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