Merge branch 'perf-fixes-for-linus-2' of git://git.kernel.org/pub/scm/linux/kernel...
[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 static 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                 raw_spin_lock_irqsave(&ctx->lock, *flags);
207                 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
208                         raw_spin_unlock_irqrestore(&ctx->lock, *flags);
209                         goto retry;
210                 }
211
212                 if (!atomic_inc_not_zero(&ctx->refcount)) {
213                         raw_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                 raw_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         raw_spin_lock_irqsave(&ctx->lock, flags);
244         --ctx->pin_count;
245         raw_spin_unlock_irqrestore(&ctx->lock, flags);
246         put_ctx(ctx);
247 }
248
249 static inline u64 perf_clock(void)
250 {
251         return cpu_clock(smp_processor_id());
252 }
253
254 /*
255  * Update the record of the current time in a context.
256  */
257 static void update_context_time(struct perf_event_context *ctx)
258 {
259         u64 now = perf_clock();
260
261         ctx->time += now - ctx->timestamp;
262         ctx->timestamp = now;
263 }
264
265 /*
266  * Update the total_time_enabled and total_time_running fields for a event.
267  */
268 static void update_event_times(struct perf_event *event)
269 {
270         struct perf_event_context *ctx = event->ctx;
271         u64 run_end;
272
273         if (event->state < PERF_EVENT_STATE_INACTIVE ||
274             event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
275                 return;
276
277         if (ctx->is_active)
278                 run_end = ctx->time;
279         else
280                 run_end = event->tstamp_stopped;
281
282         event->total_time_enabled = run_end - event->tstamp_enabled;
283
284         if (event->state == PERF_EVENT_STATE_INACTIVE)
285                 run_end = event->tstamp_stopped;
286         else
287                 run_end = ctx->time;
288
289         event->total_time_running = run_end - event->tstamp_running;
290 }
291
292 /*
293  * Add a event from the lists for its context.
294  * Must be called with ctx->mutex and ctx->lock held.
295  */
296 static void
297 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
298 {
299         struct perf_event *group_leader = event->group_leader;
300
301         /*
302          * Depending on whether it is a standalone or sibling event,
303          * add it straight to the context's event list, or to the group
304          * leader's sibling list:
305          */
306         if (group_leader == event)
307                 list_add_tail(&event->group_entry, &ctx->group_list);
308         else {
309                 list_add_tail(&event->group_entry, &group_leader->sibling_list);
310                 group_leader->nr_siblings++;
311         }
312
313         list_add_rcu(&event->event_entry, &ctx->event_list);
314         ctx->nr_events++;
315         if (event->attr.inherit_stat)
316                 ctx->nr_stat++;
317 }
318
319 /*
320  * Remove a event from the lists for its context.
321  * Must be called with ctx->mutex and ctx->lock held.
322  */
323 static void
324 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
325 {
326         struct perf_event *sibling, *tmp;
327
328         if (list_empty(&event->group_entry))
329                 return;
330         ctx->nr_events--;
331         if (event->attr.inherit_stat)
332                 ctx->nr_stat--;
333
334         list_del_init(&event->group_entry);
335         list_del_rcu(&event->event_entry);
336
337         if (event->group_leader != event)
338                 event->group_leader->nr_siblings--;
339
340         update_event_times(event);
341
342         /*
343          * If event was in error state, then keep it
344          * that way, otherwise bogus counts will be
345          * returned on read(). The only way to get out
346          * of error state is by explicit re-enabling
347          * of the event
348          */
349         if (event->state > PERF_EVENT_STATE_OFF)
350                 event->state = PERF_EVENT_STATE_OFF;
351
352         /*
353          * If this was a group event with sibling events then
354          * upgrade the siblings to singleton events by adding them
355          * to the context list directly:
356          */
357         list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
358
359                 list_move_tail(&sibling->group_entry, &ctx->group_list);
360                 sibling->group_leader = sibling;
361         }
362 }
363
364 static void
365 event_sched_out(struct perf_event *event,
366                   struct perf_cpu_context *cpuctx,
367                   struct perf_event_context *ctx)
368 {
369         if (event->state != PERF_EVENT_STATE_ACTIVE)
370                 return;
371
372         event->state = PERF_EVENT_STATE_INACTIVE;
373         if (event->pending_disable) {
374                 event->pending_disable = 0;
375                 event->state = PERF_EVENT_STATE_OFF;
376         }
377         event->tstamp_stopped = ctx->time;
378         event->pmu->disable(event);
379         event->oncpu = -1;
380
381         if (!is_software_event(event))
382                 cpuctx->active_oncpu--;
383         ctx->nr_active--;
384         if (event->attr.exclusive || !cpuctx->active_oncpu)
385                 cpuctx->exclusive = 0;
386 }
387
388 static void
389 group_sched_out(struct perf_event *group_event,
390                 struct perf_cpu_context *cpuctx,
391                 struct perf_event_context *ctx)
392 {
393         struct perf_event *event;
394
395         if (group_event->state != PERF_EVENT_STATE_ACTIVE)
396                 return;
397
398         event_sched_out(group_event, cpuctx, ctx);
399
400         /*
401          * Schedule out siblings (if any):
402          */
403         list_for_each_entry(event, &group_event->sibling_list, group_entry)
404                 event_sched_out(event, cpuctx, ctx);
405
406         if (group_event->attr.exclusive)
407                 cpuctx->exclusive = 0;
408 }
409
410 /*
411  * Cross CPU call to remove a performance event
412  *
413  * We disable the event on the hardware level first. After that we
414  * remove it from the context list.
415  */
416 static void __perf_event_remove_from_context(void *info)
417 {
418         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
419         struct perf_event *event = info;
420         struct perf_event_context *ctx = event->ctx;
421
422         /*
423          * If this is a task context, we need to check whether it is
424          * the current task context of this cpu. If not it has been
425          * scheduled out before the smp call arrived.
426          */
427         if (ctx->task && cpuctx->task_ctx != ctx)
428                 return;
429
430         raw_spin_lock(&ctx->lock);
431         /*
432          * Protect the list operation against NMI by disabling the
433          * events on a global level.
434          */
435         perf_disable();
436
437         event_sched_out(event, cpuctx, ctx);
438
439         list_del_event(event, ctx);
440
441         if (!ctx->task) {
442                 /*
443                  * Allow more per task events with respect to the
444                  * reservation:
445                  */
446                 cpuctx->max_pertask =
447                         min(perf_max_events - ctx->nr_events,
448                             perf_max_events - perf_reserved_percpu);
449         }
450
451         perf_enable();
452         raw_spin_unlock(&ctx->lock);
453 }
454
455
456 /*
457  * Remove the event from a task's (or a CPU's) list of events.
458  *
459  * Must be called with ctx->mutex held.
460  *
461  * CPU events are removed with a smp call. For task events we only
462  * call when the task is on a CPU.
463  *
464  * If event->ctx is a cloned context, callers must make sure that
465  * every task struct that event->ctx->task could possibly point to
466  * remains valid.  This is OK when called from perf_release since
467  * that only calls us on the top-level context, which can't be a clone.
468  * When called from perf_event_exit_task, it's OK because the
469  * context has been detached from its task.
470  */
471 static void perf_event_remove_from_context(struct perf_event *event)
472 {
473         struct perf_event_context *ctx = event->ctx;
474         struct task_struct *task = ctx->task;
475
476         if (!task) {
477                 /*
478                  * Per cpu events are removed via an smp call and
479                  * the removal is always successful.
480                  */
481                 smp_call_function_single(event->cpu,
482                                          __perf_event_remove_from_context,
483                                          event, 1);
484                 return;
485         }
486
487 retry:
488         task_oncpu_function_call(task, __perf_event_remove_from_context,
489                                  event);
490
491         raw_spin_lock_irq(&ctx->lock);
492         /*
493          * If the context is active we need to retry the smp call.
494          */
495         if (ctx->nr_active && !list_empty(&event->group_entry)) {
496                 raw_spin_unlock_irq(&ctx->lock);
497                 goto retry;
498         }
499
500         /*
501          * The lock prevents that this context is scheduled in so we
502          * can remove the event safely, if the call above did not
503          * succeed.
504          */
505         if (!list_empty(&event->group_entry))
506                 list_del_event(event, ctx);
507         raw_spin_unlock_irq(&ctx->lock);
508 }
509
510 /*
511  * Update total_time_enabled and total_time_running for all events in a group.
512  */
513 static void update_group_times(struct perf_event *leader)
514 {
515         struct perf_event *event;
516
517         update_event_times(leader);
518         list_for_each_entry(event, &leader->sibling_list, group_entry)
519                 update_event_times(event);
520 }
521
522 /*
523  * Cross CPU call to disable a performance event
524  */
525 static void __perf_event_disable(void *info)
526 {
527         struct perf_event *event = info;
528         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
529         struct perf_event_context *ctx = event->ctx;
530
531         /*
532          * If this is a per-task event, need to check whether this
533          * event's task is the current task on this cpu.
534          */
535         if (ctx->task && cpuctx->task_ctx != ctx)
536                 return;
537
538         raw_spin_lock(&ctx->lock);
539
540         /*
541          * If the event is on, turn it off.
542          * If it is in error state, leave it in error state.
543          */
544         if (event->state >= PERF_EVENT_STATE_INACTIVE) {
545                 update_context_time(ctx);
546                 update_group_times(event);
547                 if (event == event->group_leader)
548                         group_sched_out(event, cpuctx, ctx);
549                 else
550                         event_sched_out(event, cpuctx, ctx);
551                 event->state = PERF_EVENT_STATE_OFF;
552         }
553
554         raw_spin_unlock(&ctx->lock);
555 }
556
557 /*
558  * Disable a event.
559  *
560  * If event->ctx is a cloned context, callers must make sure that
561  * every task struct that event->ctx->task could possibly point to
562  * remains valid.  This condition is satisifed when called through
563  * perf_event_for_each_child or perf_event_for_each because they
564  * hold the top-level event's child_mutex, so any descendant that
565  * goes to exit will block in sync_child_event.
566  * When called from perf_pending_event it's OK because event->ctx
567  * is the current context on this CPU and preemption is disabled,
568  * hence we can't get into perf_event_task_sched_out for this context.
569  */
570 void perf_event_disable(struct perf_event *event)
571 {
572         struct perf_event_context *ctx = event->ctx;
573         struct task_struct *task = ctx->task;
574
575         if (!task) {
576                 /*
577                  * Disable the event on the cpu that it's on
578                  */
579                 smp_call_function_single(event->cpu, __perf_event_disable,
580                                          event, 1);
581                 return;
582         }
583
584  retry:
585         task_oncpu_function_call(task, __perf_event_disable, event);
586
587         raw_spin_lock_irq(&ctx->lock);
588         /*
589          * If the event is still active, we need to retry the cross-call.
590          */
591         if (event->state == PERF_EVENT_STATE_ACTIVE) {
592                 raw_spin_unlock_irq(&ctx->lock);
593                 goto retry;
594         }
595
596         /*
597          * Since we have the lock this context can't be scheduled
598          * in, so we can change the state safely.
599          */
600         if (event->state == PERF_EVENT_STATE_INACTIVE) {
601                 update_group_times(event);
602                 event->state = PERF_EVENT_STATE_OFF;
603         }
604
605         raw_spin_unlock_irq(&ctx->lock);
606 }
607
608 static int
609 event_sched_in(struct perf_event *event,
610                  struct perf_cpu_context *cpuctx,
611                  struct perf_event_context *ctx,
612                  int cpu)
613 {
614         if (event->state <= PERF_EVENT_STATE_OFF)
615                 return 0;
616
617         event->state = PERF_EVENT_STATE_ACTIVE;
618         event->oncpu = cpu;     /* TODO: put 'cpu' into cpuctx->cpu */
619         /*
620          * The new state must be visible before we turn it on in the hardware:
621          */
622         smp_wmb();
623
624         if (event->pmu->enable(event)) {
625                 event->state = PERF_EVENT_STATE_INACTIVE;
626                 event->oncpu = -1;
627                 return -EAGAIN;
628         }
629
630         event->tstamp_running += ctx->time - event->tstamp_stopped;
631
632         if (!is_software_event(event))
633                 cpuctx->active_oncpu++;
634         ctx->nr_active++;
635
636         if (event->attr.exclusive)
637                 cpuctx->exclusive = 1;
638
639         return 0;
640 }
641
642 static int
643 group_sched_in(struct perf_event *group_event,
644                struct perf_cpu_context *cpuctx,
645                struct perf_event_context *ctx,
646                int cpu)
647 {
648         struct perf_event *event, *partial_group;
649         int ret;
650
651         if (group_event->state == PERF_EVENT_STATE_OFF)
652                 return 0;
653
654         ret = hw_perf_group_sched_in(group_event, cpuctx, ctx, cpu);
655         if (ret)
656                 return ret < 0 ? ret : 0;
657
658         if (event_sched_in(group_event, cpuctx, ctx, cpu))
659                 return -EAGAIN;
660
661         /*
662          * Schedule in siblings as one group (if any):
663          */
664         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
665                 if (event_sched_in(event, cpuctx, ctx, cpu)) {
666                         partial_group = event;
667                         goto group_error;
668                 }
669         }
670
671         return 0;
672
673 group_error:
674         /*
675          * Groups can be scheduled in as one unit only, so undo any
676          * partial group before returning:
677          */
678         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
679                 if (event == partial_group)
680                         break;
681                 event_sched_out(event, cpuctx, ctx);
682         }
683         event_sched_out(group_event, cpuctx, ctx);
684
685         return -EAGAIN;
686 }
687
688 /*
689  * Return 1 for a group consisting entirely of software events,
690  * 0 if the group contains any hardware events.
691  */
692 static int is_software_only_group(struct perf_event *leader)
693 {
694         struct perf_event *event;
695
696         if (!is_software_event(leader))
697                 return 0;
698
699         list_for_each_entry(event, &leader->sibling_list, group_entry)
700                 if (!is_software_event(event))
701                         return 0;
702
703         return 1;
704 }
705
706 /*
707  * Work out whether we can put this event group on the CPU now.
708  */
709 static int group_can_go_on(struct perf_event *event,
710                            struct perf_cpu_context *cpuctx,
711                            int can_add_hw)
712 {
713         /*
714          * Groups consisting entirely of software events can always go on.
715          */
716         if (is_software_only_group(event))
717                 return 1;
718         /*
719          * If an exclusive group is already on, no other hardware
720          * events can go on.
721          */
722         if (cpuctx->exclusive)
723                 return 0;
724         /*
725          * If this group is exclusive and there are already
726          * events on the CPU, it can't go on.
727          */
728         if (event->attr.exclusive && cpuctx->active_oncpu)
729                 return 0;
730         /*
731          * Otherwise, try to add it if all previous groups were able
732          * to go on.
733          */
734         return can_add_hw;
735 }
736
737 static void add_event_to_ctx(struct perf_event *event,
738                                struct perf_event_context *ctx)
739 {
740         list_add_event(event, ctx);
741         event->tstamp_enabled = ctx->time;
742         event->tstamp_running = ctx->time;
743         event->tstamp_stopped = ctx->time;
744 }
745
746 /*
747  * Cross CPU call to install and enable a performance event
748  *
749  * Must be called with ctx->mutex held
750  */
751 static void __perf_install_in_context(void *info)
752 {
753         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
754         struct perf_event *event = info;
755         struct perf_event_context *ctx = event->ctx;
756         struct perf_event *leader = event->group_leader;
757         int cpu = smp_processor_id();
758         int err;
759
760         /*
761          * If this is a task context, we need to check whether it is
762          * the current task context of this cpu. If not it has been
763          * scheduled out before the smp call arrived.
764          * Or possibly this is the right context but it isn't
765          * on this cpu because it had no events.
766          */
767         if (ctx->task && cpuctx->task_ctx != ctx) {
768                 if (cpuctx->task_ctx || ctx->task != current)
769                         return;
770                 cpuctx->task_ctx = ctx;
771         }
772
773         raw_spin_lock(&ctx->lock);
774         ctx->is_active = 1;
775         update_context_time(ctx);
776
777         /*
778          * Protect the list operation against NMI by disabling the
779          * events on a global level. NOP for non NMI based events.
780          */
781         perf_disable();
782
783         add_event_to_ctx(event, ctx);
784
785         if (event->cpu != -1 && event->cpu != smp_processor_id())
786                 goto unlock;
787
788         /*
789          * Don't put the event on if it is disabled or if
790          * it is in a group and the group isn't on.
791          */
792         if (event->state != PERF_EVENT_STATE_INACTIVE ||
793             (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
794                 goto unlock;
795
796         /*
797          * An exclusive event can't go on if there are already active
798          * hardware events, and no hardware event can go on if there
799          * is already an exclusive event on.
800          */
801         if (!group_can_go_on(event, cpuctx, 1))
802                 err = -EEXIST;
803         else
804                 err = event_sched_in(event, cpuctx, ctx, cpu);
805
806         if (err) {
807                 /*
808                  * This event couldn't go on.  If it is in a group
809                  * then we have to pull the whole group off.
810                  * If the event group is pinned then put it in error state.
811                  */
812                 if (leader != event)
813                         group_sched_out(leader, cpuctx, ctx);
814                 if (leader->attr.pinned) {
815                         update_group_times(leader);
816                         leader->state = PERF_EVENT_STATE_ERROR;
817                 }
818         }
819
820         if (!err && !ctx->task && cpuctx->max_pertask)
821                 cpuctx->max_pertask--;
822
823  unlock:
824         perf_enable();
825
826         raw_spin_unlock(&ctx->lock);
827 }
828
829 /*
830  * Attach a performance event to a context
831  *
832  * First we add the event to the list with the hardware enable bit
833  * in event->hw_config cleared.
834  *
835  * If the event is attached to a task which is on a CPU we use a smp
836  * call to enable it in the task context. The task might have been
837  * scheduled away, but we check this in the smp call again.
838  *
839  * Must be called with ctx->mutex held.
840  */
841 static void
842 perf_install_in_context(struct perf_event_context *ctx,
843                         struct perf_event *event,
844                         int cpu)
845 {
846         struct task_struct *task = ctx->task;
847
848         if (!task) {
849                 /*
850                  * Per cpu events are installed via an smp call and
851                  * the install is always successful.
852                  */
853                 smp_call_function_single(cpu, __perf_install_in_context,
854                                          event, 1);
855                 return;
856         }
857
858 retry:
859         task_oncpu_function_call(task, __perf_install_in_context,
860                                  event);
861
862         raw_spin_lock_irq(&ctx->lock);
863         /*
864          * we need to retry the smp call.
865          */
866         if (ctx->is_active && list_empty(&event->group_entry)) {
867                 raw_spin_unlock_irq(&ctx->lock);
868                 goto retry;
869         }
870
871         /*
872          * The lock prevents that this context is scheduled in so we
873          * can add the event safely, if it the call above did not
874          * succeed.
875          */
876         if (list_empty(&event->group_entry))
877                 add_event_to_ctx(event, ctx);
878         raw_spin_unlock_irq(&ctx->lock);
879 }
880
881 /*
882  * Put a event into inactive state and update time fields.
883  * Enabling the leader of a group effectively enables all
884  * the group members that aren't explicitly disabled, so we
885  * have to update their ->tstamp_enabled also.
886  * Note: this works for group members as well as group leaders
887  * since the non-leader members' sibling_lists will be empty.
888  */
889 static void __perf_event_mark_enabled(struct perf_event *event,
890                                         struct perf_event_context *ctx)
891 {
892         struct perf_event *sub;
893
894         event->state = PERF_EVENT_STATE_INACTIVE;
895         event->tstamp_enabled = ctx->time - event->total_time_enabled;
896         list_for_each_entry(sub, &event->sibling_list, group_entry)
897                 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
898                         sub->tstamp_enabled =
899                                 ctx->time - sub->total_time_enabled;
900 }
901
902 /*
903  * Cross CPU call to enable a performance event
904  */
905 static void __perf_event_enable(void *info)
906 {
907         struct perf_event *event = info;
908         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
909         struct perf_event_context *ctx = event->ctx;
910         struct perf_event *leader = event->group_leader;
911         int err;
912
913         /*
914          * If this is a per-task event, need to check whether this
915          * event's task is the current task on this cpu.
916          */
917         if (ctx->task && cpuctx->task_ctx != ctx) {
918                 if (cpuctx->task_ctx || ctx->task != current)
919                         return;
920                 cpuctx->task_ctx = ctx;
921         }
922
923         raw_spin_lock(&ctx->lock);
924         ctx->is_active = 1;
925         update_context_time(ctx);
926
927         if (event->state >= PERF_EVENT_STATE_INACTIVE)
928                 goto unlock;
929         __perf_event_mark_enabled(event, ctx);
930
931         if (event->cpu != -1 && event->cpu != smp_processor_id())
932                 goto unlock;
933
934         /*
935          * If the event is in a group and isn't the group leader,
936          * then don't put it on unless the group is on.
937          */
938         if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
939                 goto unlock;
940
941         if (!group_can_go_on(event, cpuctx, 1)) {
942                 err = -EEXIST;
943         } else {
944                 perf_disable();
945                 if (event == leader)
946                         err = group_sched_in(event, cpuctx, ctx,
947                                              smp_processor_id());
948                 else
949                         err = event_sched_in(event, cpuctx, ctx,
950                                                smp_processor_id());
951                 perf_enable();
952         }
953
954         if (err) {
955                 /*
956                  * If this event can't go on and it's part of a
957                  * group, then the whole group has to come off.
958                  */
959                 if (leader != event)
960                         group_sched_out(leader, cpuctx, ctx);
961                 if (leader->attr.pinned) {
962                         update_group_times(leader);
963                         leader->state = PERF_EVENT_STATE_ERROR;
964                 }
965         }
966
967  unlock:
968         raw_spin_unlock(&ctx->lock);
969 }
970
971 /*
972  * Enable a event.
973  *
974  * If event->ctx is a cloned context, callers must make sure that
975  * every task struct that event->ctx->task could possibly point to
976  * remains valid.  This condition is satisfied when called through
977  * perf_event_for_each_child or perf_event_for_each as described
978  * for perf_event_disable.
979  */
980 void perf_event_enable(struct perf_event *event)
981 {
982         struct perf_event_context *ctx = event->ctx;
983         struct task_struct *task = ctx->task;
984
985         if (!task) {
986                 /*
987                  * Enable the event on the cpu that it's on
988                  */
989                 smp_call_function_single(event->cpu, __perf_event_enable,
990                                          event, 1);
991                 return;
992         }
993
994         raw_spin_lock_irq(&ctx->lock);
995         if (event->state >= PERF_EVENT_STATE_INACTIVE)
996                 goto out;
997
998         /*
999          * If the event is in error state, clear that first.
1000          * That way, if we see the event in error state below, we
1001          * know that it has gone back into error state, as distinct
1002          * from the task having been scheduled away before the
1003          * cross-call arrived.
1004          */
1005         if (event->state == PERF_EVENT_STATE_ERROR)
1006                 event->state = PERF_EVENT_STATE_OFF;
1007
1008  retry:
1009         raw_spin_unlock_irq(&ctx->lock);
1010         task_oncpu_function_call(task, __perf_event_enable, event);
1011
1012         raw_spin_lock_irq(&ctx->lock);
1013
1014         /*
1015          * If the context is active and the event is still off,
1016          * we need to retry the cross-call.
1017          */
1018         if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1019                 goto retry;
1020
1021         /*
1022          * Since we have the lock this context can't be scheduled
1023          * in, so we can change the state safely.
1024          */
1025         if (event->state == PERF_EVENT_STATE_OFF)
1026                 __perf_event_mark_enabled(event, ctx);
1027
1028  out:
1029         raw_spin_unlock_irq(&ctx->lock);
1030 }
1031
1032 static int perf_event_refresh(struct perf_event *event, int refresh)
1033 {
1034         /*
1035          * not supported on inherited events
1036          */
1037         if (event->attr.inherit)
1038                 return -EINVAL;
1039
1040         atomic_add(refresh, &event->event_limit);
1041         perf_event_enable(event);
1042
1043         return 0;
1044 }
1045
1046 void __perf_event_sched_out(struct perf_event_context *ctx,
1047                               struct perf_cpu_context *cpuctx)
1048 {
1049         struct perf_event *event;
1050
1051         raw_spin_lock(&ctx->lock);
1052         ctx->is_active = 0;
1053         if (likely(!ctx->nr_events))
1054                 goto out;
1055         update_context_time(ctx);
1056
1057         perf_disable();
1058         if (ctx->nr_active) {
1059                 list_for_each_entry(event, &ctx->group_list, group_entry)
1060                         group_sched_out(event, cpuctx, ctx);
1061         }
1062         perf_enable();
1063  out:
1064         raw_spin_unlock(&ctx->lock);
1065 }
1066
1067 /*
1068  * Test whether two contexts are equivalent, i.e. whether they
1069  * have both been cloned from the same version of the same context
1070  * and they both have the same number of enabled events.
1071  * If the number of enabled events is the same, then the set
1072  * of enabled events should be the same, because these are both
1073  * inherited contexts, therefore we can't access individual events
1074  * in them directly with an fd; we can only enable/disable all
1075  * events via prctl, or enable/disable all events in a family
1076  * via ioctl, which will have the same effect on both contexts.
1077  */
1078 static int context_equiv(struct perf_event_context *ctx1,
1079                          struct perf_event_context *ctx2)
1080 {
1081         return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1082                 && ctx1->parent_gen == ctx2->parent_gen
1083                 && !ctx1->pin_count && !ctx2->pin_count;
1084 }
1085
1086 static void __perf_event_sync_stat(struct perf_event *event,
1087                                      struct perf_event *next_event)
1088 {
1089         u64 value;
1090
1091         if (!event->attr.inherit_stat)
1092                 return;
1093
1094         /*
1095          * Update the event value, we cannot use perf_event_read()
1096          * because we're in the middle of a context switch and have IRQs
1097          * disabled, which upsets smp_call_function_single(), however
1098          * we know the event must be on the current CPU, therefore we
1099          * don't need to use it.
1100          */
1101         switch (event->state) {
1102         case PERF_EVENT_STATE_ACTIVE:
1103                 event->pmu->read(event);
1104                 /* fall-through */
1105
1106         case PERF_EVENT_STATE_INACTIVE:
1107                 update_event_times(event);
1108                 break;
1109
1110         default:
1111                 break;
1112         }
1113
1114         /*
1115          * In order to keep per-task stats reliable we need to flip the event
1116          * values when we flip the contexts.
1117          */
1118         value = atomic64_read(&next_event->count);
1119         value = atomic64_xchg(&event->count, value);
1120         atomic64_set(&next_event->count, value);
1121
1122         swap(event->total_time_enabled, next_event->total_time_enabled);
1123         swap(event->total_time_running, next_event->total_time_running);
1124
1125         /*
1126          * Since we swizzled the values, update the user visible data too.
1127          */
1128         perf_event_update_userpage(event);
1129         perf_event_update_userpage(next_event);
1130 }
1131
1132 #define list_next_entry(pos, member) \
1133         list_entry(pos->member.next, typeof(*pos), member)
1134
1135 static void perf_event_sync_stat(struct perf_event_context *ctx,
1136                                    struct perf_event_context *next_ctx)
1137 {
1138         struct perf_event *event, *next_event;
1139
1140         if (!ctx->nr_stat)
1141                 return;
1142
1143         update_context_time(ctx);
1144
1145         event = list_first_entry(&ctx->event_list,
1146                                    struct perf_event, event_entry);
1147
1148         next_event = list_first_entry(&next_ctx->event_list,
1149                                         struct perf_event, event_entry);
1150
1151         while (&event->event_entry != &ctx->event_list &&
1152                &next_event->event_entry != &next_ctx->event_list) {
1153
1154                 __perf_event_sync_stat(event, next_event);
1155
1156                 event = list_next_entry(event, event_entry);
1157                 next_event = list_next_entry(next_event, event_entry);
1158         }
1159 }
1160
1161 /*
1162  * Called from scheduler to remove the events of the current task,
1163  * with interrupts disabled.
1164  *
1165  * We stop each event and update the event value in event->count.
1166  *
1167  * This does not protect us against NMI, but disable()
1168  * sets the disabled bit in the control field of event _before_
1169  * accessing the event control register. If a NMI hits, then it will
1170  * not restart the event.
1171  */
1172 void perf_event_task_sched_out(struct task_struct *task,
1173                                  struct task_struct *next, int cpu)
1174 {
1175         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1176         struct perf_event_context *ctx = task->perf_event_ctxp;
1177         struct perf_event_context *next_ctx;
1178         struct perf_event_context *parent;
1179         struct pt_regs *regs;
1180         int do_switch = 1;
1181
1182         regs = task_pt_regs(task);
1183         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1184
1185         if (likely(!ctx || !cpuctx->task_ctx))
1186                 return;
1187
1188         rcu_read_lock();
1189         parent = rcu_dereference(ctx->parent_ctx);
1190         next_ctx = next->perf_event_ctxp;
1191         if (parent && next_ctx &&
1192             rcu_dereference(next_ctx->parent_ctx) == parent) {
1193                 /*
1194                  * Looks like the two contexts are clones, so we might be
1195                  * able to optimize the context switch.  We lock both
1196                  * contexts and check that they are clones under the
1197                  * lock (including re-checking that neither has been
1198                  * uncloned in the meantime).  It doesn't matter which
1199                  * order we take the locks because no other cpu could
1200                  * be trying to lock both of these tasks.
1201                  */
1202                 raw_spin_lock(&ctx->lock);
1203                 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1204                 if (context_equiv(ctx, next_ctx)) {
1205                         /*
1206                          * XXX do we need a memory barrier of sorts
1207                          * wrt to rcu_dereference() of perf_event_ctxp
1208                          */
1209                         task->perf_event_ctxp = next_ctx;
1210                         next->perf_event_ctxp = ctx;
1211                         ctx->task = next;
1212                         next_ctx->task = task;
1213                         do_switch = 0;
1214
1215                         perf_event_sync_stat(ctx, next_ctx);
1216                 }
1217                 raw_spin_unlock(&next_ctx->lock);
1218                 raw_spin_unlock(&ctx->lock);
1219         }
1220         rcu_read_unlock();
1221
1222         if (do_switch) {
1223                 __perf_event_sched_out(ctx, cpuctx);
1224                 cpuctx->task_ctx = NULL;
1225         }
1226 }
1227
1228 /*
1229  * Called with IRQs disabled
1230  */
1231 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1232 {
1233         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1234
1235         if (!cpuctx->task_ctx)
1236                 return;
1237
1238         if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1239                 return;
1240
1241         __perf_event_sched_out(ctx, cpuctx);
1242         cpuctx->task_ctx = NULL;
1243 }
1244
1245 /*
1246  * Called with IRQs disabled
1247  */
1248 static void perf_event_cpu_sched_out(struct perf_cpu_context *cpuctx)
1249 {
1250         __perf_event_sched_out(&cpuctx->ctx, cpuctx);
1251 }
1252
1253 static void
1254 __perf_event_sched_in(struct perf_event_context *ctx,
1255                         struct perf_cpu_context *cpuctx, int cpu)
1256 {
1257         struct perf_event *event;
1258         int can_add_hw = 1;
1259
1260         raw_spin_lock(&ctx->lock);
1261         ctx->is_active = 1;
1262         if (likely(!ctx->nr_events))
1263                 goto out;
1264
1265         ctx->timestamp = perf_clock();
1266
1267         perf_disable();
1268
1269         /*
1270          * First go through the list and put on any pinned groups
1271          * in order to give them the best chance of going on.
1272          */
1273         list_for_each_entry(event, &ctx->group_list, group_entry) {
1274                 if (event->state <= PERF_EVENT_STATE_OFF ||
1275                     !event->attr.pinned)
1276                         continue;
1277                 if (event->cpu != -1 && event->cpu != cpu)
1278                         continue;
1279
1280                 if (group_can_go_on(event, cpuctx, 1))
1281                         group_sched_in(event, cpuctx, ctx, cpu);
1282
1283                 /*
1284                  * If this pinned group hasn't been scheduled,
1285                  * put it in error state.
1286                  */
1287                 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1288                         update_group_times(event);
1289                         event->state = PERF_EVENT_STATE_ERROR;
1290                 }
1291         }
1292
1293         list_for_each_entry(event, &ctx->group_list, group_entry) {
1294                 /*
1295                  * Ignore events in OFF or ERROR state, and
1296                  * ignore pinned events since we did them already.
1297                  */
1298                 if (event->state <= PERF_EVENT_STATE_OFF ||
1299                     event->attr.pinned)
1300                         continue;
1301
1302                 /*
1303                  * Listen to the 'cpu' scheduling filter constraint
1304                  * of events:
1305                  */
1306                 if (event->cpu != -1 && event->cpu != cpu)
1307                         continue;
1308
1309                 if (group_can_go_on(event, cpuctx, can_add_hw))
1310                         if (group_sched_in(event, cpuctx, ctx, cpu))
1311                                 can_add_hw = 0;
1312         }
1313         perf_enable();
1314  out:
1315         raw_spin_unlock(&ctx->lock);
1316 }
1317
1318 /*
1319  * Called from scheduler to add the events of the current task
1320  * with interrupts disabled.
1321  *
1322  * We restore the event value and then enable it.
1323  *
1324  * This does not protect us against NMI, but enable()
1325  * sets the enabled bit in the control field of event _before_
1326  * accessing the event control register. If a NMI hits, then it will
1327  * keep the event running.
1328  */
1329 void perf_event_task_sched_in(struct task_struct *task, int cpu)
1330 {
1331         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1332         struct perf_event_context *ctx = task->perf_event_ctxp;
1333
1334         if (likely(!ctx))
1335                 return;
1336         if (cpuctx->task_ctx == ctx)
1337                 return;
1338         __perf_event_sched_in(ctx, cpuctx, cpu);
1339         cpuctx->task_ctx = ctx;
1340 }
1341
1342 static void perf_event_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
1343 {
1344         struct perf_event_context *ctx = &cpuctx->ctx;
1345
1346         __perf_event_sched_in(ctx, cpuctx, cpu);
1347 }
1348
1349 #define MAX_INTERRUPTS (~0ULL)
1350
1351 static void perf_log_throttle(struct perf_event *event, int enable);
1352
1353 static void perf_adjust_period(struct perf_event *event, u64 events)
1354 {
1355         struct hw_perf_event *hwc = &event->hw;
1356         u64 period, sample_period;
1357         s64 delta;
1358
1359         events *= hwc->sample_period;
1360         period = div64_u64(events, event->attr.sample_freq);
1361
1362         delta = (s64)(period - hwc->sample_period);
1363         delta = (delta + 7) / 8; /* low pass filter */
1364
1365         sample_period = hwc->sample_period + delta;
1366
1367         if (!sample_period)
1368                 sample_period = 1;
1369
1370         hwc->sample_period = sample_period;
1371 }
1372
1373 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1374 {
1375         struct perf_event *event;
1376         struct hw_perf_event *hwc;
1377         u64 interrupts, freq;
1378
1379         raw_spin_lock(&ctx->lock);
1380         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1381                 if (event->state != PERF_EVENT_STATE_ACTIVE)
1382                         continue;
1383
1384                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1385                         continue;
1386
1387                 hwc = &event->hw;
1388
1389                 interrupts = hwc->interrupts;
1390                 hwc->interrupts = 0;
1391
1392                 /*
1393                  * unthrottle events on the tick
1394                  */
1395                 if (interrupts == MAX_INTERRUPTS) {
1396                         perf_log_throttle(event, 1);
1397                         event->pmu->unthrottle(event);
1398                         interrupts = 2*sysctl_perf_event_sample_rate/HZ;
1399                 }
1400
1401                 if (!event->attr.freq || !event->attr.sample_freq)
1402                         continue;
1403
1404                 /*
1405                  * if the specified freq < HZ then we need to skip ticks
1406                  */
1407                 if (event->attr.sample_freq < HZ) {
1408                         freq = event->attr.sample_freq;
1409
1410                         hwc->freq_count += freq;
1411                         hwc->freq_interrupts += interrupts;
1412
1413                         if (hwc->freq_count < HZ)
1414                                 continue;
1415
1416                         interrupts = hwc->freq_interrupts;
1417                         hwc->freq_interrupts = 0;
1418                         hwc->freq_count -= HZ;
1419                 } else
1420                         freq = HZ;
1421
1422                 perf_adjust_period(event, freq * interrupts);
1423
1424                 /*
1425                  * In order to avoid being stalled by an (accidental) huge
1426                  * sample period, force reset the sample period if we didn't
1427                  * get any events in this freq period.
1428                  */
1429                 if (!interrupts) {
1430                         perf_disable();
1431                         event->pmu->disable(event);
1432                         atomic64_set(&hwc->period_left, 0);
1433                         event->pmu->enable(event);
1434                         perf_enable();
1435                 }
1436         }
1437         raw_spin_unlock(&ctx->lock);
1438 }
1439
1440 /*
1441  * Round-robin a context's events:
1442  */
1443 static void rotate_ctx(struct perf_event_context *ctx)
1444 {
1445         struct perf_event *event;
1446
1447         if (!ctx->nr_events)
1448                 return;
1449
1450         raw_spin_lock(&ctx->lock);
1451         /*
1452          * Rotate the first entry last (works just fine for group events too):
1453          */
1454         perf_disable();
1455         list_for_each_entry(event, &ctx->group_list, group_entry) {
1456                 list_move_tail(&event->group_entry, &ctx->group_list);
1457                 break;
1458         }
1459         perf_enable();
1460
1461         raw_spin_unlock(&ctx->lock);
1462 }
1463
1464 void perf_event_task_tick(struct task_struct *curr, int cpu)
1465 {
1466         struct perf_cpu_context *cpuctx;
1467         struct perf_event_context *ctx;
1468
1469         if (!atomic_read(&nr_events))
1470                 return;
1471
1472         cpuctx = &per_cpu(perf_cpu_context, cpu);
1473         ctx = curr->perf_event_ctxp;
1474
1475         perf_ctx_adjust_freq(&cpuctx->ctx);
1476         if (ctx)
1477                 perf_ctx_adjust_freq(ctx);
1478
1479         perf_event_cpu_sched_out(cpuctx);
1480         if (ctx)
1481                 __perf_event_task_sched_out(ctx);
1482
1483         rotate_ctx(&cpuctx->ctx);
1484         if (ctx)
1485                 rotate_ctx(ctx);
1486
1487         perf_event_cpu_sched_in(cpuctx, cpu);
1488         if (ctx)
1489                 perf_event_task_sched_in(curr, cpu);
1490 }
1491
1492 /*
1493  * Enable all of a task's events that have been marked enable-on-exec.
1494  * This expects task == current.
1495  */
1496 static void perf_event_enable_on_exec(struct task_struct *task)
1497 {
1498         struct perf_event_context *ctx;
1499         struct perf_event *event;
1500         unsigned long flags;
1501         int enabled = 0;
1502
1503         local_irq_save(flags);
1504         ctx = task->perf_event_ctxp;
1505         if (!ctx || !ctx->nr_events)
1506                 goto out;
1507
1508         __perf_event_task_sched_out(ctx);
1509
1510         raw_spin_lock(&ctx->lock);
1511
1512         list_for_each_entry(event, &ctx->group_list, group_entry) {
1513                 if (!event->attr.enable_on_exec)
1514                         continue;
1515                 event->attr.enable_on_exec = 0;
1516                 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1517                         continue;
1518                 __perf_event_mark_enabled(event, ctx);
1519                 enabled = 1;
1520         }
1521
1522         /*
1523          * Unclone this context if we enabled any event.
1524          */
1525         if (enabled)
1526                 unclone_ctx(ctx);
1527
1528         raw_spin_unlock(&ctx->lock);
1529
1530         perf_event_task_sched_in(task, smp_processor_id());
1531  out:
1532         local_irq_restore(flags);
1533 }
1534
1535 /*
1536  * Cross CPU call to read the hardware event
1537  */
1538 static void __perf_event_read(void *info)
1539 {
1540         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1541         struct perf_event *event = info;
1542         struct perf_event_context *ctx = event->ctx;
1543
1544         /*
1545          * If this is a task context, we need to check whether it is
1546          * the current task context of this cpu.  If not it has been
1547          * scheduled out before the smp call arrived.  In that case
1548          * event->count would have been updated to a recent sample
1549          * when the event was scheduled out.
1550          */
1551         if (ctx->task && cpuctx->task_ctx != ctx)
1552                 return;
1553
1554         raw_spin_lock(&ctx->lock);
1555         update_context_time(ctx);
1556         update_event_times(event);
1557         raw_spin_unlock(&ctx->lock);
1558
1559         event->pmu->read(event);
1560 }
1561
1562 static u64 perf_event_read(struct perf_event *event)
1563 {
1564         /*
1565          * If event is enabled and currently active on a CPU, update the
1566          * value in the event structure:
1567          */
1568         if (event->state == PERF_EVENT_STATE_ACTIVE) {
1569                 smp_call_function_single(event->oncpu,
1570                                          __perf_event_read, event, 1);
1571         } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1572                 struct perf_event_context *ctx = event->ctx;
1573                 unsigned long flags;
1574
1575                 raw_spin_lock_irqsave(&ctx->lock, flags);
1576                 update_context_time(ctx);
1577                 update_event_times(event);
1578                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1579         }
1580
1581         return atomic64_read(&event->count);
1582 }
1583
1584 /*
1585  * Initialize the perf_event context in a task_struct:
1586  */
1587 static void
1588 __perf_event_init_context(struct perf_event_context *ctx,
1589                             struct task_struct *task)
1590 {
1591         raw_spin_lock_init(&ctx->lock);
1592         mutex_init(&ctx->mutex);
1593         INIT_LIST_HEAD(&ctx->group_list);
1594         INIT_LIST_HEAD(&ctx->event_list);
1595         atomic_set(&ctx->refcount, 1);
1596         ctx->task = task;
1597 }
1598
1599 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1600 {
1601         struct perf_event_context *ctx;
1602         struct perf_cpu_context *cpuctx;
1603         struct task_struct *task;
1604         unsigned long flags;
1605         int err;
1606
1607         if (pid == -1 && cpu != -1) {
1608                 /* Must be root to operate on a CPU event: */
1609                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1610                         return ERR_PTR(-EACCES);
1611
1612                 if (cpu < 0 || cpu >= nr_cpumask_bits)
1613                         return ERR_PTR(-EINVAL);
1614
1615                 /*
1616                  * We could be clever and allow to attach a event to an
1617                  * offline CPU and activate it when the CPU comes up, but
1618                  * that's for later.
1619                  */
1620                 if (!cpu_online(cpu))
1621                         return ERR_PTR(-ENODEV);
1622
1623                 cpuctx = &per_cpu(perf_cpu_context, cpu);
1624                 ctx = &cpuctx->ctx;
1625                 get_ctx(ctx);
1626
1627                 return ctx;
1628         }
1629
1630         rcu_read_lock();
1631         if (!pid)
1632                 task = current;
1633         else
1634                 task = find_task_by_vpid(pid);
1635         if (task)
1636                 get_task_struct(task);
1637         rcu_read_unlock();
1638
1639         if (!task)
1640                 return ERR_PTR(-ESRCH);
1641
1642         /*
1643          * Can't attach events to a dying task.
1644          */
1645         err = -ESRCH;
1646         if (task->flags & PF_EXITING)
1647                 goto errout;
1648
1649         /* Reuse ptrace permission checks for now. */
1650         err = -EACCES;
1651         if (!ptrace_may_access(task, PTRACE_MODE_READ))
1652                 goto errout;
1653
1654  retry:
1655         ctx = perf_lock_task_context(task, &flags);
1656         if (ctx) {
1657                 unclone_ctx(ctx);
1658                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1659         }
1660
1661         if (!ctx) {
1662                 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1663                 err = -ENOMEM;
1664                 if (!ctx)
1665                         goto errout;
1666                 __perf_event_init_context(ctx, task);
1667                 get_ctx(ctx);
1668                 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1669                         /*
1670                          * We raced with some other task; use
1671                          * the context they set.
1672                          */
1673                         kfree(ctx);
1674                         goto retry;
1675                 }
1676                 get_task_struct(task);
1677         }
1678
1679         put_task_struct(task);
1680         return ctx;
1681
1682  errout:
1683         put_task_struct(task);
1684         return ERR_PTR(err);
1685 }
1686
1687 static void perf_event_free_filter(struct perf_event *event);
1688
1689 static void free_event_rcu(struct rcu_head *head)
1690 {
1691         struct perf_event *event;
1692
1693         event = container_of(head, struct perf_event, rcu_head);
1694         if (event->ns)
1695                 put_pid_ns(event->ns);
1696         perf_event_free_filter(event);
1697         kfree(event);
1698 }
1699
1700 static void perf_pending_sync(struct perf_event *event);
1701
1702 static void free_event(struct perf_event *event)
1703 {
1704         perf_pending_sync(event);
1705
1706         if (!event->parent) {
1707                 atomic_dec(&nr_events);
1708                 if (event->attr.mmap)
1709                         atomic_dec(&nr_mmap_events);
1710                 if (event->attr.comm)
1711                         atomic_dec(&nr_comm_events);
1712                 if (event->attr.task)
1713                         atomic_dec(&nr_task_events);
1714         }
1715
1716         if (event->output) {
1717                 fput(event->output->filp);
1718                 event->output = NULL;
1719         }
1720
1721         if (event->destroy)
1722                 event->destroy(event);
1723
1724         put_ctx(event->ctx);
1725         call_rcu(&event->rcu_head, free_event_rcu);
1726 }
1727
1728 int perf_event_release_kernel(struct perf_event *event)
1729 {
1730         struct perf_event_context *ctx = event->ctx;
1731
1732         WARN_ON_ONCE(ctx->parent_ctx);
1733         mutex_lock(&ctx->mutex);
1734         perf_event_remove_from_context(event);
1735         mutex_unlock(&ctx->mutex);
1736
1737         mutex_lock(&event->owner->perf_event_mutex);
1738         list_del_init(&event->owner_entry);
1739         mutex_unlock(&event->owner->perf_event_mutex);
1740         put_task_struct(event->owner);
1741
1742         free_event(event);
1743
1744         return 0;
1745 }
1746 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1747
1748 /*
1749  * Called when the last reference to the file is gone.
1750  */
1751 static int perf_release(struct inode *inode, struct file *file)
1752 {
1753         struct perf_event *event = file->private_data;
1754
1755         file->private_data = NULL;
1756
1757         return perf_event_release_kernel(event);
1758 }
1759
1760 static int perf_event_read_size(struct perf_event *event)
1761 {
1762         int entry = sizeof(u64); /* value */
1763         int size = 0;
1764         int nr = 1;
1765
1766         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1767                 size += sizeof(u64);
1768
1769         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1770                 size += sizeof(u64);
1771
1772         if (event->attr.read_format & PERF_FORMAT_ID)
1773                 entry += sizeof(u64);
1774
1775         if (event->attr.read_format & PERF_FORMAT_GROUP) {
1776                 nr += event->group_leader->nr_siblings;
1777                 size += sizeof(u64);
1778         }
1779
1780         size += entry * nr;
1781
1782         return size;
1783 }
1784
1785 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1786 {
1787         struct perf_event *child;
1788         u64 total = 0;
1789
1790         *enabled = 0;
1791         *running = 0;
1792
1793         mutex_lock(&event->child_mutex);
1794         total += perf_event_read(event);
1795         *enabled += event->total_time_enabled +
1796                         atomic64_read(&event->child_total_time_enabled);
1797         *running += event->total_time_running +
1798                         atomic64_read(&event->child_total_time_running);
1799
1800         list_for_each_entry(child, &event->child_list, child_list) {
1801                 total += perf_event_read(child);
1802                 *enabled += child->total_time_enabled;
1803                 *running += child->total_time_running;
1804         }
1805         mutex_unlock(&event->child_mutex);
1806
1807         return total;
1808 }
1809 EXPORT_SYMBOL_GPL(perf_event_read_value);
1810
1811 static int perf_event_read_group(struct perf_event *event,
1812                                    u64 read_format, char __user *buf)
1813 {
1814         struct perf_event *leader = event->group_leader, *sub;
1815         int n = 0, size = 0, ret = -EFAULT;
1816         struct perf_event_context *ctx = leader->ctx;
1817         u64 values[5];
1818         u64 count, enabled, running;
1819
1820         mutex_lock(&ctx->mutex);
1821         count = perf_event_read_value(leader, &enabled, &running);
1822
1823         values[n++] = 1 + leader->nr_siblings;
1824         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1825                 values[n++] = enabled;
1826         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1827                 values[n++] = running;
1828         values[n++] = count;
1829         if (read_format & PERF_FORMAT_ID)
1830                 values[n++] = primary_event_id(leader);
1831
1832         size = n * sizeof(u64);
1833
1834         if (copy_to_user(buf, values, size))
1835                 goto unlock;
1836
1837         ret = size;
1838
1839         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1840                 n = 0;
1841
1842                 values[n++] = perf_event_read_value(sub, &enabled, &running);
1843                 if (read_format & PERF_FORMAT_ID)
1844                         values[n++] = primary_event_id(sub);
1845
1846                 size = n * sizeof(u64);
1847
1848                 if (copy_to_user(buf + ret, values, size)) {
1849                         ret = -EFAULT;
1850                         goto unlock;
1851                 }
1852
1853                 ret += size;
1854         }
1855 unlock:
1856         mutex_unlock(&ctx->mutex);
1857
1858         return ret;
1859 }
1860
1861 static int perf_event_read_one(struct perf_event *event,
1862                                  u64 read_format, char __user *buf)
1863 {
1864         u64 enabled, running;
1865         u64 values[4];
1866         int n = 0;
1867
1868         values[n++] = perf_event_read_value(event, &enabled, &running);
1869         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1870                 values[n++] = enabled;
1871         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1872                 values[n++] = running;
1873         if (read_format & PERF_FORMAT_ID)
1874                 values[n++] = primary_event_id(event);
1875
1876         if (copy_to_user(buf, values, n * sizeof(u64)))
1877                 return -EFAULT;
1878
1879         return n * sizeof(u64);
1880 }
1881
1882 /*
1883  * Read the performance event - simple non blocking version for now
1884  */
1885 static ssize_t
1886 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
1887 {
1888         u64 read_format = event->attr.read_format;
1889         int ret;
1890
1891         /*
1892          * Return end-of-file for a read on a event that is in
1893          * error state (i.e. because it was pinned but it couldn't be
1894          * scheduled on to the CPU at some point).
1895          */
1896         if (event->state == PERF_EVENT_STATE_ERROR)
1897                 return 0;
1898
1899         if (count < perf_event_read_size(event))
1900                 return -ENOSPC;
1901
1902         WARN_ON_ONCE(event->ctx->parent_ctx);
1903         if (read_format & PERF_FORMAT_GROUP)
1904                 ret = perf_event_read_group(event, read_format, buf);
1905         else
1906                 ret = perf_event_read_one(event, read_format, buf);
1907
1908         return ret;
1909 }
1910
1911 static ssize_t
1912 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1913 {
1914         struct perf_event *event = file->private_data;
1915
1916         return perf_read_hw(event, buf, count);
1917 }
1918
1919 static unsigned int perf_poll(struct file *file, poll_table *wait)
1920 {
1921         struct perf_event *event = file->private_data;
1922         struct perf_mmap_data *data;
1923         unsigned int events = POLL_HUP;
1924
1925         rcu_read_lock();
1926         data = rcu_dereference(event->data);
1927         if (data)
1928                 events = atomic_xchg(&data->poll, 0);
1929         rcu_read_unlock();
1930
1931         poll_wait(file, &event->waitq, wait);
1932
1933         return events;
1934 }
1935
1936 static void perf_event_reset(struct perf_event *event)
1937 {
1938         (void)perf_event_read(event);
1939         atomic64_set(&event->count, 0);
1940         perf_event_update_userpage(event);
1941 }
1942
1943 /*
1944  * Holding the top-level event's child_mutex means that any
1945  * descendant process that has inherited this event will block
1946  * in sync_child_event if it goes to exit, thus satisfying the
1947  * task existence requirements of perf_event_enable/disable.
1948  */
1949 static void perf_event_for_each_child(struct perf_event *event,
1950                                         void (*func)(struct perf_event *))
1951 {
1952         struct perf_event *child;
1953
1954         WARN_ON_ONCE(event->ctx->parent_ctx);
1955         mutex_lock(&event->child_mutex);
1956         func(event);
1957         list_for_each_entry(child, &event->child_list, child_list)
1958                 func(child);
1959         mutex_unlock(&event->child_mutex);
1960 }
1961
1962 static void perf_event_for_each(struct perf_event *event,
1963                                   void (*func)(struct perf_event *))
1964 {
1965         struct perf_event_context *ctx = event->ctx;
1966         struct perf_event *sibling;
1967
1968         WARN_ON_ONCE(ctx->parent_ctx);
1969         mutex_lock(&ctx->mutex);
1970         event = event->group_leader;
1971
1972         perf_event_for_each_child(event, func);
1973         func(event);
1974         list_for_each_entry(sibling, &event->sibling_list, group_entry)
1975                 perf_event_for_each_child(event, func);
1976         mutex_unlock(&ctx->mutex);
1977 }
1978
1979 static int perf_event_period(struct perf_event *event, u64 __user *arg)
1980 {
1981         struct perf_event_context *ctx = event->ctx;
1982         unsigned long size;
1983         int ret = 0;
1984         u64 value;
1985
1986         if (!event->attr.sample_period)
1987                 return -EINVAL;
1988
1989         size = copy_from_user(&value, arg, sizeof(value));
1990         if (size != sizeof(value))
1991                 return -EFAULT;
1992
1993         if (!value)
1994                 return -EINVAL;
1995
1996         raw_spin_lock_irq(&ctx->lock);
1997         if (event->attr.freq) {
1998                 if (value > sysctl_perf_event_sample_rate) {
1999                         ret = -EINVAL;
2000                         goto unlock;
2001                 }
2002
2003                 event->attr.sample_freq = value;
2004         } else {
2005                 event->attr.sample_period = value;
2006                 event->hw.sample_period = value;
2007         }
2008 unlock:
2009         raw_spin_unlock_irq(&ctx->lock);
2010
2011         return ret;
2012 }
2013
2014 static int perf_event_set_output(struct perf_event *event, int output_fd);
2015 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2016
2017 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2018 {
2019         struct perf_event *event = file->private_data;
2020         void (*func)(struct perf_event *);
2021         u32 flags = arg;
2022
2023         switch (cmd) {
2024         case PERF_EVENT_IOC_ENABLE:
2025                 func = perf_event_enable;
2026                 break;
2027         case PERF_EVENT_IOC_DISABLE:
2028                 func = perf_event_disable;
2029                 break;
2030         case PERF_EVENT_IOC_RESET:
2031                 func = perf_event_reset;
2032                 break;
2033
2034         case PERF_EVENT_IOC_REFRESH:
2035                 return perf_event_refresh(event, arg);
2036
2037         case PERF_EVENT_IOC_PERIOD:
2038                 return perf_event_period(event, (u64 __user *)arg);
2039
2040         case PERF_EVENT_IOC_SET_OUTPUT:
2041                 return perf_event_set_output(event, arg);
2042
2043         case PERF_EVENT_IOC_SET_FILTER:
2044                 return perf_event_set_filter(event, (void __user *)arg);
2045
2046         default:
2047                 return -ENOTTY;
2048         }
2049
2050         if (flags & PERF_IOC_FLAG_GROUP)
2051                 perf_event_for_each(event, func);
2052         else
2053                 perf_event_for_each_child(event, func);
2054
2055         return 0;
2056 }
2057
2058 int perf_event_task_enable(void)
2059 {
2060         struct perf_event *event;
2061
2062         mutex_lock(&current->perf_event_mutex);
2063         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2064                 perf_event_for_each_child(event, perf_event_enable);
2065         mutex_unlock(&current->perf_event_mutex);
2066
2067         return 0;
2068 }
2069
2070 int perf_event_task_disable(void)
2071 {
2072         struct perf_event *event;
2073
2074         mutex_lock(&current->perf_event_mutex);
2075         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2076                 perf_event_for_each_child(event, perf_event_disable);
2077         mutex_unlock(&current->perf_event_mutex);
2078
2079         return 0;
2080 }
2081
2082 #ifndef PERF_EVENT_INDEX_OFFSET
2083 # define PERF_EVENT_INDEX_OFFSET 0
2084 #endif
2085
2086 static int perf_event_index(struct perf_event *event)
2087 {
2088         if (event->state != PERF_EVENT_STATE_ACTIVE)
2089                 return 0;
2090
2091         return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2092 }
2093
2094 /*
2095  * Callers need to ensure there can be no nesting of this function, otherwise
2096  * the seqlock logic goes bad. We can not serialize this because the arch
2097  * code calls this from NMI context.
2098  */
2099 void perf_event_update_userpage(struct perf_event *event)
2100 {
2101         struct perf_event_mmap_page *userpg;
2102         struct perf_mmap_data *data;
2103
2104         rcu_read_lock();
2105         data = rcu_dereference(event->data);
2106         if (!data)
2107                 goto unlock;
2108
2109         userpg = data->user_page;
2110
2111         /*
2112          * Disable preemption so as to not let the corresponding user-space
2113          * spin too long if we get preempted.
2114          */
2115         preempt_disable();
2116         ++userpg->lock;
2117         barrier();
2118         userpg->index = perf_event_index(event);
2119         userpg->offset = atomic64_read(&event->count);
2120         if (event->state == PERF_EVENT_STATE_ACTIVE)
2121                 userpg->offset -= atomic64_read(&event->hw.prev_count);
2122
2123         userpg->time_enabled = event->total_time_enabled +
2124                         atomic64_read(&event->child_total_time_enabled);
2125
2126         userpg->time_running = event->total_time_running +
2127                         atomic64_read(&event->child_total_time_running);
2128
2129         barrier();
2130         ++userpg->lock;
2131         preempt_enable();
2132 unlock:
2133         rcu_read_unlock();
2134 }
2135
2136 static unsigned long perf_data_size(struct perf_mmap_data *data)
2137 {
2138         return data->nr_pages << (PAGE_SHIFT + data->data_order);
2139 }
2140
2141 #ifndef CONFIG_PERF_USE_VMALLOC
2142
2143 /*
2144  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2145  */
2146
2147 static struct page *
2148 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2149 {
2150         if (pgoff > data->nr_pages)
2151                 return NULL;
2152
2153         if (pgoff == 0)
2154                 return virt_to_page(data->user_page);
2155
2156         return virt_to_page(data->data_pages[pgoff - 1]);
2157 }
2158
2159 static struct perf_mmap_data *
2160 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2161 {
2162         struct perf_mmap_data *data;
2163         unsigned long size;
2164         int i;
2165
2166         WARN_ON(atomic_read(&event->mmap_count));
2167
2168         size = sizeof(struct perf_mmap_data);
2169         size += nr_pages * sizeof(void *);
2170
2171         data = kzalloc(size, GFP_KERNEL);
2172         if (!data)
2173                 goto fail;
2174
2175         data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2176         if (!data->user_page)
2177                 goto fail_user_page;
2178
2179         for (i = 0; i < nr_pages; i++) {
2180                 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2181                 if (!data->data_pages[i])
2182                         goto fail_data_pages;
2183         }
2184
2185         data->data_order = 0;
2186         data->nr_pages = nr_pages;
2187
2188         return data;
2189
2190 fail_data_pages:
2191         for (i--; i >= 0; i--)
2192                 free_page((unsigned long)data->data_pages[i]);
2193
2194         free_page((unsigned long)data->user_page);
2195
2196 fail_user_page:
2197         kfree(data);
2198
2199 fail:
2200         return NULL;
2201 }
2202
2203 static void perf_mmap_free_page(unsigned long addr)
2204 {
2205         struct page *page = virt_to_page((void *)addr);
2206
2207         page->mapping = NULL;
2208         __free_page(page);
2209 }
2210
2211 static void perf_mmap_data_free(struct perf_mmap_data *data)
2212 {
2213         int i;
2214
2215         perf_mmap_free_page((unsigned long)data->user_page);
2216         for (i = 0; i < data->nr_pages; i++)
2217                 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2218         kfree(data);
2219 }
2220
2221 #else
2222
2223 /*
2224  * Back perf_mmap() with vmalloc memory.
2225  *
2226  * Required for architectures that have d-cache aliasing issues.
2227  */
2228
2229 static struct page *
2230 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2231 {
2232         if (pgoff > (1UL << data->data_order))
2233                 return NULL;
2234
2235         return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2236 }
2237
2238 static void perf_mmap_unmark_page(void *addr)
2239 {
2240         struct page *page = vmalloc_to_page(addr);
2241
2242         page->mapping = NULL;
2243 }
2244
2245 static void perf_mmap_data_free_work(struct work_struct *work)
2246 {
2247         struct perf_mmap_data *data;
2248         void *base;
2249         int i, nr;
2250
2251         data = container_of(work, struct perf_mmap_data, work);
2252         nr = 1 << data->data_order;
2253
2254         base = data->user_page;
2255         for (i = 0; i < nr + 1; i++)
2256                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2257
2258         vfree(base);
2259         kfree(data);
2260 }
2261
2262 static void perf_mmap_data_free(struct perf_mmap_data *data)
2263 {
2264         schedule_work(&data->work);
2265 }
2266
2267 static struct perf_mmap_data *
2268 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2269 {
2270         struct perf_mmap_data *data;
2271         unsigned long size;
2272         void *all_buf;
2273
2274         WARN_ON(atomic_read(&event->mmap_count));
2275
2276         size = sizeof(struct perf_mmap_data);
2277         size += sizeof(void *);
2278
2279         data = kzalloc(size, GFP_KERNEL);
2280         if (!data)
2281                 goto fail;
2282
2283         INIT_WORK(&data->work, perf_mmap_data_free_work);
2284
2285         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2286         if (!all_buf)
2287                 goto fail_all_buf;
2288
2289         data->user_page = all_buf;
2290         data->data_pages[0] = all_buf + PAGE_SIZE;
2291         data->data_order = ilog2(nr_pages);
2292         data->nr_pages = 1;
2293
2294         return data;
2295
2296 fail_all_buf:
2297         kfree(data);
2298
2299 fail:
2300         return NULL;
2301 }
2302
2303 #endif
2304
2305 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2306 {
2307         struct perf_event *event = vma->vm_file->private_data;
2308         struct perf_mmap_data *data;
2309         int ret = VM_FAULT_SIGBUS;
2310
2311         if (vmf->flags & FAULT_FLAG_MKWRITE) {
2312                 if (vmf->pgoff == 0)
2313                         ret = 0;
2314                 return ret;
2315         }
2316
2317         rcu_read_lock();
2318         data = rcu_dereference(event->data);
2319         if (!data)
2320                 goto unlock;
2321
2322         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2323                 goto unlock;
2324
2325         vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2326         if (!vmf->page)
2327                 goto unlock;
2328
2329         get_page(vmf->page);
2330         vmf->page->mapping = vma->vm_file->f_mapping;
2331         vmf->page->index   = vmf->pgoff;
2332
2333         ret = 0;
2334 unlock:
2335         rcu_read_unlock();
2336
2337         return ret;
2338 }
2339
2340 static void
2341 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2342 {
2343         long max_size = perf_data_size(data);
2344
2345         atomic_set(&data->lock, -1);
2346
2347         if (event->attr.watermark) {
2348                 data->watermark = min_t(long, max_size,
2349                                         event->attr.wakeup_watermark);
2350         }
2351
2352         if (!data->watermark)
2353                 data->watermark = max_size / 2;
2354
2355
2356         rcu_assign_pointer(event->data, data);
2357 }
2358
2359 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2360 {
2361         struct perf_mmap_data *data;
2362
2363         data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2364         perf_mmap_data_free(data);
2365 }
2366
2367 static void perf_mmap_data_release(struct perf_event *event)
2368 {
2369         struct perf_mmap_data *data = event->data;
2370
2371         WARN_ON(atomic_read(&event->mmap_count));
2372
2373         rcu_assign_pointer(event->data, NULL);
2374         call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2375 }
2376
2377 static void perf_mmap_open(struct vm_area_struct *vma)
2378 {
2379         struct perf_event *event = vma->vm_file->private_data;
2380
2381         atomic_inc(&event->mmap_count);
2382 }
2383
2384 static void perf_mmap_close(struct vm_area_struct *vma)
2385 {
2386         struct perf_event *event = vma->vm_file->private_data;
2387
2388         WARN_ON_ONCE(event->ctx->parent_ctx);
2389         if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2390                 unsigned long size = perf_data_size(event->data);
2391                 struct user_struct *user = current_user();
2392
2393                 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2394                 vma->vm_mm->locked_vm -= event->data->nr_locked;
2395                 perf_mmap_data_release(event);
2396                 mutex_unlock(&event->mmap_mutex);
2397         }
2398 }
2399
2400 static const struct vm_operations_struct perf_mmap_vmops = {
2401         .open           = perf_mmap_open,
2402         .close          = perf_mmap_close,
2403         .fault          = perf_mmap_fault,
2404         .page_mkwrite   = perf_mmap_fault,
2405 };
2406
2407 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2408 {
2409         struct perf_event *event = file->private_data;
2410         unsigned long user_locked, user_lock_limit;
2411         struct user_struct *user = current_user();
2412         unsigned long locked, lock_limit;
2413         struct perf_mmap_data *data;
2414         unsigned long vma_size;
2415         unsigned long nr_pages;
2416         long user_extra, extra;
2417         int ret = 0;
2418
2419         if (!(vma->vm_flags & VM_SHARED))
2420                 return -EINVAL;
2421
2422         vma_size = vma->vm_end - vma->vm_start;
2423         nr_pages = (vma_size / PAGE_SIZE) - 1;
2424
2425         /*
2426          * If we have data pages ensure they're a power-of-two number, so we
2427          * can do bitmasks instead of modulo.
2428          */
2429         if (nr_pages != 0 && !is_power_of_2(nr_pages))
2430                 return -EINVAL;
2431
2432         if (vma_size != PAGE_SIZE * (1 + nr_pages))
2433                 return -EINVAL;
2434
2435         if (vma->vm_pgoff != 0)
2436                 return -EINVAL;
2437
2438         WARN_ON_ONCE(event->ctx->parent_ctx);
2439         mutex_lock(&event->mmap_mutex);
2440         if (event->output) {
2441                 ret = -EINVAL;
2442                 goto unlock;
2443         }
2444
2445         if (atomic_inc_not_zero(&event->mmap_count)) {
2446                 if (nr_pages != event->data->nr_pages)
2447                         ret = -EINVAL;
2448                 goto unlock;
2449         }
2450
2451         user_extra = nr_pages + 1;
2452         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2453
2454         /*
2455          * Increase the limit linearly with more CPUs:
2456          */
2457         user_lock_limit *= num_online_cpus();
2458
2459         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2460
2461         extra = 0;
2462         if (user_locked > user_lock_limit)
2463                 extra = user_locked - user_lock_limit;
2464
2465         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2466         lock_limit >>= PAGE_SHIFT;
2467         locked = vma->vm_mm->locked_vm + extra;
2468
2469         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2470                 !capable(CAP_IPC_LOCK)) {
2471                 ret = -EPERM;
2472                 goto unlock;
2473         }
2474
2475         WARN_ON(event->data);
2476
2477         data = perf_mmap_data_alloc(event, nr_pages);
2478         ret = -ENOMEM;
2479         if (!data)
2480                 goto unlock;
2481
2482         ret = 0;
2483         perf_mmap_data_init(event, data);
2484
2485         atomic_set(&event->mmap_count, 1);
2486         atomic_long_add(user_extra, &user->locked_vm);
2487         vma->vm_mm->locked_vm += extra;
2488         event->data->nr_locked = extra;
2489         if (vma->vm_flags & VM_WRITE)
2490                 event->data->writable = 1;
2491
2492 unlock:
2493         mutex_unlock(&event->mmap_mutex);
2494
2495         vma->vm_flags |= VM_RESERVED;
2496         vma->vm_ops = &perf_mmap_vmops;
2497
2498         return ret;
2499 }
2500
2501 static int perf_fasync(int fd, struct file *filp, int on)
2502 {
2503         struct inode *inode = filp->f_path.dentry->d_inode;
2504         struct perf_event *event = filp->private_data;
2505         int retval;
2506
2507         mutex_lock(&inode->i_mutex);
2508         retval = fasync_helper(fd, filp, on, &event->fasync);
2509         mutex_unlock(&inode->i_mutex);
2510
2511         if (retval < 0)
2512                 return retval;
2513
2514         return 0;
2515 }
2516
2517 static const struct file_operations perf_fops = {
2518         .release                = perf_release,
2519         .read                   = perf_read,
2520         .poll                   = perf_poll,
2521         .unlocked_ioctl         = perf_ioctl,
2522         .compat_ioctl           = perf_ioctl,
2523         .mmap                   = perf_mmap,
2524         .fasync                 = perf_fasync,
2525 };
2526
2527 /*
2528  * Perf event wakeup
2529  *
2530  * If there's data, ensure we set the poll() state and publish everything
2531  * to user-space before waking everybody up.
2532  */
2533
2534 void perf_event_wakeup(struct perf_event *event)
2535 {
2536         wake_up_all(&event->waitq);
2537
2538         if (event->pending_kill) {
2539                 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2540                 event->pending_kill = 0;
2541         }
2542 }
2543
2544 /*
2545  * Pending wakeups
2546  *
2547  * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2548  *
2549  * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2550  * single linked list and use cmpxchg() to add entries lockless.
2551  */
2552
2553 static void perf_pending_event(struct perf_pending_entry *entry)
2554 {
2555         struct perf_event *event = container_of(entry,
2556                         struct perf_event, pending);
2557
2558         if (event->pending_disable) {
2559                 event->pending_disable = 0;
2560                 __perf_event_disable(event);
2561         }
2562
2563         if (event->pending_wakeup) {
2564                 event->pending_wakeup = 0;
2565                 perf_event_wakeup(event);
2566         }
2567 }
2568
2569 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2570
2571 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2572         PENDING_TAIL,
2573 };
2574
2575 static void perf_pending_queue(struct perf_pending_entry *entry,
2576                                void (*func)(struct perf_pending_entry *))
2577 {
2578         struct perf_pending_entry **head;
2579
2580         if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2581                 return;
2582
2583         entry->func = func;
2584
2585         head = &get_cpu_var(perf_pending_head);
2586
2587         do {
2588                 entry->next = *head;
2589         } while (cmpxchg(head, entry->next, entry) != entry->next);
2590
2591         set_perf_event_pending();
2592
2593         put_cpu_var(perf_pending_head);
2594 }
2595
2596 static int __perf_pending_run(void)
2597 {
2598         struct perf_pending_entry *list;
2599         int nr = 0;
2600
2601         list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2602         while (list != PENDING_TAIL) {
2603                 void (*func)(struct perf_pending_entry *);
2604                 struct perf_pending_entry *entry = list;
2605
2606                 list = list->next;
2607
2608                 func = entry->func;
2609                 entry->next = NULL;
2610                 /*
2611                  * Ensure we observe the unqueue before we issue the wakeup,
2612                  * so that we won't be waiting forever.
2613                  * -- see perf_not_pending().
2614                  */
2615                 smp_wmb();
2616
2617                 func(entry);
2618                 nr++;
2619         }
2620
2621         return nr;
2622 }
2623
2624 static inline int perf_not_pending(struct perf_event *event)
2625 {
2626         /*
2627          * If we flush on whatever cpu we run, there is a chance we don't
2628          * need to wait.
2629          */
2630         get_cpu();
2631         __perf_pending_run();
2632         put_cpu();
2633
2634         /*
2635          * Ensure we see the proper queue state before going to sleep
2636          * so that we do not miss the wakeup. -- see perf_pending_handle()
2637          */
2638         smp_rmb();
2639         return event->pending.next == NULL;
2640 }
2641
2642 static void perf_pending_sync(struct perf_event *event)
2643 {
2644         wait_event(event->waitq, perf_not_pending(event));
2645 }
2646
2647 void perf_event_do_pending(void)
2648 {
2649         __perf_pending_run();
2650 }
2651
2652 /*
2653  * Callchain support -- arch specific
2654  */
2655
2656 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2657 {
2658         return NULL;
2659 }
2660
2661 /*
2662  * Output
2663  */
2664 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2665                               unsigned long offset, unsigned long head)
2666 {
2667         unsigned long mask;
2668
2669         if (!data->writable)
2670                 return true;
2671
2672         mask = perf_data_size(data) - 1;
2673
2674         offset = (offset - tail) & mask;
2675         head   = (head   - tail) & mask;
2676
2677         if ((int)(head - offset) < 0)
2678                 return false;
2679
2680         return true;
2681 }
2682
2683 static void perf_output_wakeup(struct perf_output_handle *handle)
2684 {
2685         atomic_set(&handle->data->poll, POLL_IN);
2686
2687         if (handle->nmi) {
2688                 handle->event->pending_wakeup = 1;
2689                 perf_pending_queue(&handle->event->pending,
2690                                    perf_pending_event);
2691         } else
2692                 perf_event_wakeup(handle->event);
2693 }
2694
2695 /*
2696  * Curious locking construct.
2697  *
2698  * We need to ensure a later event_id doesn't publish a head when a former
2699  * event_id isn't done writing. However since we need to deal with NMIs we
2700  * cannot fully serialize things.
2701  *
2702  * What we do is serialize between CPUs so we only have to deal with NMI
2703  * nesting on a single CPU.
2704  *
2705  * We only publish the head (and generate a wakeup) when the outer-most
2706  * event_id completes.
2707  */
2708 static void perf_output_lock(struct perf_output_handle *handle)
2709 {
2710         struct perf_mmap_data *data = handle->data;
2711         int cur, cpu = get_cpu();
2712
2713         handle->locked = 0;
2714
2715         for (;;) {
2716                 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2717                 if (cur == -1) {
2718                         handle->locked = 1;
2719                         break;
2720                 }
2721                 if (cur == cpu)
2722                         break;
2723
2724                 cpu_relax();
2725         }
2726 }
2727
2728 static void perf_output_unlock(struct perf_output_handle *handle)
2729 {
2730         struct perf_mmap_data *data = handle->data;
2731         unsigned long head;
2732         int cpu;
2733
2734         data->done_head = data->head;
2735
2736         if (!handle->locked)
2737                 goto out;
2738
2739 again:
2740         /*
2741          * The xchg implies a full barrier that ensures all writes are done
2742          * before we publish the new head, matched by a rmb() in userspace when
2743          * reading this position.
2744          */
2745         while ((head = atomic_long_xchg(&data->done_head, 0)))
2746                 data->user_page->data_head = head;
2747
2748         /*
2749          * NMI can happen here, which means we can miss a done_head update.
2750          */
2751
2752         cpu = atomic_xchg(&data->lock, -1);
2753         WARN_ON_ONCE(cpu != smp_processor_id());
2754
2755         /*
2756          * Therefore we have to validate we did not indeed do so.
2757          */
2758         if (unlikely(atomic_long_read(&data->done_head))) {
2759                 /*
2760                  * Since we had it locked, we can lock it again.
2761                  */
2762                 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2763                         cpu_relax();
2764
2765                 goto again;
2766         }
2767
2768         if (atomic_xchg(&data->wakeup, 0))
2769                 perf_output_wakeup(handle);
2770 out:
2771         put_cpu();
2772 }
2773
2774 void perf_output_copy(struct perf_output_handle *handle,
2775                       const void *buf, unsigned int len)
2776 {
2777         unsigned int pages_mask;
2778         unsigned long offset;
2779         unsigned int size;
2780         void **pages;
2781
2782         offset          = handle->offset;
2783         pages_mask      = handle->data->nr_pages - 1;
2784         pages           = handle->data->data_pages;
2785
2786         do {
2787                 unsigned long page_offset;
2788                 unsigned long page_size;
2789                 int nr;
2790
2791                 nr          = (offset >> PAGE_SHIFT) & pages_mask;
2792                 page_size   = 1UL << (handle->data->data_order + PAGE_SHIFT);
2793                 page_offset = offset & (page_size - 1);
2794                 size        = min_t(unsigned int, page_size - page_offset, len);
2795
2796                 memcpy(pages[nr] + page_offset, buf, size);
2797
2798                 len         -= size;
2799                 buf         += size;
2800                 offset      += size;
2801         } while (len);
2802
2803         handle->offset = offset;
2804
2805         /*
2806          * Check we didn't copy past our reservation window, taking the
2807          * possible unsigned int wrap into account.
2808          */
2809         WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2810 }
2811
2812 int perf_output_begin(struct perf_output_handle *handle,
2813                       struct perf_event *event, unsigned int size,
2814                       int nmi, int sample)
2815 {
2816         struct perf_event *output_event;
2817         struct perf_mmap_data *data;
2818         unsigned long tail, offset, head;
2819         int have_lost;
2820         struct {
2821                 struct perf_event_header header;
2822                 u64                      id;
2823                 u64                      lost;
2824         } lost_event;
2825
2826         rcu_read_lock();
2827         /*
2828          * For inherited events we send all the output towards the parent.
2829          */
2830         if (event->parent)
2831                 event = event->parent;
2832
2833         output_event = rcu_dereference(event->output);
2834         if (output_event)
2835                 event = output_event;
2836
2837         data = rcu_dereference(event->data);
2838         if (!data)
2839                 goto out;
2840
2841         handle->data    = data;
2842         handle->event   = event;
2843         handle->nmi     = nmi;
2844         handle->sample  = sample;
2845
2846         if (!data->nr_pages)
2847                 goto fail;
2848
2849         have_lost = atomic_read(&data->lost);
2850         if (have_lost)
2851                 size += sizeof(lost_event);
2852
2853         perf_output_lock(handle);
2854
2855         do {
2856                 /*
2857                  * Userspace could choose to issue a mb() before updating the
2858                  * tail pointer. So that all reads will be completed before the
2859                  * write is issued.
2860                  */
2861                 tail = ACCESS_ONCE(data->user_page->data_tail);
2862                 smp_rmb();
2863                 offset = head = atomic_long_read(&data->head);
2864                 head += size;
2865                 if (unlikely(!perf_output_space(data, tail, offset, head)))
2866                         goto fail;
2867         } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2868
2869         handle->offset  = offset;
2870         handle->head    = head;
2871
2872         if (head - tail > data->watermark)
2873                 atomic_set(&data->wakeup, 1);
2874
2875         if (have_lost) {
2876                 lost_event.header.type = PERF_RECORD_LOST;
2877                 lost_event.header.misc = 0;
2878                 lost_event.header.size = sizeof(lost_event);
2879                 lost_event.id          = event->id;
2880                 lost_event.lost        = atomic_xchg(&data->lost, 0);
2881
2882                 perf_output_put(handle, lost_event);
2883         }
2884
2885         return 0;
2886
2887 fail:
2888         atomic_inc(&data->lost);
2889         perf_output_unlock(handle);
2890 out:
2891         rcu_read_unlock();
2892
2893         return -ENOSPC;
2894 }
2895
2896 void perf_output_end(struct perf_output_handle *handle)
2897 {
2898         struct perf_event *event = handle->event;
2899         struct perf_mmap_data *data = handle->data;
2900
2901         int wakeup_events = event->attr.wakeup_events;
2902
2903         if (handle->sample && wakeup_events) {
2904                 int events = atomic_inc_return(&data->events);
2905                 if (events >= wakeup_events) {
2906                         atomic_sub(wakeup_events, &data->events);
2907                         atomic_set(&data->wakeup, 1);
2908                 }
2909         }
2910
2911         perf_output_unlock(handle);
2912         rcu_read_unlock();
2913 }
2914
2915 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
2916 {
2917         /*
2918          * only top level events have the pid namespace they were created in
2919          */
2920         if (event->parent)
2921                 event = event->parent;
2922
2923         return task_tgid_nr_ns(p, event->ns);
2924 }
2925
2926 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
2927 {
2928         /*
2929          * only top level events have the pid namespace they were created in
2930          */
2931         if (event->parent)
2932                 event = event->parent;
2933
2934         return task_pid_nr_ns(p, event->ns);
2935 }
2936
2937 static void perf_output_read_one(struct perf_output_handle *handle,
2938                                  struct perf_event *event)
2939 {
2940         u64 read_format = event->attr.read_format;
2941         u64 values[4];
2942         int n = 0;
2943
2944         values[n++] = atomic64_read(&event->count);
2945         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2946                 values[n++] = event->total_time_enabled +
2947                         atomic64_read(&event->child_total_time_enabled);
2948         }
2949         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2950                 values[n++] = event->total_time_running +
2951                         atomic64_read(&event->child_total_time_running);
2952         }
2953         if (read_format & PERF_FORMAT_ID)
2954                 values[n++] = primary_event_id(event);
2955
2956         perf_output_copy(handle, values, n * sizeof(u64));
2957 }
2958
2959 /*
2960  * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
2961  */
2962 static void perf_output_read_group(struct perf_output_handle *handle,
2963                             struct perf_event *event)
2964 {
2965         struct perf_event *leader = event->group_leader, *sub;
2966         u64 read_format = event->attr.read_format;
2967         u64 values[5];
2968         int n = 0;
2969
2970         values[n++] = 1 + leader->nr_siblings;
2971
2972         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2973                 values[n++] = leader->total_time_enabled;
2974
2975         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2976                 values[n++] = leader->total_time_running;
2977
2978         if (leader != event)
2979                 leader->pmu->read(leader);
2980
2981         values[n++] = atomic64_read(&leader->count);
2982         if (read_format & PERF_FORMAT_ID)
2983                 values[n++] = primary_event_id(leader);
2984
2985         perf_output_copy(handle, values, n * sizeof(u64));
2986
2987         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2988                 n = 0;
2989
2990                 if (sub != event)
2991                         sub->pmu->read(sub);
2992
2993                 values[n++] = atomic64_read(&sub->count);
2994                 if (read_format & PERF_FORMAT_ID)
2995                         values[n++] = primary_event_id(sub);
2996
2997                 perf_output_copy(handle, values, n * sizeof(u64));
2998         }
2999 }
3000
3001 static void perf_output_read(struct perf_output_handle *handle,
3002                              struct perf_event *event)
3003 {
3004         if (event->attr.read_format & PERF_FORMAT_GROUP)
3005                 perf_output_read_group(handle, event);
3006         else
3007                 perf_output_read_one(handle, event);
3008 }
3009
3010 void perf_output_sample(struct perf_output_handle *handle,
3011                         struct perf_event_header *header,
3012                         struct perf_sample_data *data,
3013                         struct perf_event *event)
3014 {
3015         u64 sample_type = data->type;
3016
3017         perf_output_put(handle, *header);
3018
3019         if (sample_type & PERF_SAMPLE_IP)
3020                 perf_output_put(handle, data->ip);
3021
3022         if (sample_type & PERF_SAMPLE_TID)
3023                 perf_output_put(handle, data->tid_entry);
3024
3025         if (sample_type & PERF_SAMPLE_TIME)
3026                 perf_output_put(handle, data->time);
3027
3028         if (sample_type & PERF_SAMPLE_ADDR)
3029                 perf_output_put(handle, data->addr);
3030
3031         if (sample_type & PERF_SAMPLE_ID)
3032                 perf_output_put(handle, data->id);
3033
3034         if (sample_type & PERF_SAMPLE_STREAM_ID)
3035                 perf_output_put(handle, data->stream_id);
3036
3037         if (sample_type & PERF_SAMPLE_CPU)
3038                 perf_output_put(handle, data->cpu_entry);
3039
3040         if (sample_type & PERF_SAMPLE_PERIOD)
3041                 perf_output_put(handle, data->period);
3042
3043         if (sample_type & PERF_SAMPLE_READ)
3044                 perf_output_read(handle, event);
3045
3046         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3047                 if (data->callchain) {
3048                         int size = 1;
3049
3050                         if (data->callchain)
3051                                 size += data->callchain->nr;
3052
3053                         size *= sizeof(u64);
3054
3055                         perf_output_copy(handle, data->callchain, size);
3056                 } else {
3057                         u64 nr = 0;
3058                         perf_output_put(handle, nr);
3059                 }
3060         }
3061
3062         if (sample_type & PERF_SAMPLE_RAW) {
3063                 if (data->raw) {
3064                         perf_output_put(handle, data->raw->size);
3065                         perf_output_copy(handle, data->raw->data,
3066                                          data->raw->size);
3067                 } else {
3068                         struct {
3069                                 u32     size;
3070                                 u32     data;
3071                         } raw = {
3072                                 .size = sizeof(u32),
3073                                 .data = 0,
3074                         };
3075                         perf_output_put(handle, raw);
3076                 }
3077         }
3078 }
3079
3080 void perf_prepare_sample(struct perf_event_header *header,
3081                          struct perf_sample_data *data,
3082                          struct perf_event *event,
3083                          struct pt_regs *regs)
3084 {
3085         u64 sample_type = event->attr.sample_type;
3086
3087         data->type = sample_type;
3088
3089         header->type = PERF_RECORD_SAMPLE;
3090         header->size = sizeof(*header);
3091
3092         header->misc = 0;
3093         header->misc |= perf_misc_flags(regs);
3094
3095         if (sample_type & PERF_SAMPLE_IP) {
3096                 data->ip = perf_instruction_pointer(regs);
3097
3098                 header->size += sizeof(data->ip);
3099         }
3100
3101         if (sample_type & PERF_SAMPLE_TID) {
3102                 /* namespace issues */
3103                 data->tid_entry.pid = perf_event_pid(event, current);
3104                 data->tid_entry.tid = perf_event_tid(event, current);
3105
3106                 header->size += sizeof(data->tid_entry);
3107         }
3108
3109         if (sample_type & PERF_SAMPLE_TIME) {
3110                 data->time = perf_clock();
3111
3112                 header->size += sizeof(data->time);
3113         }
3114
3115         if (sample_type & PERF_SAMPLE_ADDR)
3116                 header->size += sizeof(data->addr);
3117
3118         if (sample_type & PERF_SAMPLE_ID) {
3119                 data->id = primary_event_id(event);
3120
3121                 header->size += sizeof(data->id);
3122         }
3123
3124         if (sample_type & PERF_SAMPLE_STREAM_ID) {
3125                 data->stream_id = event->id;
3126
3127                 header->size += sizeof(data->stream_id);
3128         }
3129
3130         if (sample_type & PERF_SAMPLE_CPU) {
3131                 data->cpu_entry.cpu             = raw_smp_processor_id();
3132                 data->cpu_entry.reserved        = 0;
3133
3134                 header->size += sizeof(data->cpu_entry);
3135         }
3136
3137         if (sample_type & PERF_SAMPLE_PERIOD)
3138                 header->size += sizeof(data->period);
3139
3140         if (sample_type & PERF_SAMPLE_READ)
3141                 header->size += perf_event_read_size(event);
3142
3143         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3144                 int size = 1;
3145
3146                 data->callchain = perf_callchain(regs);
3147
3148                 if (data->callchain)
3149                         size += data->callchain->nr;
3150
3151                 header->size += size * sizeof(u64);
3152         }
3153
3154         if (sample_type & PERF_SAMPLE_RAW) {
3155                 int size = sizeof(u32);
3156
3157                 if (data->raw)
3158                         size += data->raw->size;
3159                 else
3160                         size += sizeof(u32);
3161
3162                 WARN_ON_ONCE(size & (sizeof(u64)-1));
3163                 header->size += size;
3164         }
3165 }
3166
3167 static void perf_event_output(struct perf_event *event, int nmi,
3168                                 struct perf_sample_data *data,
3169                                 struct pt_regs *regs)
3170 {
3171         struct perf_output_handle handle;
3172         struct perf_event_header header;
3173
3174         perf_prepare_sample(&header, data, event, regs);
3175
3176         if (perf_output_begin(&handle, event, header.size, nmi, 1))
3177                 return;
3178
3179         perf_output_sample(&handle, &header, data, event);
3180
3181         perf_output_end(&handle);
3182 }
3183
3184 /*
3185  * read event_id
3186  */
3187
3188 struct perf_read_event {
3189         struct perf_event_header        header;
3190
3191         u32                             pid;
3192         u32                             tid;
3193 };
3194
3195 static void
3196 perf_event_read_event(struct perf_event *event,
3197                         struct task_struct *task)
3198 {
3199         struct perf_output_handle handle;
3200         struct perf_read_event read_event = {
3201                 .header = {
3202                         .type = PERF_RECORD_READ,
3203                         .misc = 0,
3204                         .size = sizeof(read_event) + perf_event_read_size(event),
3205                 },
3206                 .pid = perf_event_pid(event, task),
3207                 .tid = perf_event_tid(event, task),
3208         };
3209         int ret;
3210
3211         ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3212         if (ret)
3213                 return;
3214
3215         perf_output_put(&handle, read_event);
3216         perf_output_read(&handle, event);
3217
3218         perf_output_end(&handle);
3219 }
3220
3221 /*
3222  * task tracking -- fork/exit
3223  *
3224  * enabled by: attr.comm | attr.mmap | attr.task
3225  */
3226
3227 struct perf_task_event {
3228         struct task_struct              *task;
3229         struct perf_event_context       *task_ctx;
3230
3231         struct {
3232                 struct perf_event_header        header;
3233
3234                 u32                             pid;
3235                 u32                             ppid;
3236                 u32                             tid;
3237                 u32                             ptid;
3238                 u64                             time;
3239         } event_id;
3240 };
3241
3242 static void perf_event_task_output(struct perf_event *event,
3243                                      struct perf_task_event *task_event)
3244 {
3245         struct perf_output_handle handle;
3246         int size;
3247         struct task_struct *task = task_event->task;
3248         int ret;
3249
3250         size  = task_event->event_id.header.size;
3251         ret = perf_output_begin(&handle, event, size, 0, 0);
3252
3253         if (ret)
3254                 return;
3255
3256         task_event->event_id.pid = perf_event_pid(event, task);
3257         task_event->event_id.ppid = perf_event_pid(event, current);
3258
3259         task_event->event_id.tid = perf_event_tid(event, task);
3260         task_event->event_id.ptid = perf_event_tid(event, current);
3261
3262         task_event->event_id.time = perf_clock();
3263
3264         perf_output_put(&handle, task_event->event_id);
3265
3266         perf_output_end(&handle);
3267 }
3268
3269 static int perf_event_task_match(struct perf_event *event)
3270 {
3271         if (event->cpu != -1 && event->cpu != smp_processor_id())
3272                 return 0;
3273
3274         if (event->attr.comm || event->attr.mmap || event->attr.task)
3275                 return 1;
3276
3277         return 0;
3278 }
3279
3280 static void perf_event_task_ctx(struct perf_event_context *ctx,
3281                                   struct perf_task_event *task_event)
3282 {
3283         struct perf_event *event;
3284
3285         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3286                 if (perf_event_task_match(event))
3287                         perf_event_task_output(event, task_event);
3288         }
3289 }
3290
3291 static void perf_event_task_event(struct perf_task_event *task_event)
3292 {
3293         struct perf_cpu_context *cpuctx;
3294         struct perf_event_context *ctx = task_event->task_ctx;
3295
3296         rcu_read_lock();
3297         cpuctx = &get_cpu_var(perf_cpu_context);
3298         perf_event_task_ctx(&cpuctx->ctx, task_event);
3299         if (!ctx)
3300                 ctx = rcu_dereference(task_event->task->perf_event_ctxp);
3301         if (ctx)
3302                 perf_event_task_ctx(ctx, task_event);
3303         put_cpu_var(perf_cpu_context);
3304         rcu_read_unlock();
3305 }
3306
3307 static void perf_event_task(struct task_struct *task,
3308                               struct perf_event_context *task_ctx,
3309                               int new)
3310 {
3311         struct perf_task_event task_event;
3312
3313         if (!atomic_read(&nr_comm_events) &&
3314             !atomic_read(&nr_mmap_events) &&
3315             !atomic_read(&nr_task_events))
3316                 return;
3317
3318         task_event = (struct perf_task_event){
3319                 .task     = task,
3320                 .task_ctx = task_ctx,
3321                 .event_id    = {
3322                         .header = {
3323                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3324                                 .misc = 0,
3325                                 .size = sizeof(task_event.event_id),
3326                         },
3327                         /* .pid  */
3328                         /* .ppid */
3329                         /* .tid  */
3330                         /* .ptid */
3331                 },
3332         };
3333
3334         perf_event_task_event(&task_event);
3335 }
3336
3337 void perf_event_fork(struct task_struct *task)
3338 {
3339         perf_event_task(task, NULL, 1);
3340 }
3341
3342 /*
3343  * comm tracking
3344  */
3345
3346 struct perf_comm_event {
3347         struct task_struct      *task;
3348         char                    *comm;
3349         int                     comm_size;
3350
3351         struct {
3352                 struct perf_event_header        header;
3353
3354                 u32                             pid;
3355                 u32                             tid;
3356         } event_id;
3357 };
3358
3359 static void perf_event_comm_output(struct perf_event *event,
3360                                      struct perf_comm_event *comm_event)
3361 {
3362         struct perf_output_handle handle;
3363         int size = comm_event->event_id.header.size;
3364         int ret = perf_output_begin(&handle, event, size, 0, 0);
3365
3366         if (ret)
3367                 return;
3368
3369         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3370         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3371
3372         perf_output_put(&handle, comm_event->event_id);
3373         perf_output_copy(&handle, comm_event->comm,
3374                                    comm_event->comm_size);
3375         perf_output_end(&handle);
3376 }
3377
3378 static int perf_event_comm_match(struct perf_event *event)
3379 {
3380         if (event->cpu != -1 && event->cpu != smp_processor_id())
3381                 return 0;
3382
3383         if (event->attr.comm)
3384                 return 1;
3385
3386         return 0;
3387 }
3388
3389 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3390                                   struct perf_comm_event *comm_event)
3391 {
3392         struct perf_event *event;
3393
3394         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3395                 if (perf_event_comm_match(event))
3396                         perf_event_comm_output(event, comm_event);
3397         }
3398 }
3399
3400 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3401 {
3402         struct perf_cpu_context *cpuctx;
3403         struct perf_event_context *ctx;
3404         unsigned int size;
3405         char comm[TASK_COMM_LEN];
3406
3407         memset(comm, 0, sizeof(comm));
3408         strlcpy(comm, comm_event->task->comm, sizeof(comm));
3409         size = ALIGN(strlen(comm)+1, sizeof(u64));
3410
3411         comm_event->comm = comm;
3412         comm_event->comm_size = size;
3413
3414         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3415
3416         rcu_read_lock();
3417         cpuctx = &get_cpu_var(perf_cpu_context);
3418         perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3419         ctx = rcu_dereference(current->perf_event_ctxp);
3420         if (ctx)
3421                 perf_event_comm_ctx(ctx, comm_event);
3422         put_cpu_var(perf_cpu_context);
3423         rcu_read_unlock();
3424 }
3425
3426 void perf_event_comm(struct task_struct *task)
3427 {
3428         struct perf_comm_event comm_event;
3429
3430         if (task->perf_event_ctxp)
3431                 perf_event_enable_on_exec(task);
3432
3433         if (!atomic_read(&nr_comm_events))
3434                 return;
3435
3436         comm_event = (struct perf_comm_event){
3437                 .task   = task,
3438                 /* .comm      */
3439                 /* .comm_size */
3440                 .event_id  = {
3441                         .header = {
3442                                 .type = PERF_RECORD_COMM,
3443                                 .misc = 0,
3444                                 /* .size */
3445                         },
3446                         /* .pid */
3447                         /* .tid */
3448                 },
3449         };
3450
3451         perf_event_comm_event(&comm_event);
3452 }
3453
3454 /*
3455  * mmap tracking
3456  */
3457
3458 struct perf_mmap_event {
3459         struct vm_area_struct   *vma;
3460
3461         const char              *file_name;
3462         int                     file_size;
3463
3464         struct {
3465                 struct perf_event_header        header;
3466
3467                 u32                             pid;
3468                 u32                             tid;
3469                 u64                             start;
3470                 u64                             len;
3471                 u64                             pgoff;
3472         } event_id;
3473 };
3474
3475 static void perf_event_mmap_output(struct perf_event *event,
3476                                      struct perf_mmap_event *mmap_event)
3477 {
3478         struct perf_output_handle handle;
3479         int size = mmap_event->event_id.header.size;
3480         int ret = perf_output_begin(&handle, event, size, 0, 0);
3481
3482         if (ret)
3483                 return;
3484
3485         mmap_event->event_id.pid = perf_event_pid(event, current);
3486         mmap_event->event_id.tid = perf_event_tid(event, current);
3487
3488         perf_output_put(&handle, mmap_event->event_id);
3489         perf_output_copy(&handle, mmap_event->file_name,
3490                                    mmap_event->file_size);
3491         perf_output_end(&handle);
3492 }
3493
3494 static int perf_event_mmap_match(struct perf_event *event,
3495                                    struct perf_mmap_event *mmap_event)
3496 {
3497         if (event->cpu != -1 && event->cpu != smp_processor_id())
3498                 return 0;
3499
3500         if (event->attr.mmap)
3501                 return 1;
3502
3503         return 0;
3504 }
3505
3506 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3507                                   struct perf_mmap_event *mmap_event)
3508 {
3509         struct perf_event *event;
3510
3511         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3512                 if (perf_event_mmap_match(event, mmap_event))
3513                         perf_event_mmap_output(event, mmap_event);
3514         }
3515 }
3516
3517 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3518 {
3519         struct perf_cpu_context *cpuctx;
3520         struct perf_event_context *ctx;
3521         struct vm_area_struct *vma = mmap_event->vma;
3522         struct file *file = vma->vm_file;
3523         unsigned int size;
3524         char tmp[16];
3525         char *buf = NULL;
3526         const char *name;
3527
3528         memset(tmp, 0, sizeof(tmp));
3529
3530         if (file) {
3531                 /*
3532                  * d_path works from the end of the buffer backwards, so we
3533                  * need to add enough zero bytes after the string to handle
3534                  * the 64bit alignment we do later.
3535                  */
3536                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3537                 if (!buf) {
3538                         name = strncpy(tmp, "//enomem", sizeof(tmp));
3539                         goto got_name;
3540                 }
3541                 name = d_path(&file->f_path, buf, PATH_MAX);
3542                 if (IS_ERR(name)) {
3543                         name = strncpy(tmp, "//toolong", sizeof(tmp));
3544                         goto got_name;
3545                 }
3546         } else {
3547                 if (arch_vma_name(mmap_event->vma)) {
3548                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3549                                        sizeof(tmp));
3550                         goto got_name;
3551                 }
3552
3553                 if (!vma->vm_mm) {
3554                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
3555                         goto got_name;
3556                 }
3557
3558                 name = strncpy(tmp, "//anon", sizeof(tmp));
3559                 goto got_name;
3560         }
3561
3562 got_name:
3563         size = ALIGN(strlen(name)+1, sizeof(u64));
3564
3565         mmap_event->file_name = name;
3566         mmap_event->file_size = size;
3567
3568         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3569
3570         rcu_read_lock();
3571         cpuctx = &get_cpu_var(perf_cpu_context);
3572         perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3573         ctx = rcu_dereference(current->perf_event_ctxp);
3574         if (ctx)
3575                 perf_event_mmap_ctx(ctx, mmap_event);
3576         put_cpu_var(perf_cpu_context);
3577         rcu_read_unlock();
3578
3579         kfree(buf);
3580 }
3581
3582 void __perf_event_mmap(struct vm_area_struct *vma)
3583 {
3584         struct perf_mmap_event mmap_event;
3585
3586         if (!atomic_read(&nr_mmap_events))
3587                 return;
3588
3589         mmap_event = (struct perf_mmap_event){
3590                 .vma    = vma,
3591                 /* .file_name */
3592                 /* .file_size */
3593                 .event_id  = {
3594                         .header = {
3595                                 .type = PERF_RECORD_MMAP,
3596                                 .misc = 0,
3597                                 /* .size */
3598                         },
3599                         /* .pid */
3600                         /* .tid */
3601                         .start  = vma->vm_start,
3602                         .len    = vma->vm_end - vma->vm_start,
3603                         .pgoff  = vma->vm_pgoff,
3604                 },
3605         };
3606
3607         perf_event_mmap_event(&mmap_event);
3608 }
3609
3610 /*
3611  * IRQ throttle logging
3612  */
3613
3614 static void perf_log_throttle(struct perf_event *event, int enable)
3615 {
3616         struct perf_output_handle handle;
3617         int ret;
3618
3619         struct {
3620                 struct perf_event_header        header;
3621                 u64                             time;
3622                 u64                             id;
3623                 u64                             stream_id;
3624         } throttle_event = {
3625                 .header = {
3626                         .type = PERF_RECORD_THROTTLE,
3627                         .misc = 0,
3628                         .size = sizeof(throttle_event),
3629                 },
3630                 .time           = perf_clock(),
3631                 .id             = primary_event_id(event),
3632                 .stream_id      = event->id,
3633         };
3634
3635         if (enable)
3636                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3637
3638         ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3639         if (ret)
3640                 return;
3641
3642         perf_output_put(&handle, throttle_event);
3643         perf_output_end(&handle);
3644 }
3645
3646 /*
3647  * Generic event overflow handling, sampling.
3648  */
3649
3650 static int __perf_event_overflow(struct perf_event *event, int nmi,
3651                                    int throttle, struct perf_sample_data *data,
3652                                    struct pt_regs *regs)
3653 {
3654         int events = atomic_read(&event->event_limit);
3655         struct hw_perf_event *hwc = &event->hw;
3656         int ret = 0;
3657
3658         throttle = (throttle && event->pmu->unthrottle != NULL);
3659
3660         if (!throttle) {
3661                 hwc->interrupts++;
3662         } else {
3663                 if (hwc->interrupts != MAX_INTERRUPTS) {
3664                         hwc->interrupts++;
3665                         if (HZ * hwc->interrupts >
3666                                         (u64)sysctl_perf_event_sample_rate) {
3667                                 hwc->interrupts = MAX_INTERRUPTS;
3668                                 perf_log_throttle(event, 0);
3669                                 ret = 1;
3670                         }
3671                 } else {
3672                         /*
3673                          * Keep re-disabling events even though on the previous
3674                          * pass we disabled it - just in case we raced with a
3675                          * sched-in and the event got enabled again:
3676                          */
3677                         ret = 1;
3678                 }
3679         }
3680
3681         if (event->attr.freq) {
3682                 u64 now = perf_clock();
3683                 s64 delta = now - hwc->freq_stamp;
3684
3685                 hwc->freq_stamp = now;
3686
3687                 if (delta > 0 && delta < TICK_NSEC)
3688                         perf_adjust_period(event, NSEC_PER_SEC / (int)delta);
3689         }
3690
3691         /*
3692          * XXX event_limit might not quite work as expected on inherited
3693          * events
3694          */
3695
3696         event->pending_kill = POLL_IN;
3697         if (events && atomic_dec_and_test(&event->event_limit)) {
3698                 ret = 1;
3699                 event->pending_kill = POLL_HUP;
3700                 if (nmi) {
3701                         event->pending_disable = 1;
3702                         perf_pending_queue(&event->pending,
3703                                            perf_pending_event);
3704                 } else
3705                         perf_event_disable(event);
3706         }
3707
3708         if (event->overflow_handler)
3709                 event->overflow_handler(event, nmi, data, regs);
3710         else
3711                 perf_event_output(event, nmi, data, regs);
3712
3713         return ret;
3714 }
3715
3716 int perf_event_overflow(struct perf_event *event, int nmi,
3717                           struct perf_sample_data *data,
3718                           struct pt_regs *regs)
3719 {
3720         return __perf_event_overflow(event, nmi, 1, data, regs);
3721 }
3722
3723 /*
3724  * Generic software event infrastructure
3725  */
3726
3727 /*
3728  * We directly increment event->count and keep a second value in
3729  * event->hw.period_left to count intervals. This period event
3730  * is kept in the range [-sample_period, 0] so that we can use the
3731  * sign as trigger.
3732  */
3733
3734 static u64 perf_swevent_set_period(struct perf_event *event)
3735 {
3736         struct hw_perf_event *hwc = &event->hw;
3737         u64 period = hwc->last_period;
3738         u64 nr, offset;
3739         s64 old, val;
3740
3741         hwc->last_period = hwc->sample_period;
3742
3743 again:
3744         old = val = atomic64_read(&hwc->period_left);
3745         if (val < 0)
3746                 return 0;
3747
3748         nr = div64_u64(period + val, period);
3749         offset = nr * period;
3750         val -= offset;
3751         if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3752                 goto again;
3753
3754         return nr;
3755 }
3756
3757 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3758                                     int nmi, struct perf_sample_data *data,
3759                                     struct pt_regs *regs)
3760 {
3761         struct hw_perf_event *hwc = &event->hw;
3762         int throttle = 0;
3763
3764         data->period = event->hw.last_period;
3765         if (!overflow)
3766                 overflow = perf_swevent_set_period(event);
3767
3768         if (hwc->interrupts == MAX_INTERRUPTS)
3769                 return;
3770
3771         for (; overflow; overflow--) {
3772                 if (__perf_event_overflow(event, nmi, throttle,
3773                                             data, regs)) {
3774                         /*
3775                          * We inhibit the overflow from happening when
3776                          * hwc->interrupts == MAX_INTERRUPTS.
3777                          */
3778                         break;
3779                 }
3780                 throttle = 1;
3781         }
3782 }
3783
3784 static void perf_swevent_unthrottle(struct perf_event *event)
3785 {
3786         /*
3787          * Nothing to do, we already reset hwc->interrupts.
3788          */
3789 }
3790
3791 static void perf_swevent_add(struct perf_event *event, u64 nr,
3792                                int nmi, struct perf_sample_data *data,
3793                                struct pt_regs *regs)
3794 {
3795         struct hw_perf_event *hwc = &event->hw;
3796
3797         atomic64_add(nr, &event->count);
3798
3799         if (!regs)
3800                 return;
3801
3802         if (!hwc->sample_period)
3803                 return;
3804
3805         if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3806                 return perf_swevent_overflow(event, 1, nmi, data, regs);
3807
3808         if (atomic64_add_negative(nr, &hwc->period_left))
3809                 return;
3810
3811         perf_swevent_overflow(event, 0, nmi, data, regs);
3812 }
3813
3814 static int perf_swevent_is_counting(struct perf_event *event)
3815 {
3816         /*
3817          * The event is active, we're good!
3818          */
3819         if (event->state == PERF_EVENT_STATE_ACTIVE)
3820                 return 1;
3821
3822         /*
3823          * The event is off/error, not counting.
3824          */
3825         if (event->state != PERF_EVENT_STATE_INACTIVE)
3826                 return 0;
3827
3828         /*
3829          * The event is inactive, if the context is active
3830          * we're part of a group that didn't make it on the 'pmu',
3831          * not counting.
3832          */
3833         if (event->ctx->is_active)
3834                 return 0;
3835
3836         /*
3837          * We're inactive and the context is too, this means the
3838          * task is scheduled out, we're counting events that happen
3839          * to us, like migration events.
3840          */
3841         return 1;
3842 }
3843
3844 static int perf_tp_event_match(struct perf_event *event,
3845                                 struct perf_sample_data *data);
3846
3847 static int perf_exclude_event(struct perf_event *event,
3848                               struct pt_regs *regs)
3849 {
3850         if (regs) {
3851                 if (event->attr.exclude_user && user_mode(regs))
3852                         return 1;
3853
3854                 if (event->attr.exclude_kernel && !user_mode(regs))
3855                         return 1;
3856         }
3857
3858         return 0;
3859 }
3860
3861 static int perf_swevent_match(struct perf_event *event,
3862                                 enum perf_type_id type,
3863                                 u32 event_id,
3864                                 struct perf_sample_data *data,
3865                                 struct pt_regs *regs)
3866 {
3867         if (event->cpu != -1 && event->cpu != smp_processor_id())
3868                 return 0;
3869
3870         if (!perf_swevent_is_counting(event))
3871                 return 0;
3872
3873         if (event->attr.type != type)
3874                 return 0;
3875
3876         if (event->attr.config != event_id)
3877                 return 0;
3878
3879         if (perf_exclude_event(event, regs))
3880                 return 0;
3881
3882         if (event->attr.type == PERF_TYPE_TRACEPOINT &&
3883             !perf_tp_event_match(event, data))
3884                 return 0;
3885
3886         return 1;
3887 }
3888
3889 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
3890                                      enum perf_type_id type,
3891                                      u32 event_id, u64 nr, int nmi,
3892                                      struct perf_sample_data *data,
3893                                      struct pt_regs *regs)
3894 {
3895         struct perf_event *event;
3896
3897         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3898                 if (perf_swevent_match(event, type, event_id, data, regs))
3899                         perf_swevent_add(event, nr, nmi, data, regs);
3900         }
3901 }
3902
3903 int perf_swevent_get_recursion_context(void)
3904 {
3905         struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3906         int rctx;
3907
3908         if (in_nmi())
3909                 rctx = 3;
3910         else if (in_irq())
3911                 rctx = 2;
3912         else if (in_softirq())
3913                 rctx = 1;
3914         else
3915                 rctx = 0;
3916
3917         if (cpuctx->recursion[rctx]) {
3918                 put_cpu_var(perf_cpu_context);
3919                 return -1;
3920         }
3921
3922         cpuctx->recursion[rctx]++;
3923         barrier();
3924
3925         return rctx;
3926 }
3927 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
3928
3929 void perf_swevent_put_recursion_context(int rctx)
3930 {
3931         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
3932         barrier();
3933         cpuctx->recursion[rctx]--;
3934         put_cpu_var(perf_cpu_context);
3935 }
3936 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
3937
3938 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
3939                                     u64 nr, int nmi,
3940                                     struct perf_sample_data *data,
3941                                     struct pt_regs *regs)
3942 {
3943         struct perf_cpu_context *cpuctx;
3944         struct perf_event_context *ctx;
3945
3946         cpuctx = &__get_cpu_var(perf_cpu_context);
3947         rcu_read_lock();
3948         perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
3949                                  nr, nmi, data, regs);
3950         /*
3951          * doesn't really matter which of the child contexts the
3952          * events ends up in.
3953          */
3954         ctx = rcu_dereference(current->perf_event_ctxp);
3955         if (ctx)
3956                 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
3957         rcu_read_unlock();
3958 }
3959
3960 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
3961                             struct pt_regs *regs, u64 addr)
3962 {
3963         struct perf_sample_data data;
3964         int rctx;
3965
3966         rctx = perf_swevent_get_recursion_context();
3967         if (rctx < 0)
3968                 return;
3969
3970         data.addr = addr;
3971         data.raw  = NULL;
3972
3973         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
3974
3975         perf_swevent_put_recursion_context(rctx);
3976 }
3977
3978 static void perf_swevent_read(struct perf_event *event)
3979 {
3980 }
3981
3982 static int perf_swevent_enable(struct perf_event *event)
3983 {
3984         struct hw_perf_event *hwc = &event->hw;
3985
3986         if (hwc->sample_period) {
3987                 hwc->last_period = hwc->sample_period;
3988                 perf_swevent_set_period(event);
3989         }
3990         return 0;
3991 }
3992
3993 static void perf_swevent_disable(struct perf_event *event)
3994 {
3995 }
3996
3997 static const struct pmu perf_ops_generic = {
3998         .enable         = perf_swevent_enable,
3999         .disable        = perf_swevent_disable,
4000         .read           = perf_swevent_read,
4001         .unthrottle     = perf_swevent_unthrottle,
4002 };
4003
4004 /*
4005  * hrtimer based swevent callback
4006  */
4007
4008 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4009 {
4010         enum hrtimer_restart ret = HRTIMER_RESTART;
4011         struct perf_sample_data data;
4012         struct pt_regs *regs;
4013         struct perf_event *event;
4014         u64 period;
4015
4016         event   = container_of(hrtimer, struct perf_event, hw.hrtimer);
4017         event->pmu->read(event);
4018
4019         data.addr = 0;
4020         data.raw = NULL;
4021         data.period = event->hw.last_period;
4022         regs = get_irq_regs();
4023         /*
4024          * In case we exclude kernel IPs or are somehow not in interrupt
4025          * context, provide the next best thing, the user IP.
4026          */
4027         if ((event->attr.exclude_kernel || !regs) &&
4028                         !event->attr.exclude_user)
4029                 regs = task_pt_regs(current);
4030
4031         if (regs) {
4032                 if (!(event->attr.exclude_idle && current->pid == 0))
4033                         if (perf_event_overflow(event, 0, &data, regs))
4034                                 ret = HRTIMER_NORESTART;
4035         }
4036
4037         period = max_t(u64, 10000, event->hw.sample_period);
4038         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4039
4040         return ret;
4041 }
4042
4043 static void perf_swevent_start_hrtimer(struct perf_event *event)
4044 {
4045         struct hw_perf_event *hwc = &event->hw;
4046
4047         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4048         hwc->hrtimer.function = perf_swevent_hrtimer;
4049         if (hwc->sample_period) {
4050                 u64 period;
4051
4052                 if (hwc->remaining) {
4053                         if (hwc->remaining < 0)
4054                                 period = 10000;
4055                         else
4056                                 period = hwc->remaining;
4057                         hwc->remaining = 0;
4058                 } else {
4059                         period = max_t(u64, 10000, hwc->sample_period);
4060                 }
4061                 __hrtimer_start_range_ns(&hwc->hrtimer,
4062                                 ns_to_ktime(period), 0,
4063                                 HRTIMER_MODE_REL, 0);
4064         }
4065 }
4066
4067 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4068 {
4069         struct hw_perf_event *hwc = &event->hw;
4070
4071         if (hwc->sample_period) {
4072                 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4073                 hwc->remaining = ktime_to_ns(remaining);
4074
4075                 hrtimer_cancel(&hwc->hrtimer);
4076         }
4077 }
4078
4079 /*
4080  * Software event: cpu wall time clock
4081  */
4082
4083 static void cpu_clock_perf_event_update(struct perf_event *event)
4084 {
4085         int cpu = raw_smp_processor_id();
4086         s64 prev;
4087         u64 now;
4088
4089         now = cpu_clock(cpu);
4090         prev = atomic64_xchg(&event->hw.prev_count, now);
4091         atomic64_add(now - prev, &event->count);
4092 }
4093
4094 static int cpu_clock_perf_event_enable(struct perf_event *event)
4095 {
4096         struct hw_perf_event *hwc = &event->hw;
4097         int cpu = raw_smp_processor_id();
4098
4099         atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4100         perf_swevent_start_hrtimer(event);
4101
4102         return 0;
4103 }
4104
4105 static void cpu_clock_perf_event_disable(struct perf_event *event)
4106 {
4107         perf_swevent_cancel_hrtimer(event);
4108         cpu_clock_perf_event_update(event);
4109 }
4110
4111 static void cpu_clock_perf_event_read(struct perf_event *event)
4112 {
4113         cpu_clock_perf_event_update(event);
4114 }
4115
4116 static const struct pmu perf_ops_cpu_clock = {
4117         .enable         = cpu_clock_perf_event_enable,
4118         .disable        = cpu_clock_perf_event_disable,
4119         .read           = cpu_clock_perf_event_read,
4120 };
4121
4122 /*
4123  * Software event: task time clock
4124  */
4125
4126 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4127 {
4128         u64 prev;
4129         s64 delta;
4130
4131         prev = atomic64_xchg(&event->hw.prev_count, now);
4132         delta = now - prev;
4133         atomic64_add(delta, &event->count);
4134 }
4135
4136 static int task_clock_perf_event_enable(struct perf_event *event)
4137 {
4138         struct hw_perf_event *hwc = &event->hw;
4139         u64 now;
4140
4141         now = event->ctx->time;
4142
4143         atomic64_set(&hwc->prev_count, now);
4144
4145         perf_swevent_start_hrtimer(event);
4146
4147         return 0;
4148 }
4149
4150 static void task_clock_perf_event_disable(struct perf_event *event)
4151 {
4152         perf_swevent_cancel_hrtimer(event);
4153         task_clock_perf_event_update(event, event->ctx->time);
4154
4155 }
4156
4157 static void task_clock_perf_event_read(struct perf_event *event)
4158 {
4159         u64 time;
4160
4161         if (!in_nmi()) {
4162                 update_context_time(event->ctx);
4163                 time = event->ctx->time;
4164         } else {
4165                 u64 now = perf_clock();
4166                 u64 delta = now - event->ctx->timestamp;
4167                 time = event->ctx->time + delta;
4168         }
4169
4170         task_clock_perf_event_update(event, time);
4171 }
4172
4173 static const struct pmu perf_ops_task_clock = {
4174         .enable         = task_clock_perf_event_enable,
4175         .disable        = task_clock_perf_event_disable,
4176         .read           = task_clock_perf_event_read,
4177 };
4178
4179 #ifdef CONFIG_EVENT_PROFILE
4180
4181 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4182                           int entry_size)
4183 {
4184         struct perf_raw_record raw = {
4185                 .size = entry_size,
4186                 .data = record,
4187         };
4188
4189         struct perf_sample_data data = {
4190                 .addr = addr,
4191                 .raw = &raw,
4192         };
4193
4194         struct pt_regs *regs = get_irq_regs();
4195
4196         if (!regs)
4197                 regs = task_pt_regs(current);
4198
4199         /* Trace events already protected against recursion */
4200         do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4201                                 &data, regs);
4202 }
4203 EXPORT_SYMBOL_GPL(perf_tp_event);
4204
4205 static int perf_tp_event_match(struct perf_event *event,
4206                                 struct perf_sample_data *data)
4207 {
4208         void *record = data->raw->data;
4209
4210         if (likely(!event->filter) || filter_match_preds(event->filter, record))
4211                 return 1;
4212         return 0;
4213 }
4214
4215 static void tp_perf_event_destroy(struct perf_event *event)
4216 {
4217         ftrace_profile_disable(event->attr.config);
4218 }
4219
4220 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4221 {
4222         /*
4223          * Raw tracepoint data is a severe data leak, only allow root to
4224          * have these.
4225          */
4226         if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4227                         perf_paranoid_tracepoint_raw() &&
4228                         !capable(CAP_SYS_ADMIN))
4229                 return ERR_PTR(-EPERM);
4230
4231         if (ftrace_profile_enable(event->attr.config))
4232                 return NULL;
4233
4234         event->destroy = tp_perf_event_destroy;
4235
4236         return &perf_ops_generic;
4237 }
4238
4239 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4240 {
4241         char *filter_str;
4242         int ret;
4243
4244         if (event->attr.type != PERF_TYPE_TRACEPOINT)
4245                 return -EINVAL;
4246
4247         filter_str = strndup_user(arg, PAGE_SIZE);
4248         if (IS_ERR(filter_str))
4249                 return PTR_ERR(filter_str);
4250
4251         ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4252
4253         kfree(filter_str);
4254         return ret;
4255 }
4256
4257 static void perf_event_free_filter(struct perf_event *event)
4258 {
4259         ftrace_profile_free_filter(event);
4260 }
4261
4262 #else
4263
4264 static int perf_tp_event_match(struct perf_event *event,
4265                                 struct perf_sample_data *data)
4266 {
4267         return 1;
4268 }
4269
4270 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4271 {
4272         return NULL;
4273 }
4274
4275 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4276 {
4277         return -ENOENT;
4278 }
4279
4280 static void perf_event_free_filter(struct perf_event *event)
4281 {
4282 }
4283
4284 #endif /* CONFIG_EVENT_PROFILE */
4285
4286 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4287 static void bp_perf_event_destroy(struct perf_event *event)
4288 {
4289         release_bp_slot(event);
4290 }
4291
4292 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4293 {
4294         int err;
4295
4296         err = register_perf_hw_breakpoint(bp);
4297         if (err)
4298                 return ERR_PTR(err);
4299
4300         bp->destroy = bp_perf_event_destroy;
4301
4302         return &perf_ops_bp;
4303 }
4304
4305 void perf_bp_event(struct perf_event *bp, void *data)
4306 {
4307         struct perf_sample_data sample;
4308         struct pt_regs *regs = data;
4309
4310         sample.raw = NULL;
4311         sample.addr = bp->attr.bp_addr;
4312
4313         if (!perf_exclude_event(bp, regs))
4314                 perf_swevent_add(bp, 1, 1, &sample, regs);
4315 }
4316 #else
4317 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4318 {
4319         return NULL;
4320 }
4321
4322 void perf_bp_event(struct perf_event *bp, void *regs)
4323 {
4324 }
4325 #endif
4326
4327 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4328
4329 static void sw_perf_event_destroy(struct perf_event *event)
4330 {
4331         u64 event_id = event->attr.config;
4332
4333         WARN_ON(event->parent);
4334
4335         atomic_dec(&perf_swevent_enabled[event_id]);
4336 }
4337
4338 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4339 {
4340         const struct pmu *pmu = NULL;
4341         u64 event_id = event->attr.config;
4342
4343         /*
4344          * Software events (currently) can't in general distinguish
4345          * between user, kernel and hypervisor events.
4346          * However, context switches and cpu migrations are considered
4347          * to be kernel events, and page faults are never hypervisor
4348          * events.
4349          */
4350         switch (event_id) {
4351         case PERF_COUNT_SW_CPU_CLOCK:
4352                 pmu = &perf_ops_cpu_clock;
4353
4354                 break;
4355         case PERF_COUNT_SW_TASK_CLOCK:
4356                 /*
4357                  * If the user instantiates this as a per-cpu event,
4358                  * use the cpu_clock event instead.
4359                  */
4360                 if (event->ctx->task)
4361                         pmu = &perf_ops_task_clock;
4362                 else
4363                         pmu = &perf_ops_cpu_clock;
4364
4365                 break;
4366         case PERF_COUNT_SW_PAGE_FAULTS:
4367         case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4368         case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4369         case PERF_COUNT_SW_CONTEXT_SWITCHES:
4370         case PERF_COUNT_SW_CPU_MIGRATIONS:
4371         case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4372         case PERF_COUNT_SW_EMULATION_FAULTS:
4373                 if (!event->parent) {
4374                         atomic_inc(&perf_swevent_enabled[event_id]);
4375                         event->destroy = sw_perf_event_destroy;
4376                 }
4377                 pmu = &perf_ops_generic;
4378                 break;
4379         }
4380
4381         return pmu;
4382 }
4383
4384 /*
4385  * Allocate and initialize a event structure
4386  */
4387 static struct perf_event *
4388 perf_event_alloc(struct perf_event_attr *attr,
4389                    int cpu,
4390                    struct perf_event_context *ctx,
4391                    struct perf_event *group_leader,
4392                    struct perf_event *parent_event,
4393                    perf_overflow_handler_t overflow_handler,
4394                    gfp_t gfpflags)
4395 {
4396         const struct pmu *pmu;
4397         struct perf_event *event;
4398         struct hw_perf_event *hwc;
4399         long err;
4400
4401         event = kzalloc(sizeof(*event), gfpflags);
4402         if (!event)
4403                 return ERR_PTR(-ENOMEM);
4404
4405         /*
4406          * Single events are their own group leaders, with an
4407          * empty sibling list:
4408          */
4409         if (!group_leader)
4410                 group_leader = event;
4411
4412         mutex_init(&event->child_mutex);
4413         INIT_LIST_HEAD(&event->child_list);
4414
4415         INIT_LIST_HEAD(&event->group_entry);
4416         INIT_LIST_HEAD(&event->event_entry);
4417         INIT_LIST_HEAD(&event->sibling_list);
4418         init_waitqueue_head(&event->waitq);
4419
4420         mutex_init(&event->mmap_mutex);
4421
4422         event->cpu              = cpu;
4423         event->attr             = *attr;
4424         event->group_leader     = group_leader;
4425         event->pmu              = NULL;
4426         event->ctx              = ctx;
4427         event->oncpu            = -1;
4428
4429         event->parent           = parent_event;
4430
4431         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
4432         event->id               = atomic64_inc_return(&perf_event_id);
4433
4434         event->state            = PERF_EVENT_STATE_INACTIVE;
4435
4436         if (!overflow_handler && parent_event)
4437                 overflow_handler = parent_event->overflow_handler;
4438         
4439         event->overflow_handler = overflow_handler;
4440
4441         if (attr->disabled)
4442                 event->state = PERF_EVENT_STATE_OFF;
4443
4444         pmu = NULL;
4445
4446         hwc = &event->hw;
4447         hwc->sample_period = attr->sample_period;
4448         if (attr->freq && attr->sample_freq)
4449                 hwc->sample_period = 1;
4450         hwc->last_period = hwc->sample_period;
4451
4452         atomic64_set(&hwc->period_left, hwc->sample_period);
4453
4454         /*
4455          * we currently do not support PERF_FORMAT_GROUP on inherited events
4456          */
4457         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4458                 goto done;
4459
4460         switch (attr->type) {
4461         case PERF_TYPE_RAW:
4462         case PERF_TYPE_HARDWARE:
4463         case PERF_TYPE_HW_CACHE:
4464                 pmu = hw_perf_event_init(event);
4465                 break;
4466
4467         case PERF_TYPE_SOFTWARE:
4468                 pmu = sw_perf_event_init(event);
4469                 break;
4470
4471         case PERF_TYPE_TRACEPOINT:
4472                 pmu = tp_perf_event_init(event);
4473                 break;
4474
4475         case PERF_TYPE_BREAKPOINT:
4476                 pmu = bp_perf_event_init(event);
4477                 break;
4478
4479
4480         default:
4481                 break;
4482         }
4483 done:
4484         err = 0;
4485         if (!pmu)
4486                 err = -EINVAL;
4487         else if (IS_ERR(pmu))
4488                 err = PTR_ERR(pmu);
4489
4490         if (err) {
4491                 if (event->ns)
4492                         put_pid_ns(event->ns);
4493                 kfree(event);
4494                 return ERR_PTR(err);
4495         }
4496
4497         event->pmu = pmu;
4498
4499         if (!event->parent) {
4500                 atomic_inc(&nr_events);
4501                 if (event->attr.mmap)
4502                         atomic_inc(&nr_mmap_events);
4503                 if (event->attr.comm)
4504                         atomic_inc(&nr_comm_events);
4505                 if (event->attr.task)
4506                         atomic_inc(&nr_task_events);
4507         }
4508
4509         return event;
4510 }
4511
4512 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4513                           struct perf_event_attr *attr)
4514 {
4515         u32 size;
4516         int ret;
4517
4518         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4519                 return -EFAULT;
4520
4521         /*
4522          * zero the full structure, so that a short copy will be nice.
4523          */
4524         memset(attr, 0, sizeof(*attr));
4525
4526         ret = get_user(size, &uattr->size);
4527         if (ret)
4528                 return ret;
4529
4530         if (size > PAGE_SIZE)   /* silly large */
4531                 goto err_size;
4532
4533         if (!size)              /* abi compat */
4534                 size = PERF_ATTR_SIZE_VER0;
4535
4536         if (size < PERF_ATTR_SIZE_VER0)
4537                 goto err_size;
4538
4539         /*
4540          * If we're handed a bigger struct than we know of,
4541          * ensure all the unknown bits are 0 - i.e. new
4542          * user-space does not rely on any kernel feature
4543          * extensions we dont know about yet.
4544          */
4545         if (size > sizeof(*attr)) {
4546                 unsigned char __user *addr;
4547                 unsigned char __user *end;
4548                 unsigned char val;
4549
4550                 addr = (void __user *)uattr + sizeof(*attr);
4551                 end  = (void __user *)uattr + size;
4552
4553                 for (; addr < end; addr++) {
4554                         ret = get_user(val, addr);
4555                         if (ret)
4556                                 return ret;
4557                         if (val)
4558                                 goto err_size;
4559                 }
4560                 size = sizeof(*attr);
4561         }
4562
4563         ret = copy_from_user(attr, uattr, size);
4564         if (ret)
4565                 return -EFAULT;
4566
4567         /*
4568          * If the type exists, the corresponding creation will verify
4569          * the attr->config.
4570          */
4571         if (attr->type >= PERF_TYPE_MAX)
4572                 return -EINVAL;
4573
4574         if (attr->__reserved_1 || attr->__reserved_2)
4575                 return -EINVAL;
4576
4577         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4578                 return -EINVAL;
4579
4580         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4581                 return -EINVAL;
4582
4583 out:
4584         return ret;
4585
4586 err_size:
4587         put_user(sizeof(*attr), &uattr->size);
4588         ret = -E2BIG;
4589         goto out;
4590 }
4591
4592 static int perf_event_set_output(struct perf_event *event, int output_fd)
4593 {
4594         struct perf_event *output_event = NULL;
4595         struct file *output_file = NULL;
4596         struct perf_event *old_output;
4597         int fput_needed = 0;
4598         int ret = -EINVAL;
4599
4600         if (!output_fd)
4601                 goto set;
4602
4603         output_file = fget_light(output_fd, &fput_needed);
4604         if (!output_file)
4605                 return -EBADF;
4606
4607         if (output_file->f_op != &perf_fops)
4608                 goto out;
4609
4610         output_event = output_file->private_data;
4611
4612         /* Don't chain output fds */
4613         if (output_event->output)
4614                 goto out;
4615
4616         /* Don't set an output fd when we already have an output channel */
4617         if (event->data)
4618                 goto out;
4619
4620         atomic_long_inc(&output_file->f_count);
4621
4622 set:
4623         mutex_lock(&event->mmap_mutex);
4624         old_output = event->output;
4625         rcu_assign_pointer(event->output, output_event);
4626         mutex_unlock(&event->mmap_mutex);
4627
4628         if (old_output) {
4629                 /*
4630                  * we need to make sure no existing perf_output_*()
4631                  * is still referencing this event.
4632                  */
4633                 synchronize_rcu();
4634                 fput(old_output->filp);
4635         }
4636
4637         ret = 0;
4638 out:
4639         fput_light(output_file, fput_needed);
4640         return ret;
4641 }
4642
4643 /**
4644  * sys_perf_event_open - open a performance event, associate it to a task/cpu
4645  *
4646  * @attr_uptr:  event_id type attributes for monitoring/sampling
4647  * @pid:                target pid
4648  * @cpu:                target cpu
4649  * @group_fd:           group leader event fd
4650  */
4651 SYSCALL_DEFINE5(perf_event_open,
4652                 struct perf_event_attr __user *, attr_uptr,
4653                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4654 {
4655         struct perf_event *event, *group_leader;
4656         struct perf_event_attr attr;
4657         struct perf_event_context *ctx;
4658         struct file *event_file = NULL;
4659         struct file *group_file = NULL;
4660         int fput_needed = 0;
4661         int fput_needed2 = 0;
4662         int err;
4663
4664         /* for future expandability... */
4665         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4666                 return -EINVAL;
4667
4668         err = perf_copy_attr(attr_uptr, &attr);
4669         if (err)
4670                 return err;
4671
4672         if (!attr.exclude_kernel) {
4673                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4674                         return -EACCES;
4675         }
4676
4677         if (attr.freq) {
4678                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4679                         return -EINVAL;
4680         }
4681
4682         /*
4683          * Get the target context (task or percpu):
4684          */
4685         ctx = find_get_context(pid, cpu);
4686         if (IS_ERR(ctx))
4687                 return PTR_ERR(ctx);
4688
4689         /*
4690          * Look up the group leader (we will attach this event to it):
4691          */
4692         group_leader = NULL;
4693         if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4694                 err = -EINVAL;
4695                 group_file = fget_light(group_fd, &fput_needed);
4696                 if (!group_file)
4697                         goto err_put_context;
4698                 if (group_file->f_op != &perf_fops)
4699                         goto err_put_context;
4700
4701                 group_leader = group_file->private_data;
4702                 /*
4703                  * Do not allow a recursive hierarchy (this new sibling
4704                  * becoming part of another group-sibling):
4705                  */
4706                 if (group_leader->group_leader != group_leader)
4707                         goto err_put_context;
4708                 /*
4709                  * Do not allow to attach to a group in a different
4710                  * task or CPU context:
4711                  */
4712                 if (group_leader->ctx != ctx)
4713                         goto err_put_context;
4714                 /*
4715                  * Only a group leader can be exclusive or pinned
4716                  */
4717                 if (attr.exclusive || attr.pinned)
4718                         goto err_put_context;
4719         }
4720
4721         event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4722                                      NULL, NULL, GFP_KERNEL);
4723         err = PTR_ERR(event);
4724         if (IS_ERR(event))
4725                 goto err_put_context;
4726
4727         err = anon_inode_getfd("[perf_event]", &perf_fops, event, O_RDWR);
4728         if (err < 0)
4729                 goto err_free_put_context;
4730
4731         event_file = fget_light(err, &fput_needed2);
4732         if (!event_file)
4733                 goto err_free_put_context;
4734
4735         if (flags & PERF_FLAG_FD_OUTPUT) {
4736                 err = perf_event_set_output(event, group_fd);
4737                 if (err)
4738                         goto err_fput_free_put_context;
4739         }
4740
4741         event->filp = event_file;
4742         WARN_ON_ONCE(ctx->parent_ctx);
4743         mutex_lock(&ctx->mutex);
4744         perf_install_in_context(ctx, event, cpu);
4745         ++ctx->generation;
4746         mutex_unlock(&ctx->mutex);
4747
4748         event->owner = current;
4749         get_task_struct(current);
4750         mutex_lock(&current->perf_event_mutex);
4751         list_add_tail(&event->owner_entry, &current->perf_event_list);
4752         mutex_unlock(&current->perf_event_mutex);
4753
4754 err_fput_free_put_context:
4755         fput_light(event_file, fput_needed2);
4756
4757 err_free_put_context:
4758         if (err < 0)
4759                 kfree(event);
4760
4761 err_put_context:
4762         if (err < 0)
4763                 put_ctx(ctx);
4764
4765         fput_light(group_file, fput_needed);
4766
4767         return err;
4768 }
4769
4770 /**
4771  * perf_event_create_kernel_counter
4772  *
4773  * @attr: attributes of the counter to create
4774  * @cpu: cpu in which the counter is bound
4775  * @pid: task to profile
4776  */
4777 struct perf_event *
4778 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4779                                  pid_t pid,
4780                                  perf_overflow_handler_t overflow_handler)
4781 {
4782         struct perf_event *event;
4783         struct perf_event_context *ctx;
4784         int err;
4785
4786         /*
4787          * Get the target context (task or percpu):
4788          */
4789
4790         ctx = find_get_context(pid, cpu);
4791         if (IS_ERR(ctx)) {
4792                 err = PTR_ERR(ctx);
4793                 goto err_exit;
4794         }
4795
4796         event = perf_event_alloc(attr, cpu, ctx, NULL,
4797                                  NULL, overflow_handler, GFP_KERNEL);
4798         if (IS_ERR(event)) {
4799                 err = PTR_ERR(event);
4800                 goto err_put_context;
4801         }
4802
4803         event->filp = NULL;
4804         WARN_ON_ONCE(ctx->parent_ctx);
4805         mutex_lock(&ctx->mutex);
4806         perf_install_in_context(ctx, event, cpu);
4807         ++ctx->generation;
4808         mutex_unlock(&ctx->mutex);
4809
4810         event->owner = current;
4811         get_task_struct(current);
4812         mutex_lock(&current->perf_event_mutex);
4813         list_add_tail(&event->owner_entry, &current->perf_event_list);
4814         mutex_unlock(&current->perf_event_mutex);
4815
4816         return event;
4817
4818  err_put_context:
4819         put_ctx(ctx);
4820  err_exit:
4821         return ERR_PTR(err);
4822 }
4823 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4824
4825 /*
4826  * inherit a event from parent task to child task:
4827  */
4828 static struct perf_event *
4829 inherit_event(struct perf_event *parent_event,
4830               struct task_struct *parent,
4831               struct perf_event_context *parent_ctx,
4832               struct task_struct *child,
4833               struct perf_event *group_leader,
4834               struct perf_event_context *child_ctx)
4835 {
4836         struct perf_event *child_event;
4837
4838         /*
4839          * Instead of creating recursive hierarchies of events,
4840          * we link inherited events back to the original parent,
4841          * which has a filp for sure, which we use as the reference
4842          * count:
4843          */
4844         if (parent_event->parent)
4845                 parent_event = parent_event->parent;
4846
4847         child_event = perf_event_alloc(&parent_event->attr,
4848                                            parent_event->cpu, child_ctx,
4849                                            group_leader, parent_event,
4850                                            NULL, GFP_KERNEL);
4851         if (IS_ERR(child_event))
4852                 return child_event;
4853         get_ctx(child_ctx);
4854
4855         /*
4856          * Make the child state follow the state of the parent event,
4857          * not its attr.disabled bit.  We hold the parent's mutex,
4858          * so we won't race with perf_event_{en, dis}able_family.
4859          */
4860         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
4861                 child_event->state = PERF_EVENT_STATE_INACTIVE;
4862         else
4863                 child_event->state = PERF_EVENT_STATE_OFF;
4864
4865         if (parent_event->attr.freq)
4866                 child_event->hw.sample_period = parent_event->hw.sample_period;
4867
4868         child_event->overflow_handler = parent_event->overflow_handler;
4869
4870         /*
4871          * Link it up in the child's context:
4872          */
4873         add_event_to_ctx(child_event, child_ctx);
4874
4875         /*
4876          * Get a reference to the parent filp - we will fput it
4877          * when the child event exits. This is safe to do because
4878          * we are in the parent and we know that the filp still
4879          * exists and has a nonzero count:
4880          */
4881         atomic_long_inc(&parent_event->filp->f_count);
4882
4883         /*
4884          * Link this into the parent event's child list
4885          */
4886         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4887         mutex_lock(&parent_event->child_mutex);
4888         list_add_tail(&child_event->child_list, &parent_event->child_list);
4889         mutex_unlock(&parent_event->child_mutex);
4890
4891         return child_event;
4892 }
4893
4894 static int inherit_group(struct perf_event *parent_event,
4895               struct task_struct *parent,
4896               struct perf_event_context *parent_ctx,
4897               struct task_struct *child,
4898               struct perf_event_context *child_ctx)
4899 {
4900         struct perf_event *leader;
4901         struct perf_event *sub;
4902         struct perf_event *child_ctr;
4903
4904         leader = inherit_event(parent_event, parent, parent_ctx,
4905                                  child, NULL, child_ctx);
4906         if (IS_ERR(leader))
4907                 return PTR_ERR(leader);
4908         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
4909                 child_ctr = inherit_event(sub, parent, parent_ctx,
4910                                             child, leader, child_ctx);
4911                 if (IS_ERR(child_ctr))
4912                         return PTR_ERR(child_ctr);
4913         }
4914         return 0;
4915 }
4916
4917 static void sync_child_event(struct perf_event *child_event,
4918                                struct task_struct *child)
4919 {
4920         struct perf_event *parent_event = child_event->parent;
4921         u64 child_val;
4922
4923         if (child_event->attr.inherit_stat)
4924                 perf_event_read_event(child_event, child);
4925
4926         child_val = atomic64_read(&child_event->count);
4927
4928         /*
4929          * Add back the child's count to the parent's count:
4930          */
4931         atomic64_add(child_val, &parent_event->count);
4932         atomic64_add(child_event->total_time_enabled,
4933                      &parent_event->child_total_time_enabled);
4934         atomic64_add(child_event->total_time_running,
4935                      &parent_event->child_total_time_running);
4936
4937         /*
4938          * Remove this event from the parent's list
4939          */
4940         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4941         mutex_lock(&parent_event->child_mutex);
4942         list_del_init(&child_event->child_list);
4943         mutex_unlock(&parent_event->child_mutex);
4944
4945         /*
4946          * Release the parent event, if this was the last
4947          * reference to it.
4948          */
4949         fput(parent_event->filp);
4950 }
4951
4952 static void
4953 __perf_event_exit_task(struct perf_event *child_event,
4954                          struct perf_event_context *child_ctx,
4955                          struct task_struct *child)
4956 {
4957         struct perf_event *parent_event;
4958
4959         perf_event_remove_from_context(child_event);
4960
4961         parent_event = child_event->parent;
4962         /*
4963          * It can happen that parent exits first, and has events
4964          * that are still around due to the child reference. These
4965          * events need to be zapped - but otherwise linger.
4966          */
4967         if (parent_event) {
4968                 sync_child_event(child_event, child);
4969                 free_event(child_event);
4970         }
4971 }
4972
4973 /*
4974  * When a child task exits, feed back event values to parent events.
4975  */
4976 void perf_event_exit_task(struct task_struct *child)
4977 {
4978         struct perf_event *child_event, *tmp;
4979         struct perf_event_context *child_ctx;
4980         unsigned long flags;
4981
4982         if (likely(!child->perf_event_ctxp)) {
4983                 perf_event_task(child, NULL, 0);
4984                 return;
4985         }
4986
4987         local_irq_save(flags);
4988         /*
4989          * We can't reschedule here because interrupts are disabled,
4990          * and either child is current or it is a task that can't be
4991          * scheduled, so we are now safe from rescheduling changing
4992          * our context.
4993          */
4994         child_ctx = child->perf_event_ctxp;
4995         __perf_event_task_sched_out(child_ctx);
4996
4997         /*
4998          * Take the context lock here so that if find_get_context is
4999          * reading child->perf_event_ctxp, we wait until it has
5000          * incremented the context's refcount before we do put_ctx below.
5001          */
5002         raw_spin_lock(&child_ctx->lock);
5003         child->perf_event_ctxp = NULL;
5004         /*
5005          * If this context is a clone; unclone it so it can't get
5006          * swapped to another process while we're removing all
5007          * the events from it.
5008          */
5009         unclone_ctx(child_ctx);
5010         update_context_time(child_ctx);
5011         raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5012
5013         /*
5014          * Report the task dead after unscheduling the events so that we
5015          * won't get any samples after PERF_RECORD_EXIT. We can however still
5016          * get a few PERF_RECORD_READ events.
5017          */
5018         perf_event_task(child, child_ctx, 0);
5019
5020         /*
5021          * We can recurse on the same lock type through:
5022          *
5023          *   __perf_event_exit_task()
5024          *     sync_child_event()
5025          *       fput(parent_event->filp)
5026          *         perf_release()
5027          *           mutex_lock(&ctx->mutex)
5028          *
5029          * But since its the parent context it won't be the same instance.
5030          */
5031         mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
5032
5033 again:
5034         list_for_each_entry_safe(child_event, tmp, &child_ctx->group_list,
5035                                  group_entry)
5036                 __perf_event_exit_task(child_event, child_ctx, child);
5037
5038         /*
5039          * If the last event was a group event, it will have appended all
5040          * its siblings to the list, but we obtained 'tmp' before that which
5041          * will still point to the list head terminating the iteration.
5042          */
5043         if (!list_empty(&child_ctx->group_list))
5044                 goto again;
5045
5046         mutex_unlock(&child_ctx->mutex);
5047
5048         put_ctx(child_ctx);
5049 }
5050
5051 /*
5052  * free an unexposed, unused context as created by inheritance by
5053  * init_task below, used by fork() in case of fail.
5054  */
5055 void perf_event_free_task(struct task_struct *task)
5056 {
5057         struct perf_event_context *ctx = task->perf_event_ctxp;
5058         struct perf_event *event, *tmp;
5059
5060         if (!ctx)
5061                 return;
5062
5063         mutex_lock(&ctx->mutex);
5064 again:
5065         list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry) {
5066                 struct perf_event *parent = event->parent;
5067
5068                 if (WARN_ON_ONCE(!parent))
5069                         continue;
5070
5071                 mutex_lock(&parent->child_mutex);
5072                 list_del_init(&event->child_list);
5073                 mutex_unlock(&parent->child_mutex);
5074
5075                 fput(parent->filp);
5076
5077                 list_del_event(event, ctx);
5078                 free_event(event);
5079         }
5080
5081         if (!list_empty(&ctx->group_list))
5082                 goto again;
5083
5084         mutex_unlock(&ctx->mutex);
5085
5086         put_ctx(ctx);
5087 }
5088
5089 /*
5090  * Initialize the perf_event context in task_struct
5091  */
5092 int perf_event_init_task(struct task_struct *child)
5093 {
5094         struct perf_event_context *child_ctx = NULL, *parent_ctx;
5095         struct perf_event_context *cloned_ctx;
5096         struct perf_event *event;
5097         struct task_struct *parent = current;
5098         int inherited_all = 1;
5099         int ret = 0;
5100
5101         child->perf_event_ctxp = NULL;
5102
5103         mutex_init(&child->perf_event_mutex);
5104         INIT_LIST_HEAD(&child->perf_event_list);
5105
5106         if (likely(!parent->perf_event_ctxp))
5107                 return 0;
5108
5109         /*
5110          * If the parent's context is a clone, pin it so it won't get
5111          * swapped under us.
5112          */
5113         parent_ctx = perf_pin_task_context(parent);
5114
5115         /*
5116          * No need to check if parent_ctx != NULL here; since we saw
5117          * it non-NULL earlier, the only reason for it to become NULL
5118          * is if we exit, and since we're currently in the middle of
5119          * a fork we can't be exiting at the same time.
5120          */
5121
5122         /*
5123          * Lock the parent list. No need to lock the child - not PID
5124          * hashed yet and not running, so nobody can access it.
5125          */
5126         mutex_lock(&parent_ctx->mutex);
5127
5128         /*
5129          * We dont have to disable NMIs - we are only looking at
5130          * the list, not manipulating it:
5131          */
5132         list_for_each_entry(event, &parent_ctx->group_list, group_entry) {
5133
5134                 if (!event->attr.inherit) {
5135                         inherited_all = 0;
5136                         continue;
5137                 }
5138
5139                 if (!child->perf_event_ctxp) {
5140                         /*
5141                          * This is executed from the parent task context, so
5142                          * inherit events that have been marked for cloning.
5143                          * First allocate and initialize a context for the
5144                          * child.
5145                          */
5146
5147                         child_ctx = kzalloc(sizeof(struct perf_event_context),
5148                                             GFP_KERNEL);
5149                         if (!child_ctx) {
5150                                 ret = -ENOMEM;
5151                                 break;
5152                         }
5153
5154                         __perf_event_init_context(child_ctx, child);
5155                         child->perf_event_ctxp = child_ctx;
5156                         get_task_struct(child);
5157                 }
5158
5159                 ret = inherit_group(event, parent, parent_ctx,
5160                                              child, child_ctx);
5161                 if (ret) {
5162                         inherited_all = 0;
5163                         break;
5164                 }
5165         }
5166
5167         if (child_ctx && inherited_all) {
5168                 /*
5169                  * Mark the child context as a clone of the parent
5170                  * context, or of whatever the parent is a clone of.
5171                  * Note that if the parent is a clone, it could get
5172                  * uncloned at any point, but that doesn't matter
5173                  * because the list of events and the generation
5174                  * count can't have changed since we took the mutex.
5175                  */
5176                 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5177                 if (cloned_ctx) {
5178                         child_ctx->parent_ctx = cloned_ctx;
5179                         child_ctx->parent_gen = parent_ctx->parent_gen;
5180                 } else {
5181                         child_ctx->parent_ctx = parent_ctx;
5182                         child_ctx->parent_gen = parent_ctx->generation;
5183                 }
5184                 get_ctx(child_ctx->parent_ctx);
5185         }
5186
5187         mutex_unlock(&parent_ctx->mutex);
5188
5189         perf_unpin_context(parent_ctx);
5190
5191         return ret;
5192 }
5193
5194 static void __cpuinit perf_event_init_cpu(int cpu)
5195 {
5196         struct perf_cpu_context *cpuctx;
5197
5198         cpuctx = &per_cpu(perf_cpu_context, cpu);
5199         __perf_event_init_context(&cpuctx->ctx, NULL);
5200
5201         spin_lock(&perf_resource_lock);
5202         cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5203         spin_unlock(&perf_resource_lock);
5204
5205         hw_perf_event_setup(cpu);
5206 }
5207
5208 #ifdef CONFIG_HOTPLUG_CPU
5209 static void __perf_event_exit_cpu(void *info)
5210 {
5211         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5212         struct perf_event_context *ctx = &cpuctx->ctx;
5213         struct perf_event *event, *tmp;
5214
5215         list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry)
5216                 __perf_event_remove_from_context(event);
5217 }
5218 static void perf_event_exit_cpu(int cpu)
5219 {
5220         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5221         struct perf_event_context *ctx = &cpuctx->ctx;
5222
5223         mutex_lock(&ctx->mutex);
5224         smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5225         mutex_unlock(&ctx->mutex);
5226 }
5227 #else
5228 static inline void perf_event_exit_cpu(int cpu) { }
5229 #endif
5230
5231 static int __cpuinit
5232 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5233 {
5234         unsigned int cpu = (long)hcpu;
5235
5236         switch (action) {
5237
5238         case CPU_UP_PREPARE:
5239         case CPU_UP_PREPARE_FROZEN:
5240                 perf_event_init_cpu(cpu);
5241                 break;
5242
5243         case CPU_ONLINE:
5244         case CPU_ONLINE_FROZEN:
5245                 hw_perf_event_setup_online(cpu);
5246                 break;
5247
5248         case CPU_DOWN_PREPARE:
5249         case CPU_DOWN_PREPARE_FROZEN:
5250                 perf_event_exit_cpu(cpu);
5251                 break;
5252
5253         default:
5254                 break;
5255         }
5256
5257         return NOTIFY_OK;
5258 }
5259
5260 /*
5261  * This has to have a higher priority than migration_notifier in sched.c.
5262  */
5263 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5264         .notifier_call          = perf_cpu_notify,
5265         .priority               = 20,
5266 };
5267
5268 void __init perf_event_init(void)
5269 {
5270         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5271                         (void *)(long)smp_processor_id());
5272         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5273                         (void *)(long)smp_processor_id());
5274         register_cpu_notifier(&perf_cpu_nb);
5275 }
5276
5277 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
5278 {
5279         return sprintf(buf, "%d\n", perf_reserved_percpu);
5280 }
5281
5282 static ssize_t
5283 perf_set_reserve_percpu(struct sysdev_class *class,
5284                         const char *buf,
5285                         size_t count)
5286 {
5287         struct perf_cpu_context *cpuctx;
5288         unsigned long val;
5289         int err, cpu, mpt;
5290
5291         err = strict_strtoul(buf, 10, &val);
5292         if (err)
5293                 return err;
5294         if (val > perf_max_events)
5295                 return -EINVAL;
5296
5297         spin_lock(&perf_resource_lock);
5298         perf_reserved_percpu = val;
5299         for_each_online_cpu(cpu) {
5300                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5301                 raw_spin_lock_irq(&cpuctx->ctx.lock);
5302                 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5303                           perf_max_events - perf_reserved_percpu);
5304                 cpuctx->max_pertask = mpt;
5305                 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5306         }
5307         spin_unlock(&perf_resource_lock);
5308
5309         return count;
5310 }
5311
5312 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
5313 {
5314         return sprintf(buf, "%d\n", perf_overcommit);
5315 }
5316
5317 static ssize_t
5318 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
5319 {
5320         unsigned long val;
5321         int err;
5322
5323         err = strict_strtoul(buf, 10, &val);
5324         if (err)
5325                 return err;
5326         if (val > 1)
5327                 return -EINVAL;
5328
5329         spin_lock(&perf_resource_lock);
5330         perf_overcommit = val;
5331         spin_unlock(&perf_resource_lock);
5332
5333         return count;
5334 }
5335
5336 static SYSDEV_CLASS_ATTR(
5337                                 reserve_percpu,
5338                                 0644,
5339                                 perf_show_reserve_percpu,
5340                                 perf_set_reserve_percpu
5341                         );
5342
5343 static SYSDEV_CLASS_ATTR(
5344                                 overcommit,
5345                                 0644,
5346                                 perf_show_overcommit,
5347                                 perf_set_overcommit
5348                         );
5349
5350 static struct attribute *perfclass_attrs[] = {
5351         &attr_reserve_percpu.attr,
5352         &attr_overcommit.attr,
5353         NULL
5354 };
5355
5356 static struct attribute_group perfclass_attr_group = {
5357         .attrs                  = perfclass_attrs,
5358         .name                   = "perf_events",
5359 };
5360
5361 static int __init perf_event_sysfs_init(void)
5362 {
5363         return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5364                                   &perfclass_attr_group);
5365 }
5366 device_initcall(perf_event_sysfs_init);