         init_irq_work(&event->pending, perf_pending_event);
5585
5586         mutex_init(&event->mmap_mutex);
5587
5588         event->cpu              = cpu;
5589         event->attr             = *attr;
5590         event->group_leader     = group_leader;
5591         event->pmu              = NULL;
5592         event->oncpu            = -1;
5593
5594         event->parent           = parent_event;
5595
5596         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
5597         event->id               = atomic64_inc_return(&perf_event_id);
5598
5599         event->state            = PERF_EVENT_STATE_INACTIVE;
5600
5601         if (task) {
5602                 event->attach_state = PERF_ATTACH_TASK;
5603 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5604                 /*
5605                  * hw_breakpoint is a bit difficult here..
5606                  */
5607                 if (attr->type == PERF_TYPE_BREAKPOINT)
5608                         event->hw.bp_target = task;
5609 #endif
5610         }
5611
5612         if (!overflow_handler && parent_event)
5613                 overflow_handler = parent_event->overflow_handler;
5614
5615         event->overflow_handler = overflow_handler;
5616
5617         if (attr->disabled)
5618                 event->state = PERF_EVENT_STATE_OFF;
5619
5620         pmu = NULL;
5621
5622         hwc = &event->hw;
5623         hwc->sample_period = attr->sample_period;
5624         if (attr->freq && attr->sample_freq)
5625                 hwc->sample_period = 1;
5626         hwc->last_period = hwc->sample_period;
5627
5628         local64_set(&hwc->period_left, hwc->sample_period);
5629
5630         /*
5631          * we currently do not support PERF_FORMAT_GROUP on inherited events
5632          */
5633         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
5634                 goto done;
5635
5636         pmu = perf_init_event(event);
5637
5638 done:
5639         err = 0;
5640         if (!pmu)
5641                 err = -EINVAL;
5642         else if (IS_ERR(pmu))
5643                 err = PTR_ERR(pmu);
5644
5645         if (err) {
5646                 if (event->ns)
5647                         put_pid_ns(event->ns);
5648                 kfree(event);
5649                 return ERR_PTR(err);
5650         }
5651
5652         event->pmu = pmu;
5653
5654         if (!event->parent) {
5655                 if (event->attach_state & PERF_ATTACH_TASK)
5656                         jump_label_inc(&perf_task_events);
5657                 if (event->attr.mmap || event->attr.mmap_data)
5658                         atomic_inc(&nr_mmap_events);
5659                 if (event->attr.comm)
5660                         atomic_inc(&nr_comm_events);
5661                 if (event->attr.task)
5662                         atomic_inc(&nr_task_events);
5663                 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
5664                         err = get_callchain_buffers();
5665                         if (err) {
5666                                 free_event(event);
5667                                 return ERR_PTR(err);
5668                         }
5669                 }
5670         }
5671
5672         return event;
5673 }
5674
5675 static int perf_copy_attr(struct perf_event_attr __user *uattr,
5676                           struct perf_event_attr *attr)
5677 {
5678         u32 size;
5679         int ret;
5680
5681         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
5682                 return -EFAULT;
5683
5684         /*
5685          * zero the full structure, so that a short copy will be nice.
5686          */
5687         memset(attr, 0, sizeof(*attr));
5688
5689         ret = get_user(size, &uattr->size);
5690         if (ret)
5691                 return ret;
5692
5693         if (size > PAGE_SIZE)   /* silly large */
5694                 goto err_size;
5695
5696         if (!size)              /* abi compat */
5697                 size = PERF_ATTR_SIZE_VER0;
5698
5699         if (size < PERF_ATTR_SIZE_VER0)
5700                 goto err_size;
5701
5702         /*
5703          * If we're handed a bigger struct than we know of,
5704          * ensure all the unknown bits are 0 - i.e. new
5705          * user-space does not rely on any kernel feature
5706          * extensions we dont know about yet.
5707          */
5708         if (size > sizeof(*attr)) {
5709                 unsigned char __user *addr;
5710                 unsigned char __user *end;
5711                 unsigned char val;
5712
5713                 addr = (void __user *)uattr + sizeof(*attr);
5714                 end  = (void __user *)uattr + size;
5715
5716                 for (; addr < end; addr++) {
5717                         ret = get_user(val, addr);
5718                         if (ret)
5719                                 return ret;
5720                         if (val)
5721                                 goto err_size;
5722                 }
5723                 size = sizeof(*attr);
5724         }
5725
5726         ret = copy_from_user(attr, uattr, size);
5727         if (ret)
5728                 return -EFAULT;
5729
5730         /*
5731          * If the type exists, the corresponding creation will verify
5732          * the attr->config.
5733          */
5734         if (attr->type >= PERF_TYPE_MAX)
5735                 return -EINVAL;
5736
5737         if (attr->__reserved_1)
5738                 return -EINVAL;
5739
5740         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
5741                 return -EINVAL;
5742
5743         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
5744                 return -EINVAL;
5745
5746 out:
5747         return ret;
5748
5749 err_size:
5750         put_user(sizeof(*attr), &uattr->size);
5751         ret = -E2BIG;
5752         goto out;
5753 }
5754
5755 static int
5756 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
5757 {
5758         struct perf_buffer *buffer = NULL, *old_buffer = NULL;
5759         int ret = -EINVAL;
5760
5761         if (!output_event)
5762                 goto set;
5763
5764         /* don't allow circular references */
5765         if (event == output_event)
5766                 goto out;
5767
5768         /*
5769          * Don't allow cross-cpu buffers
5770          */
5771         if (output_event->cpu != event->cpu)
5772                 goto out;
5773
5774         /*
5775          * If its not a per-cpu buffer, it must be the same task.
5776          */
5777         if (output_event->cpu == -1 && output_event->ctx != event->ctx)
5778                 goto out;
5779
5780 set:
5781         mutex_lock(&event->mmap_mutex);
5782         /* Can't redirect output if we've got an active mmap() */
5783         if (atomic_read(&event->mmap_count))
5784                 goto unlock;
5785
5786         if (output_event) {
5787                 /* get the buffer we want to redirect to */
5788                 buffer = perf_buffer_get(output_event);
5789                 if (!buffer)
5790                         goto unlock;
5791         }
5792
5793         old_buffer = event->buffer;
5794         rcu_assign_pointer(event->buffer, buffer);
5795         ret = 0;
5796 unlock:
5797         mutex_unlock(&event->mmap_mutex);
5798
5799         if (old_buffer)
5800                 perf_buffer_put(old_buffer);
5801 out:
5802         return ret;
5803 }
5804
5805 /**
5806  * sys_perf_event_open - open a performance event, associate it to a task/cpu
5807  *
5808  * @attr_uptr:  event_id type attributes for monitoring/sampling
5809  * @pid:                target pid
5810  * @cpu:                target cpu
5811  * @group_fd:           group leader event fd
5812  */
5813 SYSCALL_DEFINE5(perf_event_open,
5814                 struct perf_event_attr __user *, attr_uptr,
5815                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
5816 {
5817         struct perf_event *group_leader = NULL, *output_event = NULL;
5818         struct perf_event *event, *sibling;
5819         struct perf_event_attr attr;
5820         struct perf_event_context *ctx;
5821         struct file *event_file = NULL;
5822         struct file *group_file = NULL;
5823         struct task_struct *task = NULL;
5824         struct pmu *pmu;
5825         int event_fd;
5826         int move_group = 0;
5827         int fput_needed = 0;
5828         int err;
5829
5830         /* for future expandability... */
5831         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
5832                 return -EINVAL;
5833
5834         err = perf_copy_attr(attr_uptr, &attr);
5835         if (err)
5836                 return err;
5837
5838         if (!attr.exclude_kernel) {
5839                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
5840                         return -EACCES;
5841         }
5842
5843         if (attr.freq) {
5844                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
5845                         return -EINVAL;
5846         }
5847
5848         event_fd = get_unused_fd_flags(O_RDWR);
5849         if (event_fd < 0)
5850                 return event_fd;
5851
5852         if (group_fd != -1) {
5853                 group_leader = perf_fget_light(group_fd, &fput_needed);
5854                 if (IS_ERR(group_leader)) {
5855                         err = PTR_ERR(group_leader);
5856                         goto err_fd;
5857                 }
5858                 group_file = group_leader->filp;
5859                 if (flags & PERF_FLAG_FD_OUTPUT)
5860                         output_event = group_leader;
5861                 if (flags & PERF_FLAG_FD_NO_GROUP)
5862                         group_leader = NULL;
5863         }
5864
5865         if (pid != -1) {
5866                 task = find_lively_task_by_vpid(pid);
5867                 if (IS_ERR(task)) {
5868                         err = PTR_ERR(task);
5869                         goto err_group_fd;
5870                 }
5871         }
5872
5873         event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, NULL);
5874         if (IS_ERR(event)) {
5875                 err = PTR_ERR(event);
5876                 goto err_task;
5877         }
5878
5879         /*
5880          * Special case software events and allow them to be part of
5881          * any hardware group.
5882          */
5883         pmu = event->pmu;
5884
5885         if (group_leader &&
5886             (is_software_event(event) != is_software_event(group_leader))) {
5887                 if (is_software_event(event)) {
5888                         /*
5889                          * If event and group_leader are not both a software
5890                          * event, and event is, then group leader is not.
5891                          *
5892                          * Allow the addition of software events to !software
5893                          * groups, this is safe because software events never
5894                          * fail to schedule.
5895                          */
5896                         pmu = group_leader->pmu;
5897                 } else if (is_software_event(group_leader) &&
5898                            (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
5899                         /*
5900                          * In case the group is a pure software group, and we
5901                          * try to add a hardware event, move the whole group to
5902                          * the hardware context.
5903                          */
5904                         move_group = 1;
5905                 }
5906         }
5907
5908         /*
5909          * Get the target context (task or percpu):
5910          */
5911         ctx = find_get_context(pmu, task, cpu);
5912         if (IS_ERR(ctx)) {
5913                 err = PTR_ERR(ctx);
5914                 goto err_alloc;
5915         }
5916
5917         /*
5918          * Look up the group leader (we will attach this event to it):
5919          */
5920         if (group_leader) {
5921                 err = -EINVAL;
5922
5923                 /*
5924                  * Do not allow a recursive hierarchy (this new sibling
5925                  * becoming part of another group-sibling):
5926                  */
5927                 if (group_leader->group_leader != group_leader)
5928                         goto err_context;
5929                 /*
5930                  * Do not allow to attach to a group in a different
5931                  * task or CPU context:
5932                  */
5933                 if (move_group) {
5934                         if (group_leader->ctx->type != ctx->type)
5935                                 goto err_context;
5936                 } else {
5937                         if (group_leader->ctx != ctx)
5938                                 goto err_context;
5939                 }
5940
5941                 /*
5942                  * Only a group leader can be exclusive or pinned
5943                  */
5944                 if (attr.exclusive || attr.pinned)
5945                         goto err_context;
5946         }
5947
5948         if (output_event) {
5949                 err = perf_event_set_output(event, output_event);
5950                 if (err)
5951                         goto err_context;
5952         }
5953
5954         event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
5955         if (IS_ERR(event_file)) {
5956                 err = PTR_ERR(event_file);
5957                 goto err_context;
5958         }
5959
5960         if (move_group) {
5961                 struct perf_event_context *gctx = group_leader->ctx;
5962
5963                 mutex_lock(&gctx->mutex);
5964                 perf_event_remove_from_context(group_leader);
5965                 list_for_each_entry(sibling, &group_leader->sibling_list,
5966                                     group_entry) {
5967                         perf_event_remove_from_context(sibling);
5968                         put_ctx(gctx);
5969                 }
5970                 mutex_unlock(&gctx->mutex);
5971                 put_ctx(gctx);
5972         }
5973
5974         event->filp = event_file;
5975         WARN_ON_ONCE(ctx->parent_ctx);
5976         mutex_lock(&ctx->mutex);
5977
5978         if (move_group) {
5979                 perf_install_in_context(ctx, group_leader, cpu);
5980                 get_ctx(ctx);
5981                 list_for_each_entry(sibling, &group_leader->sibling_list,
5982                                     group_entry) {
5983                         perf_install_in_context(ctx, sibling, cpu);
5984                         get_ctx(ctx);
5985                 }
5986         }
5987
5988         perf_install_in_context(ctx, event, cpu);
5989         ++ctx->generation;
5990         mutex_unlock(&ctx->mutex);
5991
5992         event->owner = current;
5993
5994         mutex_lock(&current->perf_event_mutex);
5995         list_add_tail(&event->owner_entry, &current->perf_event_list);
5996         mutex_unlock(&current->perf_event_mutex);
5997
5998         /*
5999          * Precalculate sample_data sizes
6000          */
6001         perf_event__header_size(event);
6002         perf_event__id_header_size(event);
6003
6004         /*
6005          * Drop the reference on the group_event after placing the
6006          * new event on the sibling_list. This ensures destruction
6007          * of the group leader will find the pointer to itself in
6008          * perf_group_detach().
6009          */
6010         fput_light(group_file, fput_needed);
6011         fd_install(event_fd, event_file);
6012         return event_fd;
6013
6014 err_context:
6015         put_ctx(ctx);
6016 err_alloc:
6017         free_event(event);
6018 err_task:
6019         if (task)
6020                 put_task_struct(task);
6021 err_group_fd:
6022         fput_light(group_file, fput_needed);
6023 err_fd:
6024         put_unused_fd(event_fd);
6025         return err;
6026 }
6027
6028 /**
6029  * perf_event_create_kernel_counter
6030  *
6031  * @attr: attributes of the counter to create
6032  * @cpu: cpu in which the counter is bound
6033  * @task: task to profile (NULL for percpu)
6034  */
6035 struct perf_event *
6036 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
6037                                  struct task_struct *task,
6038                                  perf_overflow_handler_t overflow_handler)
6039 {
6040         struct perf_event_context *ctx;
6041         struct perf_event *event;
6042         int err;
6043
6044         /*
6045          * Get the target context (task or percpu):
6046          */
6047
6048         event = perf_event_alloc(attr, cpu, task, NULL, NULL, overflow_handler);
6049         if (IS_ERR(event)) {
6050                 err = PTR_ERR(event);
6051                 goto err;
6052         }
6053
6054         ctx = find_get_context(event->pmu, task, cpu);
6055         if (IS_ERR(ctx)) {
6056                 err = PTR_ERR(ctx);
6057                 goto err_free;
6058         }
6059
6060         event->filp = NULL;
6061         WARN_ON_ONCE(ctx->parent_ctx);
6062         mutex_lock(&ctx->mutex);
6063         perf_install_in_context(ctx, event, cpu);
6064         ++ctx->generation;
6065         mutex_unlock(&ctx->mutex);
6066
6067         return event;
6068
6069 err_free:
6070         free_event(event);
6071 err:
6072         return ERR_PTR(err);
6073 }
6074 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
6075
6076 static void sync_child_event(struct perf_event *child_event,
6077                                struct task_struct *child)
6078 {
6079         struct perf_event *parent_event = child_event->parent;
6080         u64 child_val;
6081
6082         if (child_event->attr.inherit_stat)
6083                 perf_event_read_event(child_event, child);
6084
6085         child_val = perf_event_count(child_event);
6086
6087         /*
6088          * Add back the child's count to the parent's count:
6089          */
6090         atomic64_add(child_val, &parent_event->child_count);
6091         atomic64_add(child_event->total_time_enabled,
6092                      &parent_event->child_total_time_enabled);
6093         atomic64_add(child_event->total_time_running,
6094                      &parent_event->child_total_time_running);
6095
6096         /*
6097          * Remove this event from the parent's list
6098          */
6099         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6100         mutex_lock(&parent_event->child_mutex);
6101         list_del_init(&child_event->child_list);
6102         mutex_unlock(&parent_event->child_mutex);
6103
6104         /*
6105          * Release the parent event, if this was the last
6106          * reference to it.
6107          */
6108         fput(parent_event->filp);
6109 }
6110
6111 static void
6112 __perf_event_exit_task(struct perf_event *child_event,
6113                          struct perf_event_context *child_ctx,
6114                          struct task_struct *child)
6115 {
6116         struct perf_event *parent_event;
6117
6118         perf_event_remove_from_context(child_event);
6119
6120         parent_event = child_event->parent;
6121         /*
6122          * It can happen that parent exits first, and has events
6123          * that are still around due to the child reference. These
6124          * events need to be zapped - but otherwise linger.
6125          */
6126         if (parent_event) {
6127                 sync_child_event(child_event, child);
6128                 free_event(child_event);
6129         }
6130 }
6131
6132 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
6133 {
6134         struct perf_event *child_event, *tmp;
6135         struct perf_event_context *child_ctx;
6136         unsigned long flags;
6137
6138         if (likely(!child->perf_event_ctxp[ctxn])) {
6139                 perf_event_task(child, NULL, 0);
6140                 return;
6141         }
6142
6143         local_irq_save(flags);
6144         /*
6145          * We can't reschedule here because interrupts are disabled,
6146          * and either child is current or it is a task that can't be
6147          * scheduled, so we are now safe from rescheduling changing
6148          * our context.
6149          */
6150         child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
6151         task_ctx_sched_out(child_ctx, EVENT_ALL);
6152
6153         /*
6154          * Take the context lock here so that if find_get_context is
6155          * reading child->perf_event_ctxp, we wait until it has
6156          * incremented the context's refcount before we do put_ctx below.
6157          */
6158         raw_spin_lock(&child_ctx->lock);
6159         child->perf_event_ctxp[ctxn] = NULL;
6160         /*
6161          * If this context is a clone; unclone it so it can't get
6162          * swapped to another process while we're removing all
6163          * the events from it.
6164          */
6165         unclone_ctx(child_ctx);
6166         update_context_time(child_ctx);
6167         raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6168
6169         /*
6170          * Report the task dead after unscheduling the events so that we
6171          * won't get any samples after PERF_RECORD_EXIT. We can however still
6172          * get a few PERF_RECORD_READ events.
6173          */
6174         perf_event_task(child, child_ctx, 0);
6175
6176         /*
6177          * We can recurse on the same lock type through:
6178          *
6179          *   __perf_event_exit_task()
6180          *     sync_child_event()
6181          *       fput(parent_event->filp)
6182          *         perf_release()
6183          *           mutex_lock(&ctx->mutex)
6184          *
6185          * But since its the parent context it won't be the same instance.
6186          */
6187         mutex_lock(&child_ctx->mutex);
6188
6189 again:
6190         list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
6191                                  group_entry)
6192                 __perf_event_exit_task(child_event, child_ctx, child);
6193
6194         list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
6195                                  group_entry)
6196                 __perf_event_exit_task(child_event, child_ctx, child);
6197
6198         /*
6199          * If the last event was a group event, it will have appended all
6200          * its siblings to the list, but we obtained 'tmp' before that which
6201          * will still point to the list head terminating the iteration.
6202          */
6203         if (!list_empty(&child_ctx->pinned_groups) ||
6204             !list_empty(&child_ctx->flexible_groups))
6205                 goto again;
6206
6207         mutex_unlock(&child_ctx->mutex);
6208
6209         put_ctx(child_ctx);
6210 }
6211
6212 /*
6213  * When a child task exits, feed back event values to parent events.
6214  */
6215 void perf_event_exit_task(struct task_struct *child)
6216 {
6217         struct perf_event *event, *tmp;
6218         int ctxn;
6219
6220         mutex_lock(&child->perf_event_mutex);
6221         list_for_each_entry_safe(event, tmp, &child->perf_event_list,
6222                                  owner_entry) {
6223                 list_del_init(&event->owner_entry);
6224
6225                 /*
6226                  * Ensure the list deletion is visible before we clear
6227                  * the owner, closes a race against perf_release() where
6228                  * we need to serialize on the owner->perf_event_mutex.
6229                  */
6230                 smp_wmb();
6231                 event->owner = NULL;
6232         }
6233         mutex_unlock(&child->perf_event_mutex);
6234
6235         for_each_task_context_nr(ctxn)
6236                 perf_event_exit_task_context(child, ctxn);
6237 }
6238
6239 static void perf_free_event(struct perf_event *event,
6240                             struct perf_event_context *ctx)
6241 {
6242         struct perf_event *parent = event->parent;
6243
6244         if (WARN_ON_ONCE(!parent))
6245                 return;
6246
6247         mutex_lock(&parent->child_mutex);
6248         list_del_init(&event->child_list);
6249         mutex_unlock(&parent->child_mutex);
6250
6251         fput(parent->filp);
6252
6253         perf_group_detach(event);
6254         list_del_event(event, ctx);
6255         free_event(event);
6256 }
6257
6258 /*
6259  * free an unexposed, unused context as created by inheritance by
6260  * perf_event_init_task below, used by fork() in case of fail.
6261  */
6262 void perf_event_free_task(struct task_struct *task)
6263 {
6264         struct perf_event_context *ctx;
6265         struct perf_event *event, *tmp;
6266         int ctxn;
6267
6268         for_each_task_context_nr(ctxn) {
6269                 ctx = task->perf_event_ctxp[ctxn];
6270                 if (!ctx)
6271                         continue;
6272
6273                 mutex_lock(&ctx->mutex);
6274 again:
6275                 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
6276                                 group_entry)
6277                         perf_free_event(event, ctx);
6278
6279                 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
6280                                 group_entry)
6281                         perf_free_event(event, ctx);
6282
6283                 if (!list_empty(&ctx->pinned_groups) ||
6284                                 !list_empty(&ctx->flexible_groups))
6285                         goto again;
6286
6287                 mutex_unlock(&ctx->mutex);
6288
6289                 put_ctx(ctx);
6290         }
6291 }
6292
6293 void perf_event_delayed_put(struct task_struct *task)
6294 {
6295         int ctxn;
6296
6297         for_each_task_context_nr(ctxn)
6298                 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
6299 }
6300
6301 /*
6302  * inherit a event from parent task to child task:
6303  */
6304 static struct perf_event *
6305 inherit_event(struct perf_event *parent_event,
6306               struct task_struct *parent,
6307               struct perf_event_context *parent_ctx,
6308               struct task_struct *child,
6309               struct perf_event *group_leader,
6310               struct perf_event_context *child_ctx)
6311 {
6312         struct perf_event *child_event;
6313         unsigned long flags;
6314
6315         /*
6316          * Instead of creating recursive hierarchies of events,
6317          * we link inherited events back to the original parent,
6318          * which has a filp for sure, which we use as the reference
6319          * count:
6320          */
6321         if (parent_event->parent)
6322                 parent_event = parent_event->parent;
6323
6324         child_event = perf_event_alloc(&parent_event->attr,
6325                                            parent_event->cpu,
6326                                            child,
6327                                            group_leader, parent_event,
6328                                            NULL);
6329         if (IS_ERR(child_event))
6330                 return child_event;
6331         get_ctx(child_ctx);
6332
6333         /*
6334          * Make the child state follow the state of the parent event,
6335          * not its attr.disabled bit.  We hold the parent's mutex,
6336          * so we won't race with perf_event_{en, dis}able_family.
6337          */
6338         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
6339                 child_event->state = PERF_EVENT_STATE_INACTIVE;
6340         else
6341                 child_event->state = PERF_EVENT_STATE_OFF;
6342
6343         if (parent_event->attr.freq) {
6344                 u64 sample_period = parent_event->hw.sample_period;
6345                 struct hw_perf_event *hwc = &child_event->hw;
6346
6347                 hwc->sample_period = sample_period;
6348                 hwc->last_period   = sample_period;
6349
6350                 local64_set(&hwc->period_left, sample_period);
6351         }
6352
6353         child_event->ctx = child_ctx;
6354         child_event->overflow_handler = parent_event->overflow_handler;
6355
6356         /*
6357          * Precalculate sample_data sizes
6358          */
6359         perf_event__header_size(child_event);
6360         perf_event__id_header_size(child_event);
6361
6362         /*
6363          * Link it up in the child's context:
6364          */
6365         raw_spin_lock_irqsave(&child_ctx->lock, flags);
6366         add_event_to_ctx(child_event, child_ctx);
6367         raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6368
6369         /*
6370          * Get a reference to the parent filp - we will fput it
6371          * when the child event exits. This is safe to do because
6372          * we are in the parent and we know that the filp still
6373          * exists and has a nonzero count:
6374          */
6375         atomic_long_inc(&parent_event->filp->f_count);
6376
6377         /*
6378          * Link this into the parent event's child list
6379          */
6380         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6381         mutex_lock(&parent_event->child_mutex);
6382         list_add_tail(&child_event->child_list, &parent_event->child_list);
6383         mutex_unlock(&parent_event->child_mutex);
6384
6385         return child_event;
6386 }
6387
6388 static int inherit_group(struct perf_event *parent_event,
6389               struct task_struct *parent,
6390               struct perf_event_context *parent_ctx,
6391               struct task_struct *child,
6392               struct perf_event_context *child_ctx)
6393 {
6394         struct perf_event *leader;
6395         struct perf_event *sub;
6396         struct perf_event *child_ctr;
6397
6398         leader = inherit_event(parent_event, parent, parent_ctx,
6399                                  child, NULL, child_ctx);
6400         if (IS_ERR(leader))
6401                 return PTR_ERR(leader);
6402         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
6403                 child_ctr = inherit_event(sub, parent, parent_ctx,
6404                                             child, leader, child_ctx);
6405                 if (IS_ERR(child_ctr))
6406                         return PTR_ERR(child_ctr);
6407         }
6408         return 0;
6409 }
6410
6411 static int
6412 inherit_task_group(struct perf_event *event, struct task_struct *parent,
6413                    struct perf_event_context *parent_ctx,
6414                    struct task_struct *child, int ctxn,
6415                    int *inherited_all)
6416 {
6417         int ret;
6418         struct perf_event_context *child_ctx;
6419
6420         if (!event->attr.inherit) {
6421                 *inherited_all = 0;
6422                 return 0;
6423         }
6424
6425         child_ctx = child->perf_event_ctxp[ctxn];
6426         if (!child_ctx) {
6427                 /*
6428                  * This is executed from the parent task context, so
6429                  * inherit events that have been marked for cloning.
6430                  * First allocate and initialize a context for the
6431                  * child.
6432                  */
6433
6434                 child_ctx = alloc_perf_context(event->pmu, child);
6435                 if (!child_ctx)
6436                         return -ENOMEM;
6437
6438                 child->perf_event_ctxp[ctxn] = child_ctx;
6439         }
6440
6441         ret = inherit_group(event, parent, parent_ctx,
6442                             child, child_ctx);
6443
6444         if (ret)
6445                 *inherited_all = 0;
6446
6447         return ret;
6448 }
6449
6450 /*
6451  * Initialize the perf_event context in task_struct
6452  */
6453 int perf_event_init_context(struct task_struct *child, int ctxn)
6454 {
6455         struct perf_event_context *child_ctx, *parent_ctx;
6456         struct perf_event_context *cloned_ctx;
6457         struct perf_event *event;
6458         struct task_struct *parent = current;
6459         int inherited_all = 1;
6460         unsigned long flags;
6461         int ret = 0;
6462
6463         if (likely(!parent->perf_event_ctxp[ctxn]))
6464                 return 0;
6465
6466         /*
6467          * If the parent's context is a clone, pin it so it won't get
6468          * swapped under us.
6469          */
6470         parent_ctx = perf_pin_task_context(parent, ctxn);
6471
6472         /*
6473          * No need to check if parent_ctx != NULL here; since we saw
6474          * it non-NULL earlier, the only reason for it to become NULL
6475          * is if we exit, and since we're currently in the middle of
6476          * a fork we can't be exiting at the same time.
6477          */
6478
6479         /*
6480          * Lock the parent list. No need to lock the child - not PID
6481          * hashed yet and not running, so nobody can access it.
6482          */
6483         mutex_lock(&parent_ctx->mutex);
6484
6485         /*
6486          * We dont have to disable NMIs - we are only looking at
6487          * the list, not manipulating it:
6488          */
6489         list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
6490                 ret = inherit_task_group(event, parent, parent_ctx,
6491                                          child, ctxn, &inherited_all);
6492                 if (ret)
6493                         break;
6494         }
6495
6496         /*
6497          * We can't hold ctx->lock when iterating the ->flexible_group list due
6498          * to allocations, but we need to prevent rotation because
6499          * rotate_ctx() will change the list from interrupt context.
6500          */
6501         raw_spin_lock_irqsave(&parent_ctx->lock, flags);
6502         parent_ctx->rotate_disable = 1;
6503         raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
6504
6505         list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
6506                 ret = inherit_task_group(event, parent, parent_ctx,
6507                                          child, ctxn, &inherited_all);
6508                 if (ret)
6509                         break;
6510         }
6511
6512         raw_spin_lock_irqsave(&parent_ctx->lock, flags);
6513         parent_ctx->rotate_disable = 0;
6514
6515         child_ctx = child->perf_event_ctxp[ctxn];
6516
6517         if (child_ctx && inherited_all) {
6518                 /*
6519                  * Mark the child context as a clone of the parent
6520                  * context, or of whatever the parent is a clone of.
6521                  *
6522                  * Note that if the parent is a clone, the holding of
6523                  * parent_ctx->lock avoids it from being uncloned.
6524                  */
6525                 cloned_ctx = parent_ctx->parent_ctx;
6526                 if (cloned_ctx) {
6527                         child_ctx->parent_ctx = cloned_ctx;
6528                         child_ctx->parent_gen = parent_ctx->parent_gen;
6529                 } else {
6530                         child_ctx->parent_ctx = parent_ctx;
6531                         child_ctx->parent_gen = parent_ctx->generation;
6532                 }
6533                 get_ctx(child_ctx->parent_ctx);
6534         }
6535
6536         raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
6537         mutex_unlock(&parent_ctx->mutex);
6538
6539         perf_unpin_context(parent_ctx);
6540
6541         return ret;
6542 }
6543
6544 /*
6545  * Initialize the perf_event context in task_struct
6546  */
6547 int perf_event_init_task(struct task_struct *child)
6548 {
6549         int ctxn, ret;
6550
6551         memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
6552         mutex_init(&child->perf_event_mutex);
6553         INIT_LIST_HEAD(&child->perf_event_list);
6554
6555         for_each_task_context_nr(ctxn) {
6556                 ret = perf_event_init_context(child, ctxn);
6557                 if (ret)
6558                         return ret;
6559         }
6560
6561         return 0;
6562 }
6563
6564 static void __init perf_event_init_all_cpus(void)
6565 {
6566         struct swevent_htable *swhash;
6567         int cpu;
6568
6569         for_each_possible_cpu(cpu) {
6570                 swhash = &per_cpu(swevent_htable, cpu);
6571                 mutex_init(&swhash->hlist_mutex);
6572                 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
6573         }
6574 }
6575
6576 static void __cpuinit perf_event_init_cpu(int cpu)
6577 {
6578         struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6579
6580         mutex_lock(&swhash->hlist_mutex);
6581         if (swhash->hlist_refcount > 0) {
6582                 struct swevent_hlist *hlist;
6583
6584                 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
6585                 WARN_ON(!hlist);
6586                 rcu_assign_pointer(swhash->swevent_hlist, hlist);
6587         }
6588         mutex_unlock(&swhash->hlist_mutex);
6589 }
6590
6591 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
6592 static void perf_pmu_rotate_stop(struct pmu *pmu)
6593 {
6594         struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
6595
6596         WARN_ON(!irqs_disabled());
6597
6598         list_del_init(&cpuctx->rotation_list);
6599 }
6600
6601 static void __perf_event_exit_context(void *__info)
6602 {
6603         struct perf_event_context *ctx = __info;
6604         struct perf_event *event, *tmp;
6605
6606         perf_pmu_rotate_stop(ctx->pmu);
6607
6608         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
6609                 __perf_event_remove_from_context(event);
6610         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
6611                 __perf_event_remove_from_context(event);
6612 }
6613
6614 static void perf_event_exit_cpu_context(int cpu)
6615 {
6616         struct perf_event_context *ctx;
6617         struct pmu *pmu;
6618         int idx;
6619
6620         idx = srcu_read_lock(&pmus_srcu);
6621         list_for_each_entry_rcu(pmu, &pmus, entry) {
6622                 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
6623
6624                 mutex_lock(&ctx->mutex);
6625                 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
6626                 mutex_unlock(&ctx->mutex);
6627         }
6628         srcu_read_unlock(&pmus_srcu, idx);
6629 }
6630
6631 static void perf_event_exit_cpu(int cpu)
6632 {
6633         struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6634
6635         mutex_lock(&swhash->hlist_mutex);
6636         swevent_hlist_release(swhash);
6637         mutex_unlock(&swhash->hlist_mutex);
6638
6639         perf_event_exit_cpu_context(cpu);
6640 }
6641 #else
6642 static inline void perf_event_exit_cpu(int cpu) { }
6643 #endif
6644
6645 static int
6646 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
6647 {
6648         int cpu;
6649
6650         for_each_online_cpu(cpu)
6651                 perf_event_exit_cpu(cpu);
6652
6653         return NOTIFY_OK;
6654 }
6655
6656 /*
6657  * Run the perf reboot notifier at the very last possible moment so that
6658  * the generic watchdog code runs as long as possible.
6659  */
6660 static struct notifier_block perf_reboot_notifier = {
6661         .notifier_call = perf_reboot,
6662         .priority = INT_MIN,
6663 };
6664
6665 static int __cpuinit
6666 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
6667 {
6668         unsigned int cpu = (long)hcpu;
6669
6670         switch (action & ~CPU_TASKS_FROZEN) {
6671
6672         case CPU_UP_PREPARE:
6673         case CPU_DOWN_FAILED:
6674                 perf_event_init_cpu(cpu);
6675                 break;
6676
6677         case CPU_UP_CANCELED:
6678         case CPU_DOWN_PREPARE:
6679                 perf_event_exit_cpu(cpu);
6680                 break;
6681
6682         default:
6683                 break;
6684         }
6685
6686         return NOTIFY_OK;
6687 }
6688
6689 void __init perf_event_init(void)
6690 {
6691         int ret;
6692
6693         idr_init(&pmu_idr);
6694
6695         perf_event_init_all_cpus();
6696         init_srcu_struct(&pmus_srcu);
6697         perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
6698         perf_pmu_register(&perf_cpu_clock, NULL, -1);
6699         perf_pmu_register(&perf_task_clock, NULL, -1);
6700         perf_tp_register();
6701         perf_cpu_notifier(perf_cpu_notify);
6702         register_reboot_notifier(&perf_reboot_notifier);
6703
6704         ret = init_hw_breakpoint();
6705         WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
6706 }
6707
6708 static int __init perf_event_sysfs_init(void)
6709 {
6710         struct pmu *pmu;
6711         int ret;
6712
6713         mutex_lock(&pmus_lock);
6714
6715         ret = bus_register(&pmu_bus);
6716         if (ret)
6717                 goto unlock;
6718
6719         list_for_each_entry(pmu, &pmus, entry) {
6720                 if (!pmu->name || pmu->type < 0)
6721                         continue;
6722
6723                 ret = pmu_dev_alloc(pmu);
6724                 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
6725         }
6726         pmu_bus_running = 1;
6727         ret = 0;
6728
6729 unlock:
6730         mutex_unlock(&pmus_lock);
6731
6732         return ret;
6733 }
6734 device_initcall(perf_event_sysfs_init);