perf_events: Disable events when we detach them
[linux-2.6.git] / kernel / perf_event.c
1 /*
2  * Performance events core code:
3  *
4  *  Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5  *  Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6  *  Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7  *  Copyright  ©  2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
8  *
9  * For licensing details see kernel-base/COPYING
10  */
11
12 #include <linux/fs.h>
13 #include <linux/mm.h>
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/file.h>
17 #include <linux/poll.h>
18 #include <linux/sysfs.h>
19 #include <linux/dcache.h>
20 #include <linux/percpu.h>
21 #include <linux/ptrace.h>
22 #include <linux/vmstat.h>
23 #include <linux/vmalloc.h>
24 #include <linux/hardirq.h>
25 #include <linux/rculist.h>
26 #include <linux/uaccess.h>
27 #include <linux/syscalls.h>
28 #include <linux/anon_inodes.h>
29 #include <linux/kernel_stat.h>
30 #include <linux/perf_event.h>
31 #include <linux/ftrace_event.h>
32 #include <linux/hw_breakpoint.h>
33
34 #include <asm/irq_regs.h>
35
36 /*
37  * Each CPU has a list of per CPU events:
38  */
39 DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
40
41 int perf_max_events __read_mostly = 1;
42 static int perf_reserved_percpu __read_mostly;
43 static int perf_overcommit __read_mostly = 1;
44
45 static atomic_t nr_events __read_mostly;
46 static atomic_t nr_mmap_events __read_mostly;
47 static atomic_t nr_comm_events __read_mostly;
48 static atomic_t nr_task_events __read_mostly;
49
50 /*
51  * perf event paranoia level:
52  *  -1 - not paranoid at all
53  *   0 - disallow raw tracepoint access for unpriv
54  *   1 - disallow cpu events for unpriv
55  *   2 - disallow kernel profiling for unpriv
56  */
57 int sysctl_perf_event_paranoid __read_mostly = 1;
58
59 static inline bool perf_paranoid_tracepoint_raw(void)
60 {
61         return sysctl_perf_event_paranoid > -1;
62 }
63
64 static inline bool perf_paranoid_cpu(void)
65 {
66         return sysctl_perf_event_paranoid > 0;
67 }
68
69 static inline bool perf_paranoid_kernel(void)
70 {
71         return sysctl_perf_event_paranoid > 1;
72 }
73
74 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
75
76 /*
77  * max perf event sample rate
78  */
79 int sysctl_perf_event_sample_rate __read_mostly = 100000;
80
81 static atomic64_t perf_event_id;
82
83 /*
84  * Lock for (sysadmin-configurable) event reservations:
85  */
86 static DEFINE_SPINLOCK(perf_resource_lock);
87
88 /*
89  * Architecture provided APIs - weak aliases:
90  */
91 extern __weak const struct pmu *hw_perf_event_init(struct perf_event *event)
92 {
93         return NULL;
94 }
95
96 void __weak hw_perf_disable(void)               { barrier(); }
97 void __weak hw_perf_enable(void)                { barrier(); }
98
99 void __weak hw_perf_event_setup(int cpu)        { barrier(); }
100 void __weak hw_perf_event_setup_online(int cpu) { barrier(); }
101
102 int __weak
103 hw_perf_group_sched_in(struct perf_event *group_leader,
104                struct perf_cpu_context *cpuctx,
105                struct perf_event_context *ctx, int cpu)
106 {
107         return 0;
108 }
109
110 void __weak perf_event_print_debug(void)        { }
111
112 static DEFINE_PER_CPU(int, perf_disable_count);
113
114 void __perf_disable(void)
115 {
116         __get_cpu_var(perf_disable_count)++;
117 }
118
119 bool __perf_enable(void)
120 {
121         return !--__get_cpu_var(perf_disable_count);
122 }
123
124 void perf_disable(void)
125 {
126         __perf_disable();
127         hw_perf_disable();
128 }
129
130 void perf_enable(void)
131 {
132         if (__perf_enable())
133                 hw_perf_enable();
134 }
135
136 static void get_ctx(struct perf_event_context *ctx)
137 {
138         WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
139 }
140
141 static void free_ctx(struct rcu_head *head)
142 {
143         struct perf_event_context *ctx;
144
145         ctx = container_of(head, struct perf_event_context, rcu_head);
146         kfree(ctx);
147 }
148
149 static void put_ctx(struct perf_event_context *ctx)
150 {
151         if (atomic_dec_and_test(&ctx->refcount)) {
152                 if (ctx->parent_ctx)
153                         put_ctx(ctx->parent_ctx);
154                 if (ctx->task)
155                         put_task_struct(ctx->task);
156                 call_rcu(&ctx->rcu_head, free_ctx);
157         }
158 }
159
160 static void unclone_ctx(struct perf_event_context *ctx)
161 {
162         if (ctx->parent_ctx) {
163                 put_ctx(ctx->parent_ctx);
164                 ctx->parent_ctx = NULL;
165         }
166 }
167
168 /*
169  * If we inherit events we want to return the parent event id
170  * to userspace.
171  */
172 static u64 primary_event_id(struct perf_event *event)
173 {
174         u64 id = event->id;
175
176         if (event->parent)
177                 id = event->parent->id;
178
179         return id;
180 }
181
182 /*
183  * Get the perf_event_context for a task and lock it.
184  * This has to cope with with the fact that until it is locked,
185  * the context could get moved to another task.
186  */
187 static struct perf_event_context *
188 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
189 {
190         struct perf_event_context *ctx;
191
192         rcu_read_lock();
193  retry:
194         ctx = rcu_dereference(task->perf_event_ctxp);
195         if (ctx) {
196                 /*
197                  * If this context is a clone of another, it might
198                  * get swapped for another underneath us by
199                  * perf_event_task_sched_out, though the
200                  * rcu_read_lock() protects us from any context
201                  * getting freed.  Lock the context and check if it
202                  * got swapped before we could get the lock, and retry
203                  * if so.  If we locked the right context, then it
204                  * can't get swapped on us any more.
205                  */
206                 spin_lock_irqsave(&ctx->lock, *flags);
207                 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
208                         spin_unlock_irqrestore(&ctx->lock, *flags);
209                         goto retry;
210                 }
211
212                 if (!atomic_inc_not_zero(&ctx->refcount)) {
213                         spin_unlock_irqrestore(&ctx->lock, *flags);
214                         ctx = NULL;
215                 }
216         }
217         rcu_read_unlock();
218         return ctx;
219 }
220
221 /*
222  * Get the context for a task and increment its pin_count so it
223  * can't get swapped to another task.  This also increments its
224  * reference count so that the context can't get freed.
225  */
226 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
227 {
228         struct perf_event_context *ctx;
229         unsigned long flags;
230
231         ctx = perf_lock_task_context(task, &flags);
232         if (ctx) {
233                 ++ctx->pin_count;
234                 spin_unlock_irqrestore(&ctx->lock, flags);
235         }
236         return ctx;
237 }
238
239 static void perf_unpin_context(struct perf_event_context *ctx)
240 {
241         unsigned long flags;
242
243         spin_lock_irqsave(&ctx->lock, flags);
244         --ctx->pin_count;
245         spin_unlock_irqrestore(&ctx->lock, flags);
246         put_ctx(ctx);
247 }
248
249 /*
250  * Add a event from the lists for its context.
251  * Must be called with ctx->mutex and ctx->lock held.
252  */
253 static void
254 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
255 {
256         struct perf_event *group_leader = event->group_leader;
257
258         /*
259          * Depending on whether it is a standalone or sibling event,
260          * add it straight to the context's event list, or to the group
261          * leader's sibling list:
262          */
263         if (group_leader == event)
264                 list_add_tail(&event->group_entry, &ctx->group_list);
265         else {
266                 list_add_tail(&event->group_entry, &group_leader->sibling_list);
267                 group_leader->nr_siblings++;
268         }
269
270         list_add_rcu(&event->event_entry, &ctx->event_list);
271         ctx->nr_events++;
272         if (event->attr.inherit_stat)
273                 ctx->nr_stat++;
274 }
275
276 /*
277  * Remove a event from the lists for its context.
278  * Must be called with ctx->mutex and ctx->lock held.
279  */
280 static void
281 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
282 {
283         struct perf_event *sibling, *tmp;
284
285         if (list_empty(&event->group_entry))
286                 return;
287         ctx->nr_events--;
288         if (event->attr.inherit_stat)
289                 ctx->nr_stat--;
290
291         list_del_init(&event->group_entry);
292         list_del_rcu(&event->event_entry);
293
294         if (event->group_leader != event)
295                 event->group_leader->nr_siblings--;
296
297         event->state = PERF_EVENT_STATE_OFF;
298
299         /*
300          * If this was a group event with sibling events then
301          * upgrade the siblings to singleton events by adding them
302          * to the context list directly:
303          */
304         list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
305
306                 list_move_tail(&sibling->group_entry, &ctx->group_list);
307                 sibling->group_leader = sibling;
308         }
309 }
310
311 static void
312 event_sched_out(struct perf_event *event,
313                   struct perf_cpu_context *cpuctx,
314                   struct perf_event_context *ctx)
315 {
316         if (event->state != PERF_EVENT_STATE_ACTIVE)
317                 return;
318
319         event->state = PERF_EVENT_STATE_INACTIVE;
320         if (event->pending_disable) {
321                 event->pending_disable = 0;
322                 event->state = PERF_EVENT_STATE_OFF;
323         }
324         event->tstamp_stopped = ctx->time;
325         event->pmu->disable(event);
326         event->oncpu = -1;
327
328         if (!is_software_event(event))
329                 cpuctx->active_oncpu--;
330         ctx->nr_active--;
331         if (event->attr.exclusive || !cpuctx->active_oncpu)
332                 cpuctx->exclusive = 0;
333 }
334
335 static void
336 group_sched_out(struct perf_event *group_event,
337                 struct perf_cpu_context *cpuctx,
338                 struct perf_event_context *ctx)
339 {
340         struct perf_event *event;
341
342         if (group_event->state != PERF_EVENT_STATE_ACTIVE)
343                 return;
344
345         event_sched_out(group_event, cpuctx, ctx);
346
347         /*
348          * Schedule out siblings (if any):
349          */
350         list_for_each_entry(event, &group_event->sibling_list, group_entry)
351                 event_sched_out(event, cpuctx, ctx);
352
353         if (group_event->attr.exclusive)
354                 cpuctx->exclusive = 0;
355 }
356
357 /*
358  * Cross CPU call to remove a performance event
359  *
360  * We disable the event on the hardware level first. After that we
361  * remove it from the context list.
362  */
363 static void __perf_event_remove_from_context(void *info)
364 {
365         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
366         struct perf_event *event = info;
367         struct perf_event_context *ctx = event->ctx;
368
369         /*
370          * If this is a task context, we need to check whether it is
371          * the current task context of this cpu. If not it has been
372          * scheduled out before the smp call arrived.
373          */
374         if (ctx->task && cpuctx->task_ctx != ctx)
375                 return;
376
377         spin_lock(&ctx->lock);
378         /*
379          * Protect the list operation against NMI by disabling the
380          * events on a global level.
381          */
382         perf_disable();
383
384         event_sched_out(event, cpuctx, ctx);
385
386         list_del_event(event, ctx);
387
388         if (!ctx->task) {
389                 /*
390                  * Allow more per task events with respect to the
391                  * reservation:
392                  */
393                 cpuctx->max_pertask =
394                         min(perf_max_events - ctx->nr_events,
395                             perf_max_events - perf_reserved_percpu);
396         }
397
398         perf_enable();
399         spin_unlock(&ctx->lock);
400 }
401
402
403 /*
404  * Remove the event from a task's (or a CPU's) list of events.
405  *
406  * Must be called with ctx->mutex held.
407  *
408  * CPU events are removed with a smp call. For task events we only
409  * call when the task is on a CPU.
410  *
411  * If event->ctx is a cloned context, callers must make sure that
412  * every task struct that event->ctx->task could possibly point to
413  * remains valid.  This is OK when called from perf_release since
414  * that only calls us on the top-level context, which can't be a clone.
415  * When called from perf_event_exit_task, it's OK because the
416  * context has been detached from its task.
417  */
418 static void perf_event_remove_from_context(struct perf_event *event)
419 {
420         struct perf_event_context *ctx = event->ctx;
421         struct task_struct *task = ctx->task;
422
423         if (!task) {
424                 /*
425                  * Per cpu events are removed via an smp call and
426                  * the removal is always sucessful.
427                  */
428                 smp_call_function_single(event->cpu,
429                                          __perf_event_remove_from_context,
430                                          event, 1);
431                 return;
432         }
433
434 retry:
435         task_oncpu_function_call(task, __perf_event_remove_from_context,
436                                  event);
437
438         spin_lock_irq(&ctx->lock);
439         /*
440          * If the context is active we need to retry the smp call.
441          */
442         if (ctx->nr_active && !list_empty(&event->group_entry)) {
443                 spin_unlock_irq(&ctx->lock);
444                 goto retry;
445         }
446
447         /*
448          * The lock prevents that this context is scheduled in so we
449          * can remove the event safely, if the call above did not
450          * succeed.
451          */
452         if (!list_empty(&event->group_entry))
453                 list_del_event(event, ctx);
454         spin_unlock_irq(&ctx->lock);
455 }
456
457 static inline u64 perf_clock(void)
458 {
459         return cpu_clock(smp_processor_id());
460 }
461
462 /*
463  * Update the record of the current time in a context.
464  */
465 static void update_context_time(struct perf_event_context *ctx)
466 {
467         u64 now = perf_clock();
468
469         ctx->time += now - ctx->timestamp;
470         ctx->timestamp = now;
471 }
472
473 /*
474  * Update the total_time_enabled and total_time_running fields for a event.
475  */
476 static void update_event_times(struct perf_event *event)
477 {
478         struct perf_event_context *ctx = event->ctx;
479         u64 run_end;
480
481         if (event->state < PERF_EVENT_STATE_INACTIVE ||
482             event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
483                 return;
484
485         event->total_time_enabled = ctx->time - event->tstamp_enabled;
486
487         if (event->state == PERF_EVENT_STATE_INACTIVE)
488                 run_end = event->tstamp_stopped;
489         else
490                 run_end = ctx->time;
491
492         event->total_time_running = run_end - event->tstamp_running;
493 }
494
495 /*
496  * Update total_time_enabled and total_time_running for all events in a group.
497  */
498 static void update_group_times(struct perf_event *leader)
499 {
500         struct perf_event *event;
501
502         update_event_times(leader);
503         list_for_each_entry(event, &leader->sibling_list, group_entry)
504                 update_event_times(event);
505 }
506
507 /*
508  * Cross CPU call to disable a performance event
509  */
510 static void __perf_event_disable(void *info)
511 {
512         struct perf_event *event = info;
513         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
514         struct perf_event_context *ctx = event->ctx;
515
516         /*
517          * If this is a per-task event, need to check whether this
518          * event's task is the current task on this cpu.
519          */
520         if (ctx->task && cpuctx->task_ctx != ctx)
521                 return;
522
523         spin_lock(&ctx->lock);
524
525         /*
526          * If the event is on, turn it off.
527          * If it is in error state, leave it in error state.
528          */
529         if (event->state >= PERF_EVENT_STATE_INACTIVE) {
530                 update_context_time(ctx);
531                 update_group_times(event);
532                 if (event == event->group_leader)
533                         group_sched_out(event, cpuctx, ctx);
534                 else
535                         event_sched_out(event, cpuctx, ctx);
536                 event->state = PERF_EVENT_STATE_OFF;
537         }
538
539         spin_unlock(&ctx->lock);
540 }
541
542 /*
543  * Disable a event.
544  *
545  * If event->ctx is a cloned context, callers must make sure that
546  * every task struct that event->ctx->task could possibly point to
547  * remains valid.  This condition is satisifed when called through
548  * perf_event_for_each_child or perf_event_for_each because they
549  * hold the top-level event's child_mutex, so any descendant that
550  * goes to exit will block in sync_child_event.
551  * When called from perf_pending_event it's OK because event->ctx
552  * is the current context on this CPU and preemption is disabled,
553  * hence we can't get into perf_event_task_sched_out for this context.
554  */
555 static void perf_event_disable(struct perf_event *event)
556 {
557         struct perf_event_context *ctx = event->ctx;
558         struct task_struct *task = ctx->task;
559
560         if (!task) {
561                 /*
562                  * Disable the event on the cpu that it's on
563                  */
564                 smp_call_function_single(event->cpu, __perf_event_disable,
565                                          event, 1);
566                 return;
567         }
568
569  retry:
570         task_oncpu_function_call(task, __perf_event_disable, event);
571
572         spin_lock_irq(&ctx->lock);
573         /*
574          * If the event is still active, we need to retry the cross-call.
575          */
576         if (event->state == PERF_EVENT_STATE_ACTIVE) {
577                 spin_unlock_irq(&ctx->lock);
578                 goto retry;
579         }
580
581         /*
582          * Since we have the lock this context can't be scheduled
583          * in, so we can change the state safely.
584          */
585         if (event->state == PERF_EVENT_STATE_INACTIVE) {
586                 update_group_times(event);
587                 event->state = PERF_EVENT_STATE_OFF;
588         }
589
590         spin_unlock_irq(&ctx->lock);
591 }
592
593 static int
594 event_sched_in(struct perf_event *event,
595                  struct perf_cpu_context *cpuctx,
596                  struct perf_event_context *ctx,
597                  int cpu)
598 {
599         if (event->state <= PERF_EVENT_STATE_OFF)
600                 return 0;
601
602         event->state = PERF_EVENT_STATE_ACTIVE;
603         event->oncpu = cpu;     /* TODO: put 'cpu' into cpuctx->cpu */
604         /*
605          * The new state must be visible before we turn it on in the hardware:
606          */
607         smp_wmb();
608
609         if (event->pmu->enable(event)) {
610                 event->state = PERF_EVENT_STATE_INACTIVE;
611                 event->oncpu = -1;
612                 return -EAGAIN;
613         }
614
615         event->tstamp_running += ctx->time - event->tstamp_stopped;
616
617         if (!is_software_event(event))
618                 cpuctx->active_oncpu++;
619         ctx->nr_active++;
620
621         if (event->attr.exclusive)
622                 cpuctx->exclusive = 1;
623
624         return 0;
625 }
626
627 static int
628 group_sched_in(struct perf_event *group_event,
629                struct perf_cpu_context *cpuctx,
630                struct perf_event_context *ctx,
631                int cpu)
632 {
633         struct perf_event *event, *partial_group;
634         int ret;
635
636         if (group_event->state == PERF_EVENT_STATE_OFF)
637                 return 0;
638
639         ret = hw_perf_group_sched_in(group_event, cpuctx, ctx, cpu);
640         if (ret)
641                 return ret < 0 ? ret : 0;
642
643         if (event_sched_in(group_event, cpuctx, ctx, cpu))
644                 return -EAGAIN;
645
646         /*
647          * Schedule in siblings as one group (if any):
648          */
649         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
650                 if (event_sched_in(event, cpuctx, ctx, cpu)) {
651                         partial_group = event;
652                         goto group_error;
653                 }
654         }
655
656         return 0;
657
658 group_error:
659         /*
660          * Groups can be scheduled in as one unit only, so undo any
661          * partial group before returning:
662          */
663         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
664                 if (event == partial_group)
665                         break;
666                 event_sched_out(event, cpuctx, ctx);
667         }
668         event_sched_out(group_event, cpuctx, ctx);
669
670         return -EAGAIN;
671 }
672
673 /*
674  * Return 1 for a group consisting entirely of software events,
675  * 0 if the group contains any hardware events.
676  */
677 static int is_software_only_group(struct perf_event *leader)
678 {
679         struct perf_event *event;
680
681         if (!is_software_event(leader))
682                 return 0;
683
684         list_for_each_entry(event, &leader->sibling_list, group_entry)
685                 if (!is_software_event(event))
686                         return 0;
687
688         return 1;
689 }
690
691 /*
692  * Work out whether we can put this event group on the CPU now.
693  */
694 static int group_can_go_on(struct perf_event *event,
695                            struct perf_cpu_context *cpuctx,
696                            int can_add_hw)
697 {
698         /*
699          * Groups consisting entirely of software events can always go on.
700          */
701         if (is_software_only_group(event))
702                 return 1;
703         /*
704          * If an exclusive group is already on, no other hardware
705          * events can go on.
706          */
707         if (cpuctx->exclusive)
708                 return 0;
709         /*
710          * If this group is exclusive and there are already
711          * events on the CPU, it can't go on.
712          */
713         if (event->attr.exclusive && cpuctx->active_oncpu)
714                 return 0;
715         /*
716          * Otherwise, try to add it if all previous groups were able
717          * to go on.
718          */
719         return can_add_hw;
720 }
721
722 static void add_event_to_ctx(struct perf_event *event,
723                                struct perf_event_context *ctx)
724 {
725         list_add_event(event, ctx);
726         event->tstamp_enabled = ctx->time;
727         event->tstamp_running = ctx->time;
728         event->tstamp_stopped = ctx->time;
729 }
730
731 /*
732  * Cross CPU call to install and enable a performance event
733  *
734  * Must be called with ctx->mutex held
735  */
736 static void __perf_install_in_context(void *info)
737 {
738         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
739         struct perf_event *event = info;
740         struct perf_event_context *ctx = event->ctx;
741         struct perf_event *leader = event->group_leader;
742         int cpu = smp_processor_id();
743         int err;
744
745         /*
746          * If this is a task context, we need to check whether it is
747          * the current task context of this cpu. If not it has been
748          * scheduled out before the smp call arrived.
749          * Or possibly this is the right context but it isn't
750          * on this cpu because it had no events.
751          */
752         if (ctx->task && cpuctx->task_ctx != ctx) {
753                 if (cpuctx->task_ctx || ctx->task != current)
754                         return;
755                 cpuctx->task_ctx = ctx;
756         }
757
758         spin_lock(&ctx->lock);
759         ctx->is_active = 1;
760         update_context_time(ctx);
761
762         /*
763          * Protect the list operation against NMI by disabling the
764          * events on a global level. NOP for non NMI based events.
765          */
766         perf_disable();
767
768         add_event_to_ctx(event, ctx);
769
770         /*
771          * Don't put the event on if it is disabled or if
772          * it is in a group and the group isn't on.
773          */
774         if (event->state != PERF_EVENT_STATE_INACTIVE ||
775             (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
776                 goto unlock;
777
778         /*
779          * An exclusive event can't go on if there are already active
780          * hardware events, and no hardware event can go on if there
781          * is already an exclusive event on.
782          */
783         if (!group_can_go_on(event, cpuctx, 1))
784                 err = -EEXIST;
785         else
786                 err = event_sched_in(event, cpuctx, ctx, cpu);
787
788         if (err) {
789                 /*
790                  * This event couldn't go on.  If it is in a group
791                  * then we have to pull the whole group off.
792                  * If the event group is pinned then put it in error state.
793                  */
794                 if (leader != event)
795                         group_sched_out(leader, cpuctx, ctx);
796                 if (leader->attr.pinned) {
797                         update_group_times(leader);
798                         leader->state = PERF_EVENT_STATE_ERROR;
799                 }
800         }
801
802         if (!err && !ctx->task && cpuctx->max_pertask)
803                 cpuctx->max_pertask--;
804
805  unlock:
806         perf_enable();
807
808         spin_unlock(&ctx->lock);
809 }
810
811 /*
812  * Attach a performance event to a context
813  *
814  * First we add the event to the list with the hardware enable bit
815  * in event->hw_config cleared.
816  *
817  * If the event is attached to a task which is on a CPU we use a smp
818  * call to enable it in the task context. The task might have been
819  * scheduled away, but we check this in the smp call again.
820  *
821  * Must be called with ctx->mutex held.
822  */
823 static void
824 perf_install_in_context(struct perf_event_context *ctx,
825                         struct perf_event *event,
826                         int cpu)
827 {
828         struct task_struct *task = ctx->task;
829
830         if (!task) {
831                 /*
832                  * Per cpu events are installed via an smp call and
833                  * the install is always sucessful.
834                  */
835                 smp_call_function_single(cpu, __perf_install_in_context,
836                                          event, 1);
837                 return;
838         }
839
840 retry:
841         task_oncpu_function_call(task, __perf_install_in_context,
842                                  event);
843
844         spin_lock_irq(&ctx->lock);
845         /*
846          * we need to retry the smp call.
847          */
848         if (ctx->is_active && list_empty(&event->group_entry)) {
849                 spin_unlock_irq(&ctx->lock);
850                 goto retry;
851         }
852
853         /*
854          * The lock prevents that this context is scheduled in so we
855          * can add the event safely, if it the call above did not
856          * succeed.
857          */
858         if (list_empty(&event->group_entry))
859                 add_event_to_ctx(event, ctx);
860         spin_unlock_irq(&ctx->lock);
861 }
862
863 /*
864  * Put a event into inactive state and update time fields.
865  * Enabling the leader of a group effectively enables all
866  * the group members that aren't explicitly disabled, so we
867  * have to update their ->tstamp_enabled also.
868  * Note: this works for group members as well as group leaders
869  * since the non-leader members' sibling_lists will be empty.
870  */
871 static void __perf_event_mark_enabled(struct perf_event *event,
872                                         struct perf_event_context *ctx)
873 {
874         struct perf_event *sub;
875
876         event->state = PERF_EVENT_STATE_INACTIVE;
877         event->tstamp_enabled = ctx->time - event->total_time_enabled;
878         list_for_each_entry(sub, &event->sibling_list, group_entry)
879                 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
880                         sub->tstamp_enabled =
881                                 ctx->time - sub->total_time_enabled;
882 }
883
884 /*
885  * Cross CPU call to enable a performance event
886  */
887 static void __perf_event_enable(void *info)
888 {
889         struct perf_event *event = info;
890         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
891         struct perf_event_context *ctx = event->ctx;
892         struct perf_event *leader = event->group_leader;
893         int err;
894
895         /*
896          * If this is a per-task event, need to check whether this
897          * event's task is the current task on this cpu.
898          */
899         if (ctx->task && cpuctx->task_ctx != ctx) {
900                 if (cpuctx->task_ctx || ctx->task != current)
901                         return;
902                 cpuctx->task_ctx = ctx;
903         }
904
905         spin_lock(&ctx->lock);
906         ctx->is_active = 1;
907         update_context_time(ctx);
908
909         if (event->state >= PERF_EVENT_STATE_INACTIVE)
910                 goto unlock;
911         __perf_event_mark_enabled(event, ctx);
912
913         /*
914          * If the event is in a group and isn't the group leader,
915          * then don't put it on unless the group is on.
916          */
917         if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
918                 goto unlock;
919
920         if (!group_can_go_on(event, cpuctx, 1)) {
921                 err = -EEXIST;
922         } else {
923                 perf_disable();
924                 if (event == leader)
925                         err = group_sched_in(event, cpuctx, ctx,
926                                              smp_processor_id());
927                 else
928                         err = event_sched_in(event, cpuctx, ctx,
929                                                smp_processor_id());
930                 perf_enable();
931         }
932
933         if (err) {
934                 /*
935                  * If this event can't go on and it's part of a
936                  * group, then the whole group has to come off.
937                  */
938                 if (leader != event)
939                         group_sched_out(leader, cpuctx, ctx);
940                 if (leader->attr.pinned) {
941                         update_group_times(leader);
942                         leader->state = PERF_EVENT_STATE_ERROR;
943                 }
944         }
945
946  unlock:
947         spin_unlock(&ctx->lock);
948 }
949
950 /*
951  * Enable a event.
952  *
953  * If event->ctx is a cloned context, callers must make sure that
954  * every task struct that event->ctx->task could possibly point to
955  * remains valid.  This condition is satisfied when called through
956  * perf_event_for_each_child or perf_event_for_each as described
957  * for perf_event_disable.
958  */
959 static void perf_event_enable(struct perf_event *event)
960 {
961         struct perf_event_context *ctx = event->ctx;
962         struct task_struct *task = ctx->task;
963
964         if (!task) {
965                 /*
966                  * Enable the event on the cpu that it's on
967                  */
968                 smp_call_function_single(event->cpu, __perf_event_enable,
969                                          event, 1);
970                 return;
971         }
972
973         spin_lock_irq(&ctx->lock);
974         if (event->state >= PERF_EVENT_STATE_INACTIVE)
975                 goto out;
976
977         /*
978          * If the event is in error state, clear that first.
979          * That way, if we see the event in error state below, we
980          * know that it has gone back into error state, as distinct
981          * from the task having been scheduled away before the
982          * cross-call arrived.
983          */
984         if (event->state == PERF_EVENT_STATE_ERROR)
985                 event->state = PERF_EVENT_STATE_OFF;
986
987  retry:
988         spin_unlock_irq(&ctx->lock);
989         task_oncpu_function_call(task, __perf_event_enable, event);
990
991         spin_lock_irq(&ctx->lock);
992
993         /*
994          * If the context is active and the event is still off,
995          * we need to retry the cross-call.
996          */
997         if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
998                 goto retry;
999
1000         /*
1001          * Since we have the lock this context can't be scheduled
1002          * in, so we can change the state safely.
1003          */
1004         if (event->state == PERF_EVENT_STATE_OFF)
1005                 __perf_event_mark_enabled(event, ctx);
1006
1007  out:
1008         spin_unlock_irq(&ctx->lock);
1009 }
1010
1011 static int perf_event_refresh(struct perf_event *event, int refresh)
1012 {
1013         /*
1014          * not supported on inherited events
1015          */
1016         if (event->attr.inherit)
1017                 return -EINVAL;
1018
1019         atomic_add(refresh, &event->event_limit);
1020         perf_event_enable(event);
1021
1022         return 0;
1023 }
1024
1025 void __perf_event_sched_out(struct perf_event_context *ctx,
1026                               struct perf_cpu_context *cpuctx)
1027 {
1028         struct perf_event *event;
1029
1030         spin_lock(&ctx->lock);
1031         ctx->is_active = 0;
1032         if (likely(!ctx->nr_events))
1033                 goto out;
1034         update_context_time(ctx);
1035
1036         perf_disable();
1037         if (ctx->nr_active) {
1038                 list_for_each_entry(event, &ctx->group_list, group_entry)
1039                         group_sched_out(event, cpuctx, ctx);
1040         }
1041         perf_enable();
1042  out:
1043         spin_unlock(&ctx->lock);
1044 }
1045
1046 /*
1047  * Test whether two contexts are equivalent, i.e. whether they
1048  * have both been cloned from the same version of the same context
1049  * and they both have the same number of enabled events.
1050  * If the number of enabled events is the same, then the set
1051  * of enabled events should be the same, because these are both
1052  * inherited contexts, therefore we can't access individual events
1053  * in them directly with an fd; we can only enable/disable all
1054  * events via prctl, or enable/disable all events in a family
1055  * via ioctl, which will have the same effect on both contexts.
1056  */
1057 static int context_equiv(struct perf_event_context *ctx1,
1058                          struct perf_event_context *ctx2)
1059 {
1060         return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1061                 && ctx1->parent_gen == ctx2->parent_gen
1062                 && !ctx1->pin_count && !ctx2->pin_count;
1063 }
1064
1065 static void __perf_event_sync_stat(struct perf_event *event,
1066                                      struct perf_event *next_event)
1067 {
1068         u64 value;
1069
1070         if (!event->attr.inherit_stat)
1071                 return;
1072
1073         /*
1074          * Update the event value, we cannot use perf_event_read()
1075          * because we're in the middle of a context switch and have IRQs
1076          * disabled, which upsets smp_call_function_single(), however
1077          * we know the event must be on the current CPU, therefore we
1078          * don't need to use it.
1079          */
1080         switch (event->state) {
1081         case PERF_EVENT_STATE_ACTIVE:
1082                 event->pmu->read(event);
1083                 /* fall-through */
1084
1085         case PERF_EVENT_STATE_INACTIVE:
1086                 update_event_times(event);
1087                 break;
1088
1089         default:
1090                 break;
1091         }
1092
1093         /*
1094          * In order to keep per-task stats reliable we need to flip the event
1095          * values when we flip the contexts.
1096          */
1097         value = atomic64_read(&next_event->count);
1098         value = atomic64_xchg(&event->count, value);
1099         atomic64_set(&next_event->count, value);
1100
1101         swap(event->total_time_enabled, next_event->total_time_enabled);
1102         swap(event->total_time_running, next_event->total_time_running);
1103
1104         /*
1105          * Since we swizzled the values, update the user visible data too.
1106          */
1107         perf_event_update_userpage(event);
1108         perf_event_update_userpage(next_event);
1109 }
1110
1111 #define list_next_entry(pos, member) \
1112         list_entry(pos->member.next, typeof(*pos), member)
1113
1114 static void perf_event_sync_stat(struct perf_event_context *ctx,
1115                                    struct perf_event_context *next_ctx)
1116 {
1117         struct perf_event *event, *next_event;
1118
1119         if (!ctx->nr_stat)
1120                 return;
1121
1122         update_context_time(ctx);
1123
1124         event = list_first_entry(&ctx->event_list,
1125                                    struct perf_event, event_entry);
1126
1127         next_event = list_first_entry(&next_ctx->event_list,
1128                                         struct perf_event, event_entry);
1129
1130         while (&event->event_entry != &ctx->event_list &&
1131                &next_event->event_entry != &next_ctx->event_list) {
1132
1133                 __perf_event_sync_stat(event, next_event);
1134
1135                 event = list_next_entry(event, event_entry);
1136                 next_event = list_next_entry(next_event, event_entry);
1137         }
1138 }
1139
1140 /*
1141  * Called from scheduler to remove the events of the current task,
1142  * with interrupts disabled.
1143  *
1144  * We stop each event and update the event value in event->count.
1145  *
1146  * This does not protect us against NMI, but disable()
1147  * sets the disabled bit in the control field of event _before_
1148  * accessing the event control register. If a NMI hits, then it will
1149  * not restart the event.
1150  */
1151 void perf_event_task_sched_out(struct task_struct *task,
1152                                  struct task_struct *next, int cpu)
1153 {
1154         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1155         struct perf_event_context *ctx = task->perf_event_ctxp;
1156         struct perf_event_context *next_ctx;
1157         struct perf_event_context *parent;
1158         struct pt_regs *regs;
1159         int do_switch = 1;
1160
1161         regs = task_pt_regs(task);
1162         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1163
1164         if (likely(!ctx || !cpuctx->task_ctx))
1165                 return;
1166
1167         rcu_read_lock();
1168         parent = rcu_dereference(ctx->parent_ctx);
1169         next_ctx = next->perf_event_ctxp;
1170         if (parent && next_ctx &&
1171             rcu_dereference(next_ctx->parent_ctx) == parent) {
1172                 /*
1173                  * Looks like the two contexts are clones, so we might be
1174                  * able to optimize the context switch.  We lock both
1175                  * contexts and check that they are clones under the
1176                  * lock (including re-checking that neither has been
1177                  * uncloned in the meantime).  It doesn't matter which
1178                  * order we take the locks because no other cpu could
1179                  * be trying to lock both of these tasks.
1180                  */
1181                 spin_lock(&ctx->lock);
1182                 spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1183                 if (context_equiv(ctx, next_ctx)) {
1184                         /*
1185                          * XXX do we need a memory barrier of sorts
1186                          * wrt to rcu_dereference() of perf_event_ctxp
1187                          */
1188                         task->perf_event_ctxp = next_ctx;
1189                         next->perf_event_ctxp = ctx;
1190                         ctx->task = next;
1191                         next_ctx->task = task;
1192                         do_switch = 0;
1193
1194                         perf_event_sync_stat(ctx, next_ctx);
1195                 }
1196                 spin_unlock(&next_ctx->lock);
1197                 spin_unlock(&ctx->lock);
1198         }
1199         rcu_read_unlock();
1200
1201         if (do_switch) {
1202                 __perf_event_sched_out(ctx, cpuctx);
1203                 cpuctx->task_ctx = NULL;
1204         }
1205 }
1206
1207 /*
1208  * Called with IRQs disabled
1209  */
1210 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1211 {
1212         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1213
1214         if (!cpuctx->task_ctx)
1215                 return;
1216
1217         if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1218                 return;
1219
1220         __perf_event_sched_out(ctx, cpuctx);
1221         cpuctx->task_ctx = NULL;
1222 }
1223
1224 /*
1225  * Called with IRQs disabled
1226  */
1227 static void perf_event_cpu_sched_out(struct perf_cpu_context *cpuctx)
1228 {
1229         __perf_event_sched_out(&cpuctx->ctx, cpuctx);
1230 }
1231
1232 static void
1233 __perf_event_sched_in(struct perf_event_context *ctx,
1234                         struct perf_cpu_context *cpuctx, int cpu)
1235 {
1236         struct perf_event *event;
1237         int can_add_hw = 1;
1238
1239         spin_lock(&ctx->lock);
1240         ctx->is_active = 1;
1241         if (likely(!ctx->nr_events))
1242                 goto out;
1243
1244         ctx->timestamp = perf_clock();
1245
1246         perf_disable();
1247
1248         /*
1249          * First go through the list and put on any pinned groups
1250          * in order to give them the best chance of going on.
1251          */
1252         list_for_each_entry(event, &ctx->group_list, group_entry) {
1253                 if (event->state <= PERF_EVENT_STATE_OFF ||
1254                     !event->attr.pinned)
1255                         continue;
1256                 if (event->cpu != -1 && event->cpu != cpu)
1257                         continue;
1258
1259                 if (group_can_go_on(event, cpuctx, 1))
1260                         group_sched_in(event, cpuctx, ctx, cpu);
1261
1262                 /*
1263                  * If this pinned group hasn't been scheduled,
1264                  * put it in error state.
1265                  */
1266                 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1267                         update_group_times(event);
1268                         event->state = PERF_EVENT_STATE_ERROR;
1269                 }
1270         }
1271
1272         list_for_each_entry(event, &ctx->group_list, group_entry) {
1273                 /*
1274                  * Ignore events in OFF or ERROR state, and
1275                  * ignore pinned events since we did them already.
1276                  */
1277                 if (event->state <= PERF_EVENT_STATE_OFF ||
1278                     event->attr.pinned)
1279                         continue;
1280
1281                 /*
1282                  * Listen to the 'cpu' scheduling filter constraint
1283                  * of events:
1284                  */
1285                 if (event->cpu != -1 && event->cpu != cpu)
1286                         continue;
1287
1288                 if (group_can_go_on(event, cpuctx, can_add_hw))
1289                         if (group_sched_in(event, cpuctx, ctx, cpu))
1290                                 can_add_hw = 0;
1291         }
1292         perf_enable();
1293  out:
1294         spin_unlock(&ctx->lock);
1295 }
1296
1297 /*
1298  * Called from scheduler to add the events of the current task
1299  * with interrupts disabled.
1300  *
1301  * We restore the event value and then enable it.
1302  *
1303  * This does not protect us against NMI, but enable()
1304  * sets the enabled bit in the control field of event _before_
1305  * accessing the event control register. If a NMI hits, then it will
1306  * keep the event running.
1307  */
1308 void perf_event_task_sched_in(struct task_struct *task, int cpu)
1309 {
1310         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1311         struct perf_event_context *ctx = task->perf_event_ctxp;
1312
1313         if (likely(!ctx))
1314                 return;
1315         if (cpuctx->task_ctx == ctx)
1316                 return;
1317         __perf_event_sched_in(ctx, cpuctx, cpu);
1318         cpuctx->task_ctx = ctx;
1319 }
1320
1321 static void perf_event_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
1322 {
1323         struct perf_event_context *ctx = &cpuctx->ctx;
1324
1325         __perf_event_sched_in(ctx, cpuctx, cpu);
1326 }
1327
1328 #define MAX_INTERRUPTS (~0ULL)
1329
1330 static void perf_log_throttle(struct perf_event *event, int enable);
1331
1332 static void perf_adjust_period(struct perf_event *event, u64 events)
1333 {
1334         struct hw_perf_event *hwc = &event->hw;
1335         u64 period, sample_period;
1336         s64 delta;
1337
1338         events *= hwc->sample_period;
1339         period = div64_u64(events, event->attr.sample_freq);
1340
1341         delta = (s64)(period - hwc->sample_period);
1342         delta = (delta + 7) / 8; /* low pass filter */
1343
1344         sample_period = hwc->sample_period + delta;
1345
1346         if (!sample_period)
1347                 sample_period = 1;
1348
1349         hwc->sample_period = sample_period;
1350 }
1351
1352 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1353 {
1354         struct perf_event *event;
1355         struct hw_perf_event *hwc;
1356         u64 interrupts, freq;
1357
1358         spin_lock(&ctx->lock);
1359         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1360                 if (event->state != PERF_EVENT_STATE_ACTIVE)
1361                         continue;
1362
1363                 hwc = &event->hw;
1364
1365                 interrupts = hwc->interrupts;
1366                 hwc->interrupts = 0;
1367
1368                 /*
1369                  * unthrottle events on the tick
1370                  */
1371                 if (interrupts == MAX_INTERRUPTS) {
1372                         perf_log_throttle(event, 1);
1373                         event->pmu->unthrottle(event);
1374                         interrupts = 2*sysctl_perf_event_sample_rate/HZ;
1375                 }
1376
1377                 if (!event->attr.freq || !event->attr.sample_freq)
1378                         continue;
1379
1380                 /*
1381                  * if the specified freq < HZ then we need to skip ticks
1382                  */
1383                 if (event->attr.sample_freq < HZ) {
1384                         freq = event->attr.sample_freq;
1385
1386                         hwc->freq_count += freq;
1387                         hwc->freq_interrupts += interrupts;
1388
1389                         if (hwc->freq_count < HZ)
1390                                 continue;
1391
1392                         interrupts = hwc->freq_interrupts;
1393                         hwc->freq_interrupts = 0;
1394                         hwc->freq_count -= HZ;
1395                 } else
1396                         freq = HZ;
1397
1398                 perf_adjust_period(event, freq * interrupts);
1399
1400                 /*
1401                  * In order to avoid being stalled by an (accidental) huge
1402                  * sample period, force reset the sample period if we didn't
1403                  * get any events in this freq period.
1404                  */
1405                 if (!interrupts) {
1406                         perf_disable();
1407                         event->pmu->disable(event);
1408                         atomic64_set(&hwc->period_left, 0);
1409                         event->pmu->enable(event);
1410                         perf_enable();
1411                 }
1412         }
1413         spin_unlock(&ctx->lock);
1414 }
1415
1416 /*
1417  * Round-robin a context's events:
1418  */
1419 static void rotate_ctx(struct perf_event_context *ctx)
1420 {
1421         struct perf_event *event;
1422
1423         if (!ctx->nr_events)
1424                 return;
1425
1426         spin_lock(&ctx->lock);
1427         /*
1428          * Rotate the first entry last (works just fine for group events too):
1429          */
1430         perf_disable();
1431         list_for_each_entry(event, &ctx->group_list, group_entry) {
1432                 list_move_tail(&event->group_entry, &ctx->group_list);
1433                 break;
1434         }
1435         perf_enable();
1436
1437         spin_unlock(&ctx->lock);
1438 }
1439
1440 void perf_event_task_tick(struct task_struct *curr, int cpu)
1441 {
1442         struct perf_cpu_context *cpuctx;
1443         struct perf_event_context *ctx;
1444
1445         if (!atomic_read(&nr_events))
1446                 return;
1447
1448         cpuctx = &per_cpu(perf_cpu_context, cpu);
1449         ctx = curr->perf_event_ctxp;
1450
1451         perf_ctx_adjust_freq(&cpuctx->ctx);
1452         if (ctx)
1453                 perf_ctx_adjust_freq(ctx);
1454
1455         perf_event_cpu_sched_out(cpuctx);
1456         if (ctx)
1457                 __perf_event_task_sched_out(ctx);
1458
1459         rotate_ctx(&cpuctx->ctx);
1460         if (ctx)
1461                 rotate_ctx(ctx);
1462
1463         perf_event_cpu_sched_in(cpuctx, cpu);
1464         if (ctx)
1465                 perf_event_task_sched_in(curr, cpu);
1466 }
1467
1468 /*
1469  * Enable all of a task's events that have been marked enable-on-exec.
1470  * This expects task == current.
1471  */
1472 static void perf_event_enable_on_exec(struct task_struct *task)
1473 {
1474         struct perf_event_context *ctx;
1475         struct perf_event *event;
1476         unsigned long flags;
1477         int enabled = 0;
1478
1479         local_irq_save(flags);
1480         ctx = task->perf_event_ctxp;
1481         if (!ctx || !ctx->nr_events)
1482                 goto out;
1483
1484         __perf_event_task_sched_out(ctx);
1485
1486         spin_lock(&ctx->lock);
1487
1488         list_for_each_entry(event, &ctx->group_list, group_entry) {
1489                 if (!event->attr.enable_on_exec)
1490                         continue;
1491                 event->attr.enable_on_exec = 0;
1492                 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1493                         continue;
1494                 __perf_event_mark_enabled(event, ctx);
1495                 enabled = 1;
1496         }
1497
1498         /*
1499          * Unclone this context if we enabled any event.
1500          */
1501         if (enabled)
1502                 unclone_ctx(ctx);
1503
1504         spin_unlock(&ctx->lock);
1505
1506         perf_event_task_sched_in(task, smp_processor_id());
1507  out:
1508         local_irq_restore(flags);
1509 }
1510
1511 /*
1512  * Cross CPU call to read the hardware event
1513  */
1514 static void __perf_event_read(void *info)
1515 {
1516         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1517         struct perf_event *event = info;
1518         struct perf_event_context *ctx = event->ctx;
1519
1520         /*
1521          * If this is a task context, we need to check whether it is
1522          * the current task context of this cpu.  If not it has been
1523          * scheduled out before the smp call arrived.  In that case
1524          * event->count would have been updated to a recent sample
1525          * when the event was scheduled out.
1526          */
1527         if (ctx->task && cpuctx->task_ctx != ctx)
1528                 return;
1529
1530         spin_lock(&ctx->lock);
1531         update_context_time(ctx);
1532         update_event_times(event);
1533         spin_unlock(&ctx->lock);
1534
1535         event->pmu->read(event);
1536 }
1537
1538 static u64 perf_event_read(struct perf_event *event)
1539 {
1540         /*
1541          * If event is enabled and currently active on a CPU, update the
1542          * value in the event structure:
1543          */
1544         if (event->state == PERF_EVENT_STATE_ACTIVE) {
1545                 smp_call_function_single(event->oncpu,
1546                                          __perf_event_read, event, 1);
1547         } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1548                 struct perf_event_context *ctx = event->ctx;
1549                 unsigned long flags;
1550
1551                 spin_lock_irqsave(&ctx->lock, flags);
1552                 update_context_time(ctx);
1553                 update_event_times(event);
1554                 spin_unlock_irqrestore(&ctx->lock, flags);
1555         }
1556
1557         return atomic64_read(&event->count);
1558 }
1559
1560 /*
1561  * Initialize the perf_event context in a task_struct:
1562  */
1563 static void
1564 __perf_event_init_context(struct perf_event_context *ctx,
1565                             struct task_struct *task)
1566 {
1567         memset(ctx, 0, sizeof(*ctx));
1568         spin_lock_init(&ctx->lock);
1569         mutex_init(&ctx->mutex);
1570         INIT_LIST_HEAD(&ctx->group_list);
1571         INIT_LIST_HEAD(&ctx->event_list);
1572         atomic_set(&ctx->refcount, 1);
1573         ctx->task = task;
1574 }
1575
1576 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1577 {
1578         struct perf_event_context *ctx;
1579         struct perf_cpu_context *cpuctx;
1580         struct task_struct *task;
1581         unsigned long flags;
1582         int err;
1583
1584         /*
1585          * If cpu is not a wildcard then this is a percpu event:
1586          */
1587         if (cpu != -1) {
1588                 /* Must be root to operate on a CPU event: */
1589                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1590                         return ERR_PTR(-EACCES);
1591
1592                 if (cpu < 0 || cpu > num_possible_cpus())
1593                         return ERR_PTR(-EINVAL);
1594
1595                 /*
1596                  * We could be clever and allow to attach a event to an
1597                  * offline CPU and activate it when the CPU comes up, but
1598                  * that's for later.
1599                  */
1600                 if (!cpu_isset(cpu, cpu_online_map))
1601                         return ERR_PTR(-ENODEV);
1602
1603                 cpuctx = &per_cpu(perf_cpu_context, cpu);
1604                 ctx = &cpuctx->ctx;
1605                 get_ctx(ctx);
1606
1607                 return ctx;
1608         }
1609
1610         rcu_read_lock();
1611         if (!pid)
1612                 task = current;
1613         else
1614                 task = find_task_by_vpid(pid);
1615         if (task)
1616                 get_task_struct(task);
1617         rcu_read_unlock();
1618
1619         if (!task)
1620                 return ERR_PTR(-ESRCH);
1621
1622         /*
1623          * Can't attach events to a dying task.
1624          */
1625         err = -ESRCH;
1626         if (task->flags & PF_EXITING)
1627                 goto errout;
1628
1629         /* Reuse ptrace permission checks for now. */
1630         err = -EACCES;
1631         if (!ptrace_may_access(task, PTRACE_MODE_READ))
1632                 goto errout;
1633
1634  retry:
1635         ctx = perf_lock_task_context(task, &flags);
1636         if (ctx) {
1637                 unclone_ctx(ctx);
1638                 spin_unlock_irqrestore(&ctx->lock, flags);
1639         }
1640
1641         if (!ctx) {
1642                 ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1643                 err = -ENOMEM;
1644                 if (!ctx)
1645                         goto errout;
1646                 __perf_event_init_context(ctx, task);
1647                 get_ctx(ctx);
1648                 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1649                         /*
1650                          * We raced with some other task; use
1651                          * the context they set.
1652                          */
1653                         kfree(ctx);
1654                         goto retry;
1655                 }
1656                 get_task_struct(task);
1657         }
1658
1659         put_task_struct(task);
1660         return ctx;
1661
1662  errout:
1663         put_task_struct(task);
1664         return ERR_PTR(err);
1665 }
1666
1667 static void perf_event_free_filter(struct perf_event *event);
1668
1669 static void free_event_rcu(struct rcu_head *head)
1670 {
1671         struct perf_event *event;
1672
1673         event = container_of(head, struct perf_event, rcu_head);
1674         if (event->ns)
1675                 put_pid_ns(event->ns);
1676         perf_event_free_filter(event);
1677         kfree(event);
1678 }
1679
1680 static void perf_pending_sync(struct perf_event *event);
1681
1682 static void free_event(struct perf_event *event)
1683 {
1684         perf_pending_sync(event);
1685
1686         if (!event->parent) {
1687                 atomic_dec(&nr_events);
1688                 if (event->attr.mmap)
1689                         atomic_dec(&nr_mmap_events);
1690                 if (event->attr.comm)
1691                         atomic_dec(&nr_comm_events);
1692                 if (event->attr.task)
1693                         atomic_dec(&nr_task_events);
1694         }
1695
1696         if (event->output) {
1697                 fput(event->output->filp);
1698                 event->output = NULL;
1699         }
1700
1701         if (event->destroy)
1702                 event->destroy(event);
1703
1704         put_ctx(event->ctx);
1705         call_rcu(&event->rcu_head, free_event_rcu);
1706 }
1707
1708 int perf_event_release_kernel(struct perf_event *event)
1709 {
1710         struct perf_event_context *ctx = event->ctx;
1711
1712         WARN_ON_ONCE(ctx->parent_ctx);
1713         mutex_lock(&ctx->mutex);
1714         perf_event_remove_from_context(event);
1715         mutex_unlock(&ctx->mutex);
1716
1717         mutex_lock(&event->owner->perf_event_mutex);
1718         list_del_init(&event->owner_entry);
1719         mutex_unlock(&event->owner->perf_event_mutex);
1720         put_task_struct(event->owner);
1721
1722         free_event(event);
1723
1724         return 0;
1725 }
1726 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1727
1728 /*
1729  * Called when the last reference to the file is gone.
1730  */
1731 static int perf_release(struct inode *inode, struct file *file)
1732 {
1733         struct perf_event *event = file->private_data;
1734
1735         file->private_data = NULL;
1736
1737         return perf_event_release_kernel(event);
1738 }
1739
1740 static int perf_event_read_size(struct perf_event *event)
1741 {
1742         int entry = sizeof(u64); /* value */
1743         int size = 0;
1744         int nr = 1;
1745
1746         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1747                 size += sizeof(u64);
1748
1749         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1750                 size += sizeof(u64);
1751
1752         if (event->attr.read_format & PERF_FORMAT_ID)
1753                 entry += sizeof(u64);
1754
1755         if (event->attr.read_format & PERF_FORMAT_GROUP) {
1756                 nr += event->group_leader->nr_siblings;
1757                 size += sizeof(u64);
1758         }
1759
1760         size += entry * nr;
1761
1762         return size;
1763 }
1764
1765 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1766 {
1767         struct perf_event *child;
1768         u64 total = 0;
1769
1770         *enabled = 0;
1771         *running = 0;
1772
1773         mutex_lock(&event->child_mutex);
1774         total += perf_event_read(event);
1775         *enabled += event->total_time_enabled +
1776                         atomic64_read(&event->child_total_time_enabled);
1777         *running += event->total_time_running +
1778                         atomic64_read(&event->child_total_time_running);
1779
1780         list_for_each_entry(child, &event->child_list, child_list) {
1781                 total += perf_event_read(child);
1782                 *enabled += child->total_time_enabled;
1783                 *running += child->total_time_running;
1784         }
1785         mutex_unlock(&event->child_mutex);
1786
1787         return total;
1788 }
1789 EXPORT_SYMBOL_GPL(perf_event_read_value);
1790
1791 static int perf_event_read_group(struct perf_event *event,
1792                                    u64 read_format, char __user *buf)
1793 {
1794         struct perf_event *leader = event->group_leader, *sub;
1795         int n = 0, size = 0, ret = -EFAULT;
1796         struct perf_event_context *ctx = leader->ctx;
1797         u64 values[5];
1798         u64 count, enabled, running;
1799
1800         mutex_lock(&ctx->mutex);
1801         count = perf_event_read_value(leader, &enabled, &running);
1802
1803         values[n++] = 1 + leader->nr_siblings;
1804         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1805                 values[n++] = enabled;
1806         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1807                 values[n++] = running;
1808         values[n++] = count;
1809         if (read_format & PERF_FORMAT_ID)
1810                 values[n++] = primary_event_id(leader);
1811
1812         size = n * sizeof(u64);
1813
1814         if (copy_to_user(buf, values, size))
1815                 goto unlock;
1816
1817         ret = size;
1818
1819         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1820                 n = 0;
1821
1822                 values[n++] = perf_event_read_value(sub, &enabled, &running);
1823                 if (read_format & PERF_FORMAT_ID)
1824                         values[n++] = primary_event_id(sub);
1825
1826                 size = n * sizeof(u64);
1827
1828                 if (copy_to_user(buf + size, values, size)) {
1829                         ret = -EFAULT;
1830                         goto unlock;
1831                 }
1832
1833                 ret += size;
1834         }
1835 unlock:
1836         mutex_unlock(&ctx->mutex);
1837
1838         return ret;
1839 }
1840
1841 static int perf_event_read_one(struct perf_event *event,
1842                                  u64 read_format, char __user *buf)
1843 {
1844         u64 enabled, running;
1845         u64 values[4];
1846         int n = 0;
1847
1848         values[n++] = perf_event_read_value(event, &enabled, &running);
1849         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1850                 values[n++] = enabled;
1851         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1852                 values[n++] = running;
1853         if (read_format & PERF_FORMAT_ID)
1854                 values[n++] = primary_event_id(event);
1855
1856         if (copy_to_user(buf, values, n * sizeof(u64)))
1857                 return -EFAULT;
1858
1859         return n * sizeof(u64);
1860 }
1861
1862 /*
1863  * Read the performance event - simple non blocking version for now
1864  */
1865 static ssize_t
1866 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
1867 {
1868         u64 read_format = event->attr.read_format;
1869         int ret;
1870
1871         /*
1872          * Return end-of-file for a read on a event that is in
1873          * error state (i.e. because it was pinned but it couldn't be
1874          * scheduled on to the CPU at some point).
1875          */
1876         if (event->state == PERF_EVENT_STATE_ERROR)
1877                 return 0;
1878
1879         if (count < perf_event_read_size(event))
1880                 return -ENOSPC;
1881
1882         WARN_ON_ONCE(event->ctx->parent_ctx);
1883         if (read_format & PERF_FORMAT_GROUP)
1884                 ret = perf_event_read_group(event, read_format, buf);
1885         else
1886                 ret = perf_event_read_one(event, read_format, buf);
1887
1888         return ret;
1889 }
1890
1891 static ssize_t
1892 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1893 {
1894         struct perf_event *event = file->private_data;
1895
1896         return perf_read_hw(event, buf, count);
1897 }
1898
1899 static unsigned int perf_poll(struct file *file, poll_table *wait)
1900 {
1901         struct perf_event *event = file->private_data;
1902         struct perf_mmap_data *data;
1903         unsigned int events = POLL_HUP;
1904
1905         rcu_read_lock();
1906         data = rcu_dereference(event->data);
1907         if (data)
1908                 events = atomic_xchg(&data->poll, 0);
1909         rcu_read_unlock();
1910
1911         poll_wait(file, &event->waitq, wait);
1912
1913         return events;
1914 }
1915
1916 static void perf_event_reset(struct perf_event *event)
1917 {
1918         (void)perf_event_read(event);
1919         atomic64_set(&event->count, 0);
1920         perf_event_update_userpage(event);
1921 }
1922
1923 /*
1924  * Holding the top-level event's child_mutex means that any
1925  * descendant process that has inherited this event will block
1926  * in sync_child_event if it goes to exit, thus satisfying the
1927  * task existence requirements of perf_event_enable/disable.
1928  */
1929 static void perf_event_for_each_child(struct perf_event *event,
1930                                         void (*func)(struct perf_event *))
1931 {
1932         struct perf_event *child;
1933
1934         WARN_ON_ONCE(event->ctx->parent_ctx);
1935         mutex_lock(&event->child_mutex);
1936         func(event);
1937         list_for_each_entry(child, &event->child_list, child_list)
1938                 func(child);
1939         mutex_unlock(&event->child_mutex);
1940 }
1941
1942 static void perf_event_for_each(struct perf_event *event,
1943                                   void (*func)(struct perf_event *))
1944 {
1945         struct perf_event_context *ctx = event->ctx;
1946         struct perf_event *sibling;
1947
1948         WARN_ON_ONCE(ctx->parent_ctx);
1949         mutex_lock(&ctx->mutex);
1950         event = event->group_leader;
1951
1952         perf_event_for_each_child(event, func);
1953         func(event);
1954         list_for_each_entry(sibling, &event->sibling_list, group_entry)
1955                 perf_event_for_each_child(event, func);
1956         mutex_unlock(&ctx->mutex);
1957 }
1958
1959 static int perf_event_period(struct perf_event *event, u64 __user *arg)
1960 {
1961         struct perf_event_context *ctx = event->ctx;
1962         unsigned long size;
1963         int ret = 0;
1964         u64 value;
1965
1966         if (!event->attr.sample_period)
1967                 return -EINVAL;
1968
1969         size = copy_from_user(&value, arg, sizeof(value));
1970         if (size != sizeof(value))
1971                 return -EFAULT;
1972
1973         if (!value)
1974                 return -EINVAL;
1975
1976         spin_lock_irq(&ctx->lock);
1977         if (event->attr.freq) {
1978                 if (value > sysctl_perf_event_sample_rate) {
1979                         ret = -EINVAL;
1980                         goto unlock;
1981                 }
1982
1983                 event->attr.sample_freq = value;
1984         } else {
1985                 event->attr.sample_period = value;
1986                 event->hw.sample_period = value;
1987         }
1988 unlock:
1989         spin_unlock_irq(&ctx->lock);
1990
1991         return ret;
1992 }
1993
1994 static int perf_event_set_output(struct perf_event *event, int output_fd);
1995 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
1996
1997 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1998 {
1999         struct perf_event *event = file->private_data;
2000         void (*func)(struct perf_event *);
2001         u32 flags = arg;
2002
2003         switch (cmd) {
2004         case PERF_EVENT_IOC_ENABLE:
2005                 func = perf_event_enable;
2006                 break;
2007         case PERF_EVENT_IOC_DISABLE:
2008                 func = perf_event_disable;
2009                 break;
2010         case PERF_EVENT_IOC_RESET:
2011                 func = perf_event_reset;
2012                 break;
2013
2014         case PERF_EVENT_IOC_REFRESH:
2015                 return perf_event_refresh(event, arg);
2016
2017         case PERF_EVENT_IOC_PERIOD:
2018                 return perf_event_period(event, (u64 __user *)arg);
2019
2020         case PERF_EVENT_IOC_SET_OUTPUT:
2021                 return perf_event_set_output(event, arg);
2022
2023         case PERF_EVENT_IOC_SET_FILTER:
2024                 return perf_event_set_filter(event, (void __user *)arg);
2025
2026         default:
2027                 return -ENOTTY;
2028         }
2029
2030         if (flags & PERF_IOC_FLAG_GROUP)
2031                 perf_event_for_each(event, func);
2032         else
2033                 perf_event_for_each_child(event, func);
2034
2035         return 0;
2036 }
2037
2038 int perf_event_task_enable(void)
2039 {
2040         struct perf_event *event;
2041
2042         mutex_lock(&current->perf_event_mutex);
2043         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2044                 perf_event_for_each_child(event, perf_event_enable);
2045         mutex_unlock(&current->perf_event_mutex);
2046
2047         return 0;
2048 }
2049
2050 int perf_event_task_disable(void)
2051 {
2052         struct perf_event *event;
2053
2054         mutex_lock(&current->perf_event_mutex);
2055         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2056                 perf_event_for_each_child(event, perf_event_disable);
2057         mutex_unlock(&current->perf_event_mutex);
2058
2059         return 0;
2060 }
2061
2062 #ifndef PERF_EVENT_INDEX_OFFSET
2063 # define PERF_EVENT_INDEX_OFFSET 0
2064 #endif
2065
2066 static int perf_event_index(struct perf_event *event)
2067 {
2068         if (event->state != PERF_EVENT_STATE_ACTIVE)
2069                 return 0;
2070
2071         return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2072 }
2073
2074 /*
2075  * Callers need to ensure there can be no nesting of this function, otherwise
2076  * the seqlock logic goes bad. We can not serialize this because the arch
2077  * code calls this from NMI context.
2078  */
2079 void perf_event_update_userpage(struct perf_event *event)
2080 {
2081         struct perf_event_mmap_page *userpg;
2082         struct perf_mmap_data *data;
2083
2084         rcu_read_lock();
2085         data = rcu_dereference(event->data);
2086         if (!data)
2087                 goto unlock;
2088
2089         userpg = data->user_page;
2090
2091         /*
2092          * Disable preemption so as to not let the corresponding user-space
2093          * spin too long if we get preempted.
2094          */
2095         preempt_disable();
2096         ++userpg->lock;
2097         barrier();
2098         userpg->index = perf_event_index(event);
2099         userpg->offset = atomic64_read(&event->count);
2100         if (event->state == PERF_EVENT_STATE_ACTIVE)
2101                 userpg->offset -= atomic64_read(&event->hw.prev_count);
2102
2103         userpg->time_enabled = event->total_time_enabled +
2104                         atomic64_read(&event->child_total_time_enabled);
2105
2106         userpg->time_running = event->total_time_running +
2107                         atomic64_read(&event->child_total_time_running);
2108
2109         barrier();
2110         ++userpg->lock;
2111         preempt_enable();
2112 unlock:
2113         rcu_read_unlock();
2114 }
2115
2116 static unsigned long perf_data_size(struct perf_mmap_data *data)
2117 {
2118         return data->nr_pages << (PAGE_SHIFT + data->data_order);
2119 }
2120
2121 #ifndef CONFIG_PERF_USE_VMALLOC
2122
2123 /*
2124  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2125  */
2126
2127 static struct page *
2128 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2129 {
2130         if (pgoff > data->nr_pages)
2131                 return NULL;
2132
2133         if (pgoff == 0)
2134                 return virt_to_page(data->user_page);
2135
2136         return virt_to_page(data->data_pages[pgoff - 1]);
2137 }
2138
2139 static struct perf_mmap_data *
2140 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2141 {
2142         struct perf_mmap_data *data;
2143         unsigned long size;
2144         int i;
2145
2146         WARN_ON(atomic_read(&event->mmap_count));
2147
2148         size = sizeof(struct perf_mmap_data);
2149         size += nr_pages * sizeof(void *);
2150
2151         data = kzalloc(size, GFP_KERNEL);
2152         if (!data)
2153                 goto fail;
2154
2155         data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2156         if (!data->user_page)
2157                 goto fail_user_page;
2158
2159         for (i = 0; i < nr_pages; i++) {
2160                 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2161                 if (!data->data_pages[i])
2162                         goto fail_data_pages;
2163         }
2164
2165         data->data_order = 0;
2166         data->nr_pages = nr_pages;
2167
2168         return data;
2169
2170 fail_data_pages:
2171         for (i--; i >= 0; i--)
2172                 free_page((unsigned long)data->data_pages[i]);
2173
2174         free_page((unsigned long)data->user_page);
2175
2176 fail_user_page:
2177         kfree(data);
2178
2179 fail:
2180         return NULL;
2181 }
2182
2183 static void perf_mmap_free_page(unsigned long addr)
2184 {
2185         struct page *page = virt_to_page((void *)addr);
2186
2187         page->mapping = NULL;
2188         __free_page(page);
2189 }
2190
2191 static void perf_mmap_data_free(struct perf_mmap_data *data)
2192 {
2193         int i;
2194
2195         perf_mmap_free_page((unsigned long)data->user_page);
2196         for (i = 0; i < data->nr_pages; i++)
2197                 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2198 }
2199
2200 #else
2201
2202 /*
2203  * Back perf_mmap() with vmalloc memory.
2204  *
2205  * Required for architectures that have d-cache aliasing issues.
2206  */
2207
2208 static struct page *
2209 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2210 {
2211         if (pgoff > (1UL << data->data_order))
2212                 return NULL;
2213
2214         return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2215 }
2216
2217 static void perf_mmap_unmark_page(void *addr)
2218 {
2219         struct page *page = vmalloc_to_page(addr);
2220
2221         page->mapping = NULL;
2222 }
2223
2224 static void perf_mmap_data_free_work(struct work_struct *work)
2225 {
2226         struct perf_mmap_data *data;
2227         void *base;
2228         int i, nr;
2229
2230         data = container_of(work, struct perf_mmap_data, work);
2231         nr = 1 << data->data_order;
2232
2233         base = data->user_page;
2234         for (i = 0; i < nr + 1; i++)
2235                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2236
2237         vfree(base);
2238 }
2239
2240 static void perf_mmap_data_free(struct perf_mmap_data *data)
2241 {
2242         schedule_work(&data->work);
2243 }
2244
2245 static struct perf_mmap_data *
2246 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2247 {
2248         struct perf_mmap_data *data;
2249         unsigned long size;
2250         void *all_buf;
2251
2252         WARN_ON(atomic_read(&event->mmap_count));
2253
2254         size = sizeof(struct perf_mmap_data);
2255         size += sizeof(void *);
2256
2257         data = kzalloc(size, GFP_KERNEL);
2258         if (!data)
2259                 goto fail;
2260
2261         INIT_WORK(&data->work, perf_mmap_data_free_work);
2262
2263         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2264         if (!all_buf)
2265                 goto fail_all_buf;
2266
2267         data->user_page = all_buf;
2268         data->data_pages[0] = all_buf + PAGE_SIZE;
2269         data->data_order = ilog2(nr_pages);
2270         data->nr_pages = 1;
2271
2272         return data;
2273
2274 fail_all_buf:
2275         kfree(data);
2276
2277 fail:
2278         return NULL;
2279 }
2280
2281 #endif
2282
2283 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2284 {
2285         struct perf_event *event = vma->vm_file->private_data;
2286         struct perf_mmap_data *data;
2287         int ret = VM_FAULT_SIGBUS;
2288
2289         if (vmf->flags & FAULT_FLAG_MKWRITE) {
2290                 if (vmf->pgoff == 0)
2291                         ret = 0;
2292                 return ret;
2293         }
2294
2295         rcu_read_lock();
2296         data = rcu_dereference(event->data);
2297         if (!data)
2298                 goto unlock;
2299
2300         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2301                 goto unlock;
2302
2303         vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2304         if (!vmf->page)
2305                 goto unlock;
2306
2307         get_page(vmf->page);
2308         vmf->page->mapping = vma->vm_file->f_mapping;
2309         vmf->page->index   = vmf->pgoff;
2310
2311         ret = 0;
2312 unlock:
2313         rcu_read_unlock();
2314
2315         return ret;
2316 }
2317
2318 static void
2319 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2320 {
2321         long max_size = perf_data_size(data);
2322
2323         atomic_set(&data->lock, -1);
2324
2325         if (event->attr.watermark) {
2326                 data->watermark = min_t(long, max_size,
2327                                         event->attr.wakeup_watermark);
2328         }
2329
2330         if (!data->watermark)
2331                 data->watermark = max_size / 2;
2332
2333
2334         rcu_assign_pointer(event->data, data);
2335 }
2336
2337 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2338 {
2339         struct perf_mmap_data *data;
2340
2341         data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2342         perf_mmap_data_free(data);
2343         kfree(data);
2344 }
2345
2346 static void perf_mmap_data_release(struct perf_event *event)
2347 {
2348         struct perf_mmap_data *data = event->data;
2349
2350         WARN_ON(atomic_read(&event->mmap_count));
2351
2352         rcu_assign_pointer(event->data, NULL);
2353         call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2354 }
2355
2356 static void perf_mmap_open(struct vm_area_struct *vma)
2357 {
2358         struct perf_event *event = vma->vm_file->private_data;
2359
2360         atomic_inc(&event->mmap_count);
2361 }
2362
2363 static void perf_mmap_close(struct vm_area_struct *vma)
2364 {
2365         struct perf_event *event = vma->vm_file->private_data;
2366
2367         WARN_ON_ONCE(event->ctx->parent_ctx);
2368         if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2369                 unsigned long size = perf_data_size(event->data);
2370                 struct user_struct *user = current_user();
2371
2372                 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2373                 vma->vm_mm->locked_vm -= event->data->nr_locked;
2374                 perf_mmap_data_release(event);
2375                 mutex_unlock(&event->mmap_mutex);
2376         }
2377 }
2378
2379 static const struct vm_operations_struct perf_mmap_vmops = {
2380         .open           = perf_mmap_open,
2381         .close          = perf_mmap_close,
2382         .fault          = perf_mmap_fault,
2383         .page_mkwrite   = perf_mmap_fault,
2384 };
2385
2386 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2387 {
2388         struct perf_event *event = file->private_data;
2389         unsigned long user_locked, user_lock_limit;
2390         struct user_struct *user = current_user();
2391         unsigned long locked, lock_limit;
2392         struct perf_mmap_data *data;
2393         unsigned long vma_size;
2394         unsigned long nr_pages;
2395         long user_extra, extra;
2396         int ret = 0;
2397
2398         if (!(vma->vm_flags & VM_SHARED))
2399                 return -EINVAL;
2400
2401         vma_size = vma->vm_end - vma->vm_start;
2402         nr_pages = (vma_size / PAGE_SIZE) - 1;
2403
2404         /*
2405          * If we have data pages ensure they're a power-of-two number, so we
2406          * can do bitmasks instead of modulo.
2407          */
2408         if (nr_pages != 0 && !is_power_of_2(nr_pages))
2409                 return -EINVAL;
2410
2411         if (vma_size != PAGE_SIZE * (1 + nr_pages))
2412                 return -EINVAL;
2413
2414         if (vma->vm_pgoff != 0)
2415                 return -EINVAL;
2416
2417         WARN_ON_ONCE(event->ctx->parent_ctx);
2418         mutex_lock(&event->mmap_mutex);
2419         if (event->output) {
2420                 ret = -EINVAL;
2421                 goto unlock;
2422         }
2423
2424         if (atomic_inc_not_zero(&event->mmap_count)) {
2425                 if (nr_pages != event->data->nr_pages)
2426                         ret = -EINVAL;
2427                 goto unlock;
2428         }
2429
2430         user_extra = nr_pages + 1;
2431         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2432
2433         /*
2434          * Increase the limit linearly with more CPUs:
2435          */
2436         user_lock_limit *= num_online_cpus();
2437
2438         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2439
2440         extra = 0;
2441         if (user_locked > user_lock_limit)
2442                 extra = user_locked - user_lock_limit;
2443
2444         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2445         lock_limit >>= PAGE_SHIFT;
2446         locked = vma->vm_mm->locked_vm + extra;
2447
2448         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2449                 !capable(CAP_IPC_LOCK)) {
2450                 ret = -EPERM;
2451                 goto unlock;
2452         }
2453
2454         WARN_ON(event->data);
2455
2456         data = perf_mmap_data_alloc(event, nr_pages);
2457         ret = -ENOMEM;
2458         if (!data)
2459                 goto unlock;
2460
2461         ret = 0;
2462         perf_mmap_data_init(event, data);
2463
2464         atomic_set(&event->mmap_count, 1);
2465         atomic_long_add(user_extra, &user->locked_vm);
2466         vma->vm_mm->locked_vm += extra;
2467         event->data->nr_locked = extra;
2468         if (vma->vm_flags & VM_WRITE)
2469                 event->data->writable = 1;
2470
2471 unlock:
2472         mutex_unlock(&event->mmap_mutex);
2473
2474         vma->vm_flags |= VM_RESERVED;
2475         vma->vm_ops = &perf_mmap_vmops;
2476
2477         return ret;
2478 }
2479
2480 static int perf_fasync(int fd, struct file *filp, int on)
2481 {
2482         struct inode *inode = filp->f_path.dentry->d_inode;
2483         struct perf_event *event = filp->private_data;
2484         int retval;
2485
2486         mutex_lock(&inode->i_mutex);
2487         retval = fasync_helper(fd, filp, on, &event->fasync);
2488         mutex_unlock(&inode->i_mutex);
2489
2490         if (retval < 0)
2491                 return retval;
2492
2493         return 0;
2494 }
2495
2496 static const struct file_operations perf_fops = {
2497         .release                = perf_release,
2498         .read                   = perf_read,
2499         .poll                   = perf_poll,
2500         .unlocked_ioctl         = perf_ioctl,
2501         .compat_ioctl           = perf_ioctl,
2502         .mmap                   = perf_mmap,
2503         .fasync                 = perf_fasync,
2504 };
2505
2506 /*
2507  * Perf event wakeup
2508  *
2509  * If there's data, ensure we set the poll() state and publish everything
2510  * to user-space before waking everybody up.
2511  */
2512
2513 void perf_event_wakeup(struct perf_event *event)
2514 {
2515         wake_up_all(&event->waitq);
2516
2517         if (event->pending_kill) {
2518                 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2519                 event->pending_kill = 0;
2520         }
2521 }
2522
2523 /*
2524  * Pending wakeups
2525  *
2526  * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2527  *
2528  * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2529  * single linked list and use cmpxchg() to add entries lockless.
2530  */
2531
2532 static void perf_pending_event(struct perf_pending_entry *entry)
2533 {
2534         struct perf_event *event = container_of(entry,
2535                         struct perf_event, pending);
2536
2537         if (event->pending_disable) {
2538                 event->pending_disable = 0;
2539                 __perf_event_disable(event);
2540         }
2541
2542         if (event->pending_wakeup) {
2543                 event->pending_wakeup = 0;
2544                 perf_event_wakeup(event);
2545         }
2546 }
2547
2548 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2549
2550 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2551         PENDING_TAIL,
2552 };
2553
2554 static void perf_pending_queue(struct perf_pending_entry *entry,
2555                                void (*func)(struct perf_pending_entry *))
2556 {
2557         struct perf_pending_entry **head;
2558
2559         if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2560                 return;
2561
2562         entry->func = func;
2563
2564         head = &get_cpu_var(perf_pending_head);
2565
2566         do {
2567                 entry->next = *head;
2568         } while (cmpxchg(head, entry->next, entry) != entry->next);
2569
2570         set_perf_event_pending();
2571
2572         put_cpu_var(perf_pending_head);
2573 }
2574
2575 static int __perf_pending_run(void)
2576 {
2577         struct perf_pending_entry *list;
2578         int nr = 0;
2579
2580         list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2581         while (list != PENDING_TAIL) {
2582                 void (*func)(struct perf_pending_entry *);
2583                 struct perf_pending_entry *entry = list;
2584
2585                 list = list->next;
2586
2587                 func = entry->func;
2588                 entry->next = NULL;
2589                 /*
2590                  * Ensure we observe the unqueue before we issue the wakeup,
2591                  * so that we won't be waiting forever.
2592                  * -- see perf_not_pending().
2593                  */
2594                 smp_wmb();
2595
2596                 func(entry);
2597                 nr++;
2598         }
2599
2600         return nr;
2601 }
2602
2603 static inline int perf_not_pending(struct perf_event *event)
2604 {
2605         /*
2606          * If we flush on whatever cpu we run, there is a chance we don't
2607          * need to wait.
2608          */
2609         get_cpu();
2610         __perf_pending_run();
2611         put_cpu();
2612
2613         /*
2614          * Ensure we see the proper queue state before going to sleep
2615          * so that we do not miss the wakeup. -- see perf_pending_handle()
2616          */
2617         smp_rmb();
2618         return event->pending.next == NULL;
2619 }
2620
2621 static void perf_pending_sync(struct perf_event *event)
2622 {
2623         wait_event(event->waitq, perf_not_pending(event));
2624 }
2625
2626 void perf_event_do_pending(void)
2627 {
2628         __perf_pending_run();
2629 }
2630
2631 /*
2632  * Callchain support -- arch specific
2633  */
2634
2635 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2636 {
2637         return NULL;
2638 }
2639
2640 /*
2641  * Output
2642  */
2643 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2644                               unsigned long offset, unsigned long head)
2645 {
2646         unsigned long mask;
2647
2648         if (!data->writable)
2649                 return true;
2650
2651         mask = perf_data_size(data) - 1;
2652
2653         offset = (offset - tail) & mask;
2654         head   = (head   - tail) & mask;
2655
2656         if ((int)(head - offset) < 0)
2657                 return false;
2658
2659         return true;
2660 }
2661
2662 static void perf_output_wakeup(struct perf_output_handle *handle)
2663 {
2664         atomic_set(&handle->data->poll, POLL_IN);
2665
2666         if (handle->nmi) {
2667                 handle->event->pending_wakeup = 1;
2668                 perf_pending_queue(&handle->event->pending,
2669                                    perf_pending_event);
2670         } else
2671                 perf_event_wakeup(handle->event);
2672 }
2673
2674 /*
2675  * Curious locking construct.
2676  *
2677  * We need to ensure a later event_id doesn't publish a head when a former
2678  * event_id isn't done writing. However since we need to deal with NMIs we
2679  * cannot fully serialize things.
2680  *
2681  * What we do is serialize between CPUs so we only have to deal with NMI
2682  * nesting on a single CPU.
2683  *
2684  * We only publish the head (and generate a wakeup) when the outer-most
2685  * event_id completes.
2686  */
2687 static void perf_output_lock(struct perf_output_handle *handle)
2688 {
2689         struct perf_mmap_data *data = handle->data;
2690         int cur, cpu = get_cpu();
2691
2692         handle->locked = 0;
2693
2694         for (;;) {
2695                 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2696                 if (cur == -1) {
2697                         handle->locked = 1;
2698                         break;
2699                 }
2700                 if (cur == cpu)
2701                         break;
2702
2703                 cpu_relax();
2704         }
2705 }
2706
2707 static void perf_output_unlock(struct perf_output_handle *handle)
2708 {
2709         struct perf_mmap_data *data = handle->data;
2710         unsigned long head;
2711         int cpu;
2712
2713         data->done_head = data->head;
2714
2715         if (!handle->locked)
2716                 goto out;
2717
2718 again:
2719         /*
2720          * The xchg implies a full barrier that ensures all writes are done
2721          * before we publish the new head, matched by a rmb() in userspace when
2722          * reading this position.
2723          */
2724         while ((head = atomic_long_xchg(&data->done_head, 0)))
2725                 data->user_page->data_head = head;
2726
2727         /*
2728          * NMI can happen here, which means we can miss a done_head update.
2729          */
2730
2731         cpu = atomic_xchg(&data->lock, -1);
2732         WARN_ON_ONCE(cpu != smp_processor_id());
2733
2734         /*
2735          * Therefore we have to validate we did not indeed do so.
2736          */
2737         if (unlikely(atomic_long_read(&data->done_head))) {
2738                 /*
2739                  * Since we had it locked, we can lock it again.
2740                  */
2741                 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2742                         cpu_relax();
2743
2744                 goto again;
2745         }
2746
2747         if (atomic_xchg(&data->wakeup, 0))
2748                 perf_output_wakeup(handle);
2749 out:
2750         put_cpu();
2751 }
2752
2753 void perf_output_copy(struct perf_output_handle *handle,
2754                       const void *buf, unsigned int len)
2755 {
2756         unsigned int pages_mask;
2757         unsigned long offset;
2758         unsigned int size;
2759         void **pages;
2760
2761         offset          = handle->offset;
2762         pages_mask      = handle->data->nr_pages - 1;
2763         pages           = handle->data->data_pages;
2764
2765         do {
2766                 unsigned long page_offset;
2767                 unsigned long page_size;
2768                 int nr;
2769
2770                 nr          = (offset >> PAGE_SHIFT) & pages_mask;
2771                 page_size   = 1UL << (handle->data->data_order + PAGE_SHIFT);
2772                 page_offset = offset & (page_size - 1);
2773                 size        = min_t(unsigned int, page_size - page_offset, len);
2774
2775                 memcpy(pages[nr] + page_offset, buf, size);
2776
2777                 len         -= size;
2778                 buf         += size;
2779                 offset      += size;
2780         } while (len);
2781
2782         handle->offset = offset;
2783
2784         /*
2785          * Check we didn't copy past our reservation window, taking the
2786          * possible unsigned int wrap into account.
2787          */
2788         WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2789 }
2790
2791 int perf_output_begin(struct perf_output_handle *handle,
2792                       struct perf_event *event, unsigned int size,
2793                       int nmi, int sample)
2794 {
2795         struct perf_event *output_event;
2796         struct perf_mmap_data *data;
2797         unsigned long tail, offset, head;
2798         int have_lost;
2799         struct {
2800                 struct perf_event_header header;
2801                 u64                      id;
2802                 u64                      lost;
2803         } lost_event;
2804
2805         rcu_read_lock();
2806         /*
2807          * For inherited events we send all the output towards the parent.
2808          */
2809         if (event->parent)
2810                 event = event->parent;
2811
2812         output_event = rcu_dereference(event->output);
2813         if (output_event)
2814                 event = output_event;
2815
2816         data = rcu_dereference(event->data);
2817         if (!data)
2818                 goto out;
2819
2820         handle->data    = data;
2821         handle->event   = event;
2822         handle->nmi     = nmi;
2823         handle->sample  = sample;
2824
2825         if (!data->nr_pages)
2826                 goto fail;
2827
2828         have_lost = atomic_read(&data->lost);
2829         if (have_lost)
2830                 size += sizeof(lost_event);
2831
2832         perf_output_lock(handle);
2833
2834         do {
2835                 /*
2836                  * Userspace could choose to issue a mb() before updating the
2837                  * tail pointer. So that all reads will be completed before the
2838                  * write is issued.
2839                  */
2840                 tail = ACCESS_ONCE(data->user_page->data_tail);
2841                 smp_rmb();
2842                 offset = head = atomic_long_read(&data->head);
2843                 head += size;
2844                 if (unlikely(!perf_output_space(data, tail, offset, head)))
2845                         goto fail;
2846         } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2847
2848         handle->offset  = offset;
2849         handle->head    = head;
2850
2851         if (head - tail > data->watermark)
2852                 atomic_set(&data->wakeup, 1);
2853
2854         if (have_lost) {
2855                 lost_event.header.type = PERF_RECORD_LOST;
2856                 lost_event.header.misc = 0;
2857                 lost_event.header.size = sizeof(lost_event);
2858                 lost_event.id          = event->id;
2859                 lost_event.lost        = atomic_xchg(&data->lost, 0);
2860
2861                 perf_output_put(handle, lost_event);
2862         }
2863
2864         return 0;
2865
2866 fail:
2867         atomic_inc(&data->lost);
2868         perf_output_unlock(handle);
2869 out:
2870         rcu_read_unlock();
2871
2872         return -ENOSPC;
2873 }
2874
2875 void perf_output_end(struct perf_output_handle *handle)
2876 {
2877         struct perf_event *event = handle->event;
2878         struct perf_mmap_data *data = handle->data;
2879
2880         int wakeup_events = event->attr.wakeup_events;
2881
2882         if (handle->sample && wakeup_events) {
2883                 int events = atomic_inc_return(&data->events);
2884                 if (events >= wakeup_events) {
2885                         atomic_sub(wakeup_events, &data->events);
2886                         atomic_set(&data->wakeup, 1);
2887                 }
2888         }
2889
2890         perf_output_unlock(handle);
2891         rcu_read_unlock();
2892 }
2893
2894 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
2895 {
2896         /*
2897          * only top level events have the pid namespace they were created in
2898          */
2899         if (event->parent)
2900                 event = event->parent;
2901
2902         return task_tgid_nr_ns(p, event->ns);
2903 }
2904
2905 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
2906 {
2907         /*
2908          * only top level events have the pid namespace they were created in
2909          */
2910         if (event->parent)
2911                 event = event->parent;
2912
2913         return task_pid_nr_ns(p, event->ns);
2914 }
2915
2916 static void perf_output_read_one(struct perf_output_handle *handle,
2917                                  struct perf_event *event)
2918 {
2919         u64 read_format = event->attr.read_format;
2920         u64 values[4];
2921         int n = 0;
2922
2923         values[n++] = atomic64_read(&event->count);
2924         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2925                 values[n++] = event->total_time_enabled +
2926                         atomic64_read(&event->child_total_time_enabled);
2927         }
2928         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2929                 values[n++] = event->total_time_running +
2930                         atomic64_read(&event->child_total_time_running);
2931         }
2932         if (read_format & PERF_FORMAT_ID)
2933                 values[n++] = primary_event_id(event);
2934
2935         perf_output_copy(handle, values, n * sizeof(u64));
2936 }
2937
2938 /*
2939  * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
2940  */
2941 static void perf_output_read_group(struct perf_output_handle *handle,
2942                             struct perf_event *event)
2943 {
2944         struct perf_event *leader = event->group_leader, *sub;
2945         u64 read_format = event->attr.read_format;
2946         u64 values[5];
2947         int n = 0;
2948
2949         values[n++] = 1 + leader->nr_siblings;
2950
2951         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2952                 values[n++] = leader->total_time_enabled;
2953
2954         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2955                 values[n++] = leader->total_time_running;
2956
2957         if (leader != event)
2958                 leader->pmu->read(leader);
2959
2960         values[n++] = atomic64_read(&leader->count);
2961         if (read_format & PERF_FORMAT_ID)
2962                 values[n++] = primary_event_id(leader);
2963
2964         perf_output_copy(handle, values, n * sizeof(u64));
2965
2966         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2967                 n = 0;
2968
2969                 if (sub != event)
2970                         sub->pmu->read(sub);
2971
2972                 values[n++] = atomic64_read(&sub->count);
2973                 if (read_format & PERF_FORMAT_ID)
2974                         values[n++] = primary_event_id(sub);
2975
2976                 perf_output_copy(handle, values, n * sizeof(u64));
2977         }
2978 }
2979
2980 static void perf_output_read(struct perf_output_handle *handle,
2981                              struct perf_event *event)
2982 {
2983         if (event->attr.read_format & PERF_FORMAT_GROUP)
2984                 perf_output_read_group(handle, event);
2985         else
2986                 perf_output_read_one(handle, event);
2987 }
2988
2989 void perf_output_sample(struct perf_output_handle *handle,
2990                         struct perf_event_header *header,
2991                         struct perf_sample_data *data,
2992                         struct perf_event *event)
2993 {
2994         u64 sample_type = data->type;
2995
2996         perf_output_put(handle, *header);
2997
2998         if (sample_type & PERF_SAMPLE_IP)
2999                 perf_output_put(handle, data->ip);
3000
3001         if (sample_type & PERF_SAMPLE_TID)
3002                 perf_output_put(handle, data->tid_entry);
3003
3004         if (sample_type & PERF_SAMPLE_TIME)
3005                 perf_output_put(handle, data->time);
3006
3007         if (sample_type & PERF_SAMPLE_ADDR)
3008                 perf_output_put(handle, data->addr);
3009
3010         if (sample_type & PERF_SAMPLE_ID)
3011                 perf_output_put(handle, data->id);
3012
3013         if (sample_type & PERF_SAMPLE_STREAM_ID)
3014                 perf_output_put(handle, data->stream_id);
3015
3016         if (sample_type & PERF_SAMPLE_CPU)
3017                 perf_output_put(handle, data->cpu_entry);
3018
3019         if (sample_type & PERF_SAMPLE_PERIOD)
3020                 perf_output_put(handle, data->period);
3021
3022         if (sample_type & PERF_SAMPLE_READ)
3023                 perf_output_read(handle, event);
3024
3025         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3026                 if (data->callchain) {
3027                         int size = 1;
3028
3029                         if (data->callchain)
3030                                 size += data->callchain->nr;
3031
3032                         size *= sizeof(u64);
3033
3034                         perf_output_copy(handle, data->callchain, size);
3035                 } else {
3036                         u64 nr = 0;
3037                         perf_output_put(handle, nr);
3038                 }
3039         }
3040
3041         if (sample_type & PERF_SAMPLE_RAW) {
3042                 if (data->raw) {
3043                         perf_output_put(handle, data->raw->size);
3044                         perf_output_copy(handle, data->raw->data,
3045                                          data->raw->size);
3046                 } else {
3047                         struct {
3048                                 u32     size;
3049                                 u32     data;
3050                         } raw = {
3051                                 .size = sizeof(u32),
3052                                 .data = 0,
3053                         };
3054                         perf_output_put(handle, raw);
3055                 }
3056         }
3057 }
3058
3059 void perf_prepare_sample(struct perf_event_header *header,
3060                          struct perf_sample_data *data,
3061                          struct perf_event *event,
3062                          struct pt_regs *regs)
3063 {
3064         u64 sample_type = event->attr.sample_type;
3065
3066         data->type = sample_type;
3067
3068         header->type = PERF_RECORD_SAMPLE;
3069         header->size = sizeof(*header);
3070
3071         header->misc = 0;
3072         header->misc |= perf_misc_flags(regs);
3073
3074         if (sample_type & PERF_SAMPLE_IP) {
3075                 data->ip = perf_instruction_pointer(regs);
3076
3077                 header->size += sizeof(data->ip);
3078         }
3079
3080         if (sample_type & PERF_SAMPLE_TID) {
3081                 /* namespace issues */
3082                 data->tid_entry.pid = perf_event_pid(event, current);
3083                 data->tid_entry.tid = perf_event_tid(event, current);
3084
3085                 header->size += sizeof(data->tid_entry);
3086         }
3087
3088         if (sample_type & PERF_SAMPLE_TIME) {
3089                 data->time = perf_clock();
3090
3091                 header->size += sizeof(data->time);
3092         }
3093
3094         if (sample_type & PERF_SAMPLE_ADDR)
3095                 header->size += sizeof(data->addr);
3096
3097         if (sample_type & PERF_SAMPLE_ID) {
3098                 data->id = primary_event_id(event);
3099
3100                 header->size += sizeof(data->id);
3101         }
3102
3103         if (sample_type & PERF_SAMPLE_STREAM_ID) {
3104                 data->stream_id = event->id;
3105
3106                 header->size += sizeof(data->stream_id);
3107         }
3108
3109         if (sample_type & PERF_SAMPLE_CPU) {
3110                 data->cpu_entry.cpu             = raw_smp_processor_id();
3111                 data->cpu_entry.reserved        = 0;
3112
3113                 header->size += sizeof(data->cpu_entry);
3114         }
3115
3116         if (sample_type & PERF_SAMPLE_PERIOD)
3117                 header->size += sizeof(data->period);
3118
3119         if (sample_type & PERF_SAMPLE_READ)
3120                 header->size += perf_event_read_size(event);
3121
3122         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3123                 int size = 1;
3124
3125                 data->callchain = perf_callchain(regs);
3126
3127                 if (data->callchain)
3128                         size += data->callchain->nr;
3129
3130                 header->size += size * sizeof(u64);
3131         }
3132
3133         if (sample_type & PERF_SAMPLE_RAW) {
3134                 int size = sizeof(u32);
3135
3136                 if (data->raw)
3137                         size += data->raw->size;
3138                 else
3139                         size += sizeof(u32);
3140
3141                 WARN_ON_ONCE(size & (sizeof(u64)-1));
3142                 header->size += size;
3143         }
3144 }
3145
3146 static void perf_event_output(struct perf_event *event, int nmi,
3147                                 struct perf_sample_data *data,
3148                                 struct pt_regs *regs)
3149 {
3150         struct perf_output_handle handle;
3151         struct perf_event_header header;
3152
3153         perf_prepare_sample(&header, data, event, regs);
3154
3155         if (perf_output_begin(&handle, event, header.size, nmi, 1))
3156                 return;
3157
3158         perf_output_sample(&handle, &header, data, event);
3159
3160         perf_output_end(&handle);
3161 }
3162
3163 /*
3164  * read event_id
3165  */
3166
3167 struct perf_read_event {
3168         struct perf_event_header        header;
3169
3170         u32                             pid;
3171         u32                             tid;
3172 };
3173
3174 static void
3175 perf_event_read_event(struct perf_event *event,
3176                         struct task_struct *task)
3177 {
3178         struct perf_output_handle handle;
3179         struct perf_read_event read_event = {
3180                 .header = {
3181                         .type = PERF_RECORD_READ,
3182                         .misc = 0,
3183                         .size = sizeof(read_event) + perf_event_read_size(event),
3184                 },
3185                 .pid = perf_event_pid(event, task),
3186                 .tid = perf_event_tid(event, task),
3187         };
3188         int ret;
3189
3190         ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3191         if (ret)
3192                 return;
3193
3194         perf_output_put(&handle, read_event);
3195         perf_output_read(&handle, event);
3196
3197         perf_output_end(&handle);
3198 }
3199
3200 /*
3201  * task tracking -- fork/exit
3202  *
3203  * enabled by: attr.comm | attr.mmap | attr.task
3204  */
3205
3206 struct perf_task_event {
3207         struct task_struct              *task;
3208         struct perf_event_context       *task_ctx;
3209
3210         struct {
3211                 struct perf_event_header        header;
3212
3213                 u32                             pid;
3214                 u32                             ppid;
3215                 u32                             tid;
3216                 u32                             ptid;
3217                 u64                             time;
3218         } event_id;
3219 };
3220
3221 static void perf_event_task_output(struct perf_event *event,
3222                                      struct perf_task_event *task_event)
3223 {
3224         struct perf_output_handle handle;
3225         int size;
3226         struct task_struct *task = task_event->task;
3227         int ret;
3228
3229         size  = task_event->event_id.header.size;
3230         ret = perf_output_begin(&handle, event, size, 0, 0);
3231
3232         if (ret)
3233                 return;
3234
3235         task_event->event_id.pid = perf_event_pid(event, task);
3236         task_event->event_id.ppid = perf_event_pid(event, current);
3237
3238         task_event->event_id.tid = perf_event_tid(event, task);
3239         task_event->event_id.ptid = perf_event_tid(event, current);
3240
3241         task_event->event_id.time = perf_clock();
3242
3243         perf_output_put(&handle, task_event->event_id);
3244
3245         perf_output_end(&handle);
3246 }
3247
3248 static int perf_event_task_match(struct perf_event *event)
3249 {
3250         if (event->attr.comm || event->attr.mmap || event->attr.task)
3251                 return 1;
3252
3253         return 0;
3254 }
3255
3256 static void perf_event_task_ctx(struct perf_event_context *ctx,
3257                                   struct perf_task_event *task_event)
3258 {
3259         struct perf_event *event;
3260
3261         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3262                 if (perf_event_task_match(event))
3263                         perf_event_task_output(event, task_event);
3264         }
3265 }
3266
3267 static void perf_event_task_event(struct perf_task_event *task_event)
3268 {
3269         struct perf_cpu_context *cpuctx;
3270         struct perf_event_context *ctx = task_event->task_ctx;
3271
3272         rcu_read_lock();
3273         cpuctx = &get_cpu_var(perf_cpu_context);
3274         perf_event_task_ctx(&cpuctx->ctx, task_event);
3275         put_cpu_var(perf_cpu_context);
3276
3277         if (!ctx)
3278                 ctx = rcu_dereference(task_event->task->perf_event_ctxp);
3279         if (ctx)
3280                 perf_event_task_ctx(ctx, task_event);
3281         rcu_read_unlock();
3282 }
3283
3284 static void perf_event_task(struct task_struct *task,
3285                               struct perf_event_context *task_ctx,
3286                               int new)
3287 {
3288         struct perf_task_event task_event;
3289
3290         if (!atomic_read(&nr_comm_events) &&
3291             !atomic_read(&nr_mmap_events) &&
3292             !atomic_read(&nr_task_events))
3293                 return;
3294
3295         task_event = (struct perf_task_event){
3296                 .task     = task,
3297                 .task_ctx = task_ctx,
3298                 .event_id    = {
3299                         .header = {
3300                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3301                                 .misc = 0,
3302                                 .size = sizeof(task_event.event_id),
3303                         },
3304                         /* .pid  */
3305                         /* .ppid */
3306                         /* .tid  */
3307                         /* .ptid */
3308                 },
3309         };
3310
3311         perf_event_task_event(&task_event);
3312 }
3313
3314 void perf_event_fork(struct task_struct *task)
3315 {
3316         perf_event_task(task, NULL, 1);
3317 }
3318
3319 /*
3320  * comm tracking
3321  */
3322
3323 struct perf_comm_event {
3324         struct task_struct      *task;
3325         char                    *comm;
3326         int                     comm_size;
3327
3328         struct {
3329                 struct perf_event_header        header;
3330
3331                 u32                             pid;
3332                 u32                             tid;
3333         } event_id;
3334 };
3335
3336 static void perf_event_comm_output(struct perf_event *event,
3337                                      struct perf_comm_event *comm_event)
3338 {
3339         struct perf_output_handle handle;
3340         int size = comm_event->event_id.header.size;
3341         int ret = perf_output_begin(&handle, event, size, 0, 0);
3342
3343         if (ret)
3344                 return;
3345
3346         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3347         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3348
3349         perf_output_put(&handle, comm_event->event_id);
3350         perf_output_copy(&handle, comm_event->comm,
3351                                    comm_event->comm_size);
3352         perf_output_end(&handle);
3353 }
3354
3355 static int perf_event_comm_match(struct perf_event *event)
3356 {
3357         if (event->attr.comm)
3358                 return 1;
3359
3360         return 0;
3361 }
3362
3363 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3364                                   struct perf_comm_event *comm_event)
3365 {
3366         struct perf_event *event;
3367
3368         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3369                 if (perf_event_comm_match(event))
3370                         perf_event_comm_output(event, comm_event);
3371         }
3372 }
3373
3374 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3375 {
3376         struct perf_cpu_context *cpuctx;
3377         struct perf_event_context *ctx;
3378         unsigned int size;
3379         char comm[TASK_COMM_LEN];
3380
3381         memset(comm, 0, sizeof(comm));
3382         strlcpy(comm, comm_event->task->comm, sizeof(comm));
3383         size = ALIGN(strlen(comm)+1, sizeof(u64));
3384
3385         comm_event->comm = comm;
3386         comm_event->comm_size = size;
3387
3388         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3389
3390         rcu_read_lock();
3391         cpuctx = &get_cpu_var(perf_cpu_context);
3392         perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3393         put_cpu_var(perf_cpu_context);
3394
3395         /*
3396          * doesn't really matter which of the child contexts the
3397          * events ends up in.
3398          */
3399         ctx = rcu_dereference(current->perf_event_ctxp);
3400         if (ctx)
3401                 perf_event_comm_ctx(ctx, comm_event);
3402         rcu_read_unlock();
3403 }
3404
3405 void perf_event_comm(struct task_struct *task)
3406 {
3407         struct perf_comm_event comm_event;
3408
3409         if (task->perf_event_ctxp)
3410                 perf_event_enable_on_exec(task);
3411
3412         if (!atomic_read(&nr_comm_events))
3413                 return;
3414
3415         comm_event = (struct perf_comm_event){
3416                 .task   = task,
3417                 /* .comm      */
3418                 /* .comm_size */
3419                 .event_id  = {
3420                         .header = {
3421                                 .type = PERF_RECORD_COMM,
3422                                 .misc = 0,
3423                                 /* .size */
3424                         },
3425                         /* .pid */
3426                         /* .tid */
3427                 },
3428         };
3429
3430         perf_event_comm_event(&comm_event);
3431 }
3432
3433 /*
3434  * mmap tracking
3435  */
3436
3437 struct perf_mmap_event {
3438         struct vm_area_struct   *vma;
3439
3440         const char              *file_name;
3441         int                     file_size;
3442
3443         struct {
3444                 struct perf_event_header        header;
3445
3446                 u32                             pid;
3447                 u32                             tid;
3448                 u64                             start;
3449                 u64                             len;
3450                 u64                             pgoff;
3451         } event_id;
3452 };
3453
3454 static void perf_event_mmap_output(struct perf_event *event,
3455                                      struct perf_mmap_event *mmap_event)
3456 {
3457         struct perf_output_handle handle;
3458         int size = mmap_event->event_id.header.size;
3459         int ret = perf_output_begin(&handle, event, size, 0, 0);
3460
3461         if (ret)
3462                 return;
3463
3464         mmap_event->event_id.pid = perf_event_pid(event, current);
3465         mmap_event->event_id.tid = perf_event_tid(event, current);
3466
3467         perf_output_put(&handle, mmap_event->event_id);
3468         perf_output_copy(&handle, mmap_event->file_name,
3469                                    mmap_event->file_size);
3470         perf_output_end(&handle);
3471 }
3472
3473 static int perf_event_mmap_match(struct perf_event *event,
3474                                    struct perf_mmap_event *mmap_event)
3475 {
3476         if (event->attr.mmap)
3477                 return 1;
3478
3479         return 0;
3480 }
3481
3482 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3483                                   struct perf_mmap_event *mmap_event)
3484 {
3485         struct perf_event *event;
3486
3487         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3488                 if (perf_event_mmap_match(event, mmap_event))
3489                         perf_event_mmap_output(event, mmap_event);
3490         }
3491 }
3492
3493 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3494 {
3495         struct perf_cpu_context *cpuctx;
3496         struct perf_event_context *ctx;
3497         struct vm_area_struct *vma = mmap_event->vma;
3498         struct file *file = vma->vm_file;
3499         unsigned int size;
3500         char tmp[16];
3501         char *buf = NULL;
3502         const char *name;
3503
3504         memset(tmp, 0, sizeof(tmp));
3505
3506         if (file) {
3507                 /*
3508                  * d_path works from the end of the buffer backwards, so we
3509                  * need to add enough zero bytes after the string to handle
3510                  * the 64bit alignment we do later.
3511                  */
3512                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3513                 if (!buf) {
3514                         name = strncpy(tmp, "//enomem", sizeof(tmp));
3515                         goto got_name;
3516                 }
3517                 name = d_path(&file->f_path, buf, PATH_MAX);
3518                 if (IS_ERR(name)) {
3519                         name = strncpy(tmp, "//toolong", sizeof(tmp));
3520                         goto got_name;
3521                 }
3522         } else {
3523                 if (arch_vma_name(mmap_event->vma)) {
3524                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3525                                        sizeof(tmp));
3526                         goto got_name;
3527                 }
3528
3529                 if (!vma->vm_mm) {
3530                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
3531                         goto got_name;
3532                 }
3533
3534                 name = strncpy(tmp, "//anon", sizeof(tmp));
3535                 goto got_name;
3536         }
3537
3538 got_name:
3539         size = ALIGN(strlen(name)+1, sizeof(u64));
3540
3541         mmap_event->file_name = name;
3542         mmap_event->file_size = size;
3543
3544         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3545
3546         rcu_read_lock();
3547         cpuctx = &get_cpu_var(perf_cpu_context);
3548         perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3549         put_cpu_var(perf_cpu_context);
3550
3551         /*
3552          * doesn't really matter which of the child contexts the
3553          * events ends up in.
3554          */
3555         ctx = rcu_dereference(current->perf_event_ctxp);
3556         if (ctx)
3557                 perf_event_mmap_ctx(ctx, mmap_event);
3558         rcu_read_unlock();
3559
3560         kfree(buf);
3561 }
3562
3563 void __perf_event_mmap(struct vm_area_struct *vma)
3564 {
3565         struct perf_mmap_event mmap_event;
3566
3567         if (!atomic_read(&nr_mmap_events))
3568                 return;
3569
3570         mmap_event = (struct perf_mmap_event){
3571                 .vma    = vma,
3572                 /* .file_name */
3573                 /* .file_size */
3574                 .event_id  = {
3575                         .header = {
3576                                 .type = PERF_RECORD_MMAP,
3577                                 .misc = 0,
3578                                 /* .size */
3579                         },
3580                         /* .pid */
3581                         /* .tid */
3582                         .start  = vma->vm_start,
3583                         .len    = vma->vm_end - vma->vm_start,
3584                         .pgoff  = vma->vm_pgoff,
3585                 },
3586         };
3587
3588         perf_event_mmap_event(&mmap_event);
3589 }
3590
3591 /*
3592  * IRQ throttle logging
3593  */
3594
3595 static void perf_log_throttle(struct perf_event *event, int enable)
3596 {
3597         struct perf_output_handle handle;
3598         int ret;
3599
3600         struct {
3601                 struct perf_event_header        header;
3602                 u64                             time;
3603                 u64                             id;
3604                 u64                             stream_id;
3605         } throttle_event = {
3606                 .header = {
3607                         .type = PERF_RECORD_THROTTLE,
3608                         .misc = 0,
3609                         .size = sizeof(throttle_event),
3610                 },
3611                 .time           = perf_clock(),
3612                 .id             = primary_event_id(event),
3613                 .stream_id      = event->id,
3614         };
3615
3616         if (enable)
3617                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3618
3619         ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3620         if (ret)
3621                 return;
3622
3623         perf_output_put(&handle, throttle_event);
3624         perf_output_end(&handle);
3625 }
3626
3627 /*
3628  * Generic event overflow handling, sampling.
3629  */
3630
3631 static int __perf_event_overflow(struct perf_event *event, int nmi,
3632                                    int throttle, struct perf_sample_data *data,
3633                                    struct pt_regs *regs)
3634 {
3635         int events = atomic_read(&event->event_limit);
3636         struct hw_perf_event *hwc = &event->hw;
3637         int ret = 0;
3638
3639         throttle = (throttle && event->pmu->unthrottle != NULL);
3640
3641         if (!throttle) {
3642                 hwc->interrupts++;
3643         } else {
3644                 if (hwc->interrupts != MAX_INTERRUPTS) {
3645                         hwc->interrupts++;
3646                         if (HZ * hwc->interrupts >
3647                                         (u64)sysctl_perf_event_sample_rate) {
3648                                 hwc->interrupts = MAX_INTERRUPTS;
3649                                 perf_log_throttle(event, 0);
3650                                 ret = 1;
3651                         }
3652                 } else {
3653                         /*
3654                          * Keep re-disabling events even though on the previous
3655                          * pass we disabled it - just in case we raced with a
3656                          * sched-in and the event got enabled again:
3657                          */
3658                         ret = 1;
3659                 }
3660         }
3661
3662         if (event->attr.freq) {
3663                 u64 now = perf_clock();
3664                 s64 delta = now - hwc->freq_stamp;
3665
3666                 hwc->freq_stamp = now;
3667
3668                 if (delta > 0 && delta < TICK_NSEC)
3669                         perf_adjust_period(event, NSEC_PER_SEC / (int)delta);
3670         }
3671
3672         /*
3673          * XXX event_limit might not quite work as expected on inherited
3674          * events
3675          */
3676
3677         event->pending_kill = POLL_IN;
3678         if (events && atomic_dec_and_test(&event->event_limit)) {
3679                 ret = 1;
3680                 event->pending_kill = POLL_HUP;
3681                 if (nmi) {
3682                         event->pending_disable = 1;
3683                         perf_pending_queue(&event->pending,
3684                                            perf_pending_event);
3685                 } else
3686                         perf_event_disable(event);
3687         }
3688
3689         if (event->overflow_handler)
3690                 event->overflow_handler(event, nmi, data, regs);
3691         else
3692                 perf_event_output(event, nmi, data, regs);
3693
3694         return ret;
3695 }
3696
3697 int perf_event_overflow(struct perf_event *event, int nmi,
3698                           struct perf_sample_data *data,
3699                           struct pt_regs *regs)
3700 {
3701         return __perf_event_overflow(event, nmi, 1, data, regs);
3702 }
3703
3704 /*
3705  * Generic software event infrastructure
3706  */
3707
3708 /*
3709  * We directly increment event->count and keep a second value in
3710  * event->hw.period_left to count intervals. This period event
3711  * is kept in the range [-sample_period, 0] so that we can use the
3712  * sign as trigger.
3713  */
3714
3715 static u64 perf_swevent_set_period(struct perf_event *event)
3716 {
3717         struct hw_perf_event *hwc = &event->hw;
3718         u64 period = hwc->last_period;
3719         u64 nr, offset;
3720         s64 old, val;
3721
3722         hwc->last_period = hwc->sample_period;
3723
3724 again:
3725         old = val = atomic64_read(&hwc->period_left);
3726         if (val < 0)
3727                 return 0;
3728
3729         nr = div64_u64(period + val, period);
3730         offset = nr * period;
3731         val -= offset;
3732         if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3733                 goto again;
3734
3735         return nr;
3736 }
3737
3738 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3739                                     int nmi, struct perf_sample_data *data,
3740                                     struct pt_regs *regs)
3741 {
3742         struct hw_perf_event *hwc = &event->hw;
3743         int throttle = 0;
3744
3745         data->period = event->hw.last_period;
3746         if (!overflow)
3747                 overflow = perf_swevent_set_period(event);
3748
3749         if (hwc->interrupts == MAX_INTERRUPTS)
3750                 return;
3751
3752         for (; overflow; overflow--) {
3753                 if (__perf_event_overflow(event, nmi, throttle,
3754                                             data, regs)) {
3755                         /*
3756                          * We inhibit the overflow from happening when
3757                          * hwc->interrupts == MAX_INTERRUPTS.
3758                          */
3759                         break;
3760                 }
3761                 throttle = 1;
3762         }
3763 }
3764
3765 static void perf_swevent_unthrottle(struct perf_event *event)
3766 {
3767         /*
3768          * Nothing to do, we already reset hwc->interrupts.
3769          */
3770 }
3771
3772 static void perf_swevent_add(struct perf_event *event, u64 nr,
3773                                int nmi, struct perf_sample_data *data,
3774                                struct pt_regs *regs)
3775 {
3776         struct hw_perf_event *hwc = &event->hw;
3777
3778         atomic64_add(nr, &event->count);
3779
3780         if (!regs)
3781                 return;
3782
3783         if (!hwc->sample_period)
3784                 return;
3785
3786         if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3787                 return perf_swevent_overflow(event, 1, nmi, data, regs);
3788
3789         if (atomic64_add_negative(nr, &hwc->period_left))
3790                 return;
3791
3792         perf_swevent_overflow(event, 0, nmi, data, regs);
3793 }
3794
3795 static int perf_swevent_is_counting(struct perf_event *event)
3796 {
3797         /*
3798          * The event is active, we're good!
3799          */
3800         if (event->state == PERF_EVENT_STATE_ACTIVE)
3801                 return 1;
3802
3803         /*
3804          * The event is off/error, not counting.
3805          */
3806         if (event->state != PERF_EVENT_STATE_INACTIVE)
3807                 return 0;
3808
3809         /*
3810          * The event is inactive, if the context is active
3811          * we're part of a group that didn't make it on the 'pmu',
3812          * not counting.
3813          */
3814         if (event->ctx->is_active)
3815                 return 0;
3816
3817         /*
3818          * We're inactive and the context is too, this means the
3819          * task is scheduled out, we're counting events that happen
3820          * to us, like migration events.
3821          */
3822         return 1;
3823 }
3824
3825 static int perf_tp_event_match(struct perf_event *event,
3826                                 struct perf_sample_data *data);
3827
3828 static int perf_swevent_match(struct perf_event *event,
3829                                 enum perf_type_id type,
3830                                 u32 event_id,
3831                                 struct perf_sample_data *data,
3832                                 struct pt_regs *regs)
3833 {
3834         if (!perf_swevent_is_counting(event))
3835                 return 0;
3836
3837         if (event->attr.type != type)
3838                 return 0;
3839         if (event->attr.config != event_id)
3840                 return 0;
3841
3842         if (regs) {
3843                 if (event->attr.exclude_user && user_mode(regs))
3844                         return 0;
3845
3846                 if (event->attr.exclude_kernel && !user_mode(regs))
3847                         return 0;
3848         }
3849
3850         if (event->attr.type == PERF_TYPE_TRACEPOINT &&
3851             !perf_tp_event_match(event, data))
3852                 return 0;
3853
3854         return 1;
3855 }
3856
3857 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
3858                                      enum perf_type_id type,
3859                                      u32 event_id, u64 nr, int nmi,
3860                                      struct perf_sample_data *data,
3861                                      struct pt_regs *regs)
3862 {
3863         struct perf_event *event;
3864
3865         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3866                 if (perf_swevent_match(event, type, event_id, data, regs))
3867                         perf_swevent_add(event, nr, nmi, data, regs);
3868         }
3869 }
3870
3871 /*
3872  * Must be called with preemption disabled
3873  */
3874 int perf_swevent_get_recursion_context(int **recursion)
3875 {
3876         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
3877
3878         if (in_nmi())
3879                 *recursion = &cpuctx->recursion[3];
3880         else if (in_irq())
3881                 *recursion = &cpuctx->recursion[2];
3882         else if (in_softirq())
3883                 *recursion = &cpuctx->recursion[1];
3884         else
3885                 *recursion = &cpuctx->recursion[0];
3886
3887         if (**recursion)
3888                 return -1;
3889
3890         (**recursion)++;
3891
3892         return 0;
3893 }
3894 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
3895
3896 void perf_swevent_put_recursion_context(int *recursion)
3897 {
3898         (*recursion)--;
3899 }
3900 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
3901
3902 static void __do_perf_sw_event(enum perf_type_id type, u32 event_id,
3903                                u64 nr, int nmi,
3904                                struct perf_sample_data *data,
3905                                struct pt_regs *regs)
3906 {
3907         struct perf_event_context *ctx;
3908         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
3909
3910         rcu_read_lock();
3911         perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
3912                                  nr, nmi, data, regs);
3913         /*
3914          * doesn't really matter which of the child contexts the
3915          * events ends up in.
3916          */
3917         ctx = rcu_dereference(current->perf_event_ctxp);
3918         if (ctx)
3919                 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
3920         rcu_read_unlock();
3921 }
3922
3923 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
3924                                     u64 nr, int nmi,
3925                                     struct perf_sample_data *data,
3926                                     struct pt_regs *regs)
3927 {
3928         int *recursion;
3929
3930         preempt_disable();
3931
3932         if (perf_swevent_get_recursion_context(&recursion))
3933                 goto out;
3934
3935         __do_perf_sw_event(type, event_id, nr, nmi, data, regs);
3936
3937         perf_swevent_put_recursion_context(recursion);
3938 out:
3939         preempt_enable();
3940 }
3941
3942 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
3943                             struct pt_regs *regs, u64 addr)
3944 {
3945         struct perf_sample_data data;
3946
3947         data.addr = addr;
3948         data.raw  = NULL;
3949
3950         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
3951 }
3952
3953 static void perf_swevent_read(struct perf_event *event)
3954 {
3955 }
3956
3957 static int perf_swevent_enable(struct perf_event *event)
3958 {
3959         struct hw_perf_event *hwc = &event->hw;
3960
3961         if (hwc->sample_period) {
3962                 hwc->last_period = hwc->sample_period;
3963                 perf_swevent_set_period(event);
3964         }
3965         return 0;
3966 }
3967
3968 static void perf_swevent_disable(struct perf_event *event)
3969 {
3970 }
3971
3972 static const struct pmu perf_ops_generic = {
3973         .enable         = perf_swevent_enable,
3974         .disable        = perf_swevent_disable,
3975         .read           = perf_swevent_read,
3976         .unthrottle     = perf_swevent_unthrottle,
3977 };
3978
3979 /*
3980  * hrtimer based swevent callback
3981  */
3982
3983 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
3984 {
3985         enum hrtimer_restart ret = HRTIMER_RESTART;
3986         struct perf_sample_data data;
3987         struct pt_regs *regs;
3988         struct perf_event *event;
3989         u64 period;
3990
3991         event   = container_of(hrtimer, struct perf_event, hw.hrtimer);
3992         event->pmu->read(event);
3993
3994         data.addr = 0;
3995         regs = get_irq_regs();
3996         /*
3997          * In case we exclude kernel IPs or are somehow not in interrupt
3998          * context, provide the next best thing, the user IP.
3999          */
4000         if ((event->attr.exclude_kernel || !regs) &&
4001                         !event->attr.exclude_user)
4002                 regs = task_pt_regs(current);
4003
4004         if (regs) {
4005                 if (!(event->attr.exclude_idle && current->pid == 0))
4006                         if (perf_event_overflow(event, 0, &data, regs))
4007                                 ret = HRTIMER_NORESTART;
4008         }
4009
4010         period = max_t(u64, 10000, event->hw.sample_period);
4011         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4012
4013         return ret;
4014 }
4015
4016 static void perf_swevent_start_hrtimer(struct perf_event *event)
4017 {
4018         struct hw_perf_event *hwc = &event->hw;
4019
4020         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4021         hwc->hrtimer.function = perf_swevent_hrtimer;
4022         if (hwc->sample_period) {
4023                 u64 period;
4024
4025                 if (hwc->remaining) {
4026                         if (hwc->remaining < 0)
4027                                 period = 10000;
4028                         else
4029                                 period = hwc->remaining;
4030                         hwc->remaining = 0;
4031                 } else {
4032                         period = max_t(u64, 10000, hwc->sample_period);
4033                 }
4034                 __hrtimer_start_range_ns(&hwc->hrtimer,
4035                                 ns_to_ktime(period), 0,
4036                                 HRTIMER_MODE_REL, 0);
4037         }
4038 }
4039
4040 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4041 {
4042         struct hw_perf_event *hwc = &event->hw;
4043
4044         if (hwc->sample_period) {
4045                 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4046                 hwc->remaining = ktime_to_ns(remaining);
4047
4048                 hrtimer_cancel(&hwc->hrtimer);
4049         }
4050 }
4051
4052 /*
4053  * Software event: cpu wall time clock
4054  */
4055
4056 static void cpu_clock_perf_event_update(struct perf_event *event)
4057 {
4058         int cpu = raw_smp_processor_id();
4059         s64 prev;
4060         u64 now;
4061
4062         now = cpu_clock(cpu);
4063         prev = atomic64_read(&event->hw.prev_count);
4064         atomic64_set(&event->hw.prev_count, now);
4065         atomic64_add(now - prev, &event->count);
4066 }
4067
4068 static int cpu_clock_perf_event_enable(struct perf_event *event)
4069 {
4070         struct hw_perf_event *hwc = &event->hw;
4071         int cpu = raw_smp_processor_id();
4072
4073         atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4074         perf_swevent_start_hrtimer(event);
4075
4076         return 0;
4077 }
4078
4079 static void cpu_clock_perf_event_disable(struct perf_event *event)
4080 {
4081         perf_swevent_cancel_hrtimer(event);
4082         cpu_clock_perf_event_update(event);
4083 }
4084
4085 static void cpu_clock_perf_event_read(struct perf_event *event)
4086 {
4087         cpu_clock_perf_event_update(event);
4088 }
4089
4090 static const struct pmu perf_ops_cpu_clock = {
4091         .enable         = cpu_clock_perf_event_enable,
4092         .disable        = cpu_clock_perf_event_disable,
4093         .read           = cpu_clock_perf_event_read,
4094 };
4095
4096 /*
4097  * Software event: task time clock
4098  */
4099
4100 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4101 {
4102         u64 prev;
4103         s64 delta;
4104
4105         prev = atomic64_xchg(&event->hw.prev_count, now);
4106         delta = now - prev;
4107         atomic64_add(delta, &event->count);
4108 }
4109
4110 static int task_clock_perf_event_enable(struct perf_event *event)
4111 {
4112         struct hw_perf_event *hwc = &event->hw;
4113         u64 now;
4114
4115         now = event->ctx->time;
4116
4117         atomic64_set(&hwc->prev_count, now);
4118
4119         perf_swevent_start_hrtimer(event);
4120
4121         return 0;
4122 }
4123
4124 static void task_clock_perf_event_disable(struct perf_event *event)
4125 {
4126         perf_swevent_cancel_hrtimer(event);
4127         task_clock_perf_event_update(event, event->ctx->time);
4128
4129 }
4130
4131 static void task_clock_perf_event_read(struct perf_event *event)
4132 {
4133         u64 time;
4134
4135         if (!in_nmi()) {
4136                 update_context_time(event->ctx);
4137                 time = event->ctx->time;
4138         } else {
4139                 u64 now = perf_clock();
4140                 u64 delta = now - event->ctx->timestamp;
4141                 time = event->ctx->time + delta;
4142         }
4143
4144         task_clock_perf_event_update(event, time);
4145 }
4146
4147 static const struct pmu perf_ops_task_clock = {
4148         .enable         = task_clock_perf_event_enable,
4149         .disable        = task_clock_perf_event_disable,
4150         .read           = task_clock_perf_event_read,
4151 };
4152
4153 #ifdef CONFIG_EVENT_PROFILE
4154
4155 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4156                           int entry_size)
4157 {
4158         struct perf_raw_record raw = {
4159                 .size = entry_size,
4160                 .data = record,
4161         };
4162
4163         struct perf_sample_data data = {
4164                 .addr = addr,
4165                 .raw = &raw,
4166         };
4167
4168         struct pt_regs *regs = get_irq_regs();
4169
4170         if (!regs)
4171                 regs = task_pt_regs(current);
4172
4173         /* Trace events already protected against recursion */
4174         __do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4175                                 &data, regs);
4176 }
4177 EXPORT_SYMBOL_GPL(perf_tp_event);
4178
4179 static int perf_tp_event_match(struct perf_event *event,
4180                                 struct perf_sample_data *data)
4181 {
4182         void *record = data->raw->data;
4183
4184         if (likely(!event->filter) || filter_match_preds(event->filter, record))
4185                 return 1;
4186         return 0;
4187 }
4188
4189 static void tp_perf_event_destroy(struct perf_event *event)
4190 {
4191         ftrace_profile_disable(event->attr.config);
4192 }
4193
4194 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4195 {
4196         /*
4197          * Raw tracepoint data is a severe data leak, only allow root to
4198          * have these.
4199          */
4200         if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4201                         perf_paranoid_tracepoint_raw() &&
4202                         !capable(CAP_SYS_ADMIN))
4203                 return ERR_PTR(-EPERM);
4204
4205         if (ftrace_profile_enable(event->attr.config))
4206                 return NULL;
4207
4208         event->destroy = tp_perf_event_destroy;
4209
4210         return &perf_ops_generic;
4211 }
4212
4213 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4214 {
4215         char *filter_str;
4216         int ret;
4217
4218         if (event->attr.type != PERF_TYPE_TRACEPOINT)
4219                 return -EINVAL;
4220
4221         filter_str = strndup_user(arg, PAGE_SIZE);
4222         if (IS_ERR(filter_str))
4223                 return PTR_ERR(filter_str);
4224
4225         ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4226
4227         kfree(filter_str);
4228         return ret;
4229 }
4230
4231 static void perf_event_free_filter(struct perf_event *event)
4232 {
4233         ftrace_profile_free_filter(event);
4234 }
4235
4236 #else
4237
4238 static int perf_tp_event_match(struct perf_event *event,
4239                                 struct perf_sample_data *data)
4240 {
4241         return 1;
4242 }
4243
4244 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4245 {
4246         return NULL;
4247 }
4248
4249 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4250 {
4251         return -ENOENT;
4252 }
4253
4254 static void perf_event_free_filter(struct perf_event *event)
4255 {
4256 }
4257
4258 #endif /* CONFIG_EVENT_PROFILE */
4259
4260 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4261 static void bp_perf_event_destroy(struct perf_event *event)
4262 {
4263         release_bp_slot(event);
4264 }
4265
4266 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4267 {
4268         int err;
4269         /*
4270          * The breakpoint is already filled if we haven't created the counter
4271          * through perf syscall
4272          * FIXME: manage to get trigerred to NULL if it comes from syscalls
4273          */
4274         if (!bp->callback)
4275                 err = register_perf_hw_breakpoint(bp);
4276         else
4277                 err = __register_perf_hw_breakpoint(bp);
4278         if (err)
4279                 return ERR_PTR(err);
4280
4281         bp->destroy = bp_perf_event_destroy;
4282
4283         return &perf_ops_bp;
4284 }
4285
4286 void perf_bp_event(struct perf_event *bp, void *regs)
4287 {
4288         /* TODO */
4289 }
4290 #else
4291 static void bp_perf_event_destroy(struct perf_event *event)
4292 {
4293 }
4294
4295 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4296 {
4297         return NULL;
4298 }
4299
4300 void perf_bp_event(struct perf_event *bp, void *regs)
4301 {
4302 }
4303 #endif
4304
4305 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4306
4307 static void sw_perf_event_destroy(struct perf_event *event)
4308 {
4309         u64 event_id = event->attr.config;
4310
4311         WARN_ON(event->parent);
4312
4313         atomic_dec(&perf_swevent_enabled[event_id]);
4314 }
4315
4316 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4317 {
4318         const struct pmu *pmu = NULL;
4319         u64 event_id = event->attr.config;
4320
4321         /*
4322          * Software events (currently) can't in general distinguish
4323          * between user, kernel and hypervisor events.
4324          * However, context switches and cpu migrations are considered
4325          * to be kernel events, and page faults are never hypervisor
4326          * events.
4327          */
4328         switch (event_id) {
4329         case PERF_COUNT_SW_CPU_CLOCK:
4330                 pmu = &perf_ops_cpu_clock;
4331
4332                 break;
4333         case PERF_COUNT_SW_TASK_CLOCK:
4334                 /*
4335                  * If the user instantiates this as a per-cpu event,
4336                  * use the cpu_clock event instead.
4337                  */
4338                 if (event->ctx->task)
4339                         pmu = &perf_ops_task_clock;
4340                 else
4341                         pmu = &perf_ops_cpu_clock;
4342
4343                 break;
4344         case PERF_COUNT_SW_PAGE_FAULTS:
4345         case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4346         case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4347         case PERF_COUNT_SW_CONTEXT_SWITCHES:
4348         case PERF_COUNT_SW_CPU_MIGRATIONS:
4349         case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4350         case PERF_COUNT_SW_EMULATION_FAULTS:
4351                 if (!event->parent) {
4352                         atomic_inc(&perf_swevent_enabled[event_id]);
4353                         event->destroy = sw_perf_event_destroy;
4354                 }
4355                 pmu = &perf_ops_generic;
4356                 break;
4357         }
4358
4359         return pmu;
4360 }
4361
4362 /*
4363  * Allocate and initialize a event structure
4364  */
4365 static struct perf_event *
4366 perf_event_alloc(struct perf_event_attr *attr,
4367                    int cpu,
4368                    struct perf_event_context *ctx,
4369                    struct perf_event *group_leader,
4370                    struct perf_event *parent_event,
4371                    perf_callback_t callback,
4372                    gfp_t gfpflags)
4373 {
4374         const struct pmu *pmu;
4375         struct perf_event *event;
4376         struct hw_perf_event *hwc;
4377         long err;
4378
4379         event = kzalloc(sizeof(*event), gfpflags);
4380         if (!event)
4381                 return ERR_PTR(-ENOMEM);
4382
4383         /*
4384          * Single events are their own group leaders, with an
4385          * empty sibling list:
4386          */
4387         if (!group_leader)
4388                 group_leader = event;
4389
4390         mutex_init(&event->child_mutex);
4391         INIT_LIST_HEAD(&event->child_list);
4392
4393         INIT_LIST_HEAD(&event->group_entry);
4394         INIT_LIST_HEAD(&event->event_entry);
4395         INIT_LIST_HEAD(&event->sibling_list);
4396         init_waitqueue_head(&event->waitq);
4397
4398         mutex_init(&event->mmap_mutex);
4399
4400         event->cpu              = cpu;
4401         event->attr             = *attr;
4402         event->group_leader     = group_leader;
4403         event->pmu              = NULL;
4404         event->ctx              = ctx;
4405         event->oncpu            = -1;
4406
4407         event->parent           = parent_event;
4408
4409         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
4410         event->id               = atomic64_inc_return(&perf_event_id);
4411
4412         event->state            = PERF_EVENT_STATE_INACTIVE;
4413
4414         if (!callback && parent_event)
4415                 callback = parent_event->callback;
4416         
4417         event->callback = callback;
4418
4419         if (attr->disabled)
4420                 event->state = PERF_EVENT_STATE_OFF;
4421
4422         pmu = NULL;
4423
4424         hwc = &event->hw;
4425         hwc->sample_period = attr->sample_period;
4426         if (attr->freq && attr->sample_freq)
4427                 hwc->sample_period = 1;
4428         hwc->last_period = hwc->sample_period;
4429
4430         atomic64_set(&hwc->period_left, hwc->sample_period);
4431
4432         /*
4433          * we currently do not support PERF_FORMAT_GROUP on inherited events
4434          */
4435         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4436                 goto done;
4437
4438         switch (attr->type) {
4439         case PERF_TYPE_RAW:
4440         case PERF_TYPE_HARDWARE:
4441         case PERF_TYPE_HW_CACHE:
4442                 pmu = hw_perf_event_init(event);
4443                 break;
4444
4445         case PERF_TYPE_SOFTWARE:
4446                 pmu = sw_perf_event_init(event);
4447                 break;
4448
4449         case PERF_TYPE_TRACEPOINT:
4450                 pmu = tp_perf_event_init(event);
4451                 break;
4452
4453         case PERF_TYPE_BREAKPOINT:
4454                 pmu = bp_perf_event_init(event);
4455                 break;
4456
4457
4458         default:
4459                 break;
4460         }
4461 done:
4462         err = 0;
4463         if (!pmu)
4464                 err = -EINVAL;
4465         else if (IS_ERR(pmu))
4466                 err = PTR_ERR(pmu);
4467
4468         if (err) {
4469                 if (event->ns)
4470                         put_pid_ns(event->ns);
4471                 kfree(event);
4472                 return ERR_PTR(err);
4473         }
4474
4475         event->pmu = pmu;
4476
4477         if (!event->parent) {
4478                 atomic_inc(&nr_events);
4479                 if (event->attr.mmap)
4480                         atomic_inc(&nr_mmap_events);
4481                 if (event->attr.comm)
4482                         atomic_inc(&nr_comm_events);
4483                 if (event->attr.task)
4484                         atomic_inc(&nr_task_events);
4485         }
4486
4487         return event;
4488 }
4489
4490 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4491                           struct perf_event_attr *attr)
4492 {
4493         u32 size;
4494         int ret;
4495
4496         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4497                 return -EFAULT;
4498
4499         /*
4500          * zero the full structure, so that a short copy will be nice.
4501          */
4502         memset(attr, 0, sizeof(*attr));
4503
4504         ret = get_user(size, &uattr->size);
4505         if (ret)
4506                 return ret;
4507
4508         if (size > PAGE_SIZE)   /* silly large */
4509                 goto err_size;
4510
4511         if (!size)              /* abi compat */
4512                 size = PERF_ATTR_SIZE_VER0;
4513
4514         if (size < PERF_ATTR_SIZE_VER0)
4515                 goto err_size;
4516
4517         /*
4518          * If we're handed a bigger struct than we know of,
4519          * ensure all the unknown bits are 0 - i.e. new
4520          * user-space does not rely on any kernel feature
4521          * extensions we dont know about yet.
4522          */
4523         if (size > sizeof(*attr)) {
4524                 unsigned char __user *addr;
4525                 unsigned char __user *end;
4526                 unsigned char val;
4527
4528                 addr = (void __user *)uattr + sizeof(*attr);
4529                 end  = (void __user *)uattr + size;
4530
4531                 for (; addr < end; addr++) {
4532                         ret = get_user(val, addr);
4533                         if (ret)
4534                                 return ret;
4535                         if (val)
4536                                 goto err_size;
4537                 }
4538                 size = sizeof(*attr);
4539         }
4540
4541         ret = copy_from_user(attr, uattr, size);
4542         if (ret)
4543                 return -EFAULT;
4544
4545         /*
4546          * If the type exists, the corresponding creation will verify
4547          * the attr->config.
4548          */
4549         if (attr->type >= PERF_TYPE_MAX)
4550                 return -EINVAL;
4551
4552         if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3)
4553                 return -EINVAL;
4554
4555         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4556                 return -EINVAL;
4557
4558         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4559                 return -EINVAL;
4560
4561 out:
4562         return ret;
4563
4564 err_size:
4565         put_user(sizeof(*attr), &uattr->size);
4566         ret = -E2BIG;
4567         goto out;
4568 }
4569
4570 static int perf_event_set_output(struct perf_event *event, int output_fd)
4571 {
4572         struct perf_event *output_event = NULL;
4573         struct file *output_file = NULL;
4574         struct perf_event *old_output;
4575         int fput_needed = 0;
4576         int ret = -EINVAL;
4577
4578         if (!output_fd)
4579                 goto set;
4580
4581         output_file = fget_light(output_fd, &fput_needed);
4582         if (!output_file)
4583                 return -EBADF;
4584
4585         if (output_file->f_op != &perf_fops)
4586                 goto out;
4587
4588         output_event = output_file->private_data;
4589
4590         /* Don't chain output fds */
4591         if (output_event->output)
4592                 goto out;
4593
4594         /* Don't set an output fd when we already have an output channel */
4595         if (event->data)
4596                 goto out;
4597
4598         atomic_long_inc(&output_file->f_count);
4599
4600 set:
4601         mutex_lock(&event->mmap_mutex);
4602         old_output = event->output;
4603         rcu_assign_pointer(event->output, output_event);
4604         mutex_unlock(&event->mmap_mutex);
4605
4606         if (old_output) {
4607                 /*
4608                  * we need to make sure no existing perf_output_*()
4609                  * is still referencing this event.
4610                  */
4611                 synchronize_rcu();
4612                 fput(old_output->filp);
4613         }
4614
4615         ret = 0;
4616 out:
4617         fput_light(output_file, fput_needed);
4618         return ret;
4619 }
4620
4621 /**
4622  * sys_perf_event_open - open a performance event, associate it to a task/cpu
4623  *
4624  * @attr_uptr:  event_id type attributes for monitoring/sampling
4625  * @pid:                target pid
4626  * @cpu:                target cpu
4627  * @group_fd:           group leader event fd
4628  */
4629 SYSCALL_DEFINE5(perf_event_open,
4630                 struct perf_event_attr __user *, attr_uptr,
4631                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4632 {
4633         struct perf_event *event, *group_leader;
4634         struct perf_event_attr attr;
4635         struct perf_event_context *ctx;
4636         struct file *event_file = NULL;
4637         struct file *group_file = NULL;
4638         int fput_needed = 0;
4639         int fput_needed2 = 0;
4640         int err;
4641
4642         /* for future expandability... */
4643         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4644                 return -EINVAL;
4645
4646         err = perf_copy_attr(attr_uptr, &attr);
4647         if (err)
4648                 return err;
4649
4650         if (!attr.exclude_kernel) {
4651                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4652                         return -EACCES;
4653         }
4654
4655         if (attr.freq) {
4656                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4657                         return -EINVAL;
4658         }
4659
4660         /*
4661          * Get the target context (task or percpu):
4662          */
4663         ctx = find_get_context(pid, cpu);
4664         if (IS_ERR(ctx))
4665                 return PTR_ERR(ctx);
4666
4667         /*
4668          * Look up the group leader (we will attach this event to it):
4669          */
4670         group_leader = NULL;
4671         if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4672                 err = -EINVAL;
4673                 group_file = fget_light(group_fd, &fput_needed);
4674                 if (!group_file)
4675                         goto err_put_context;
4676                 if (group_file->f_op != &perf_fops)
4677                         goto err_put_context;
4678
4679                 group_leader = group_file->private_data;
4680                 /*
4681                  * Do not allow a recursive hierarchy (this new sibling
4682                  * becoming part of another group-sibling):
4683                  */
4684                 if (group_leader->group_leader != group_leader)
4685                         goto err_put_context;
4686                 /*
4687                  * Do not allow to attach to a group in a different
4688                  * task or CPU context:
4689                  */
4690                 if (group_leader->ctx != ctx)
4691                         goto err_put_context;
4692                 /*
4693                  * Only a group leader can be exclusive or pinned
4694                  */
4695                 if (attr.exclusive || attr.pinned)
4696                         goto err_put_context;
4697         }
4698
4699         event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4700                                      NULL, NULL, GFP_KERNEL);
4701         err = PTR_ERR(event);
4702         if (IS_ERR(event))
4703                 goto err_put_context;
4704
4705         err = anon_inode_getfd("[perf_event]", &perf_fops, event, 0);
4706         if (err < 0)
4707                 goto err_free_put_context;
4708
4709         event_file = fget_light(err, &fput_needed2);
4710         if (!event_file)
4711                 goto err_free_put_context;
4712
4713         if (flags & PERF_FLAG_FD_OUTPUT) {
4714                 err = perf_event_set_output(event, group_fd);
4715                 if (err)
4716                         goto err_fput_free_put_context;
4717         }
4718
4719         event->filp = event_file;
4720         WARN_ON_ONCE(ctx->parent_ctx);
4721         mutex_lock(&ctx->mutex);
4722         perf_install_in_context(ctx, event, cpu);
4723         ++ctx->generation;
4724         mutex_unlock(&ctx->mutex);
4725
4726         event->owner = current;
4727         get_task_struct(current);
4728         mutex_lock(&current->perf_event_mutex);
4729         list_add_tail(&event->owner_entry, &current->perf_event_list);
4730         mutex_unlock(&current->perf_event_mutex);
4731
4732 err_fput_free_put_context:
4733         fput_light(event_file, fput_needed2);
4734
4735 err_free_put_context:
4736         if (err < 0)
4737                 kfree(event);
4738
4739 err_put_context:
4740         if (err < 0)
4741                 put_ctx(ctx);
4742
4743         fput_light(group_file, fput_needed);
4744
4745         return err;
4746 }
4747
4748 /**
4749  * perf_event_create_kernel_counter
4750  *
4751  * @attr: attributes of the counter to create
4752  * @cpu: cpu in which the counter is bound
4753  * @pid: task to profile
4754  */
4755 struct perf_event *
4756 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4757                                  pid_t pid, perf_callback_t callback)
4758 {
4759         struct perf_event *event;
4760         struct perf_event_context *ctx;
4761         int err;
4762
4763         /*
4764          * Get the target context (task or percpu):
4765          */
4766
4767         ctx = find_get_context(pid, cpu);
4768         if (IS_ERR(ctx))
4769                 return NULL;
4770
4771         event = perf_event_alloc(attr, cpu, ctx, NULL,
4772                                      NULL, callback, GFP_KERNEL);
4773         err = PTR_ERR(event);
4774         if (IS_ERR(event))
4775                 goto err_put_context;
4776
4777         event->filp = NULL;
4778         WARN_ON_ONCE(ctx->parent_ctx);
4779         mutex_lock(&ctx->mutex);
4780         perf_install_in_context(ctx, event, cpu);
4781         ++ctx->generation;
4782         mutex_unlock(&ctx->mutex);
4783
4784         event->owner = current;
4785         get_task_struct(current);
4786         mutex_lock(&current->perf_event_mutex);
4787         list_add_tail(&event->owner_entry, &current->perf_event_list);
4788         mutex_unlock(&current->perf_event_mutex);
4789
4790         return event;
4791
4792 err_put_context:
4793         if (err < 0)
4794                 put_ctx(ctx);
4795
4796         return NULL;
4797 }
4798 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4799
4800 /*
4801  * inherit a event from parent task to child task:
4802  */
4803 static struct perf_event *
4804 inherit_event(struct perf_event *parent_event,
4805               struct task_struct *parent,
4806               struct perf_event_context *parent_ctx,
4807               struct task_struct *child,
4808               struct perf_event *group_leader,
4809               struct perf_event_context *child_ctx)
4810 {
4811         struct perf_event *child_event;
4812
4813         /*
4814          * Instead of creating recursive hierarchies of events,
4815          * we link inherited events back to the original parent,
4816          * which has a filp for sure, which we use as the reference
4817          * count:
4818          */
4819         if (parent_event->parent)
4820                 parent_event = parent_event->parent;
4821
4822         child_event = perf_event_alloc(&parent_event->attr,
4823                                            parent_event->cpu, child_ctx,
4824                                            group_leader, parent_event,
4825                                            NULL, GFP_KERNEL);
4826         if (IS_ERR(child_event))
4827                 return child_event;
4828         get_ctx(child_ctx);
4829
4830         /*
4831          * Make the child state follow the state of the parent event,
4832          * not its attr.disabled bit.  We hold the parent's mutex,
4833          * so we won't race with perf_event_{en, dis}able_family.
4834          */
4835         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
4836                 child_event->state = PERF_EVENT_STATE_INACTIVE;
4837         else
4838                 child_event->state = PERF_EVENT_STATE_OFF;
4839
4840         if (parent_event->attr.freq)
4841                 child_event->hw.sample_period = parent_event->hw.sample_period;
4842
4843         child_event->overflow_handler = parent_event->overflow_handler;
4844
4845         /*
4846          * Link it up in the child's context:
4847          */
4848         add_event_to_ctx(child_event, child_ctx);
4849
4850         /*
4851          * Get a reference to the parent filp - we will fput it
4852          * when the child event exits. This is safe to do because
4853          * we are in the parent and we know that the filp still
4854          * exists and has a nonzero count:
4855          */
4856         atomic_long_inc(&parent_event->filp->f_count);
4857
4858         /*
4859          * Link this into the parent event's child list
4860          */
4861         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4862         mutex_lock(&parent_event->child_mutex);
4863         list_add_tail(&child_event->child_list, &parent_event->child_list);
4864         mutex_unlock(&parent_event->child_mutex);
4865
4866         return child_event;
4867 }
4868
4869 static int inherit_group(struct perf_event *parent_event,
4870               struct task_struct *parent,
4871               struct perf_event_context *parent_ctx,
4872               struct task_struct *child,
4873               struct perf_event_context *child_ctx)
4874 {
4875         struct perf_event *leader;
4876         struct perf_event *sub;
4877         struct perf_event *child_ctr;
4878
4879         leader = inherit_event(parent_event, parent, parent_ctx,
4880                                  child, NULL, child_ctx);
4881         if (IS_ERR(leader))
4882                 return PTR_ERR(leader);
4883         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
4884                 child_ctr = inherit_event(sub, parent, parent_ctx,
4885                                             child, leader, child_ctx);
4886                 if (IS_ERR(child_ctr))
4887                         return PTR_ERR(child_ctr);
4888         }
4889         return 0;
4890 }
4891
4892 static void sync_child_event(struct perf_event *child_event,
4893                                struct task_struct *child)
4894 {
4895         struct perf_event *parent_event = child_event->parent;
4896         u64 child_val;
4897
4898         if (child_event->attr.inherit_stat)
4899                 perf_event_read_event(child_event, child);
4900
4901         child_val = atomic64_read(&child_event->count);
4902
4903         /*
4904          * Add back the child's count to the parent's count:
4905          */
4906         atomic64_add(child_val, &parent_event->count);
4907         atomic64_add(child_event->total_time_enabled,
4908                      &parent_event->child_total_time_enabled);
4909         atomic64_add(child_event->total_time_running,
4910                      &parent_event->child_total_time_running);
4911
4912         /*
4913          * Remove this event from the parent's list
4914          */
4915         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4916         mutex_lock(&parent_event->child_mutex);
4917         list_del_init(&child_event->child_list);
4918         mutex_unlock(&parent_event->child_mutex);
4919
4920         /*
4921          * Release the parent event, if this was the last
4922          * reference to it.
4923          */
4924         fput(parent_event->filp);
4925 }
4926
4927 static void
4928 __perf_event_exit_task(struct perf_event *child_event,
4929                          struct perf_event_context *child_ctx,
4930                          struct task_struct *child)
4931 {
4932         struct perf_event *parent_event;
4933
4934         update_event_times(child_event);
4935         perf_event_remove_from_context(child_event);
4936
4937         parent_event = child_event->parent;
4938         /*
4939          * It can happen that parent exits first, and has events
4940          * that are still around due to the child reference. These
4941          * events need to be zapped - but otherwise linger.
4942          */
4943         if (parent_event) {
4944                 sync_child_event(child_event, child);
4945                 free_event(child_event);
4946         }
4947 }
4948
4949 /*
4950  * When a child task exits, feed back event values to parent events.
4951  */
4952 void perf_event_exit_task(struct task_struct *child)
4953 {
4954         struct perf_event *child_event, *tmp;
4955         struct perf_event_context *child_ctx;
4956         unsigned long flags;
4957
4958         if (likely(!child->perf_event_ctxp)) {
4959                 perf_event_task(child, NULL, 0);
4960                 return;
4961         }
4962
4963         local_irq_save(flags);
4964         /*
4965          * We can't reschedule here because interrupts are disabled,
4966          * and either child is current or it is a task that can't be
4967          * scheduled, so we are now safe from rescheduling changing
4968          * our context.
4969          */
4970         child_ctx = child->perf_event_ctxp;
4971         __perf_event_task_sched_out(child_ctx);
4972
4973         /*
4974          * Take the context lock here so that if find_get_context is
4975          * reading child->perf_event_ctxp, we wait until it has
4976          * incremented the context's refcount before we do put_ctx below.
4977          */
4978         spin_lock(&child_ctx->lock);
4979         child->perf_event_ctxp = NULL;
4980         /*
4981          * If this context is a clone; unclone it so it can't get
4982          * swapped to another process while we're removing all
4983          * the events from it.
4984          */
4985         unclone_ctx(child_ctx);
4986         spin_unlock_irqrestore(&child_ctx->lock, flags);
4987
4988         /*
4989          * Report the task dead after unscheduling the events so that we
4990          * won't get any samples after PERF_RECORD_EXIT. We can however still
4991          * get a few PERF_RECORD_READ events.
4992          */
4993         perf_event_task(child, child_ctx, 0);
4994
4995         /*
4996          * We can recurse on the same lock type through:
4997          *
4998          *   __perf_event_exit_task()
4999          *     sync_child_event()
5000          *       fput(parent_event->filp)
5001          *         perf_release()
5002          *           mutex_lock(&ctx->mutex)
5003          *
5004          * But since its the parent context it won't be the same instance.
5005          */
5006         mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
5007
5008 again:
5009         list_for_each_entry_safe(child_event, tmp, &child_ctx->group_list,
5010                                  group_entry)
5011                 __perf_event_exit_task(child_event, child_ctx, child);
5012
5013         /*
5014          * If the last event was a group event, it will have appended all
5015          * its siblings to the list, but we obtained 'tmp' before that which
5016          * will still point to the list head terminating the iteration.
5017          */
5018         if (!list_empty(&child_ctx->group_list))
5019                 goto again;
5020
5021         mutex_unlock(&child_ctx->mutex);
5022
5023         put_ctx(child_ctx);
5024 }
5025
5026 /*
5027  * free an unexposed, unused context as created by inheritance by
5028  * init_task below, used by fork() in case of fail.
5029  */
5030 void perf_event_free_task(struct task_struct *task)
5031 {
5032         struct perf_event_context *ctx = task->perf_event_ctxp;
5033         struct perf_event *event, *tmp;
5034
5035         if (!ctx)
5036                 return;
5037
5038         mutex_lock(&ctx->mutex);
5039 again:
5040         list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry) {
5041                 struct perf_event *parent = event->parent;
5042
5043                 if (WARN_ON_ONCE(!parent))
5044                         continue;
5045
5046                 mutex_lock(&parent->child_mutex);
5047                 list_del_init(&event->child_list);
5048                 mutex_unlock(&parent->child_mutex);
5049
5050                 fput(parent->filp);
5051
5052                 list_del_event(event, ctx);
5053                 free_event(event);
5054         }
5055
5056         if (!list_empty(&ctx->group_list))
5057                 goto again;
5058
5059         mutex_unlock(&ctx->mutex);
5060
5061         put_ctx(ctx);
5062 }
5063
5064 /*
5065  * Initialize the perf_event context in task_struct
5066  */
5067 int perf_event_init_task(struct task_struct *child)
5068 {
5069         struct perf_event_context *child_ctx, *parent_ctx;
5070         struct perf_event_context *cloned_ctx;
5071         struct perf_event *event;
5072         struct task_struct *parent = current;
5073         int inherited_all = 1;
5074         int ret = 0;
5075
5076         child->perf_event_ctxp = NULL;
5077
5078         mutex_init(&child->perf_event_mutex);
5079         INIT_LIST_HEAD(&child->perf_event_list);
5080
5081         if (likely(!parent->perf_event_ctxp))
5082                 return 0;
5083
5084         /*
5085          * This is executed from the parent task context, so inherit
5086          * events that have been marked for cloning.
5087          * First allocate and initialize a context for the child.
5088          */
5089
5090         child_ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
5091         if (!child_ctx)
5092                 return -ENOMEM;
5093
5094         __perf_event_init_context(child_ctx, child);
5095         child->perf_event_ctxp = child_ctx;
5096         get_task_struct(child);
5097
5098         /*
5099          * If the parent's context is a clone, pin it so it won't get
5100          * swapped under us.
5101          */
5102         parent_ctx = perf_pin_task_context(parent);
5103
5104         /*
5105          * No need to check if parent_ctx != NULL here; since we saw
5106          * it non-NULL earlier, the only reason for it to become NULL
5107          * is if we exit, and since we're currently in the middle of
5108          * a fork we can't be exiting at the same time.
5109          */
5110
5111         /*
5112          * Lock the parent list. No need to lock the child - not PID
5113          * hashed yet and not running, so nobody can access it.
5114          */
5115         mutex_lock(&parent_ctx->mutex);
5116
5117         /*
5118          * We dont have to disable NMIs - we are only looking at
5119          * the list, not manipulating it:
5120          */
5121         list_for_each_entry(event, &parent_ctx->group_list, group_entry) {
5122
5123                 if (!event->attr.inherit) {
5124                         inherited_all = 0;
5125                         continue;
5126                 }
5127
5128                 ret = inherit_group(event, parent, parent_ctx,
5129                                              child, child_ctx);
5130                 if (ret) {
5131                         inherited_all = 0;
5132                         break;
5133                 }
5134         }
5135
5136         if (inherited_all) {
5137                 /*
5138                  * Mark the child context as a clone of the parent
5139                  * context, or of whatever the parent is a clone of.
5140                  * Note that if the parent is a clone, it could get
5141                  * uncloned at any point, but that doesn't matter
5142                  * because the list of events and the generation
5143                  * count can't have changed since we took the mutex.
5144                  */
5145                 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5146                 if (cloned_ctx) {
5147                         child_ctx->parent_ctx = cloned_ctx;
5148                         child_ctx->parent_gen = parent_ctx->parent_gen;
5149                 } else {
5150                         child_ctx->parent_ctx = parent_ctx;
5151                         child_ctx->parent_gen = parent_ctx->generation;
5152                 }
5153                 get_ctx(child_ctx->parent_ctx);
5154         }
5155
5156         mutex_unlock(&parent_ctx->mutex);
5157
5158         perf_unpin_context(parent_ctx);
5159
5160         return ret;
5161 }
5162
5163 static void __cpuinit perf_event_init_cpu(int cpu)
5164 {
5165         struct perf_cpu_context *cpuctx;
5166
5167         cpuctx = &per_cpu(perf_cpu_context, cpu);
5168         __perf_event_init_context(&cpuctx->ctx, NULL);
5169
5170         spin_lock(&perf_resource_lock);
5171         cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5172         spin_unlock(&perf_resource_lock);
5173
5174         hw_perf_event_setup(cpu);
5175 }
5176
5177 #ifdef CONFIG_HOTPLUG_CPU
5178 static void __perf_event_exit_cpu(void *info)
5179 {
5180         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5181         struct perf_event_context *ctx = &cpuctx->ctx;
5182         struct perf_event *event, *tmp;
5183
5184         list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry)
5185                 __perf_event_remove_from_context(event);
5186 }
5187 static void perf_event_exit_cpu(int cpu)
5188 {
5189         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5190         struct perf_event_context *ctx = &cpuctx->ctx;
5191
5192         mutex_lock(&ctx->mutex);
5193         smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5194         mutex_unlock(&ctx->mutex);
5195 }
5196 #else
5197 static inline void perf_event_exit_cpu(int cpu) { }
5198 #endif
5199
5200 static int __cpuinit
5201 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5202 {
5203         unsigned int cpu = (long)hcpu;
5204
5205         switch (action) {
5206
5207         case CPU_UP_PREPARE:
5208         case CPU_UP_PREPARE_FROZEN:
5209                 perf_event_init_cpu(cpu);
5210                 break;
5211
5212         case CPU_ONLINE:
5213         case CPU_ONLINE_FROZEN:
5214                 hw_perf_event_setup_online(cpu);
5215                 break;
5216
5217         case CPU_DOWN_PREPARE:
5218         case CPU_DOWN_PREPARE_FROZEN:
5219                 perf_event_exit_cpu(cpu);
5220                 break;
5221
5222         default:
5223                 break;
5224         }
5225
5226         return NOTIFY_OK;
5227 }
5228
5229 /*
5230  * This has to have a higher priority than migration_notifier in sched.c.
5231  */
5232 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5233         .notifier_call          = perf_cpu_notify,
5234         .priority               = 20,
5235 };
5236
5237 void __init perf_event_init(void)
5238 {
5239         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5240                         (void *)(long)smp_processor_id());
5241         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5242                         (void *)(long)smp_processor_id());
5243         register_cpu_notifier(&perf_cpu_nb);
5244 }
5245
5246 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
5247 {
5248         return sprintf(buf, "%d\n", perf_reserved_percpu);
5249 }
5250
5251 static ssize_t
5252 perf_set_reserve_percpu(struct sysdev_class *class,
5253                         const char *buf,
5254                         size_t count)
5255 {
5256         struct perf_cpu_context *cpuctx;
5257         unsigned long val;
5258         int err, cpu, mpt;
5259
5260         err = strict_strtoul(buf, 10, &val);
5261         if (err)
5262                 return err;
5263         if (val > perf_max_events)
5264                 return -EINVAL;
5265
5266         spin_lock(&perf_resource_lock);
5267         perf_reserved_percpu = val;
5268         for_each_online_cpu(cpu) {
5269                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5270                 spin_lock_irq(&cpuctx->ctx.lock);
5271                 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5272                           perf_max_events - perf_reserved_percpu);
5273                 cpuctx->max_pertask = mpt;
5274                 spin_unlock_irq(&cpuctx->ctx.lock);
5275         }
5276         spin_unlock(&perf_resource_lock);
5277
5278         return count;
5279 }
5280
5281 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
5282 {
5283         return sprintf(buf, "%d\n", perf_overcommit);
5284 }
5285
5286 static ssize_t
5287 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
5288 {
5289         unsigned long val;
5290         int err;
5291
5292         err = strict_strtoul(buf, 10, &val);
5293         if (err)
5294                 return err;
5295         if (val > 1)
5296                 return -EINVAL;
5297
5298         spin_lock(&perf_resource_lock);
5299         perf_overcommit = val;
5300         spin_unlock(&perf_resource_lock);
5301
5302         return count;
5303 }
5304
5305 static SYSDEV_CLASS_ATTR(
5306                                 reserve_percpu,
5307                                 0644,
5308                                 perf_show_reserve_percpu,
5309                                 perf_set_reserve_percpu
5310                         );
5311
5312 static SYSDEV_CLASS_ATTR(
5313                                 overcommit,
5314                                 0644,
5315                                 perf_show_overcommit,
5316                                 perf_set_overcommit
5317                         );
5318
5319 static struct attribute *perfclass_attrs[] = {
5320         &attr_reserve_percpu.attr,
5321         &attr_overcommit.attr,
5322         NULL
5323 };
5324
5325 static struct attribute_group perfclass_attr_group = {
5326         .attrs                  = perfclass_attrs,
5327         .name                   = "perf_events",
5328 };
5329
5330 static int __init perf_event_sysfs_init(void)
5331 {
5332         return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5333                                   &perfclass_attr_group);
5334 }
5335 device_initcall(perf_event_sysfs_init);